Predictive Maintenance: Going Next Level With the Ninja Foodi Air Fryer

I love my air fryer.

Our Ninja Foodi Air Fryer became our family’s go-to gadget during the stay-at-home order. When you order takeout, you probably noticed that the food isn’t the same. The plastic containers are nice for containing during transport, but they have a side effect of steaming your food. Pop your sides into an air fryer and they are back to their former glory… and sometimes even better. We also found a bunch a recipes made for the air fryer. This is going to change your whole existence and I never, ever use hyperbole.

A few years ago, I worked on a predictive maintenance project for duct fans. An air flyer is like a normal oven with the addition of a fan. There are fans in almost every major system and when they fail, the system that they are attached to will eventually fail. Fan blades can get bent, dusty, covered in food particles and oils, and damaged over time. The idea behind predictive maintenance is to predict the remaining useful life (RUL) of a machine and perform maintenance and repairs just-in-time. When applied at scale, predictive maintenance strategies provide a substantial return on investment. Studies show that condition-monitoring and predictive maintenance pays for itself and can be developed by anyone using MATLAB and the Predictive Maintenance Toolbox.

Air fryer with vibration sensor connected to ThingSpeak

Check out my Condition-based Maintenance of a Duct Fan using ThingSpeak over at Hackster for a complete tutorial, video, and the source code. I adapted this project for an air fryer, but you could literally apply this technique to any machine that you want to monitor or to predict its remaining useful life. The idea is that fans and monitors vibrate when they are operational. As they degrade, the vibration changes over time. If you train a model with the data, you can use the model to predict the machine’s state, condition, and remaining lifespan. The predicted information can be used to change the maintenance schedule of the machine and optimize repair timing. It’s better to replace a part when the machine is scheduled to be down versus when the machine is in production.

I am using the Particle Internet Button for this project. The button gives me a nice convenient packaging of a Wi-Fi device for connectivity to ThingSpeak, an accelerometer to measure vibration, and LEDs for status lights.

The system uses ThingSpeak channels to store the vibration data, MATLAB / Predictive Maintenance Toolbox to build our machine learning model, MATLAB to analyze live data, and ThingSpeak to visualize our machine state and predictions.

Predictive Maintenance System Overview

Here’s an overview of the steps involved:

  • Data Collection – Your first challenge is to collect data from the system. Attach the vibration sensor package to the machine, connect it to the network, and send the data to an IoT service like ThingSpeak.
  • Data Parsing, Exploration, Feature Extraction, and Training – Now that you have data, you have to figure out what it all means. Use MATLAB, the Predictive Maintenance Toolbox, and the Statistics and Machine Learning Toolbox to parse the data, extract interesting features, and to train a machine learning model.
  • Visualize and Report – Present your findings with charts and gauges. Use ThingSpeak Alerts to send automated email messages about the machine’s health.
  • Profit.

Here’s what some raw data from air fyer looks like.

Air Fryer Vibration Data on ThingSpeak

Vibration data is hard to interpret, but with machine learning, it’s possible to classify what the vibration data means and understand what the changes mean over time. You may even have noticed the sound of these appliances change over time. This is your brain picking up on the changes and telling you, “IT IS TIME TO CLEAN YOUR NASTY AIR FRYER.” Imagine a machine learning model always monitoring the vibration data and alerting you on the changes. Things will be working when they are supposed to be working.

Run two experiments with your fan and collect the data:

  • Experiment 1: Simulating failure due to blocking the airflow (1000 data points) – Collect 1000 data points: 250 – OFF, 250 –ON, and 500 – BLOCKAGE
  • Experiment 2: Simulating failure due to rotor imbalance (1000 data points) – Collect 1000 data points: = 250 – OFF, 250 –ON, 500 – IMBALANCED
MATLAB Data Exploration

When we explore the vibration data, you may notice some things:

  • The variance of the data in the time domain increases as the fan is turned on and as the failure develops in both experiments.
  • In the frequency domain for the experiment 1 data, the peak shifts from 30 Hz to 25 Hz.
  • In experiment 2, the peak in the frequency domain remains at the same point but the amplitude at that frequency increases.

Using MATLAB’s built in Classification Leaner App in the Statistics and Machine Learning Toolbox makes it easy to train a machine learning model with your data.

Classification Leaner App in MATLAB

When everything is running in operation, the system will monitor the vibration data, analyze using the MATLAB analytics that we developed in the tutorial, and predict the time to service. As you use appliances and machines they degrade over time. You don’t always need to clean and maintain. The trick is to do it when you need to do it before catastrophic failure… or going to Amazon and buying a new one.

Time To Service

If you are looking to try this out for yourself, check out the Condition-based Maintenance of a Duct Fan project published on on Hackster. We have included the source code for the device as well as the MATLAB code to make the system to work. This can be applied to other systems and the process applies to many applications of predictive maintenance. Now that you know that these techniques exist, you might be able to find some useful ways to apply them.

Apparently, I like connecting appliances to the internet.