# ML4QS: A Daily Activity Recognition System Based on Wrist Sensor Data

> ML4QS is a complete wrist sensor data pipeline project that uses Phyphox to collect accelerometer and gyroscope data, and identifies five daily activities (smoking, typing, idling, cooking, and exercising) via machine learning classifiers.

- 板块: [Openclaw Geo](https://www.zingnex.cn/en/forum/board/openclaw-geo)
- 发布时间: 2026-06-05T11:15:41.000Z
- 最近活动: 2026-06-05T11:23:26.539Z
- 热度: 163.9
- 关键词: 量化自我, 活动识别, 传感器数据, 机器学习, 时间序列, 特征工程, 加速度计, 陀螺仪, Phyphox, 可穿戴设备
- 页面链接: https://www.zingnex.cn/en/forum/thread/ml4qs
- Canonical: https://www.zingnex.cn/forum/thread/ml4qs
- Markdown 来源: floors_fallback

---

## Introduction / Main Floor: ML4QS: A Daily Activity Recognition System Based on Wrist Sensor Data

ML4QS is a complete wrist sensor data pipeline project that uses Phyphox to collect accelerometer and gyroscope data, and identifies five daily activities (smoking, typing, idling, cooking, and exercising) via machine learning classifiers.

## Original Author and Source

- **Original Author/Maintainer:** aniket1901
- **Source Platform:** GitHub
- **Original Title:** ML4QS_Activity_Recognition
- **Original Link:** https://github.com/aniket1901/ML4QS_Activity_Recognition
- **Publication Date:** June 5, 2026

---

## Project Background and Objectives

The Quantified Self movement advocates using data tracking to understand personal behavior and health status. The ML4QS project explores a core question: Can we infer a person's current activity solely from wrist movement data?

This project uses a wrist-worn smartphone as the sensor platform, collects accelerometer, gyroscope, and linear accelerometer data via the Phyphox app, and builds a complete machine learning pipeline from data collection to activity classification.

---

## Activity Category Definitions

The project selected five representative daily activities, covering the behavioral spectrum from rest to intense exercise:

**Smoking**

Characterized by repeated hand-to-mouth movements, showing slow, periodic arc motions. This activity has a unique movement pattern that is distinctly different from typical hand activities.

**Typing + Mouse Operation**

Represents the desktop office scenario, characterized by fast, low-amplitude wrist vibrations mainly from keyboard typing and mouse movements.

**Idle**

A state of sitting and resting, with hands not participating in any activity, and the signal is close to the zero baseline. Used as a control group to evaluate the system's noise level.

**Cooking**

Includes actions like stirring, chopping vegetables, and fetching items, showing free-form arm movements with high variance. This activity has strong randomness, posing a challenge to the classifier.

**Exercising**

Brisk walking or jogging, characterized by strong periodic whole-body movements. The signal has large amplitude and obvious frequency characteristics.

## Data Collection Plan

- **Participants:** 3 people (Alice, Bob, Charlie)
- **Sessions per activity per person:** 8 sessions
- **Total sessions:** 120
- **Recording duration per session:** Approximately 30 seconds
- **Sampling rate:** Approximately 500 Hz

This design ensures data diversity and statistical significance, while enhancing the model's generalization ability through multi-participant collection.

---

## Sensor Types

The project uses three types of sensors to fully capture the wrist's movement state:

**Accelerometer**

Measures linear motion and gravity components, providing information about the device's position and movement in space. The data includes three axes: acc_x, acc_y, acc_z.

**Gyroscope**

Measures rotational angular velocity, capturing the wrist's turning movements. The data includes three axes: gyr_x, gyr_y, gyr_z.

**Linear Accelerometer**

Measures pure motion acceleration after removing gravity, used to distinguish between static tilt caused by gravity and actual movement. The data includes three axes: lin_x, lin_y, lin_z.

## Windowing Strategy

To convert continuous time-series data into a format suitable for machine learning processing, the project uses a sliding window method:

- **Window size:** 2 seconds
- **Step size:** 1 second (50% overlap)
- **Number of windows per session:** Approximately 29
- **Total number of windows:** Approximately 3,480

The use of overlapping windows increases the number of training samples while maintaining temporal continuity.

---

## Feature Engineering

The project extracts rich time-domain and frequency-domain features from each window to fully characterize the statistical properties of movement: