Time-series anomaly detection is vital for analyzing industrial and biomedical sensor data. Its proper use is crucial for maintaining system reliability and monitoring health. Traditional unsupervised anomaly detection methods often exhibit degraded performance when the training data is contaminated with anomalies. To address this limitation, active learning approaches have been explored, but they are frequently constrained by simplistic query strategies that hinder the selection of high-value samples for labeling. Motivated by the preceding observations, this paper introduces a novel multi-scoring active learning (MSAL) framework specifically developed to achieve highly robust anomaly detection in time-series data. The model is first trained within an active-learning paradigm on a small, high-quality labeled subset to avoid the weaknesses of fully unsupervised learning.  Next, a composite multi-metric scoring mechanism identifies and queries highly informative samples for manual annotation, thereby refining the dataset iteratively. The proposed method was evaluated on three benchmark datasets spanning biological and sensor signals. Experimental results indicate that the introduced framework MSAL achieves state-of-the-art performance compared to both conventional unsupervised and competing active learning methods, with its superiority being particularly significant in scenarios with a high proportion of anomalies.

time series anomaly detection, active learning, transformer, biomedical signals, sensor measurement