基于LSTM编码器解码器的无监督健康指数多传感器预测
需积分: 10 118 浏览量
更新于2024-09-08
收藏 1.7MB PDF 举报
"这篇论文提出了一种基于LSTM编码器-解码器的无监督健康指数方法，用于多传感器预测性维护。这种方法适用于工业环境，能够处理无模式的系统退化问题，通过训练LSTM模型来重构系统的健康状态时间序列，并利用重建误差计算健康指数，进而估计剩余使用寿命。论文对涡轮风扇发动机和铣削机床的公开数据集进行了评估，并展示了来自实际工业环境的数据集的结果。"
在工业4.0和物联网(IoT)时代，设备的健康监测和预测性维护变得至关重要。传统的故障诊断方法依赖于人为设定的规则或对已知故障模式的匹配，而这种方法往往无法适应复杂的系统退化行为。多传感器数据提供了一种更全面的设备状态监测方式，但如何有效地利用这些数据进行预测性维护仍是一个挑战。
论文提出的LSTM编码器-解码器模型是一种递归神经网络结构，特别适合处理序列数据，如时间序列的传感器数据。LSTM网络能够捕捉时间序列中的长期依赖关系，这对于理解和预测系统的健康状态非常有用。在这个模型中，编码器部分负责将输入的时间序列数据压缩成一个紧凑的表示（也称为潜空间表示），而解码器则尝试从这个表示中重建原始序列。
在训练过程中，模型的目标是尽可能准确地重构健康的系统状态时间序列。当模型的重构误差增大时，这通常意味着系统的健康状态正在恶化，因此可以利用这个误差来计算无监督的健康指数(HI)。HI值的变化趋势可以反映设备的退化程度，进一步用于预测设备的剩余使用寿命(RUL)。
在实验部分，该方法在两个公开数据集上进行了验证：一个是NASA的涡轮风扇发动机数据集，另一个是铣削机床数据集。这两个数据集包含了多传感器数据，模拟了真实世界的复杂工况。此外，作者还展示了一个来自实际工业环境的数据集的结果，进一步证明了该方法的实用性和有效性。
通过与传统方法的对比，该研究展示了LSTM-ED在处理无规律的系统退化问题上的优势，其无监督特性使得它无需大量的有标签数据，降低了数据预处理和标记的成本。这种方法对于提高工业设备的运行效率、降低维护成本以及预防意外停机具有重要的实际意义。
Multi-Sensor Prognostics using an Unsupervised Health
Index based on LSTM Encoder-Decoder
Pankaj Malhotra, Vishnu TV, Anusha Ramakrishnan
Gaurangi Anand, Lovekesh Vig, Puneet Agarwal, Gautam Shro↵
TCS Research, New Delhi, India
{malhotra.pankaj, vishnu.tv, anusha.ramakrishnan}@tcs.com
{gaurangi.anand, lovekesh.vig, puneet.a, gautam.shro↵}@tcs.com
ABSTRACT
Many approaches for estimation of Remaining Useful Life
(RUL) of a machine, using its operational sensor data, make
assumptions ab out how a system degrades or a fault evolves,
e.g., exponential degradation. However, in many domains
degradation may not follow a pattern. We prop ose a Long
Short Term Memory based Encoder-Decoder (LSTM-ED)
scheme to obtain an unsupervised health index (HI) for
a system using multi-sensor time-series data. LSTM-ED
is trained to reconstruct the time-series corresponding to
healthy state of a system. The reconstruction error is used
to compute HI which is then used for RUL estimation. We
evaluate our approach on publicly available Turbofan Engine
and Milling Machine datasets. We also present results on a
real-world industry dataset from a pulverizer mill where we
find significant correlation between LSTM-ED based HI and
maintenance costs.
1. INTRODUCTION
Industrial Internet has given rise to availability of sensor
data from numerous machines belonging to various domains
such as agriculture, energy, manufacturing etc. These sensor
readings can indicate health of the machines. This has led
to increased business desire to p erform maintenance of these
machines based on their condition rather than following the
current industry practice of time-based maintenance. It has
also been shown that condition-based maintenance can lead
to significant financial savings. Such goal s can be achieved
by building models for prediction of remaining useful life
(RUL) of the machines, based on their sensor readings.
Traditional approach for RUL prediction is based on an
assumption that the health degradation curves (drawn w.r.t.
time) follow specific shape such as exponential or linear.
Under this assumption we can build a model for health
index (HI) prediction, as a function of sensor readings.
