IoT predictive maintenance for airplanes, bringing deep learning to the edge

The internet of things (IoT) has brought new opportunities to create a positive impact in the world around us; it has provided us with new ways to measure the world, and with them, we have gained knowledge to make our lives better.

However, we cannot take advantage of everything that the Internet of Things has to offer without intelligence and machine learning. IoT is not only about collecting the data, but also about gaining knowledge.

Using sensor data becomes worthy when we can predict, control, and make valuable decisions. This is where machine learning comes into the picture. Deep learning models can provide useful information using large amounts of data. Therefore, they fit as a perfect solution for providing valuable insights when processing the sensor data gathered by the IoT devices.

One of the many ways IoT has created a positive impact in the world is by making industry sector smarter. One example of this is how, thanks to IoT, we can avoid failures in machines using Predictive Maintenance.

Failure prediction aims to provide early warnings for potential failures. In the past, to avoid failures, companies used schedule-based maintenance. However, this didn’t solve the problem since critical issues often arise between maintenance intervals.

Nowadays, companies can take advantage of using sensor data to shift maintenance into a predictive approach. They can prevent unplanned reactive maintenance due to unexpected failures and eliminate unnecessary costs that come from doing unnecessary preventative maintenance.

Many benefits come from intelligent, real-time decision-making, such as fixing machines before they fail. Nevertheless, real-time decision-making in IoT systems is still challenging due to latency, connectivity, power consumption, etc. A way to mitigate these difficulties is to bring the predictive models to the point where the IoT devices connect to the network, also known as the “edge.”

However, bringing machine learning models to the edge raises a new challenge: deep learning models often require significant amounts of computational resources, memory and power to train and run. IoT devices are not able to meet the hardware requirements that deep learning models need. Then, how can we bring deep learning models to the edge without compromising their performance? In our session, we will address this question.

We will present an end to end solution for aeroplane turbine predictive maintenance in the edge and we will explain how to overcome the challenges previously mentioned.

Moreover, the session will cover the following points:

– Introduction to predictive maintenance.
– How to use Deep Learning and Recurrent Neural Networks to solve predictive maintenance problems.
– Machine Learning workflow for the solution proposed. We will provide all the details of our model’s implementation and the technology used: Databricks, Python and the Deep Learning Framework, Tensorflow:

o Data Pre-processing and Feature selection
o Normalization
o Model selection
o Training & Testing.
o Evaluation Metrics.
o Deployment of the model using a lightweight version of Tensorflow: TFLite. This new framework provides the tools needed to deploy a lightweight model in the edge without compromising its accuracy.

– Development of different IoT edge modules and how to deploy them in a Raspberry Pi.

o Sensor engine module
o Predictive module using a pre-compiled tf-nightly version of Tensorflow for ARM32 architecture.

– How to send telemetry in real time to the Cloud from our Raspberry Pi.
– Analytics & BI: Visualize data and model predictions in real time. Monitor possible failures to plan the technical repairs.