Using techniques such as machine learning and data analysis, predictive maintenance can detect early warning signs of failure or degradation in equipment, enabling maintenance teams to intervene proactively to avoid unplanned downtime and associated costs. However, two major difficulties arise: firstly, the construction of predictive models for predictive maintenance, and secondly, data confidentiality, given its private nature.
The Internet of Things (IoT), or the Internet of Connected Objects, involves a network of physical objects integrating sensors, software and other technologies enabling the connection and exchange of data via the Internet. This data reflects behavior in specific contexts of use. For example, in the case of two connected objects such as baggage conveyors in an airport, their behavior may differ according to the frequency of daily use, linked to the airport’s affluence, which constitutes the context.
However, this perspective raises challenges for the use of this data in various applications, including predictive maintenance. Deployed to predict potential equipment or system failures before they occur, predictive maintenance is based on the analysis of data collected in real time. So how do you build a predictive model that is both unique and applicable to the diversity of data behaviors of each object?
Meta-modeling the real-time statistical heterogeneity of a massive data stream is a major challenge for any data scientist! With this in mind, the simplest route would be to build a predictive model for each connected object. This may seem the best approach, but it poses a major problem in terms of model management and maintenance once the number of connected objects starts to grow. What’s more, this approach can also come up against a major obstacle linked to data sharing. In many cases, data is private and isolated in silos, which can limit access to the data needed to build high-performance models.
An “all-inclusive” model for reliable maintenance and secure data
Faced with these limitations, a new approach proposed by Google in 2017 is federated learning. This method aims to solve data privacy issues by allowing machine learning models to be distributed across local devices, avoiding the need to centralize data on a central server. With this approach, data remains on users’ devices, while allowing a centralized model to be updated by aggregating updates from local models. This privacy-friendly vision ensures data confidentiality while exploiting the benefits of machine learning on distributed data.
Federated learning represents a great opportunity for predictive maintenance, especially given the limitations of data access. This collaborative approach enables us to build a single global model for a group of connected objects. Each connected object drives its model locally with its own data. Then, these local models can be sent to the servers for aggregation, forming a global model for all participants. With this method, we can not only overcome the problems of limited data access, but also share the experiences of each monitored device with the other participants via their local model, thus enriching the overall experience. This is particularly relevant in the case of predictive maintenance, which requires data over the entire lifespan of devices, enabling better anticipation and management of failures.
Smart Transport application with Alstef: equipping luggage conveyors
In BLPredict, our commitment to federated learning (FL) began in 2021. In partnership with Alstef Group, we explored a specific case study in the field of predictive maintenance, focusing on determining the condition of baggage conveyors in airports. The initial results of our work have highlighted the ability of federated learning to deal effectively with this type of problem, particularly in its distributed dimension. During the course of this collaboration, we also identified several challenges inherent in the application of federated learning in real IIoT environments, such as resource constraints and the exchange of updates between servers and participants. To overcome these challenges, we developed and implemented a platform dedicated to federated learning, supporting aspects ranging from participant management to global model deployment.
More robust, secure and precise models with high industrial potential
Federated learning presents both opportunities and challenges for predictive maintenance. On the opportunity side, it can leverage distributed data across different devices to build more robust and accurate global models. This collaborative approach paves the way for better anticipation of equipment failures, thus reducing unplanned downtime and maintenance costs. In addition, it offers a solution to data confidentiality issues, allowing data to remain on local devices while contributing to the improvement of global models.
However, federated learning also poses major challenges, particularly in terms of managing communication between devices, aggregating local models and ensuring the quality and representativeness of the data used. In addition, the diversity of devices and environments can make the harmonization of global models more complex. Nevertheless, by overcoming these challenges, this approach offers considerable potential for transforming predictive maintenance into a more efficient and proactive practice, helping to improve equipment reliability and user satisfaction.
Our work in this area is continuing, notably by tackling issues such as the selection and grouping of participants based on their statistical characteristics to improve model performance.
