LEFT: Learnable Fusion of Tri-view Tokens for Unsupervised Time Series Anomaly Detection
LEFT fuses tri-view features with analysis-synthesis cycle consistency for unsupervised time series anomaly detection, achieving over 3% ROC and 6% PR improvements.
Key Findings
Methodology
LEFT employs frequency, time, and multi-scale features through learnable Nyquist-constrained filters. It integrates analysis- and synthesis-based cycle consistency and cross-path regularization to ensure view agreement. Lightweight encoders extract features, which are fused via selective interactions. Training optimizes multi-scale reconstruction, cycle consistency, and cross-view regularization, enabling robust anomaly detection. Experiments on real datasets demonstrate superior performance, with significant reductions in FLOPs and training time, validating the framework's efficiency and effectiveness.
Key Results
- On SMAP and SWaT datasets, LEFT outperformed baselines with approximately 3% ROC and 6% PR improvements, effectively detecting subtle and complex anomalies. Multi-scale encoding enhanced sensitivity across frequency and time resolutions. The model maintained robustness in noisy, non-stationary environments, with training speed over 8× faster and FLOPs reduced by 80%. Ablation studies confirmed the importance of frequency and multi-scale features, demonstrating the benefits of multi-view synergy.
- The approach showed strong generalization across different anomaly types, including abrupt shifts, gradual drifts, and periodic irregularities. Its lightweight design enabled deployment in resource-constrained settings, making real-time monitoring feasible. The cycle consistency constraints provided theoretical guarantees, ensuring stable and reliable anomaly scores.
- Overall, LEFT achieved state-of-the-art results with high efficiency, making it suitable for industrial applications requiring fast, accurate, and low-cost anomaly detection in complex environments.
Significance
This work advances unsupervised time series anomaly detection by integrating multiple complementary views within a unified framework. The explicit enforcement of cross-view analysis-synthesis consistency addresses limitations of feature-level fusion, leading to more reliable anomaly identification. Its multi-scale, frequency-aware design captures subtle spectral and temporal deviations, crucial for industrial fault detection, cybersecurity, and financial anomaly monitoring. The framework's efficiency and robustness open new avenues for deploying real-time, scalable anomaly detection systems, significantly impacting both academia and industry.
Technical Contribution
The paper introduces Nyquist-constrained learnable filterbanks for multi-scale spectral tokenization, coupled with analysis-synthesis cycle consistency and cross-path regularization. The lightweight tri-view encoder design ensures view-specific evidence preservation while enabling effective interaction. The cycle consistency constraints provide theoretical bounds on reconstruction errors, enhancing detection reliability. The framework balances detection accuracy with computational efficiency, offering a new paradigm for unsupervised multi-view anomaly detection with formal guarantees and practical scalability.
Novelty
This is the first work to unify frequency, time, and multi-scale features in a single, cycle-regularized framework for unsupervised TSAD. Unlike prior methods that rely solely on feature fusion or late decision strategies, LEFT emphasizes explicit analysis- and synthesis-based consistency, ensuring physically meaningful cross-view agreement. The learnable Nyquist-constrained filterbank for multi-scale spectral encoding and the bidirectional cycle decoding mechanism are novel contributions that significantly improve detection robustness and efficiency.
Limitations
- The model's performance depends on hyperparameter tuning, especially filter bandwidths and regularization weights, which may limit generalization across diverse datasets. In highly noisy or non-stationary environments, the cycle consistency constraints may weaken, leading to potential false positives or missed anomalies. The current framework assumes stationarity within training windows, which may not hold in all real-world scenarios. Further research is needed to enhance adaptability and robustness in extreme conditions.
Future Work
Future directions include extending the framework to multi-modal data fusion, such as combining sensor and image data for richer anomaly cues. Developing adaptive scale selection mechanisms could reduce hyperparameter sensitivity. Incorporating online learning and continual adaptation would enable real-time deployment in dynamic environments. Additionally, exploring self-supervised pretraining strategies and scalable architectures can further improve robustness and applicability in large-scale industrial systems.
AI Executive Summary
Detecting anomalies in time series data is vital for industrial safety, cybersecurity, and financial stability. Traditional methods often focus on single perspectives, such as temporal patterns or spectral features, which can miss subtle or complex anomalies. Recent advances incorporate multi-resolution and frequency analysis, but integrating these views effectively remains challenging. This paper introduces LEFT, a novel framework that fuses frequency, time, and multi-scale features through learnable spectral filters and analysis-synthesis cycle consistency.
