Rob Purser
Development Manager for IoT and Hardware Interfacing for MATLAB at MathWorks. Leads the creation of software that connects MATLAB with the real world
Lots Collected, Little Analyzed
- Collection without analysis has negative ROI
Most IoT data are not used currently. For example, only 1 percent of data from an oil rig with 30,000 sensors is examined.
-- McKinsey & Company
- As difficult as it is to collect and capture data,
gaining insights via analysis takes just as much effort
- Dependable, real-world on-line analytics take time to develop, tune, and test.
Not a Product Pitch...but
MATLAB: High-level language and interactive environment
Simulink: Block diagram environment for simulation and design. Simulate, generate code, and verify embedded systems
ThingSpeak: Free web service for storing sensor data and developing IoT applications
- Online analysis using the MATLAB Language
- Sensor data can be sent to ThingSpeak from any internet capable device, including Arduino and Raspberry Pi.
A Project Focused Tour
Tide Prediction - Online Analytics
Nighttime Noise - Historical Analytics
Weather - IoT Analytic Workflow
Counting Cars - Edge Node Analytics
A Project Focused Tour
Tide Prediction - Online Analytics
Nighttime Noise - Historical Analytics
Weather - IoT Analytic Workflow
Counting Cars - Edge Node Analytics
Weather - IoT Analytic Workflow
Weather - IoT Analytic Workflow
MathWorks Weather Station -- Revisited
Classic IoT Maker Project with Machine Learning
Deep Dive: Google "mathworks weather revisit"
Weather - IoT Analytic Workflow
Weather - IoT Analytic Workflow
Weather - IoT Analytic Workflow
Weather - IoT Analytic Workflow
Weather - IoT Analytic Workflow
Weather - IoT Analytic Workflow
Weather - IoT Analytic Workflow
1
2
3
Weather - IoT Analytic Workflow
Weather - IoT Analytic Workflow
Weather - IoT Analytic Workflow
Weather - IoT Analytic Workflow
A Project Focused Tour
Tide Prediction - Online Analytics
Nighttime Noise - Historical Analytics
Weather - IoT Analytic Workflow
Counting Cars - Edge Node Analytics
Car Counter - Edge Node Analytics
Goal: Build a traffic monitor using a Raspberry Pi & Webcam
- Workflow
- Foreground Detection
- Median Filter
- Blob Centroid Detection
- Car Direction & Counter
- Send to ThingSpeak
- Automatic Code Generation and Deployment
to Raspberry Pi
Car Counter - Edge Node Analytics
Counting Cars
Traffic monitor using a Raspberry Pi & Webcam
Deep Dive: Google "count cars thingspeak"
Car Counter - Edge Node Analytics
Counting Cars
Traffic monitor using a Raspberry Pi & Webcam
Deep Dive: Google "count cars thingspeak"
Car Counter - Edge Node Analytics
Embedded devices have always used:
- Signal processing
- Matrix math
- Closed loop control
Powerful mobile processors enable:
- Sensor fusion & analytics
- Computer vision & image processing
- Trained machine learning
- Statistical modeling
- Small vocabulary voice recognition
Car Counter - Edge Node Analytics
Car Counter - Edge Node Analytics
Foreground Detection, 2-D Median Filter, Blob Analysis
Feeds custom counting block and Transmit to ThingSpeak
Car Counter - Edge Node Analytics
Foreground Detection, 2-D Median Filter, Blob Analysis
Feeds custom counting block and Transmit to ThingSpeak
Car Counter - Edge Node Analytics
Foreground Detection, 2-D Median Filter, Blob Analysis
Feeds custom counting block and Transmit to ThingSpeak
Car Counter - Edge Node Analytics
Foreground Detection, 2-D Median Filter, Blob Analysis
Feeds custom counting block and Transmit to ThingSpeak
Car Counter - Edge Node Analytics
Foreground Detection, 2-D Median Filter, Blob Analysis
Feeds custom counting block and Transmit to ThingSpeak
Car Counter - Edge Node Analytics
Foreground Detection, 2-D Median Filter, Blob Analysis
Feeds custom counting block and Transmit to ThingSpeak
Want a deeper dive? Freescale Smart Stovetop
Using machine learning in embedded applications
Car Counter - Edge Node Analytics
Source: Freescale
Embedded devices now have enough compute power to use algorithms that previously were once only practical on desktop class machines or FPGAs
