Machine learning has been researched for decades, but its use in applying artificial intelligence in industrial plants and asset operations and maintenance is now advancing exponentially. This is due to the growth in Big Data, the expansion of the internet of things (IoT), the ability to provide the processing power needed to analyze larger data sets, the availability of machine-learning methods, and the need for superior predictive and prescriptive capabilities required to manage today’s complex assets.
While machine learning has typically been linked with industries such as transportation and banking (think self-driving cars and fraud monitoring, respectively), there are many uses for machine learning within the industrial sector. We’ll focus here on some of the principles within machine learning and on industries that are primed to take advantage of the application of machine learning to maximize the benefits it can bring to improve situational intelligence, performance, and reliability.
The route to deeper understanding
Machine learning makes complex processes and data easier to understand and is ideal for industries that are asset- and data-rich. A great deal of data from various data sources is required in machine learning, and a data scientist or analyst may be needed to help set up and interpret the results. While it is possible to build your own ML platform, this design takes time, specific skills, and investment in a tool such as Microsoft Azure for a secure, private cloud platform for developers and data scientists. Alternatively, purchasing machine-learning capabilities off the shelf as part of an asset performance management software solution or outsourcing ML to a third party are options, provided you ensure that you can supply input from in-house skills.
Supervised machine learning
Supervised learning is the most common technique. There are many steps involved – uploading data sets, training the data, choosing the algorithms, visualizing the data, and more – to lead you to finish with the desired level of accuracy that can then be applied to new data to start the predictive stage. Supervised learning encompasses two techniques:
Classification: Classification is typically used for data that can be categorized – for instance, in determining whether an email can be classified as worthwhile or spam. Common algorithms used within classification techniques include logistic regression, decision-tree, and neural networks.
Predictive maintenance falls into the classification approach because it has several possible outcomes that can be categorized as potential equipment problems generated by various parameters, such as levels of risk, health indices, reasons for failure, etc.
This is the same way machine learning is used to predict a medical diagnosis: by identifying symptoms and issuing a diagnosis. Predictive maintenance has a multiclass classification because there are multiple categories or reasons why a piece of equipment will fail, whereas with the email example, there are only two binary states, genuine or spam.
Regression: Regression is used when data has a range, such as with sensor or device-driven data, and it is used to estimate or predict a response from one or more continuous values. The most common algorithm for regression is linear regression. This is one of the most easily applied algorithms because it is easy to interpret and quick to implement. Measuring and predicting a temperature is an example of linear regression because it has a continuous value where the estimate would be easy to train.
Clustering: While supervised learning usually has an expected outcome to work from and can be trained, unsupervised learning is usually applied when the specific goal is not yet known or the information of the data is unknown. This means that the data is grouped or clustered together and then meanings are deduced from patterns hidden in the input data by putting the data into similar groups.
Neural networks: A neural network, which can be both supervised and unsupervised, is one group of algorithms used for machine learning that models the data using graphs of artificial neurons (i.e., neurons that are a mathematical model that simulate approximately how a neuron in the brain operates). For instance, if a brain were a city, neural networks would be the transportation routes, or networks, within it.