| 摘要: | 本論文主要針對紡織產業和空調系統中耗電量較大的織布定型機與冰水主機做資料分析與建立機器學習模型，以達到節能之目標。在紡織工廠中，熟悉機台參數的老師傅凋零，經驗無法累積，烘乾參數無法反應少量多樣接單模式。因此，本研究目標是建立織品烘乾過程的精確輸入/輸出關係模型，以常見機器學習演算法如決策樹、隨機森林、AdaBoost、KNN、XGBoost、神經網路、Voting regressor和Stacking regressor建立模型。通過Grid Search的方式尋找各種機器學習模型的最佳參數，並以整合學習方法來降低變異量、偏差，建立準確率更高的模型。也建立倒傳遞神經網路(BPNN)，加入Embeddings，透過One-cycle policy的學習方法，訓練神經網路達到滿意的含水率預測。最後，比較不同機器學習模型，以最佳的模型XGBoost來預測含水率，並計算SHAP(Shapley Additive exPlanations)值解釋模型輸入和輸出的正向或負向關聯。當給予一組含水率較大的測試資料，以前述最佳模型模擬定型機的運轉，藉此來調整某些輸入特徵，以求達到滿意的織物含水率，因而達到節能的目的。
在空調系統中，冰水主機約占總耗電量之60%，以目前200噸的主機而言，使用三年後的冰機效率約只達到新機的70%，因此，目標在以K-means將不同能耗的冰機分群，分析同類型(同噸數)冰機在不同聚類下造成低效肇因的輸入特徵(參數)。本文透過建立XGBoost模型，計算SHAP值，選擇和輸出關聯性最大的四個特徵，以平均輪廓分數(Silhouette score)最高來決定耗能冰機的聚類，比較和分析聚類結果和實際冰機資料，視覺化呈現造成冰機低效的肇因。最後，建立圖形使用者介面的冰機診斷系統，進行線上(On-line)診斷，以獨立樣本t檢定來檢驗當前冰機效率是否比模型估測效率還差，當檢定為真說明冰機效率開始變差，透過改善因子分析，告訴使用者可能的輸入肇因，藉此來調整設備，讓效率提升。
;Based on the machine learning algorithm and data analysis technique, this thesis studies the energy saving for the cloth setting machine in the textile industry and chiller in the air conditioning system. In the textile factories, the master who is familiar with the machine parameters is reducing year by year, hence, the useful and helpful experience could not pass on. On the other hand, the same parameters cannot reflect small-volume, large-variety production. The first purpose of this study is to build an accurate input-output model in the fabric drying process, therefore, a series of familiar machine learning algorithms, such as decision tree, AdaBoost, KNN (k-nearest neighbors), XGBoost, neural network, voting regressor, and stacking regressor are applied. The grid search is used to find the best parameters of machine learning models, and the integrated learning methods are used to reduce variation, bias and deviation such that the model with higher accuracy can be established. A back-propagation network with embeddings is trained with one-cycle policy to achieve satisfactory moisture content prediction. After comparing the performance of machine learning models, the best trained model is XGBoost. The model will predict the moisture of the fabric, and the SHAP (Shapley Additive exPlanations) value will be calculated to explain the positive or negative relationship between the input and output of model. When a set of test data with a larger moisture content is given, we use the aforementioned best model to simulate the operation of the stenter machine so that we can know how to adjust input characteristics to achieve a satisfactory fabric moisture content. Thus the purpose of energy saving is achieved.
In the air condition system, the chiller accounts for about 60% of total power consumption. Actually, for the 200-ton chiller, the performance is only about 70% in comparison with the new one after it runs for three years. Therefore, the second purpose of this study is to cluster different efficient chillers and look for the input characteristics (parameters) of inefficient causes. An XGBoost model is chosen and SHAP value is calculated to select the four features which have the highest correlation with the output. The highest average silhouette score is used to determine the clustering of energy-consuming chillers. The clustering results and actual chiller data are compared and analyzed to visualize the causes of the inefficiency of the chiller. Finally, a chiller diagnostic system with a graphical user interface is established to perform on-line diagnosis. An independent sample t test is used to test whether the current chiller efficiency is worse than the model estimated efficiency. When the test is true, it indicates the chiller efficiency starts to deteriorate. Then through improved factor analysis, we can find the possible input characteristics (parameters) of inefficient causes. |
