李俊賢

Jyun-SianLi

碩士論文（2020）

基於機器學習之食品供應鏈異常偵測方法與技術研發

Development of Machine Learning based Method and Technology for Food Supply Chain Anomaly Detection

關鍵字 Keywords

食品供應鏈、食品安全、食安監控與管理、安全食品防護、資料科學、機器學習、區塊鏈

Food supply chain、Food safety、Food safety monitoring and management、Safe food protection、Data Science、Machine Learning、Blockchain

摘要

食品供應鏈的複雜化導致食安事件頻繁發生，不僅造成社會不安，也直接或間接危害人們健康與生活，故食品安全的維護已成為世界各國重視之議題。隨著資料科學(Data Science)、機器學習(Machine Learning)與區塊鏈(Blockchain)的興起，系統智能化的理想已逐漸能夠實現，本研究參考資料科學之觀念與方法，以及機器學習與區塊鏈之技術，期為食品供應鏈之食品安全管理提供解方案，為人類帶來福祉。

本研究以資料科學之概念，設計一個「食安監控與管理模式」，針對「食安監控與管理模式」分析系統之需求，並參考區塊鏈之概念以及機器學習之原理，規劃「安全食品防護系統」之架構。依據該架構界定資料分析之需求、設計「食安稽查資料模型」，運用資料探勘(Data Mining)與機器學習技術，分析食品異常之影響因子與異常之模式，以開發與建置「異常偵測技術與機制」。

本研究以公開之供銷資料進行測試，驗證所提之技術的正確性與有效性。由於異常偵測之重點在將異常挑出越多越好，因此以召回率作為評估指標，最終供銷異常偵測模型之召回率由0.75027上升至0.86638。針對設備異常偵測，本研究同樣以公開之設備資料進行測試，模型之誤差率由0.006077下降至0.004112。上述之模型皆以評估指標進行評估，而評估指標能夠反映模型準確度，因此能夠驗證本研究之異常偵測方法與技術的有效性。

Abstract

The complexity of the food supply chain leads to frequent food security incidents, which not only causes social unrest, but also directly or indirectly endangers people's health and life. Therefore, the maintenance of food safety has become an important factor for countries all over the world. With the rise of Data Science, Machine Learning and Blockchain, the ideal of system intelligence has gradually been realized. This study looks forward to solving food safety problems through the support of intelligent systems to thereby promote human welfare.

This study designed a food safety monitoring and management model based on the concept of data science. For this food safety monitoring and management mode, the functional requirements of its system are analyzed, and the functional architecture of the safe food protection system is planned and designed with reference to the concept of blockchain and the principles of machine learning. According to the functional framework, define the needs of data analysis, design the food safety inspection data model, use machine learning technology to analyze the impact factors and abnormal patterns of food anomalies to build anomaly detection mechanism.

In order to verify the validity and correctness, this study uses public data to detect. First, detect supply and sales anomalies. Since anomaly detection focuses on picking out as many anomalies as possible, the recall rate is used as an validation index. The final model's recall rate increases from 0.75027 to 0.86638. Then it detects the abnormality of the supplier's equipment, and also uses public data to detect the error rate of the model from 0.006077 to 0.004112. The above models are all evaluated with validation index, and the validation index can reflect the accuracy of the model, so it can verify the effectiveness of the anomaly detection methods and techniques in this study.