Technology Deep Dive
The engineering behind the platform — from signal processing and multi-sensor fusion to machine learning models and private network architecture.
Behavioral Signal Processing
Raw accelerometer data is noisy. The Herdwize signal processing pipeline transforms 3-axis accelerometer readings into classified behavioral states — rumination, feeding, walking, resting, and social interaction — using a combination of frequency-domain analysis, peak detection algorithms, and time-series windowing. This classification runs continuously, enabling the AI engine to work with semantic behavior data rather than raw sensor values.
- Sampling rate: 10 Hz continuous, with adaptive duty cycling for power optimization
- Classification accuracy: 94% across five behavioral states in field validation
- Latency: Under 30 seconds from raw data to classified behavior state
Multi-Sensor Fusion
Individual sensor readings tell an incomplete story. The Herdwize platform fuses temperature, accelerometer, light, and GPS data streams to build a comprehensive picture of each animal's state. Temperature alone might indicate fever — but temperature combined with reduced activity, decreased rumination, and isolation from the herd creates a high-confidence health risk assessment.
- Cross-sensor correlation analysis for health risk scoring
- Individual animal baseline modeling (adapts to each animal's normal patterns)
- Herd-level pattern detection for environmental and management event identification
Machine Learning Models
Herdwize ML models are trained on multi-species behavioral datasets collected from active farm deployments. The estrus detection model analyzes 72-hour activity windows to identify pre-standing-heat behavioral patterns. The health risk model scores animals on a continuous 0-100 scale based on deviation from individually established baselines.
- Estrus detection: 92% sensitivity, 95% specificity in field validation
- Health risk scoring: continuous 0-100 scale with configurable alert thresholds
- Models retrained quarterly with new field data for continuous improvement
- Species-specific model variants for cattle, sheep, and goats
Predictive Analytics Engine
Beyond real-time monitoring, the Herdwize platform generates predictive insights. Calving date estimation based on breeding records and behavioral pattern changes. Feed efficiency scoring based on weight gain correlation with activity and rumination data. Seasonal health risk forecasting based on historical herd data and environmental conditions.
- Calving prediction: ±3 day accuracy based on behavioral pattern analysis
- Feed efficiency scoring for individual animals and cohort comparison
- Seasonal risk forecasting with historical trend analysis
Data Security & Privacy
Farm data is sensitive operational information. The Herdwize platform implements end-to-end encryption from sensor to cloud. LoRaWAN network traffic is encrypted with AES-128 at both network and application layers. Cloud data is stored in SOC 2-compliant infrastructure with role-based access controls. Data ownership remains with the producer — Herdwize does not sell, share, or aggregate customer data.
- AES-128 encryption at LoRaWAN network and application layers
- TLS 1.3 for all cloud API communications
- SOC 2 Type II compliant cloud infrastructure
- Data ownership: producer retains full ownership and export rights
- GDPR and Canadian PIPEDA compliant data handling
Private Farm-Level Network
Unlike solutions that depend on cellular connectivity, Herdwize deploys a private LoRaWAN network on each farm. This provides reliable coverage in areas with poor or no cellular service, eliminates ongoing cellular data costs, and keeps sensitive farm data on-premise by default. The private network architecture also means the producer controls their data infrastructure — not a third-party carrier.
- No cellular dependency — operates on private LoRaWAN infrastructure
- Zero recurring cellular data costs
- Data remains on-premise until explicitly synced to cloud
- Producer controls network infrastructure and data flow
How Behavioral Baselines Are Built
Every animal is different. A Holstein in a 300-cow freestall dairy behaves differently from an Angus on 5,000 acres of rangeland. Herdwize's per-animal baseline modeling is what enables the system to distinguish between "this cow is sick" and "this cow always ruminated less than average."
Days 1-3: Raw Data Collection
The system collects continuous sensor data — temperature, activity, rumination, GPS position — without generating alerts. This raw data populates the initial observation window.
Days 4-7: Pattern Extraction
The AI engine identifies daily rhythms: when does this animal typically eat, ruminate, rest, and move? What is her normal temperature range across the day-night cycle? These patterns form the initial behavioral profile.
Days 8-14: Baseline Refinement
The system cross-references individual patterns with herd-level norms and environmental data (weather, season). Confidence intervals narrow as more data accumulates. By day 14, the individual baseline is stable enough for high-accuracy anomaly detection.
Ongoing: Adaptive Learning
Baselines are not static. The model continuously adapts to seasonal changes, aging effects, lactation stages (dairy), and management changes. This adaptive learning prevents false alerts that plague fixed-threshold systems.
Model Training and Validation Process
Herdwize ML models are trained on real-world farm data — not laboratory datasets. This distinction matters because farm conditions introduce variability that lab-controlled studies cannot replicate: weather events, management changes, feed quality variations, and seasonal behavioral shifts.
- Training data sourced from active farm deployments across multiple geographic regions
- Ground truth labels provided by producer reports, veterinary diagnoses, and progesterone testing
- Models validated using hold-out datasets (never trained on test data)
- Quarterly retraining cycle incorporating new field data and corrected predictions
- Species-specific model variants: separate models for dairy cattle, beef cattle, sheep, and goats
- Transfer learning enables faster baseline establishment for new species or breed types
LoRaWAN vs. Cellular vs. WiFi for Livestock Monitoring
Connectivity technology choice is one of the most consequential decisions in a livestock monitoring system. Each technology has distinct trade-offs in range, power consumption, cost, and reliability. Here is why Herdwize chose LoRaWAN as its primary network technology.
| Feature | LoRaWAN | Cellular | WiFi |
|---|---|---|---|
| Range | 10 km line-of-sight | Depends on tower (0-30 km) | 50-100 meters |
| Battery life (sensor) | 5+ years | 3-12 months | Days to weeks |
| Monthly data cost | $0 (private network) | $2-10/device/month | Broadband cost |
| Coverage gaps | You control it | Carrier dependent | Building/barn only |
| Capacity per base station | 1,000+ devices | Varies by tower | 50-100 devices |
| Infrastructure cost | $500-2,000 per gateway | $0 (carrier owned) | $50-200 per AP |
| Data rate | Low (sufficient for sensors) | High | High |
| Best for | Remote, large-area, low-power | Urban, high-bandwidth | Indoor, high-density |
Ready to Deploy Livestock Intelligence?
Schedule a platform briefing to discuss deployment scope, infrastructure requirements, and expected ROI for your operation.