Car Counter - Edge Node Analytics
A Project Focused Tour
Tide Prediction - Online Analytics
Nighttime Noise - Historical Analytics
Weather - IoT Analytic Workflow
Counting Cars - Edge Node Analytics
Night Noise - Exploratory Analysis
Night Noise - Exploratory Analysis
Night Noise Analysis
Identify unusual changes in ambient noise level at night
Night Noise - Exploratory Analysis
Median Filter
Night Noise - Exploratory Analysis
Overlaid with Expected Levels
Determined expected based on median at a given time
Night Noise - Exploratory Analysis
Normalized based on first 3 hours
Polynomial fit of expected noise levels
Night Noise - Exploratory Analysis
Set Expected Levels for a given night
Move polynomial up or down to compensate for variation
Night Noise - Exploratory Analysis
Identify alarm conditions
> 1 std deviation above expected
Night Noise - Exploratory Analysis
Identify alarm conditions
> 1 std deviation above expected
Night Noise - Exploratory Analysis
To do: Investigate what happened on September 10th...
Side Trip into Median Filters
Raw data from a single night
Windowed median filter with 15 minute window
Side Trip into Median Filters
Contrast with mean filter with same window
Side Trip into Median Filters
Things to keep in mind:
- Median filter is your friend
- You haven't seen all the scenarios
- Always consider how you will convert to on-line analytics
- Pilot on-line analytics and alarms
- One false alarm will destroy trust
Night Noise - Exploratory Analysis
A Project Focused Tour
Tide Prediction - Online Analytics
Nighttime Noise - Historical Analytics
Weather - IoT Analytic Workflow
Counting Cars - Edge Node Analytics
Tide Prediction - Online Analytics
Tide Prediction - Online Analytics
On Line Tide Alerts
Tide Measurement and Prediction
Deep Dive: Google "tide matlab thingspeak"
Tide Prediction - Online Analytics
Source: SUNY StoneyBrook
23 astronomical components in tidal harmonics...
plus geography and weather -- it's always an approximation!
Tide Prediction - Online Analytics
Tidal variation between bays 6 miles apart
Edge Node
- Arduino
- Ultrasonic Rangefinder and Temperature Sensor
- Data Reduction
- De-noising and "De-batting"
- Sends Data to ThingSpeak
Tide Prediction - Online Analytics
On-Line Analysis
- Median Filter
- Temperature Compensation
- Convert to Water Depth
- Tweet @tidealerts
- Hysteresis and state machine to avoid being too "tweekative"
Tide Prediction - Online Analytics
On-Line Analysis
- Median Filter
- Temperature Compensation
- Convert to Water Depth
- Tweet @tidealerts
- Hysteresis and state machine to avoid being too "tweekative"
Tide Prediction - Online Analytics
Thank You!
- Collection without analysis has negative ROI
- Analysis happens at every stage
- Dependable, real-world on-line analytics take time to develop, tune, and test
- Trust can be lost in a single moment
- Median filtering is your friend
Rob Purser -- rob.purser@mathworks.com, @rpurser47
http://slides.com/rpurser/environmental_analysis
Greater Boston CMG: Analysis and Visualization Techniques for Environmental IoT Systems
By Rob Purser
Greater Boston CMG: Analysis and Visualization Techniques for Environmental IoT Systems
Data from our environment surrounds us: sensors and inexpensive edge nodes make it easier than ever to collect lots of data on our surroundings, both indoors & outdoors. From weather data and noise levels to counting cars on the highway outside our window, we can correlate this data with external sources of data to gain deeper insights into our surroundings. Working with live environmental time series data brings unique challenges that span traditional instrumentation, embedded programming, signal processing, and statistics. We’ll discuss some of the environmental Internet of Things data collection and analysis that we’ve done, and three kinds of analytics that we’ve applied: sensor analytics at the edge node, historical data analytics, and online analytics at the data aggregator as the data comes in. You’ll see practical techniques for working with this data, tools, and visualizations.
