Contact us
Our friendly team would love to hear from you.
A machine learning module development for Industry 4.0.
A Polish company turned to us with the task of a machine learning module development for Industry 4.0. This smart solution should ensure that the production line runs continuously and minimize the downtime by introducing predictive maintenance features. Customer receives real-time analytics and anticipates working anomalies to correct the line operation in time and schedule repair works if needed.
Manufacturing
Mass production
3 AI/ML specialists
1 Backend developer
1 Frontend developer
8 months
Poland
While working on the project, our team encountered several problems. The main one was that every machine in the production line could be in a different state, based on the task it’s processing at the moment. Based on this – some vibration readings which were fine for one state could be anomalies for another and vice-versa. Solution which was already implemented in the system was capable of detecting only two states – whether the machine is working or not – which was totally insufficient.
As sensors could be attached to various types/models of production line machines, the solution should be able to detect and classify the states automatically.
The main task was to create a universal model that can work with almost any type of sensor that provides time series data, for example 3-axis vibration sensor, temperature sensor and many others.
The next point was the amount of data gathered in the state and anomaly prediction system. Each measurement point provides one to three measurements per second. The number of sensors per machine can be as high as ten, and the number of machines on a single production line very often exceeds ten pieces. So an entire production line may have hundreds of measurement points that need to be analyzed in real time.
Besides, the work process was hampered by the fact that the sensors were on the customer’s side, and we only could work with the data that was sent to us.
Moreover our client had a very small amount of data that was not labeled and required additional steps to prepare the correct data to build the models.
The last and probably most important task was to deliver a solution that would be simple enough that non-machine learning specialists could analyze the results obtained from the predictive engines.
Therefore, our team began researching supervised machine learning methods to select the most effective classification models. This approach helped to narrow the scope of the potential use of regression models to determine the state of a machine in real time.
In the beginning, the data sources provided by the client had no labels and were quite small. We offered the client to integrate the system with a cloud infrastructure with all the necessary anonymous data stored in Azure DataLake (the solution was economical and allowed unlimited storage). To work with data, we used our open source library CogniPy.
In addition, thanks to Politechnika Świętokrzyska, with whom we work closely on many other projects, we gained access to their laboratory production line to run several tests. This allowed us to collect valuable data to train and validate our models and deliver the project on time and according to the agreed requirements at the beginning.
Previously, the client worked with an LTE connection to send the latest measurements from sensors. This entailed several problems, since measurements could be delayed, some of them could not be available at all due to less network stability. In such scenarios, data was sent without confirmation that the target service had received it. When building regression models, delays and missing samples required special attention to ensure satisfactory results. Standardization and appropriate interpolation algorithms were used. In addition, the client plans to switch from LTE to 5G in the near future. The system will be implemented in such a way that it can handle a much larger data stream.
As a result, the client got an IoT solution that shows easy-to-understand state prediction parameters and does not require special skills to work with. The web app warns him about anomalies and machine states in production within 1-5 seconds. Thanks to flexible settings he can choose the notification method himself (sms, e-mail, others). It is intuitive and easy to use, with an admin panel where anyone gets access to real-time analytics (gathered data for each machine and sensor, their state and downtime period if any). Besides, one can add other people and assign them roles with different rights permission (Admin/Super Admin/Operator), change settings in accordance with the data being collected, and much more.
The client was satisfied with the result of our cooperation and is going to continue it. There are plans to launch new versions of the web application with deeper and more detailed analytics and more specific machine states on the line.
Let’s talk about your project, just click the button and we will reach you as soon as possible.
Contact Us
