# Industrial Problem Solution

Total Posts 25
25
1. Company introduction Ensearch LAB operates a service that provides direct or processed data produced and collected by various instruments (water level and water quality, etc.) and solar inverters. 2. Problem Background and Summary Photovoltaic power generation is a light-based generation method that affects power generation efficiency according to the amount of light • Conditions should be provided for the installation site in order to increase the efficiency of solar power generation • We need a mathematical model that can provide altitude and azimuth angles for various conditions such as latitude, longitude, and time, so we ask for questions.   3. Solving Process • Develop an algorithm that calculates the position of the sun by entering latitude and longitude and date and time, resulting in the azimuth and inclination of the optimal solar module. • Comparison of power generation efficiency between regional fixed installation angles and variable installation depending on the position of the sun.   4. Ripple effects and future plans • We will conduct modeling to reduce the margin of error by comparing and analyzing the initial estimated data with the actual developed data through R&D or demonstration. • The altitude and azimuth calculation algorithms will be implemented as a program (calculator) to provide premium services to customers using our monitoring services.

24
1. Company introduction DAP Co., Ltd., a company specializing in artificial intelligence solutions based in Uiwang-si, Gyeonggi-do, founded in 2017. The goal is to solve public problems by reducing fine dust. There are smart air quality management system for underground stations, development of artificial intelligence measurement technology for fine dust/dissipation dust, and development of air purification system in the business area   2. Problem Background and Summary In order to improve the deterioration of air quality in subway stations and inconvenience to subway users, DAP Co., Ltd. installed fine dust measuring devices in station to analyze the observed fine dust concentration data and to create an artificial intelligence model that automatically controls air conditioning facilities and fine dust reduction devices. To predict the concentration of fine dust in subway stations in an hour. Numericalize the actual effect of fine dust reduction devices installed in station   3. Solving Process Present two methods of prediction using ARIMA model, a traditional method of analyzing time series data, to predict fine dust concentration, and by converting time series data into two-dimensional image data, using synthetic triple neural network (CNN) To identify the effects of the reduction device, a comparative analysis with the concentration of fine dust in the atmosphere suggests a method to statistically quantify the actual effects of the reduction device.   4. Ripple effects and future plans As a future task, we suggest ways to increase the accuracy of fine dust concentration levels by directly using the data inside the fine dust meter and introducing a new machine learning method that uses both the fine dust concentration data and the external atmospheric fine dust concentration data observed by the measuring instrument.

23
1. Company introduction SmartSocial Co., Ltd. was founded in 2012 and is taking the lead in solving social problems by utilizing big data. Major business areas include resolution of job mismatching, public works, youth employee tomorrow deduction, small business exploration project, youth tomorrow morning deduction operation support, student employment management, etc.   2. Problem Background and Summary Request to resolve the issue of job recommendation using the National Competency Standards (NCS) The goal is to provide an objective solution by utilizing national competency standards rather than recommendations based on subjective experience.      3. Solving Process Methods for collecting and refining NCS data, traditional machine learning based on statistics, as well as deep learning to present matching algorithms suitable for web application environments It implements matching algorithms directly to be mounted directly on the system operated by SmartSocial Co., Ltd. and provides solutions in the form of middleware web servers.   4. Ripple effects and future plans Present engineering technology research required to implement algorithm with introduction of new machine learning as a future task It also suggests ways to find and utilize data linked to NCS data.

22
1. Company introduction DN Co., Ltd.is a company that develops ophthalmic diagnostic medical equipment and software, and is currently conducting research aiming to develop a strabismus automatic diagnostic program.   2. Problem Background and Summary The number of patients undergoing strabismus tends to increase gradually. In particular, strabismus was found in about 2% of children. The strabismus specialist who diagnoses strabismus is only 5-10% of all ophthalmologists. Request the development of an algorithm that automatically determines strabismus without the help of a specialist 3. Solving Process It is difficult to distinguish strabismus and extra when implementing the algorithm as it is, referring to the related paper. After normal translation of the 2D data extracted from the eye-tracking device, normal and strabismus are classified based on the Pearson correlation coefficient. Use features obtained in various ways to apply algorithms such as random forest to classify normal and squint   4. Ripple effects and future plans It is better to present the actual strabismus diagnostic method directly for the development of strabismus diagnostic software. After collecting enough data, it is better to apply deep running method such as multilayer perceptron or convolution neural network. Expected to obtain a model of performance The provided algorithm will be installed in software manufactured by the current company in the future, and it will be used in client applications actually used by ophthalmologists and pediatricians.

