Machine Learning Applications for PLC-Based Industrial Automation

In recent years, efforts have been made towards the integration of smart digital technologies into traditional industrial processes, allowing for the collection of vast amounts of data across all aspects of the manufacturing cycle, in what has been referred to as the “Fourth Industrial Revolution”, or “Industry 4.0”. Machine learning tools, particularly neural networks, have been playing an increasingly relevant role within this context, thanks to their ability to handle complex multidimensional data, uncover hidden relationships extrapolating patterns, predict the future evolution of the process and thus aid decision-making, ultimately improving efficiency. In this work, we aim to showcase how a next-generation programmable logic controller like the Finder Opta - which integrates traditional PLC features with a powerful processor, several connectivity options and access to the Arduino language and ecosystem - enables breakthrough applications within the “Industry 4.0” framework, even supporting small neural networks to handle on-device classification or regression tasks. While deploying a neural network to a resource- and memory-constrained device poses significant challenges, it can be very advantageous as it eliminates the need for a central cloud infrastructure, at least for simpler tasks, with many benefits in terms of latency, security and privacy. This trend, known as edge computing, is indeed a very active area of research. The thesis is structured around three case studies, where we demonstrate how to address the hardware limitations while still deploying meaningful machine learning algorithms, and how the Arduino platform allows for straightforward software development, facilitating the transition towards intelligent manufacturing even for smaller entities. To lay the groundwork and explore the tools at our disposal, the first application involves training an MLP network to classify an input waveform and deploying such a model to the Opta PLC with TensorFlow Lite for Microcontrollers. Then, we transition to a more complex convolutional neural network to infer the rotational speed of two rolling bearings, starting from the spectrogram of an audio recording of the test bench in operation; this represents a preliminary study in the field of anomalous sound detection for predictive maintenance. Shifting focus away from machine learning for the last application, the Opta is tasked with controlling a wave energy converter, exploiting its internet connection capabilities to analyze forecast data about wave height, thereby protecting the system from damage due to excessive stress on the floating device.