用於機器學習進行半導體機台異常檢測的資料清理方法設計

Equipment Sensor Data Cleansing Algorithm Design for ML-Based Anomaly Detection

謝允哲(Yun-Che Hsieh)

指導教授 ： 張時中

國立臺灣大學/電機資訊學院/電機工程學研究所/碩士(2022年)

https://doi.org/10.6342/NTU202200346

摘要

在半導體晶圓製造領域中，針對機台的異常檢測可以提早發現機台的異常，而機台的異常會影響最終的成品，因此盡早的檢測出異常可以降低因機台異常導致的損失，並提高良率。常見的異常檢測主要是利用感測器收集到的數據來做分析，並判斷機台狀況是否正常。由於半導體機台感測器數量增加且感測器數值特徵不明，越來越多異常檢測採用機器學習的方式來做預測。在機器學習領域中，大部分的模型其神經網路的架構為固定的，因此做訓練學習時，對輸入資料的格式有一定的限制，例如一筆資料項目的數量、每筆資料的長度等等。
本論文研究探討半導體機台異常檢測所需的感測器資料清理問題，著重如何設計出一套資料清理演算法，對原始資料影響最低，並滿足異常檢測對輸入資料長度一致的要求。研究問題P)以及相應的挑戰C)如下:
P1)週期與步驟資料長度定義問題:週期資料為機台重複於不同的晶圓上執行同一製程感測器取樣收集到的資料，而週期資料對應到製程步驟的序列為步驟資料。不同週期中的步驟資料因工程師微調時間製程參數與資料傳輸過程資料遺失，導致週期與步驟資料長度(取樣點數量)不同。但機器學習模型對輸入資料每個週期與步驟的長度要求一致，如何決定清理後周期與步驟資料的固定長度?
C1)資料中取樣點數目與各取樣點數值變化關係不明。
P2)取樣點重要程度指標定義問題 : 如何定義指標量化每個取樣點重要程度，用於資料清理時決定取樣點刪補的優先順序?
C2)清理後的資料用於檢測異常，因此取樣點刪補應對各感測項目原始取樣資料的特徵影響最小。如何定義項目眾多且差異大的個別項目原始資料特徵，又如何進而設計指標來量化取樣點重要程度都是挑戰。
P3)資料刪補問題:如何搭配取樣點重要程度指標，選擇各項目一個週期資料中需要刪去或填補取樣點的位置與值?
C3)週期資料的型態(pattern)是取樣點數值依時序所形成，為檢測判斷的重要依據。搭配P2定義的指標進行刪、補取樣點時，如何以避免破壞原始資料型態為原則，決定取樣點的位置或數值是一大挑戰。
P4)資料清理對異常檢測效能影響的評估問題:如何評估並量化資料清理對異常檢測效能的影響是減損還是提升?
C4)機台感測器資料有限的情形下，如何根據真實的情境模擬資料出現異常的情形，來評估資料清理對異常檢測造成的影響是一項挑戰。
針對以上問題與挑戰，本論文新提出與解決方案如下:
M1)針對需更動步驟資料中相同步驟數量最少定義步驟資料中單一步驟長度。以所有週期下步驟資料中相同步驟數量的眾數，決定單一步驟固定的長度。基於眾數為出現次數最多的數量，可以達到需更動步驟資料中相同步驟數量最少。依序將步驟資料中每個步驟固定的長度相加做為週期資料固定長度。
M2)創新設計以取樣點序列中，相鄰兩點差距比例的前點大小做為取樣點重要程度指標。根據分析實務機台資料，觀察到單一項目相同製程參數下，不同週期資料會有相似的型態，而型態的變化是由取樣點數值的變化所形成，因此將取樣點數值變化定義為特徵。為了保留原始資料特徵，本研究設計以取樣點序列中，相鄰兩點差距比例的前點大小做為取樣點重要程度指標，比例越大，重要程度越高。
M3)新設計製程步驟導向資料清理(step based data cleansing)演算法。為了避免僅依靠特徵指標刪補取樣點破壞型態，設計利用群界點與熵選擇資料中相對多數取樣點相鄰兩點差距比例小的群刪補取樣點。群界點為序列中相鄰兩點的數值差距比例大於三倍標準差，分布為全序列相鄰兩點數值差距比例。群為資料拿掉群界點後連續的取樣點序列。計算每個群取樣點數值分布的熵，熵越小代表取樣點數值分布較集中，取樣點相鄰兩點差距比例小，選擇熵小的群刪補取樣點。刪去取樣點按照群中取樣點重要程度指標，由小到大做刪去。填補取樣點時，根據群中取樣點的數量與需要填補的取樣點數量決定填補的間隔，利用此間隔平均的於群中決定填補的位置，並根據填補位置前一點的值決定數值，達到保留原始資料型態且不增加M2定義的特徵。
M4)運用實務資料，設計資料清理對異常檢測效能影響的實驗架構並與現行方法進行實測比較。根據分析實務機台資料，觀察實際資料中常出現的幾種異常類型，包含飄移、移位與脈衝。設計異常模組將異常新增於原始資料，並以異常檢測模型STALAD為例，根據靈敏度、誤警率及檢測速度評估異常檢測的效能。比較的資料清理方式為實務上常用的軌跡對齊方法與經驗法則。
本論文的研究發現與貢獻包含:
(1)證明製程步驟導向資料清理更動步驟資料中相同步驟數量最少。所有週期下步驟資料中相同步驟數量出現最多次的為眾數，利用眾數做為步驟資料相同步驟數量固定的長度可以達到需更動步驟資料中相同步驟數量最少。
(2)透過實驗不同的異常類型發現異常檢測對取樣點數值變化敏感的特性。設計以取樣點序列中，相鄰兩點差距比例的前點大小做為取樣點重要程度指標。
(3)設計並實做製程步驟導向資料清理演算法於實務機台資料上，利用群界點與熵選擇資料中多數取樣點相鄰兩點差距比例小的群刪補取樣點，避免破壞原始資料型態。
(4)於實務情境下製程步驟導向資料清理搭配STALAD，相比軌跡對齊方法，在飄移異常大小1的情境可以提高STALAD檢測異常的靈敏度從0.2到0.98，降低STALAD檢測異常漏檢率從0.8到0.02，並提升STALAD檢測異常的檢測速度從第58筆週期資料到第19筆，且不改變誤警率。製程步驟導向資料清理將製程步驟對齊，提高正常週期資料間的相似度，降低異常檢測門檻值，使異常更容易被檢測。而異常資料的特徵為數值出現異常，透過差距比例指標將其保留。