21
1. Company introduction Korea Trade Information Communication is a company that provides trade automation services through the Internet and establishes electronic trade infrastructure for the automation of complex import and export processes. 2. Problem Background and Summary Advanced algorithms to match product item names to numerical codes using textual information about products related to trade Ask to explore mathematical methods of expressing item name words in numerical vectors and develop a deep learning model for text classification 3. Solving Process Explain the process of tokenization, word embedding algorithms, which are necessary steps in quantifying item names, and deliver comparative analysis of the data. Using embedded vectors of item names, various deep learning model networks matching to numerical codes were designed to calculate model accuracy. 4. Ripple effects and future plans More than 80% accuracy was obtained in the prediction of top3 or top5 rather than single numerical code classification. It will be used as a result to determine the feasibility of the project whether the trade-related product names can be matched to the number code. It is expected to improve results by comparing and analyzing existing results by applying the recent Google-developed Bidirectional Encoder Presentations from Transformer (BERT).

20
1. Company introduction Seoul Metro, which used to operate subway lines 1 to 4, and Seoul Metropolitan Rapid Transit Corporation, which operated lines 5 to 8, were established by integrating them.   2. Problem Background and Summary Developing and upgrading algorithms to monitor the abnormal conditions and major component conditions of air conditioning installations installed and operated in each station for passenger safety Identify the data characteristics that can extract the characteristics of failure and part condition Develops a method of dividing sensing data collected in real time over 24 hours into point-of-action units. How to efficiently apply algorithms developed in more than 7,000 different standards of air conditioners and analyze data analysis results   3. Solving Process Develop data preprocessing methods, such as time series synchronization techniques, to divide current data of air conditioning motors into point of operation units. Comparative analysis of current data and vibration data provides a method of data analysis for the characteristics and type of failure of the equipment Developing a health monitoring algorithm that uses current data to predict and informs the air-conditioner V-belt -out detection model and when to replace key parts (V-belt, bearing) Development of a methodology that can integrate key component status monitoring algorithms into individual air conditioners Sharing the results of applying non-map clustering techniques to find replacement points for key components without additional replacement timing   4. Ripple effects and future plans Jointly developed integrated models are tested in various subway stations to detect component abnormalities with 95% accuracy on air-conditioner V-belt (7) and bearing (5). The test will be carried out by applying it to air conditioners of different station and specifications. The Seoul Metropolitan Transportation Corporation's Machinery Department plans to build a system that analyzes the operation data of the air conditioner to determine and inform the problem by installing a model jointly developed with the institute on the server. Expect increased air conditioning operation rate by timely informing of life expectancy or replacement timing for major components Mean Time to Repair Reduction Effect by enabling repair preparation by failure prediction alarm The stable operation of the subway ventilation system is expected to improve customer satisfaction by maintaining pleasant air quality.

19
1. Company introduction EHR&C Co., Ltd. has technology for ecological and chemical toxicity in the environmental field and risk assessment of products, and is a research firm that conducts R&D and academic research in the field. 2. Problem Background and Summary Explore appropriate mathematical methods to derive a distribution of collective exposure that reflects individual usage patterns (time, frequency of use, usage) and usage of individuals' products Learn the distribution of personal exposure data for the product and use it to generate data to create a usage pattern matrix Appropriate methods of filling missing values of exposure data by product for 3,000 people who responded to the survey and methods of interpreting generated data 3. Solving Process Provides a statistical review of the parametric estimation methods used by existing companies estimating the usage pattern survey response distribution for each product Using the deep learning algorithm Variational Autoencoder, it presents a nonparametric way to learn the distribution of exposure and to generate data Provide a deion of the distribution distances that can measure the different degrees of data distribution and share information on measuring the distance between the data generated and the existing data Suggest how to interpret the generated distribution and test its independence from other product exposure distributions 4. Ripple effects and future plans Generates an individual product usage pattern matrix by learning the distribution of data more easily than traditional parametric methods used by the enterprise. This matrix has been used immediately for the research the enterprise is conducting. Additional discussions can be made to think about the problem of exploring which matrix converges when a product usage pattern matrix is created in a bootstrap manner.

