AI & ML for IoT Systems

1

Why AI + IoT Now

2

IoT Stack Overview

3

From Raw Signals to Usable Data

4

Control and Actuation Loops

5

Feature Engineering for Sensor Data

• Windowing (fixed, sliding, adaptive)
• Stats: mean, variance, kurtosis
• Frequency domain: FFT, spectrograms
• Domain features: vibration signatures

6

Why ML in IoT?

• Predictive maintenance
• Anomaly detection in industrial systems
• Activity / gesture recognition
• Occupancy-aware energy optimization
• Autonomous navigation / robotics

7

Edge vs Cloud vs Hybrid

8

What is Tiny Machine Learning (TinyML)?

9

TinyML

Fastest-growing field of ML

Algorithms, hardware, software

Low power consumption

On-device sensor analytics

Always-on ML

Battery-operated

Why Edge Intelligence?

• Privacy: data stays local
• Low latency: sub-100 ms actuation
• Reliability offline / flaky network
• Bandwidth savings: send insights, not raw video

10

Edge Hardware Landscape

11

TinyML Concepts

12

Reference AIoT Architecture

13

EdgeML (P )

14

TinyML (P )

KeyWord Spotting

Motion & biometric

Environmental Control

Image Spot

Image Recognition

Autonomous Car Control

15

16

EdgeML

17

46

Rpi-Pico (Cortex-M0+)

Arduino Nano (Cortex-M4)

Arduino Pro (Cortex-M7)

RaspberryPi SmartPhone (Cortex-A)

Jetson Nano (Cortex-A + GPU)

Object Detection Complex Voice Processing

1 MB+

Image

Classification 250 KB+

KeyWord Spotting Audio Classification 50 KB

Anomaly Detection Sensor Classification 20 KB

Video Classification 2 MB+

Source: Edge Impulse

TinyML

Hardware

Software

18

(TFL Micro)

Example: KeyWord Spotting (KWS)

19

https://mjrobot.org/2021/01/27/building-an-intelligent-voice-assistant-from-scratch/

20

Sound

Image

KeyWord Spotting (KWS) - Inference

21

Digital Mic

“Yes”

Obtains an input

16KHz / 16 bits Sample: [1s]

KeyWord Spotting (KWS) - Inference

22

Digital Mic

“Yes”

MFCC

Feature Converter

Obtains an input

Pre-Process

16KHz / 16 bits Sample: [1s]

Output: Image [49, 40, 1]

KeyWord Spotting (KWS) - Inference

23

Digital Mic

“Yes”

MFCC

Feature Converter

Obtains an input

Pre-Process

Runs model

16KHz / 16 bits Sample: [1s]

Output: Image [49, 40, 1]

Output Dim [1, 4]

• Prob ‘Silence’
• Prob ‘Unknown’
• Prob ‘Yes’
• Prob ‘No’

KeyWord Spotting (KWS) - Inference

24

Digital Mic

“Yes”

MFCC

Feature Converter

Obtains an input

Pre-Process

Runs model

Post-Processes

16KHz / 16 bits Sample: [1s]

Output: Image [49, 40, 1]

Output Dim [1, 4]

• Prob ‘Silence’
• Prob ‘Unknown’
• Prob ‘Yes’
• Prob ‘No’

If Probability of YES is greater than 80% Take actions

KeyWord Spotting (KWS) - Inference

25

Digital Mic

“Yes”

MFCC

Feature Converter

Obtains an input

Pre-Process

Runs model

Post-Processes

Make things happen

16KHz / 16 bits Sample: [1s]

Output: Image [49, 40, 1]

Output Dim [1, 4]

• Prob ‘Silence’
• Prob ‘Unknown’
• Prob ‘Yes’
• Prob ‘No’

If Probability of YES is greater than 80% Take actions

KeyWord Spotting (KWS) - Model

26

Digital Mic

“Yes

”

MFCC

Feature Converter

Obtains an input

Pre-Process

Runs model

postprocesses

Make things happen

16KHz / 16

bits

Sample: [1s]

