Business Challenges
- COVID-19 pandemic amplified the need for reliable cold storage for vaccine protection
- A cold storage freezer’s failure can lead to the loss of valuable research and assets
- Existing practices were obsolete
- The need was to get rid of the waiting period to fix a failure


Objective
- To research & implement state-of-the-art AI/ML solutions & classify Normal Refrigerator Cycles from Refrigerator Short Cycles
- To detect the number of Refrigerator Short Cycles behavior per day
Solution
AN ANALYSIS OF TIME-SERIES DATA
The subsequent solution pipeline is based on research & initial data visualization on Tableau
01
Exploratory Data Analysis
- Convert time series data from text/JSON to CSV format, making a data frame through pandas
- Analysis of every feature received in the data frame
02
Feature Engineering & Extraction Process
- Check skewness in the data 12% for a short cycle 88% for a normal cycle
- Remove unwanted refrigerator types, feature selection based on feature importance histogram analysis
- Used clustering techniques to differentiate short cycles vs normal cycles and make labels
OBSERVATION
Clustering techniques could not differentiate between short and normal cycles
03
Feature Engineering - Further Analysis
- Manuel labeling of each unique refrigerator type (mac_ids)
- Converted original time-series data patterns into binarized data-patterns
- Created 30 new features
04
ML Model Training And Validation
- An unsupervised approach
- clusters were developed based before labelling
- Implemented manual clustering & advanced clustering techniques
- Split data into Training,Validation, and Test Set
- Trained model on state-of-the-art Ensemble ML Algorithm
- Model evaluation analysis and hypert-uned model parameters
OBSERVATION
Short cycles & normal cycles weren't identified in 2 separate clusters. One cluster capturing bad data was removed as post-processing.
05
Current Model Performance and Deployment Strategies
- Successfully able to classify refrigerator short cycles from normal cycles
- Current Model Accuracy on Test Set: 97%
- Ongoing improvements - Deciding probability score for short cycles before production, reducing False Positives and Negatives
- Initial Model is deployed via Flask

Architecture Diagram

State-Of-The-Art ML Libraries & Packages

Result
Achieved a 20% increase in accuracy from the initial 65%
Discuss Your Idea!

+1 (646) 203-1075

Request a call back