What is Feature Engineering?

Feature engineering is recognized by data scientists as a critically important part of designing machine learning algorithms.

To better understand feature engineering in manufacturing, let’s start by defining what a feature is:

On the most basic level, if we take as an example a generic dataset of machinery sensor data like the one below, a feature is simply one of the columns.

Feature engineering: A column of features within a dataset
A column of features within a dataset.

In other words:

a feature is a measurable attribute of the asset or production process that we’re aiming to analyze.

Features can be as simple as temperature, vibration, and pressure readings, but in many cases we will want to add contextual information to the data. In manufacturing, for example, adding categorical information is one technique used by data scientists to make data more contextually meaningful.


Feature Engineering in Manufacturing

In Industry 4.0 applications, machine learning algorithms are used to solve specific sub-optimum issues in maintenance, quality, and overall performance. The features determine the focal point of the algorithm’s learning process.

For this reason, the definition of features greatly affects the performance of a machine learning model, and most importantly, how that model will help us solve a manufacturing problem.

While this explanation might seem simple, successfully engineering features is one of the most challenging and time-consuming parts of constructing a successful machine learning application.

Feature engineering is extremely challenging because:

When trying to represent the physical world as a virtual model, there’s always going to be a gap.

What makes feature engineering inherently complex is that it involves the task of representing real-life manufacturing characteristics as purely numerical data.

The goal in feature engineering is to build a bridge that closes this gap between the physical and virtual versions of a manufacturing environment in as accurate a way as possible.


Why feature engineering is critical to machine learning that works

Since data is the fuel of any AI system, there’s a tendency to emphasize the importance of screening, aggregating, and categorizing data so that it’s “clean” enough to allow an ML algorithm to operate successfully.

The problem is that these processes demand time and significant development.

When feature engineering is performed properly, a sub-optimal model can still produce excellent results, even when the data being fed into the algorithm is considered messy.


Using feature engineering to solve manufacturing problems

Applying feature engineering to manufacturing problems is an iterative process that progresses by cycling through a number of activities including:

Research – learning about how features have been used to transform data in other similar manufacturing problems.

Construction – defining the features to be included in the model. This can be done manually, through automation, or by using a hybrid of both approaches.

Focus – comparing the value of certain features and combinations of features by employing importance scorings and a variety of feature selection methods.

Evaluation – monitoring the accuracy of the model as it works on new data, each time using a different set and configuration of features.

An important step in successfully applying machine learning to manufacturing is to target a well-defined problem. This could be to reduce the number of times a pump’s pressure exceeds a given threshold in a production line. Or, to predict whether the number of quality rejects is likely to exceed 2% in the next 30 minutes.

Sticking to a well-defined problem will keep you on track when it comes to comparing the effectiveness of various models so that you don’t get locked onto trying to optimize one specific model.

Another critical factor is to have a test harness in place. This will allow for the objective testing of models, and will be the only way to gauge your feature engineering efforts.


Manual Vs. Automated Feature Engineering

When performed manually, the process of feature engineering can be prone to error. In addition, manual feature engineering is problem-specific – the algorithm cannot be applied to solve other manufacturing issues.

Another shortcoming of manual feature engineering is that it’s limited by the imagination of those working on the problem. You can only think up so many features, before feeling that you’ve done enough.

With automated feature engineering, none of these obstacles exist. A group of related tables can provide data for the automatic construction of hundreds, if not thousands, of useful and interpretable features that can be applied across a range of manufacturing problems.

The ability of automated feature engineering to produce more meaningful features that can also be applied to a large number of manufacturing problems makes it an extremely valuable tool for improving maintenance and overall plant/factory efficiency.

Cut your production losses with machine learning built for manufacturing:

Get a 1-on-1 demo of the Seebo platform and see how accurate and timely alerts can significantly improve maintenance, product quality, and profitability in manufacturing.

New call-to-action