TCN for Anomaly Detection in TS (2019,34)Permalink

ContentsPermalink

1. Abstract
2. Introduction
3. Convolutional Sequence Model
   1. TCN
   2. Dilated Convolutions
   3. Residual Connections
   4. Multi-scale Feature Maps for Prediction
4. Anomaly Detection

AbstractPermalink

TCN = causal convolutions & dilations

$\rightarrow$ apply TCN for “anomaly detection”

Steps

• 1) apply TCN to predict trend

• 2) prediction errors are fitted by Multivariate Gaussian distribution &

  used to calculate the anomaly scores

1. IntroductionPermalink

propose anomaly detection algorithm in unsupervised way

• 1) TCN : predictor model
• 2) Multivariate Gaussian : identify anomaly points in TS

2. Convolutional Sequence ModelPermalink

Notation

• $X$ : time series
• $x_{t} \in R^{m}$ : each time point

Prediction model

• predict next $l$ values, with window of length $L$

Residuals between

• 1) prediction values
• 2) real values

are calculated & fit a Multivariate Gaussian distribution model

(1) TCNPermalink

2 constraints

• 1) output of the network should have the same length as input
• 2) can only use information of past time steps

$\rightarrow$ zero padding & no cheating

(2) Dilated ConvolutionsPermalink

to memorize LONG term

( should make larger receptive field)

(3) Residual ConnectionsPermalink

• pass

(4) Multi-scale Feature Maps for PredictionPermalink

to capture different scale patterns

• not only one last layer,
• but also use multiple layers!

3. Anomaly DetectionPermalink

implemented in point-wise

prediction errors distribution on training data, is modeled with Multivariate Gaussian

Anomaly score :

• $e_t$ : observation prediction error

$\rightarrow$ $x^{(t)}$ is classified as “anomalous”, if $a_t > \tau$