"""Dataset baselines: screening evidence, never autonomous maintenance orders."""
from __future__ import annotations
import hashlib
import json
from pathlib import Path
import numpy as np
import pandas as pd
from PIL import Image
from sklearn.dummy import DummyClassifier
from sklearn.ensemble import RandomForestClassifier
from sklearn.metrics import classification_report, confusion_matrix
from sklearn.model_selection import train_test_split
AI4I_FEATURES = [
"Air temperature [K]",
"Process temperature [K]",
"Rotational speed [rpm]",
"Torque [Nm]",
"Tool wear [min]",
]
[docs]
def load_ai4i(path: Path):
"""Exclude IDs and failure-mode targets to prevent direct target leakage."""
frame = pd.read_csv(path)
x = frame[AI4I_FEATURES].to_numpy(dtype=float)
y = frame["Machine failure"].to_numpy(dtype=int)
if set(np.unique(y)) != {0, 1} or not np.isfinite(x).all():
raise ValueError("Expected finite measurements and binary failure labels")
return x, y, frame["UDI"].astype(str).to_numpy()
[docs]
def load_infrared(root: Path):
"""Load original metadata; de-duplicate identical decoded images before splitting."""
metadata = json.loads((root / "module_metadata.json").read_text())
records, ambiguous = {}, set()
root = root.resolve()
for identifier, item in sorted(metadata.items()):
path = (root / item["image_filepath"]).resolve()
if not path.is_relative_to(root):
raise ValueError("Image path leaves dataset directory")
with Image.open(path) as image:
if image.size != (24, 40):
raise ValueError(f"Unexpected image dimensions: {image.size}")
pixels = np.asarray(image.convert("L"), dtype=np.uint8)
digest = hashlib.sha256(pixels.tobytes()).hexdigest()
label = item["anomaly_class"]
if digest in ambiguous:
continue
if digest in records:
if records[digest][1] != label:
del records[digest]
ambiguous.add(digest)
continue
records[digest] = (pixels.ravel() / 255.0, label, str(identifier))
if not records:
raise ValueError("No unambiguous images remain")
xs, ys, ids = zip(*records.values(), strict=True)
return np.asarray(xs), np.asarray(ys), np.asarray(ids)
[docs]
def evaluate(x, y, identifiers, *, dataset: str, output: Path):
"""Fixed stratified holdout, fixed model settings, no test-driven tuning."""
if len(x) != len(y) or len(y) != len(identifiers) or not np.isfinite(x).all():
raise ValueError("Invalid sample arrays")
train, test = train_test_split(np.arange(len(y)), test_size=0.25, random_state=42, stratify=y)
model = RandomForestClassifier(
n_estimators=100,
max_depth=16,
class_weight="balanced",
random_state=42,
n_jobs=2,
)
model.fit(x[train], y[train])
prediction = model.predict(x[test])
baseline = DummyClassifier(strategy="most_frequent").fit(x[train], y[train])
labels = np.unique(y)
report = {
"schema_version": "1.0",
"dataset": dataset,
"input_basis": "synthetic_industrial" if dataset == "ai4i" else "observed_infrared",
"seed": 42,
"train_rows": len(train),
"test_rows": len(test),
"model": "RandomForestClassifier(n_estimators=100,max_depth=16,class_weight=balanced)",
"classification": classification_report(
y[test], prediction, output_dict=True, zero_division=0
),
"baseline": classification_report(
y[test], baseline.predict(x[test]), output_dict=True, zero_division=0
),
"labels": labels.tolist(),
"confusion_matrix": confusion_matrix(y[test], prediction, labels=labels).tolist(),
"maintenance_action": "human_review",
"dispatch_authorized": False,
"limitations": (
"Random sample holdout; no site/time generalization or causal diagnosis established. "
"No recoverable energy or failure-time forecast."
),
}
output.mkdir(parents=True, exist_ok=True)
(output / "metrics.json").write_text(json.dumps(report, indent=2, allow_nan=False) + "\n")
pd.DataFrame(
{
"source_id": identifiers,
"split": np.where(np.isin(np.arange(len(y)), test), "test", "train"),
}
).to_csv(output / "split.csv", index=False)
pd.DataFrame(
{"source_id": identifiers[test], "actual": y[test], "predicted": prediction}
).to_csv(output / "predictions.csv", index=False)
return report