Abstract:
Cold chains are strict logistical requirements set up to protect certain products during the
manufacturing and delivery phases. These products include food, medicine, and vaccines.
For medicine, the cold chains end when it is dispensed to patients. Previous research shows
that the medicine is frequently exposed to non-ideal temperatures once it is dispensed.
In most cases, this exposure is due to domestic refrigerators not maintaining an ideal
temperature range for medicine. This research aims to solve this issue by providing a
data-driven solution and establishing a classification model that can warn medicine users
if their refrigerators are going out of the ideal temperature range.
Medangel provided temperature sensor data from 14 different refrigerators. To develop the
model, we looked into different model characteristics: labeling method, classifier testing,
feature window length, prediction horizon length, and hyperparameter optimization.
Further data engineering was done to extract derivative features from the data set, but
these were not beneficial. We extracted features from overlapping three-hour feature
windows to train the final classification model. We labeled each window high if the
temperatures within the next one hour will be too high, low if the temperatures within
the next one hour will be too low, and ideal if the following one-hour temperatures are
suitable.
The classification model achieved a Matthews Correlation Coefficient score of 0.95 when
tested on unseen data. This score shows that the classifier works exceptionally well
even in unbalanced data sets. Investigation of the feature importance scores shows that
minimum and maximum temperatures were the most significant features in classifying
different windows. We also found less common statistical features such as energy ratio
and Benford’s correlation coefficient. Finally, we created a dashboard concept that could
be used to implement the classification model within the Medangel sensor system.
