IoT AI and Machine Learning

IoT systems generate enormous volumes of sensor data continuously. Rules-based systems handle predictable conditions well — "if temperature exceeds 80°C, send an alert" — but they fail when patterns are complex, variable, or too subtle for a human to specify as a rule. Machine learning finds those patterns automatically by training on historical data. Combining IoT with AI turns passive monitoring into intelligent prediction and autonomous decision-making.

Where AI Fits in the IoT Stack

[IoT Sensors] --> [Data Collection] --> [Storage]
                                            |
                                            v
                                    [ML Training]
                                   (uses historical data
                                    to build a model)
                                            |
                                            v
                                    [Trained Model]
                                            |
                          +-----------------+-----------------+
                          |                                   |
                    [Cloud Inference]                  [Edge Inference]
                   (complex models,                  (fast local decisions,
                    batch analysis)                   no cloud needed)
                          |                                   |
                          v                                   v
                    [Alerts, Reports,               [Immediate Actions,
                     Recommendations]                Local Alerts]

Key AI and ML Applications in IoT

Predictive Maintenance

This is the most widely deployed ML application in industrial IoT. A machine learning model trains on months or years of vibration, temperature, and current draw data from motors, pumps, and bearings — including data from periods that ended in equipment failures. The model learns the signature patterns that precede failure.

Once deployed, the model scores incoming sensor data in real time. When it detects a pattern that historically led to failure within two weeks, it triggers a maintenance alert — before any visible symptom exists. Factories using predictive maintenance report 50–70% fewer unexpected breakdowns compared to calendar-based maintenance schedules.

Historical Training Data:
  Vibration: normal --> slightly elevated --> high --> FAILURE
  Temperature: normal --> slightly warm --> overheating --> FAILURE

Trained Model learns: "elevated vibration + slight warmth = failure in 8-14 days"

Live Monitoring:
  Day 1: Vibration = 4.2g (normal)
  Day 4: Vibration = 6.1g (slightly elevated)
  Day 5: Temperature +3°C (slightly warm)
  Day 5: Model predicts FAILURE risk HIGH --> Alert sent --> Maintenance scheduled

Anomaly Detection

Anomaly detection models learn what "normal" looks like for a system and flag anything that deviates significantly. Unlike rules-based alerts (which require knowing the threshold in advance), anomaly detection handles complex multi-variable patterns.

Example: A water pump normally draws 4.2 amps, runs at 1,450 RPM, and produces 12 PSI output. The anomaly detection model flags an event where current draw rises to 5.1 amps while RPM drops to 1,280 — because that combination has never occurred in normal operation and suggests a blockage or impeller damage, even though no individual measurement crossed a fixed threshold.

Computer Vision on IoT Cameras

Cameras paired with computer vision AI perform automated visual inspection tasks that previously required human attention. Convolutional neural networks (CNNs) analyze image data and recognize specific patterns — product defects, intrusion events, parking violations, or human poses.

Real deployments:

Manufacturing quality control: Cameras on production lines detect surface cracks, missing components, label misalignment, and fill level errors at full production speed — hundreds of items per minute — with higher consistency than human inspectors.
Retail analytics: Store cameras with computer vision count footfall, track customer paths through a store, measure queue lengths, and detect when shelves need restocking.
Smart parking: Overhead cameras identify empty and occupied parking spaces in real time and update a public availability map without requiring any in-ground sensors.
Agricultural crop scouting: Drone cameras with image classification models identify pest damage, disease symptoms, and weed infestations in field crops at the early stages when treatment is most effective.

Natural Language Processing (NLP) in IoT

Voice-controlled IoT devices — smart speakers, voice-activated industrial controls, healthcare voice assistants — use NLP models to convert spoken language into commands. Modern on-device NLP chips (like Apple's Neural Engine or Amazon's AZ2) process voice commands locally without sending audio to the cloud, protecting privacy.

Time-Series Forecasting

ML models trained on IoT sensor data predict future values based on historical patterns. This enables proactive action rather than reactive response.

