QuickAI™

Challenge

• Companies developing endpoint devices often lack the resources to work effectively with the complexities of AI development: data collection and training, feature extraction, and classifier and AI model generation.
• Industrial and manufacturing OEMs, who are working to address the Industrial 4.0 initiative, lack the data science and firmware engineering resources needed to develop AI models and implement them in a low power SoC platform.
• The diversity and uniqueness of endpoint use cases drive the need to develop application- specific algorithms and models, thus requiring large SW teams to address numerous product lines, and even derivative products, that exhibit different sensor characteristics.
• Once endpoint devices are deployed, manufacturers must have a plan to manage the distributed endpoints, leverage the information they collect, and often update AI models remotely.

Solution

QuickAI: An integrated HW/SW solution that enables AI for IoT devices

• Low power sensor processing
• Data capture lab for data collection, labeling and training
• Data analytics SW for AI model and classifier generation
• Neuron development for AI computing
• Automatic firmware development for the S3AI multi-core SoC platform
• eFPGA low power implementation for HW acceleration and feature extraction

SensiML Analytics Toolkit for AI

• An end-to-end software suite that provides OEMs with a straightforward process for developing pattern matching sensor algorithms. It uses machine learning technology and is optimized for ultra-low power consumption.
• Enables OEMs to quickly and easily leverage the power of local AI in edge, endpoint and wearable designs without the need for significant data science or firmware engineering resources.
• Automatically optimizes AI models to minimize power consumption in targeted SoCs, and is designed specifically to leverage the inherent benefits of heterogeneous multi-core SoC architectures and eFPGA technology.
• For more info: www.sensiml.com

Industrial Predictive Maintenance

• Unique model doesn’t scale across similar motor differences in mounting or loading
• Endpoint AI decreases system bandwidth, latency, power

• Data collection, segmenting, labeling
• Sensor input: motion, audio, pressure, temp/humidity, other time series data
• Feature extraction
• Model building

• Sensor Data Creation → Feature Extraction → Feature Vector
• Hardware accelerator (FFT & MFCC)

• Event trigger (segmentation)
• Feature extraction for simpler features
• Ultra-low power AON function

SHM (Structural Health Monitoring)

• Damage detection
• Structural Integrity reporting

• Data collection, segmenting, labeling Sensor input: motion, audio
• Feature extraction
• Model building

• Sensor Data Creation → Feature Extraction → Feature Vector
• Hardware accelerator (FFT)

• Event trigger (segmentation)
• Feature extraction for simpler features
• Vibration (high precision accel) analysis at 200Hz ODR

Smart Watch Application

Voice + AI Algorithms

• Wearable, IoT
• Elderly Care
• Animal Tracker

• Fixed Trigger and KWD
• Low Power Sound Detection

• Activity Classifier (Walk, Run, Bike, Still, Transport)
• Gestures: Bring-to-See, Double-Tap , Rotate
• Context: Fall Detection

• FPGA
• FFE (uDSP) Programmability
• Other Custom Algos on M4

IoT Application

Voice + AI Algorithms

• Wearable, IoT
• Elderly Care
• Animal Tracker

• Fixed Trigger and KWD
• Low Power Sound Detection

• Activity Classifier (Walk, Run, Bike, Still, Transport)
• Gestures: Bring-to-See, Double-Tap , Rotate
• Context: Fall Detection

• FPGA
• FFE (uDSP) Programmability
• Other Custom Algos on M4