An Online Anomaly Detection Approach for Fault Detection on Fire Alarm Systems.

Area of Science:

• Engineering
• Computer Science
• Safety Systems

Background:

• Early fire detection is critical for mitigating risks to human lives and economic assets.
• Current fire alarm systems suffer from failures and false alarms, necessitating improved reliability.
• Traditional periodic maintenance schedules for smoke detectors are often inefficient and not condition-based.

Purpose of the Study:

• To develop and evaluate an online, data-driven anomaly detection approach for smoke sensors.
• To enable predictive maintenance by modeling sensor behavior and identifying abnormal patterns indicative of failure.
• To enhance the reliability and reduce false alarms in fire alarm sensory systems.

Main Methods:

• An online, data-driven anomaly detection technique was employed to model smoke sensor behavior over time.
• The approach focused on identifying abnormal patterns that could signal impending sensor malfunctions.
• The method was tested using approximately three years of data from independent fire alarm systems across four customers.

Main Results:

• The proposed anomaly detection method demonstrated promising results for one customer, achieving a precision score of 1 with no false positives for 3 out of 4 fault types.
• Analysis of data from other customers provided insights into potential reasons for performance variations and areas for improvement.
• The study successfully identified specific failure patterns in smoke detectors.

Conclusions:

• The data-driven anomaly detection approach shows potential for predictive maintenance of smoke sensors.
• Further research and refinement are needed to address variations across different customer systems and improve overall performance.
• The findings offer valuable insights for enhancing the reliability and reducing false alarms in fire detection systems.