What does machine learning really mean and how does it apply to manufacturing?

Not that long-ago machine learning was a term we only heard in sci-fi movies and in conversations about how the far future of the human race will look like with intelligent robots. By today, however, it has become part of our everyday reality. Just some of the most typical application areas of machine learning are:

  • Face recognition (on airports for example)
  • Text and voice recognition
  • Personalized contents (Facebook feed)
  • Spam filter of your inbox
  • Viewing or shopping recommendations (Netflix, Amazon)
  • Credit card fraud detection
  • Detecting skin cancer
  • Sorting vegetables and fruits

We couldn’t imagine our everyday modern life anymore without the benefits provided by machine learning. But what about manufacturing? Is machine learning (ML) having a crucial role there as well? The short answer is yes. Manufacturing is one of the areas where ML is taking over the role of manual calculations and human decision making. However, there are no machine learning methods that could be universally applied in manufacturing environments. Every company represents a standalone ecosystem of resources, processes and business strategies. There is no simple equation to how to apply ML, everyone has to find their own unique way of applying ML in order to gain real business advantages from it.

The concept of machine learning is inseparable from the phenomena of industry 4.0, big data and smart manufacturing or smart factory as it is called in South Korea. It is an umbrella term for many broad areas that are standalone disciplines on their own, like data mining (DM), artificial intelligence (AI) and knowledge discovery (KD).

So, what is machine learning exactly? The expression is very descriptive since it is about enabling machines to learn in a way human does, to teach them to learn in an autonomous fashion, from the data they are fed. Machine learning is a group of algorithms that are capable of improving their own performance as they learn from the increasing amount of available data. They do all that without relying on a predefined equation. It uses statistical analysis to recognize patterns in the data and to derive predictions, detections, classifications and forecastings from it. These predictions are being continuously fine-tuned as more data is being fed to the algorithm. So, the core strength of machine learning solutions is to automatically learn from and adapt to the changing environments they were designed to operate in.

Machine learning technologies are unique compared to all other process improvement methods and technologies available in manufacturing where a lot of time is spent on extracting information from the available data. Unlike these traditional approaches, ML has practically no data processing limitations. Its bottleneck lies in the process of collecting and processing the data that it is going to learn from. In order to use ML, the system designers have to have a deep understanding of the nature of the data they need, in order to be able to pre-process it for the ML algorithms. The way the data is pre-processed has a critical impact on the outcome of the machine learning environment.

 

Classification of ML techniques

ML has become a wide field of research with many different sub-domains, theories and application areas, so there is not one widely accepted way of structuring the different ML approaches, but many different structures depending on the approach of the researcher. Here I’m going to briefly introduce you one of the most widely used structurings of ML algorithms. It doesn’t contain all available algorithms though, but it gives a good overview of 3 main machine learning categories. When reviewing the different available algorithms in any ML structure, it is important to keep in mind that these algorithms can be combined with each other creating a so-called hybrid algorithm that can have a significantly more powerful and accurate performance when compared to individual algorithms.

Machine learning

Unsupervised learning

Unsupervised learning is a type of ML where the system only has input data but no output variables. The main goal of unsupervised learning is to discover and learn about the hidden structure, patterns, and groupings of data. It describes the structure of ‘unlabelled’ data (data that hasn’t been classified or categorized) without the feedback of an external teacher or knowledge expert.

Supervised learning

Supervised learning is a widely used ML type in data rich but knowledge-spare environments, where the data is labeled. This system uses a set of training examples to learn from. Based on the labeled training material that contains a known set of input and corresponding output data, the algorithm learns to map input data to the output. These models can produce predictions in the presence of uncertainty.

Reinforcement learning

Many theorists consider reinforcement learning as a special type of supervised learning, but the main difference is that in the case of RL there are no labeled examples of ‘good’ and ‘bad’ behavior. This area of ML has been inspired by behaviorist psychology, the system learns how to behave in an environment by performing certain actions and recording the results. So, the idea behind this approach is that an agent is learning from its environment by interacting with it and receives rewards when performing the ‘right’ actions.

 

How can we apply machine learning in manufacturing?

  • With the help of machine learning, you can automate production planning based on the maximum performance of the production machines. It can optimize timing and cost by considering the impact of changeovers.
  • Predictive maintenance is realized by ML where data is collected from the machinery through sensors. Patterns in the performance data are recognized and therefore predictions can be made on when each individual machine is going to need repair. This way, serious and costly faults can be prevented through timely and preventive maintenance. Also, regular maintenance can be scheduled automatically, selecting the least impactful downtimes.
  • ML can allow you to recognize sales trends and patterns that are not noticeable by humans. Foreseeing market trends and sales patterns allows you to maximize the profitability of your business by setting up the most lucrative product portfolio that includes the products with the highest turnover and lowest production cost.
  • Certain algorithms can allow you to classify your customers. For human eyes, often the best customer seems to be the one who pays the most, but in reality, costly changeovers, product configurations, and late payments will define who are the best (most economical) customers.
  • ML algorithms can be of a huge help when it comes to material forecasting. It will be a lot more accurate than any kind of human planning and will eliminate the need for storing surplus material for the sake of security. Precise material forecasting allows you to save significant warehousing costs.
  • Your delivery routes can be optimized by using real-time traffic data and by automating the calculation of the safest, cheapest and fastest routes for each and every delivery.
  • Machine learning is able to capture the productivity and output of each individual production worker. Based on the collected data, ML algorithms are capable of assigning the right people to the right task to provide maximum efficiency.

Conclusion

The above use cases are just some of the examples of how machine learning is shaping today’s manufacturing industries. We have already seen many real-life examples of these application areas but it is clear that there is still a lot to come in manufacturing due to the continued spread and evolution of this field of science. We see a trend in the application of ML which suggests that ML will be applied in several many ways and not just by large, tier one manufacturers but also by middle and small size companies.

The main prerequisites for successfully implementing ML environments are the availability of substantial amounts of data and data storage. Here at Actify, we support and equip manufacturers on their journey to becoming a data-driven smart factory.  Unlike traditional systems that are simply informational, we provide a truly transformational data infrastructure to manage the data handling requirements of tomorrow. Our Centro solution is a scalable platform that connects your enterprise and allows you to effectively collect, share, manage and analyze data.

 

Did you like this article? Read more!

Release the power of your data! Product data intelligence for manufacturers.
Read more

Please complete the form to request more information.