Anomaly detection models based on autoencoders often suffer from unstable training, inconsistent reproducibility, and sensitivity to manually selected hyperparameters. These challenges limit their reliability across domains with varying data scales and feature distributions. This paper proposes an adaptive anomaly detection framework that combines a dual autoencoder architecture with a fully automated training-stability pipeline. The dual encoders capture global structural patterns and fine-grained local variations, enabling robust detection of subtle and heterogeneous anomalies. Unlike existing automated approaches which focus primarily on architecture search, our method targets the overlooked problem of training reproducibility and dataset-adaptive optimization. The proposed framework integrates (i) dynamic seed initialization for bias-resistant reproducibility, (ii) adaptive batch size estimation through a statistically derived logarithmic scaling function, (iii) iterative epoch optimization with multi-attempt checkpointing, and (iv) automated dropout tuning via Keras Tuner. To improve decision reliability, we further incorporate Monte Carlo dropout to estimate predictive uncertainty and reduce false-positive classifications. Extensive experiments on medical imaging, industrial inspection, 2D/3D textures, and surveillance datasets demonstrate consistent improvements over 11 state-of-the-art anomaly detection methods, yielding AUROC gains of up to 12.4% and significantly higher recall–precision stability. Ablation studies confirm that each optimization component contributes meaningfully to performance and reproducibility. • The framework eliminates manual tuning by integrating dynamic seed initialization, empirical batch size estimation, and attempt-level early stopping with model checkpointing. • Incorporates input gradient heatmaps and reconstruction error visualizations to explain model predictions, helping domain experts understand anomaly decisions. • Uses a min-distance ROC method to automatically determine optimal classification thresholds. • Evaluated on diverse datasets from medical, industrial, and surveillance domains. • Achieves better or comparable results to existing methods, while being more reproducible, scalable, and less sensitive to hyperparameter selection.
Pillai et al. (2026) studied this question.