Industrial equipment energy consumption anomaly detection and optimization based on vision transformer

We present an integrated framework for multivariate energy time series that unifies anomaly detection and executable scheduling. First, a standardized preprocessing pipeline enforces cross-channel consistency and min–max normalization before Gramian Angular Field (GAF) imaging, yielding physically coherent time-to-image representations. Next, a Vision Transformer (ViT) detector with discrimination –reconstruction co-training outputs calibrated anomaly scores while regularizing features via a reconstruction head, improving robustness to weak and spurious anomalies and enabling visual interpretability. Finally, a differentiable weight mapping injects the scores into an optimization module that penalizes energy-usage deviations relative to baselines, embeds total-variation smoothing to avoid oscillatory set-points, and enforces production/safety constraints. This closes the perception-to-action loop: detected events trigger cost-aware, constraint-respecting schedule adjustments rather than isolated alarms. Experiments on real-world and public benchmarks demonstrate consistent gains over strong time-series and imaging baselines in AUROC/F1 and downstream value (peak shaving and cost reduction), with sensitivity analyses covering windowing, imaging hyperparameters, and score calibration. The framework offers a reproducible, industry-oriented path from detection to decision.