Assembling AI transformation in manufacturing

Assembling AI transformation in manufacturing

Benjamin Klöpper, Felix Balhorn

Sep 11, 2024

• Facebook
• Linkedin

Analyzing the limits and possibilities of machine learning in industry

Recent technological advancements in Artificial Intelligence (AI), coupled with user-friendly applications featuring text prompts, such as image-generating AIs and chatbots, have garnered widespread attention in both the media and corporate discussions. While AI in manufacturing has been integrated into manufacturing processes for quite some time, its transformative impact has not yet realized its full potential.

What follows is a guide for plant managers and production engineers on how to strategically harness AI's power. Incorporating the latest technological developments, including foundation models and generative AI, can unlock AI's true potential and provide a competitive advantage.

AI use cases in manufacturing

In most manufacturing processes, an important part of the job of operators and engineers is to monitor the process, analyze problems, find root-causes, identify the right corrective actions, and implement them efficiently. Figure 1 outlines different levels of AI capabilities that support each of the levels in this workflow: Perceive, analyze, prescribe, and closed loop AI. Moreover, it illustrates how AI can revolutionize manufacturing processes to make them more efficient, sustainable, and safe across different industries. Furthermore, it highlights how these AI capabilities can be built upon one another, facilitating a stepwise development and introduction of AI functionalities.

It is important for business leaders to be intimately acquainted with each level of the workflow and their correlating areas of application. Let's take a more detailed look at each level.

The AI ecosystem in manufacturing

The introduction of AI requires more than just expertise in ML and software engineering. Manufacturing is already a complex socio-technical system, consisting of manufacturing and automation technology and many human actors in different roles. Figure 2 illustrates how AI systems operate across the major elements of this system.

AI and the manufacturing workforce

As AI transforms manufacturing operations, the role of the workforce undergoes a parallel transformation. In the ML development phase, engineers play a crucial role in teaching the AI by providing data annotations and assisting ML developers in validating model performance. Explainable AI (XAI) empowers manufacturing experts to assess ML models beyond statistical performance, enabling them to evaluate whether the model has learned meaningful concepts from the data.

During AI deployment and operation, workers and engineers become gatekeepers, deciding whether or not to act on AI predictions and translating AI insights into corrective actions. Fully autonomous systems in manufacturing, where AI directly controls automation systems, remains feasible and economically viable only for a limited set of use cases - it is not expected to become common in the near future. AI models can be significantly enhanced with valuable data by simply recording corrective actions and analyzing the manufacturing system's responses. Here, workers and engineers can provide feedback, offering additional labels for supervised learning or evaluating ML model outputs.

Automation and the IIoT

Crucial data for AI modelling is generated at all levels of automation systems (field, control, supervisory, and planning). Sensors, instruments, analyzers, and quality systems measure physical quantities and monitor manufacturing system states. The Manufacturing Execution System (MES) guides resource allocation in production. Industrial Internet of Things (IIoT) systems and sensors complement data collected by the automation system, enhancing the overall information pool. The key manufacturing shifts and an ageing installed base of automation systems are crying out for green- and brownfield projects. Their introduction would update the manufacturing processing, offering an opportunity to create AI-ready manufacturing systems.

"Companies that use this opportunity to create the foundations for the large-scale application of AI in manufacturing will get ahead of the competition and maintain their advantage by accelerating their AI journey."

Development and operations of AI systems

The data generated by automation and IIoT systems serves as the foundation for AI, encompassing both ML and traditional symbolic AI techniques. Access to real-time manufacturing process data is essential for AI models to perform online tasks, such as pattern recognition, anomaly detection, and predictive capabilities. The development and operation of AI systems in manufacturing are intricately tied to the continuous flow of data from the complex socio-technical ecosystem.

By understanding and leveraging the interactions within this ecosystem, organizations can maximize the benefits of AI in manufacturing, fostering a collaborative environment where human expertise and AI capabilities synergize for optimal results.

AI breakthroughs open new opportunities

The integration of cutting-edge AI technologies into businesses and processes has paved the way for unprecedented possibilities and new opportunities. Let's examine an overview of the pivotal breakthroughs in the world of AI and the emerging opportunities that have the potential to disrupt the manufacturing sector.

Data-centric AI

If sufficient training data is available, deep learning capabilities that extract meaningful features from raw data can completely replace the tedious and error-prone human effort of feature engineering in ML. This capability has also enabled the shift to data-centric AI, which prioritizes the effective use of available data against optimizing model parameters. This approach mitigates typical challenges in AI for manufacturing, such as lack of labels and low data variance. It also supports tedious and expensive data curation activities. Certain methods, such as self-supervised learning or contrastive learning, facilitate the extraction of semantically meaningful features from raw data without the huge labelling effort of early deep learning models. This new paradigm places data center stage and calls for a change in the design and operations of manufacturing systems. Such systems need to be designed and upgraded with the data requirements of modern AI technology in mind. Moreover, it is imperative for the workforce to recognize data recording and documentation as essential components that add value to their work, rather than viewing them as mere inconveniences.

