Fast and accurate recognition of control chart patterns using a time delay neural network

Abstract

Pattern recognition is a critical issue in statistical process control because unnatural patterns displayed by control charts can be associated with specific causes that adversely impact on the manufacturing process. Recently, neural networks have been widely investigated as an effective approach to control chart pattern (CCP) recognition. However, most of the research in this field has used traditional back propagation networks (BPNs) that cannot deal with patterns that vary over time in an online CCP recognition scheme. This causes a pattern misclassification problem in nearly all neural network-based studies in the field of online CCP recognition. The present article presents a novel application of utilizing a time delay neural network (TDNN)-based model to address this problem. The TDNN, with its special architecture, can represent relationships between patterns in a time sequence, and is, therefore, suitable to be trained with dynamic patterns that change over time. Numerical results indicate that the pattern misclassification problem has been addressed effectively by the proposed TDNN-based model. When compared with traditional BPNs, the TDNN model has better performance in both recognition accuracy and speed. In comparison with traditional control chart approaches, the proposed model is capable of superior performance of average run length, while the category of the unnatural CCP can also be accurately identified.

管制圖的異常形狀可被用來當做尋找引發製程異常的原因(assignable cause) ， 因此有效且及時地辦識異常管制圖形狀(control chart pattern ， CCP)是統計製程管制(statistical process control, SPC)中的重要任務。 近年來 ， 類神經網路(neural network, NN)在辦識CCP的領域中有很好的績效表現 ， 但此領域中 ， 大部份的研究皆採用倒傳遞網路(back propagation network, BPN)為主要的偵測網路 ， BPN本身並不適合應用在與時間有關的應用上 ， 但管制圖本身卻是一個時間序列(time series) ， 這個問題使得NN為基本的CCP辨識器在辨識相似的CCP(如跳動(shift)及趨勢(trend))時 ， 產生極大的誤辨識率。 本研究應用時間延遲神經網路(time delay neural network, TDNN)來解決這個問題 ， TDNN用空間的架構來代表訓練樣本間的時間關係 ， 因此可用來學習動態的CCP。 本研究採用模擬的方式 ， 並以辨識率及平均串連長度(average run length, ARL)作為績效評估指標。 模擬結果顯示 ， 相似CCP之間的誤辨識(false recognition)問題可被此以TDNN為基的模式有效解決。 與BPN比較 ， TDNN具有較佳的辨識精度及速度 ， 與傳統的辨識工具相比 ， TDNN有較佳的ARL表現 ， 並具備分辨CCP種類的能力。

(*聯絡人: [email protected])

Keywords:

• time delay neural network
• pattern recognition
• control chart
• statistical process control

Keywords:

• 關鍵詞： 時間延遲神經網路 ; 形狀辨識 ; 管制圖 ; 統計製程管制

Acknowledgment

This work was partially supported by funding from the National Science Council of the Republic of China (NSC 96-2221-E-150-021-MY2, NSC. 96-2221-E-211-008).

Notes

(*聯絡人: [email protected])