旋轉機械滾動軸承智慧故障診斷模型

Intelligent Fault Diagnosis Models for Rolling Bearing in Rotating Machinery

黎長安(Le Truong An)

指導教授 ： 李俊耀

中原大學/電機資訊學院/電機工程研究所/博士(2022年)

https://doi.org/10.6840/cycu202100304

摘要

根據測量信號的旋轉機械軸承故障的準確診斷仍然是一個引起廣泛關注的主要問題。目前，越來越多基於機器學習或深度學習理論的智慧故障診斷模型已被開發。這些模型預期能減少對人工的依賴，並增強診斷模型的自動故障檢測。構建智慧故障診斷模型有兩種方法：基於機器學習方法和基於深度學習方法。然而，這兩種方法的有效性仍是一個需要考慮的問題。因此，本研究提出了基於這兩種方法的模型應用於檢測旋轉機械的軸承故障。
第一種方法是基於機器學習的智慧軸承故障診斷模型（intelligent bearing fault diagnosis model based on machine learning, IBFDM based on ML）。此模型包括三個主要部分：特徵提取、特徵選取和特徵分類。旋轉機械的測量信號通過包絡線分析和希爾伯特-黃轉換技術處理以提取潛在特徵。通過基於特徵權重的群體初始化策略、新的群體更新機制以及群體篩選和替換過程對二進制粒子群最佳化進行了增強，創建了一種新的有效特徵選取方法，可提高分類精度並減少數據大小。最優特徵子集分別提供給人工神經網路以及支撐向量機作為最終識別任務。
第二種方法是基於深度學習的智慧軸承故障診斷模型（intelligent bearing fault diagnosis model based on deep learning, IBFDM based on DL）。此模型有兩個主要部分：第一部分是根據每個信號幀的持久性光譜構建圖像數據集。具有殘差網路（residual network, ResNet）結構的卷積類神經網路（convolutional neural network, CNN）被設計用於基於輸入數據的分類是第二部分。持久性光譜是從原始信號的包絡線中提取的。然後，基於短時傅立葉變換構建持久性光譜圖像，呈現出傳統頻譜分析方法未曾給出的每個信號的頻率、振幅和能量隨時間變化的新關係。具有 ResNet 結構的改進 CNN 允許從較低層到較高層直接連接特徵圖，以從包絡信號的持久性光譜圖像中探索判別特徵。這有助於利用低級層中的粒度特徵，這些特徵在傳統 CNN 中前饋通過相鄰層時可能會遺失。
因此，所提出的軸承故障診斷模型的性能在電流信號和振動信號的不同測試平台上得到驗證。模型的效率在軸承電流數據集上實現超過96%的辨識率，在軸承振動數據集上實現超過99%的辨識率。此外，IBFDM based on ML中的新特徵選取方法根據七個基準數據集進行評估，顯示出與其他同級競爭者相當的性能。此外，與其他類型的二維圖像（頻譜圖和尺度圖）和其他最先進的診斷模型相比，IBFDM based on DL的性能更佳。
綜上所述，所提出的兩種模型在自動識別旋轉機械健康狀態領域具有很高的可行性。

關鍵字

軸承故障診斷 ； 特徵提取 ； 特徵選取 ； 二進制粒子群最佳化 ； 卷積類神經網路 ； 持久性光譜 ； 殘差網路

並列摘要

An accurately bearing fault diagnosis of the rotary machinery from the measured signal remains a major problem that has been attracting a lot of attention. Currently, intelligent fault diagnosis models based on machine learning or deep learning theories are increasingly being developed. They are expected to reduce the reliance on human labor and enhance the automatic fault detection of diagnostic models. There are two approaches to building intelligent fault diagnosis models: intelligent fault diagnosis model based on machine learning method and intelligent fault diagnosis model based on deep learning method. The effectiveness of these two approaches remains an issue to be considered. This study proposes two models to be applied to detect bearing failures of rotating machines based on both approaches.
The first one is an intelligent bearing fault diagnosis model based on machine learning (IBFDM based on ML). This model has three main parts: feature extraction, feature selection, and feature classification. The measured signal of the electric motor is processed by envelope analysis and Hilbert-Huang transform techniques to extract the potential features. An enhancement of the binary particle swarm optimization algorithm through population initialization strategy based on feature weights, new updating mechanism, and the screening and replacing process create a new and effective feature selection method that improves classification accuracy and reduces data size. The optimal feature subset is provided separately for artificial neural networks, and support vector machine classifier for the final recognition task.
The second model is based on deep learning (IBFDM based on DL). This model has two main parts: the first part is to construct an image data set based on the persistence spectrum of each signal frame. The second part, the convolutional neural network (CNN) with residual network (ResNet) structure is designed for classification based on the input data. The persistence spectrum is extracted from the envelope of the raw signal. Then, the persistence spectrum image is constructed based on short-time Fourier transform, which presents a new relationship between the frequency, magnitude, and energy of each signal with time, which the traditional spectrum analysis methods have not been given before. An improved CNN with ResNet structure allows direct connection feature maps from the lower-level layer to the higher-level layer to explore the discriminant features from the persistence spectrum image of the envelope signal. That helps to exploit the granularity features in the low-level layer which can be lost when feedforward through adjacent layers like in a traditional CNN.
As a result, the performance of the proposed bearing fault diagnostic models is verified on the different testbeds with the current signal and the vibration signal. The efficiency of the models achieves high accuracy of over 96% on the bearing current data set and 99% on the bearing vibration data set. Besides, the new feature selection method in the IBFDM based on ML is evaluated against the seven benchmark data sets and shows performance comparable to other peer competitors. Moreover, the performance of IBFDM based on DL is higher than compared to other types of two-dimensional images (spectrogram and scalogram) and other state-of-the-art diagnosis models.
In conclusion, both proposed models are highly feasible in the field of automatically recognizing the health status of machines.

並列關鍵字

Bearing fault diagnosis ； feature extraction ； feature selection ； binary particle swarm optimization ； convolutional neural network ； persistence spectrum ； ResNet

參考文獻

[1] J. Carrera, R. Franceschi, and G. Gonzalez, “Induction Motor Fault Modeling Based on the Winding Function,” KnE Engineering, pp. 758-767, Jan. 2018.

[2] Y. Lei, B. Yang, X. Jiang, F. Jia, N. Li, and A. K. Nandi, “Applications of machine learning to machine fault diagnosis: A review and roadmap,” Mechanical Systems and Signal Processing, vol. 138, p. 106587, Apr. 2020.

[3] S. Kumar, D. Mukherjee, P. K. Guchhait, R. Banerjee, A. K. Srivastava, D. N. Vishwakarma, and R. K. Saket, “A Comprehensive Review of Condition Based Prognostic Maintenance (CBPM) for Induction Motor,” IEEE Access, vol. 7, pp. 90690-90704, Jul. 2019.

[4] J. Zhang, Y. Sun, L. Guo, H. Gao, X. Hong, and H. Song, “A new bearing fault diagnosis method based on modified convolutional neural networks,” Chinese Journal of Aeronautics, vol. 33, no. 2, pp. 439-447, Feb. 2020.

[5] P. Konar, and P. Chattopadhyay, “Multi-class fault diagnosis of induction motor using Hilbert and Wavelet Transform,” Applied Soft Computing, vol. 30, pp. 341-352, May 2015.