vG.0.1/benchmark_classical.py

import pandas as pd
import numpy as np
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import StandardScaler
from sklearn.decomposition import PCA
from sklearn.ensemble import IsolationForest
from sklearn.svm import OneClassSVM
from sklearn.impute import SimpleImputer
from sklearn.metrics import roc_auc_score

# 1. Load Data
print("Loading data...")
df = pd.read_csv('real_tokamak_data.csv')

# 2. Preprocessing & Cleaning
# Separate label and remove metadata
y = df['label']
X = df.drop(['label', 'shot_id'], axis=1)

# --- SAFETY BLOCK: CLEAN THE DATA ---
print(f"Original Shape: {X.shape}")

# A. Drop columns that are completely constant (StandardScaler hates these)
# If a sensor reads '0.0' for every single shot, it gives 0 variance -> NaN
X = X.loc[:, X.std() > 0]
print(f"Shape after dropping dead sensors: {X.shape}")

# B. Fill any remaining NaNs with 0
imputer = SimpleImputer(strategy='constant', fill_value=0)
X = pd.DataFrame(imputer.fit_transform(X), columns=X.columns)

# Check if infinite values exist and replace them
X.replace([np.inf, -np.inf], 0, inplace=True)
# ------------------------------------

# Split: Train ONLY on Healthy data (Label 0)
X_healthy = X[y == 0]
X_train, X_test_healthy = train_test_split(X_healthy, test_size=0.2, random_state=42)

# Add Disruptive shots to Test Set
X_disruptive = X[y == 1]
X_test = pd.concat([X_test_healthy, X_disruptive])
y_test = np.concatenate([np.zeros(len(X_test_healthy)), np.ones(len(X_disruptive))])

# Normalize
scaler = StandardScaler()
X_train_scaled = scaler.fit_transform(X_train)
X_test_scaled = scaler.transform(X_test)

# PCA (Reduce to 10 features)
# We use 0.95 variance or 10 components, whichever is safer for your data size
n_components = min(10, X_train_scaled.shape[1]) 
pca = PCA(n_components=n_components)
X_train_pca = pca.fit_transform(X_train_scaled)
X_test_pca = pca.transform(X_test_scaled)

print(f"Data Prepared. Training on {len(X_train_pca)} healthy shots.")
print(f"Testing on {len(X_test_pca)} shots ({len(X_disruptive)} disruptions).")

# --- MODEL 1: Isolation Forest ---
print("\nRunning Isolation Forest...")
iso = IsolationForest(contamination=0.1, random_state=42)
iso.fit(X_train_pca)
y_pred_iso = -iso.score_samples(X_test_pca) 
auc_iso = roc_auc_score(y_test, y_pred_iso)
print(f" >> Isolation Forest AUC: {auc_iso:.4f}")

# --- MODEL 2: One-Class SVM ---
print("Running One-Class SVM...")
svm = OneClassSVM(nu=0.1)
svm.fit(X_train_pca)
y_pred_svm = -svm.score_samples(X_test_pca)
auc_svm = roc_auc_score(y_test, y_pred_svm)
print(f" >> One-Class SVM AUC:    {auc_svm:.4f}")

# Save the PCA data for QGAN
np.savez('pca_dataset.npz', X_train=X_train_pca, X_test=X_test_pca, y_test=y_test)
print("\n✅ PCA Data saved to 'pca_dataset.npz'.")