Anomaly Detection in Time Series Data using Data-Centric AI

Abstract

Detecting the anomalous data points in the time-series data is a crucial task in most of the industrial applications where time is a key component. As time-series data is used for forecasting/predicting the values, building a most accurate model is essential. If the input data consists of anomalies, then the model fails to perform well and so does the future prediction. The conventional method of building a good predictive model suggests to improve the model performance by applying regularization techniques, performing feature engineering or by experimenting with different combinations of activation functions and/or loss functions along with number of neurons and hidden layers in a neural network. But, such a model-centric approach fails miserably in real-time applications. Data-centric approach where the input data itself must be updated and corrected is a novel technique in solving the issues faced by model-centric approach. This paper proposes a technique of using data-centric approach to detect anomalies in time series data. Several models using model-centric approach are demonstrated and proved to be underperforming with high False Negatives. Whereas, the data-centric approach proved to achieve 100% performance in correctly identifying the anomalous data points.