Transformer‐Based Contrastive Learning With Dynamic Masking and Adaptive Pathways for Time Series Anomaly Detection

作者
Qian Liang,Xiang Yin
出处
期刊:Expert Systems [Wiley]
卷期号:42 (8) 被引量:3
标识
DOI:10.1111/exsy.70102
摘要

ABSTRACT Time Series Anomaly Detection (TSAD) has demonstrated broad applicability across various industries, including manufacturing, healthcare, and finance. Its primary objective is to identify unusual deviations in the test set by capturing the typical behavioral patterns of timing data. Despite their strong detection capabilities when labeled data is not available, current reconstruction‐based approaches still struggle with anomalous interference and inadequate semantic information extraction at higher time series levels. To tackle these problems, we provide a multi‐scale dual‐domain patch attention contrast learning model (DMAP‐DDCL) that incorporates adaptive path selection and adaptive dynamic context‐aware masking. Dynamic context‐aware masks are specifically used by DMAP‐DDCL to improve the model's generalization ability and mitigate bias resulting from the influence of anomalous data during training. Multi‐scale patch segmentation and dual attention to the segmented patches are introduced to capture local details and global correlations as time dependencies. By enlarging the contrast between the two data perspectives, global and local, DMAP‐DDCL improves the capacity to differentiate between normal and abnormal patterns. In addition, we enhance the adaptive path of the multi‐scale bi‐domain attention network, which adapts the multi‐scale modeling process to the temporal dynamics of the inputs and enhances the model's accuracy. According to experimental results, DMAP‐DDCL performs better on five real datasets from various domains than eight state‐of‐the‐art baselines. Specifically, our model enhances F1 and R_AUC_ROC by an average of 7.5% and 16.67%.

科研通智能强力驱动
Strongly Powered by AbleSci AI
科研通是完全免费的文献互助平台,具备全网最快的应助速度,最高的求助完成率。 对每一个文献求助,科研通都将尽心尽力,给求助人一个满意的交代。
实时播报
qian发布了新的文献求助10
刚刚
赘婿的应助被氨气采纳,获得10
刚刚
刚刚
唐正皓发布了新的文献求助10
刚刚
勤恳的老姆完成签到,获得积分10
刚刚
蚂蚁完成签到,获得积分10
刚刚
1秒前
大模型的应助被江sir采纳,获得50
1秒前
1秒前
Zhoushuang完成签到,获得积分10
2秒前
2秒前
2秒前
马宇飞发布了新的文献求助10
2秒前
BEMJ发布了新的文献求助10
3秒前
Wang发布了新的文献求助10
3秒前
dwass完成签到,获得积分10
3秒前
3秒前
3秒前
褚褚完成签到,获得积分10
3秒前
可爱的函函的应助被tang采纳,获得30
4秒前
沉静白凡完成签到,获得积分10
4秒前
dj完成签到,获得积分10
4秒前
小马甲的应助被Lee采纳,获得10
4秒前
zdx完成签到,获得积分10
5秒前
Jyouang发布了新的文献求助10
5秒前
一百完成签到,获得积分10
5秒前
qq完成签到,获得积分10
6秒前
莽兽鳞上最黑的皮完成签到,获得积分10
6秒前
6秒前
Gabriel1116发布了新的文献求助10
6秒前
科研通AI65的应助被xh采纳,获得10
7秒前
秋风的应助被yj采纳,获得10
7秒前
7秒前
WLM发布了新的文献求助10
7秒前
小魏小魏发布了新的文献求助10
7秒前
7秒前
7秒前
Sandclub发布了新的文献求助10
7秒前
ding的应助被Code_king采纳,获得50
7秒前
无限的鸵鸟完成签到,获得积分20
8秒前
高分求助中
(应助此贴封号)【重要!!请各用户(尤其是新用户)详细阅读】【科研通的精品贴汇总】 10000
拉伸高分子熔体中缠结行为的模拟和小角中子散射研究 1700
奈特基础组织学:配套组织病理学(第4版)(Netter's Essential Histology With Correlated Histopathology) 800
John McDowells Bildungstheorie 800
八旗制度の研究 600
plantTFDB 508
Resource Recovery Handbook 500
热门求助领域 (近24小时)
化学 材料科学 医学 生物 纳米技术 计算机科学 工程类 有机化学 化学工程 生物化学 物理 复合材料 内科学 催化作用 光电子学 人工智能 细胞生物学 无机化学 物理化学 数学
热门帖子
关注 科研通微信公众号,转发送积分 8175943
求助须知:如何正确求助?哪些是违规求助? 9567501
关于积分的说明 21793186
捐赠科研通 7770490
什么是DOI,文献DOI怎么找? 3431935
关于科研通互助平台的介绍 2556294
邀请新用户注册赠送积分活动 2477485