当前位置:网站首页>A NOVEL DEEP PARALLEL TIME-SERIES RELATION NETWORK FOR FAULT DIAGNOSIS
A NOVEL DEEP PARALLEL TIME-SERIES RELATION NETWORK FOR FAULT DIAGNOSIS
2022-07-28 05:23:00 【山上的小酒馆】
介绍一篇2022年来自电子科大的论文,深度并行时序网络
0.摘要
考虑到应用时序数据的上下文信息的模型能够提高故障诊断的表现,一些网络(RNN,LSTM,GRU)对诊断结果比较有效。然后,这些模型受限于计算复杂度,难以达到很高的诊断效率。由于大尺寸的卷积核或深层的架构,并行CNN要处理长期的信息也难以达到较高的效率。BERT模型采用绝对位置编码为模型引入了上下文信息,但同时也带入了噪音。为了处理上述问题,提出deep parallel time-series relation network(DPTRN)。有以下三个优点:1. DPTRN基于MLP架构,可以十分有效的提高计算效率。2.改善绝对位置编码,使用解耦位置编码单元可以学习上下文信息。3.在特征可解释性方面有着明显的优点。
1.引言
PCA,ICA,RVM,DBN,等诊断方法。然而工业过程是连续变化的,因此,利用时序特征为工业过程的每个阶段建模是很有必要的。

模型架构如下:
1.建立时序数据集
2.划分当前时间节点D(T)与历史时间节点D(k),k=1,2,…,T-1.
3.将每一个历史时间节点与当前时间节点输入时序单元。
4.生成绝对位置编码,并利用位置查询query矩阵和位置key矩阵映射为解耦位置编码。
5.将时序信息单元与解耦位置编码单元的输出相加得到每个历史节点的时间关系权重。
6.历史信息向量与当前时间节点的特征拼接(contact)作为分类层的输入获取故障诊断结果。
边栏推荐
- Geek challenge 2019-sql injection five questions PW
- 硬件电路设计学习笔记1--温升设计
- ESXi社区版网卡驱动2022年3月更新
- CString转char[]函数
- EXFO 730c optical time domain reflectometer only has IOLm optical eye to upgrade OTDR (open OTDR permission)
- 论福禄克DTX-1800如何测试CAT7网线?
- ESXi 7.0 Update 1c中加入的systemMediaSize启动选项
- Reversible watermarking method based on difference expansion
- (PHP graduation project) based on PHP student daily behavior management system access
- Clustering of machine learning
猜你喜欢

压敏电阻设计参数及经典电路记录 硬件学习笔记5

Difference between shallow copy and deep copy

Tensorflow2.1 basic concepts and common functions

D2sc-gan: low resolution face recognition of classroom scenes based on dual depth and shallow channel generation confrontation network

Reversible watermarking method based on difference expansion

Reinforcement learning - proximal policy optimization algorithms

(PHP graduation project) based on PHP online travel website management system to obtain

关于Fusion on Apple Silicon的谨慎猜测

The difference and relation between TCP and UDP

Overview of unconstrained low resolution face recognition I: data sets for low resolution face recognition
随机推荐
Deep learning (incremental learning) -- iccv2021:ss-il: separated softmax for incremental learning
福禄克DSX2-5000、DSX2-8000模块如何找到校准到期日期?
Uniapp problem: "navigationbartextstyle" error: invalid prop: custom validator check failed for prop "Navigator
Model Inversion Attacks that Exploit Confidence Informati on and Basic Countermeasures 阅读心得
Adaface: quality adaptive margin for face recognition image quality adaptive edge loss for face recognition
Reinforcement learning - proximal policy optimization algorithms
VB-ocx应用于Web
Reading experience of protecting against DNN model steaming attacks
用于快速低分辨率人脸识别模型训练的改进知识蒸馏《Improved Knowledge Distillation for Training Fast LR_FR》
开关电源电路EMI设计在layout过程中注意事项
Scenario solution of distributed cluster architecture: cluster clock synchronization
Difference between shallow copy and deep copy
Tf.keras build neural network function expansion
Reversible watermarking method based on difference expansion
福禄克DSX2-5000 网络线缆测试仪为什么每年都要校准一次?
《On Low-Resolution Face Recognition in the Wild:Comparisons and New Techniques》低分辨率人脸识别论文解读
8类网线测试仪AEM testpro CV100 和FLUKE DSX-8000哪些事?
Agilent安捷伦 E5071测试阻抗、衰减均正常,惟独串扰NG?---修复方案
将GrilView中的数据转换成DataTable
短跳线DSX-8000测试正常,但是DSX-5000测试无长度显示?