Rapid Deployment with Existing Data
Analyze historical CSV logs locally without first setting up complex live streaming pipelines.
Technology
Deterministic edge AI for explainable
condition intelligence
SF2 transforms complex sensor and process data into sparse binary fingerprints, comparing live system states against verified reference conditions in real time. This results in a transparent workflow from raw signal to actionable trigger for Condition Monitoring, Anomaly Detection, and Predictive Maintenance – deployed locally on workstations, industrial servers, or edge devices.
Transparent ProcessingReproducible ResultsFully Traceable System States
Technological fundamentals
The evolution of Semantic Folding Sensor Fusion
SF2’s underlying technology is based on two decades of advancement in Natural Language Processing (NLP). Originally designed to map linguistic relationships as compact semantic fingerprints, Semantic Folding has now been adapted by SF2 Systems for numerical industrial sensor and process data.
By translating linguistic semantics into industrial machine semantics, complex system conditions become instantly comparable, fully explainable, and directly executable at the edge.
Proven technological foundations from semantic language processing – further developed for industrial condition intelligence.
Core principle
Direct overlap vs. statistical probabilities
-
01
Semantic Representation
Heterogeneous signals across a defined time window are condensed into a sparse, 2D binary matrix (e.g., 64×64) representing the current machine state.
-
02
Reference Library
Operational, transitional, and fault states are saved as named reference fingerprints that remain permanently inspectable.
-
03
Direct Overlap
The system continuously checks current fingerprints against the reference library, outputting normalized similarity scores (0–100%) rather than black-box probabilities.
Practical benefits
Why this matters in practice
Application, maintenance, and OT teams receive a transparent path from existing sensor logs to productive condition monitoring. Modeling, reference creation, and trigger logic remain technically traceable and can be operated without cloud infrastructure.
Complete Data Sovereignty
Perform processing and runtime analysis entirely inside your local network—ideal for air-gapped or critical infrastructure (KRITIS) environments.
Seamless Scalability
Export validated models as versioned runtime packages for local servers, industrial PCs, or dedicated SF2 Edge Devices.
Processing Pipeline
From raw signal to actionable trigger
The SF2 architecture strictly separates model creation from live runtime execution. This guarantees full traceability—making it explicitly clear which processing steps build the model and which calculations execute at the edge.
UDTAT — Reproducible Data Preparation
The Universal Data Transformation and Analytic Tool selects relevant sensor channels, applies defined transformations, and generates computed signatures. This produces a consistent, clean data view engineered specifically for semantic modeling.
UDM — Topological State Mapping
The Universal Data Mapper arranges processed input patterns across a 2D topological matrix—the model's semantic space. Similar operating states are automatically mapped adjacent to one another.
UFP — Fingerprint Conversion
The Universal Finger Printer translates binned values and sensor channels into semantic support fingerprints, establishing the baseline dictionary for live runtime processing.
Overlap & Triggers — Operational Signal Generation
SF2 continuously compares live fingerprints against stored reference states to output normalized similarity channels. These channels directly feed warning thresholds and triggers into your PLC, SCADA, or MES.
Model creation in SF2 Suite SE
Define model
Sensor LogsUDTAT · Universal Data Transformation and Analytic ToolUDM · Universal Data MapperUFP · Universal Finger PrinterReference library and versioned model package
These pipeline steps define and version the core model. Data ingestion remains entirely within your local infrastructure.
PRODUCTION DEPLOYMENT WITH SF2 ENGINE RT
Execute model
Running Data windowcurrent fingerprintOverlap with referencesCondition channel or triggerPLC / SCADA / MES / Backend
SF2 handles transformation, fingerprint creation, comparison, and recognition. The operational response takes place within the customer infrastructure.
Data preparation
UDTAT Workflow: 3 Core Processing Steps
Raw data view & preprocessing pipeline
Displays unaltered sensor channels alongside their baseline signal distributions. A pre-configured processing pipeline is loaded and ready to apply reproducibly across the current dataset or any structural equivalent.
- Unfiltered technical reference: Original raw values and distributions remain permanently visible for verification.
- Repeatable configuration: Every pipeline step is explicitly defined, fully versionable, and 100% deterministic.
