Total Engines
—
Registered fleet
FLEET
Operational
—
All nominal
STATUS
Warning
—
Elevated risk
STATUS
Critical
—
Immediate action
STATUS
Avg P10 Risk
—
Fleet failure prob
AGENT 2
Avg RUL
—
Remaining cycles
AGENT 3
Avg Health
—
Fleet health score
AGENT 3
Early Warnings
—
Active warnings
AGENTS
Risk Trajectory Distribution
AGENT 2 · BiLSTM
Degradation Phase Distribution
AGENT 3 · TCN
FL RMSE Trend — Both Agents
FEDERATED LEARNING
Action Source Breakdown
AGENT 2 + 3 + SUPERVISOR
⚡ Engines Requiring Attention
UNIFIED DUAL-AGENT VIEW
| Engine | Aircraft | Fleet | P10 | Trajectory | RUL | CI Range | Phase | Health | Flags |
|---|
Recent Actions (24h)
⬡
AGENT 2 — FAILURE RISK REASONING
BiLSTM + Monte Carlo Dropout · Multi-Horizon P10/P30/P50
Failure Probability Trend — P10 / P30 / P50
Risk Velocity & Acceleration
Trajectory State History
Alert History — False Alarm Rate
◎
AGENT 3 — RUL REASONING & LIFECYCLE INTELLIGENCE
TCN + Monte Carlo Dropout · CI Bands · Self-Correction · Federated Learning
RUL Trajectory with Confidence Interval Band
TCN + MC DROPOUT
Health Score Over Time
Degradation Speed (RUL/cycle)
⚠ Lifecycle Reasoning — Instability Flags
AGENT 3 MODULE 4
| Engine | Cycle | CI Collapse | Sudden Drop | Inconsistency | Total | Model | Phase |
|---|
⟳
LIFECYCLE INTELLIGENCE — FLEET RUL COMPARISON
All Engines · Elbow Detection · Early Warning · Phase Classification
Fleet RUL Comparison
AGENT 3 OUTPUT
Degradation Phase Distribution
FL Performance — RUL Agent RMSE
Complete Fleet Lifecycle Table
| Engine | Type | Cycle | RUL | CI Range | CI Width | Phase | Health | Speed | Elbow | Warning | Flags | Model | P10 |
|---|
⚡ Unified Alert Feed
AGENT 2 + AGENT 3 + SUPERVISOR
Risk Trend — P10/P30/P50
AGENT 2
RUL Trend + CI Band
AGENT 3
False Alarm Rate Trend
EWA SELF-ADAPTATION
FL RMSE — Both Agents
FEDERATED LEARNING
Combined Engine Report — Risk + RUL Overlay
DUAL AGENT
Risk Prediction (Agent 2)
RUL Prediction (Agent 3)
Sensor Data
Raw JSON Feed
API Docs
Submit Risk Prediction — Agent 2
Submit RUL Prediction — Agent 3
Bulk Sensor Data Upload
Raw JSON — Python Agent Feed
API Reference — Dual-Agent Endpoints
BASE: http://localhost/aircraft_maintenance/api/index.php?path=
POST /feed — Risk agent (sensors + BiLSTM prediction)
POST /rul-feed — RUL agent (sensors + TCN prediction)
GET /engines/{id} — Engine + latest risk + RUL
GET /rul?engine_id=X&limit=60 — RUL history (Agent 3)
GET /predictions?engine_id=X — Risk history (Agent 2)
PUT /actions/{id} — Feedback signal → EWA adaptation
GET /analytics?type=rul_trend&engine_id=X&days=30
GET /analytics?type=combined_engine&engine_id=X
GET /analytics?type=fl_performance — Both agents FL RMSE
R26-DS-003 · SLIIT FACULTY OF COMPUTING
AI SYSTEM FOR AUTONOMOUS
PREDICTIVE MAINTENANCE
PREDICTIVE MAINTENANCE
A federated multi-agent AI system targeting SriLankan Airlines' Airbus A330/A320 turbofan fleet.
Two specialized agents collaborate: Agent 2 predicts failure probability; Agent 3 predicts remaining useful life.
Raw sensor data never leaves the facility — protected by Flower FedAvg federated learning.
Two specialized agents collaborate: Agent 2 predicts failure probability; Agent 3 predicts remaining useful life.
Raw sensor data never leaves the facility — protected by Flower FedAvg federated learning.
BiLSTMTCNMonte Carlo DropoutFlower FedAvgEWA Self-AdaptationNASA C-MAPSSMySQLPHP REST API
RESEARCHERS
M B T D Samarasinghe
IT22106056 · Agent 2 (Risk Reasoning)
W.B. Shohani
Agent 3 (RUL Lifecycle Intelligence)
Supervisor: MR. Samadhi Rathnayaka
⬡
AGENT 2 — FAILURE RISK REASONING
IT22106056 · M B T D Samarasinghe
MODULE 01
Multi-Horizon Forecaster
BiLSTM + MC Dropout outputs P10/P30/P50 failure probability across three time horizons with Bayesian confidence intervals.
MODULE 02
Risk Trajectory Reasoning
Computes risk velocity & acceleration from P10 history. Classifies state: Stable → Rising → Accelerating → Critical.
MODULE 03
Agent Memory
Episodic store of 200 risk episodes. Cosine similarity retrieval matches current profile to past failure patterns.
MODULE 04
State Vector Packaging
Assembles standardized output contract for Agent 4 supervisor: trajectory, P10/P30/P50, uncertainty, velocity.
MODULE 05
EWA Self-Adaptation
Exponentially Weighted Average threshold updates from outcome feedback. Targets ≥30% FAR reduction in 50 cycles.
MODULE 06
Federated Learning Client
Flower FedAvg client shares BiLSTM weights only. Raw sensor data never leaves facility. FL accuracy gap target ≤5%.
◎
AGENT 3 — RUL REASONING & LIFECYCLE INTELLIGENCE
W.B. Shohani
MODULE 01
TCN + MC Dropout
Temporal Convolutional Network predicts RUL with 50 MC passes. Outputs point estimate + confidence interval [lower, upper].
MODULE 02
Degradation Trajectory
Tracks degradation speed, acceleration, and elbow point detection (onset of rapid decline). 5-phase classification.
MODULE 03
Agent Memory
Stores 200 RUL episodes per engine. Cosine similarity retrieval enables early warning from historical pattern matching.
MODULE 04
Lifecycle Reasoning Engine
Detects CI collapse, sudden RUL drops, prediction inconsistency. Counts instability flags (0–4) for supervisor signal.
MODULE 05
Self-Correction
Auto-switches to fallback TCN model when instability ≥2 flags. Widens CI bounds. Logs correction reason to database.
MODULE 06
Lifecycle Output Contract
Standardized state vector: RUL + CI + phase + health + flags — consumed by Agent 4 for final maintenance decision.
Agent 2 — Performance Targets
RMSE
≤ 15 (FD001)
Failure probability forecast accuracy
Early Warning
≥ 80%
Failures caught ≥20 cycles ahead
FAR Reduction
≥ 30%
False alarm rate after 50 EWA cycles
Trajectory Accuracy
≥ 85%
State classification vs expert labels
FL Accuracy Gap
≤ 5%
Federated vs centralized RMSE
Agent 3 — Performance Targets
RMSE
≤ 15 (FD001)
RUL point estimate accuracy
CI Coverage
≥ 90%
True RUL falls inside predicted CI
Early Warning
≥ 80%
Failures caught ≥30 cycles ahead
Flag Accuracy
≥ 85%
Instability flag precision
Self-Correction Rate
≥ 75%
Corrections improve accuracy