Integration of Principal Component Analysis and Integer Linear Programming for Complete Must-Link Constrained Clustering Problem (2011.10~2012,7)
Funded by National Science Council (currently, Ministry of Science and Technology) （NSC 100-2218-E-011-026）, PI
This research aims at developing an integrated method to solve a special but not uncommon constrained clustering problem constructed by Complete Must-Link (CML) constraints. Constrained clustering analysis is a semi-supervised learning to accommodate the information while it is available, to improve efficiency and purity of clustering. According to the domain knowledge or prior information, the CML constraints in which each of the data points is specified within a particular partition may be available in the dataset. For example, in defect detection analysis, the categorical feature such as the manufacturing batch number or quality measuring group can generate CML constraints. Due to similar manufacturing or measuring characteristics, the data points with the same batch number or measuring group might require to be grouped in the same clusters. In this research, CML constrained problem is defined as performing data clustering on data points with complete partition information to search for the interesting patterns. While performing the traditional clustering method such as a constrained hierarchical on a CML constrained problem, the information missing issue is caused by the centroid replacement process where all data points within a block are replaced by a centroid. This issue affects the clustering accuracy because the clustering only concerns on the centriod data without considering the characteristic of data blocks. In this research, a two-step clustering method is proposed to solve CML clustering problem. First, in order to alleviate the information missing issue, in the similarity matrix, the principal component eigenvectors which can represent the characteristics of the partition block is added. Based on the new similarity matrix, a set of candidate clusters can be generated. In the second step, an integer linear programming optimization is formulated to select the resulting clusters from the candidate clusters. The optimization can guarantee the resulting clusters to have minimum within-cluster distances.
A Framework of Clustering and Classification for Data Analysis on Multiple-component Prototype Testing (2012.8~2013,7)
Funded by National Science Council (currently, Ministry of Science and Technology)（NSC 101-2221-E-011-057）, PI
This research aims to combine two major data mining approaches: clustering and classification to analyze the prototype testing dataset. Due to the model assumption, traditional statistical methods have the limitation on analyzing the prototype testing dataset. Applying data mining techniques such as clustering and classification algorithms is promising to reveal and analyze the prototype testing dataset which usually has fewer prototype samples, complicated attributes, and frequent design change characteristics. In the proposed framework, the clustering was applied to analyze the numerical measures ( dataset). The clustering results, labels, are then combined with categorical feature dataset as the inputs of the classification model to classify the clustering labels by using dataset. In order to maintain the balanced performance of clustering and classification learning simultaneously, Clustering Classification Evaluation plot (CCE) plot was proposed to determine the number of clusters which is an important parameter affecting the performance of the propose framework. The result shows combining clustering and classification methods is able to analyze the prototype testing dataset without considering the model assumptions.
Data Analysis and Operation Innovation for Plant Factory (I) (II) (2013.8~2015,7)
Funded by National Science Council (currently, Ministry of Science and Technology)（NSC 102-2221-E-011-052，NSC 103-2221-E-011-116）, PI
A plant factory is a crop production facility in which all the environmental elements such as temperature, humidity, lighting, and nutrition for plant growth are artificially controlled. This project aims at improving the profitability of plant factory by proposing the 3-stage research plans to investigate the needed data analysis method, the adaptive scheduling, and the innovative rental service of plant factory. The first stage will collect 1) the market price of crops which are suitable for cultivating in plant factory, 2) relevant climate information, and 3) the corresponding yield rate data subject to certain environmental control factors. The data analysis tool by using functional principal component analysis (FPCA) will be developed to study the data and provide the better decision-making for plant factory operations. The second stage of this project will study the adaptive scheduling problem which complies with the crop marketing price and the yield rate. In order to enhance the profitability, the crop scheduling of a plant factory should consider higher-price crops as higher planting priority. Also, the scheduling should select an appropriate cultivating space for a crop based on it’s the associating yield rate. In this stage, the mixed integer programming (MIP) model will be utilized to formulate the scheduling problem. The third stage of this project will integrate the research achievements of the first and second stages, and propose an innovative rental service for the profitability enhancement. The rental service basically rents the space to customers for crop cultivation and also provides the professional planting service to improve the quality of crop cultivation and service satisfaction. This research will focus on developing a dynamic pricing strategy to execute the marketing positioning. By combining the scheduling model and the rental service, this project will investigate the resource allocation problem and analyzing the profitability of plant factory. In short, this project expects to contribute to developing a profitable plant factory operation by delivering the better decision-making.
Combining Sale Forecasting and Transshipment of Perishable Goods for Retailing Franchisee (2016.8~2017,7)
Funded by Ministry of Science and Technology（105-2221-E-011-108）, PI
This research aims to develop a sale forecasting model combined with a transshipment model for the perishable good franchisees to enhance the total sale in the franchisee system. Due to the nature of franchise, the franchisees order the materials or goods from the franchiser and sell the branded products under the franchise agreement. Especially, if the selling goods are perishable such as bread or food, the franchisees usually have relatively conservative ordering strategy to avoid the lost caused by scraping the overdue products. This conservative ordering strategy will cause the misleading on predicting sales. In order to provide the better sale forecasting, in this research, the Point of Sales (POS) data are collected and analyzed. The prediction on the perishable good’s sold-out time is used to predict the daily sale performance. Also, in order to reduce the loss caused by the sale forecasting discrepancy, the transshipment among franchises is suggested to transship the surplus goods which might not be sold in regular business hour to the stores which might face the shortage of the same goods. In this research, the decision making model of the transshipment will be developed. Based on the given probability of the predicted sale, the transshipment model first evaluates the need of transshipment for each franchisee. The transshipment price of goods will be investigated to setting up the optimal transshipment condition for enhancing the mutual benefit. The model considering economic benefits of the transshipment among franchisee system will be also studied.
Development of Heterogeneous Manufacturing Framework Technologies for Legacy Systems and Their CPS Application and Verification (2016.10~2019,9)
Funded by Ministry of Science and Technology（105-2218-E-011-015）, CoPI
Cyber-physical systems (CPS) are core technologies for industry 4.0. The key of CPS is to collect and feedback extensive information from the floor, conduct real-time control on physical systems so that they can meet the target performance. Most of the small and medium industry cannot build up CPS because the legacy systems do not have the network capability. Replacing the legacy systems just because of this reason is impossible. Therefore, to upgrade the systems to become CPS-enabled is necessary. The project will develop the heterogeneous manufacturing framework, build the interface for legacy systems, and establish the knowledge repository. The objective of this project is to make legacy systems compatible with the latest CPS-enabled systems. In the end, an automatic production line based on the heterogeneous manufacturing framework will be demonstrated to provide the industry a reference of industry 4.0.
The project consists of three subprojects. In subproject I, we will develop IoT and M2M interface and communication protocol for various legacy systems, and apply the analyzed results of the data collected from the systems on precision machining. In subproject II, we will develop the modeling, design, simulation, analysis of the production line. Also, we will implement smart monitoring, on-line sensing and diagnosis on the machines. In subproject III, we will introduce the big data cloud computing, planning, management, data analysis, service and lean management.
The project will focus on the realization of the CPS technology. We believe that the results will help upgrade the small and medium manufacturing industry, and provide CPS experts to our industry.
