Jeff Fischer - Python and IoT: From Chips and Bits to Data Science
2 likes444 views
This document summarizes a project that uses IoT devices, Python, and machine learning to predict and replay home lighting patterns. Sensor nodes with ESP8266 microcontrollers capture light sensor data and send it to a Raspberry Pi for analysis. Hidden Markov models are trained on the data and used to generate predicted lighting states. A player script then controls smart lights to match the predictions. The project demonstrates an end-to-end IoT and machine learning workflow for an autonomous home lighting application.
Jeff Fischer - Python and IoT: From Chips and Bits to Data Science
- 1. IOT, PYTHON, AND ML: From Chips and Bits to Data Science Jeff Fischer Data-Ken Research jeff@data-ken.org https://data-ken.org Sunnyvale, California, USA PyData Seattle July 6, 2017
- 2. Agenda ¨ Project overview ¨ Hardware ¨ Data capture ¨ Data analysis ¨ Player ¨ Parting thoughts © 2016, 2017 Jeff Fischer
- 3. Why Python for IoT? ¨ High-level, easy to prototype ideas and explore options ¨ Runs on embedded devices ¨ Python data analysis ecosystem © 2016, 2017 Jeff Fischer Array and matrix processing High level data analysis tools Numerical analysis routines Machine learning Raspberry Pi • Linux “workstation” • Can run CPython and full data science stack • Not battery friendly ESP8266 • System-on-a-chip with 32-bit CPU, WiFi, I/O • Low power consumption • Only 96K data memory! • MicroPython to the rescure
- 4. Project Motivation ¨ First thought about smart thermostat, but too dangerous ¨ Lighting is “safe” ¨ If out of town for the weekend, don’t want to leave the house dark ¨ Timers are flakey and predictable ¨ Would like a self-contained solution ¨ “Wouldn’t it be cool to use machine learning?” © 2016, 2017 Jeff Fischer
- 5. Lighting Replay Application Data Capture Lux Sensors ESP8266 remote nodes + Raspberry Pi Analysis and Machine Learning Offline analysis using Jupyter, Pandas, HMMlearn Captured sensor data Smart Lights Player Application Simple script using HMMlearn and Phue to control Philips Hue lights HMM state machines © 2016, 2017 Jeff Fischer
- 6. Hardware
- 7. ESP8266 © 2016, 2017 Jeff Fischer TSL2591 lux sensor breakout board Lithium Ion Polymer Battery 3.7v 350mAh MicroUSB to USB cable ½ Size breadboard Adafruit Feather HUZZAH ESP8266 breakout board
- 8. ESP8266: Wiring Diagram © 2016, 2017 Jeff Fischer SDA SCL GND 3V
- 9. Raspberry Pi © 2016, 2017 Jeff Fischer Raspberry Pi 2 Breakout cable “Pi Cobbler Plus” Solderless Breadboards Resistor LED TSL2591 lux sensor breakout board
- 10. Raspberry Pi: Wiring Diagram © 2016, 2017 Jeff Fischer Resistor LED Anode (long lead) Cathode (short lead) 10k GND 3.3V SDA SCL GPIO 0
- 11. Data Capture
- 12. Lighting Replay Application: Capture Lux Sensor ESP8266 Front Bedroom Sensor Node Lux Sensor ESP8266 Back Bedroom Sensor Node Raspberry Pi (Dining Room) MQTT Broker Data Capture App Lux Sensor Flat Files © 2016, 2017 Jeff Fischer MQTT MQTT
- 13. Event-driven IoT Code Can Be Ugly def sample_and_process(sensor, mqtt_writer, xducer, completion_cb, error_cb): try: sample = sensor.sample() except StopIteration: final_event = xducer.complete() if final_event: mqtt_writer.send(final_event, lambda: mqtt_writer.disconnect(lambda: completion_cb(False), error_cb), error_cb) else: mqtt_writer.disconnect(lambda: completion_cb(False), error_cb) return except Exception as e: error_cb(e) mqtt_writer.disconnect(lambda: pass, error_cb) return event = SensorEvent(sensor_id=sensor.sensor_id, ts=time.time(), val=sample) csv_writer(event) median_event = xducer.step(event) if median_event: mqtt_writer.send(median_event, lambda: completion_cb(True), error_cb) else: completion_cb(True) def loop(): def completion_cb(more): if more: event_loop.call_later(0.5, loop) else: print("all done, no more callbacks to schedule") event_loop.stop() def error_cb(e): print("Got error: %s" % e) event_loop.stop() event_loop.call_soon(lambda: sample_and_process(sensor, mqtt_writer, transducer, completion_cb, error_cb)) Problems 1. Callback hell 2. Connecting of event streams intermixed with handling of runtime situations: normal flow, error, and end-of-stream conditions. 3. Low-level scheduling 4. async/await helps, but not much © 2016, 2017 Jeff Fischer
- 14. My Solution: ThingFlow ¨ What is ThingFlow? ¤ A Domain Specific Language for IoT event processing ¤ Runs on Python3 and MicroPython ¨ Co-creator ¤ Rupak Majumdar, Scientific Director at Max Planck Institute for Software Systems ¨ Why did we create ThingFlow? ¤ IoT event processing code can be very convoluted ¤ No standardization of sensors, adapters, and transformations ¤ Different frameworks for microcontrollers, edge processing, analytics © 2016, 2017 Jeff Fischer
- 15. Simple ThingFlow Example o Periodically sample a light sensor o Write the sensed value to a local file o Every 5 samples, send the moving average to MQTT Broker © 2016, 2017 Jeff Fischer Lux Sensor Write to File Event Scheduler Send to MQTT Moving Avg Graphical Representation sensor.connect(file_writer(’file’)) sensor.transduce(MovingAvg(5)).connect(mqtt_writer) scheduler.schedule_periodic(sensor, 5) Code
