Case study

Smart Temperature Anomaly Detection

Monitoring and ensuring appropriate temperatures of stocks such as food items is critical. Variations in temperature can end up damaging the stock. Celsi, a company that provides smart temperature monitoring services, wanted a system for cold storage areas and refrigerators that could automatically raise alarms if the temperature is different than what it should be. This is called anomaly detection.

Client: Celsi
Industry: Logistics
AI capability: Predictive Analytics

Visit Celsi

News article titled Helping hand for restaurants: a restaurant owner at his bar holding a phone that runs the Celsi food-safety monitoring app — The demo video loads from YouTube only when you press play.

The challenge

Every freezer has its own rhythm, so fixed thresholds fail

Temperatures in cold storage areas often follow a frosting-defrosting cycle which can vary across different refrigerators. Furthermore, temperature cycles also depend on the environment. So simply raising an alarm if the temperature went beyond a certain range of value would have either raised too many false alarms or have missed out on a lot of different anomalies. Instead, we required a dynamic system that can adapt to different refrigerator cycles and environmental changes.

What we built

A forecast per unit, and alarms on the deviation

The model learns what normal looks like for that refrigerator.

Stage 1
Learn each unit's own cycle
Cold storage runs a frosting and defrosting cycle that differs between refrigerators and moves with the environment around them. The model is fitted per unit rather than to one global rule.
Per-unit temperature profile
Stage 2
Forecast the next readings
A machine learning time series model predicts what the temperature should do next, given that unit's cycle and its surroundings.
Expected temperature curve
Stage 3
Alarm on the deviation, not the value
An alarm is raised when the actual reading departs from the forecast. Normal defrost peaks stay quiet, while a genuine fault shows up as a gap between the two lines.
Anomaly alerts

Ultra-low freezer: a spike in the real readings departs from the forecast, which is what triggers an alarm.

The impact

Alarms that mean something, with nobody watching

A temperature model per unit
Normal defrost peaks stay quiet
No continuous manual supervision

Our AI system was able to automatically raise alarms whenever temperatures were anomalous thereby reducing chances of spoilage and ensuring freshness of stock. Automating this process also eliminated the need for continuous manual supervision.

Project showcase

More from the project

Discuss a similar project

Freezer with regular defrost peaks: the model learns the cycle, so normal peaks are not flagged.
Irregular cycles are forecast closely, so the alarm threshold follows each unit's own behaviour.
A cold room during a long warm period, with the forecast tracking real readings through the change.