Part 4 of 11

Tracking and AI

Raw GPS is noisy. This layer turns a stream of position reports into things staff can act on: which zone a unit is in, where it is in its rental, whether its behaviour looks wrong, and when it will be back.

Location ingestion

Each report carries the device ID, a sequence number, timestamp, position, speed, heading, satellite count and the receiver’s own accuracy estimate. Before a fix is allowed to move a unit on the map, it passes these checks in order.

Rejected fixes are kept for diagnostics, never silently dropped. A tracker that suddenly produces many rejects is flagged for inspection.

  1. Authenticate

    Device certificate checked; unknown or revoked trackers are refused.

  2. Validate format

    Schema, ranges and units. Malformed messages go to a quarantine queue.

  3. Remove duplicates and reorder

    Sequence numbers handle resends and fixes arriving late from the store-and-forward buffer.

  4. Sanity check

    Reject impossible speeds, positions on land far from the dock, and timestamps in the future.

  5. Quality and authenticity gate

    Correction state, accuracy estimate and QZNMA authentication set each fix’s weight; a spoofing flag escalates.

  6. Smooth and snap

    A Kalman filter smooths the track; units at the dock snap to their berth.

Grey: raw fixes from a moored unit over 10 minutes Magenta: snapped dock berth, shown to staff

Handling drift near the dock

Piers, moored hulls and buildings bounce satellite signals, so a parked unit can appear to wander 20 to 30 metres. Left alone, that drift would fire false “left dock” alerts all day.

  • QZSS MADOCA-PPP corrections cut drift from metres to decimetres before any filtering
  • Units in the dock zone with no IMU motion are snapped to their assigned berth
  • Leaving the dock needs both GNSS movement and IMU motion, or a check-out
  • Zone edges use hysteresis: an exit is confirmed only after two consecutive fixes beyond the boundary by more than the accuracy radius
  • Geofence sizes are tuned in the pilot against surveyed ground-truth markers

Geofencing

Zones are drawn on the chart in the dashboard and can be changed in minutes, for example to close an area for an event or a jellyfish warning. Each zone produces the events below.

EventFires whenExample
ApproachUnit enters the warning buffer around a restricted zoneJS-04 is within 50 m of the coral reserve and closing
EnterConfirmed inside a zoneUnit inside the swim zone
ExitConfirmed outside a zone it should be inUnit has left the riding zone toward the channel
DwellInside a zone longer than its limitStopped at the recovery point for 10 minutes
Path violationBreaks a rule tied to a zoneAbove the speed limit in the launch corridor
ScheduleA zone rule depends on time of dayRiding zone closes at 7pm; any unit still out is overdue

Trip lifecycle

Every unit is always in exactly one state. Movement moves it along, but the start and end of a rental are always confirmed by staff, so billing never depends on GPS alone.

At dockmoored, in hours Checked outstaff confirmed Launchingin corridor Ridingin riding zone Returningheading to dock Checked instaff confirmed unit cleaned and ready for the next rental check-out tap leaves dock zone enters riding zone heading + time left docked + “Mark returned” Overduetime window passed Offlineno report for 2 min Alarmgeofence, tamper, anomaly from any on-water state

Each state change is stored with its time and position, which gives the rental timeline, route segments, stop durations and the trip report without extra work from staff.

AI anomaly detection

Rules catch what we can define in advance. The anomaly model catches what we can’t: movement that is unusual for this fleet, at this beach, at this time of day.

SignalWhat it looks like
Teleport jumpPosition leaps further than any jet ski could travel between fixes. Usually a bad fix; repeated jumps, or a failed QZNMA check, suggest interference or spoofing.
Speed spikeSustained speed well above the fleet norm for first-time riders, or acceleration a rental unit can’t produce.
Towing signatureSteady low speed with ignition off and little hull vibration.
Out-of-window movementAny movement outside operating hours, or a unit leaving with no active rental.
Erratic patternRepeated tight circles near swimmers or boats, a sign of reckless riding.
Unusual stillnessLong stop mid-bay with no recovery request: breakdown or rider in difficulty.

How the model works

Features are computed over a rolling window for every unit: speed and acceleration, distance between fixes versus the distance physically possible, heading variance, ignition and vibration state, zone history and time of day. An isolation forest trained on the site’s own normal riding gives each window an anomaly score from 0 to 1.

Every score comes with its top contributing factors, so a supervisor sees “speed 78 km/h, fleet norm 42” rather than a bare number. When staff mark an alert as false, that feedback is used to retune thresholds each week.

Practical first. Phase 1 runs on rules alone. The model starts in shadow mode during the pilot, learning the site’s normal patterns before it is allowed to raise alerts in Phase 3.

Try the anomaly score

A simplified version of the scoring used in the platform. Change the readings for one 10-second window and see how the score and the resulting action change.

The production model weighs many more features and learns each site’s baseline. This explorer uses fixed weights for illustration.

Ignition

0.00Normal

Speed above norm0
Impossible jump0
Out of hours0
Towing signature0

Predicted return time

Knowing a unit will be back at 3:42pm rather than “around 3:40” lets staff brief and kit up the next customer in advance, and spot a late return before it happens.

A gradient-boosted model trained on the site’s past trips predicts arrival at the dock, with a confidence range. It improves as the history grows; until there is enough data, a distance-and-speed estimate is used.

Inputs

Time left on the rental, distance and heading to the dock, current speed, how this rider has ridden so far, wind and sea state, time of day, and how similar trips ended.

Outputs

Predicted arrival time, a range such as plus or minus 2 minutes, and the probability of a late return.

Actions

If late return is likely, dock staff get a nudge 5 minutes before the rental ends, and the rider buzzer (if fitted) signals “return now”.

Rules engine and alert routing

Deterministic rules and the anomaly score feed one decision step. It sets severity, chooses who is told and how, and stops the same alert from repeating.

SituationSeverityWho is toldChannel
Approaching a restricted zoneWarningDock staff on shiftApp push, rider buzzer
Entered swim zone or restricted areaCriticalDock staff, supervisor, patrol boatPush, WhatsApp, SMS
Left the riding zoneWarningDock staffApp push
Overdue by 5 minutesWarningDock staffApp push
Overdue by 15 minutesCriticalSupervisorWhatsApp
Over the corridor speed limitInfoLogged on the trip reportDashboard
Immobile on the water 3 minutesWarningDock staff, patrol boatApp push
Movement after hours or towing signatureCriticalOwner, supervisor, night securityWhatsApp, SMS, call
Tamper: case opened or power cutCriticalSupervisor, ownerWhatsApp, SMS
Anomaly score 0.6 to 0.8WarningDock staffApp push
Anomaly score 0.8 or aboveCriticalSupervisorWhatsApp, SMS

Critical alerts need acknowledgement in the app. If nobody acknowledges within 2 minutes, the alert escalates to the next person on the rota. Repeats of the same alert for the same unit are suppressed for 2 minutes so phones don’t flood during an incident.