Paul: Welcome! In this edition of The AI Interview, we are pleased to welcome Gerhard Kreß and Dr. Danny Krautz. We will look at applications, barriers, and the practical dos and don'ts of implementing AI in the industrial sector. We will also look at analytics applications, examine the difficulties in manufacturing environments, and discuss the necessity for effective analytics applications. To kick things off, Danny, what are analytics applications in the industrial sector that excite you the most?

Danny: Thank you, Paul. What excites me the most is the opportunity to enhance productivity and improve products using data and analytics. At CeramTec, where we manufacture complex products, the potential to translate expert knowledge into even more valuable outcomes is truly exciting. We can unlock a lot of potential within the company through data and analytics.

Paul: That’s an insightful perspective, Danny. Now, Gerhard, can you share the most exciting AI application your team at Siemens has developed?

Gerhard: One of the most exciting applications we worked on was in the rail industry. We tackled the challenge of predicting point machine failures, a leading cause of rail delays. Instead of using sensors, we developed an algorithm that relied solely on existing data from interlockings. This innovation eliminated the need for additional sensors, offering a more cost-effective solution. What's remarkable is that our algorithm achieved predictions as accurate as, or even better than, sensor-based methods. We also developed a pre-trained algorithm that self-adapts to the different point machines, streamlining deployment. This project was particularly cool because it addressed a real-world problem at scale, eliminated the need for sensors, and used a unique data-driven approach.

Paul: The complexity and timeline of developing such a precise algorithm must have been quite intense, right Gerhard?

Gerhard: Definitely. Developing a highly accurate prediction model in an industrial setting presents various challenges. The accuracy needed was in the upper 98-99% range, which demanded a deep understanding of the data and the mechanical aspects of point machines. Point machines fail infrequently, making historical data limited. Moreover, extracting data from safety-critical systems like interlockings required meticulous handling. The project took around two to three man years to achieve sufficiently high accuracy, including data engineering, model development, and integration. It involved engineers with expertise in mechanics, electrical systems, and data analysis. Getting contracts and buy-in from customers like SBB or Deutsche Bahn added another layer of complexity. In the end, the results were impressive, but the journey was arduous.

Paul: Thank you for sharing those insights, Gerhard. Now, let's shift our focus to the barriers faced when implementing AI in industrial settings. If you had to identify the top three barriers, what would they be?

Gerhard: Sure, first, acquiring sufficient and high-quality data can be a struggle due to diverse, often older equipment. Balancing scarce and abundant data sources is crucial to avoid creating separate models for each generation of equipment. Second, these systems are designed to often last for over a decade, resulting in limited historical failure data. This necessitates precise predictions to avoid excessive false alarms. Third, gaining customer trust and convincing them to adapt their processes is a significant hurdle. Deploying AI in industrial settings requires rigorous accuracy and long-term commitment.

Paul: Those are indeed significant challenges, Gerhard. Now, Danny, based on your experience at CeramTec, what success factors do you consider crucial for AI projects in industrial domains?

Danny: First, it's essential to have a strong top management commitment from the beginning. AI projects in the industrial sector are long-term endeavors that require strategic support. Second, transparency and realistic expectations are vital. Explaining that initial accuracy might be low and that AI will complement existing processes, not replace them, helps set the right expectations. Data engineering is also fundamental and often underestimated, so educating stakeholders is essential. Lastly, focusing on solving real problems with a clear understanding of the impact and involving the right people is critical.

Paul: Thank you, Danny. From identifying problems to selecting and validating AI use cases at CeramTec, how do you navigate through this process?

Danny: At CeramTec, having strong board commitment is a starting point. AI projects require a strategic commitment. To identify use cases, we begin by thinking from a customer perspective and considering holistic problem understanding. We engage stakeholders from different business units to ensure a comprehensive view. Using design sprints and collaborative problem iteration, we prioritize ideas and ensure everyone understands the potential impact. We aim to be transparent about our approach and set realistic expectations. In smaller companies like CeramTec, resources are limited, so it's crucial to select use cases wisely and focus on creating tangible, short-term impact while keeping long-term goals in mind.

Paul: Thank you, Danny. We've been talking about AI in the industrial context, and I'd like to start by discussing the importance of having a clear process and responsibilities in place before implementing AI solutions. Danny, would you like to share your thoughts on this?

Danny: I think it's crucial to emphasize that AI is a tool, not the ultimate solution. Before diving into AI or data-driven approaches, it's essential to establish a clear process and assign responsibilities. Often, the problem isn't the lack of AI or data; it's the absence of a well-defined process or understanding of who should do what. Start with clarity in your operations, and then identify where AI can support these processes, not the other way around.

Gerhard: I completely agree with Danny on this point. AI should improve existing processes and decisions, not complicate them unnecessarily. It's about making sure you use AI where it's really needed, not just adopting technology for technology adoption reasons.

Paul: That's a great point. It's essential to avoid using technology just because it's available and instead focus on its value within existing processes. Another crucial aspect you touched upon is the importance of involving experienced engineers. How do you see the role of experienced engineers in AI projects?

Danny: From my experience, engineers with years of hands-on experience are incredibly valuable assets. They might not speak the same language as data scientists, but they possess deep domain knowledge. It's often more efficient to have engineers identify situations that need attention rather than relying solely on automated predictions. We've had success with this approach, where engineers provide insights and validate the relevance of alerts, leading to a high hit rate.

Gerhard: I couldn't agree more. AI isn't always about full automation; sometimes, it's about guiding engineers' attention to the right areas. Experienced engineers can distinguish between mundane data and critical insights, ensuring their valuable time is well-spent.

Paul: That's an excellent perspective. Now, let's move on to the concept of AI as a decision-support tool rather than an automation tool, as you both mentioned. Gerhard, could you elaborate on this idea?

Gerhard: Often, there's a misconception that AI should fully automate decisions. However, in many industrial contexts, AI works best as a decision-support tool. It helps in requesting the right resources, providing insights, and guiding human decision-making. It doesn't replace human expertise but augments it.

Danny: I agree. AI should assist in making informed decisions by presenting relevant data and insights, especially in complex industrial settings. It's about enhancing human capabilities, not replacing them entirely.

Paul: Let's move on to the use cases. Gerhard, could you share an example of an "evergreen" use case that never leads to successful results?

Gerhard: In the industrial sector, predictive maintenance is often hailed as a game-changer, but it can be challenging to implement successfully. Many proof-of-concepts (POCs) in predictive maintenance struggle due to high false positive rates. It's a case where the promise of AI doesn't always align with reality. I've rarely seen the cost savings that some reports claim. There are more valuable use cases out there.

Paul: Before we finish, Danny, could you shed some light on how organizations should assemble project teams for AI initiatives, taking into account both internal and external resources?

Danny: When it comes to AI project teams, it's essential to strike a balance between internal and external resources. You should internally own the problem, understand the process, and manage the strategy. Data scientists and data engineers can be brought in externally, especially when you need specific technical expertise. Small, dedicated teams with diverse skills work best, typically consisting of process experts, data engineers, data scientists, and possibly UX specialists. Collaboration and communication within the team are key to success.

Paul: Thank you both for sharing your valuable insights today. It was a pleasure to have you with us. Goodbye.