Skip to main content

DataRPM Analytics 50 Submission


Sundeep Sanghavi, Chief Executive Officer




Predictive Analytics

Business Challenge

The challenge that most of DataRPM’s customers—asset-based businesses ranging from manufacturing to utilities to healthcare and smart cities, among many others—commonly face is how to do Predictive Maintenance (PdM) and scale on production from the huge amount of data generated by the sensors in the Industrial IoT. The key challenges in doing PdM for IoT are:

  • The traditional approach of monitoring one sensor at a time and manually writing alert rules doesn’t scale
  • The sensor data is extremely noisy, as different sensors generate data-different instances and at different frequencies
  • There is no labeled data to train machine learning models for supervised learning
  • Statically generating models does not work, as they become obsolete by the time they are operationalized

Analytics Solution

DataRPM’s solution delivers a cognitive predictive maintenance platform that automates data science using meta-learning, which is an actively researched area and the next frontier for machine learning. Meta-learning is about training machines to do machine learning, just like data scientists work, by running multiple different experiments and from each of those experiments extracting meta-data about the characteristics of the dataset, the features used, the algorithm applied, the hyper-parameters selected and the result of the objective function (error, accuracy, precision, recall, etc.). The machine learns from these experiments continuously and becomes smarter over time to be able to identify which combinations of features, algorithms and hyper-parameters are going to deliver the best ensemble of predictive models for predicting everything from asset failures/breakdowns, inventory and resource optimization, quality and warranty issues, and risks.


Across various customers, DataRPM has seen on an average:

  • 30% cost savings for predictive maintenance
  • 300% increase in prediction quality
  • One-thirtieth the time it takes to do analysis manually