Editing AI for Manufacturing and Industry 4.0 (section)

== <span style="color: #FFFFFF;">Applying</span> ==
'''Vibration-based predictive maintenance:'''
<syntaxhighlight lang="python">
import numpy as np
import pandas as pd
from scipy.fft import fft, fftfreq
from sklearn.ensemble import IsolationForest
import torch
import torch.nn as nn

# Load vibration sensor data (time series from accelerometer)
# Sample rate 25.6 kHz, 1-second windows
def extract_spectral_features(vibration_signal: np.ndarray, sample_rate: int = 25600) -> np.ndarray:
    """Extract frequency-domain features for ML."""
    N = len(vibration_signal)
    freqs = fftfreq(N, 1/sample_rate)[:N//2]
    fft_vals = np.abs(fft(vibration_signal))[:N//2]
    # RMS and peak features
    rms = np.sqrt(np.mean(vibration_signal**2))
    peak = np.max(np.abs(vibration_signal))
    crest_factor = peak / rms
    # Frequency band energies (diagnostic frequency bands for bearing faults)
    bands = [(0, 1000), (1000, 5000), (5000, 10000), (10000, 12800)]
    band_energies = []
    for f_low, f_high in bands:
        mask = (freqs >= f_low) & (freqs < f_high)
        band_energies.append(np.sum(fft_vals[mask]**2))
    return np.array([rms, peak, crest_factor] + band_energies)

# Fit anomaly detector on normal operation data
normal_windows = load_normal_vibration_data()  # shape: (N, window_size)
features = np.array([extract_spectral_features(w) for w in normal_windows])
detector = IsolationForest(contamination=0.01, n_estimators=200)
detector.fit(features)

# Real-time monitoring
def monitor_bearing(live_window: np.ndarray) -> dict:
    feats = extract_spectral_features(live_window).reshape(1, -1)
    score = detector.score_samples(feats)[0]  # More negative = more anomalous
    is_anomaly = detector.predict(feats)[0] == -1
    return {'anomaly': is_anomaly, 'score': score,
            'alert': 'MAINTENANCE REQUIRED' if is_anomaly else 'NORMAL'}
</syntaxhighlight>

; Manufacturing AI application map
: '''Predictive maintenance''' → Vibration + FFT + Isolation Forest / LSTM autoencoder
: '''Visual inspection''' → EfficientNet/ResNet fine-tuned on defect images; anomaly-based (PatchCore)
: '''Process optimization''' → Bayesian optimization (Ax/BoTorch), RLHF for process control
: '''Digital twins''' → NVIDIA Omniverse, Siemens Tecnomatix, Ansys Twin Builder
: '''Energy optimization''' → DeepMind's RL for data center cooling (40% savings)
</div>

<div style="background-color: #8B4500; color: #FFFFFF; padding: 20px; border-radius: 8px; margin-bottom: 15px;">