關鍵字

資料清理 ； 資料特徵 ； 取樣點重要程度指標 ； 資料填補 ； 資料刪去 ； 異常檢測

並列摘要

In the field of semiconductor fabrication, equipment anomaly detection finds the abnormality of the semiconductor equipment, which will affect the final product. Therefore, early anomaly detection can not only reduce the loss caused by the anomaly of the semiconductor equipment, but also improve yield. Anomaly detection usually identify the condition of equipment by utilizing the data collected from the sensor. Due to the increased use of sensors in semiconductor equipment and the unknown l characteristics of the sensor readings, more and more anomaly detection uses machine learning to make predictions. In the field of machine learning, most models have a fixed neural network architecture, so when training and learning, there are certain restrictions on the format of the input data, such as the number of data items, the length of each data, etc. .
This research discusses sensor data cleansing for equipment anomaly detection, focusing on how to design a data cleaning algorithm which not only have the lowest impact on the original data but also meet the need of equal length data input . The main research problems P)and corresponding challenges C) are:
P1)Cycle and step data length definition problem: Cycle data is the equipment sensory data (ESD) collected from processing a specific recipe over a wafer by an equipment. Step data.is the sequence of the cycle data corresponding to the process steps.
C1)The relationship between the number of sampling points in the data and the value changes of each sampling point is unclear.
P2)Sampling point importance indicator definition problem: How to design an indicator to quantify the importance of each sampling point, and to determine the priority of sampling point discard and imputation?
C2)The cleansed data is used for anomaly detection, so the discard of sampling points should contain characteristics of the original sampling data . Challenges lie in defining the characteristics of the original data with numerous items and large differences. It’s also another challenge to design indicators quantifying the importance of sampling points.
P3)Sampling point discard and imputation problem: How to utilize importance indicator of the sampling points to select the positions and values of the sampling points that need to be deleted or imputed?
C3)Equipment anomaly detection identifies unusual behaviors in the pattern of cycle data. Challenges lie in utilizing sampling point importance indicator while doing sample discard and imputation.
P4)Evaluation of how data cleaning affect the performance of anomaly detection: How to evaluate and quantify the impact of data cleaning on anomaly detection performance?
C4)Due to the lack of ESD, it’s a challenge to evaluate the impact of data cleaning on the anomaly detection based on the real situation.
For the above problems and challenges, this study proposes and designs the following new solutions.
M1)Determine the fixed length of a single step based on the mode of sample points in the same step data in all cycles. Based on the mode as the number with the most occurrences, it can adjust least number of the same step data. Sequentially add the fixed length of each step in the step data as the fixed length of the cycle data.
M2)Propose the rate of the difference between two adjacent points in the sequence as indicator of sampling point importance. According to the analysis of real ESD, it is observed that under the same recipe, different cycle data have similar patterns. We define the change of the value of the sampling point as a characteristic, because the pattern is formed by it .In order to retain the characteristics of the original data, this study designed the rate of the difference between two adjacent points in the sequence as indicator of sampling point importance . The larger the rate, the higher the importance.
