<img height="1" width="1" style="display:none;" alt="" src="https://dc.ads.linkedin.com/collect/?pid=402977&amp;fmt=gif">

What is machine learning and how can it help improve performance?

Author: Oliver Jobson

machine-learning-2This article is written for those with little or no knowledge of machine learning, who seek to develop an understanding of the key technologies involved and the potential benefits it has for the organisation.

Machine Learning ≠ Artificial Intelligence

Using the right terminology is a useful first step. Machine learning (ML) is a key technology in the ‘4th Industrial Revolution’, (also referred to as 'Industry 4.0') that is 'blurring the lines between the physical, digital and biological spheres'.

ML is often used interchangeably with the term ‘Artificial Intelligence’ (AI). This has arisen as learning is one of the key cognitive functions that can be mimicked by machines in order to demonstrate AI. As well as learning, the other cognitive functions that may be mimicked are reasoning and problem solving. ML can help an AI system to build an understanding of its environment through learning.

machine-learning

Types of machine learning

There are two main types of ML that are named after the objectives, or purpose, to which it is deployed:

  • Supervised Learning - for making better predictions
  • Unsupervised Learning - discovering patterns and associations
  • Reinforcement Learning - learning to improve on past performance

ML for making predictions

Accurate forecasts and predictions can provide the organisation with a competitive advantage. ML can help the organisation to understand when, and under what circumstance, a production line might produce scrap, or used to produce more accurate forecasts of when your customers will make a purchase. ML can even help an organisation to convert opportunities in the sales pipeline by recommending the best course of action to take next. To do all this ML relies on a computer model (typically using statistical modelling techniques), which must first be trained and tested on a labelled dataset.

A labelled dataset could comprise of historical data from the process in question. This data would contain ‘features’ i.e the measurements that were recorded (sometimes referred to as ‘independent variables’), as well as labels to identify when the desired ‘target’ outcome was achieved (the ‘dependent variable’). For example, for a company producing widgets, a manufacturing scrap prediction model could be trained on historical records of scrapped and not scrapped widgets, using a scrap status of No/Yes (or 0/1 in binary) as labels and the production process parameters as the features.

Once trained, the model is fed live data from the production process to estimate the likelihood, at any one time, of a newly produced widget being scrap. Significantly, the model could also predict a time in the future that scrap may start to be produced. This would be based on how quickly production parameters are changing and when they are likely to be exceeded. Using ML in this way to make predictions is called supervised learning. 

Being able to predict the type of situations where scrap will occur allows organisation to put in place the suitable process steps that will eliminate or significantly reduce these situations from occurring, thereby lowering the cost of quality. For the manufacturer this could mean reducing scrap and increasing machine uptime (and lowering maintenance costs) through optimised service intervals, or even introducing condition-based servicing.

ML is also being used to develop more complex machine vision systems that allow defects to be identified without the need for human inspection; further driving down the cost of quality.

big-data

ML for discovering patterns

On the other hand, unsupervised learning is used for interrogating data sets where there isn’t necessarily a target outcome. Instead ML in this situation uses unlabelled data sets to discover underlying patterns and groupings. Clustering techniques are used to detect natural groupings that exist in the data set. Visualisation of such clusters can help to identify previously unknown patterns.

Pattern discovery can be used in anomaly detection where a system, trained with normal instances, may detect unusual readings - such as sensor readings that correlate to when defects are created in a manufacturing process, or patterns of network traffic across a firewall that may be an indication of a cyber attack. Machine learning may be deployed on 'Big Data' - another popular Industry 4.0 term which refers to the manipulation and use of very large data sets.

ML for improving outcomes

You may already be familiar with the idea of reinforcement learning if you are familiar with Pavlov's dog's. This ML approach requires a certain level of labelling in regards to the data as it is the labelling that provides the opportunity to reinforce the systems 'behaviour'. An untrained reinforcement learning algorithm will likely make a lot of mistakes at first, but the system will learn to understand how the ideal outcomes can be produced and how to reduce the chance of sub-optimal performance.

The rise in popularity of ML

ML has become very fashionable right now due in-part to the cost-effective access the organisation now has to ML's enabling technologies; such as Internet-of-Things (IOT) devices and cloud computing power. The continual stream of rich data and insights produced couldn’t otherwise be easily or cheaply obtained by using traditional methods. Using inexpensive and small network-connected IOT devices and low-cost sensors makes ML possible by gathering data to be processed in the cloud (or other central location). 

The profound impact that ML can have on business outcomes is also driving interest; however, data science and machine learning need to be used correctly. It helps to have the capability to know which questions to ask in order to unearth the real value in the data. Unipart’s wealth of experience in developing problem identification and problem solving techniques, which includes Six Sigma, means that we are well positioned to ask the right questions to uncover valuable insight.

Unipart’s data science team use machine learning techniques within a wide array of scenarios across our own operations and those of our clients in order to understand, visualise and improve business processes and outcomes for customers.

The Unipart Way: performance improvement that sustains

Get in touch

Please fill in the form below and a member of our team will respond to you shortly.