Gen AI for operations

Generative AI (Gen AI) is rapidly becoming an era-defining topic. Recent Capgemini research found that 97% of executives are discussing its applications. Gen AI refers to AI models that are capable of generating text, images, videos, or other types of data, often based on user input. In the manufacturing context, this capability enables many different use cases. One of the most noteworthy benefits is Gen AI's ability to liberate workforces from the completion of tedious manual tasks. For instance, with recorded inputs (photos, audio recordings, or signal data), Gen AI in manufacturing can help operators and maintenance personnel document observations on the shopfloor. This leads to improved data quality for downstream and value-adding AI tasks. Similarly, generative AI in manufacturing can synthesize training data for rare cases (i.e., quality defects or process breakdowns) and thereby improve the performance and robustness of industrial AI systems. Gen AI models can even create working program code to make automation engineers more effective and help to extend the coverage of automation in industrial plants.

Explainable AI helps to overcome black-box syndrome

Deep neural networks and large ensemble models (e.g., Gradient Boosting Trees) have outperformed traditional approaches in various manufacturing applications. However, their opacity has led to a lack of trust. This is largely due to the inherent "black-box" nature of such high-performing models. Explainable AI (XAI) methods shed light on these black boxes, providing insights into how ML models arrive at decisions. Certain techniques, such as counterfactual explanations, help to debug ML models and gain new domain insights. XAI is essential in overcoming trust and acceptance issues faced by AI in manufacturing operations, empowering the workforce to effectively fulfill their roles as teachers and gatekeepers.

Deep reinforcement learning enabled by digital twins

Deep Reinforcement Learning (DRL) takes a unique approach by learning through interactions with the environment, making it possible to derive an optimal control strategy. DRL's deployment in real-life manufacturing settings is facilitated by training within a digital twin-a digital counterpart replicating physical products or processes. Digital twins collect and present data from their physical counterparts, facilitating advanced applications in robotics, scheduling optimization, and process control. AI in manufacturing operations and digital twins share infrastructure and functionality, creating a symbiotic relationship that enhances predictive capabilities.

"To a certain extent, AI in manufacturing and digital twins share similar technology stack and overlap in functionality. But the relationship is symbiotic, rather than competitive. On one hand, AI models trained on actual process data are valuable within the digital twin concept to predict behavior, detect anomalies and simulate what-if scenarios. On the other hand, AI training can benefit from the ability to generate artificial training data with sophisticated digital twins."

AI for manufacturing will span both edge and cloud

Designing an AI environment for manufacturing encounters challenges posed by the data-rich and information-poor nature of the most industrial data. It is not cost effective to send all data, which can often contain little information, to the cloud. Instead, the most economically viable option necessitates filtering data on the edge. Not all machine learning models can run in the cloud. This is because low-latency predictions are vital. Questions revolve around the timing, quality, and sample rate of data required for training or running AI models. Striking a balance between complexity and data efficiency remains a central challenge for organizations in AI implementation. The concept of fog computing, which extends established concepts from cloud computing to devices located on the edge, is an important enabler for AI in manufacturing operations.

Data products as a governance model

Data products (reusable data assets providing trusted data) offer a structured approach to managing the flow of data in AI for manufacturing. Defining freshness and data quality based on consumer requirements is crucial in the designing of data products. Cloud infrastructure is ideal for creating ML models, while edge resources are essential for smartly filtering and aggregating data, providing low-latency runtime for ML models near manufacturing processes.

"Data products are a great way to get answers to the question that data needs to be processed, to structure the flow of data and facilitate reuse across AI development projects."

Democratization of AI with no/low-code platforms and AutoML

While AI tools like ChatGPT and Midjourney have made AI approachable, AI's application in complex manufacturing tasks can still be technically challenging. No/low-code platforms empower domain experts to develop, test, and deploy AI applications efficiently. AutoML takes this further by automating many ML steps, enabling non-ML experts to create models efficiently. By addressing skill shortages in ML and adapting to changing operating conditions, AutoML plays a crucial role in making AI more accessible and adaptable for a wide range of applications.

Mastering AI and ML in manufacturing

To succeed in the AI transformation of manufacturing, three guiding principles are essential:

• Value-centric approach: Prioritize value over technology, avoiding siloed and costly solutions.
• Evolutionary approach: Follow an evolutionary roadmap, gradually increasing AI complexity while delivering value early and allowing a gradual adaptation within organizations.
• User and developer journeys: Align strategies with user and developer experiences, maintaining clarity on value, development efforts, and infrastructure requirements.

Phased approach to AI transformation

The figure below outlines the essential stages for methodically and sustainably integrating AI into manufacturing. Starting from a strong and robust vision for AI in the manufacturing process, the first phase(assess), captures the current state and identifies gaps in the vision. During the second phase(define and design), the future organization and operating models are defined as well as technological blueprints. In the third phase(implement), the organization is ramped-up while refining the operating and governance model. The fourth and final phase combines assessment, definition & design, and implementation. These phases are organized into four workstreams: Transformation, people and AI, data and AI, and automation and IIoT.

Conclusion

Manufacturing companies are yet to fully exploit the potential of AI in manufacturing in reshaping their operational model. AI presents significant opportunities to expand in scale and scope and to accelerate learning. Companies that achieve a robust understanding of AI's technological advancements and leverage them to support their workforce, invest in leadership and organizational training, and adopt a systematic approach to exploring and leveraging AI (supported by a comprehensive AI data strategy) will not only outperform but potentially replace their less AI-integrated counterparts.

Stay informed

Subscribe to get notified about the latest articles and reports from our experts at Capgemini Invent