Attention

This project use features selecting method to improve the accuracy of an exist HAR CNN(Reference https://github.com/healthDataScience/deep-learning-HAR)

And the result as bellowings with extends features combination are acc_x_fft、acc_y_fft and acc_z_fft

Method	Test accuracy	Test accuracy(extend_features + tanh)
CNN	93%	94.9583%

We must mentioned here, the tanh activation play a key role in improving the accuracy. But the features extends is very useful in some situation like that you only accelerator meters. Just try it.

training loss
training accuracy

Human Activity Recognition (HAR)

In this part of the repo, we discuss the human activity recognition problem using deep learning algorithms and compare the results with standard machine learning algorithms that use engineered features.

The data can be downloaded from the UCI repository.

Also see G.Chevalier's repo and A. Saeed's blog where I have got lots of inspiration.

Contents

The utils folder contains the code for reading and processing the data into a tensor form. The generated tensors have the dimensions

(batch, seq_len, n_channels)

where batch is the number of training examples in each batch, seq_len is the number of steps in the time series (128) and n_channels is the number of channels where observations are made (9).

The aim is to clasify the activities correctly, which are

1 WALKING
2 WALKING_UPSTAIRS
3 WALKING_DOWNSTAIRS
4 SITTING
5 STANDING
6 LAYING

Below are the architectures used for training

Notebook	Description
explore_data	Data exploration
HAR-LSTM	LSTM network
HAR-CNN	Convolutional neural network(CNN)
HAR-CNN-LSTM	CNN + LSTM hybrid
HAR-CNN-Inception	CNN with inception module

Results

Method	Test accuracy
CNN	93%
LSTM	88%
CNN+LSTM	88%
CNN+Inception	89%
Xgboost	96%

CNN architecture

LSTM architecture