Research · Sensor TelemetryData Science Intern · Noodle.aiJul 2019 – Oct 2019
Spurious Outlier Detection for High-Frequency Time Series
An integrated, scalable approach to detecting spurious outliers in high-frequency sensor streams, differentiating contextual anomalies from noisy outliers to improve the reliability of analytical models on manufacturing IoT data. Subject of a published US patent (US20210294787A1).
View patent filing on Google Patents ↗US20210294787A1
Patent
Published
Reduced
False Shutdowns
Sensor noise vs. true drift
// THE PROBLEM
Problem formulation
Industrial IoT sensors emit spurious spikes and electrical noise that corrupt model training and trigger false plant shutdowns. Distinguishing a real process anomaly from a faulty sensor reading is the core challenge.
// DATA FOUNDATIONS
Data engineering & ingestion
- High-frequency time series from raw manufacturing sensor streams
- Windowed feature extraction capturing local distributional context
- Label strategy separating ground-truth process drift from instrumentation faults
// MODELING STRATEGY
Mathematical & algorithmic approach
- Statistical baselines (z-score, moving median, rolling volatility) for noise detection
- Unsupervised ML models (isolation forest, clustering) for contextual anomaly separation
- Ensemble decision logic combining local context with global distribution
// EXPLAINABILITY
Interpretability & governance
- Clear delineation between spurious outlier vs. contextual process anomaly
- Alert metadata preserving reason codes for downstream operator triage
- Human-interpretable thresholds tuned per sensor channel
// PRODUCTION
Deployment & integration
- Designed for high-throughput streaming environments
- Pre-processing stage installed ahead of downstream quality models
- Prevents model degradation and false positive plant shutdowns
// TECH STACK
PythonNumPySciPyStatistical ModelingIsolation ForestTime Series