Anomaly Detection

Pattern Recognition Based Machine Anomaly Detection

Using SensiML generated algorithms and the QuickLogic QuickAI HDK, developers and integrators of programmable sensor devices can offer industrial end-users rapidly customizable smart sensors conforming to a variety outputs.  As machine-specific anomaly detectors, sensors can be built with custom trained responses programmed on-site during system commissioning or retrofit.  With low data rate output of only insights of interest and power efficient microcontroller-based signal processing, long-life battery powered sensors are achievable avoiding high initial installation costs of wired/powered sensor points.

Smart anomaly detection sensors powered by SensiML algorithms can readily feature:

  • Continuous or triggered high frequency raw sampling (up to 100kHz)
  • Multi-channel analog sensor inputs (4 channels x 16kHz)
  • Realtime processing of vibration, audio, acoustic emission, load cell, strain gauge, and piezo sensors
  • Anomaly severity level/quality assessment with event-driven excursion metrics
  • Long-life battery powered operation with efficient signal processing and hardware accelerated feature extractors
  • Local real-time alert/control or remote interfacing over wireless or wired network interfaces
Predictive Maintenance

Pattern Recognition Based Behavior Classification

A step beyond basic anomaly detection, predictive maintenance seeks to provide greater insight by classifying observed patterns from one or more sensors against a model with multiple defined states.  The benefit is greater insight when addressing possible faults.  The challenge is need for either (a) sufficient historical data with examples of each excursion / failure mode to properly train the algorithm or (b) an ability to translate theoretical expectations of as-yet-unobserved classes into a functioning algorithm.

Either way, the SensiML toolkit makes it possible to quickly create embedded predictive classification algorithms capable of executing locally in real-time on the sensor microcontroller.  Application developers can utilize the knowledge contained in existing datasets to automatically generate code or have access to our entire library of transforms, feature extractors and classifiers to customize behaviors where sufficient data for ML training lacks.   SensiML provides both a highly automated process and a fully user managed algorithm build flow and the ability to quickly generate fully functioning firmware quickly that implements accurate classifiers best suited to the application.  This allows expert users the potential for extending beyond basic anomaly detection even in cases where historical fault data is insufficient to develop empirically trained models.

  • Models that can be trained in as few as 30-50 samples for conceptual testing
  • Support of a variety of classifiers ranging from kNN, RBF distance based classification to tree, and hierarchical models to binarized neural networks
  • Continually growing library of over 80 supported features
  • Server-based code generator that optimizes feature vector selection, event segmentation and triggering, and classifier selection to best suit the problem
  • Platform specific auto code-generation for a variety of platforms spanning ARM and x86 architecture with HW accelerators for DSP and FPGA capable SoCs
  • Executable code that as small as 40KB SRAM
Process Control and Robotics

High Speed Inspection and Sorting Processes

Nearly every high volume manufacturing process has one or more critical sort/inspection stages that must operate in real-time at high line speeds.  These control points invariably use complex and critical sensor measurements to make decisions directly impacting product yields and overall quality control.  Measurement involves one or more of optical, audio, profile, ultrasonic, or electrode based sensor signals used to compare some physical property against established reference patterns for classification and subsequent action (i.e. sorting, rework, reporting, and upstream process adaptations).

As important as the physical measurement itself, effective sort/inspect systems must possess the ability to analyze the rich sensor signals to detect often subtle differences before they become major excursions leading to rework or scrap.  SensiML generated algorithms combined with such sensors provide the intelligence and real-time responsiveness to support high performance inspection and sort.  Industrial sensor vendors and systems integrators selling solutions to end-user manufacturers can truly differentiate their offerings working with SensiML to build adaptive machine learning into their products and projects.

Industrial Wearables

Worker Safety / Work-aid Augmentation Sensors

There are numerous industrial applications for smart sensing human wearable devices aiding either worker safety or improving day-to-day job efficiency through real-time feedback on quantitative measures readily classified on operator worn or tooling integrated sensors.  SensiML originated from work in advanced motion sensing for next-gen wearable computing and has proven examples where human motion analysis either alone or in combination with other sensor types can inform workers and operations managers in ways to improve job safety, quality, and productivity.  Examples include:

  • Operator wearables for real-time status monitoring of defined activities, gestures, and environmental conditions for first responders and industrial workers.
  • Smart tooling with sensors and SensiML algorithms capable of alerting for conformance to specific motion, exertion, power/torque, and rotational patterns.