Many technology buzzwords currently are being discussed in our industry: IoT, IIoT, Industry 4.0, edge devices, artificial intelligence and machine learning. The current wave of technology solutions transforms manufacturing from Industry 3.0, or mainly paper-based processes and isolated digital systems, to Industry 4.0, or fully connected, predictive and adaptive plants. I have discussed some of the general technology trends in previous articles in Process Heating.
This article focuses on machine learning (ML) and integrating it into the broader technology landscape in process heating applications. Before I begin, however, it is helpful to define machine learning. Wikipedia describes machine learning as “the study of computer algorithms that can improve automatically through experience and by the use of data.” Thus, machine learning is a subset of artificial intelligence.
Current Industry 4.0 system architectures consist of control devices connected to an edge device to leverage modern protocols to communicate with a cloud-based platform. Streams of data flow to a cloud solution. Powerful computer systems crunch the data and spit out observations and insights into a user-friendly interface.
To minimize the risk of network infiltration by cybercriminals, most companies demand that the connection back to the control device is allowed only via the internal network. For the same reason, companies are hesitant to provide both read and write access to the equipment controller from the cloud solution. Some modern protocols have moved to a publish/subscribe communication pattern to aid network security. (One example is MQTT, which Wikipedia notes is a publish/subscribe network communication protocol.) I discussed the importance of cybersecurity in smart temperature controls in Process Heating in 2018.
A common question is why — with the associated transmission, computing and storage costs, along with cellular links and those associated charges — we need to move data.
In the consumer marketplace, typical artificial intelligence (AI) systems with big data are developed with large amounts of data. They require giant storage sets and massive computing power. The AI system utilizes those resources to identify patterns within the data to provide valuable connections. The e-commerce and internet giants leverage these insights to better understand customer preferences, provide more tailored advertising and recommendations and, ultimately, gain more sales. Larger datasets equal better financial results. Companies with millions of customers and tons of data seem to win the race.
By contrast, manufacturing industries — including process heating — do not always have large datasets available. The expansive teams in the consumer world and allowed extended periods of research are not necessary for many of the small projects and lighter datasets found in manufacturing.
The AI model/code and neural network appear to be a solved problem. The focus is now on the quality of data needed to provide valuable results. Machine learning proofs-of-concept can go from a 12-to-18-month timescale to a few months with a small team, provided that the team is focused on a specific problem and so the scale of investment and return is lower than in the consumer space. American computer scientist Andrew Ng describes this as a “data-centric” perspective of artificial intelligence.
Practical Experiment with Small Datasets and AI/ML
To test out this data-centric theory and deepen my knowledge on machine learning — to see whether this was a candidate for process heating applications — I built a simple proof-of-concept using modern AI/ML prototyping tools. I had the constraints of out-of-office time and needed to generate my data. (There was limited data publicly available.) I researched the technology space and decided to use the Edge Impulse development platform for embedded machine learning and an associated course.
The results shown below are from a standard smartphone using its three-axis accelerometer to generate vibration data. Any device/sensor arrangement can produce data. However, the quality of the data is essential.
(Click on the image to enlarge.)
On a practical level, I used a home appliance as the test machine. Three categories of data were collected (figure 1) to simulate normal flow (green), impeded flow (orange) and no flow (blue). The data capture rate was 100 ms with two minutes of samples per category. The minimum suggested quantity of samples was captured with standard data filters, and I chose the suggested neural network model and its default settings.
A useful output from the machine learning platform toolset includes a clear visualization of the data. (The 3D chart can be rotated and viewed from different angles.) Three distinct regions are easily identifiable although some mixed data is present.
The model achieved an initial accuracy of 90.4 percent based on the validation set. A helpful feature of the embedded machine learning platform is its ability to display a confusion matrix to help users better understand the model’s performance across categories.
To provide a deeper analysis of the dataset, I added an anomaly detection model. This addition effectively bounds the existing data, and anything outside the working envelope is an anomaly. The bounding dimension is shown in the anomaly explorer image (figure 2).
(Click on the image to enlarge.)
To test the model accuracy in a live situation, the application was ported to a smartphone for direct test and verification.
Starting with the home appliance in the rest state, it accurately labeled the no-flow state. Switching the machine on immediately changed the status classification to normal flow. The impeded flow scenario did indicate some misclassification with the labeling, but most results were classified as impeded flow. The model also accurately classified an anomaly simulation by correctly labeling a significant deviation from no flow/normal/impeded flow.
