Acoustic Event Detection
Technologies
Technologies Overview
Acoustic Event Detection
Activity Recognition
Anomaly Detection
Gesture Recognition
Keyword Spotting
Vibration Classification
Products
Industries
Resources
Support
Company
Contact Us
ML Models Optimized for Embedded IoT
Embedded systems and IoT devices present a unique set of challenges for AI/ML given limited memory size, smaller computing cores, competing processor workloads, and often battery-powered energy constraints.
SensiML Knowledge Packs are optimized AI/ML inference models built using tools designed specifically for such devices. So whether your application runs on a 64-bit app processor or an 8-bit microcontroller, chances are high that a SensiML Knowledge Pack can be constructed to provide an autonomous, IoT edge sensor prediction algorithm.
SensiML Knowledge Packs are optimized AI/ML inference models built using tools designed specifically for such devices. So whether your application runs on a 64-bit app processor or an 8-bit microcontroller, chances are high that a SensiML Knowledge Pack can be constructed to provide an autonomous, IoT edge sensor prediction algorithm.
Self-Contained, Simple to Integrate, Fully Transparent
SensiML Knowledge Packs are provided as fully standalone code with no dependencies, runtime interpreters, bloated libraries, or other external needs. Simply include the Knowledge Pack in your firmware device build and with a couple of lines of code, the inference model can be called to feed in sensor streaming data with callback functions to report out recognition events of interest.
Knowledge Packs come either as:
- Binary code - Useful for rapid development and model testing on supported popular IoT eval boards and HDKs
- Library code - Convenient integration using compiled object code
- C Source code - For ultimate flexibility, full non-obfuscated C source code can be generated that is readily modifiable
Knowledge Pack Model Pipeline
SensiML Knowledge Packs integrate a comprehensive pipeline for processing, feature extraction, and classification of time-series signals to provide predictive results that are accurate, compact, and explainable. Where deep learning frameworks when used alone rely on unexplainable abstract data transformation accomplished within hidden layers, SensiML utilizes a mix of traditional DSP pre-processing, feature engineering, and classification. The benefit is code that is readily understandable, explainable, and often able to use simpler classification networks or methods.
Raw Sensor Data
- Time-series data
- Digital or ADC sources
- <1 Hz to 1 MHz sample rates
- 1 to many channels
- Mixed sensor types
Signal Pre-Processing
- Filtering
- Segmentation
- Downsampling
- Averaging
- Vector Magnitude
- Scaling
- Normalization
Feature Extraction
- Fully automated
selection - 80+ feature transforms
- Option for manual
definition / tuning
Classification
- Classic ML
(PME, trees, ensemble) - Neural Network
(TensorFlow Lite) - Hierarchical models
- Fully automated
or manual tuning
Inference Result
- Ordinal class value
- Interim feature vector
- Associated raw data buffer
- ML processing task control
SensiML offers complete Knowledge Pack development services allowing you to focus on your application and not on machine learning and data science tasks. Working in close collaboration with your team, SensiML can devise a customized project plan to take your supplied application dataset(s) and turn them into intelligent embedded firmware recognition libraries ready to drop into your application.
For project teams with basic machine learning familiarity, desire to learn, or needing to undertake the tasks entirely in-house, SensiML offers its ML Analytics Toolkit suite. A true end-to-end workflow, SensiML Analytics Toolkit supports the complete process from data collection and labeling to Knowledge Pack generation and testing.
Common Knowledge Pack Applications
Activity Recognition
Anomaly Detection
Gesture Recognition
Keyword Spotting
Vibration Classification