Self Learning Anomaly Detection on a Multidimensional Feature Space

Early detection of material degradation is critical for ensuring structural reliability. The majority of conventional evaluation methods require both normal and defective reference samples. This increases the effort needed for calibration since producing defective samples is challenging, costly, and impractical. Therefore, in this work, an algorithm for learning the normal behavior directly from the data, without requiring predefined anomalous samples was developed. A self-learning anomaly detection algorithm is implemented. It is based on a dense autoencoder trained solely on micromagnetic data representing frequently occurring conditions. The autoencoder learns a data-driven representation of the undamaged state and detects deviations during the measurement. A dynamic thresholding strategy is applied to adapt to changes in the feature distribution, enabling the detection of early anomalies without manual limits. The results show that the method can detect subtle variations in micromagnetic responses before any visible failure or load reduction occurs. This capability enables continuous monitoring and allows the approach to adapt to evolving material behavior.