Output:

Image

[49, 40, 1]

Output Dim [1, 4]

• Prob ‘Silence’
• Prob ‘Unknown’
• Prob ‘Yes’
• Prob ‘No’

If Probability of YES is greater than 80%

Take actions

KeyWord Spotting (KWS) – Create Model (Training)

27

Digital Mic

MFCC

Feature Converter

Obtains data

Pre-Process

16KHz / 16 bits Sample: [1s]

Output: Image [49, 40, 1]

KeyWord Spotting (KWS) – Create Model (Training)

28

Digital Mic

MFCC

Feature Converter

Obtains data

Pre-Process

Train model

Evaluate Model

16KHz / 16 bits Sample: [1s]

Output: Image [49, 40, 1]

KeyWord Spotting (KWS) – Create Model (Training)

29

Digital Mic

MFCC

Feature Converter

Obtains data

Pre-Process

Train model

Evaluate Model

Deploy

16KHz / 16 bits Sample: [1s]

Output: Image [49, 40, 1]

Supervised Learning Refresher

• Regression vs classification
• Temporal train/val/test splits matter
• Labeling is expensive and noisy
• Concept drift over time

30

Supervised Learning

• Majority of practical ML uses supervised learning
• Mapping function approximated from past experience
• Regression f(x)=y, y is a real number
• Classification f(x)=y, y is a category label
• Training
• Labeled positive and negative examples
• From unseen input predict corresponding output
• Learning until acceptable performance is achieved

31

Unsupervised & Self-Supervised IoT

32

Unsupervised Learning

• Discover hidden relations and learn about the data
• Clustering f(X) = [X₁, …, X_k], k disjoint subsets
• Association f(X_i, X_j) = R, relation
• Training
• All examples are positive
• No labeling, no teacher
• No single correct answer
• Practical usage
• Derive groups, not explicitly labeled
• Market basket analysis (association among items)

33

Time Series Modeling Approaches

34

Activity Recognition Example

• Input: accelerometer / gyroscope
• Sliding window segmentation
• Feature extraction vs end-to-end deep model
• Output: walk / run / fall / idle
• Use case: elder fall detection

35

Predictive Maintenance Example

• Vibration + temperature from motors
• Estimate Remaining Useful Life (RUL)
• Early fault signatures are subtle
• Class imbalance: failures are ~0.1%
• Metric: precision at very low false alarms

36

Anomaly Detection in IoT

• Why anomalies matter: safety, cost, security
• Thresholding simple stats
• Isolation Forest / density methods
• Autoencoder reconstruction error
• Online streaming anomaly scores

37

Computer Vision in IoT

• Smart cameras: traffic, quality inspection
• Object detection, pose estimation
• PPE compliance / safety zones
• Privacy: blur faces, process on-device

38

Multimodal Sensor Fusion

• Fusing IMU + camera + GPS + audio
• Early fusion vs late fusion
• Attention-based fusion architectures
• Robustness if one sensor drops

39

ML Decision Tree

40

Diagnose Steps (part 1)

1. Is it a ML task? Are you sure ML is the best solution?

• Hard: X is independent of Y: X<name, age, income>, Y = height
• Easy: X is a set with limited variations. Configure Y = F(X)

2. Appropriate ML scenario?

• Supervised learningclassification, regression, anomaly detection)
• Unsupervised learning (clustering, pattern learning)

3. Appropriate model?

• Data size (small data -> linear model, large data -> consider nonlinear)
• Sparse data (require normalization to perform better)
• Imbalanced data (special treatment of the minority class required)
• Data quality (noise and missing values require loss function)

4. Enough training data?

• Investigate how precision improves with more data

41

Diagnose Steps (part 2)

5. Model overly complicated?

• Start simple first, increase complexity and evaluate performance
• Avoid overfitting to training set

6. Feature quality

• Have you identified all useful features?
• Use domain knowledge of an expert to start
• Include any feature that could be found and investigate model performance

