Based on 3 predictive interventions × $42,000 per event
Active scenario
Flow A — gradual degradation
Vessels monitored
3
All online
Active alerts
2
1 urgent · 1 watch
False alarms suppressed
1
Model filtered
Avg oil health score
68
↓ Down from 74
Parameter trend — MV Horizon Star
Current value
98 cSt
Watch threshold
> 80 cSt
Replace threshold
> 90 cSt
Actual
Forecast
Why is VSL-001 flagged? — model explanation
Sensor readings — all vessels
Live feed · updates every 6 hours
5 sensors reporting
|
Last updated: just now
|
Update frequency: every 6 hours (4×/day per vessel)
|
Protocol: Modbus TCP → MQTT broker
|
Next simulated reading in: 12s
Demo mode: simulating new readings every ~12 seconds to illustrate the 6-hourly production cadence. Toggle off to freeze the feed.
Sensor ID
Vessel ID
Vessel name
Timestamp
Viscosity
Temp °C
TAN
Fe ppm
Eng hrs
Scenario
Status
Select a day
Su
Mo
Tu
We
Th
Fr
Sa
Selected
Has data
No data
Filter by vessel
Filter by status
Select a day on the calendar to view readings
Select a month
Bold = months with data
Filter by vessel
Select a month to view aggregated daily readings
Select a year
Bold = years with data
Filter by vessel
Select a year to view monthly aggregated data
Drop a CSV file here or click to upload
Supported: .csv · Max 50 MB
Add a manual sensor reading
Inject a single reading into the live data stream to trigger a specific event on demand during a walkthrough.
Bulk CSV upload
Upload a pre-built scenario CSV file to replay across the dashboard. Uploaded files also appear in Data — CSV files.
Training data — future retrain dataset v2 dataset
Upload additional sensor readings to build the second training dataset (viscqai_retrain_v2). Files tagged here are reserved exclusively for model retraining — they do not appear in the live feed or CSV files view.
What makes good retrain data: CSV with the same column schema as the Phase 1 training data (vessel_id, timestamp, viscosity_cst, temperature_c, tan, fe_ppm, cu_ppm, water_pct, base_number, engine_hours, label). Label column is optional — the retrain agent will auto-label using the existing model before training.
Dataset v2 will be used automatically by the retrain agent when drift is detected, or you can select it manually in the Models page.
Oil degradation classifier
Gradient Boosting (XGBoost). Predicts whether oil will require replacement within 14 days. Trained on 5 vessels × 18 months of synthetic data across all three scenario types.
F1 score0.91
Precision0.93
Recall0.89
Last trainedToday
Model confidence
Remaining useful life — regression model
LightGBM regressor. Predicts days-to-replacement as a continuous output. Directly feeds the day-count chips on vessel cards.
MAE1.8 days
RMSE2.4 days
R²0.94
Last trainedToday
Regression accuracy
False alarm detector — anomaly model
Isolation Forest trained on normal-operation data only. Suppresses single-parameter spikes when cross-parameter signals remain stable — powers the alert suppression layer.
Catch rate89%
Contamination0.05
Last trainedToday
Suppression accuracy
Reference — marine lubricant brands in scope
ViscQ.ai's thresholds are aligned to general OEM service guidance for the engine classes monitored, not to a single proprietary oil chemistry. For context, here is how three widely used marine lubricant brands position their products against the engines in this fleet.
Brand
Product line
Typical viscosity grade
Positioned for
Shell
Rimula / Alexia
SAE 40 / 15W-40
Medium- and low-speed trunk piston engines
ExxonMobil
Mobilgard
SAE 30 / 40
Two-stroke crosshead and trunk piston engines
Chevron
Delo Marine / Taro
SAE 30 / 40
Medium-speed diesel and dual-fuel engines
This table reflects each brand's publicly marketed product positioning, not a lab-validated comparison. ViscQ.ai's actual replacement thresholds (Section above) are based on general OEM service-interval guidance and require SME sign-off before being presented as technically authoritative to any specific client fleet.
Selected: Flow A — gradual degradation
Replay speed
Adjust how fast synthetic data streams into the dashboard. Slow is best for live walkthroughs with visitors.
SlowFast1×
VR
Arjun M.
Product Manager · ViscQ Labs
Account details
Demo configuration
⏳ Pending approval: MV Horizon Star job is complete — approve below or wait for auto-approval (30s).
Auto-approving in 30 seconds — or approve now to immediately confirm vessel recovery.
✅ VSL-001 recovered and showing Healthy. To run the full demo loop again:
MV Pacific Endurance
VSL-002 · Job #J-2024-046 · Wärtsilä 6RT-flex50
In progress
📅 Scheduled: Jul 4, 2024 · 13:00–15:00📍 Port: Port Klang, Malaysia👷 Crew: Anton Lebediev, Priya Krishnan🟠 Priority: High
Checklist:6/10 completed2 photos submitted
MV Horizon Star
VSL-001 · Job #J-2024-041 · MAN B&W 6G60ME-C9
Approved ✓
📅 Completed: Jun 18, 2024📍 Port: Colombo, Sri Lanka👷 Crew: Rajan Subramaniam✅ Approved by: Demo admin
VSL-001 returned to Healthy status after this job. Cost avoidance recorded: +$42,000.