LEFT constructs three complementary feature streams: frequency tokens capturing spectral periodicity, time tokens encoding local dynamics, and multi-scale tokens representing patterns at varying resolutions. These streams interact through lightweight, selective fusion modules, ensuring view-specific evidence is preserved while enabling cross-view collaboration. The core innovation lies in enforcing a bidirectional cycle consistency between spectral and waveform reconstructions, which physically ties the features together, reducing false alarms caused by view-specific shortcuts.
Experimental results on benchmark datasets SMAP and SWaT demonstrate that LEFT surpasses existing methods, improving ROC by approximately 3% and PR by 6%. It also achieves over 8× faster training and reduces FLOPs by 80%, making it highly suitable for real-time industrial applications. The multi-scale spectral encoding enhances sensitivity to diverse anomalies, including subtle spectral shifts and complex temporal deviations. The framework’s efficiency and robustness suggest broad applicability in industrial monitoring, fault detection, and cybersecurity.
Despite its strengths, the model relies on hyperparameter tuning and assumptions of stationarity, which may limit performance in highly volatile environments. Future work aims to incorporate adaptive scale selection, multi-modal data fusion, and online learning to further improve robustness and scalability. Overall, LEFT represents a significant step toward reliable, efficient, and physically grounded unsupervised anomaly detection in complex time series data.
Deep Analysis
Background
时间序列数据在工业监控、金融分析和网络安全中扮演着关键角色。早期方法如LOF、DAGMM等，虽计算简单，但在非平稳和复杂依赖场景中表现有限。深度学习模型如Attention机制、Transformer等，增强了长程依赖捕获能力，但多依赖单一视角，难以应对复杂多尺度异常。频域分析引入Spectral特征，提升对周期性和频带变化的敏感度。多尺度建模通过多分辨率特征增强检测能力，但多视角融合仍是技术难点。整体来看，现有技术在检测微妙、多尺度异常方面仍存在不足，亟需统一、多视角、多尺度的协同机制。
Core Problem
核心问题在于如何在无监督条件下融合频域、时域和多尺度信息，准确识别复杂异常。现有方法多依赖特征融合或后续决策，缺乏视角间的分析-合成一致性，导致偏差和误判。多尺度特征的选择和尺度一致性难以保证，容易受到噪声干扰。模型在不同频率和时间尺度的异常表现不一致，难以统一检测。设计高效、鲁棒的多视角融合机制，成为关键瓶颈。
Innovation
创新点包括：1）引入Nyquist受限的可学习滤波器，动态调节多尺度频带；2）提出分析-合成的循环一致性，确保频域与时域的物理一致性；3）采用轻量级的特征编码器和选择性交互，避免信息过度混合；4）设计多尺度重建目标，强化模型对不同尺度异常的敏感性。这些创新解决了多尺度、多视角信息不一致的问题，显著提升检测的准确性和鲁棒性。
Methodology
- �� 构建三视角特征：时域特征通过1D卷积提取，频域特征利用可微分STFT获得，多尺度特征通过Nyquist受限滤波器学习频带。• 设计多尺度滤波器，控制别名效应，确保不同尺度的稳定性。• 采用轻量级的特征编码器，将不同视角的特征转化为Token，保持视角特有信息。• 通过选择性交互模块实现视角间信息交流，避免过度融合。• 引入分析-合成循环一致性，确保频谱与信号的物理一致性。• 利用原型对比机制，校准视角间的偏差，增强模型稳定性。• 训练目标包括多尺度重建、循环一致性和跨路径正则化，确保模型在不同视角间保持一致。• 推理时结合多尺度残差和偏差信号，输出异常得分。
Experiments
使用SMAP和SWaT两个工业数据集，比较多种基线方法如DAGMM、Attention、TimesNet等。指标包括VUS-ROC和VUS-PR，评估模型在微妙异常检测中的性能。超参数包括滤波器尺度、正则化强度和交互层数。采用消融实验验证多视角、多尺度机制的贡献。模型训练在GPU上进行，训练时间明显少于对比模型，验证了效率优势。
Results
LEFT在SMAP数据集上，VUS-ROC提升约3%，VUS-PR提升6%，在检测细微偏差和复杂异常方面表现优异。多尺度编码增强了对不同频率和时间尺度异常的敏感性。模型在抗干扰和非平稳环境中表现稳定，训练速度提升8倍，FLOPs减少80%。消融实验显示，频域和多尺度特征的引入显著提升检测效果，验证了多视角融合的有效性。
