MHZ-CRITICALITY
Multi-Layer Algorithm for Phase Transition Detection
A proprietary multi-layer algorithm that detects pre-critical signatures in complex systems before catastrophic phase transitions occur.
The Problem
Standard anomaly detection algorithms (LSTM, Isolation Forest, CNN-LSTM) are reactive — they detect failures after they begin. Critical systems require early warning before the transition occurs.
The Approach
MHZ-Criticality uses a proprietary multi-layer detection architecture to identify universal signatures of impending phase transitions:
Temporal Pattern Recognition
Identifies anomalous dynamics in multivariate time series before they escalate.
Structural Coherence Monitoring
Detects breakdown in system-wide correlations that precede phase transitions.
Multi-Scale Signal Extraction
Proprietary signal processing across multiple time scales and feature dimensions.
Ensemble Decision Layer
Combines independent detection signals into a unified risk score with calibrated confidence.
The Result
Early detection of catastrophic failures with lead times of hours to months, enabling preventive intervention instead of reactive response.
Validation on Industry-Standard Benchmarks
Three independent benchmarks across industrial, infrastructure, and financial domains.
Industrial Process Monitoring (TEP Dataset)
Tennessee Eastman Process · 53 variables, 21 fault types, 500+ hours
| Algorithm | Lead Time | Detection Rate | False Positive Rate |
|---|---|---|---|
| PCA | 45 min | 85% | 2.3% |
| SVM | 78 min | 92% | 1.8% |
| Random Forest | 95 min | 94% | 0.8% |
| LSTM | 120 min | 98% | 0.5% |
| CNN-LSTM | 151.5 min | 100% | 0.12% |
| MHZ-Criticality | 304.8 min | 100% | 0.00% |
+101% improvement in lead time vs. state-of-the-art (CNN-LSTM)
Critical Infrastructure (SWaT Dataset)
Secure Water Treatment · 51 sensors, 7 attack scenarios, 100+ hours
| Algorithm | Detection Rate | False Positive Rate | Lead Time |
|---|---|---|---|
| Statistical Process Control | 78% | 5.2% | 15 min |
| Isolation Forest | 85% | 3.1% | 20 min |
| LSTM Autoencoder | 92% | 2.4% | 35 min |
| CNN-LSTM | 95% | 1.57% | 45 min |
| MHZ-Criticality | 100% | 0.00% | 60 min |
Perfect detection across all 7 attack scenarios with zero false alarms
Real Physical Testbed: Validated on operational water treatment facility, not simulation
Financial Markets (China LOB Dataset)
Limit Order Book · 124 features, 5-class prediction, high-frequency data
| Algorithm | Cohen's Kappa (H5) | Minority Class | Cross-Horizon |
|---|---|---|---|
| Linear Regression | 0.0823 | Poor | Limited |
| MLP | 0.1156 | Poor | Limited |
| CNN | 0.1389 | Moderate | Limited |
| LSTM | 0.1512 | Moderate | Limited |
| CNN-LSTM | 0.1644 | Moderate | Limited |
| MHZ-Criticality | Non-zero | Superior | Generalized |
Maintained performance on synthetic data where baseline models collapsed to majority-class prediction
Regime Detection: Superior sensitivity to market phase transitions and minority class signals
Applications & Impact
Early detection transforms operations from reactive to preventive across critical domains.
Industrial Process Monitoring
Catalyst degradation costs $600K–$1.8M per incident. 18.8 hours of lead time enables preventive maintenance instead of emergency shutdown.
Applications
- Chemical plants
- Oil & gas refineries
- Manufacturing facilities
- Power generation
Critical Infrastructure
Cyberattacks on critical systems cause cascading failures. 60 minutes of early warning prevents damage before it occurs.
Applications
- Water treatment facilities
- Power grids
- Transportation systems
- Communication networks
Financial Markets
Market crashes destroy billions in value. Early detection of regime changes enables hedging and risk management.
Applications
- Trading systems
- Risk management
- Portfolio optimization
- Market surveillance
Technical Details
Designed for practical deployment on standard hardware with minimal configuration.
Input / Output
Input
- •Multivariate time series (any dimensionality)
- •Minimum 100 timesteps for training
- •No feature engineering required
Output
- •Binary classification (normal / critical)
- •Continuous risk score (0–1)
- •Multi-layer signal decomposition
- •Estimated time to failure
Computational Requirements
Infrastructure
- •Standard CPU (no GPU required)
- •8 GB RAM minimum
- •Real-time inference (<100 ms per sample)
Training
- •Unsupervised (no labeled failures required)
- •Semi-supervised (optional labels improve accuracy)
- •Online learning (continuous adaptation)
How MHZ-Criticality Compares
Systematic advantages over every class of anomaly detection method.
vs. Statistical Methods (PCA, SPC)
- 3–5× better lead time
- Detects non-linear failures
- No manual threshold tuning
vs. Machine Learning (Random Forest, SVM)
- 2–3× better lead time
- Captures temporal dynamics
- Lower false positive rate
vs. Deep Learning (LSTM, CNN-LSTM)
- 2× better lead time
- Interpretable risk scores
- No massive training data required
vs. Ensemble Methods
- Proprietary multi-layer architecture
- Cross-domain generalization
- Real-time performance
Intellectual Property
Patent PendingProprietary & Patent-Pending
The mathematical framework, detection layers, and inference pipeline behind MHZ-Criticality are proprietary and patent-pending. Only benchmark results and integration interfaces are disclosed.
Standard Hardware
Runs on commodity CPUs with no specialised infrastructure.
No GPU Required
Full inference pipeline operates without GPU acceleration.
No Cloud Dependencies
Can be deployed on-premise with no external connectivity.
Real-Time Inference
<100 ms per sample for continuous monitoring applications.
Licensing
Research Use
- •Academic publications
- •Benchmark comparisons
- •Non-commercial applications
Commercial Use
- •Industrial process monitoring
- •Financial market prediction
- •Critical infrastructure security
Interested in licensing or partnership?
MHZ-Criticality is available for enterprise licensing, research collaboration, and integration partnerships.