- Controlled transformation: Signal transformations are executed only when explicitly triggered by the user.
Normalization & channel scaling
Scales selected sensor channels to a uniform numerical range. This transformation aligns signal amplitudes across diverse sensor types while strictly preserving relative signal dynamics and distribution shapes.
- Cross-channel consistency: Normalizes disparate sensor inputs into a single, unified value range.
- Signal integrity: Retains underlying trends, peaks, and relative dynamics without distorting signal character.
- Deterministic foundation: Establishes a clean, reproducible baseline for subsequent quantization and topological mapping.
Quantization & binning
Converts continuous time-series data into discrete, well-defined operational bins. Quantization filters out high-frequency noise and minor signal drift, establishing stable input states for semantic modeling while deliberately accentuating structural state changes.
- Noise-resilient inputs: Maps continuous signal values into stable discrete states to eliminate minor numerical fluctuations.
- Configurable boundaries: Allows precise, fully repeatable definition of bin thresholds across all channels.
- Foundation for mapping: Outputs clean, discretized signal states directly to UDM and UFP for topological state generation.
State representation
Mapping and conversion into semantic fingerprints
UDM mapping: construction of the two-dimensional state map
After data preparation, UDM arranges the observed input patterns in a two-dimensional topology. The animation shows the mapping run over 100 epochs and makes it visible how the map occupancy develops step by step.
- The lower graph shows the course of empty and occupied map positions.
- Adjacent positions represent similar input contexts within the generated mapping.
- The stabilization of the occupancy supports the technical review of the mapping run.
UFP: From mapping to semantic fingerprints
For each discretized value range of a sensor channel, UFP generates an associated semantic fingerprint. These fingerprints form the basis for representing the current system condition.
- Each bin and channel has a uniquely assigned binary representation (fingerprint).
- The fingerprints can be directly viewed in UFP and inspected with exact positioning.
- Generation typically occurs once per finalized model.
- Changes to channels, transformations, binning, or mapping require new generation and versioning.
Live analysis
Live analysis: Semantic fingerprints in operation
SF2 carries raw data, processed signals, the current condition fingerprint, and the comparison with stored references together in a common view.
Current Status
The current status fingerprint represents the currently evaluated data window as a common system condition.
Semantic focus
The center-of-gravity rendering serves as a visual orientation. It does not provide an error probability and does not replace the direct overlap.
Raw data and conditioned inputs
Raw data remains a technical reference; conditioned inputs show the channels after the stored UDTAT pipeline.
Overlap channels and trigger logic
Overlap channels show how strongly the current condition matches selected reference fingerprints.
Technology comparison
A different approach to industrial condition analysis
SF2 does not universally replace every machine learning method. The following comparison shows how the reference-based Semantic Folding approach differs from typical supervised ML projects.
| Aspect | Classical ML Project | SF2 Semantic Folding |
|---|---|---|
| State representation | Model-dependent features, latent representations, or learned decision boundaries | Sparse binary fingerprints that represent concrete system conditions |
| Decision Logic | Statistical prediction or classification; dependent on the chosen method | Direct, reproducible overlap with verified reference conditions |
| Traceability | Depending on the model and Explanation procedures vary in intensity | Agreement with named reference fingerprints and overlap channels directly visible |
| Data basis | For supervised methods, often extensive labeled training data is required | Reference states can be built from observed and verified operating phases; known fault states require reliable references |
| Operating Model | Depending on architecture, local, hybrid, or cloud-based | Designed for local workstation, server, and edge operation; no cloud requirement |
| Computational principle | Procedure-dependent, often matrix-based inference | Compact binary representations and efficient bit operations for state comparison |
| Change management | Dependent on data pipeline, model, training, and MLOps process | Pipeline, mapping, and reference library are versionable; changes to channels or mapping require a new model generation |
| Entry barrier | Data science and MLOps focus | Engineering-first, domain-driven modeling |
| Interpretability | Often black-box behavior | Deterministic overlap-based approach |
| Time-to-value | Long lead time | Fast pilot and rollout path |
Next Step
Understand the technology. Validate your data.
Scale at the edge.
Get started today with the free SF2 Suite SE to analyze local logs, or talk with our team to design your edge deployment.