Applications:

Energy demand forecasting: predict electricity demand for the next 24 hours so the grid operator adjusts generation capacity in advance
Inventory prediction: forecast when parts or raw materials will run out based on consumption rate trends so purchase orders arrive before stockouts occur
Crop yield prediction: use soil, weather, and crop health sensor data to forecast harvest volume weeks in advance so logistics can be arranged

Reinforcement Learning for Autonomous Control

Reinforcement learning (RL) trains a control agent by letting it interact with an environment and learn which actions produce the best outcomes over time. In IoT, this powers systems that continuously optimize themselves.

Example: Google's DeepMind used reinforcement learning to control the cooling systems of Google's data centers. The RL agent receives real-time sensor readings from thousands of IoT sensors tracking temperature, airflow, and power. It continuously adjusts cooling parameters to minimize energy use while staying within safe temperature limits. The result was a 40% reduction in cooling energy — far beyond what human-designed rules achieved.

TinyML: Running AI on Microcontrollers

Traditional machine learning models require substantial computing power to run inference (make predictions). Running a model on a cloud server is straightforward. Running one on a microcontroller with 256 KB of RAM and no operating system was considered impossible until recently.

TinyML is the field of optimizing machine learning models to run on resource-constrained devices. Techniques used include:

Model quantization: Reduce the precision of model weights from 32-bit floats to 8-bit integers, shrinking model size by 4× with minimal accuracy loss
Pruning: Remove neural network connections that contribute little to the output, making the model smaller and faster
Knowledge distillation: Train a large "teacher" model, then train a smaller "student" model to mimic it — compressing capability into a fraction of the size

Frameworks enabling TinyML:

TensorFlow Lite for Microcontrollers: Google's framework for running quantized TensorFlow models on microcontrollers like Arduino and ESP32
Edge Impulse: A no-code/low-code platform for training, optimizing, and deploying ML models directly to edge devices — including microcontrollers, Raspberry Pi, and Android
ONNX Runtime: Runs models from any framework on edge hardware with hardware-specific optimizations

What TinyML enables on microcontrollers today:

Keyword detection (wake words like "Hey Alexa") on a coin-cell battery
Gesture and activity recognition from accelerometer data
Anomaly detection for machine health monitoring without any cloud connection
Image classification on cameras that cost under $10

Federated Learning: Training Without Sharing Data

Standard ML training requires collecting all sensor data in one place — a cloud server — to train the model. For sensitive applications (healthcare, private homes, industrial secrets), this creates privacy and security concerns.

Federated learning trains the model differently. Each device trains a local version of the model on its own data. Only the model's weight updates — not the raw data — are sent to the cloud. The cloud aggregates the weight updates from thousands of devices to improve the global model. The raw sensor data never leaves the device.

STANDARD ML TRAINING:
Device A --[raw data]--> Cloud trains model
Device B --[raw data]-->

FEDERATED LEARNING:
Device A trains locally --> sends [model update only] --> Cloud
Device B trains locally --> sends [model update only] --> Cloud
Cloud merges updates into improved global model
Raw data stays on each device

AI + IoT in Practice: Smart Building Energy Optimization

IoT Sensors:
  - Occupancy sensors (each room)
  - Temperature sensors (each zone)
  - External weather station
  - Energy meters (per floor)
  - CO2 sensors (air quality)

ML Models:
  - Occupancy prediction model: forecasts when each room will be occupied
    based on calendar data and historical patterns
  - Thermal model: predicts how quickly each zone heats or cools
    given current weather and HVAC settings
  - Optimization model: sets HVAC schedules that pre-condition spaces
    30 minutes before predicted occupancy using the least energy

Result:
  Building reaches comfort temperature exactly when needed
  No energy wasted heating empty rooms
  15-25% energy reduction vs. fixed schedule thermostat

Summary

AI and machine learning transform IoT from a passive data collection system into an intelligent one that predicts, detects, and decides. Predictive maintenance prevents equipment failures. Anomaly detection surfaces problems no fixed rule would catch. Computer vision automates visual inspection. TinyML brings AI inference to microcontrollers with no cloud dependency. Federated learning enables model training across sensitive data without centralizing that data. Together, AI and IoT shift industrial and consumer systems from reactive to proactive operation.

Previous lesson

Back to course

Next lesson