Background:
Our client is a ‘battery as a service’ operator, who collaborated with us, with an objective to predict a battery’s failure in advance. Data captured on battery vitals using IoT devices installed in the batteries & pre-determined threshold limits form the foundation of this project.
Approach:
In the preprocessing phase of the project, we identified various metrics such as number of recharge cycles, number of times the vitals have breached the threshold, time intervals between successive breaches, duration of each breach etc. which were used in labeling the rows.
The rows were thus labelled as ‘Warnings’/ ‘Healthy’ / ‘Breached’ based on the vitals of each datapoint. These labels were used to train a classifier model, with the intention of being able to identify faulty rows in real-time data.
Successes and Challenges:
After visualizing the breaches, it was evident that a malfunction is likely to happen only at a particular battery voltage range. This reinforced our hypothesis that the ‘Warnings’ class has a quantifiable relationship with the ‘Breached’ class, which is a conclusive proof of the predictability of the breaches.
Since the number of samples in the ‘Warning’ class are far lesser than ‘Healthy’ samples, the dataset was highly imbalanced. To overcome this, we implemented Synthetic Minority Oversampling Technique (SMOTE),, which artificially populates the minority class (‘Warnings’) with data points that are derived from the already available samples.
Results:
The classifier we built yielded an accuracy of 98% + (which is typical for imbalanced datasets), precision, recall and F1 scores in the range of 0.78-0.82, which are extremely good figures and conclusive proof of our model’s ability to predict breaches.