7. Feature engineering

• The best strategy to improve performance and reveal important input
• Encode features, normalize [0:1], combine features

8. Combine models

• If multiple models have similar performance there is a chance of improvement
• Use one model for one subset of data and another model for the other

42

Diagnose Steps (part 3)

9. Model validation

• Use appropriate performance indicator (accuracy, precision, recall, F1, etc.)
• How well does the model describe data? (AUC)
• Data typically divide into Training and Validation
• Evaluate accuracy on disjoint dataset (other than training dataset)
• Tune model hyper parameters (i.e. number of iterations)

43

Types of Algorithms

• Linear Algorithms
• Classification – classes separated by straight line
• Support Vector Machine – wide gap instead of line
• Regression – linear relation between variables and label
• Non-Linear Algorithms
• Decision Trees and Jungles – divide space into regions
• Neural Networks – complex and irregular boundaries
• Special Algorithms
• Ordinal Regression – ranked values (i.e. race)
• Poisson – discrete distribution (i.e. count of events)
• Bayesian – normal distribution of errors (bell curve)

​

44

Model Performance (Classification)

• Binary classification outcomes {positive; negative}
• ROC curve
• TP Rate = True Positives / All Positives
• FP Rate = False Positives / All Negatives
• Example:

​

​

• AUC (Area Under Curve)
• KPI for model performance and comparison
• 0.5 = random prediction, 1 = perfect match
• For Multiclass – average from all RoC curves

45

Threshold Selection (Binary)

• Probability Threshold
• Cost of one error could be much higher than cost of other
• E.g. spam filter – it is more expensive to miss a real mail
• Accuracy
• For symmetric 50/50 data
• Precision
• E.g. 1000 devices, 5 fails, 8 predicted, 4 true failures
• Correct positives (e.g. 4/8 = 0.5, FP are expensive)
• Recall
• Correctly predicted positives (e.g. 4/5=0.8, FN are expensive)
• F1(balanced error cost)
• Balanced cost of Precision/Recall

46

Model Performance (Regression)

• Coefficient of Determination (R²)
• Single numeric KPI – how well data fits model
• R² > 0.6 – good, R² > 0.8 – very good, R² = 1 – perfect
• Mean Absoute Error (MAE) / Root Mean Squared Error
• Deviation of estimates from observed real values
• Compare model errors measure in the SAME units
• Relative Absolute Error Relative Squared Error
• % deviation from real value
• Compare model errors measure in the DIFFERENT units

​

47

Evaluation Metrics for IoT ML

• Latency (ms)
• Power draw (mW)
• Model size / memory footprint
• False alarm vs miss cost
• Closed-loop control stability

48

Model Compression Techniques

• Quantization (8-bit, 4-bit)
• Pruning weights / neurons
• Knowledge distillation
• Neural architecture search for low FLOPs

49

On-Device Inference Pipeline

• Capture → preprocess → inference → decision
• Memory budgeting for buffers + weights
• Real-time scheduling with firmware tasks
• Interrupt-driven ML (e.g. wake word)

50

Federated Learning for IoT

• Devices train locally on private data
• Only model updates are shared
• Benefits: personalization + privacy
• Challenges: non-IID data, stragglers

51

Digital Twins

• Virtual replica of a physical asset
• Live sync with sensor data
• Simulate 'what if' scenarios
• Uses: predictive maintenance, optimization

52

Privacy & Responsible AI in IoT

• Always-on sensing = surveillance risk
• Behavior inference from 'harmless' signals
• Data minimization and edge filtering
• Transparency + consent for end users

53

Regulatory & Compliance

• Safety standards for industrial monitoring
• Data protection (GDPR-style thinking)
• AI accountability + audit trails
• Explainability for automated decisions

54

LLMs + IoT

• Natural language interfaces ('Why did the alarm trigger?')
• Edge copilots for technicians
• Summarizing fleets of sensor logs
• Risk: hallucination in safety-critical loops

55

Energy-Aware Intelligence

56

Edge Swarms & Collective Intelligence

57

Where This Field Is Going

58