refineryguard-lstm-ae

RefineryGuard LSTM-AE

CPU-friendly process-anomaly bundle for the Tennessee Eastman Process:

• lstm_ae.onnx — LSTM autoencoder (reconstruction error = anomaly score)
• isolation_forest.joblib — window mean/std baseline
• pca.joblib — 90% variance reconstruction baseline (operational primary detector)
• fault_classifier.joblib — 22-class TEP fault id (HistGradientBoosting)
• scaler.joblib / normal_stats.joblib / thresholds.joblib
• eval_results.json — held-out Braatz test protocol

Data honesty

Trained only on TEP simulation (Downs & Vogel 1993; Braatz evaluation files). Not a live refinery model. Thresholds are 99th percentiles of normal reconstruction / IF scores on IDV(0) train windows. Recalibrate on each plant before any operational use.

Operational alerts use PCA reconstruction (lowest usable false-alarm rate among LSTM / PCA / Isolation Forest). LSTM-AE is kept for scoring and per-tag attribution; it is not the primary alarm because its pre-fault FAR on this protocol is high (~0.58).

Held-out protocol (fault after 8 h)

Mean alarm reduction (PCA incidents vs raw 3σ tag alarms): 93%. IDV(3), IDV(9), IDV(15) are weakly observable in the TEP literature.

How to score a window

import joblib, numpy as np, onnxruntime as ort
scaler = joblib.load("scaler.joblib")
thr = joblib.load("thresholds.joblib")
sess = ort.InferenceSession("lstm_ae.onnx")
# x: (T, 52) raw tags in xmeas_1..xmv_11 order, T>=20
from numpy.lib.stride_tricks import sliding_window_view
scaled = scaler.transform(x)
windows = sliding_window_view(scaled, (20, 52))[:, 0]
recon = sess.run(None, {"window": windows.astype("float32")})[0]
score = ((windows - recon) ** 2).mean(axis=(1, 2))
alert = score > thr["lstm"]

Related

• Dataset: alirezaaminzadeh/refineryguard-tep-features
• Space: alirezaaminzadeh/refineryguard-process-anomaly
• Collection: RefineryGuard
• Product: aria-ai.ir

MIT · Aria AI Engineering Team