M3)Propose step based data cleaning algorithm. In order to retain data patterns, we utilize group boundary points and entropy to select the cluster for sampling point discard and imputation. The cluster boundary point is the ratio of the numerical difference between two adjacent points in the sequence greater than three times the standard deviation. The cluster is a sequence of consecutive sampling points after the data is removed from the cluster boundary. Calculate the entropy of the numerical distribution of the sampling points of each group. The smaller the entropy, the more concentrated the numerical distribution of the sampling points, the smaller the ratio of the difference between the two adjacent points of the sampling point, and the small entropy group is selected to delete the sampling points. Delete the sampling points according to the importance index of the sampling points in the group, and delete them from small to large. When filling sampling points, determine the filling interval according to the number of sampling points in the group and the number of sampling points that need to be filled, use this interval to determine the filling position in the group on average, and determine the value according to the value of the point before the filling position to achieve retention The original data type does not increase the characteristics defined by M2.
M4)Design the experimental framework of data cleaning on the anomaly detection performance and compare it with the current method. Based on the analysis of the actual machine data, observe several types of abnormalities that often appear in the actual data, including drift, shift, and pulse. The design anomaly module adds anomalies to the original data, and takes the anomaly detection model STALAD as an example to evaluate the effectiveness of anomaly detection based on sensitivity, false alarm rate and detection speed. The comparative data cleaning method is trajectory alignment and heuristic method, which is commonly used in practice.
The contributions of this research are as follows:
(1)Prove the step based data cleaning change the least step data. The most frequent occurrence of the same number of steps in the step data in all cycles is the mode. Using the mode as the step data, the fixed length of the same number of steps can achieve the least number of the same steps in the step data that needs to be changed.
(2)According to the analysis of different types anomalies, we discovered that anomaly detection is sensitive to the variation of sampling point. Propose the rate of the difference between two adjacent points in the sequence as indicator of sampling point importance.
(3)Design and implement the step based data cleaning algorithm on the real ESD. We use the cluster boundary point and entropy to select most of the sampling points in the data. The difference between the two adjacent points in the data is small. The group deletes the sampling points to avoid destroying the original data type.
(4)Compared with trajectory alignment, the sensitivity of STALAD's detection of abnormalities can be increased from 0.2 to 0.98 in the context of drift anomaly size 1, and the missed rate of STALAD's detection of abnormalities has been reduced from 0.8 to 0.02, and increase the detection speed of STALAD to detect abnormalities from the 58th cycle to the 19th, without changing the false alarm rate. Step based data cleaning aligns the process steps, improves the similarity between normal cycle data, lowers the threshold of abnormality detection, and makes abnormalities easier to be detected. The abnormal data is characterized by abnormal values, which are retained through indicator of sampling point importance.

並列關鍵字

data cleaning ； data characteristics ； indicator of sampling point importance ； data imputation ； data discard ； anomaly detection

參考文獻

[Che21] 資料探勘 (Data Mining) - 聯合大學資管系陳士杰[Online]. Available: http://debussy.im.nuu.edu.tw/sjchen/DataMining_final.html

[ChB09] Argon Chen and Jakey Blue , “Recipe-Independent Indicator for Tool HealthDiagnosis and Predictive Maintenance”, IEEE Transactions on Semiconductor Manufacturing,vol. 22 , November,2009

[RBY16] Hamideh Rostami, Jakey Blue, Claude Yugma, “Equipment condition diagnosis and fault fingerprint extraction in semiconductor manufacturing”, in Proc. of 15th IEEE International Conference on Machine Learning and Applications pp.534-539,2016.

[HeW11] Q. Peter He, Jin Wang ,“Statistics pattern analysis: A new process monitoring framework and its application to semiconductor batch processes”, in Proc. of AIChE Journal, vol. 57, No.1, pp.107-121,January, 2011.

[WaW19] Xi Wang; Chen Wang,“Time Series Data Cleaning: A Survey,”, in Proc. of IEEE Access,vol.8, pp.1866 – 1881,December,2019