You may have heard the terms machine learning and AI being used interchangeably, but what are they exactly? Machine learning is broadly speaking the use of statistics to make predictions, while AI is more concerned with behaviour (e.g. playing games (like AlphaGO), visual perception, speech recognition, decision-making, and translation between languages).

Machine learning is a subset of AI. Data Science uses machine learning/AI to perform the necessary predictive tasks. To illustrate the relationships between ML, AI, Data Science, Deep Learning and maths/statistics:

There are many potential machine learning applications within industry, particularly as traditional manufacturing pivots to Industry 4.0. Examples of such applications include pattern and image recognition, natural language processing, operations optimisation, anomaly detection data mining, and knowledge discovery.  As the scale of data available in manufacturing environments increases rapidly, machine learning will become a vital element in the everyday industrial workflow.

Types of machine learning

There are two main categories of machine learning:

• Supervised learning and unsupervised learning

How does supervised learning work?

1. We train a machine learning model using labelled data (the "response"). The “machine learning model” learns the relationship between the features and the response.
2. We make predictions on new data for which the response is unknown.

The primary goal of supervised learning is to build a model that “generalises” — i.e., accurately predicts the future rather than the past.

How does unsupervised learning work?

• Extracts structure from data.
• Attempts to represent.
• Does not require past data on the element we want to predict.

Examples of Supervised and Unsupervised Learning

Supervised Learning: Parts Classifier

• Observations: Parts.
• Features: Size and mass.
• Response or target variable: Hand-labelled part type.

Train a machine learning model using labelled data.

• The model learns the relationship between the features and the part type.

Make predictions on new data for which the response is unknown.

• Give the model a new part and it will predict the part type automatically.

Unsupervised Learning:  Anomaly Detection

• Observations: Machine sensor event data, IT web/network events etc.
• Features: Temperature, humidity, electricity usage, network events, measured vibration etc.
• Response or target variable: There isn’t one: instead, we group similar observations together.

Train a machine learning model using unlabelled data.

• The model learns the relationship/similarity between the features and labels them.

Make predictions on new data.

• Give the model a new observation and it will predict the label based on its similarity level to the training data. If it is significantly different, it is classed as an anomaly.

Limitations

Machine learning is not a panacea to solve your problem: there is no free lunch. To build good, actionable models,

• Domain knowledge of the problem is required to frame it correctly and to be able to interpret whether the models are providing high quality predictions for the problem involved. 
• Data, and specifically, lack of good data: if you don’t have enough data, and of high quality, to approximate the problem being observed, any model created will make poor predictions.
• Lack of interpretability: if it’s hard to discern what the models are telling you, and how they made that decision, their usefulness is limited.

Conclusions

Machine learning, if appropriately used, is a powerful tool that can be used glean insights and make predictions for a range of industrial applications. It will allow manufacturers to optimise and improve on processes, workflows and final products, becoming central to the everyday manufacturing workflow in the Industry 4.0 era.