Industrial Anomaly Detecting in Spectral Data with Energy Consistency

Abstract

Anomaly detection is critically important in ensuring the required stability and reliability of industrial equipment. Conventional methods for it typically assume the availability of sufficient samples from both normal and faulty operating conditions. In practice, however, real-world industrial scenarios rarely provide adequate abnormal samples, and fault data are extremely scarce. This scarcity makes it challenging to train robust and generalizable anomaly detection models. Although many studies have attempted to resolve this challenge by generating artificial anomaly samples, such synthetic samples often deviate from genuine fault patterns. As a result, models trained on them may overfit to unrealistic features and lower their generalization capability in practical applications. To address such issue, we propose Spectral-Regularized Denoising AutoEncoder (SR-DAE), that is designed for anomaly detection in high-dimensional spectral signals under anomaly-scarce conditions. Specifically, it introduces a bandstructured spectral regularization strategy that guides the reconstruction process to preserve the coarse-grained energy distribution across different frequency sub-bands. SR-DAE improves sensitivity to subtle abnormal patterns while enhancing overall detection performance. We evaluate it on two turbocharger vibration datasets and compare it with several representative baseline methods. Experimental results demonstrate that SR-DAE outperforms its peers in terms of several commonly-used evaluation metrics. It can well balance between reducing false alarms and accurately identifying anomalous samples.