Bouchra Ananou

Fault prognosis for Discrete Manufacturing Processes

Abstract This paper deals with a fault prognosis method, based on the extraction of a health indicator (HI) from a large amount of raw sensors data, applied to Discrete Manufacturing Processes (DMP). The HI is extracted by locating the significant points of machine which are related to the degradation. The dynamics of HI is then analysed and modelled using an appropriate stochastic process. The adaptive aspect of the prediction model allows the updating of the Remaining Usesul Life (RUL) estimation. The developed approach is applied on a real case provided by ST-Microelectronics, where experimental result shows its efficiency.

Health Index Extraction Methods for Batch Processes in Semiconductor Manufacturing

This paper deals with a study of three methods for health index (HI) extraction in semiconductor manufacturing equipments. The first method uses degradation reconstruction-based identification with basic principal component analysis (PCA), the second one uses multiway PCA and the last one extracts HI from the significant points related to degradation. A comparison of these methods are made discussing about their efficiency and shortcoming for the implementation. The studied methods are applied on two data sets: 1) a simulation case and 2) a real case provided by ST-Microelectronics, where experimental results highlight the advantages and limits of each one.

Product quality prediction using alarm data : Application to the semiconductor manufacturing process

In the complex manufacturing processes, high quantity of products might be rejected. This can be due to the no detected failures. To evaluate the processing of manufacturing steps, alarms are setting off to indicate failures. However, industrial plant operators often receive many more alarms than they can manage, which include correlation. A poor alarm system may cause nuisance alarms and thus alarm floods, which reduces the ability of operators to take actions. This paper aims to identify unnecessary alarms within a large amount of event data. We prove the equivalence between similarity approaches in case of sparse binary data. The second purpose of this paper is the product quality prediction based on historical alarm events by using a regularized regression method. To demonstrate the effectiveness of these tools and their utility in the product quality prediction, we present an industrial case study based on alarm and scrap data collected from a semiconductor manufacturing process. Application results show the practicality and utility of the proposed methodology for both alarm management and product quality prediction.

Analysis of similarities between alarm events in the semiconductor manufacturing process

In the complex industry processes, failures occurred during manufacturing may deteriorate the final product quality, which results in a large number of scrapped products and thus a huge loss of the manufacturing Yield. To control the production steps, alarms go off to indicate occurred problems. However, alarms may be frequently raised which reduces the ability of operators to manage plant abnormalities. This paper deals with the study of correlation and the visualization of historical alarm data. Alarm events are mathematically represented as binary sequences, and two indices for identifying no critical alarms and similarities between alarms are proposed. An application on real industrial alarm data from the semiconductor manufacturing process has highlighted the practicability of the proposed approach in terms of reducing nuisance alarms, in order to decrease the operator overload.

Degradation modelling with operating mode changes

Remaining Useful Life (RUL) estimation is one of the main objectives of the fault prognosis, as it is the key element of the Condition-based Maintenance. This paper proposes a method of estimating the RUL in the presence of operating mode changes of the system by modeling the trend Health Index. The model does not require the knowledge of failure occurrence time and takes into account the current state of the system. Two application scenarios where the run to failure data is available and unavailable are given to show the efficiency of the developed method.