From this simple experiment, the AI model/code and neural network appear to be pretty robust (using software tools like the Edge Impulse platform) and enable quick prototyping and testing of ideas. Due to the lightweight code, the application does not need to reside in the cloud in the live environment. The cloud compute power is used to train the model, but you have options where that model is deployed. (For more information on machine learning at the edge, please visit tinyml.org.)
The development focus can now move from heavy research on model types to investigating the applications worth analyzing with machine learning technology and generating quality datasets.
Experts at McKinsey, a global management consulting firm, suggest that there will be a significant gap in economic gains among those companies absorbing artificial intelligence within the next few years (front-runners) compared to the followers (those that will incorporate AI by 2030) and the laggards (nonadopters).
In addition, Capgemini Research Institute cites three prominent use cases that stand out to kickstart a journey in artificial intelligence or machine learning:
- Intelligent maintenance.
- Product quality control.
- Demand planning.
Of course, for readers of Process Heating, the primary use case would be intelligent maintenance in process heating.
(Click on the image to enlarge.)
Intelligent Maintenance in Process Heating Applications
Boilers, dryers and process ovens are essential to manufacturing productivity.
Fuel-fired equipment types require a valve safety train to manage the correct flow and pressure to the combustion chamber and provide safety protection. In addition, regular periodic checks should be in place to ensure employee safety, equipment longevity, fuel efficiency and process control.
Four key operational issues should be considered:
- Fuel flow can displace piping scale (or other foreign materials), causing a blockage over time.
- Diaphragm devices can be vulnerable to aging, embrittlement and rupture.
- Control valves and linkages can wear and loosen with continuous cycling.
- Debris buildup on fans, burners and air piping can affect air-to-fuel ratios and reduce energy efficiency.
Most of the above issues develop over time, and the trends may not be evident with just periodic inspection checks. Adding new sensors and machine learning intelligence, however, can detect abnormal running patterns. This awareness can assist with regularly planned maintenance activities and reduce unplanned downtime.
(Click on the image to enlarge.)
An emerging trend is to leverage the complete equipment support network (OEM, third-party service provider and end user) to maintain vigilance on all aspects of the equipment performance. As a result, newer IIoT collaboration platforms have been developed to enable these connections.
In some applications, for instance, digital tracking tape offers a low risk way to start the digitization journey by enabling transparency of equipment issues and checks across the organization as well as service providers.
In conclusion, artificial intelligence and machine-learning models can improve existing periodic maintenance plans and achieve operational savings. There will be continued development of new tools and technologies, and several proofs-of-concept studies to validate potential savings before volume rollout. If you are a leader in your sector (OEM, third-party service provider or end user), you likely have already started your journey.
1Sherwin, P. How Digital Trends and Extended Reality Will Impact Process Heating. Process Heating. (2020, October 27). https://www.process-heating.com/articles/93536-how-digital-trends-and-extended-reality-will-impact-process-heating
2Wikipedia. Machine Learning. https://en.wikipedia.org/wiki/Machine_learning. (Retrieved 2021, August 19).
3Wikipedia. MQQT. https://en.wikipedia.org/wiki/MQTT. (Retrieved 2021, October 1).
4Sherwin, P. The Importance of Cybersecurity in Smart Temperature Controls. Process Heating. (2018, August 7). https://www.process-heating.com/articles/92744-the-importance-of-cybersecurity-in-smart-temperature-controls
5Gordon, N. Don’t buy the ‘big data’ hype, says cofounder of Google Brain. Fortune. (2021, July 29). https://fortune.com/2021/07/30/ai-adoption-big-data-andrew-ng-consumer-internet/
6Machine learning development platform, edgeimpulse.com
7Shelby, Z. Announcing Intro to Embedded Machine Learning on Coursera. Edge Impulse. (2021, February 8).
8Bughin, J., et al. Notes from the A.I. frontier: Modeling the impact of A.I. on the world economy. (2018, September 4). https://www.mckinsey.com/featured-insights/artificial-intelligence/notes-from-the-ai-frontier-modeling-the-impact-of-ai-on-the-world-economy
9Brosset, P., et al. Scaling A.I. in Manufacturing Operations: A Practitioners’ Perspective. Capgemini Research Institute. (2019, July 18). https://www.capgemini.com/wp-content/uploads/2019/12/AI-in-manufacturing-operations.pdf
10ISHN. If left unmaintained, combustion systems can be catastrophic. Industrial Safety & Hygiene News. (2019, October 1). https://www.ishn.com/articles/111602-if-left-unmaintained-combustion-systems-can-be-catastrophic
Author’s Note: Opinions expressed in this article are solely the author’s and do not necessarily represent the views or opinions of Eurotherm by Schneider Electric.