- 16. ESP8266 ThingFlow Code © 2016, 2017 Jeff Fischer from thingflow import Scheduler, SensorAsOutputThing from tsl2591 import Tsl2591 from mqtt_writer import MQTTWriter from wifi import wifi_connect import os # Params to set WIFI_SID= … WIFI_PW= … SENSOR_ID="front-room" BROKER='192.168.11.153' wifi_connect(WIFI_SID, WIFI_PW) sensor = SensorAsOutputThing(Tsl2591()) writer = MQTTWriter(SENSOR_ID, BROKER, 1883, 'remote-sensors') sched = Scheduler() sched.schedule_sensor(sensor, SENSOR_ID, 60, writer) sched.run_forever() https://github.com/jfischer/micropython-tsl2591 Sample at 60 second intervals The MQTT writer is connected to the lux sensor. See https://github.com/mpi-sws-rse/thingflow-examples/blob/master/lighting_replay_app/capture/esp8266_main.py
- 17. Raspberry Pi Code © 2016, 2017 Jeff Fischer Lux Sensor MQTT Adapter Map to UTF8 Parse JSON Map to events Dispatch CSV File Writer (front room) CSV File Writer (back room) CSV File Writer (dining room) https://github.com/mpi-sws-rse/thingflow-examples/blob/master/lighting_replay_app/capture/sensor_capture.py
- 18. Data Analysis
- 19. Lighting Replay Application: Analysis Raspberry Pi (Dining Room) Flat Files HMM definitions Laptop Jupyter Notebook file copy © 2016, 2017 Jeff Fischer
- 20. Preprocessing the Data (ThingFlow running in a Jupyter Notebook) © 2016, 2017 Jeff Fischer CSV File Reader Fill in missing times Sliding Mean Round values Output Event Count Capture NaN Indexes Pandas Writer (raw series) Pandas Writer (smoothed series) reader.fill_in_missing_times() .passthrough(raw_series_writer) .transduce(SensorSlidingMeanPassNaNs(5)).select(round_event_val).passthrough(smoothed_series_writer) .passthrough(capture_nan_indexes).output_count()
- 21. Data Processing: Raw Data © 2016, 2017 Jeff Fischer Front room, last day Data gaps
- 22. Data Processing: Smoothed Data © 2016, 2017 Jeff Fischer Front room, last day
- 23. Data Processing: K-Means Clustering © 2016, 2017 Jeff Fischer Front room, last day
- 24. Data Processing: Mapping to on-off values © 2016, 2017 Jeff Fischer Front room, last day
- 25. Applying “Machine Learning” ¨ Apply a supervised learning to create predictions for the light ¤ Regression => predict light value ¤ Classification => Light “on” or “off” ¤ Features = time of day; time relative to sunrise, sunset; history ¨ Challenges ¤ Transitions more important than individual samples (200 vs. 25,000) ¤ Different class sizes: light is mostly off ¤ Really a random process ¨ Solution: Hidden Markov Models © 2016, 2017 Jeff Fischer
- 26. Hidden Markov Models (HMMs) ¨ Markov process ¤ State machine with probability associated with each outgoing transition ¤ Probabilities determined only by the current state, not on history ¨ Hidden Markov Model ¤ The states are not visible to the observer, only the outputs (“emissions”). ¨ In a machine learning context: ¤ (Sequence of emissions, # states) => inferred HMM ¨ The hmmlearn library will do this for us. ¤ https://github.com/hmmlearn/hmmlearn ¨ But, no way to account for time of day, etc. © 2016, 2017 Jeff Fischer Example Markov process (from Wikipedia)
- 27. Slicing Data into Time-based “Zones” © 2016, 2017 Jeff Fischer Sunrise 30 Minutes before sunset Max(sunset+60m, 9:30 pm) 0 1 2 3 0
- 28. HMM Training and Prediction Process Training 1. Build a list of sample subsequences for each zone 2. Guess a number of states (e.g. 5) 3. For each zone, create an HMM and call fit() with the subsequences Prediction For each zone of a given day: n Run the associated HMM to generate N samples for an N minute zone duration n Associated a computed timestamp with each sample © 2016, 2017 Jeff Fischer
- 29. HMM Predicted Data © 2016, 2017 Jeff Fischer Front room, one week predicted data Front room, one day predicted data
- 31. Lighting Replay Application: Replay Front Room Smart Light Raspberry Pi (Dining Room) HMM definitions Player Script Back Room Smart Light Philips Hue Bridge WiFi Router and Switch ZigBee HTTP © 2016, 2017 Jeff Fischer
- 32. Logic of the Replay Script ¨ Use phue library to control lights ¨ Reuse time zone logic and HMMs from analysis ¨ Pseudo-code: Initial testing of lights while True: compute predicted values for rest of day organize predictions into a time-sorted list of on/off events for each event: sleep until event time send control message for event wait until next day © 2016, 2017 Jeff Fischer https://github.com/mpi-sws-rse/thingflow-examples/blob/master/lighting_replay_app/player/lux_player.py
- 33. Parting Thoughts
- 34. Lessons Learned ¨ End-to-end projects great for learning ¨ Machine learning involves trial-and-error ¨ Visualization is key ¨ Python ecosystem is great for both runtime IoT and offline analytics © 2016, 2017 Jeff Fischer
- 35. Thank You Contact Me Email: jeff@data-ken.org Twitter: @fischer_jeff Website and blog: https://data-ken.org More Information ThingFlow: https://thingflow.io Examples (including lighting replay app): https://github.com/mpi-sws-rse/thingflow-examples Hardware tutorial: http://micropython-iot-hackathon.readthedocs.io/en/latest/