Machine Learning for Molten Aluminum Dross Prediction Accuracy

Overview of Technical Issues:

The machine learning prediction model insufficiently converts measured process parameters into accurate molten aluminum dross quantity forecasts, failing to capture complex formation mechanisms and resulting in prediction errors that undermine process optimization and material recovery decisions; the goal is to improve model accuracy to enable reliable dross forecasting for operational planning.

Solution directions generated for this problem

Problem Direction 1 :

ImproveModel feature representation capability
VS
ConstraintPrediction computation time

Inspiration 1 : Cross-domain reference

Application Principle: #10 Preliminary action
Cross-domain applicability Assess applicability
Item recommendations using image feature data
Innovative Solution Refine solution

Pre-computed feature library with runtime pattern matching for dross prediction

Offline feature library with real-time matching
How to solve :
  • During furnace idle periods, pre-compute a comprehensive feature interaction library covering temperature-composition-oxide dynamics across 10,000+ parameter combinations using physics-based models (thermal diffusion equations, oxidation kinetics)
  • store as indexed lookup tables with interpolation coefficients
Expected Effect : At runtime, deploy a <strong>fast pattern matching engine</strong> that compares incoming sensor data (temperature gradient ±1°C, Mg content 0.1-3%, oxygen partial pressure 0.01-0.5 atm) against the pre-computed library using k-nearest neighbor search (k=5); retrieve closest feature vectors in <15 seconds
Risk Control :
  • Feed matched feature vectors into a lightweight neural network (3 hidden layers, 128 neurons each) pre-trained on the library
  • network outputs dross quantity prediction in <30 seconds total runtime while maintaining <5% error through ensemble of 3 parallel matching paths with weighted voting

Problem Direction 2 :

ImproveModel feature representation capability
VS
ConstraintSystem implementation complexity

Inspiration 1 : Cross-domain reference

Application Principle: #26 Copying
Cross-domain applicability Assess applicability
Virtual height filter for reflected sound rendering using upward firing drivers
Innovative Solution Refine solution

Physics-based virtual sensor network for dross prediction feature enrichment

Deploy virtual sensors to replicate complex features without physical hardware
How to solve :
  • Implement thermal diffusion virtual sensors using Fourier heat equation solvers to compute local temperature gradients (±1°C precision) from 3–5 bulk thermocouples, eliminating need for 20+ thermal imaging arrays
  • Deploy oxide growth kinetic models (Pilling-Bedworth ratio, parabolic rate law k_p=2–8×10⁻⁸ cm²/s at 700°C) to infer oxide layer thickness from surface temperature and oxygen partial pressure, replacing spectroscopy sensors
  • Integrate edge computing modules (ARM Cortex-A72, 1.5GHz) at furnace control units to execute virtual sensor calculations in real-time (<500ms per cycle), outputting 15–20 enriched features (temperature gradients, oxide thickness, Mg depletion rate) directly to ML model input
Expected Effect : Feature count +250%, system cost +30% vs full sensor array, prediction error <8%
Risk Control :
  • virtual sensor calibration drift over time
  • thermal model boundary condition accuracy
  • edge processor computational stability under high temperature

Problem Direction 3 :

ImproveMeasurement precision of process parameters
VS
ConstraintPrediction computation time

Inspiration 1 : Cross-domain reference

Application Principle: #28 Mechanics substitution (Replace mechanical system)
Cross-domain applicability Assess applicability
Monitoring fitness using a mobile device
Innovative Solution Refine solution

Edge-processed infrared gradient extraction for real-time dross prediction

Replace sensor arrays with edge-computed thermal maps
How to solve :
  • Deploy infrared thermography camera with embedded FPGA processor at furnace top — camera captures 640×480 thermal image at 5Hz, FPGA extracts 10 spatial gradient features (±1°C precision) via real-time convolution kernels and transmits only compressed gradient vectors (200 bytes/frame) instead of raw images (1.2MB/frame)
  • Implement firmware-level gradient computation using pre-trained 3×3 Sobel filters for horizontal/vertical temperature derivatives, followed by zone-based averaging (furnace divided into 10 regions) — processing latency <50ms per frame, outputs include surface max gradient, center-edge delta, and quadrant variance
  • Integrate gradient features directly into lightweight XGBoost model (50 trees, max depth 6) running on edge server — model trained offline on 500 historical batches, accepts 10 gradient values + 5 bulk parameters, predicts dross quantity in 45 seconds with <5% error for 85% of batches
Expected Effect : Prediction time reduced to <1 min; precision ±1°C; error <5% in 85% cases; data transmission reduced 99.98%
Risk Control :
  • FPGA firmware calibration drift over 6 months
  • infrared lens contamination in high-dust environment
  • model retraining required every 200 batches

Problem Direction 4 :

ImproveMeasurement precision of process parameters
VS
ConstraintSystem implementation complexity

Inspiration 1 : Cross-domain reference

Application Principle: #26 Copying
Cross-domain applicability Assess applicability
Correlation of stack segment intensity in emergent relationships
Innovative Solution Refine solution

Physics-based virtual sensor network for gradient inference without hardware expansion

Deploy virtual sensors via thermal models to infer gradients from sparse measurements
How to solve :
  • Install sparse physical sensor array (3–5 thermocouples at furnace walls, 2 spectroscopy units at surface) measuring bulk temperature ±0.5°C and surface oxide reflectance
  • implement physics-based thermal diffusion model (Fourier heat equation solver, 10Hz update rate) on edge processor to compute local temperature gradients across 20+ virtual points with ±1°C precision from boundary measurements
  • use optical reflectance-to-thickness calibration curve (pre-trained on 50 sample batches, R²>0.92) to infer oxide layer thickness ±0.15mm at 8 virtual zones from 2 physical spectroscopy readings
Expected Effect : Precision ±1°C gradient, ±0.15mm oxide; hardware cost +40% vs +300%; maintenance complexity unchanged
Risk Control :
  • thermal model drift over furnace aging cycles
  • calibration curve degradation with alloy composition shifts
  • edge processor computational stability under high-temperature EMI

Problem Direction 5 :

ImprovePrediction accuracy level
VS
ConstraintPrediction computation time

Inspiration 1 : Cross-domain reference

Application Principle: #1 Segmentation
Cross-domain applicability Assess applicability
Method for coding and an apparatus
Innovative Solution Refine solution

Hierarchical prediction model with fast screening and selective deep computation

Deploy two-stage model: fast screening then selective deep analysis
How to solve :
  • Implement fast linear regression model (30-second runtime) to screen all batches and classify risk levels based on predicted dross quantity thresholds (low-risk <8%, medium 8-12%, high >12%)
  • Route only medium and high-risk batches (estimated 30% of total) to full deep learning model with complete feature set (temperature gradients ±1°C, oxide layer dynamics, composition interactions) for 15-minute detailed prediction
  • Low-risk batches receive direct approval with linear model output (tolerance ±3% verified against historical data), achieving <5% error target through risk-stratified computation allocation
Expected Effect : 70% batches predicted in <1 min; overall accuracy <5% error; average computation time reduced to 5 min
Risk Control :
  • risk classification threshold calibration drift
  • linear model degradation over time
  • batch misclassification rate exceeding 10%
Patsnap Eureka Solution