【作者】 张浩；

【导师】 孙洪华； 赵艳丽；

【作者基本信息】 内蒙古工业大学 ， 机械硕士（专业学位）， 2025， 硕士

【摘要】 随着工业智能化与技术现代化的快速发展,工业生产对机械设备的预测与健康管理(prognostic and health management,PHM)提出了更高的要求。滚动轴承的性能退化直接影响工业设备的运行效率与可靠性,实时监测滚动轴承运行状态并精准预测剩余使用寿命(Remaining Useful Life,RUL)有助于降低设备故障风险,提高维护决策的科学性,进而减少生产损失。近年来,深度学习模型出色的数据特征提取能力可以帮助滚动轴承RUL进行高效预测。本文围绕滚动轴承的RUL预测展开研究,主要内容包括:(1)滚动轴承在实际工况中运行受到多种噪声干扰,未经降噪处理的振动信号会影响后续的预测结果。为降低噪声对预测结果的影响,设计了一种改进的麻雀搜索算法,通过引入Halton序列初始化种群和拉普拉斯交叉策略,自适应寻优变分模态分解(Variational Mode Decomposition,VMD)重要参数组合。VMD分解后的本征模态分量利用相关系数筛选法进行分类,对含噪模态进行小波阈值降噪处理,重构信号后实现有效降噪。结果表明,该方法能够在去除噪声的同时保留关键信息,提取高质量信号,为后续RUL预测提供可靠的数据支持。(2)针对现有RUL预测模型在不同型号轴承或多变工况下泛化能力不足、预测精度下降的问题,设计了一种基于卷积神经网络(Convolutional Neural Networks,CNN)与双向长短期记忆网络(Bi-directional Long Short-Term Memory,BiLSTM)结合的滚动轴承RUL预测模型并加入多头注意力机制(Multi-Head Attention,MA)进行动态权重分配。采用时域与频域特征提取技术,筛选出单调性和鲁棒性较强的特征,将得到的敏感特征集输入CNN-BiLSTM-MA网络模型。在工况三下CNN-BiLSTM-MA模型与GRU模型、CNN-LSTM模型进行对比实验,CNN-BiLSTM-MA模型的RMSE平均值为0.1139,MAE平均值为0.0634,明显优于GRU模型的RMSE平均值0.1595,MAE的平均值0.1367和CNN-LSTM模型RMSE的平均值0.1600,MAE的平均值0.1413,结果表明在工况三下所设计模型的预测精度较高。(3)针对传统手动调参方法依赖专家经验需要通过人工调试寻找最优超参数组合,导致计算资源浪费和优化效率低下的问题,设计了一种基于贝叶斯的超参数优化方法对CNN-BiLSTM-MA网络模型进行超参数优化。贝叶斯优化通过构建目标函数的概率模型,根据历次的优化结果选出最优的超参数组合,减少手动调参的弊端提升优化效率,将得到的超参数组合输入CNN-BiLSTM-MA模型中,滚动轴承RUL预测的精度和稳定性得到进一步提升。更多还原

【Abstract】 With the rapid development of industrial intelligence and technological modernization,industrial production has put forward higher requirements for predictive and health management(PHM)of mechanical equipment.The performance degradation of rolling bearings directly affects the operational efficiency and reliability of industrial equipment.Real-time monitoring of rolling bearing operational status and accurate prediction of Remaining Useful Life(RUL)can help to reduce the risk of equipment failure,improve the scientific nature of maintenance decision-making,and thus reduce production losses.In recent years,the excellent data feature extraction capability of deep learning models can help to predict the RUL of rolling bearings efficiently.This paper centers on the RUL prediction of rolling bearings,and the main contents include:(1)Rolling bearings are subject to many kinds of noise disturbances in actual working conditions,and the vibration signals without noise reduction processing will affect the subsequent prediction results.In order to reduce the influence of noise on the prediction results,an improved sparrow search algorithm is designed,through the introduction of Halton sequence initialization population and Laplace crossover strategy,adaptive search for the optimization of the Variational Mode Decomposition(VMD)important parameter combinations.The intrinsic modal components of the VMD decomposition are classified by using correlation coefficient filtering,and the noisy modes are subjected to wavelet threshold noise reduction.After the VMD decomposition,the eigenmode components are classified using the correlation coefficient screening method,and the noise-containing modes are subjected to wavelet thresholding for noise reduction,and the signal is reconstructed to achieve effective noise reduction.The results show that this method can retain the key information while removing the noise,extract high-quality signals,and provide reliable data support for subsequent RUL prediction.(2)Aiming at the problem that the existing RUL prediction model has insufficient generalization ability and decreasing prediction accuracy under different types of bearings or variable working conditions,a rolling bearing RUL prediction model based on the combination of Convolutional Neural Networks(CNN)and Bi-directional Long Short-Term Memory(BiLSTM)is designed and incorporates the Multi-Head Attention(MA)mechanism for dynamic weight allocation.A rolling bearing RUL prediction model based on the combination of CNN and BiLSTM is designed with the addition of MA mechanism for dynamic weight allocation.The time domain and frequency domain feature extraction techniques are used to screen out the features with strong monotonicity and robustness,and the obtained sensitive feature set is input into the CNN-BiLSTM-MA network model.Comparison experiments are conducted between the CNN-BiLSTM-MA model and the GRU model and the CNN-LSTM model under working condition three,and the average value of RMSE of the CNN-BiLSTM-MA model is 0.1139,and the average value of MAE is 0.0634,which is significantly better than the average value of RMSE of the GRU model,which is 0.1595,and the average value of MAE,which is 0.1367,and the average value of CNN-LSTM model with an average value of RMSE of 0.1600 and an average value of MAE of 0.1413,the results show that the prediction accuracy of the designed model is higher under operating condition III.(3)Aiming at the problem that the traditional manual parameter tuning method relies on experts’experience and needs to find the optimal hyper-parameter combinations through manual debugging,which leads to the waste of computational resources and low efficiency of optimization,a Bayesian hyper-parameter optimization method based on Bayesian optimization is designed to optimize the hyper-parameters of the CNN-BiLSTM-MA network model.Bayesian optimization constructs the probability model of the objective function and selects the optimal hyperparameter combinations according to the successive optimization results,which reduces the drawbacks of manual debugging and improves the optimization efficiency,and the accuracy and stability of the rolling bearing RUL prediction are further improved by inputting the obtained hyperparameter combinations into the CNN-BiLSTM-MA model.更多还原