基于自适应权重时间卷积网络的剩余使用寿命预测方法
Remaining useful life prediction based on an adaptive weight temporal convolutional network
宋浏阳 1金烨 2郭旭东 3王华庆1
作者信息
- 1. 北京化工大学 高端机械装备健康监控与自愈化北京市重点实验室,北京 100029;北京化工大学 高端压缩机及系统技术全国重点实验室,北京 100029
- 2. 北京化工大学 高端机械装备健康监控与自愈化北京市重点实验室,北京 100029
- 3. 长鑫集电(北京)存储技术有限公司,北京 100176
- 折叠
摘要
随着人工智能和大数据技术的发展,基于深度学习的高端装备剩余使用寿命预测技术倍受关注,准确的剩余使用寿命预测对于高端装备安全运行意义重大.大多数基于深度学习的预测方法在构建主体预测结构通常利用循环神经网络、长短期记忆网络和门控单元等手段,但存在并行计算能力较差、预测精度不足的问题.针对上述问题,提出一种基于自适应权重时间卷积网络的寿命预测方法.采用时间卷积网络搭建预测模型,利用空洞因果卷积和残差连接结构增强并行计算能力并避免信息泄露;构建基于自注意力机制的自适应权重模块,实现时序权重自动分配来提高预测精度;利用非对称损失函数增强提前预测倾向性,避免因滞后预测带来的安全和经济问题.选取发动机数据集和轴承数据集进行实验验证,结果表明,与其他深度学习方法相比,提出的自适应权重时间卷积网络提升了预测准确率并缩短了训练时间.
Abstract
With the development of artificial intelligence and big data,remaining useful life prediction technology based on deep learning for high-end equipment has attracted much attention.Accurate prediction of remaining use-ful life is a significant parameter affecting the safe operation of high-end equipment.Deep learning-based methods commonly use recurrent neural networks,and long and short-term memory networks and gating units when construc-ting the main prediction structure.However,there are problems of poor parallel computing capability and insuffi-cient prediction accuracy.In order to address these problems,we propose a useful life prediction method based on adaptive weight temporal convolutional networks.The prediction model is built using a temporal convolutional net-work,and the parallel computing capability is enhanced and information leakage is avoided by using dilated causal convolution and a residual connection structure.An adaptive weight module based on a self-attention mechanism is constructed to realize the automatic assignment of temporal weights to improve the prediction accuracy.The asym-metric loss function is used to enhance the tendency of advance prediction and avoid the safety and economic prob-lems caused by lagging prediction.Engine datasets and bearing datasets were selected for experimental validation.The results showed that the proposed adaptive weight temporal convolutional network improved the prediction accu-racy and reduced the training time compared with other deep learning methods.
关键词
剩余使用寿命/深度学习/时间卷积网络/自注意力机制Key words
remaining useful life/deep learning/temporal convolutional network/self-attention mechanism引用本文复制引用
基金项目
国家重点研发计划(2022YFB3303603)
国家自然科学基金(52375076)
出版年
2024
NETL
NSTL
万方数据