<p>Business Process Management (BPM) plays a crucial role in overseeing organizational operations. BPM solutions generate vast quantities of data in the form of process event logs. These logs act as digital footprints of real-world activities. As deviations occur, they are captured inside event logs. Given that these anomalies can range anywhere from simple operational mishaps to fraud, and may lead to significant monetary loss. Consequently, the detection, analysis, and prevention of these instances are of utmost importance. In this paper, we introduce SPECTRE, the Spectral-channel Parallel Encoder with Contextual Attention for Reconstruction and Enhancement, an unsupervised and multi-purpose autoencoder framework developed for decontamination and anomaly detection in business process traces by reconstructing clean sequences from contaminated event logs. Our study comprehensively evaluates the proposed approach on 11 datasets: 7 public and 4 private. To push the model further on datasets with scarce traces, we employ a data augmentation method referred to as Reduction Window. Experimental results validate SPECTRE’s effectiveness, achieving up to 0.9994 in decontamination accuracy and an F1-score of 0.996 in event-level anomaly detection. This is achieved while preserving near-linear scalability in processing large volumes of events. Furthermore, Reduction Window approach noticeably boosted performance, yielding up to 2% improvement in decontamination accuracy and up to 45% gain in anomaly detection performance on challenging logs.</p>