Download as pdf or txt
Download as pdf or txt
You are on page 1of 32

Machine Learning with Microchip

A Leading Provider of Smart, Connected and Secure Embedded Control Solutions


What is Machine Learning?
• A program or system that builds and trains a predictive model from input data. The
system uses the learned model to make useful predictions from new data.

• Typically implemented using “neural networks” and generally classified as:


• Supervised Learning (Classification, Diagnostics, Regression)
• Unsupervised Learning (Cluster Analysis)
• Reinforced Learning (Skill acquisition)

• What is artificial intelligence


• A non-human program or model that can solve sophisticated tasks.
• Formally, machine learning is a sub-field of artificial intelligence. However, in recent years,
sometime artificial intelligence and machine learning are used interchangeably.

2
How does Supervised Learning work?
The basic machine learning training process

Server, Infrastructure
Tensorflow, Caffe etc.

App
Integration
“inference”

33
ML vs Rules-Based Algorithm
! Benefits
• Much faster development time
• Days vs Month
• Scalable Accuracy of the Prediction
• More Processing Power and more Data variety = More accuracy
• Local Learning
• Decrease false positives or negatives
• Create new possibilities
• Can find “hard to see” pattern

! Challenges
• Needs data to train the model
• Labelled / Curated
• Cover all possible scenarios
• Could be complex to develop the right model/inference
• May need Data Scientist and ML expertise
• Bigger memory footprint
4
Challenges of data processing
Cloud, Big Data, IoT, Edge – What does this all mean?

Methods needed to More devices, sending


process large amounts Push the more data resulted in
of noisy data resulted in large cost and power
data processing
Machine Learning/ Increases at data
Artificial Intelligence to the “Edge” centers

5 5
Why ML at the Edge?
Edge is the location of Data Processing near the source of Data Generation
• Reduce Cost
• Less Communication

• Low Power
• No need for fast communication

• Better Customer Experience


• Faster response time

• Privacy and Security


• Less data transit

• Real Time Operation


• Low Latency
• Safety operation

• Local Learning
• Better performance by learning each product individually
6
Machine Learning with Microchip

Highly efficient, low ML based audio and High performance,


Implement ML based
power, low cost HMI like gesture high port count PCIe
predictive maintenance
implementation of recognition on MCU. switch enable
using our MCU solutions
neural networks as well interconnection of
specialized for motor
as image / video MemBrain silicon IP for GPUs for machine
control, power
processing solutions for NN accelerator in low learning training and
management and
embedded vision cost silicon platforms. inference.
thermal management.
applications.
7
Silicon for Each Stage of ML Data Flow

8
Machine Learning with Microchip
MCU/MPUs

9
Who is it for?

• Embedded Designer

• Does not require deep knowledge of machine learning

• Any application with a Microchip Arm® Cortex®-based 32-bit


microcontrollers and microprocessors

• Any application with sensors/data

10
Why ML with Microchip MCU/MPUs?
• Easy to implement
• Complete flow from Data generation to Chip programming
• Partners’ solution will help with Data needs
• Partners integration within MPLAB
• Accessible to embedded engineers

• Wide Portfolio of MCUs and MPUs


• Complete Application integration

• Low Power

• Expertise in Communication and Security

• Technical Support

11
ML Flow with Partners

12
Target Areas for MCU/MPU
Smart Predictive Maintenance (SPM)
Smart HMI (SHMI)

13 13
Industry 4.0

14
Different Maintenance Strategies

Reactive Maintenance Preventive Maintenance Predictive maintenance


Only repair an asset once failure Inspecting and performing Work that is scheduled as-needed
occurs maintenance, regardless of whether based on real time conditions of
• Lower Cost / Less Staff the equipment was in need of assets
maintenance
• Increase added cost due to
unplanned downtime • Easy to implement
• Increased labor cost ( Overtime • Risk of over maintaining
maybe needed ) • Failure still likely to occur
• Inefficient use of staff resource • Labor Intensive

15
Smart Predictive Maintenance ( PdM )
• Optimize asset availability and lifetime
• Reduce Maintenance cost up to 30%*
• Eliminate breakdown up to 75%*
• Reduce downtime up to 45%*
• Minimize unplanned downtime
• Increase production by up to 25%*
• Extend equipment life
PwC Source

• Increase Investment in Diagnostic Equipment


• Saving potential not easy to measure

*Department of Energy, O&M Best Practices Guide, Release 3.0

16
Smart Predictive Maintenance
• A single hour of IT downtime costs
• >$100,000 for 98% of firms polled
• >$300,000 for 86% of firms polled
• >$1 Million for 34% of firms polled
• Excludes lawsuits, fines, and penalties
• Source: Information Technology Intelligence Consulting

• An hour of factory downtime?


