tinyML Talks on January 30, 2024 “NanoEdge AI Studio: All On-Device Anomaly Detection for Industry 4.0” by He Huang and Pierrick Autret

We held our next tinyML Talks webcast. He Huang and Pierrick Autret from STMicroelectronics presented NanoEdge AI Studio: All On-Device Anomaly Detection for Industry 4.0 on January 30, 2024.

Anomaly detection is a critical aspect of predictive maintenance in Industry 4.0, requiring the ability to detect unusual patterns or behaviors in data collected from sensors. NanoEdge AI Studio, an auto-ML software developed by STMicroelectronics, enables the creation and training of machine learning models that can perform anomaly detection fully on-device, without the need for any pre-trained models. In this presentation, we will study the algorithms behind NanoEdge AI Studio and demonstrate its potential applications of predictive maintenance in Industry 4.0.

He Huang is a data scientist and researcher with expertise in embedded Machine Learning. He holds a Ph.D. in electronics from INSA Toulouse in France, earned in 2015 and completed a post-doctoral research fellowship at Politecnico di Torino in Italy in 2016. As the former chief data scientist at French start-up Cartesiam from 2017, He played a pivotal role in the development of NanoEdge AI Studio, an innovative platform for on-device machine learning at the edge. Currently, he serves as senior staff data scientist at STMicroelectronics, where he continues to advance the field of machine learning and develop cutting-edge solutions for Industry 4.0 applications.

Pierrick Autret has been Artificial Intelligence Solutions Manager in STMicroelectronics’ microcontroller division since 2019. With a master’s degree in engineering specializing in Embedded Systems from the Ecole Polytechnique and a master’s degree in Marketing, Economics and Management from the IAE d’Aix-en-Provence, he began his career as Hardware Product Owner on the STM32 series before developing ST’s software offering.

=========================

Download presentation slides:
He Huang and Pierrick Autret

Watch on YouTube:
He Huang and Pierrick Autret

Feel free to ask your questions on this thread and keep the conversation going!

=========================

Q: Very good presentation! Are there any videos of real applications at customer locations for On Device Learning Best
A: Please find more examples in this playlist: https://www.youtube.com/playlist?list=PLnMKNibPkDnG9jgzxGxxfcmTXzJGeNCKY

Q: Can you talk about the factors that would affect the size of a model on the edge device?
A: The size of a model can be influenced by various factors, such as the signal buffet size and the complexity of the application, including the number of categories. To address this issue, NanoEdge AI Studio conducts benchmarks to identify the most optimal solution within the limited memory resources defined by the users.

Q: If it’s possible for you, can you comment on the actual model used for on-device learning? Is it a common anomaly detection model or is it a new proposal? Does it construct an “average example” of the time series from train data and then compute the distance of the newly available data from that?
A: Our tool provides five on-device learning methods that utilize various statistical features, including the “average example” method you mentioned earlier. It’s worth noting that most of these methods are patented, and you can find more information about them in our documentation.

Q: How about flash reliability after multiple writes.
A: Same as other kinds of calculations.

Q: It may take a longtime for anomalies to occur in the working equipment and for the this edge device to learn and be effective.
A: Only normal signals are used for the model training.

Q: Do you provide on-device training capabilities for the classification models also or just for the anomaly detection?
A: We support only anomaly detection today.

Q: Can you predict failure in the future?
A: Our model is primarily designed for detecting failures. However, in most cases, system failures do not occur suddenly but rather progress over time. Therefore, our model proposes the use of an anomaly score as the inference result. By monitoring the anomaly score over time, we can predict potential failures before they occur.

Q: Could some MCUs run many models in the same core by combining TinyML with RTOS. ie. is it possible to implement many models or a multi-task model?
A: NanoEdge AI studio support multi librairy https://wiki.st.com/stm32mcu/wiki/AI:NanoEdge_AI_Studio#Multi-library, and they can run with RTOS.

Q: Also please send us an email later containing the recording and the presentation pdf or pptx.
A: https://cms.tinyml.org/wp-content/uploads/talks2023/tinyML_Talks_He_Huang_and_Pierrick_Autret_240130.pdf