From machine learning to generative AI - innovation for maintenance

Artificial intelligence or artificial intelligences?

Artificial intelligence in a nutshell

Artificial Intelligence (AI) is the field of computer science that focuses on creating systems capable of performing tasks that require or mimics human intelligence. These tasks include learning, reasoning, problem-solving, understanding language, perception and even creativity. AI encompasses a wide range of techniques, spanning from rule-based systems like heuristics and meta-heuristics to data-based systems like machine learning and deep learning driven by neural networks.

Today, machine Machine Learning is the most common technique to build artificial intelligence. Machine learning operates by analyzing large sets of data, identifying patterns, and using this information to make decisions or predictions. Over the decades, early AI systems were largely deterministic, meaning they followed pre-defined rules to make decisions. For instance, a rule-based AI system might automatically shut down a machine if the temperature rises above a certain threshold.

However, over the years, AI has evolved into more adaptive and sophisticated systems, where machines not only execute tasks but can also learn and improve over time. The learning process is what distinguishes modern AI from its earlier forms, making it more versatile, adaptable and intelligent. But don’t be deceived, even though it may appear somewhat magical, the process actually requires a substantial number of probabilistic and statistical calculations.

Is generative AI just another AI model?

L’IA générative (AIGen) refers to a new form of AI models that are capable of generating content, such as text, images, or even music, autonomously based on learned patterns from vast datasets. These generative models differ from traditional AI in that they do not simply follow rules or classify data, but create new information by drawing on their understanding of existing data. This makes them more flexible and creative in their outputs.

According to Google CEO Sundar Pichai, a generative model aims to convert any input into an output. There is a wide range of generative models, each based on different technologies. For example, MidJourney can produce highly realistic images from a text description. Google Lens uses computer vision to identify objects or plants from a photo. Or DeepL, which uses certain models to translate or summarize texts. The potential of these emerging computing technologies goes beyond the simple analysis of structured data in databases; they enable us to process any type of information.

From Generative AI to LLM: when man dialogues with machine

The most widely recognized category of generative models is Large Language Models (LLMs). LLMs are a subset of generative AI (GenAI) that specifically target the transformation of natural language into natural language. This capability allows for the development of systems that can perform tasks such as translating text, engaging in conversations with humans, writing code, and more. These models, exemplified by GPT-4 or Mistral Large, undergo training on extensive text datasets and are engineered to comprehend, produce, and manipulate human language with exceptional sophistication. Leveraging neural networks architectures like transformers, LLMs are capable of managing long-range dependencies within text and sustaining context across prolonged dialogues or passages.

The key capabilities of LLMs include:

• Text Generation: Creating human-like responses to prompts, writing essays, or drafting emails.
• Contextual Understanding: LLMs can understand nuanced requests, take context into account, and provide answers or suggestions tailored to specific situations.
• Knowledge Integration: They can synthesize knowledge from diverse sources, making them useful in technical fields like maintenance, where understanding complex documentation or historical data is crucial.
• Natural Language Interaction: LLMs can engage in dynamic conversations with users, adapting their responses based on previous inputs and offering explanations or clarifications when needed.

In contrast to earlier machine learning models, which required structured input (like sensor data), LLMs excel in unstructured environments. They can analyze both structured and unstructured data (such as maintenance reports, logs, and technical manuals) and integrate these diverse sources into a single output that is actionable and contextually aware.

From Machine Learning to generative AI: The Evolution Driven by Demand

The AI technologies applied to maintenance have evolved dramatically. Initially based on traditional machine learning (ML) models, they now also incorporate generative AI (GenAI) models. This evolution was not just a leap in technological advancement but a response to the growing demand for more intelligent, adaptive solutions that could handle increasing data complexity and user needs.

Machine learning: a first step towards predictive maintenance

In the early stages, supervised ML was used for predictive maintenance—training models with labeled datasets to forecast equipment failures based on historical data. For example, predictive maintenance in manufacturing utilized ML algorithms to anticipate the failure of machinery, reducing downtime and improving operational efficiency. These early models excelled in identifying known patterns, but their limitations became evident as systems grew more complex and data sources more diverse. They struggled to adapt to changing environments and required significant manual intervention to manage data labeling and model retraining.

Unsupervised and semi-supervised learning techniques followed, enabling systems to extract information from unlabeled data and learn autonomously. These models have extended predictive capabilities, detecting anomalies or groups without predefined labels. However, even as these models have advanced, the challenge of making sense of unstructured data – such as technical documents, logs or real-time operational reports – has remained an obstacle.

