Three industrial projects show how data, features and compact models make edge AI hold up on the factory floor.

Industrial edge AI often works in the lab, then struggles on the factory floor. Background noise, seasonal change and scarce fault data are common reasons.

This on-demand webinar shows how three real projects solved those problems: welding fault detection with audio, fault detection on outdoor pumps, and motor degradation estimation with KTH Royal Institute of Technology. You will learn how data strategy, feature engineering and ensembles of small, specialized models keep edge AI reliable in industrial settings.

The session was first presented as "Real-time Edge AI: Reliability and Resiliency through Design Strategy" at Embedded Computing Design's AI at the Edge Day. Speakers from Infineon and Imagimob also show how PSOC™ Edge E84 microcontrollers and the DEEPCRAFT™ AI Suite support each step, from data collection to deployment.

Who should watch this industrial edge AI webinar?

Embedded engineers, machine learning engineers, system architects and technical leaders who are evaluating or building edge AI for industrial monitoring. It is most useful if you work on predictive maintenance or on audio- and sensor-based fault detection.

Register once, and we'll email you the link to the recording.


What will you learn about industrial edge AI?:

  • Turn scarce fault data into a robust training set. Mix lab-simulated faults with recorded factory noise, then augment distance and loudness so the model learns the fault signature, not the background.
  • Keep a deployed model improving. An end-to-end pipeline that records audio while it runs inference captures misclassifications for retraining.
  • Make a model generalize across seasons. Season-agnostic features, designed with the customer's engineers, let a model trained on a few months of data target year-round use.
  • Catch more fault types with small models. An ensemble of specialized compact models detected faults that a single compact model missed.
  • Separate unusual readings from real anomalies. Percentile thresholds at 95, 99 and 99.9 grade deviations; only events above 99.9 counted as real anomalies.
  • Find degradation in low-rate data. A small CNN+LSTM model tracked motor winding degradation from 4 kHz current and voltage data.
  • Know your validation limits. Single-motor test data and mixed stress and wear effects restrict what results prove.
  • Choose between noisy training data and audio enhancement. Train on realistic noise for fault detection; use audio enhancement for voice control.
  • Run several models on one MCU. PSOC™ Edge ran five models at once in a customer use case, with capacity left.
Documents

Speakers

Eduardo Montanez
Eduardo Montanez

Vice President, Head of PSOC Edge Products & Edge AI Solutions, Infineon

Eduardo is an engineering executive with more than 20 years in the semiconductor industry. He focuses on IoT edge processors with low power, machine learning, security and wireless connectivity. He has held roles in systems and architecture, applications engineering, business development and marketing. Today he leads Infineon's PSOC™ Edge microcontroller product line and complementary edge AI solutions for voice, vision and sensing.

Sam Al-Attiyah
Sam Al-Attiyah

Head of Machine Learning, Imagimob, An Infineon Technologies Company

Sam has spent nearly a decade at Imagimob, an Infineon Technologies company that develops the DEEPCRAFT™ brand of edge AI software. He holds a master's degree in electrical engineering, earned through studies in Australia and Sweden. He has led multiple end-to-end edge AI projects to production. He now heads Imagimob's machine learning team, which helps customers accelerate their machine learning journey.

Industrial Edge AI: Design for Reliability and Resiliency
Register to watch
Online | Webinars | On-demand

Additional Resources:

PSOC™ Edge – The Next Generation of ML Enabled MCU
DEEPCRAFT™ Studio
Infineon Developer Community: PSOC™ Edge
Infineon Developer Community: DEEPCRAFT™

Frequently asked questions about industrial edge AI

It shows how three industrial projects made edge AI reliable on real equipment: welding fault detection with audio, fault detection on outdoor pumps, and motor degradation estimation with KTH. The focus is data strategy, feature engineering, compact model ensembles and validation, running on PSOC™ Edge microcontrollers.

Click "Register to watch" and complete the short form. We'll email you the link to the on-demand recording. The webinar is free and runs about 60 minutes: a 45-minute presentation followed by Q&A with both speakers.

In the welding case, the team simulated faults in the lab, then mixed those recordings with factory background noise and varied distance and loudness. The model learned the fault signature rather than lab conditions. Recording in the field during inference added real errors to later training rounds.

Yes. In a customer use case, five models ran at the same time on a PSOC™ Edge microcontroller, with capacity left over. That headroom makes ensembles of small, specialized models practical, as the pump-monitoring case shows.

Not always. PSOC™ Edge E84 pairs an Arm® Ethos™-U55 NPU with a Cortex®-M55 for heavier models. According to the speakers, PSOC™ Control and PSOC™ Multi-Sense microcontrollers can run some models on a Cortex®-M33 alone. It depends on the inference performance your use case needs.

For fault detection, the speakers prefer training on realistic noisy data, because a filter cannot anticipate every sound, from voices to dropped tools or alarms. For voice control, audio enhancement that cleans the microphone signal is the better fit.

It shows how three industrial projects made edge AI reliable on real equipment: welding fault detection with audio, fault detection on outdoor pumps, and motor degradation estimation with KTH. The focus is data strategy, feature engineering, compact model ensembles and validation, running on PSOC™ Edge microcontrollers.

Click "Register to watch" and complete the short form. We'll email you the link to the on-demand recording. The webinar is free and runs about 60 minutes: a 45-minute presentation followed by Q&A with both speakers.

In the welding case, the team simulated faults in the lab, then mixed those recordings with factory background noise and varied distance and loudness. The model learned the fault signature rather than lab conditions. Recording in the field during inference added real errors to later training rounds.

Yes. In a customer use case, five models ran at the same time on a PSOC™ Edge microcontroller, with capacity left over. That headroom makes ensembles of small, specialized models practical, as the pump-monitoring case shows.

Not always. PSOC™ Edge E84 pairs an Arm® Ethos™-U55 NPU with a Cortex®-M55 for heavier models. According to the speakers, PSOC™ Control and PSOC™ Multi-Sense microcontrollers can run some models on a Cortex®-M33 alone. It depends on the inference performance your use case needs.

For fault detection, the speakers prefer training on realistic noisy data, because a filter cannot anticipate every sound, from voices to dropped tools or alarms. For voice control, audio enhancement that cleans the microphone signal is the better fit.