Machine Learning techniques for time series classification

This repository shows different approaches to time series classification using machine learning techniques. In addition to comparing some of the most used machine learning models (Deep learning, Neural network, Support Vector Machine ...), we oppose two different types of input for some of these models: using the raw time series or using features extracted from the time series (statistical measures, frequency domain features, geometrical features... features detailed below).

In this repository we will use python's scientific package NumPy, some scikit-learn features and TensorFlow for convolutional neural networks and neural networks.

Machine Learning Approaches in this repository

The 6 different approaches in this repository are:

• Convolutional Neural Network with raw data

• k Nearest Neighbors (kNN) with raw data

• k Nearest Neighbors with features

• Neural Network with raw data

• Neural Network with features

• Support Vector Machine (SVM) with features

Time series data: Human Activity Recognition (HAR data)

The data set we use in this repository is a Human Activity Recognition database available for free here on UCI Machine Learning Repository. This data set consists in 3-axial accelerometer and gyroscope signals captured with a smartphone worn on the waist while performing 6 different activities:

• Walking
• Walking upstairs
• Walking downstairs
• Sitting
• Standing
• Laying

Each sample in the dataset is a 2.56s window sampled at a 50Hz rate which makes 6 x 128 readings per sample (3 accelerometer axis x,y,z and 3 gyroscope axis x,y,z). In this repository we will only use the 3 accelerometer axis x, y and z.

Below we plot examples of the time series for each activity:

The classification task here consists in recognizing the 6 activities given above. As explained in the introduction, we will either use the raw signals (plots above) or features extracted from these signals.

Features extraction

In some of the notebooks in this repository we will extract features from the time series and use them as input for the Machine Learning models. In these notebooks we will build helper functions to extract the desired features. We extract statistical and geometrical features from raw signals and jerk signals (acceleration first derivative), frequency domain features from raw signals and jerk signals. The full list of features is:

• x,y and z raw signals : mean, max, min, standard deviation, skewness, kurtosis, interquartile range, median absolute deviation, area under curve, area under squared curve
• x,y and z jerk signals (first derivative) : mean, max, min, standard deviation, skewness, kurtosis, interquartile range, median absolute deviation, area under curve, area under squared curve
• x,y and z raw signals Discrete Fourrier Transform: mean, max, min, standard deviation, skewness, kurtosis, interquartile range, median absolute deviation, area under curve, area under squared curve, weighted mean frequency, 5 first DFT coefficients, 5 first local maxima of DFT coefficients and their corresponding frequencies.
• x,y and z jerk signals Discrete Fourrier Transform: mean, max, min, standard deviation, skewness, kurtosis, interquartile range, median absolute deviation, area under curve, area under squared curve, weighted mean frequency, 5 first DFT coefficients, 5 first local maxima of DFT coefficients and their corresponding frequencies.
• x,y and z correlation coefficients