機器學習於機台設備健康狀態評估之應用

Abstract

近年來，隨著工業物聯網(industrial internet of things, IIOT)技術不斷的演進以及在工業4.0的影響下，預測性維護(predictive maintenance, PdM)逐漸取代部分傳統的維護方式，且隨著工業技術的不斷進步，機台設備相關運行狀態之數據量也呈指數級增長，單純的人工分析已無法有效地處理如此龐大的數據量，因此必須仰賴先進的數據科學和機器學習的技術以實現預測性維護。本研究旨在探討機器學習在機台設備健康狀態評估以及管理維護應用上的其他可能，期望能提出一種較有效率的機器學習模型應用於機台設備的管理維護。本研究先對數據資料進行預處理，包含重新編碼、資料特徵縮放、數據集劃分以及解決資料不平衡問題等，接著，再運用整理過之數據建構模型，而本研究採用三種演算法建構預測模型，分別為決策樹(decision tree, DF)、隨機森林(random forest, RF)以及循環神經網路(recurrent neural network, RNN)來進行預測，最後根據多個評估指標對各個模型的成效進行綜合比較與分析。本論文應用上述之研究流程分析一不具時間序列的單一時間點數據集，該數據集為機台設備在不同環境條件下，如大氣溫度(air temperature)、製程溫度(process temperature)等不同環境，透過感測器所偵測到的模擬狀態數據，將該數據集經由前述之研究流程可發現，隨機森林模型更加適合運用於不具時間序列的數據集。根據此研究成果可知，機台設備的維護管理可以藉由建立一套隨機森林模型來提高管理維護上的效率，而運用單一時間點的數據資料來訓練模型，也可以降低對數據完整度之需求，提高其實用性，更好的應用於機台設備的管理維護。

In recent years, with the continuous evolution of Industrial Internet of Things (IIoT) and the influence of Industry 4.0, predictive maintenance (PdM) has gradually replaced some parts of traditional maintenance methods. Additionally, with the ongoing advancements of industrial technology, the amount of data related to the operational status of machinery and equipment has grown exponentially. Simple manual analysis can no longer effectively handle such large volumes of data, thus advanced data science and machine learning techniques are required to implement predictive maintenance. This research aims to explore the applications of machine learning in state-of-health assessment and maintenance management for equipment, with the goal of proposing a more efficient machine learning model for equipment management and maintenance. The research first preprocesses the raw data through encoding, data scaling, data splitting, and addressing data imbalance issues. Then, the processed data is used to construct models, including decision tree (DT), random forest (RF), and recurrent neural network (RNN). Finally, comprehensively comparing and analyzing the models based on various evaluation indicators and figure out the most proper model for this type of dataset. The thesis applies the above research process to analyze a dataset with single-point-in-time data. The dataset simulates data captured by sensors from equipment under various environmental conditions, such as air temperature and process temperature. Following the mentioned research process, the result showed that the random forest model is more suitable for dataset without complete time series. Based on these research findings, implementing a random forest model can increase efficiency in management and maintenance. Besides, training the predictive model with single-point-in-time dataset can reduce the need for complete datasets, thereby enhancing its practicality and improving its application in machinery and equipment management and maintenance.