基于卷积注意力长短时记忆网络的轴承寿命预测方法

Bearing life prediction method based on convolutional attention long-short term memory network

中文关键词  卷积注意力网络  长短时记忆网络  轴承  剩余使用寿命预测

英文关键词  convolution attention network  long-short term memory network  bearing  remaining useful life prediction

中文摘要

      传统的滚动轴承寿命预测方法缺乏明确的学习机制, 无法有效识别不同时序特征之间的差异并突出重要 特征, 影响其预测精度. 为克服上述缺点, 本文提出了一种基于卷积注意力长短时记忆网络(CAN-LSTM) 的剩余使 用寿命预测模型. 该模型主要由两部分组成: 前端为卷积注意力网络(CAN), 学习通道和时间维度中的深层故障特 征, 提高特征的表征能力; 后端为改进LSTM网络, 基于退化特征对轴承进行寿命预测. 归一化健康指标至[0,1]区间 内, 得到相同的失效阈值; 使用五点平滑法对预测结果进行处理, 实现预测结果的输出; 利用留一法对轴承全寿命 试验数据进行验证, 测试模型的准确性和适应性. 试验结果表明: 所提模型的平均均方根误差和平均绝对值误差比 仅用CNN 模型预测值低54.12% 和59.05%, 比仅用LSTM模型预测值低39.06% 和43.42%, 比卷积长短时记忆网络 (CNN-LSTM)低20.41% 和25.86%.

英文摘要

      The traditional rolling bearing life prediction method lacks an explicit learning mechanism to effectively identify the differences between different time series features and highlight important features, which affects the prediction accuracy. To overcome the above shortcomings, a model of remaining useful life prediction based on the convolutional attention long-short term memory network (CAN-LSTM) is proposed in this paper. The model is mainly composed of two parts: the front end is the convolution attention network (CAN) to learn deep fault features in channels and time dimensions to improve feature representation ability; the back end is the LSTM network, the bearing life is predicted based on degradation characteristics. The health index is normalized to the interval [0,1], and the same failure threshold is obtained. The five-point smoothing method is used to process the prediction results and realize the output of the prediction results; the accuracy and adaptability of the test model are verified by using the leave-one-out method to test the whole life test data of bearings. The experimental results show that the root mean square error and mean absolute error of the proposed model are 54.12% and 59.05% lower than those of CNN, 39.06% and 43.42% lower than those of LSTM, 20.41% and 25.86% lower than the convolutional long-short term memory network (CNN-LSTM).