Demo admin panel
Manage attendees, service team, notifications and the retrain agent
Push active
Unlocked — click to lock
Demo attendees
Alert · Retrain · Recovery
Receive real-time notifications during the demo — alerts, false alarms, retraining events, recovery confirmations.
SK
Suresh Kumar
suresh.kumar@shipping.com
Subscribed
AV
Anika Vogt
a.vogt@maritimegroup.de
Pending
Add attendee
Service team
Schedule · Jobs · Recovery
Receive job assignment emails + push notifications. Access crew job card at /crew?job=ID to log action taken.
RS
Rajan Subramaniam
r.subramaniam@viscqai.com
Chief Engineer
MT
Mohammed Tahir
m.tahir@viscqai.com
2nd Engineer
PK
Priya Krishnan
p.krishnan@viscqai.com
Engine Room A
AL
Anton Lebediev
a.lebediev@viscqai.com
Engine Room B
Add team member
Retrain agent
Running in Phase 4 ML service · polls every 60s · triggers on drift, new readings, schedule, or force
Primary trigger
Drift score
0.81 vs 0.92 baseline
New readings
47 / 50
Approaching threshold
Last retrain
3h 12m ago
F1: 0.91 → v1
Next scheduled
20h 48m
24h watchdog
Notification settings
Email provider
Send notifications for
Push notifications
Push requires attendees/crew to open the demo URL in their browser and click Allow when prompted. Works on Chrome, Firefox, Edge. iOS Safari requires PWA install.
📝 Supervisor notes: Ensure OEM-grade 40W marine lubricant pre-ordered. Coordinate with port agent for docking slot. Use Shell Rimula R4X if Mobilgard is unavailable.
3
Done
7
Remaining
10
Total steps
Replace photo
Engine room — pre-inspection
Replace photo
Oil drain valve — open
Replace photo
Oil filter — before replacement
Replace photo
Fresh lubricant — batch label
Add photo
Supervisor notes: Ensure OEM-grade 40W marine lubricant pre-ordered. Coordinate with port agent for docking slot.
Field notes
Admin access
Enter the 4-digit demo PIN, or use ?admin=1 in the URL as a developer shortcut.
Demo PIN: 1234
Schedule oil replacement — MV Horizon Star (VSL-001)
MV Horizon Star
VSL-001 · Engine A · Replace within 7 days
Priority
Select date
Su
Mo
Tu
We
Th
Fr
Sa
Selected: Not selected
Time slot
Estimated duration
Port / location
Assign to crew
Notes / instructions
Booking summary
Complete the form above to preview the booking summary.
Viscosity spiked to 96 cSt on Apr 11. Anomaly model cross-checked all parameters — temperature and TAN both stable. Classified as sensor glitch. Alert suppressed. No action taken.
Spike value
96 cSt
Confidence
89%
Duration
2 readings
Timestamp
Viscosity
Temp °C
TAN
Fe ppm
Anomaly score
Decision
Apr 11 12:10
83.2
77.9
1.9
26
0.12
Normal
Apr 11 13:10
96.0
77.8
1.9
26
0.81
Suppressed
Apr 11 14:10
94.1
77.9
2.0
27
0.76
Suppressed
Apr 11 15:10
84.0
78.0
2.0
27
0.14
Normal
Apr 11 16:10
83.8
78.1
2.0
27
0.11
Normal
Why suppressed: Viscosity spiked while temperature, TAN, and Fe particles all held steady. Genuine degradation creates correlated multi-parameter signals. A single-parameter spike is a characteristic sensor fault signature.
Viscosity spike — Apr 11 hourly readings
Viscosity
Suppressed window
Re-train model
The model will be retrained on the selected dataset. This typically takes 30–90 seconds.
Training dataset
What happens during retraining: 1. Dataset loaded from models/ (Phase 4 ML service) 2. Features re-engineered (rolling averages, rate-of-change, cross-parameter terms) 3. Model trained with pre-configured hyperparameters 4. Evaluation metrics updated 5. New model artefact saved and hot-swapped — no downtime
Loading dataset…
Confusion matrix — oil degradation classifier
Results on the held-out test set (20% of synthetic data, stratified across all scenario types).
Predicted: Replace
Predicted: OK
Actual: Replace
142True positive
16False negative
Actual: OK
11False positive
231True negative
Precision
0.93
Recall
0.89
Accuracy
0.94
The model correctly identified 142 of 158 genuine replacement cases (89% recall). 11 false positives means 11 unnecessary alerts — acceptable since the cost of a missed replacement far exceeds an unnecessary one.
Residual analysis — remaining useful life model
Residual = predicted days − actual days. A well-behaved model has residuals clustered near zero with no systematic bias.
MAE
1.8d
RMSE
2.4d
R²
0.94
0–7 days
-1.2d
8–14 days
-0.8d
15–21 days
+0.4d
22–30 days
+1.1d
30+ days
+2.1d
The model slightly under-predicts remaining life when oil is near end-of-life (0–14 days) — the conservative and safe direction. For longer horizons (30+ days) there is modest over-estimation of +2.1 days on average, which is acceptable.
Adjust false alarm threshold — anomaly model
The contamination parameter controls how aggressively the Isolation Forest flags anomalies. Lower = more conservative. Higher = more aggressive suppression.
ConservativeAggressive0.05
Catch rate
89%
False suppression
3%
Net reduction
86%
Current setting (0.05) is well-balanced for a demo environment. Good for live walkthroughs.
