Click Image to Enlarge

As pointed out in the previous article, current AI solutions are still based on so-called weak AI. Weak AI technologies include those related to machine learning, deep learning, neural networks, speech recognition and machine vision. AI can also be used to enrich the data required by these technologies and to train models.

Generative artificial intelligence can learn

A new and much-hyped technology has emerged: generative artificial intelligence (AI), which can create new content such as text, images, video and audio based on what it learns. Even generative AI is still classed as weak AI. So, we are still a long way from a machine thinking like a human. This does not mean that even today's weak AI solutions cannot significantly improve maintenance efficiency. In most companies, maintenance is still based on time-based preventive maintenance rather than on the actual maintenance needs of the equipment.

Asking AI itself, for example using ChatGPT, how AI can currently be used in maintenance, results in solutions that specifically use weak AI technologies. These can be used to increase the maturity of maintenance and to achieve significant benefits in terms of equipment availability, reliability and lifetime extension. However, building on past solutions and existing experience is an essential part of the development process.

The following describes the most common weak AI-based solution areas that can be used to increase the maturity of maintenance. It should be noted that the terminology is quite diverse and solutions are referred to under different and overlapping names.

Condition Based Maintenance (CBM)

The Industrial Internet of Things (IoT) has made it possible to collect huge amounts of data on machines and analyse their condition using various algorithms. Machine learning algorithms can be used to identify the occurrence of certain types of failures and react promptly. Anomaly detection algorithms can be used to generally detect abnormal operation of equipment and to investigate abnormal operations before a potential failure occurs. In this way, maintenance can be based on the actual condition of the equipment, rather than on a time-based approach, whether or not the equipment needs maintenance. As the importance of data-driven algorithms grows, AI can also be used to train the algorithms themselves.

Predictive Maintenance (PM)

Future equipment failures can be predicted before they occur by using predictive maintenance algorithms. Predictive maintenance not only examines the current condition of the equipment but also its failure history, and can be used to predict when the equipment will next fail. This information is useful not only for failure prevention but also for the optimisation of time-based predictive maintenance programmes. AI also helps in data enrichment, where measurements and the features that can be developed from them are combined with failure history.

Optimisation of planning and scheduling (Optimization)

Maintenance work can be optimally scheduled using various optimisation algorithms, taking into account equipment maintenance schedules, staff availability, and access to spare parts and tools. Similarly, the use of human resources can be optimised based on staff availability and skills. Optimisation algorithms can also be used to optimise the supply chain of spare parts for maintenance and to predict future needs.

Exploiting machine vision (Machine Vision)

Machine vision has typically been used for quality control of production lines, but it can also be used in maintenance for any activity that involves visual inspection. In particular, objects that are difficult to inspect such as structures at height, can be imaged by a drone and machine vision can be used to inspect the images. Interpretation of the images requires the ability to distinguish features and, above all, to detect changes in them.

Natural Language Processing (NLP)

Natural language processing solutions can be used, for example, to review maintenance logs and find relevant information in free text or classify information. This is where generative AI comes in. Numerical interpretation of log texts is also in its infancy.
Data often needs to be refined through data analytics to make it more useful.

Artificial intelligence has brought new opportunities to make the work of maintenance staff more efficient. In particular, generative AI has made it possible to create different types of assistants for maintenance staff. For example, they allow technicians to make queries in natural language, to which the assistant retrieves answers from a defined set of data, such as technical documents, manuals or service manuals. For example, an installer can ask for repair instructions for a specific fault code on a particular piece of equipment.

Click Image to Enlarge

The wizards can also be used to leverage the knowledge of more experienced installers. For example, a technician can show a picture of the device being repaired on a mobile device, and a more experienced technician can add annotations to the picture to guide the technician through the repair.

The challenges of using AI today

The idea of being able to detect equipment failures before the equipment breaks down of course sounds great. Ready-made solutions in this area are available from several suppliers. But why are these solutions not more widely used?

First of all, the area is vast and difficult to define. The terminology is varied and there are many different solutions. Maximising the benefits would require a top-level strategy, raising the threshold for starting a project. More easily implemented piecemeal solutions, on the other hand, will not bring major overall benefits. However, the reliability of point solutions is easier to verify, so large-scale solutions can also be built through their integration.

One reason is certainly the lack of clarity about who is responsible for what in maintenance organisations? While there are clear benefits for maintenance, data collection, storage and analysis is largely an ICT issue. Traditional operational level maintenance (EAM) systems are end-user applications. On the other hand, for advanced AI-based systems, the only interface with maintenance may be the creation of a defect chain and everything else happens elsewhere. In this case, maintenance may see it as an ICT issue, but the ICT does not see it as their own, so collaboration is essential.

Despite the huge amount of data generated by today's devices, collecting and storing it is often perceived as a problem. Similarly, data quality is often perceived as poor. Data often needs to be refined through data analytics to make it more useful. AI works best in small entities, so it is worthwhile refining data and building solutions around clear use cases.

In-depth knowledge of equipment failure mechanisms is needed when designing different types of codes to detect failures. Such knowledge may not exist, and defining failure mechanisms from scratch can be a huge task in itself. Off-the-shelf failure libraries can be a big help here, but application expertise is essential.

How to move forward with AI?

Exploiting existing AI solutions requires, first and foremost, data. So, the strategy for collecting, storing and using the data itself should be thought through. The development of maintenance maturity should also be planned at business level and the objectives should be clear. Appropriate sub-segments and their expertise are key here.

At the same time, it would also make sense to gain practical experience in the use and application areas of AI. To increase understanding, data quality and user experience, limited proof of concept trials for the most critical equipment categories will provide valuable insights for the wider deployment of AI. 

– Utilising previously used solutions and existing experience is an essential part of the development work," says Esko Juuso, author of the article and associate professor.

TEXT: Esko Juuso, Docent, Emeritus University of oulu
PHOTOS: Companies, shutterstock