Neural Learning Methods TRIZ Cases
- Laser Weld Fault Monitoring via CNN Height Profile Reconstruction
A CNN reconstructs weld seam height profiles from 2D coaxial camera images, avoiding slow 3D scans while preserving fault detection accuracy.
- Gated Multi-Branch ML Models to Prevent Negative Transfer
A gating network clusters domains and routes inputs to specialized neural blocks, cutting compute while preserving multi-task inference accuracy.
- Machine Operation Anomaly Detection Using Hyperbolic Embeddings
Hyperbolic embeddings cut dimensionality and improve anomaly detection robustness to domain shifts in machine operation signals.
- Neural Network Quantile Normalization for Domain-Shifted Measurement Data
Quantile-based neuron normalization preserves rank while reshaping activations, improving measurement-data robustness under domain shifts.
- Building Sustainability Management Using LLM Recommendations
Generative LLMs turn unstructured building models, specs, and operational data into actionable sustainability recommendations with less manual effort.
- Composite Curing Temperature Profile Control Using Thermal Stack ML
Machine learning infers thermal stack behavior from vessel and tool temperatures to control composite curing profiles when direct measurements are unreliable.
- Neural ODE State Sensing for Intermittent Continuous-Time Tracking
A dynamic autoencoder with neural ODEs reconstructs device states across different state spaces despite intermittent measurements and imbalanced data.
- Edge ML Vibration Monitoring for Low-Bandwidth Fault Detection
Local ML on edge devices analyzes equipment vibration trends and sends only abnormal data, cutting bandwidth while improving fault review.
- Injection Molding Operation Control Using State Expression Maps
Observation data is turned into a state expression map to set molding operation quantities that stay usable under changing environments.
- Autoencoder Reduced-Order Modeling for Nonlinear PDE Control
An autoencoder with physics-informed loss captures nonlinear PDE dynamics in latent space, enabling reduced-order control without full solvers.
- AI-Generated Test Data for Image-Dependent Control Algorithms
Trained AI generates realistic image and sensor test records with action regions, expanding coverage while reducing manual test design.
- Sensor Time-Series Correction Using GAN Fault Classification
GAN autoencoders correct faulty or missing motor sensor signals and classify error types to improve fault detection and data trust.
- Edge Vibration Anomaly Detection for Low-Bandwidth Building Equipment
Processes high-frequency building equipment vibration data at the edge to flag anomalies while cutting bandwidth and analyst workload.
- Autonomous Cleaner Route Segmentation for Partial Path Editing
Pre-registered route segments let operators replace only changed path sections in an autonomous cleaner, cutting re-teaching time and workload.
- Adaptive Quality Monitoring Using Value-Based Sample Selection
Value-based selection of unlabeled manufacturing data cuts labeling effort and speeds quality model adaptation for new production scenarios.
- Physics-Informed Neural Network Control for PDE System Dynamics
A Koopman-based autoencoder captures PDE-governed nonlinear dynamics from limited time-series data for stable real-time control.
- Laser Ablation Monitoring Using Scattered-Light Volume Estimation
Scattered-light images and deep learning estimate ablation volume during laser processing, enabling real-time beam control without slow 3D measurement.
- Lubricating Oil Volume Control Using Motor Current and Temperature Prediction
Motor current data feeds a machine learning model to predict machine temperature and adjust lubricating oil volume in real time, cutting waste.
- Machine Control Policy Tuning for Robust LQR Under Uncertainty
Cone-program policy tuning makes LQR machine control more robust to sensor noise and environmental perturbations while meeting safety constraints.
- Deep Learning Process Control via Non-Invasive Closed-Loop Exploration
A linear MPC safely perturbs an industrial process to gather training data, enabling deep learning control of nonlinear plant behavior.