Deep Learning for Machine Manufacturers

Deep Learning (DL) has come a long way since its revival in 2012. This was the year when AlexNet won the ImageNet challenge. Today, small microcontrollers and single board computers can support DL inferences at the edge. The possibility of hosting DL models at the edge has opened up application possibilities for machine manufacturers.

Why use Deep Learning for Machines ?

Vision and acoustics are two areas where DL has had considerable impact. Come to think of it, this provides eyes and ears to any machine! Traditionally, it has been challenge to create systems which can recognize images and sound. DL has made this task feasible and simple to implement. As a team, vision and sound can be used in multiple ways. They can be machine sensors (such as detecting weeds in an agricultural machine), distinguish between normal operating and faulty conditions (based on sound coming from a gearbox) or assist the machine operator perform a complex task (replacing a part). The list of ways in which DL can assist in designing better machines is quite long.

Today, we stand at the cusp of a DL revolution. The way microcontrollers changed machine control from discrete analog components to single chip digital controller, DL is likely to have a similar impact on machines.

Non-contact sensors are good

Adding contact type sensors (such as load cell, fluid level...) to an existing machine is not a trivial task. DL sensors (vision and acoustic) are non-contact sensors. As a result it is very simple to install them on a new machine or to retrofit them on an existing machine. Everyone loves an upgrade that does not involve downtime !

DL Models for Machines

Convolutional Neural Network (CNN) and Recurrent Neural Network (RNN) are two DL model types which provide a starting point into the world of DL. CNN provides the ability to discern images, while RNN helps detect sequences. The table below shows examples of some machines and the models they use.

As you see, all machines can make use of vision and sound to become better. A street light can use sound to detect and record car crash scenes. A factory robot can use sound/vision to understand and confirm sequence of operations. A car can use a sound sensor to detect issues with the transmission.

Who programs the DL system for the machine manufacturer ?

For many machines, it may be best that a controls engineer decides on how to incorporate DL functionality. Once this can been done, the controls engineer (with basic DL training on data acquisition and transfer learning) can go about putting together a basic DL system for the machine. For more complex systems, external help can be sought.

Interfacing DL with existing control system

Most machines use programmable logic controllers (PLCs) or single board computers as their brains. The output of a DL system can interface with these existing machine controllers via Digital Input (DI) or Analog Input (AI). For a classification application, digital inputs work best (connected to Softmax) and regression applications would interface with analog input of machine controller.

Watching the machine remotely

The DL system we discussed is deployed as an edge systems on machines. If there is an abnormal condition or an issue that needs someone to take a look, the machine communicates to a cloud based IoT hub. For a machine manufacturer, the ability to remotely interact with a machine is like parents using a baby monitor to keep watch on their newborn.

---------------------------------------------------------------------------------------------------------------