本論文針對進給系統中的雙螺帽式滾珠螺桿(Double-nut Ball Screw, DBS)提出故障診斷(Fault Diagnosis)方法，在搭配財團法人工業技術研究院南創園區所研發之預壓感測器(Preload Sensor)下，結合門控循環單元II型(Gate Recurrent Unit-II, GRU-II)，來判斷雙螺帽式滾珠螺桿是否發生故障。
本論文之研究目標在於: 藉由觀察預壓感測器之訊號之變化來(i) 判斷雙螺帽式滾珠螺桿是否發生故障; (ii) 建立故障預測系統。 為了達到上述目標，本論文首先對雙螺帽式滾珠螺桿建立故障資料庫，由於感測器所得之時序資料(Time Series)具有序列性及長度不一致，因此本論文採重新取樣(Resampling)來對訊號進行前處理，搭配GRU-II對訊號進行識別以判斷故障狀況為何。 相較於其他現行診斷方法，GRU-II特點在於: (i) 有效提升鑑別準確率、(ii) 運算過程中所需之參數較少、(iii) 針對噪音更具容錯性。
為了驗證本論文提出之故障診斷系統之效用，本論文使用深度學習套件PyTorch與Python建立一循環神經網路模型(GRU-II)，針對先前實驗所得之故障資料庫進行模擬分析，除此之外，並將雜訊所造成之影響一併考入。 由模擬結果可得知: (i) 在無噪音干擾下，本論文提出之GRU-II相較於其他神經網路: 前饋神經網路(Feed-forward Neural Network)、長短期記憶模型(Long Short Term Memory Model)與門控循環單元模型(Gate Recurrent Unit Model)，於準確率方面，分別優約13.3%、4.4%、4.4%。 (ii) 若考量訊號雜訊，GRU-II仍能維持準確率84.4%，依舊優於其他上述三種模型。 此外，於故障迴流道之召回率則可保持在100%，驗證了本論文提出之故障診斷方法有卓越的效能。

In this thesis, an innovative fault diagnosis method for Double-nut Ball Screw (DBS) systems is proposed. Different from traditional approaches based on accelerometers to examine the health status of ball screw and nuts, the methodology employed by this thesis is to use a preload sensor developed by Industry Technology Research Institute (ITRI) instead. With the major aim at diagnosing the failure of DBS system correctly and then establishing health management system, the combination of signal resampling and Gate Recurrent Uint-II (GRU-II) is established. This thesis at first creates a database which collects the failures that often occur in DBS systems, and then utilizes signal resampling subsequently to resolve the problems that the time series obtained by preload sensor is inconsistent in sequence and length. Following the pre-processing signals, GRU-II is manipulated to identify which failures are. In contrast to most current diagnosis method, GRU-II can achieve great accuracy as well as be insusceptible to noise. To evaluate the performance of GRU-II, it is set up and verified under Python environment. According to simulation results, it demonstrates that excellent accuracy can be achieved by the proposed GRU-II either with signal noise or not. Compared to the optimal approaches via other neural networks(Feed-forward, Long short Term Memory, Gate Recurrent Unit), the accuracy of GRU-II is better than those by 13.3%, 4.4% and 4.4% respectively. Furthermore, GRU-II can even maintain recall of recirculating mechanism at 100%. As mentioned above, the overall prognosis method in this thesis can be potentially applied to the real-world machine tools.

[1] 博畯實業有限公司, http://www.bearing.com.tw/?f=BallScrew-Preloading
[2] P. M. Frank, “Fault diagnosis in dynamic systems using analytical and knowledge-based redundancy: A survey and some new results.” Automatica, Volume 26, Pages 459-474, 1990.
[3] Y. C. Shin, Y. C. Huang, J. A. Chao, “Diagnosis of Ball Screw Preload Loss by Vibration Signals through the Hilbert-Huang Transform and Multiscale Measure.” Applied Mechanics and Materials, Volume 58, Pages 636-641, 2010.
[4] 許哲瑋，滾珠螺桿進給系統之健康診斷與監測，國立中正大學機械工程系研究所碩士論文，2010
[5] S. Frey, M. Walther, A. Verl, “Periodic Variation of Preloading in Ball Screws.” Production Engineering, Volume 4, Pages 261-267, 2010.
[6] Hans-Christian Möhring, Oliver Bertram, “Integrated Autonomous Monitoring of Ball Screw Drives.” CIRP Annals- Manufacturing Technology, Volume 61, Pages 355-358, 2012.
[7] 陳一愷，基於預壓感測器對雙螺帽式滾珠螺桿系統之故障診斷，國立成功大學機械工程系研究所碩士論文，2020
[8] Yang et al, “A Comparative Study on Time Series Classification.” Chinese Journal of Computers, Volume 30, Pages 1259-1266, 2007.
[9] F.Itakura, “Minimum Prediction Residual Principle Applied to Speech Recognition.” IEEE Transactions on Acoustics Speech and Signal Processing, Volume 23, Pages 67-72, 1975.
[10] P.J. Werbos, “Beyond Regression: New Tools for Prediction and Analysis in the Behavioral Science.” PhD thesis, Harvard University, 1974.
[11] J.J. Hopfield, “Neural Networks and Physical Systems with Emergent Collective Computational Abilities.” Proceedings of the National Academy of Sciences of the USA, Volume 49, Pages 2554-2558, 1982.
[12] S. Hochreiter, J. Schmidhuber, “Long Short-Term Memory.” Neural Computation, Volume 9, Pages 1735-1780, 1997.
[13] Cho et al., “Learning Phrase Representations using RNN Encoder-Decoder for Statistical Machine Translation.” Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing, Pages 1724-1734, 2014.
[14] S.J. Russekk, P. Norvig, “Artificial Intelligence: A Modern Approach.” Third Edition, Prentice Hall, 2010.
[15] C.E. Shannon, “A Mathematical Theory of Communication.” The Bell System Technical Journal, Volume 27, Pages 379-423 and 623-656, 1948.
[16] Wan et al, “Regularization of Neural Networks Using Dropconnect.” Proceedings of the 30th International Conference on Machine Learning, Pages 1058-1066, 2013.