Anomaly Detection System for an IOT Company
About FreshAir Sensor
The challenge: Improving anomaly detection system using AI
The non-AI system used previously was raising a lot of false alarms. It was cumbersome and error-prone to review all the alarms manually, and it was making the entire solution unscalable.
Cumbersome review process
FreshAir Sensor’s system raised many alarms, many of which turned out to be false positives. Manual review of these alarms was time-consuming and prone to errors.
Data inconsistency
Data pipelines suffered from low-quality or missing data, hampering accurate model training and reducing trust in the alarm system.
Limited scalability
As the volume of time-series data grew, the existing approach could not scale effectively without risking both missed events and alarm fatigue.
Reducing alarm fatigue with AI-based anomaly detection
To enhance detection accuracy, we focused on advanced feature engineering and a dual-model approach, ensuring both interpretability and precision in anomaly identification.
Robust feature engineering
- We extracted signal-based features such as time lag, moving averages, and standard deviations to capture meaningful patterns in the time-series data.
- These features served as inputs for two parallel models—a statistical XGBoost model for interpretability and a convolution-based model for automatic feature extraction.
Automated anomaly detection
- Both models were trained to classify sensor readings as normal or anomalous, drastically reducing the reliance on human intervention.
- The convolutional approach further refined detection by leveraging advanced pattern recognition.
Data pipeline improvements
- We identified critical gaps in FreshAir Sensor’s data engineering processes.
- By recommending and implementing best practices (e.g., consistent data labeling, and improved data collection pipelines), we ensured future data sets would be more reliable and actionable.
The measurable impact
Reduced false positives
By accurately distinguishing between genuine alarms and noise, FreshAir Sensor cut back on expensive and time-consuming manual reviews.
Faster detection
High-confidence alerts now move forward without delay, allowing the team to respond more quickly to actual smoking events.
Cost savings & scalability
Eliminating unnecessary manual checks freed up resources, enabling the platform to handle a growing number of sensors and events without compromising quality.
Improvement in Data Quality
Improved data pipelines and enhanced model accuracy laid a strong foundation for future analytics and potential product expansions.
Tools & Technologies
XGBoost
For statistical modeling and feature-based classification
Convolutional Neural Networks
For deep feature extraction and pattern recognition
Python & Time-Series Libraries
(NumPy, Pandas) for data preprocessing and analysis
Data Pipeline Optimization
To identify and resolve data quality issues
The engagement
Throughout the engagement, we collaborated closely with FreshAir Sensor’s engineering and operations teams:
Discovery and strategy
Audited existing data workflows and identified key bottlenecks.
Implementation and validation
Deployed two parallel models (XGBoost and convolution-based) and ran performance tests on historical and live data.
Continuous improvements
Recommended best practices for data pipeline management, ensuring ongoing performance gains and more reliable sensor data.
Through this partnership, FreshAir Sensor now benefits from an automated, scalable anomaly detection system that improves accuracy, lowers operational costs, and sets the stage for further innovation in air-quality monitoring.
CASE STUDIES
See more case studies
- Gain insights from experts
- You may be eligible for a free consultation
(10-hour business consultation and 10 hours technical)*
What we offer?
- AI consultation & development
- Get proper product roadmaps
- Validate your product ideas
- Add meaning to product designs (UI/UX)
- Custom software development
- Develop first AI products
We schedule a call at your convenience
A discovery and consulting meeting is held
We prepare the proposal and timelines