以遞歸神經網路之自編碼器應用於馬達故障預測

Motor Fault Detection bu Using Recurrent Neural Network Autoencoder

黃麒瑞(Chi-Jui Huang)

指導教授 ： 陳俊勳

國立交通大學/工學院/機械工程系所/碩士(2019年)

摘要

本研究提出一個基於機器學習及深度學習之雙層分析架構來進行馬達故障預測，所使用的實驗數據來自於自行建置的馬達實驗平台。第一層分析模型首先結合遞歸神經網路(Recurrent Neural Network)改善自編碼器(Autoencoder)對馬達振動資料進行分析與降維，其過程是將資料循序式地將輸入模型，分別以RNN、LSTM(Long Short-Term Memory)及GRU(Gated Recurrent Unit)三種神經元結合自編碼器進行比較分析，確定神經元選擇後，為了調節不同的超參數(Hyperparameters)使模型最佳化，因此針對個別超參數進行實驗以得到最適超參數來建置模型。第二層採用類神經網路、支援向量機(Support Vector Machine)、隨機叢林(Random Forest)及XGBoost等演算法，對已降維資料進行錯誤類別分類，並使用主成分分析(Principal Components Analysis)及線性判別分析(Linear Discriminant Analysis)進行第二次的降維，使最後在平面上目視即可辨識錯誤。經過第一層模型所得之降維資料集，在通過以單層神經網路建立的十五類別分類模型中，測試資料集的準確率可達99%以上，並且在經過LDA進行第二次降維後，可在圖片中清楚判別正常情境與錯誤情境的區域，表示本研究所發展出之第一層模型對資料降維後，仍可以維持高維度數據資訊，且第二層模型有極佳的預測性能，顯示此架構可適用於時間序列數據分析。

關鍵字

馬達 ； 故障檢測 ； 數據降維 ； 遞歸神經網路 ； LSTM ； GRU ； 自編碼器

並列摘要

This research proposes a two-layer analysis architecture of machine learning and deep learning to predict the motor failure modes. The data were obtained from a self-built motor testing platform. The first layer analysis model integrates Recurrent Neural Network (RNN) with Autoencoder (AE) to analyze the data and perform the corresponding dimension reduction. The procedure is to input the data into the model sequentially, then, make the comparisons by using three different neurons, which are Basic RNN, Long Short-Term Memory, and Gated Recurrent Unit, respectively, combined with AE. As the specific neuron is determined, it carries out the experiments by using the various Hyperparameters in order to get the most suitable one to optimize the model. The second layer one adopts Artificial Neural Network, Support Vector Machine, Random Forest, and XGBoost algorithms to classify the dimension-reduction data into the corresponding fault categories. In the meantime, the Principal Components and Linear Discriminant Analyses are used to further perform the second dimension reduction, such that the different fault types of data can be visualized on a plane. The accuracy of the testing data via the fifteen-category fault detection model by using the single-layer ANN can reach 99%. After the second dimension reduction through LDA, the different fault types of data can be clearly identified in a picture. It indicates that the data after dimension reduction via the first layer model developed by this research still can maintain the high-dimensional data information. The second layer model can provide excellent prediction performance. Those demonstrate that the architecture proposed by this study is good enough to be applied to time-series data analysis.

並列關鍵字

motor ； fault detection ； data dimension reduction ； RNN ； LSTM ； GRU ； Autoencoder

參考文獻

[1] S. Wang, W. Huang, and Z. Zhu, "Transient modeling and parameter identification based on wavelet and correlation filtering for rotating machine fault diagnosis," Mechanical systems and signal processing, vol. 25, no. 4, pp. 1299-1320, 2011.

[2] G. G. Barna, S. W. Butler, D. A. Sofge, and D. A. White, "Virtual sensor based monitoring and fault detection/classification system and method for semiconductor processing equipment," ed: Google Patents, 1999.

[3] J. Manyika et al., "Big data: The next frontier for innovation, competition, and productivity," 2011.

[4] A. Arnaiz et al., "Information and communication technologies within E-maintenance," in E-maintenance: Springer, 2010, pp. 39-60.

[5] M. C. Comstock, J. E. Braun, and R. Bernhard, Experimental data from fault detection and diagnostic studies on a centrifugal chiller. Purdue University, 1999.