01
Camera AI at the Scale (In Development)
An edge node at the station is being designed to read the licence plate and check it against Vietnamese plate formats, classify the truck into six classes from light truck to tractor unit, count axles from a side camera and read ISO 6346 container numbers. Its vision models are not trained yet.
02
Camera Tamper Detection
A covered or sprayed lens shows up as a sudden brightness drop; a camera turned away shows up as a large difference from the learned background. Either one raises a tamper alert on its own channel.
03
Fraud Signals per Ticket
Twenty features per ticket: tare against the plate's own recent tares, net weight against declared cargo, weighing time against the station's norm for that cargo, off-hours and weekend weighing, how often the truck came today and this week, and whether the same plate was just weighed at another station.
04
Two Models, One Score
An Isolation Forest judges the ticket on its own; an LSTM autoencoder judges it as the next step in that truck's recent sequence of tickets. A weighted blend gives a 0-100 risk score, four risk levels and a recommended action for each.
05
Reasons, Not Just a Number
Every score lists the three features that pushed it up, with the value and the normal range next to it. With a trained model the ranking comes from SHAP; before that, from deviation against baseline.
06
Tiered Alerts & Reviewer Feedback
Medium risk goes to the dashboard only, high risk also to Zalo OA, critical risk also to email and SMS. Reviewers acknowledge alerts and mark each one as true fraud or false positive, and those labels are what the next model is evaluated against.
07
Operator & Station Analytics
Anomaly rate per operator over the last seven days, a heatmap of alerts by hour of day and risk level, and a 24-hour traffic forecast per station with peak hours and a suggested number of operators per shift block.
08
Predictive Scale Maintenance
Loadcell telemetry - zero-tracking drift, calibration age, overload events - is trended per loadcell to project the day it will drift out of tolerance, with a warning well before that day. A drifting scale shortchanges one side of every deal without anyone meaning to.
