 Hello, I'm part of the Prospective and Innovation Team at FIFER, a vacuum based in ANSI, and I've been working for several years on an embedded H-I solution. On our ANSI site, we design and manufacture pumps, particularly for semiconductor industry. We are a major global player in the vacuum technology industry. We are part of the ongoing digitization of this domain, with our latest generation of products aiming to offer advanced predictive maintenance services. We attach great importance to creating more sensible products through digital solutions. Our goal was to create an onboard predictive maintenance solution that could be developed and validated quickly in the field. The solution has to be able to detect a malfunction on our vacuum dry pump and characterize this anomaly via an onboard classification tool. It should be able to estimate a remaining useful lifetime before machines shut down, with a warning time of at least four days and with a confidence rate of over 90%. This solution should be integrated on pumps in a non-intrusive and repeatable manner. It has to be plug-and-play cost-effective, with low power consumption and have IoT-reheable connectivity. Also, it should not impact existing customer network if it's secure. We started to collaborate with ST on edge computing, working with the STM32 MCU and their three-axis MEMS accelerometer. Important to us was the ability of cost-effective solutions and the capability to run machine learning algorithms on their MCU. Also, development tools such as STM32 QBID and development board like STWIN with MCU plus sensor allowed us to concentrate on the software design. We worked with ST, NanoHIE Studio, which allowed us to switch from static embedded code to a dynamic anomaly detection solution. Progress was fast and development type for embedded eye code was reduced from months to days. We became more autonomous in AI library, building and evaluation and were able to integrate our vacuum pump expertise in the code at each development step. Our vacuum pumps are perfectly suited for non-board learning, allowing for adjustment to be made through the pump life cycle, even during repair cycle where minor modification to the vibration spectrum must be interpreted into the model. This tuning is also necessary for difference in vibration spectrum among the same pump family. As the customer process environment evolved, the dynamic adjustment on an onboard AI model had even more value. Our technology has been applied in two ways today. A new multi-connectivity IOT card for installed base market and an intelligent sensor for new generation of dry pump that will directly integrate sensing and HII capability. Our collaboration with ST on this project has been characterized by enthusiasm, exchange and trust. We are currently testing and optimizing the solution we have been deployed at ST site. The first results are very encouraging, achieving the needs, reliability and precision for failure prediction. FIFER or vacuum LED team already integrate advanced name sensor and an STM32 microcontroller technology. They are actively researching new use cases for IOT-enabled pumps that benefit from the possibility of embedded AI. Thanks ST technology and their support, we have been moved from a concept to what we see a main full innovation for our customer.