Qualcomm
Predictive Maintenance model on QCX217 Board
Solution Summary
Adaptable and Robust Anomaly Detection – Time Series (ARAD-T):
Aizip time-series foundation model: pretrained on large dataset from diverse sensors, excellent generalization ability
Easily adapted to different equipment (motors, engines, HVAC, servers) & sensors (accelerometer, pressure sensor, EEG)
On-device registration, no training needed
Only need normal data. No need to collect large anomaly dataset
Efficient model can run locally on variety of device (MCUs, CPUs, GPUs, etc.)
Advantages and Values to the Customers:
More accurate & robust than traditional DSP based solutions
Significantly shorten development time compared to conventional deep learning workflow
Edge based solution enhances scalability for large equipment fleets
Cost saving
Predictive Maintenance demo:
Predictive maintenance for industrial machines using vibration data.
Registers “normal operation” of the machine during a registration phase.
Detects deviations from registered operation patterns and triggers an alert.
Adaptability: The model adapts to various industrial settings through an easy and fast registration phase.
Real-Time Response: Detects anomalies quickly.
Runs on a standard ARM Cortex-M3 processor.