6-layer deterministic pipeline
·
No black-box AI — every decision traceable
·
Closed-loop — 24h recalibration
·
Auditable output log per particle event
System Architecture

ρ-MATRIQS™
System Architecture

A six-layer pipeline that converts chaotic mixed scrap into classified, value-weighted sorted fractions — in under 10ms per particle. Every layer is documented, every decision is logged, every failure mode is handled.

Download Integration Schema View Sensor Stack →
Pipeline

The Complete 6-Layer Processing Stack

From physical material chaos to sorted, value-captured output — each layer is defined, specified, and failure-handled.

L1
Material Chaos Layer
Active
Mixed scrap stream with size/density/composition variability. Chaotic particle distribution on belt. This is where the yield gap originates — uncontrolled distribution creates sorting error before any sensor is active.
Chaotic distribution No pre-sort required 80–800 TPD capacity Belt width: 600–2400mm
L2
Multi-Sensor Acquisition
Active
Four sensing modalities sample every particle at belt speed. All sensor data is timestamped and position-synced via belt encoder. Raw signals are pre-processed at the edge node — no cloud upload.
Hyperspectral: 400–2500nm RGB optical: 4MP @ 120fps Acoustic: 8kHz–80kHz Optional: XRF / EM induction 15,000 pts/sec aggregate IP69K dust/water rated
L3
Feature Extraction Engine
Active
Raw sensor signals are transformed into feature vectors per particle. This is the critical missing layer in most AI sorting systems. Without it, classification is pattern-matching on noise.
Spectral signature vectors Density proxy estimation Shape/surface classification Oxidation/contamination markers Particle overlap detection Feature computation: <2ms
L4
Stochastic Inference Engine
Active
Bayesian classification with mass-balance constraint. The model maintains Σ(input) ≈ Σ(classified output) — preventing grade drift over a shift. Output per particle: class label + confidence score + economic value weight.
Bayesian P(material|signal) Mass-balance constraint Confidence: 92–97% Economic value weighting Model size: <180MB on-device Inference: <8ms per particle
L5
Real-Time Actuation
Active
Decision signal fires to air jet or mechanical diverter. Actuation only triggers when P(material) > configurable threshold AND economic value exceeds cut-off. Belt encoder drives position tracking — jet fires at the exact coordinate.
Decision latency: <10ms Position tracking: encoder-based Air jet: 4-bar compressed air Configurable confidence threshold Belt speed: 0.5–3.5 m/s Fail-safe: pass-through on fault
L6
Yield Feedback Loop
Active
Sorted output is weighed and composition-sampled. Recovery % is computed per shift. Delta vs expected is fed back to recalibrate classification thresholds. Every 24h cycle improves system accuracy. Drift correction is automatic.
24h recalibration cycle OTA model updates Recovery delta tracking Shift-level audit log CPCB-compliant event log Remote monitoring: included
Layer 3 Deep Dive

Feature Extraction — The Missing Layer

Most AI sorting systems go Sensor → AI → Output. This is why they underperform on real-world dirty, oxidised scrap. Feature extraction bridges raw signal to meaningful classification input.

🌈
Spectral Signature Vectors
Hyperspectral data is decomposed into material-specific absorption/reflection bands. Each particle gets a 128-dimensional spectral fingerprint compared against our calibrated material library.
⚖️
Density Proxy Estimation
Acoustic impact sensing combined with optical volumetric estimation produces a density proxy without direct weighing. Separates ferrous fines from dense non-ferrous chunks before Bayesian classification.
🔲
Shape & Surface Classification
Edge detection + surface texture analysis from RGB data. Differentiates wire/cable from sheet/chunk morphology. Relevant for Cu recovery from ASR where form factor correlates with alloy grade.
🔴
Oxidation / Contamination Markers
Spectral bands around 1400–1700nm detect oxide layers and surface contamination. The model compensates — sorting is based on sub-surface material, not surface state.
Layer 4 Deep Dive

Stochastic Inference Engine

Bayesian classification with mass-balance constraint. Not a neural network black box — a probabilistic model where every output has an explicit confidence score.

rho-matriqs::inference_engine — CLASSIFICATION PSEUDOCODE
// Per-particle inference — runs in <8ms
function classify_particle(feature_vector f, material_priors π) {
  // Bayesian posterior: P(material | signal)
  for each material_class m in [Cu, Al, Zn, Fe, Pb, Unknown]:
    P(m|f) = P(f|m) × π(m) / P(f)

  // Mass-balance constraint: Σ classified = Σ input (per shift)
  if shift_composition_drift > threshold:
    recalibrate_priors(π)

  // Actuation decision
  if max(P(m|f)) > confidence_threshold AND value(m) > economic_cutoff:
    fire_actuator(particle_position, m)
  else:
    pass_through() // conservative — do not eject uncertain particles
}
Mass-Balance Constraint

Every sort session maintains Σ(input) ≈ Σ(classified output). If grade drift is detected across a shift — the model recalibrates priors. This prevents systematic under/over-sorting of a specific material class.

Confidence Threshold Logic

Default confidence threshold: 0.78. Configurable per material class. When confidence is below threshold, particle passes through — not ejected. This bounds false positive rate and prevents fraction contamination.

Layer 6

Closed-Loop Yield Optimization

Most sorting systems stop at actuation. ρ-MATRIQS™ measures the output, computes the recovery delta, and recalibrates the model — every 24 hours, automatically.

01
Sort Event
Air jet fires. Particle routed to fraction bin.
02
Output Weighing
Fraction bins weighed at shift end.
03
Recovery %
Sorted recovery vs input mass computed.
04
Delta vs Expected
Compare to model prediction.
05
Model Recalibration
Threshold + prior adjustment via OTA.
06
Next Shift
Improved accuracy from cycle N+1.
24h recalibration is automatic. No engineer intervention required. Model drift is detected via shift-level recovery delta. OTA update is pushed from our operations centre if drift exceeds ±0.5% across 3 consecutive shifts.

Ready to Integrate?

Download the full integration schema including PLC interface spec, I/O mapping, mounting requirements, and edge compute footprint.

Download Integration Schema View Deployment Guide →