Predictive Maintenance
Aircraft engine failures are costly and dangerous. Maintenance schedules today rely on fixed intervals or reactive repairs, leaving operators blind to each engine's true condition until failure is imminent. Our transformer forecaster ingests multivariate sensor time series and predicts Remaining Useful Life (RUL) with calibrated uncertainty bounds, enabling operators to replace engines based on actual degradation rather than arbitrary schedules.
Engine-Level Forecast Explorer
Inspect RUL forecasts across 8 reliable test engines from the NASA C-MAPSS FD001 dataset. Select an engine from the dropdown to view its performance and forecast visualisation.
Engine 017
126 windows · Conformal ±11.4 cycles
Model Validation Metrics
Summary metrics compiled from out-of-sample evaluations across 8 reliable test engines with quality forecast samples, from the NASA C-MAPSS FD001 benchmark.
Average RUL prediction error across reliable engines
Proportion of actuals within the P10–P90 band
Average uncertainty band width for RUL
RMSE across reliable engines
With quality forecast samples
Across NASA C-MAPSS FD001
From 21 total measurements
Conformal calibration offset applied
Forecast Horizon & Calibration
How RUL forecast accuracy evolves across the 30-step prediction horizon, alongside quantile calibration analysis comparing expected versus empirical coverage.
Transformer Quantile Forecaster
Our architecture processes 40-cycle windows of 19 sensor signals through transformer encoder layers with multi-head self-attention, producing 3 quantile outputs (P10, P50, P90) with conformal calibration for reliable uncertainty quantification.
Patch Embedding
40-cycle windows of 19 selected sensor measurements (temperatures, pressures, fan speeds) are projected into patch embeddings through learned linear projections with sinusoidal positional encodings.
Transformer Encoder
Multi-layer transformer encoder with multi-head self-attention processes temporal dependencies across the window, capturing long-range degradation patterns in engine performance.
Quantile Prediction Head
MLP (192→64→3) produces 3 quantile outputs (P10, P50, P90) trained with pinball loss, optimising for interval calibration over point accuracy.
Conformal Calibration
Validation-set residuals compute a ±11.4 cycle calibration margin applied to widen raw quantile intervals, achieving target coverage reliability.
196 Engines, 21 Sensors
Trained on 80 engines (80%) with 20 held out for validation and 96 for testing. The model uses AdamW optimiser with cosine annealing, gradient clipping, and mixed precision.
Dataset Composition
NASA C-MAPSS FD001 uses 100 training engines (80 train, 20 validation) and 100 test engines (96 retained after windowing). 21 sensor signals sampled per cycle.
Training Configuration
AdamW optimiser (lr=1e-4) with cosine annealing + restarts, 5% linear warmup, gradient clipping at 1.0, mixed precision, trained for 20 epochs with early stopping.
Context & Prediction Windows
Sliding window sampling with 40 cycles of context to predict RUL at next cycle. 13,659 training windows, 3,072 validation windows, 9,211 test windows.
Performance Context
The pinball-loss regime optimises for quantile calibration over point accuracy. PICP reaches 78.8% against a 90% target, with engines having more data showing stronger coverage.
Incorporate predictions
API access for pilot partners is planned as the platform matures.
curl -X POST https://api.purnabodha.ai/v1/predict/engine \
-H "Authorization: Bearer $API_TOKEN" \
-H "Content-Type: application/json" \
-d '{
"engine_id": 62,
"context_length": 40,
"quantiles": [0.10, 0.50, 0.90]
}'Where We're Taking This Next
Every deployed system raises new questions. These research programs explore the next generation of capabilities for predictive maintenance.
Large-Scale Representation Learning for Industrial Telemetry
Pretrained on diverse sensor modalities, adaptable to new equipment types and operating conditions with minimal fine-tuning data.
Multimodal Maintenance Intelligence
Integrating vibration, thermal, acoustic, and operational logs for holistic equipment health assessment across disparate sensor networks.
Zero-Shot Fault Diagnosis
Identifying novel failure modes without prior examples, enabling early detection of unforeseen degradation patterns in field-deployed fleets.
Continual Learning for Evolving Fleets
Adapting to fleet-wide degradation trends and equipment refresh cycles without catastrophic forgetting of previously learned failure modes.
Interested in piloting this on your fleet?
We're looking for aerospace operators, maintenance providers, and industrial IoT teams interested in piloting this approach on real fleet data.