Intelligent quality assurance system based on deep learning and statistical analysis

Abstract

An explainable deep learning modeling method according to an embodiment of the present invention for achieving the above objective comprises: a step of designing a deep learning workflow with said algorithm when an algorithm suitable for a region of interest is selected among previously stored algorithms; and a step of performing a simulation of the deep learning workflow model on the data when data is input.

Description

Technology Field

[0001] The present invention relates to a method and system for determining quality assurance using learned deep learning and statistical analysis. Background Technology

[0002] The development of artificial intelligence algorithms is conducted in a somewhat closed manner based on individual capabilities using languages ​​such as R and Python, and performance is evaluated and verified according to individual standards. Consequently, there are numerous instances where AI requires continuous human management and control to maintain expected performance and operate stably, which presents limitations to unmanned intelligence.

[0003] Furthermore, the technology development process for existing AI algorithms has followed a somewhat personalized pattern in which analysts design, develop, and verify algorithms in their personal development environments to meet unit requirements. Additionally, the processes of understanding performance changes depending on the environment, as well as improvement, deployment, and disposal, have been carried out manually, which presents a disadvantage in that improving and sharing the technology is difficult.

[0004] In addition, while open-source-based deep learning algorithms have recently become a trend and numerous algorithms are being developed, in order to apply them to the real world, the suitability of the AI ​​judgment must be continuously verified directly to determine whether to distribute them.

[0005] However, due to the large number of AI algorithms across different domains and requirements, and the difficulty of individually verifying those algorithms that require periodic remodeling to reflect changing manufacturing data, there are instances where optimization is abandoned. In other words, the failure to manage algorithm changes leads to frequent performance faults. The problem to be solved

[0006] The present invention relates to a method for ensuring quality related to microorganisms, and the objective of the present invention is to provide a method for visualizing the connection of algorithms and verifying and simulating deep learning analysis results through integrated simulation to compare mutual performance. means of solving the problem

[0007] An explainable deep learning modeling method according to an embodiment of the present invention for achieving the above objective comprises: a step of designing a deep learning workflow with said algorithm when an algorithm suitable for a region of interest is selected among previously stored algorithms; and a step of performing a simulation of the deep learning workflow model on the data when data is input. Effects of the invention

[0008] As described above, according to the embodiments of the present invention, the connection of algorithms can be visualized and performance can be automatically verified through integrated simulation, and multiple deep learning algorithms can be simulated simultaneously to compare their performance. Brief explanation of the drawing

[0009] FIG. 1 is a flowchart of an analysis quality assurance method according to an embodiment of the present disclosure. FIG. 2 is a block diagram of a system providing an analysis quality assurance service using deep learning according to one embodiment. Figure 3 is a diagram for explaining the operation of a computer program illustrated in Figure 2. Specific details for implementing the invention

[0010] FIG. 1 is a flowchart provided for explaining an explainable artificial intelligence modeling and simulation method according to one embodiment of the present invention.

[0011] An explainable artificial intelligence modeling and simulation method according to one embodiment of the present invention allows an artificial intelligence modeling and simulation system to design an artificial intelligence workflow model (S120) by assembling the selected algorithms when algorithms suitable for a workflow written / edited to fit a subject area are selected from among the algorithms stored in advance (S110).

[0012] At this time, the algorithms may be multiple deep learning algorithms or machine learning algorithms.

[0013] And when multiple AI workflow models are designed, the AI ​​modeling and simulation system allows for the registration of simulations for each AI model (S130), and when input information is entered, it can perform each registered simulation based on the input information (S140).

[0014] In addition, when each simulation is performed simultaneously under the same conditions, the artificial intelligence modeling and simulation system continuously compares and evaluates the performance of each artificial intelligence workflow model designed based on the results of each simulation (S150), and continuously deploys or changes the optimal artificial intelligence workflow model according to the results of the comparison evaluation so that it can be applied to the process.

[0015] An explainable artificial intelligence modeling and simulation system according to one embodiment of the present invention relates to an artificial intelligence operation simulation platform technology that automates the development-verification-distribution-operation process of an artificial intelligence algorithm and increases operational efficiency. It can support key processes such as modeling of learning classification algorithms, algorithm asset management, execution of parallel simulations and automatic algorithm performance verification, and verification of the execution results of deployed algorithms using image and shape data, which are essential for the unmanned / automated operation of various types of inspection processes, including the interpretation and classification of visual information.