18
1. Company introduction Define Co., Ltd. is a company that develops and supplies system software, develops and supplies optimal collection systems for household waste, eradicates harmful tides, and smart parking systems based on deep learning. 2. Problem Background and Summary Deep Learning-based smart parking system, which is being piloted by Define Co., Ltd., provides services for entry and exit time, vehicle number, real-time parking surface usage, total number of parking spaces, etc. There is a problem with CCTV images that are being collected, depending on the degree of sunlight or the angle of the vehicle when entering and leaving the vehicle, the recognition rate of the license plate is reduced. High-accuracy algorithms are required to locate vehicle license plates in CCTV footage for improved service and corporate competitiveness of smart parking systems In addition, a system is required to extract vehicle numbers so that the parking lot can accurately recognize the time and number of vehicles entering and leaving the vehicle 3. Solving Process tensorflow object detection API model is compared and analyzed to locate the vehicle license plate. A new algorithm is needed to extract the vehicle number because we tried to extract the text using the OCR algorithm but could not see the performance as high. Extracting only the number plate part from the vehicle image using object detection will inevitably cause problems with poor image quality, so a super-resolution algorithm is required to improve image quality. 4. Ripple effects and future plans In order to locate the license plate of the vehicle, a fast and accurate model was selected and studied among the tensorflow object detection API models. The test images were selected by comparing the speed and accuracy of each model. Among algorithms that can extract vehicle numbers, Samsung Electronics presented algorithms that are suitable for image and text analysis. A super-resolution algorithm for improving image quality of vehicle license plate is presented. Object detection algorithm and super-resolution algorithm that can be applied for system advancement were proposed. Analysis based on continuous images rather than single image-based license plate analysis eliminated errors.

17
1. Company introduction IntelliCode Co., Ltd. is a system solution provider that uses log data from computer systems to detect anomalies. 2. Problem Background and Summary A study of methodology that enables humans to understand and interpret patterns found by machines in machine learning models. 3. Solving Process Certain methods, such as regression analysis and decision tree learning methods, have some potential for interpretation of the learning process or learning result itself due to the nature of the model. Provides a deion of the interpretability of the above model and examples of its application In case of artificial neural networks such as deep learning, the internal process of learning makes it difficult to obtain interpretability through the model itself and the learning process. Provides a deion and application example of the model-agnostic method of permutation feature import and local interoperable model-organization (LIME) among model-specific methods that are applicable regardless of model 4. Ripple effects and future plans Visualization and UX implementation will be used for practical use. A joint study on the analysis and implementation of the relevant latest paper, which continues to be published, will be conducted.

16
1. Company introduction As of 2019, it has supplied city gas to about 1.43 million households in Busan, and won the first place in the city gas category for 16 consecutive years from 2003 to 2018 with a focus on customer satisfaction management. 2. Problem Background and Summary Customer continuously complains about current meter-based methods, such as meter failure or billing errors, while gas usage is being inspected/charged every month. To solve the problem of determining meter error and abnormal generation through historical data analysis, and obtaining reliable estimates for non-detection generation by predicting future usage   3. Solving Process In the data preprocessing process for the development of predictive algorithms, the analysis was conducted by calculating the distance of the probability distribution, excluding those with large standard deviations or missing values. It is almost impossible to predict by creating individual models for each generation, and predicting all 12 months for the entire generation with one model is expected to be unreliable, developing a monthly model that takes time and temperature into account to predict future usage Developed abnormal data classification models through the application of rule-based algorithms, such as those with rapidly increasing usage, those with equal or less than ±10% usage for two years, those with less average usage and less usage, and those with higher usage overall.   4. Ripple effects and future plans It is expected that customers' complaints will be reduced by developing reliable usage prediction algorithms through analysis of historical data patterns for each generation, rather than an integrated usage prediction method such as the previous year-on-year method. Development of Logic using 1D CNN Deep Learning to select new estimated values. It has improved accuracy by 28% compared to current method of same month last year and obtained faster processing speed compared to other time series deep learning models such as LSTM. Rule-based algorithms that classify abnormal generations can easily be controlled by corporate officials, enabling them to classify according to the situation. Some of the abnormal data found during pre-processing of data will be improved after checking whether it is actually abnormal. The developed 1D CNN Deep Learning model / rule-based algorithm will be reflected in the business system later.