基於多輸出時間卷積神經網路之齒輪箱振動缺陷分析系統研究

迴轉機械透過齒輪箱將動力傳送至生產設備在各類的機械設備中已被廣泛的使用，但是因為經常性的碰撞及振動導致零件的損耗破壞，進而導致整個機械設備出現故障，影響生產效率的機會提高。因此本研究利用深度學習方式，以直接通過時序振動資料來開發一套齒輪缺陷檢測系統用於機械故障的診斷，希望能透過振動訊號持續監測，提前發現故障訊號並進行故障判斷以提供使用者進行預防保養。本文首先利用更改後的時間卷積網路(TCN)架構結合改良後的異常檢測算法(TCN-AE)並使用無監督學習的方式進行缺陷檢測，以檢測時序齒輪振動資料中的異常值來確認目前機械系統是否正常。接著以更改的時間卷積網路架構，在其後添加壓縮及激勵網路(SE-Net)再加上自動編碼器的架構完成本研究之SE-TCN-AE網路模型結構。最後分別將SE-TCN-AE網路模型結合分類損失函數交叉熵(Cross-Entropy)進行故障分類；結合模型回歸損失函數中的均方誤差(MSE)函數進行磨耗程度判斷。實驗結果顯示使用無監督學習的異常檢測算法及使用SE-TCN-AE網路模型結合交叉熵進行故障分類之系統均能在有限樣本條件下就達到100%的準確度；使用SE-TCN-AE網路模型結合均方誤差函數進行磨耗程度判斷在三種磨耗程度訓練後之判斷MSE值約為 1.4×10^(-7)，若以輕重兩種磨耗程度訓練後去判斷中度磨耗齒輪組，其MSE為9.8×10^(-6)。證明了此方法在齒輪缺陷異常診斷以及缺陷分類上使用時序資料有著優異的結果。

關鍵字

時序振動資料；齒輪缺陷檢測；時間卷積網路；壓縮及激勵網路；自動編碼器

並列摘要

Rotary machines which transmit power to production equipment through gearboxes have been widely used in various types of machinery and equipment. However, due to frequent collisions and vibrations, the parts are damaged and destroyed, resulting in the failure of the entire mechanical equipment, which increases the chance of affecting production efficiency. Therefore, this study uses deep learning to develop a gear defect detection system for mechanical fault diagnosis directly through time series vibration data. It is hoped that through the continuous monitoring of vibration signals, fault signals can be detected in advance and fault judgment can be carried out to provide users with preventive maintenance. This paper firstly utilizes the modified temporal convolutional network (TCN) architecture combined with the improved anomaly detection algorithm (TCN-AE) and uses unsupervised learning for defect detection to detect outliers in timing gear vibration data to confirm the current mechanical system is normal or not. Then, with the modified temporal convolutional network architecture, followed by a compression and excitation network (SE-Net) plus an auto-encoder architecture to complete the SE-TCN-AE network model structure of this study. Finally, the SE-TCN-AE network model is combined with the classification loss function cross-entropy for fault classification, and is combined the mean square error (MSE) function in the model regression loss function to judge the degree of wear, individually.The experimental results show that both the anomaly detection algorithm using unsupervised learning and the system using SE-TCN-AE network model combined with cross-entropy for fault classification can achieve 100% accuracy under the condition of limited samples. The SE-TCN-AE network model combined with the mean square error function based on the training results of three wear degrees to judge the wear degree with a MSE value of 1.4×10^(-7). I If the gear set with moderate wear level is judged after training with two wear levels of light wear and heavy wear, the MSE value is 9.8×10^(-6). It is proved that this method has excellent results in the abnormal diagnosis of gear defects and the use of time series data in defect classification.

並列關鍵字

Time Series Vibration Data ； Gear Defect Diagnosis ； Temporal Convolutional Network (TCN) ； Squeeze and Excitation Network (SE-Net) ； Auto-Encoder

參考文獻

[1] Pappas, S. S., Ekonomou, L., Karampelas, P., Karamousantas, D. C., Katsikas, S. K., Chatzarakis, G. E., & Skafidas, P. D. (2010). Electricity demand load forecasting of the Hellenic power system using an ARMA model. Electric Power Systems Research, 80(3), 256-264.

Google Scholar

[2] Huang, S. J., & Shih, K. R. (2003). Short-term load forecasting via ARMA model identification including non-Gaussian process considerations. IEEE Transactions on power systems, 18(2), 673-679.

Google Scholar

[3] Guangming, C., & Yudong, X. (2010, June). A statistical model for gears noise prediction in gearbox. In 2010 International Conference on Electrical and Control Engineering (pp. 270-272). IEEE.

Google Scholar

[4] Zhao, D., Cheng, W., Gao, R. X., Yan, R., & Wang, P. (2019). Generalized Vold–Kalman filtering for nonstationary compound faults feature extraction of bearing and gear. IEEE Transactions on Instrumentation and Measurement, 69(2), 401-410.

Google Scholar

[5] Yang, D., Liu, Y., Li, S., Li, X., & Ma, L. (2015). Gear fault diagnosis based on support vector machine optimized by artificial bee colony algorithm. Mechanism and Machine Theory, 90, 219-229.

Google Scholar

主題瀏覽