Notes on Enabling Continuous Learning for ML-Based Anomaly Detection at the Edge

Introduction

• Speaker: Journy Plunket, Senior Developer Relations Engineer at Edge Impulse
• Topic: Continuous learning for ML-based anomaly detection using Edge Impulse
• Key areas of discussion:
  • Overview of Edge Impulse
  • Advanced anomaly detection and its applications
  • Live demo of Edge Impulse Studio

What is Edge Impulse?

• Leading embedded machine learning platform
• Functionality includes:
  • Collecting and labeling training/testing data
  • Designing and training ML models
  • Testing and validating model performance
  • Deploying models to Edge devices
• Applications span various industries:
  • Predictive maintenance
  • Asset tracking
  • Quality assurance
  • Human interface systems (occupancy detection, speech recognition, predictive healthcare)

Continuous Learning for Edge ML Projects

• Continuous learning enables ML models to improve over time:
  • Sending anomaly detection data to the cloud
  • Retraining and redeploying models to Edge devices
• Benefits include rapid detection of machine failures (e.g., industrial machines)
• Advanced MLOps infrastructure:
  • Data protection, model management, user access control
  • Integration with platforms like Google Cloud and AWS
  • Advanced data set building with quality control

Advanced Anomaly Detection

• Definition: Use of machine learning on the edge to identify machine failures preemptively
• Importance:
  • Monitors equipment health to prevent costly repairs
  • Uses data (e.g., vibration, sound) for anomaly detection
• Workflow:
  1. Collect data for nominal operation and off state
  2. Process data using Digital Signal Processing (DSP)
  3. Train ML model on extracted features
  4. Validate and deploy model

Live Demo Overview

• Use case: Vibration anomaly detection for HVAC systems
• Data collection for two classes:
  • Nominal operation
  • Off state
• Feature extraction using DSP blocks in Edge Impulse Studio
• Importance of feature extraction for efficient models

DSP Blocks in Edge Impulse Studio

• Pre-written DSP blocks available for immediate use
• Custom DSP blocks can be created and integrated
• Feature importance tool suggests key features for training models

Model Training and Testing

• Classifiers and anomaly detection blocks:
  • K-means anomaly detection block for sensor data projects
• Model testing without deployment:
  • Cross-validation of classification results
  • Anomaly score calculations
• Deployment options:
  • C++ library for embedded firmware
  • No black boxes, full source code access

Other Use Cases for Anomaly Detection

• Predictive maintenance applications include:
  • Air conditioner filter monitoring
  • Circuit breaker aging detection
  • Leak detection (water/hazardous)
  • Motor control and condition monitoring

Resources for Further Learning

• Public projects available for cloning in Edge Impulse account
• Documentation and tutorials for:
  • Advanced anomaly detection
  • DSP blocks
  • Custom processing blocks

Live Demo Highlights

• Project dashboard overview in Edge Impulse Studio
• Data acquisition and testing setup
• Designing and training models with feature importance insights
• Continuous learning through retraining models
• Deployment of anomaly detection models to Edge devices

Conclusion

• Successful implementation of anomaly detection models in embedded environments
• Continuous learning and model refinement are key to effective machine learning applications at the edge.