TinyOL-HITL

Tiny Online Learning with Human-in-the-Loop for Industrial Condition Monitoring

One-Liner

Deploy predictive maintenance on Day 1. No training data. No ML expertise. System learns faults as operators discover them.

The Problem

graph LR
    A[27% PdM Adoption] --> B[Expertise Gap]
    A --> C[Vendor Lock-in]
    A --> D[Integration Cost]

    B --> E[TinyOL-HITL]
    C --> E
    D --> E

    E --> F[Deploy Immediately<br/>K=1 → K=N]

Loading

How It Works

stateDiagram-v2
    [*] --> NORMAL: K=1 baseline
    NORMAL --> FROZEN: Outlier detected
    FROZEN --> NORMAL: Operator labels → K++
    FROZEN --> NORMAL: Operator discards

Loading

1. Day 1: Everything = "normal" (K=1)
2. Fault occurs: Device freezes, alarms SCADA
3. Operator inspects: Labels fault type
4. System learns: K=2, correct predictions next time

Quick Start

# 1. Clone
git clone https://github.com/leekaize/tinyol-hitl.git
cd tinyol-hitl

# 2. Configure
cp core/config.template.h core/config.h
# Edit WiFi credentials

# 3. Upload (Arduino IDE)
# Board: ESP32 Dev Module or Raspberry Pi Pico 2 W
# Open: core/core.ino
# Upload

# 4. Wire sensor
# ADXL345/MPU6050 → I2C (SDA=21, SCL=22 for ESP32)

API (9 functions)

// Core
kmeans_init(&model, 3, 0.2f);
int8_t cluster = kmeans_update(&model, features);
kmeans_add_cluster(&model, "fault_name");

// Persistence (NEW)
storage.begin();
storage.load(&model);   // On startup
storage.save(&model);   // After add_cluster
storage.clear();        // On reset command

See docs/API.md for complete reference.

Architecture

flowchart TB
    subgraph Device[MCU - 2.5KB]
        A[Accelerometer] --> B[Feature Extraction]
        B --> C[Streaming K-means]
        C --> D{Outlier?}
        D -->|No| E[Update Centroid]
        D -->|Yes| F[FREEZE]
    end

    subgraph SCADA[FUXA Dashboard]
        G[Alarm Banner]
        H[Label Button]
        I[Gauges]
    end

    F --> G
    H --> F
    E --> I

Loading

Persistent Storage (NEW)

Trained clusters survive power cycles.

flowchart LR
    A[Label Fault] --> B[K++]
    B --> C[💾 Save to Flash]
    C --> D[Power Off]
    D --> E[Power On]
    E --> F[Load from Flash]
    F --> G[Resume K clusters]

Loading

When saves happen: Immediately after each new cluster

Reset model: mosquitto_pub -t "tinyol/{device_id}/reset" -m '{"reset":true}'

Platform storage:

• ESP32: NVS (Preferences library)
• RP2350: LittleFS

Memory Footprint

Component	Size
Clusters (16 max)	1.0 KB
Ring buffer (100 samples)	1.2 KB
Metadata	0.3 KB
Total	2.5 KB

Validation

Dataset	Method
CWRU Bearing	4 fault classes, streaming via Serial
Real Motor	0.5HP 3-phase, eccentric weight faults

TO-DO: Actual benchmark results in progress. See docs/SPRINT_2DAY.md.

Project Structure

core/                    # Arduino sketch
├── core.ino             # Main loop
├── streaming_kmeans.c   # Algorithm (200 lines)
└── config.template.h    # WiFi/MQTT settings

docs/
├── presentation/        # sli.dev slides
├── API.md              # Function reference
└── SPRINT_2DAY.md      # Current sprint

data/datasets/cwru/      # Benchmark dataset pipeline
integrations/            # FUXA SCADA setup

Presentation

cd docs/presentation
npm init slidev@latest
npm run dev
# http://localhost:3030

License

Apache-2.0

Citation

@misc{tinyol-hitl2025,
  author = {Lee, Kai Ze},
  title = {TinyOL-HITL: Unsupervised TinyML Fault Discovery},
  year = {2025},
  publisher = {GitHub},
  url = {https://github.com/leekaize/tinyol-hitl}
}