You probably have heard or read about artificial intelligence (AI), machine learning (ML), and Industry 4.0. We often start to imagine big supercomputers, complex programming algorithms and software connected with various sensors.
Over the last decade, ML techniques have made a huge leap forward as demonstrated by deep learning (DL) algorithms implemented by autonomous driving cars, or by electronic strategy games. Hence, researchers have started to consider ML also for applications within the industrial field, and many works indicate ML as one of the main enablers to evolving a traditional manufacturing system up to the Industry 4.0 level. Nonetheless, industrial applications are still but few and limited to a small cluster of international companies, and it seems even further from compressed air where we need to first eliminate leaks and ensure optimal compressor operation.
What is ML and how it can be applied? Will it make our life easier or more complicated?
Murphy defines ML as set of methods that can automatically detect patterns in data, and then use the uncovered patterns to predict future data or to perform other kinds of decision making under uncertainty.
Unlike the traditional way, engineers analyze and monitor the system using their own experience, software, and calculation algorithms, ML algorithms can process more data many times faster, learning, adjusting and analysing along the way. In the meantime, you can enjoy your cup of coffee and think about creative solutions for your production.
Probably, in a couple of years algorithms will be learning from each other so we do not need to even think about this question. But today we need to carefully develop, test and train a diverse model to achieve good results. The number of methods and models is vast from “Ant colony optimization” to “Regression” and “Genetic algorithms”.
The good news is that machine learning algorithms for application in Industry often do not require powerful computers: they can be integrated on a very simple microprocessor or your Motorola phone from 2000.
System design using digital twin
Design of the compressed air network can be a hard task while talking about the project of the industrial plant. It is not only about good pipes, accessibility for maintenance, and heat recovery, but also correct network topology, sizing, correct pressure levels and assembly components with the lowest pressure drops. The full procedure can lead to a complex problem and might be very time-consuming.
Thanks to known calculations and existing standards it is possible to calculate and choose the right components and operating pressure levels. In that case, modelling software will certainly help there. One of the known capabilities of the software is integrated machine learning algorithms for predicting behaviour or in this case, predicting missing concentrated parameters in a large-scale compressed air network.
Similar to image recognition, this problem can be solved using an autoencoder. It is a type of artificial neural network used to learn data encodings in an unsupervised manner. This type of method is also used for quality control and can be applied in filling the information gap about optimal system design.
System installation and maintenance
Maintenance is the second-highest cost of a compressed air system after energy. Pneumatics are robust, which makes them more reliable, but at the same time require more attention to maintenance due to the challenging localization of the problem. When we see the pressure drop inside the machine, we usually cannot identify whether it is a leak in the network, problems at the actuator, valves or pipes. High maintenance costs can be reduced by monitoring each component. But in this case, we need to install more sensors on each component which would lead to higher maintenance costs and downtime.
ML can solve this problem by learning sensor curves from each other thereby reducing the number of required sensors. In the example of the cylinder, it is possible to teach the algorithm how the position curve looks for various actuators to obtain the information without the need for measuring. In this case, the regression method was used to train the model and recognize positioning patterns using pressure curves. It is also possible to obtain pressure curves from the positioning.
As with other possible scenarios, it can be applied to other components such as blowers, vacuum ejectors or entire systems to learn about flow curves based on the pressure sensor. Based on this data, monitoring of failures can be achieved more easily and faster rather than in conventional applications.
Connecting various actuators and machines can also provide a cross-learning algorithm where one cylinder, for example, learns from another.
All in all, ML also finds its application in failure detection. Large pipes nowadays are inspected using wireless sensors with integrated ML algorithms for problem localization. It can be downscaled to failure detection inside the components and provide valuable data to reduce cost, energy consumption and CO2 emission of the production.
Elvira, CEO of DirektIn