Intelligent agents: the solution to the diversity of our data

The breakthrough came with the rise of GenAI and LLMs, with which many more intelligent systems can be implemented. These systems can not only understand and generate human language, but also engage in more sophisticated reasoning about structured and unstructured data. With LLMs, products could go beyond pattern recognition to interpret, adapt and generate knowledge autonomously, enabling companies to learn from complex and sparse information.

This shift from ML to GenAI has enabled more interactive intelligent agents, capable of dynamically responding to natural language queries, generating explanations and offering actionable recommendations, fundamentally transforming maintenance workflows.

This transition from ML to generative AI reflects industries’ growing demand for AI solutions that don’t just predict problems, but adapt and reason autonomously. As data becomes more varied and systems more complex, the need for intelligent assistants capable of handling uncertainty and context-specific queries has never been greater, driving forward innovation in AI for maintenance.

Our vision: from functionalities to assistants

For many years, AI in industrial maintenance was designed to perform specific tasks. These tasks were focused on automation – scheduling maintenance, predicting failures or optimizing resource allocation. These features, though useful, were isolated in their functionality and often required human operators to interpret and integrate outputs manually.

However, with the rise of IAGen and LLM, the vision has evolved. We are moving from creating isolated, task-specific functions to designing full-fledged AI assistants that can engage with humans in real-time, understand context, and provide holistic, dynamic support. These assistants are no longer just tools; they become intelligent partners capable of both technical understanding and adaptive behavior.

Portrait of tomorrow’s intelligent assistant

CARL Berger-Levrault is currently working on a number of use cases to enable its customers to take advantage of the wide range of possibilities offered by artificial intelligence.

For us, the intelligent assistants of the future will make it possible to:

• Understanding context: Instead of merely processing raw data, AI assistants will grasp the wider context of an industrial environment – understanding operational schedules, prioritizing critical tasks and even aligning with corporate goals.
• Predict and propose: AI assistants will not only predict when maintenance is needed, but also suggest the most effective strategies for repairs, taking into account real-time data such as machine workload, available parts and technician schedules.
• Explain and justify: These systems will provide users with explanations for their recommendations, thus reducing the “black box” nature of AI. This transparency will reinforce trust and make AI a more reliable partner, recognized for its quality.
• Adapt and learn: Leveraging LLMs, AI assistants will be able to learn from user interactions, improving over time as they gather more information about the environments in which they operate. This will enable them to anticipate needs before they arise, streamlining operations and reducing downtime and maintenance costs (parts changes are optimized).

Find out more about IA CARL

AI in the service of maintenance, when technology serves efficiency

Let’s imagine a production facility where traditional preventive maintenance is used. A machine learning model monitors sensor data (temperature, vibration, pressure) and predicts a potential machine failure, sending an alert to the maintenance team. In the old paradigm, the technician would then analyze the alert, collect additional data, consult maintenance records and manually determine the next steps.

Let’s now consider an AI assistant powered by LLMs and Gen AIs in the same installation. Upon detecting a possible failure, the AI assistant takes the following steps:

• Contextualized analysis: It not only reports the problem, but draws additional information from historical maintenance logs, real-time sensor data and even operational programs to provide a complete picture of the situation.
• Dynamic recommendations: The wizard suggests several solutions, ranging from immediate repair to adjusting machine parameters to prolong operation until a more favorable repair period.
• Natural language interaction: a technician can ask the assistant for clarification, for example, “Why is this breakdown happening now?” or “What happens if we delay repairs for 24 hours?” The AI responds in a conversational tone, explaining the underlying technical reasons and potential risks.
• Proactive adjustments: AI automatically checks spare parts stocks, technician availability and production schedules to recommend the optimum repair time, balancing machine performance and production needs. It can even suggest reorganizing the tasks of other technicians to accommodate urgent repairs.
• Collaboration: Throughout the process, the AI assistant learns from the technician’s decisions and contributions, enabling it to improve its future recommendations and adapt to changes in the operating environment.

In this future state, AI assistants not only anticipate problems, but actively collaborate with human teams, offering insights, context and recommendations in real time. The result is a more efficient, less disruptive maintenance process that helps minimize downtime, improve resource allocation and optimize long-term operational performance.

CARL Berger-Levrault has always been committed to innovation, as demonstrated by these solutions based on various artificial intelligence models. Integrated into industrial maintenance processes, they promote digital transformation and pave the way for the era of maintenance 5.0.

CARL’s AI solutions are an excellent example of how AI, when applied effectively, transforms maintenance processes from reactive to proactive, while making operational management more efficient and intelligent.

Discover also