The South African manufacturing sector is diverse and contributes 11.8% to gross domestic product and 10.3% to employment. Although these figures are small, compared with the figures of other countries, embracing machine learning could enhance the sector and assist in its growth, says University of Pretoria (UP) Department of Engineering and Technology Management senior lecturer Schalk Grobbelaar.
The largest manufacturing spheres in South Africa are food and beverage, basic iron and steel, nonferrous metal products, metal products and machinery, petroleum and chemicals products, as well as rubber and plastic products.
Machine learning is a subset of artificial intelligence (AI) and consists of supervised learning, unsupervised learning and reinforcement learning, explains Grobbelaar.
Supervised machine learning entails the recognition of patterns and automating decision-making, with many potential applications in the manufacturing industry.
The manufacturing sector can benefit from its ability to sort through material or goods such as identifying defects on manufactured goods and automatically sorting it into separate bins.
“It is especially useful in manufacturing applications with very high unit-per-time capacity,” says Grobbelaar.
Machine learning is also useful in condition-based monitoring in maintenance activities. The AI system can learn to identify patterns that indicate that a component is about to fail.
Grobbelaar says machine learning is mostly useful in data-rich environments and where it is possible to measure a specific aspect.
The application of computer vision –the ability of computers and systems to derive meaning from digital images, videos and other visual inputs – is growing in popularity.
“The combination of photogrammetry and machine learning allows manufacturers to measure, grade, identify, sort and even adapt machining functionalities at speeds beyond what is possible with the human eye,” Grobbelaar says.
Machine learning also allows for fast decision-making. This is especially useful in situations where the speed with which a decision must be made is not possible for humans.
“Think, for example, of fast-moving consumer goods. In many cases, technology suppliers are replacing rules-based programming with machine learning,” he adds.
Historically, programmers had to clearly define specific rules. However, machine learning programmers require only a training dataset, consequently allowing for the system to develop its own rules.
In the sawmill industry, for example, manufacturers must grade their lumber to ensure that it complies with strength requirements. They can manually develop a training dataset, and then use machines to grade several boards.
“The machine learning system can then use this as an input and automatically grade new boards based on the patterns identified with the training dataset,” Grobbelaar explains.
While the introduction of machine learning is beneficial to the sector, there are challenges that prevent its implementation.
Data science skills is most probably the biggest challenge, says Grobbelaar.
He suggests that developing these skills can be addressed through training programmes.
Several universities in South Africa have developed and are developing data science as a specific area of study, such as UP, which is providing undergraduate and postgraduate courses and degrees in Data Science.
Public and private funding for such training initiatives is critical to ensuring that the initiatives can be developed, he says.
“The biggest threat regarding the implementation of machine learning to South Africa is to not implement it,” Grobbelaar declares.
Failure to implement machine learning in South Africa will threaten manufacturing competitiveness and, consequently, job security.
If South Africa fails to remain competitive, it will fall behind other countries that have implemented machine learning, he points out.
However, it is important to note when using machine learning that the decisions made by the system are highly influenced by the training dataset. Therefore, if there are biases in the training dataset, the system will continue with those biases.
Grobbelaar cites the automation of recruitment decision-making as an example. With machine learning, the curriculum vitae of candidates are automatically classified. It is, therefore, conceivable that most companies would use their best performing employees to develop a training dataset.
However, depending on the data included in the training dataset, the system might apply biases, especially when there are small representations of specific groups within a training dataset.
Therefore, when using machine learning, it is important for the implementers or companies to consider whether the decisions being made by the system will have any moral or ethical implications, and how such implications will be managed, says Grobbelaar.
Machine learning is highly dependent on the data inputs, which include the original training dataset and new data on which decisions must be based. If this data is unreliable, then the decisions by the system will also be unreliable, he adds.
Owing to the potential for biases and unreliability, implementers should not be over-confident, but rather approach machine learning with a full understanding of how it works and its potential pitfalls, Grobbelaar concludes.