[0016] In addition, the explainable artificial intelligence modeling and simulation system according to one embodiment of the present invention is a technology for establishing a management system for the development, operation, verification, and distribution of artificial intelligence algorithms, which will continuously increase as the number of cases promoting the advancement of AI-based analysis increases, and for explaining the suitability of developed artificial intelligence algorithms with data. It enables central management that visualizes the connection of algorithms based on a workflow, automatically verifies performance through integrated simulation, and compares the performance of multiple deep learning algorithms and machine learning algorithms simultaneously under the same conditions based on the workflow. Based on the results of simulating deep learning algorithms and machine learning algorithms simultaneously under the same conditions, it selects a deep learning algorithm with optimal accuracy and selects / applies a machine learning algorithm that has the same result value as the selected deep learning algorithm to implement explainable artificial intelligence. Furthermore, it visualizes the connection of algorithms based on a workflow and automatically verifies performance through integrated simulation, and the results of classification analysis can explain the cause / reason using images and standardized feature data.

[0017] To perform these functions, an explainable artificial intelligence modeling and simulation system according to one embodiment of the present invention can design an artificial intelligence workflow model by first selecting algorithms suitable for a workflow written / edited to fit a subject area from among the stored algorithms, and assembling the selected algorithms.

[0018] For the image normalization process, any one of the Load Object, Normalize, Rotate Image, and Working Domain algorithms can be selected.

[0019] For the learning process, either GMM Learning or NCC Learning can be selected.

[0020] For image feature analysis, any one of the following algorithms may be selected: Find Defect, Validate Defect, GMM Classification, NCC Classification, Monitor Defect, and Extract Feature.

[0021] Artificial intelligence learning / classification algorithms include algorithms for data preprocessing, algorithms for learning / classification, algorithms for data postprocessing, and algorithms for iterative learning / classification.

[0022] Referring to FIG. 2, a common controller (110) that performs interfacing (or communication) with an external device receives a data-based analysis result and divides each of the data packets included in the analysis result and transmits them to a preprocessor (122) packet by packet.

[0023] A REST API interface (120) that performs the function of a handler includes a preprocessor (122) and a postprocessor (124).

[0024] The preprocessor (122) receives each analysis result included in the packet from the common controller (110). The preprocessor (122) converts the parameters constituting each data included in the packet into vector values ​​using a word embedding that has already been learned, in order to use them as input values ​​for the artificial neural network (130). The preprocessor (122) vectorizes each parameter by parameter type and outputs the parameters indexed by type as result values.

[0025] The dimension (IVD) of the input vector values ​​of the TCR (topic classification recognition) model is calculated according to Equation 1.

[0026] [Mathematical Formula 1]

[0027] IVD=NP*MN*DS

[0028] Here, NP is the number of data inputs at once, MN is the maximum number of parameters included in a single data, and DS is the dimension of the embedding of a parameter.

[0029] Data vectorized by the preprocessor (122) is input into each layer (142, 144, and 150). The TCR first inappropriate layer 1 (142) is a model that checks for inappropriate parameters included in each data. The TCR first inappropriate layer 2 (144) is a model that verifies the value output as an inappropriate representation by the TCR first inappropriate layer 1 (142).

[0030] The NER (Named Entity Recognition; 170) model is a named entity recognition model that classifies which category a specific parameter included in a single piece of data has based on machine learning. The NER (170) model further examines the data classified by the TCR section layer (150) by classifying it into categories by parameter.

[0031] The post processor (124) receives the output of the TCR second inappropriate layer (160) and the output of the NER model (170), and checks whether there is data containing inappropriate result values ​​for each parameter based on a unique key value and whether there are missing parameters that must be provided.

[0032] Figure 3 is a diagram for explaining the operation of a computer program illustrated in Figure 2.

[0033] The common controller (110) of FIG. 2 corresponds to the interface (245) of FIG. 3, the preprocessor (122) of FIG. 2 performs the functions of the modules (310, 320, and 330), and the postprocessor (124) of FIG. 2 performs the functions of the post-processing module (340) of FIG. 3.

[0034] Drawing number (130) of FIG. 2 is a part that more specifically represents the essential keyword recognition module (332) and the inappropriate parameter recognition module (334) of FIG. 3, and the essential keyword recognition module (332) of FIG. 3 can perform the functions of the configurations (150 and 170) of FIG. 2, and the inappropriate expression recognition module (334) of FIG. 3 can perform the functions of the configurations (140 and 160) of FIG. 1.

[0035] Each TCR (Topic Classification Recognize; 142, 144, 162, 164, and 166) in Fig. 2 is a deep learning model that distinguishes what type of parameter an input parameter is.

Claims

Claim 1 In an intelligent quality assurance method based on deep learning and statistical analysis, the explainable deep learning modeling method comprises the steps of: designing a deep learning workflow with said algorithm when an algorithm suitable for a region of interest is selected among previously stored algorithms; and performing a simulation of the deep learning workflow model on the data when data is input.

Images