- 學位論文
基於射頻訊號特徵萃取之電漿蝕刻設備故障偵測
Feature Extraction of Radio Frequency Signal for Fault Detection of Plasma Etching Equipment
摘要
蝕刻製程是半導體製程中相當重要且不可或缺的一環，扮演將光罩定義出的圖案刻畫出來的重責大任，且大部分的蝕刻製程都是不可逆的，不管是線寬、蝕刻深度以及產品均勻度的控制，都需精密的設備將製程環境做嚴格控制下才能達到，而大量機構及元件所組成之設備，每一機構元件都有衰退損壞的 可能，如何能夠偵測發現異常故障的發生，是一項相當重要的課題。 自組特徵映射圖網路是具有非監督式競爭式學習功能的網路，輸入資料項會映射到相同的自組映射輸出單元，最後輸出層的神經元會依據輸入向量的特徵以有意義的拓樸結構展現在輸出空間中。因此，自組映射可以將高維度的輸入資料，以低維度的空間圖形表現出來，藉此以半導體蝕刻製程中射頻功率時序波形的變化來進行模擬驗證，期能自類神經演算的方法，結合半導體設備輸出訊號，來進行設備的故障偵測，避免因突發性故障造成龐大人力、物力以及 時間的損失，進而能夠提升半導體產業發展的競爭力。 在實驗過程中，先在製程記錄中擷取正常10筆射頻輸出訊號進行離線訓練，定義出正常橢圓界限初始值，再以設備輸出時序波形訊號400筆輸入自主織映射圖網路進行比對，找出6組異常訊號輸出，比對機台故障分析結果發現有5種射頻系統故障類別，證實以特徵權重值群聚佐以橢圓界限區分，能夠達到故障偵測的效果。
並列摘要
Etch process is an important and indispensable process in the semiconductor manufacturing for removing the pattern defined by photolithography process. Most etched profiles cannot be reworked. Thus, the accuracy of critical dimension, etch depth and etch uniformity depends on the stringently control of etch equipment. However, every mechanism has a chance to decline or damage. Fault detection becomes an important subject in the semiconductor manufacturing. Self organizing feature map network is an artificial neural network with unsupervising competition learning. Input data will be mapped to the output units by self organizing and, finally, assigned in the output space with meaningful topological structure by output neurons according to input data’s vector characteristics. Thus, self organizing feature map network transfers higher dimensional input data to lower dimensional spacial figures. In this thesis, the self organizing feature map network was used for fault detection of radio frequency (RF) power of etch equipment. The waveform of RF power is monitored by self organizing feature map network to prevent faulty process so as to reduce the cost loss in the semiconductor manufacturing. The first step of experiments was to collect 10 sets of normal signals of RF power for offline training. Through the self organizing feature map network, an ellipse was calculated to cover all outputs of neural network and defined as threshold limit. Then, this trained result was implemented to fault detection of etch equipment. A total of 400 waveforms of RF power, i.e. 400 runs of etch processes, were monitored. Among them, 6 abnormal waveform signals were detected and classified into 5 kinds of faults of RF system. The results show that this proposed approach of fault detection has excellent performance.