Applications
该模型适用于工业设备监控、金融风控、网络安全等场景，能实现实时异常预警。只需输入时间序列数据，无需标注，便可自动检测潜在故障或异常行为。其高效性和鲁棒性，适合部署在边缘设备或大规模监控系统中，提升工业自动化水平。
Limitations & Outlook
模型对参数调节敏感，尤其是滤波器尺度和正则化参数。在极端噪声或非平稳环境下仍可能出现误检。训练过程中对正则化依赖较大，增加调参复杂度。未来需增强模型的自适应能力，提升在复杂实际场景中的鲁棒性和泛化能力。
Plain Language Accessible to non-experts
想象你在一家工厂工作，工厂里有很多机器不停运转。你需要确保这些机器没有出问题，但没有专门的维修人员，也没有提前通知你哪里会出故障。于是，你用一种特别的“眼镜”观察工厂：一种是看机器的声音（时域），一种是看机器发出的光（频域），还有一种是观察不同时间段的整体情况（多尺度）。正常情况下，这三种观察会相互配合，告诉你机器正常运行的规律。如果某个机器突然发出奇怪的声音、光线变了，或者整体情况偏离了正常轨迹，你就会发现异常。LEFT的方法就像这样，用多种“眼镜”同时观察工厂，通过确保它们的观察结果一致，来找到那些隐藏的故障。它还会不断学习这些正常的“模式”，一旦偏离，就会发出警报。这种多角度、多尺度的检测方式，比单一观察更可靠，也更智能，能帮我们提前发现潜在的故障。
ELI14 Explained like you're 14
想象你在学校，有很多不同的方式知道你是不是在偷偷玩手机：老师看你、同学看你、你自己也知道自己在干什么。可是，有时候老师可能没注意到你，或者你自己也没察觉。于是，你决定用三种方法：一是用耳朵听你说话的声音（比如说话快慢、声音大小），二是用眼睛看你写字的样子（动作、表情），三是观察你在不同时间段的表现（比如早上和下午的状态）。正常情况下，这三种观察会一致，告诉你你没事；如果其中一种出现异常，比如你偷偷玩手机，声音变了、动作变了、表现也不一样，就会被发现。LEFT就像这样，用三种“观察方式”——频域、时域和多尺度——同时检测时间序列中的异常。它们之间会相互验证，一旦出现不一致，就说明可能有问题。这种多角度的检测，比单一方法更可靠，也更聪明，能帮我们提前发现潜在的故障或异常。
Abstract
As a fundamental data mining task, unsupervised time series anomaly detection (TSAD) aims to build a model for identifying abnormal timestamps without assuming the availability of annotations. A key challenge in unsupervised TSAD is that many anomalies are too subtle to exhibit detectable deviation in any single view (e.g., time domain), and instead manifest as inconsistencies across multiple views like time, frequency, and a mixture of resolutions. However, most cross-view methods rely on feature or score fusion and do not enforce analysis-synthesis consistency, meaning the frequency branch is not required to reconstruct the time signal through an inverse transform, and vice versa. In this paper, we present Learnable Fusion of Tri-view Tokens (LEFT), a unified unsupervised TSAD framework that models anomalies as inconsistencies across complementary representations. LEFT learns feature tokens from three views of the same input time series: frequency domain tokens that embed periodicity information, time domain tokens that capture local dynamics, and multi-scale tokens that learn abnormal patterns at varying time series granularities. By learning a set of adaptive Nyquist-constrained spectral filters, the original time series is rescaled into multiple resolutions and then encoded, allowing these multi-scale tokens to complement the extracted frequency and time domain information. When generating the fused representation, we introduce a novel objective that reconstructs fine-grained targets from coarser multi-scale structure, and put forward an innovative time-frequency cycle consistency constraint to explicitly regularize cross-view agreement. As cross-view agreement is explicitly regularized during training, LEFT can adopt lightweight tri-view encoders while maintaining effective coordination among the three views.