Breaking News

Exploring the benefits of AI and machine learning

Many algorithms match machine learning and artificial intelligence goals, but these can help a user begin. Courtesy: Inductive Automation

Learning Objectives

  • Artificial intelligence (AI) and machine learning (ML) are being used more by manufacturers as they realize their benefits.
  • AI/ML software can help spot inefficiencies and improve manufacturing operations.
  • There are many types of AI/ML software, and not all require a data scientist, though a strong understanding of the data is required by the person using it.

Just a short time ago, machine learning (ML) and artificial intelligence (AI) were fairly new concepts for a lot of people in the manufacturing industry. Many companies are getting up to speed and adopting these technologies slowly and methodically. Fortunately, offerings in this space have drastically improved in recent years. The basic concepts, however, are the same.

Artificial intelligence (AI) is a catch-all term that covers any technologies where a computer or system seems intelligent. This can be anything from image recognition to airplane autopilot systems, which started appearing way back in 1914. Machine learning (ML) is a subset of AI and is focused on a machine’s ability to extract data insights. The study of machine learning is often about common ML algorithms, which are used to develop insights around data.

Four machine learning outcomes, benefits for manufacturers

In recent years, there’s been a change in focus from ML algorithms themselves to ML result categories. Here are the four main types of ML outcomes a lot of folks are seeking today:

1. Predictive failure and alarming can be one of the most significant areas of cost savings for a company. If a manufacturing line is about to go down, it can be much more cost-effective to shut it down early and do maintenance than for it to go down midstream. Predicting failures and generating alarms based on algorithms can save a lot of money.

2. Process optimization is another popular area. This normally takes the form of letting a system provide recommendations for tuning setpoints and variables for systems. There are two main types of process optimization: Open-loop and closed-loop. Open-loop optimization involves user interaction, where the system may recommend changes to optimize a process, and an engineer or other expert reviews the recommendations and chooses whether to apply them. Closed-loop optimization takes any human intervention out of the process entirely and tuning recommendations are automatically applied.

Often, a company will run an optimization algorithm in an open-loop fashion, and after becoming comfortable with the results, will then switch to closed-loop. Although closed-loop optimization is normally only done once every few minutes, hours, or days, there are a few cases where running faster may be useful. This would be to provide constant tuning of a process where tuning with traditional means, like a proportional-integral-derivative (PID) loop, may be very difficult or impossible. Most people who do high-speed ML process tuning end up running the ML algorithms for the tuning on an embedded PC or similar hardware right next to the PLC.

3. Anomaly detection finds deviations from normal operating conditions. This can provide insight into when a process is running sub-optimally, or sometimes this can predict bad production runs or mistakes by equipment. These systems generally provide a number indicating how close to normal operating conditions a process is. Results can be considered as an additional “sensor,” and the results of the anomaly detection algorithm can be used for anything from alarming on an abnormal condition to shutting down a process. Although similar to predictive failure and alarming, anomaly detection provides information about how a process is doing right now, rather than predicting how it may be doing in the future.

4. Defect analysis is normally done through image-recognition algorithms and can be very useful for classifying parts and detecting abnormalities.

Four initial steps to every machine-learning project

While all this all sounds great in words, how can someone begin? Here are some initial steps that apply to every ML project:

  1. Identify a system that could benefit from one of the outcomes mentioned above.
  2. Define what you want to analyze with that system and what results you’re looking to achieve.
  3. Verify you have plenty of historical data collected for that system. Most ML algorithms take mountains of data to be effective (to “learn” from). If you don’t have that available on the system, add historical logging and re-visit in a few weeks or months.

Three steps to make a machine-learning model

The next step is generating a machine-learning model. Models are the programs that are executed to get needed results. To generate a model:

  1. Prepare the data. All ML models need data to be generated. Most data exported from a historian or database isn’t perfect and needs to be cleaned. This normally means tossing bad rows, identifying data from bad sensors and excluding it and making sure the data looks reasonable. A process engineer familiar with the system can look at the data and help determine if it’s a clean data set.
  2. The model needs to be trained. This is done by feeding it the cleaned data and choosing some training options.
  3. The model is then scored to see how well it works. A model will always be generated if you go through training, but the model may be very bad at predicting things. Look at the score to see how well a model does and make sure the model scores well.

Three types of AI/ML software

While users might know how to prepare the data, they might not have any idea how to train a model or score it. If that’s the case, it’s time to choose AI/ML software. There’s a wide range of options.

1. AI software vendors 

A lot of companies have popped up here, with many who will build models, or offer pre-built models for certain types of equipment.

2. Low-code or no-code ML platforms 

A growing number of platforms help people begin creating their own models. Google Cloud AutoML, AWS ML, and Microsoft Azure Machine Learning Studio are examples. A basic knowledge of machine learning algorithms is normally suggested, but getting started isn’t too tricky. A little reading on clustering, regression and classification algorithms is a good starting point for beginners.

3. Coding ML platforms 

These are the most common. If you have a data scientist or an information technology (IT) department already doing ML, you may just lean on them to point you in the right direction. Google, AWS, and Azure all have offerings. Additionally, many free options, like TensorFlow and Scikit-learn, can be run locally.

There’s nothing wrong with starting with a low-code ML platform and moving to a full coding platform later. The coding platforms are more complex to use and are often used by data scientists, but the added flexibility often leads to higher scoring models. The better the model, the better the results, and the more likely stated objectives can be achieved.

After initial exploration, many companies hire a data scientist to help with ML efforts, but it isn’t required. These technologies are available and accessible to anyone who wants to try this new technology and pursue the promise it brings.

There’s no question AI and ML will continue to be an important part of the controls and automation landscape in the future. Companies looking for an advantage in operating costs or efficiencies may find a little research looking at possible options to be well worth their time.

Kevin McClusky is co-director of sales engineering at Inductive Automation. Edited by Chris Vavra, associate editor, Control Engineering, CFE Media and Technology, [email protected].


Keywords: artificial intelligence, machine learning

Artificial intelligence (AI) and machine learning (ML) are being used more by manufacturers as they realize their benefits.

AI/ML software can help spot inefficiencies and improve manufacturing operations.

There are many types of AI/ML software, and not all require a data scientist, though a strong understanding of the data is required by the person using it.


Learn more about AI and machine learning at


What applications on your plant floor would benefit from AI/ML software?