SmartScale AI - Weighing Station Anti-Fraud
Every closed weigh ticket scored for fraud against the truck's, the operator's and the station's own history, with a camera edge module for the scale in development
Overview
SmartScale AI is the anti-fraud layer for truck weighing stations. For the station side, an edge node on an NVIDIA Jetson is in development: it is designed to classify the truck and count its axles with YOLOv8, read the licence plate and container number with PaddleOCR, and notice a covered or turned camera; its vision models are not trained yet and the MQTT link to the cloud is still to be wired.
When a ticket closes, the backend builds twenty features from it - tare against the plate's recent tares, net against declared cargo, weighing time against the station's norm, off-hours, visit frequency, the same plate at another station, the operator's recent anomaly rate - using rolling baselines kept in Redis.
An Isolation Forest scores the ticket on its own and an LSTM autoencoder scores it as part of the truck's recent sequence; the blend becomes a 0-100 risk score with the three reasons behind it. Alerts are routed by risk level to the dashboard, Zalo OA, email and SMS, and reviewer verdicts are stored as labels that the next training run uses to calibrate and evaluate the model.
ClickHouse holds tickets, scores and loadcell telemetry, which also feed operator statistics, a per-station traffic forecast and loadcell drift warnings.