Magnetic neural Network for predictive maintenance

Project Information

MAG.NET

Grant agreement ID: 101113180

DOI

10.3030/101113180

EC signature date 30 March 2023

Start date 1 May 2023

End date 31 August 2026

Funded under

• The European Innovation Council (EIC)

Total cost

€ 2 499 999,00

EU contribution

€ 2 499 999,00

Coordinated by

GOLANA COMPUTING
France

Periodic Reporting for period 1 - MAG.NET (Magnetic neural Network for predictive maintenance)

Reporting period: 2023-05-01 to 2024-04-30

Summary of the context and overall objectives of the project

Golana Computing, a new startup, is pioneering a new magnetization reversal scheme through the MAG.Net project. This innovative approach uses domain wall (DW) depinning from geometrical chambers to mimic the behavior of biological spiking neurons. This breakthrough enables the design and fabrication of bio-mimicking magnetic neurons capable of recognizing analog signals without the typical requirement of feature extraction.

Our initial work on magnetic spiking neurons began with a proof of concept that demonstrated real-time speech recognition and speaker identification. Following this success, we secured EIC Transition funding in May 2023 under the theme "Green Digital Devices." This European grant has allowed us to advance the development of our innovative magnetic neuronal system.

Our goal is to develop an innovative system for predictive maintenance. Manufacturing industries' maintenance departments need predictive solutions to prevent downtimes, unpredictable or high costs, catastrophic failures, and address production or quality issues. They require a straightforward solution tailored to their machines. This is especially true in the case of SMEs, which lack the data scientists and ML/AI experts able to build such applications. In fact, most manufacturing facilities do not have technological solutions for maintenance or dedicated departments to advance towards Industry 4.0. Our system aims to fill this gap by providing an accessible and effective predictive maintenance solution to enhance operational efficiency and reliability.

The MAG.Net project allows us to develop a technology and create a prototype device with a unique aptitude for predictive maintenance in the manufacturing industry. Indeed, the diversity of different machines and processes encountered in the manufacturing industry, requires a solution that is extremely versatile, based on a device that can be “task agnostic”. In that purpose, our magnetic neuronal network has already demonstrated its ability to process different analog signals on equal footing. Moreover, by operating on the edge, it will reduce waste and enable energy savings, which is particularly crucial in today’s context. This system will also minimize the resources required for data transmission, computation, and security. By the end of the project, we aim to extend this technology to the largest possible range of applications, increasing competitiveness and promoting production relocation to Europe across multiple industries.

Our long term goal is to extend even further the versatility of our solution, for other applications requiring low energy consumption and full privacy. By expanding the application field beyond the analog signals from industrial machines, we will be able to address the needs of other sectors that require innovative, agnostic edge solutions.

By the end of the first year of the MAG.Net project, we have consolidated a large database of analog signals from a diverse set of machines, based on which we have developed our technology. Our first prototype is already capable of detecting errors, failures, anomalous behaviors, and drifts in industrial machines.

Work performed from the beginning of the project to the end of the period covered by the report and main results achieved so far

During the first year of the Mag.Net project we conducted experiments in five different Romanian industrial facilities, monitoring five different types of machines with our data acquisition (DAQ) system. This system, connected to multiple sensors, such as temperature, vibration, current, and sound, was developed in-house with a signal preprocessing circuit. Feedback from our partners allowed us to fine-tune parameters, optimizing the performance of our magnetic neuronal system.

We developed the “waveform-to-spike” conversion algorithm and simultaneously worked on designing and optimizing the performance of our magnetic neuronal architecture. We optimized DW depinning, enhancing the system's response and sensitivity to the relevant information contained in the analog signal.
To avoid potential delays inherent to the development of the physical magnetic device and to permit faster iterations tests, we created a simulated version of the magnetic neuronal network. This allowed us to mature a promising machine learning algorithm able to identify unusual patterns in analog signals issued from the monitored industrial machines.

Our most important achievement in the first year, consists in the establishment of the compatibility of the “full chain” of component operations, from the first step of sensor installation on the machine and the data acquisition, to the last step of alerting of potential machine malfunction. Most tests have been conducted using the simulated version of our system, while a smaller subset has been carried using our physical magnetic system, with promising results.

Progress beyond the state of the art and expected potential impact (including the socio-economic impact and the wider societal implications of the project so far)

Our first year yielded several promising results. Our magnetic system successfully detected real-time anomalous behaviors in signals without requiring prior feature extraction, demonstrating the effectiveness of our optimized magnetic architecture. Additionally, we completed the testing of the full chain operation using the simulated version, and we are nearing the completion of testing with the magnetic system. Additionally, our machine learning algorithm has also proven effective in detecting anomalous events and signal drifts from various sensors placed on industrial machines, showcasing its versatility across different machine types and machining processes.

The potential impacts of our work are significant. Environmentally, our system's efficiency can reduce waste and save energy, contributing to sustainability efforts. Economically, we anticipate both direct and indirect benefits, including reduced downtime, lower maintenance costs, and enhanced productivity. Scientifically, our innovations contribute to advancements in intellectual property and research, driving progress in magnetic systems and machine learning. Furthermore, engineers in maintenance departments will gain new skills, enhancing their capabilities in predictive maintenance and advanced technologies.

To ensure further uptake and success, we have identified several key needs. First, we need to develop a prototype suitable for commercialization to engage potential customers. Regulatory approval and standardization, such as achieving CE marking will be necessary upon developing a minimum viable product (MVP). Finally, ensuring market readiness by overcoming barriers such as costs and facilitating a frictionless adoption process will be essential for successful entering the market.