Key points
Understanding the opportunities and challenges of artificial intelligence paves the way for significant improvements in the generative AI era.
Artificial intelligence is transforming industry with promises of greater efficiency and innovation. With the release of OpenAI’s ChatGPT earlier this year, there has been a huge amount of excitement about AI’s potential use.
Yet, there is a reality behind the enthusiasm. AI brings a mix of opportunities and challenges to industrial operations. Understanding its capabilities and limitations ensures AI adoption is not just driven by trends, but by a realistic assessment of how it will address industry needs.
What Is AI for Industry?
At its core, AI refers to machines or systems that mimic human intelligence to perform tasks. They can improve themselves based on the information they collect. Machine learning, which is a subset of AI, uses algorithms to learn from and make decisions based on data.
There are generally four categories that represent AI for industry, as demonstrated in Figure 1:
- Assisted Intelligence is the most basic form of AI. It helps engineers by making their tasks easier. For instance, assistive intelligence might use operational data to suggest an improvement. However, the final choice is up to the engineer. Assistive intelligence devices cannot make decisions on their own.
- Automation intelligence solutions, on the other hand, are suitable for repetitive tasks that do not require much human interaction. An example is a machine that sorts products on a conveyor belt.
- Augmented Intelligence works with humans to perform tasks better than either could alone. An augmented intelligence solution could analyse sensor-generated data to help engineers make better and quicker decisions.
- Autonomous intelligence is the most advanced form of AI. It can make decisions and act on its own without human help. A self-driving vehicle in a warehouse uses autonomous AI. Autonomous solutions, however, are not suitable for managing manufacturing processes.
The Rise of Generative AI and LLMs
Large Language Models, such as Open AI’s ChatGPT, have recently taken AI into the mainstream. Also referred to as generative AI, these systems understand both natural and computer languages. They are designed to generate contextually relevant content.
Research companies, such as Gartner, have followed the evolution and expectations of AI solutions. According to the 2023 Gartner Hype Cycle, Generative AI is at the peak of its inflated expectations.
Forrester also said in July 2023 that 67% of businesses said they are finding ways to incorporate generative AI into their overall strategy. Furthermore, Forrester noted that technical experimentation—such as the recent introduction of customisable ChatGPT models—would further advance AI as companies move into 2024.
Many process manufacturing companies are still experimenting with generative AI. Forrester said these companies should remain cautious as they head into next year. One reason is that workflows require complex interactions between machines and operators.
Testing generative AI in these environments is too risky. Another concern is what are called generative AI “hallucinations.” These are mistakes inserted into the generated information that could lead to damaged equipment, serious injury, or even death.
Still, engineers already can use generative AI to optimise operations. For example, generative AI could be used as a coding copilot for engineers. It could also store use cases, prescriptions for solving process anomalies, and even offer suggestions for the best course of action to take based on its training.
Using Retrieval Augmented Generation (RAG) to create chatbots that tap into these databases, and with API calls, it’s possible to access the Home Page of advanced industrial analytics software and literally ask it, “What did I miss since my shift last night?” The RAG enabled GPT then would generate a detailed summary of any events it had been monitoring overnight.
Approaching the Challenges of AI
However, setting up the models probably won’t happen overnight. Moving in small steps and promoting transparency could go a long way in a successful AI program. As with any technological advancement, many of the challenges preventing success can be overcome with change management solutions.
These include organisational buy-in, support from users, a strong communications strategy, and staying power backed by foreseen value and clear benefits.
In an AI project, perhaps the most critical challenge is the availability of quality data. Insights suggest that 60-80% of data engineers and data scientists spend their time connecting data sources, cleaning, or otherwise preparing data for AI projects. When data integrity or accessibility is an issue, the AI initiative has little chance of success.
In the case of manufacturing, many years of time-series data are available for gaining valuable insights. The idea is to repeat patterns of good performance, such as ideal operational zones and golden batches. The use of advanced industrial analytics software to prepare operational data makes it accessible and ready for AI projects.
Preparing for the Data-Driven Journey
When a factory is built, it is delivered as an Automated Factory. Here, sensors are placed throughout the plant and a control room can monitor operations (in Figure 2, this is identified as the Controlled Factory). But automation is just the starting point.
Users of advanced industrial analytics software may have used spreadsheets for years and could be resistant to change. Thus, an organisation must cross The Mindset Gap before it becomes data driven.
At the Data Driven Factory level, the focus shifts to gaining deeper insights with the availability of new solutions. For more on process events, sensor-generated data needs to be considered within its operational context.
This means ensuring data is made available to everyone who needs it. Achieving this democratisation of data yields a Connected Factory, but getting there requires crossing the OT/IT Gap.
Traditionally, contextual data was kept in separate systems that are managed by separate departments. Examples include maintenance logs, shift reports, laboratory information systems, and even weather updates.
If a maintenance log shows that a system is taken offline, its sensors are also not generating data. Operational experts can use this information to filter those periods from the analysis.
With the rise of data lakes and cloud-based solutions for storing operational data, the availability of data across sites and to remote workers has improved. To be a fully Augmented Factory, however, companies must bridge The Organisational Gap.
This aligns people, processes, and technology from local operational teams with those at central data teams. Collaboration is essential to achieve higher efficiency in solving complex use cases. In these cases, engineers can work with data scientists on a machine learning exercise.
The Augmented Factory begins to see the culmination of its efforts. It uses augmented analytics and machine learning to not just understand but also predict and optimise processes in real-time. Beyond the horizon is the ultimate smart factory, but it requires harmony between AI solutions and its users.
Bridging the AI/HI Gap means that AI-driven analytics and ML techniques are accessible and actionable by users and decision makers. This requires solutions that can translate complex data into insights. Engineers then use their expertise as the final step in the process.
On the Path to AI Use Cases
Anomaly detection finds unexpected events in operational data. These events could be from a performance issue or from faulty data, such as a sensor malfunctioning. Anomaly detection models learn periods of normal behavior to be able to detect when things go wrong.
There are many ways to detect anomalies. For example, an operational expert can perform a value-based search of time-series data. Outliers from this search are shown in a way that is easy to understand.
Creating a golden fingerprint by analysing contextual data also determines what qualifies as an anomaly in those situations.
For advanced cases, a self-organising map (SOM) applies machine learning techniques to develop models that detect global and local outliers in a multivariate setting.
These ML models are like those that are being employed by the specialty chemical company Clariant. Clariant had planned for augmented factories from the moment it started its digitalisation journey.
The company determined that it needed to use more robust solutions to find the root cause of process anomalies, but that it also should apply its team’s machine learning capabilities to its processes for advanced monitoring.
Clariant’s corporate data science teams generate purpose-built machine learning based models that are made available to the organisation. These aid in building soft sensors, anomaly detection scoring, and predictive maintenance alerts.
They allow Clariant to improve its operational efficiency, reduce downtime, and make more informed decisions based on operational data enhanced with these ML capabilities.
The improvements led to a 10% batch time reduction, which allowed the company to run an extra batch every day. At the same time, energy consumption dropped 9% because of the operational improvements.
The Road Ahead
Artificial intelligence is gaining traction and technology is showing improvement every day. But AI still takes a lot of effort for little return in value.
By tapping into the data that’s already available, engineers can find immediate insights on how to optimise operational performance. At the same time, they are laying the foundations for the future of industrial AI.
