[Paper Review] Time Series Anomaly Detection Using Convolutional Neural Networks and Transfer Learning

Tailai Wen, Roy Keyes|arXiv (Cornell University)|May 31, 2019

Anomaly Detection Techniques and ApplicationsComputer Science22 references100 citations

TL;DR

This paper introduces a CNN-based time series segmentation approach (U-Net) for anomaly detection and a transfer learning framework (MU-Net) to adapt univariate pretraining to multivariate tasks, validated on synthetic and real datasets.

ABSTRACT

Time series anomaly detection plays a critical role in automated monitoring systems. Most previous deep learning efforts related to time series anomaly detection were based on recurrent neural networks (RNN). In this paper, we propose a time series segmentation approach based on convolutional neural networks (CNN) for anomaly detection. Moreover, we propose a transfer learning framework that pretrains a model on a large-scale synthetic univariate time series data set and then fine-tunes its weights on small-scale, univariate or multivariate data sets with previously unseen classes of anomalies. For the multivariate case, we introduce a novel network architecture. The approach was tested on multiple synthetic and real data sets successfully.

Motivation & Objective

• Motivate anomaly detection in time series for IoT/industrial monitoring where anomalies are sparse.
• Propose a CNN-based time series segmentation architecture inspired by U-Net to detect anomalous segments.
• Introduce a transfer learning framework to overcome data sparsity by pretraining on synthetic univariate data.
• Develop MU-Net to transfer univariate pretrained features to multivariate time series tasks.
• Evaluate performance on both univariate and multivariate datasets, including transfer learning scenarios.

Proposed method

• Adapt 1D CNNs into a U-Net-like architecture for time series segmentation with encoder–decoder symmetry and skip connections.
• Handle multivariate inputs by treating channels as separate dimensions and using appropriate last-layer configurations (multi-class/multi-label or nominal class).
• Use soft Dice loss and Adam optimizer for training.
• Apply input normalization with a user-specified magnitude scale; optionally add a per-snapshot normalization layer for dynamic scaling.
• Augment training data with time-series specific augmentations (zooming, trends, reversing, etc.) to improve robustness.
• Develop MU-Net architecture to transfer univariate pretrained weights to multivariate tasks by channel-wise encoding followed by a joint encoding stage and standard decoding stages.

Experimental results

Research questions

• RQ1 Can a CNN-based segmentation model (U-Net) effectively detect anomalous segments in univariate and multivariate time series?
• RQ2 How effective is transfer learning from a large synthetic univariate time series dataset to real-world, smaller multivariate and univariate anomaly detection tasks?
• RQ3 Does MU-Net provide a viable mechanism for transferring learned features from univariate to multivariate time series anomaly detection?
• RQ4 What is the impact of data augmentation and input normalization on detection performance in time series anomaly segmentation?

Key findings

• On Dodgers Loop sensor data (univariate), the model trained on 42 events achieved detection for 81 events with only 3 of 39 test events missed, with some false positives near missing values.
• On Gasoil plant data (multivariate, 19 variables), only 1 of 22 attacks was missed and there were 3 false alarms across 18 test sequences.
• Synthetic univariate curves with unusual shapes showed IoU of 71.95% after univariate transfer learning versus 50.96% from scratch.
• EMG dataset results improved from 56.61% IoU (scratch U-Net) to 64.10% IoU (MU-Net) and 70.04% IoU after transferring pretrained U-Net to MU-Net and fine-tuning.
• The proposed segmentation-based approach enables high-granularity localization of anomalies in streaming data, addressing issues with sliding-window classification methods.