| 摘要: | 近年來，隨著工業程度的進步，在工業4.0的革命下，傳統的製造生產方式逐漸轉型成智慧製造。因此，各國企業紛紛導入像是工業物聯網（Industrial Internet of Things, IIOT）、大數據分析（Big Data）、感測器（Sensor）及人工智慧（Artificial Intelligence, AI）等等智慧製造之技術。利用設備整合人工智慧之技術達成準確、快速、省時與省力的目標，為求在競爭激烈下能夠脫穎而出。
達成智慧製造的一大關鍵在於減少設備之停機及非計劃性維護的次數等等的風險。設備透過感測器進行資料收集，同時藉由預測性維護技術達成提早預測異常狀態或停機，進而提升整體生產線之設備效率。
本研究以A公司之塗佈機為例，以其感測器所收集的數據作為資料來源，其中數據包括張力、扭力、速度及電流等變數。由於數據為時間序列資料，因此本研究透過長短期記憶網路（Long Short Term Memory, LSTM）及門控循環單元（Gated Recurrent Unit, GRU）方法進行異常檢測，並且透過搭配不同的非飽和激勵函數進行建模與分析。實驗結果顯示本研究所建置的24種模型皆有極高的準確率，且召回率皆達100%，又以LSTM模型，搭配激勵函數Leaky ReLU、隱藏層層數為兩層、神經元個數為128個的配置最佳，模型準確率達99.77%、特異度為99.76%、F1-Score為82.86%，與實際張力異常紀錄相比，模型能夠在異常發生前12秒有效預測機台異常狀況，有助於降低設備非計劃性之異常或停機，進而達成降低整體生產線之成本。;In recent years, with the progress of the industrial level, under the revolution of Industry 4.0, the traditional manufacturing and production methods have gradually transformed into smart factory. Therefore, companies from various countries have introduced smart factory technologies such as Industrial Internet of Things (IIOT), Big Data Analysis, Sensor and Artificial Intelligence (AI). Use the technology of equipment integration AI to achieve accurate, fast, time-saving and labor-saving goals, in order to stand out in the fierce competition.
One of the keys to achieving smart factory is to reduce the risks of equipment downtime and unplanned maintenance. The equipment collects data through sensors, and at the same time uses predictive maintenance technology to achieve early prediction of abnormal conditions or shutdowns, thereby improving the equipment efficiency of the overall production line.
This study takes the coating machine of company A as an example, and uses the data collected by its sensors as the source of data. The data includes variables such as tension, torque, speed and current. Since the data is time series data, this study uses Long Short Term Memory (LSTM) and Gated Recurrent Unit (GRU) methods for anomaly detection, and by using different unsaturated excitation functions to model and analyze. The experimental results show that the 24 models built in this research all have extremely high accuracy, and the recall rate reaches 100%. The LSTM model with the following hyperparameter settings is the best: the activation function is Leaky ReLU, the number of hidden layers is 2, the number of neurons is 128. The accuracy rate of this model is 99.77%, the specificity is 99.76%, and the F1-Score is 82.86%. Compared with the actual tension abnormal record, the model can effectively predict the abnormal condition of the machine 12 seconds before the abnormality occurs, which helps to reduce unplanned abnormalities or shutdowns of equipment, thereby reducing the cost of the overall production line. |