Extrap olation of HI is used for prediction of RUL [29, 24,
25]. However, we observed that such assumptions do not
hold in the real-world datasets, mak ing the problem harder
to solve. Some of the important challenges in solving the
prognostics problem are: i) health degradation curve may
not necessarily follow a fixed shape, ii) time to reach s ame
level of degradation by machines of same specifications is
often di↵erent, iii) each instance has a slightly di↵erent
initial health or wear, iv) sensor readings if available are
Presented at 1st ACM SIGKDD Workshop on Machine Learning for Prognostics
and Health Management, San Francisco, CA, USA, 2016. Copyright 2016 Tata
Consultancy Services Ltd.
noisy, v) sensor data till end-of-life is not easily available
b ecause in practice periodic maintenance is performed.
Apart from the health index (HI) based approach as
describ ed above, mathematical models of the underlying
physical system, fault propagation models and conventional
reliability models have also been used for RUL estimation
[5, 26]. Data-driven models which use readings of sensors
carrying degradation or wear information such as vibration
in a b earing have been e↵ect ively used to build RUL
estimation models [28, 29, 37]. Typically, sensor readings
over the entire operational life of multiple instances of a
system from start till failure are used to obtain c ommon
degradation behavior trends or to build models of how a
system degrades by estimating health in terms of HI. Any
new instance is then compared with these trends and the
most similar trends are used to estimate the RUL [40].
LSTM networks are recurrent neural network models
that have been successfully used for many sequence
learning and temporal modeling tasks [12, 2] such as
handwriting recognition, speech recognition, sentiment
analysis, and c ustomer behavior prediction. A variant
of LSTM networks, LSTM encoder-decoder (LSTM-ED)
mo del has been successfully used for sequence-to-sequence
learning tasks [8, 34, 4] like machine translation, natural
language generation and reconstruction, parsing, and
image captioning. LSTM-ED works as follows: An
LSTM-based encoder is used to map a multivariate input
sequence to a fixed-dimensional vector representation. The
deco der is another LSTM network which uses this vector
representation to produce the target sequence. We provide
further details on LSTM-ED in Sections 4.1 and 4.2.
LSTM Encoder-decoder based approaches have been
prop osed for anomaly detection [21, 23]. These approaches
learn a model to reconstruct the normal data (e.g. when
machine is in perfect health) such that the learned model
could reconstruct the subsequences which belong to normal
b ehavior. The learnt model leads to high reconstruction
error for anomalous or novel subsequences, since it has not
seen such data during training. Based on similar ideas,
we use Long Short-Term Memory [14] Encoder-Decoder
(LSTM-ED) for RUL estimation. In this paper, we propose
an unsupervised technique to obtain a health index (HI)
for a system using multi-sensor time-series data, which does
not make any assumption on the shape of the degradation
curve. We use LSTM-ED to learn a model of normal
b ehavior of a system, which is trained to reconstruct
multivariate time-series corresponding to normal behavior.
The reconstruction error at a point in a time-series is used
arXiv:1608.06154v1 [cs.LG] 22 Aug 2016
下载后可阅读完整内容，剩余9页未读，立即下载
2020-03-29 上传
2022-08-03 上传
2022-08-03 上传
2023-06-01 上传
2023-06-01 上传
2023-06-01 上传
2023-06-03 上传
2023-06-03 上传
2023-06-01 上传
yushuyanggyx
- 粉丝: 0
- 资源: 1
我的内容管理
展开
最新资源
- 高清艺术文字图标资源，PNG和ICO格式免费下载
- mui框架HTML5应用界面组件使用示例教程
- Vue.js开发利器：chrome-vue-devtools插件解析
- 掌握ElectronBrowserJS：打造跨平台电子应用
- 前端导师教程：构建与部署社交证明页面
- Java多线程与线程安全在断点续传中的实现
- 免Root一键卸载安卓预装应用教程
- 易语言实现高级表格滚动条完美控制技巧
- 超声波测距尺的源码实现
- 数据可视化与交互：构建易用的数据界面
- 实现Discourse外聘回复自动标记的简易插件
- 链表的头插法与尾插法实现及长度计算
- Playwright与Typescript及Mocha集成：自动化UI测试实践指南
- 128x128像素线性工具图标下载集合
- 易语言安装包程序增强版：智能导入与重复库过滤
- 利用AJAX与Spotify API在Google地图中探索世界音乐排行榜
