How to Select Machine Learning for Molten Glass Viscosity Control
Overview of Technical Issues:
The machine learning algorithm module insufficiently models the complex nonlinear relationship between temperature, composition, and molten glass viscosity, resulting in poor prediction accuracy and unstable control performance that compromises product quality consistency; the goal is to select an appropriate machine learning approach that accurately captures viscosity dynamics for precise real-time control.
Solution directions generated for this problem
Problem Direction 1 :
ImproveModel complexity capability
VSConstraintComputational resource demand
Inspiration 1 : Cross-domain reference
Application Principle: #15 Dynamics
Cross-domain applicability
Context-specific user interfaces
Innovative Solution Refine solution
Adaptive Model Complexity Switching for Viscosity Prediction
Context-aware model switching reduces computation while preserving accuracy
How to solve :
- Deploy dual-model architecture: lightweight polynomial regression (8-12 parameters) for steady-state operation, deep neural network (3 hidden layers, 64-128 neurons) activated only during transient conditions
- Implement real-time context detector monitoring temperature rate-of-change (>2°C/min threshold) and composition variance (>1.5% threshold) every 500ms to trigger model switching within 200ms latency
- Configure seamless transition protocol using 3-second overlap period where both models run in parallel, weighted blending (exponential decay factor 0.3) ensures smooth handoff without control discontinuity
Expected Effect : Computation load -65% in steady-state; prediction accuracy ±2.8% maintained; switching latency <200ms
Risk Control :
- false trigger during noise spikes
- transition instability at threshold boundary
- model synchronization drift over extended operation
Problem Direction 2 :
ImprovePrediction accuracy
VSConstraintReal-time processing speed
Inspiration 1 : Cross-domain reference
Application Principle: #10 Preliminary action
Cross-domain applicability
Method and apparatus for efficiently transmitting information acquired by a terminal to a base station
Innovative Solution Refine solution
Pre-computed viscosity lookup table with real-time interpolation for molten glass control
Offline pre-compute viscosity predictions for all operating scenarios
How to solve :
- Offline train deep neural network on historical data (temperature 1000-1600°C, composition SiO₂ 65-75%, Na₂O 10-18%) to achieve <2% viscosity prediction error
- generate 3D lookup table with 50°C temperature intervals, 1% composition steps, storing 5000-8000 pre-computed viscosity values in memory (≤10MB)
- during production, use trilinear interpolation between nearest 8 table points based on current sensor readings, achieving prediction in <0.5ms
Expected Effect : Prediction accuracy ±2-3%, response time <1ms, 95% computational load reduction
Risk Control :
- lookup table coverage gaps in edge cases
- interpolation error accumulation during rapid transients
- memory access latency variation
Problem Direction 3 :
ImproveControl stability
VSConstraintComputational resource demand
Inspiration 1 : Cross-domain reference
Application Principle: #35 Parameter changes
Cross-domain applicability
Aerosol generation device, control method and storage medium
Innovative Solution Refine solution
Adaptive precision and update frequency control for viscosity prediction stability
Adjust model precision and frequency based on viscosity variance
How to solve :
- Monitor real-time viscosity variance using sliding window (50 samples, 5-second span)
- when variance <3%, switch to 16-bit floating-point precision and 2-second update intervals, reducing computational load by 50%
- when variance ≥5%, activate 32-bit precision and 0.5-second updates for full stability control
- Implement variance threshold detector with hysteresis (switch-up at 5%, switch-down at 2.5%) to prevent oscillation between modes, ensuring smooth transitions with <200ms detection latency
- Deploy dual-buffer prediction architecture: low-precision model (8-layer neural network, 2000 parameters) runs continuously
- high-precision model (12-layer, 8000 parameters) pre-loads when variance reaches 4%, achieving seamless mode switching without control interruption
Expected Effect : Computational demand -45% average; stability variance <2.5%; mode switch time <300ms
Risk Control :
- variance threshold calibration drift over time
- mode switching transient causing brief control deviation
- floating-point precision reduction accumulating numerical errors
Problem Direction 4 :
ImproveControl stability
VSConstraintReal-time processing speed
Inspiration 1 : Cross-domain reference
Application Principle: #10 Preliminary action
Cross-domain applicability
Fast-acting insulin formulation comprising a substituted anionic compound
Innovative Solution Refine solution
Pre-computed viscosity lookup table with real-time interpolation for molten glass control
Pre-compute viscosity predictions offline for rapid real-time retrieval
How to solve :
- Offline pre-calculate viscosity values for a 3D grid covering temperature (50°C intervals, 800–1400°C), composition (2% SiO₂ steps, 65–80%), and additive concentration (1% steps) using the complex deep neural network — store in indexed lookup table (≈15,000 entries, 2MB memory)
- During production, measure current temperature/composition via sensors, retrieve four nearest grid points in 0.5ms, apply trilinear interpolation to compute viscosity with <1ms total latency
- Implement adaptive grid refinement: detect high-gradient regions (∂viscosity/∂T > 0.5 Pa·s/°C) during initial operation, add intermediate grid points offline within 10 minutes, achieving ±1.5% prediction accuracy in critical zones
- Quality control: validate interpolated predictions against direct viscometer readings every 30 seconds (acceptance: error <3%), trigger table recalibration if 5 consecutive errors exceed threshold, maintain prediction consistency via version-controlled table updates during scheduled maintenance windows
Expected Effect : Processing speed <1ms; prediction accuracy ±2%; control stability improved 40% vs real-time neural network
Risk Control :
- grid resolution insufficient in nonlinear regions
- sensor drift causing lookup mismatch
- composition changes outside pre-computed range
