Skip to main content
Planet Zephyr

Using Blues and Edge Impulse to Increase the Accuracy of Your Machine Learning Models

By March 16, 2024No Comments

Banner image credit @theblowup on Unsplash.

Recently, I built a project as part of Hackster’s Impact Summit that demonstrated a predictive maintenance solution for a pump. As a small prototype intended to show a simple implementation of how a system could work in a smart city use case, my TinyML model and consumer-grade pond pump did the trick. However, after the event, I was asked by several industrial pump and compressor manufacturers how this could work on very large, enterprise-grade infrastructure pumps, at scale.

stock photo of industrial pumps via unsplash

Image credit @bvanbreukelen on Unsplash.

The problem they all faced, is that there is no dataset they can use to build an accurate and representative machine learning model for their equipment. Each type of pump that they build is unique, ranging from small (relatively speaking!) to massive, and literally every individual unit they produce has it’s own small variations and characteristics that make it one-of-a-kind. With no “general purpose” dataset, and inconsistency in final products, they couldn’t find a way to deploy accurate predictive maintenance and anomaly detection applications for their customers.

The combined power of Blues and Edge Impulse can help to solve this. Blues, of course, makes connectivity easy by adding cellular, Wi-Fi, LoRa, and even satellite network options to devices.

With that capability, and turnkey hardware such as the Blues Starter Kit plus an accelerometer, you can then simply capture and pass along IMU values directly from a pump, to the Edge Impulse Data Ingestion API .

edge impulse screenshot

Edge Impulse Studio

This gives you individual, pump-specific motion, vibration, or other patterns of movement, and the data will flow directly into the Edge Impulse machine learning platform. Having pump-specific data will allow you to quickly build an algorithm that understands what “normal” is for that exact pump, no matter its size and shape. This can help to dramatically increase the accuracy of your machine learning model, versus using a generic dataset taken from other pump brands, sizes, or SKUs.

Once the neural network is built using Edge Impulse, the model can then be integrated into a monitoring application or other existing infrastructure software, and deployed back to the Blues Swan (or other hardware device running a Blues Notecard). If the pump is still in the manufacturing facility, this can be done easy enough via USB and a laptop. However, if the pump is already deployed in the field at an end-customer site, then the connectivity afforded by Blues is invaluable, as there is an over-the-air process that allows the Notecard to update firmware on a Swan with a DFU process .

Follow along in this video to perform OTA ML model updates using the Blues Notecard, Edge Impulse, and Zephyr.



This end-to-end pipeline of dataset collection, model creation, and application deployment is one cycle in an MLOps loop. With continual data and training customized on a per-pump basis, Edge Impulse and Blues can help to increase the accuracy of edge machine learning algorithms by using your data, from your sensor, at your location.

For more detail, be sure to check out the full implementation, documented here on the Blues developer blog, and sign up for your free Edge Impulse account .

Benjamin Cabé