Machine learning is a category of artificial intelligence that includes a number of algorithmic approaches. In manufacturing, the 2 most common approaches are supervised and unsupervised learning.

There are endless application opportunities in industry for machine learning. Here are some examples:

  • Predicting process disturbance in chemical production
  • Predicting quality failures in a production line
  • Predicting production waste in food or PCB manufacturing
  • Predicting asset failure in a power plant
  • Determining the parameters of a “golden batch” for optimal production throughput

The topic of supervised vs. unsupervised machine learning is actually a somewhat contested one in the Industry 4.0 domain. The reality is that there is no one-size-fits-all machine learning technique that can meet the requirements of every type of manufacturing application.

So, what’s the plan of action?

Process engineers should aim to improve their understanding of machine learning techniques to support decisions regarding the use of AI in production optimization for their specific manufacturing challenges.

While this topic cannot possibly be covered completely in a single blog post, this article aims to touch upon the basics of common ML approaches as they pertain to manufacturing, and to describe when each might be appropriate.

Supervised Machine Learning

“Supervised Learning” describes a relatively didactic process by which predictive machine learning models are developed. For this type of machine learning, historical input and output data are made available to the model.

The method used to create an algorithm from a training dataset resembles a teacher guiding a student to reach a specific goal. The “student” algorithm progresses by making iterative predictions based on the training data, and is corrected by the “teacher”.

Supervised Learning problems can be split into 2 main types:

a. Classification – used when the output is categorical such as “normal” or “warning”.

An example of a classification algorithm is one that receives sensor information as input, e.g. pressure, flow rate, and vibration velocity, acceleration, and displacement, and determines the asset health of a machine.

b. Regression – used when the output is a continuous value such as temperature, voltage, or rpm.

An example of a regression algorithm could be one that receives a component’s code number and performance history as input, and predicts the component’s next malfunction. (An algorithm like this could be used to inform maintenance scheduling.)

Classification vs. regression in machine learning

Unsupervised Machine Learning

In unsupervised learning only input data is required. The goal is for the algorithm to do the work and discover the innate structure of the dataset – to model the distribution of the data and automatically provide insight into correlations.

Like supervised machine learning, unsupervised machine learning problems can be split into 2 main types:

a. Clustering – used to discover groupings found in the input data.

In manufacturing, clustering is used to detect behavior anomalies in the production process and equipment. Using measurements from sensors on a production line, clustering can detect and analyze anomalies/outliers, in turn identifying the root causes of process malfunctions or equipment failure.

Machine learning algorithms - Clustering

b. Association – used to discover rules that can describe relations in the distribution of the input data.

An example of association can be any instance of pattern/behavior detection such as the rise in a pump’s pressure as a result of a temperature increase in a cooling vessel earlier in the process.

Supervised Vs. Unsupervised Machine Learning

Semi-Supervised Machine Learning

In semi-supervised machine learning, labelled and unlabelled data are used together to train the algorithm.

Labelled data significantly improves the learning process of an algorithm. The problem is that large labelled datasets are labor-intensive to create.

This is why semi-supervised machine learning can be very advantageous. Data scientists have found that even when a small group of labelled data is used for training in conjunction with a large unlabelled group, learning accuracy is greatly improved.

 

Reinforcement Learning

Unlike cases where the input is formally fed to an algorithm, in reinforcement learning the algorithm receives input based upon experience.

For example, a robot (agent) can be given the task of learning how to connect two components together (reward). The robot can start off without any data about the task, but through experimentation (actions), will start to collect data about its movement, surroundings, and how the two components interact (observations).

When an action is taken that leads to the two components connecting, or coming close, the data related to that action is labelled accordingly and analyzed. As the robot continues to take more actions and record more data, it improves its knowledge about its task.

Human in the Loop

Unfortunately, the above scenario can be very challenging when it comes to real-world problems like the ones we see in manufacturing.

An algorithm can only perform at the level of our input definitions – how we define the reward, the methods of analysis, and other feature engineering attributes. In manufacturing, with an abundance of parameters affecting one another, it’s extremely difficult to account for everything when building this type of model.

Sometimes, it’s easy to see what an algorithm is doing wrong when you’re observing it objectively. This is the idea behind Human in the Loop (HITL).

With HITL, machine learning applications leverage human knowledge to rule out the obvious “bad ideas”. Instead of investing endless time in attempting to perfect a model, human-sourced experience can be consolidated with the algorithm’s process. This leads to improved results and a more efficient learning process.

 

Choosing the Right Machine Learning Algorithm

In manufacturing, a large number of factors affect which machine learning approach is best for any given task. And, since every machine learning problem is different, deciding on which technique to use is a complex process.

In general, a good strategy for honing in on the right machine learning approach is to:

Evaluate the data. Is it labeled/unlabelled? Is there available expert knowledge to support additional labelling? This will help to determine whether a supervised, unsupervised, semi-supervised or reinforced learning approach should be used.

Define the goal. Is the problem a recurring, defined one? Or, will the algorithm be expected to predict new problems?

Review available algorithms that may suit the problem with regards to dimensionality (number of features, attributes or characteristics). Candidate algorithms should be suited to the overall volume of data and its structure.

Study successful applications of the algorithm type on similar problems.

 

Interpretation is Key

How we interpret the algorithm’s output is crucial to how that algorithm helps us solve real-world manufacturing problems.

It’s important to keep in mind that the output is the result of how the algorithm was defined, how the data was collected and aggregated, and how the output is presented.

The interpretation stage holds a number of risks such as overfitting, which can distort our understanding of the results.

 

Machine Learning Approaches Used in Manufacturing

As in many cases of applied mathematical theory, the answer of which machine learning algorithm to use in manufacturing is the unsatisfactory “it depends”.

Every industry, facility, production line, and problem has its own characteristics. Accounting for as many of these factors as possible will improve the chances of building a system that can provide the desired results.

The decision is also affected by business factors, industry regulations, and the availability of expertise. Keeping sight of all of these parameters, and being able to come up with a machine learning solution that can meet the respective demands, will generate the most value.

 

Getting Started with Machine Learning in Manufacturing

For many manufacturers, the diversity of machine learning – the variety of theories, algorithms, methods and platforms – actually presents a barrier in the path to adoption.

It’s important to note that taking advantage of the benefits of machine learning doesn’t necessarily require a huge investment or major changes to the production floor.

The fact is that in many plants and factories, such as in the chemical processing industry, data is already being captured and stored in a well-structured way. Simply by gaining a better understanding of the type of problems machine learning can solve, manufacturers can begin to explore how their data can drive significant improvements.

 

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.

 

Leverage your data to cut production losses.

Get Demo