| 摘要: | 本論文提出一種旋轉機械的軸承故障診斷方法，運用電機電流特徵分析(Motor Current Signature Analysis, MCSA)結合時頻分析(Time-Frequency Analysis, TFA)和深度學習(Deep Learning, DL)技術，目的是提升診斷的準確性和效率，同時降低檢測成本。由於實際故障數據取得不易，我們使用開源資料庫中的數據集進行分析，該資料集以25.6 kHz的取樣頻率收集溫度和電流數據，包含內環故障、外環故障、正常狀態、轉軸錯位和轉子不平衡五種狀態。在數據預處理階段，我們使用低通濾波器濾除高頻雜訊，並應用短時傅立葉轉換(Short-Time Fourier Transform, STFT)、連續小波轉換(Continuous Wavelet Transform, CWT)和希爾伯特-黃轉換(Hilbert-Huang Transform, HHT)等時頻分析方法，提取與故障相關的低頻特徵頻率，這些時頻分析方法提供更全面的信號時域和頻域信息，使故障特徵的識別更加準確。在模型訓練方面，我們採用英國牛津大學視覺幾何研究群(Visual Geometry Group, VGG)提出的VGG-16及VGG-19模型架構，還有由微軟提出的殘差神經網絡(Residual Neural Network, ResNet)架構ResNet-50，以及自定義的Modified VGG-16進行分類。通過混淆矩陣和多種評估指標對模型進行全面評估，如準確度、F1-Score以及自定義評估指標，該指標採用權重加權的方式，對模型的分類錯誤進行加權，更準確地反映不同類型錯誤的嚴重程度，目的是更全面地反映模型在實際應用中的性能和可靠性。研究結果表明，我們的優化模型在保持高準確度的同時，大幅減少模型參數量和大小，並且在某些情況下達到更好的分類效果，這使得我們的模型在資源受限的硬體環境中更具適應性。;This paper proposes a method for diagnosing bearing faults in rotating machine using motor current signature analysis (MCSA) combined with time-frequency analysis (TFA) and deep learning (DL) techniques. The purpose is to improve diagnostic accuracy and efficiency while reducing detection costs. Due to the difficulty in obtaining actual fault data, we utilize a dataset from an open-source repository, which collects temperature and current data at a sampling frequency of 25.6 kHz, including five states: inner race fault, outer race fault, normal, shaft misalignment, and rotor Imbalance. In the data preprocessing stage, we employ a low-pass filter to remove high-frequency noise and apply short-time Fourier transform (STFT), continuous wavelet transform (CWT), and Hilbert-Huang transform (HHT) for time-frequency analysis. These methods extract low-frequency characteristic frequencies related to faults, providing comprehensive information in both time and frequency domains, enhancing fault feature identification accuracy. For model training, we adopt VGG-16 and VGG-19 architectures proposed by the Visual Geometry Group (VGG) at the University of Oxford, as well as ResNet-50 architecture proposed by Microsoft, and a Customized Modified VGG-16 for classification. Through confusion matrix and various evaluation metrics such as accuracy, F1-Score, and a Customized evaluation metric weighted to reflect the severity of different types of classification errors, we comprehensively evaluate the models, aiming to accurately reflect the performance and reliability of the models in real-world applications. The results demonstrate that our optimized model maintains high accuracy while significantly reducing model parameters and size, and in some cases, achieves better classification performance, making our model more adaptable to resource-constrained hardware environments. |
