Use case · Predictive maintenance
Autonomous AI predictive maintenance — RUL, failure probability, anomaly, end-to-end.
Drop your sensor history, telemetry, or maintenance log. OctOpus picks the right framing — RUL regression, fixed-window failure classifier, or anomaly detection — runs the experiments, validates on real out-of-asset holdout, and deploys a scoring endpoint your SCADA, MES, or fleet ops can call.
TL;DR. Predictive maintenance is a multi-modal ML problem: sometimes regression (RUL), sometimes classification (will it fail in 7 days?), sometimes anomaly detection (is this asset drifting?). OctOpus picks the right framing per dataset, rotates through GBMs, NeuralForecast, foundation models, and autoencoders, and ships an inspectable train.py and a low-latency scoring endpoint.
Predictive-maintenance problems OctOpus handles well
- Remaining useful life (RUL) — regression with right-censoring, survival loss when censoring is significant.
- Failure-window classifier — will this asset fail in the next N hours / days / weeks?
- Condition-based monitoring — anomaly score per asset over rolling windows of telemetry.
- Fleet RUL with hierarchies — per-asset RUL inside per-line, per-plant, per-region hierarchies.
- Energy / consumption drift — sensor drift detection on motors, pumps, HVAC, compressors.
- Vehicle / fleet telematics — engine, brake, battery, drivetrain prognostics from CAN bus.
Models the agent rotates through
| Tier | Family | When the agent picks it |
|---|---|---|
| 1 · Baseline | LightGBM with lag, rolling, FFT/spectral features (regression or classification) | Almost always — fast and strong on engineered sensor features. |
| 2 · Tuned GBM | XGBoost / CatBoost with Optuna, time-based CV | When tier 1 has headroom on RUL MAE or failure recall. |
| 3 · Deep / modern | NeuralForecast (NBEATS, PatchTST, TFT), LSTM-based RUL regression | Long sequences, multi-channel telemetry, multi-asset panels. |
| 4 · Foundation | Chronos, TiRex, TimesFM (zero-shot anomaly + forecast) | Cold-start sensors with no labeled failures. |
| 5 · Anomaly | IsolationForest, autoencoder reconstruction | Unsupervised drift or rare-fault detection where labels are scarce. |
How a predictive-maintenance run looks
- Profile. Detects time column, asset ID column, sample rate, sensor channels, failure-label column (if any), and decides the framing — regression, classification, or anomaly.
- Plan. Writes a research spec: MAE / RMSE / quantile loss for RUL, PR-AUC / recall-at-FPR for failure classification, reconstruction error for anomaly. Lag, rolling, and FFT feature pipeline.
- Run. Generates a fresh
train.pyper experiment, executes in sandbox, captures per-asset error. - Diagnose. When something fails (sensor-channel NaN cascade, sampling-rate mismatch, censored-data math error), the agent writes a targeted fix and retries.
- Validate. Out-of-asset holdout the LLM never sees — guards against per-asset overfit.
- Deploy. Low-latency scoring endpoint, or a deploy bundle for edge / on-prem inference inside your plant.
What enterprise operations and reliability engineering get back
- RUL estimate per asset, with confidence bounds.
- Failure probability per asset within a configurable window.
- Per-asset anomaly score over rolling windows of telemetry.
- Feature importance per channel — which sensors actually drove the prediction.
- The exact
train.pythe agent wrote — fully inspectable, edge-deployable. - Deploy bundle for SCADA, MES, fleet ops, or any custom inference stack.
Edge, air-gap, and on-prem
OctOpus Desktop trains entirely on-device — no outbound connections required beyond your chosen LLM endpoint. The Enterprise VPC deployment runs alongside plant historians (PI System, Aspen, Wonderware, OSIsoft) and pushes scores to SCADA / MES. Air-gapped operation is supported when paired with a locally hosted LLM. See Enterprise for deployment details.