Machine Learning and Maintenance Prediction in Manufacturing

Jan Říha

Senior HW Developer

How to prevent breakdowns, and therefore save time and money?

Machine learning and artificial intelligence can help us monitor and, more importantly, correctly diagnose device deviations.

No items found.

In the early afternoon, the assembly hall of a manufacturer in the automotive sector suddenly faces an unexpected issue. The gradual loss of pressure in the compressed air piping system indicates more serious problems, which escalate a few minutes later in a complete loss of pressure in the pipeline, which will prevent the use of all pneumatic tools. The production stops, and a race against time begins. Every hour costs the manufacturer a significant amount of money. This unenviable situation is caused by the air compressor failure.

Early maintenance could have prevented all this

The signals indicating this failure occurred weeks before this event. But for this part of the equipment, no tools for early failure detection were used. You might think about dozens of similar situations that might cost your business considerable financial and time expenses. What measures could have been taken to prevent this? Condition monitoring is a process of measuring the parameters of machines such as the temperature, vibrations, pressure, el. current etc. to detect and prevent failures.

Examples of parameters that can be used for condition monitoring.

El. current
El. current
No items found.

Not just monitoring but proper diagnostics

You might argue that you use a bunch of sensors to monitor your critical infrastructure. Well, that is a good approach. But in such a large amount of data, looking for various outliers is like looking for a needle in a haystack. To overcome this problem, traditional condition monitoring must be combined with data science and machine learning.

What anomaly means

In the context of condition monitoring, we must also mention anomalies. A classic definition of an anomaly was given by Douglas Hawkins: “An anomaly is an observation which deviates so much from other observations as to arouse suspicions that it was generated by a different mechanism.”

There are a number of diagnostic approaches that can help reveal anomalies. A few of them are listed below.

1. Decision trees

A systematic approach to identifying the main cause of an event, with the use of a fault tree diagram.Needs a lot of domain expertise - Expensive to build and maintain.

2. Rule-based

Use IF/THEN statements applied to the data to distinguish between normal operation and fault state.Needs a lot of domain expertise - Expensive to build and maintain.

3. Model-based

Machine learning techniques using trained models for normal operation. In case of anomalies, the model output indicates the deviation.Needs high computing resources - Requires a good amount of historical data.

No items found.

The use of machine learning techniques in anomaly detection has become quite popular in recent years.

However, not all of the known ML methods are suitable for this. In the real world we typically miss enough data representing the fault state. For this reason, using supervised learning methods that map input to output based on many examples in the learning process is not a good choice.

Amore suitable method for anomaly detection is the autoencoder.

The autoencoder (typically represented by the neural network) learns how the normal operation is represented within the input dataset. When an anomaly occurs, the autoencoder output displays a large reconstruction error.

Nowadays, condition monitoring is typically used on larger and more expensive machines.

For smaller applications, often no precautions are taken. This approach is sometimes called "run to failure." However, this does not mean that the consequences of such a failure may not be severe.

Edge computing or TinyML are examples of technologies that aim at these cases.

Edge computing is an approach where data processing takes place as close to the end device as possible – this is the opposite of cloud computing. When edge computing runs on low-power hardware (e.g. a microcontroller) and involves machine learning, we often call it TinyML.

No items found.

This approach can provide several key advantages useful for smaller or remote applications.

  • It can run in the offline mode 
  • It has a lower implementation cost 
  • It has low processing latency 
  • It provides enhanced security

Chipmakers realize that TinyML will become a fast-growing segment in the upcoming years. Therefore, many have started offering specialized AI processors supporting the energy-efficient running of ML algorithms.

To provide a more practical guide to the TinyML world, let us have a look at several examples of development boards supporting low-power ML algorithms processing:

max 78000 fthr

Dev. Board type: MAX78000FTHR
Manufacturer: Maxim Integrated
Chip type: MAX78000
Built-in sensors: Camera, Microphone

The MAX78000FTHR is a rapid development platform to help engineers in quick development of ultra-low-power, artificial intelligence (AI) solutions using the MAX78000 Arm Cortex-M4F processor with an integrated Convolutional Neural Network accelerator

Google Coreal micro

Dev. Board type: Dev Board Micro
Manufacturer: Google
Chip type: Coral Edge TPU
Built-in sensors: Camera, Microphone

The Coral Dev Board Micro is a microcontroller board with a built-in camera, microphone, and Coral Edge TPU, allowing you to quickly prototype and deploy low-power embedded systems with on-device ML inferencing

Syntiant TinyML Board front

Dev. Board type: TinyML board
Manufacturer: Syntiant
Chip type: NDP100
Built-in sensors: IMU, Microphone

Syntiant’s Tiny Machine Learning (TinyML) Development Board is the ideal platform for building low-power voice, acoustic event detection (AED), and sensor ML applications

AI Sensor board

Dev. Board type: AI Sensor Board
Manufacturer: Eta Computer
Chip type: ECM3532
Built-in sensors: Camera, Microphone, light sensor

The ECM3532 AI Sensor board is an ultra-low-power AI platform with sensors that can run many algorithms: sound classification, keyword spotting, activity classification, context awareness, defect detection, and others.

Platforms listed above contain a low-power microcontroller and a hardware accelerator supporting low-power Convolutional Neural Networks inference. It can be built as an all-in-one chip (MAX78000, ECM3532) or split into a control MCU and a separate accelerator chip.

All these boards already include sensors such as a camera, microphone, or accelerometer, and of course support connection of additional external sensors.

Due to the low power consumption, it is able to power these boards only from a battery and/or combine them with energy harvesting. We can easily build various always-on smart sensors, which will also find use in maintenance prediction. Such hardware enables easy deployment of AI and machine learning in applications where cost was a major barrier in the past. Together with the Internet of Things, it forms one of the cornerstones of Industry 4.0.

Here at Consilia, we also deal with the AI/ML deployment in industrial, medical, and other applications. We can offer case studies, the proof of concept design, or directly support customer product development in this field. You can read more about our services in hardware and software development.

Next Interesting Topics

A Different Way of Working from Home

PCB designer from Consilia works from his garden house. The walk from his family house takes just about 10 seconds, but, symbolically, it divides his two worlds.

“It is a real job, just without daily commuting, which would take me about three hours a day.”

What determines a PCB design for the automotive industry? Robustness first and then price.

What makes working for global automotive leaders challenging if you are a senior PCB designer? We asked Petr Horák, the PCB designer of the Consilia company, who has been specializing in the PCB design for the last twelve years.

“The automotive industry is specific as all your products have to last at least fifteen or even twenty years.”

I need to heat up my brain. I would not enjoy making the same boards over and over again.

Petr Horák doesn't like to go down smooth ways.

“I like project thinking and focusing on one specific thing. But at the same time, after a while, I need to switch gears and start focusing on a completely different segment. That's why I'm very keen on working on a variety of project types.” says the PCB designer, who has now been working for Consilia for six years.

TETRA base station that goes to the edge of physical limits

The ground TETRA base station, whose development we worked on between 2012-2017, thanks to our oscillators combined with excellent energetic efficiency, is still at the top among the products which can be made with current technologies.