An Improved ARIMA Based Anomaly Detection Method for Time Series Data

Currently, IoT data processing standards in China are somewhat vague, and monitoring data is typically used in subsequent stages for analysis, prediction modeling, and applications. However, monitoring and control data are often affected by various factors that can introduce anomalies, which in turn impacts intelligent control systems. To improve the quality of monitoring data, this paper addresses the limitations of traditional anomaly detection methods, which often require high similarity in data and face challenges with highly diverse monitoring data. Additionally, traditional ARIMA models are limited by difficulties in parameter selection, threshold determination, and high false-positive rates. Thus, we propose an improved ARIMA-based anomaly detection method for time-series data. This approach encompasses the complete process of monitoring data fitting and prediction, anomaly detection, and anomaly validation. The method has three main features:1.Optimized Parameter Selection: For varied batches of monitoring data, this method combines ACF and PACF plots with AIC and BIC heatmaps to determine optimal ARIMA parameters for each data batch.2.Incorporation of a Sliding Window: A sliding window is used to analyze residual data characteristics within the window, enabling precise anomaly detection for individual data points.3.Anomaly Validation: A K-means clustering algorithm is incorporated to validate initial anomaly detection results.The method was verified in a real-world cigar leaf curing barn, and experimental results demonstrate that it effectively detects anomalies in the IoT monitoring system data, contributing to improved IoT data quality.