Time Series Anomaly Detection in Radio Test

Sammanfattning: Production tests evaluate products with automated systems enabling swift data collection, while anomaly detection in the gathered data is widely employed in industries for damage prediction and issue prevention. Ericsson, a leader in the telecommunications industry, has a Temperature Quality Test (TQT) platform which involves precise performance measurements on radios, gathering abundant data in both single value and time series formats to evaluate and improve tested products. This study focuses on identifying unusual patterns within time series data obtained from four TQT test measurements, utilizing the Isolation Forest algorithm as the prediction tool. Beyond the algorithm, our approach involves careful feature extraction and selection, capturing essential information which is crucial for detecting atypical patterns. The feature extraction covers various time series data features, including three categories: statistical, temporal and spectral features, which represent data behavior from different angles. Feature selection is driven by domain expertise and filtering techniques. Utilizing these refined features, the Isolation Forest algorithm effectively identifies anomalies, capturing deviations such as sudden shifts, distinct trends or pronounced peaks. The overall prediction model performance is assessed by five evaluation metrics. For example, the F1 score yields values of 0.80, 0.94, 0.95, and 1.00 for the four time series measurements.