- How Driverless AI automates feature engineering for predictive maintenance, including target encoding and cluster based features, even on a highly imbalanced failure dataset
- How Shapley values and machine learning interpretability explain a prediction at both the global level, which features matter most overall, and the local level, why this specific machine
- How the computer vision workflow auto selects among pretrained models like Xception, ResNet, and Inception, then uses Grad-CAM to visualize the defect region driving each classification
- How to deploy the resulting model as a portable scoring artifact in Java, Python, or C++, on AWS, GCP, on premises, or directly on edge devices with no internet connection required
- Real production results, including a European truck manufacturer's air pressure sensor program and use cases at Stanley Black & Decker, Petronas, and Cisco
- How the team handles related questions in practice: labeling downtime data for predictive maintenance, anomaly detection on logs with isolation forest or autoencoders, and where transfer learning does and doesn't help
Overview
H2O.ai's Sai, drawing on a mechanical engineering background, walks through two common manufacturing AI use cases, predictive maintenance and computer vision defect inspection, both built live in H2O Driverless AI. The predictive maintenance demo uses a public sensor dataset of voltage, pressure, and vibration readings to predict which machine component is likely to fail, with automatic feature engineering and Shapley based interpretability explaining each prediction. The computer vision demo classifies casting images as defective or okay, auto selecting among pretrained architectures and visualizing which region of the image drove the call. Customer examples from Stanley Black & Decker, Petronas, and a European truck manufacturer ground the discussion in production results.