How do you implement offline anomaly detection in industrial IoT settings?

Industrial offline anomaly detection uses lightweight models deployed on edge gateways or PLCs. Architecture: collect sensor data (vibration, temperature, current, pressure) at high frequency, apply signal processing (FFT, wavelet transforms) for feature extraction, run a compact autoencoder or isolation forest for anomaly scoring. Model choices by complexity: statistical methods (Z-score, IQR) for simple thresholds, One-Class SVM for moderate complexity, LSTM autoencoders for temporal patterns, and transformer-based models for multi-sensor correlation. Deployment: quantize models to INT8, implement sliding window processing, maintain a rolling baseline of normal behavior, and use adaptive thresholds that account for seasonal/operational variations. Alert management: implement a severity hierarchy (warning, critical, emergency), include cooldown periods to prevent alert storms, and log all detections with sensor context for root cause analysis. Typical systems process 100+ sensors at 1kHz with sub-millisecond inference.