• 82% of firms polled experienced an average factory shutdown of 4 hours
• Average cost per shutdown was $2 Million ($500,000 per hour)
• Predictive maintenance decreases unplanned downtime
• Source: Vanson Bourne Research Study

• A failing bearing on a cooling fan or electric motor can cost your IT or


Industry 4.0 client’s customers $300,000 - $500,000 per hour of shutdown
17
Smart Predictive Maintenance
• Microchip can help designing the best solution:
• Easy and Inexpensive to install
• Wireless: LoRa, Wifi, Security
• Long Battery Life: Low Power MCU/MPU
• Ease of use
• Local ML Learning option
• Performance
• ML integration at the Edge
• Analog performance

18
ML in Smart HMI
• Enhance Customer experience through gesture-based interaction

• Increase ease of use for operators by making it more intuitive,


increase efficiency and decrease downtime

• Help to create a bond with the product or the brand

• Create safer environment

19
Automotive

20
Automotive Trend
• Increased amount of sensors in a car
• Sensor data requires higher data bandwidth
• The current network bandwidth may not be sufficient

• Car makers are evaluating to upgrade networks to support higher


bandwidth
• Increased development cost with changed network architectures

21
How to save cost and enhance feature?
• Reduce network bandwidth by implementing AI/ML at
sensors / nodes
• Reduce unnecessary data transmit to the ECU
• Reduce overall system power

• Car makers could keep the current network architecture


without paying more for upgrade

• Enhance Features:
• Better predicted maintenance
• Reduce maintenance cost
• Enhance safety by properly notifying driver on failures
22
Improving Sensor Performance
• What if we can improve sensors performance?
• What is the cost of a false negative or positive?
• Driver needs to go to a shop
• Shop will change a perfectly functioning unit
• Shop will charge the OEM

• Machine Learning can help:


• Can be tuned and train to get better
• Can learn at the edge

23
Partners

24
Cartesiam
Intelligence of Things
• NanoEdge AI Studio
• Your embedded developer becomes your AI developer
• Just need to provide small sample of typical Data
• Self select algorithm for accuracy and confidence
• Small footprint algorithm (4kB to 32kB RAM)
• Simplified testing with integrated emulator
• Sensors options
• Vibrations, Pressure, (VOC) Volatile organic compound, Current, Magnetic,
Temperature, Ultrasound, Multi signal sensor
• Does not work with Video or Voice

25
Edge Impulse
Making Things smarter
• Full tinyML pipeline
• Collect Data
• Build a model
• Deploy and Monitor

• Leverage TensorFlow Lite for MCU

• All code output is open-source, royalty free

26
Motion Gestures Solutions
Gesture Recognition in minutes!
• Easy to create your own gesture

• No data collection required

• Reduce software development by 10x

• Increase gesture recognition to nearly 100%

• Can use Touch or Motion ( i.e IMU ) sensors


27
Getting Started

28
Machine Learning Evaluation Kit
• EV18H79A: SAMD21 ML Evaluation Kit with TDK 6-axis MEMS
• ATSAMD21
• A light sensor
• A high-accuracy temperature sensor - MCP9808
• WiFi - WINC1510
• TDK ICM-42688-P

• EV45Y33A: SAMD21 ML Evaluation Kit with BOSCH IMU


• ATSAMD21
• A light sensor
• A high-accuracy temperature sensor - MCP9808
• WiFi - WINC1510
• Bosch BMI160

• Easy to evaluate Microchip Solutions and our partners SDK


29
Smart HMI
• Motion Gestures Demo
• SAM C21 Xplained Pro
• QT8 Touch pad
• HEX file available with limited set of Gestures
• Fastest way to test the accuracy
• Create your own Gesture
• By sending a drawing directly to Motion Gestures SDK
• Simulation in the cloud
• Contact Motion Gestures for Library

30
Adafruit Edge Badge
TensorFlow Lite for Microcontrollers
• Based on SAMD51
• 1.8” TFT Display
• 2MB of SPI Flash
• Microphone, Accelerometer, Light Sensor

• Quick tool-chain setup using Arduino

• Example code for microspeech detection

• Videos, and step-by-step documentation.


31
Microchip.com/ml

32

You might also like