How to Develop RNN Models for Human Activity Recognition Time Series Classification

Human activity recognition is the problem of classifying sequences of accelerometer data recorded by specialized harnesses or smart phones into known well-defined movements.

Classical approaches to the problem involve hand crafting features from the time series data based on fixed-sized windows and training machine learning models, such as ensembles of decision trees. The difficulty is that this feature engineering requires strong expertise in the field.

Recently, deep learning methods such as recurrent neural networks like as LSTMs and variations that make use of one-dimensional convolutional neural networks or CNNs have been shown to provide state-of-the-art results on challenging activity recognition tasks with little or no data feature engineering, instead using feature learning on raw data.

In this tutorial, you will discover three recurrent neural network architectures for modeling an activity recognition time series classification problem.

After completing this tutorial, you will know:

  • How to develop a Long Short-Term Memory Recurrent Neural Network for human activity recognition.
  • How to develop a one-dimensional Convolutional Neural Network LSTM, or CNN-LSTM, model.
  • How to develop a one-dimensional Convolutional LSTM, or ConvLSTM, model for the same problem.

Let’s get started.

How to Develop RNN Models for Human Activity Recognition Time Series Classification

How to Develop RNN Models for Human Activity Recognition Time Series Classification
Photo by Bonnie Moreland, some rights reserved.

Tutorial Overview

This tutorial is divided into four parts; they are:

  1. Activity Recognition Using Smartphones Dataset
  2. Develop an LSTM Network Model
  3. Develop a CNN-LSTM Network Model
  4. Develop a ConvLSTM Network Model

Activity Recognition Using Smartphones Dataset

Human Activity Recognition, or HAR for short, is the problem of predicting what a person is doing based on a trace of their movement using sensors.

A standard human activity recognition dataset is the ‘Activity Recognition Using Smart Phones Dataset’ made available in 2012.

It was prepared and made available by Davide Anguita, et al. from the University of Genova, Italy and is described in full in their 2013 paper “A Public Domain Dataset for Human Activity Recognition Using Smartphones.” The dataset was modeled with machine learning algorithms in their 2012 paper titled “Human Activity Recognition on Smartphones using a Multiclass Hardware-Friendly Support Vector Machine.”

The dataset was made available and can be downloaded for free from the UCI Machine Learning Repository:

The data was collected from 30 subjects aged between 19 and 48 years old performing one of six standard activities while wearing a waist-mounted smartphone that recorded the movement data. Video was recorded of each subject performing the activities and the movement data was labeled manually from these videos.

Below is an example video of a subject performing the activities while their movement data is being recorded.

The six activities performed were as follows:

  1. Walking
  2. Walking Upstairs
  3. Walking Downstairs
  4. Sitting
  5. Standing
  6. Laying

The movement data recorded was the x, y, and z accelerometer data (linear acceleration) and gyroscopic data (angular velocity) from the smart phone, specifically a Samsung Galaxy S II. Observations were recorded at 50 Hz (i.e. 50 data points per second). Each subject performed the sequence of activities twice; once with the device on their left-hand-side and once with the device on their right-hand side.

The raw data is not available. Instead, a pre-processed version of the dataset was made available. The pre-processing steps included:

  • Pre-processing accelerometer and gyroscope using noise filters.
  • Splitting data into fixed windows of 2.56 seconds (128 data points) with 50% overlap.Splitting of accelerometer data into gravitational (total) and body motion components.

Feature engineering was applied to the window data, and a copy of the data with these engineered features was made available.

A number of time and frequency features commonly used in the field of human activity recognition were extracted from each window. The result was a 561 element vector of features.

The dataset was split into train (70%) and test (30%) sets based on data for subjects, e.g. 21 subjects for train and nine for test.

Experiment results with a support vector machine intended for use on a smartphone (e.g. fixed-point arithmetic) resulted in a predictive accuracy of 89% on the test dataset, achieving similar results as an unmodified SVM implementation.

The dataset is freely available and can be downloaded from the UCI Machine Learning repository.

The data is provided as a single zip file that is about 58 megabytes in size. The direct link for this download is below:

Download the dataset and unzip all files into a new directory in your current working directory named “HARDataset”.

Need help with Deep Learning for Time Series?

Take my free 7-day email crash course now (with sample code).

Click to sign-up and also get a free PDF Ebook version of the course.

Download Your FREE Mini-Course

Develop an LSTM Network Model

In this section, we will develop a Long Short-Term Memory network model (LSTM) for the human activity recognition dataset.

LSTM network models are a type of recurrent neural network that are able to learn and remember over long sequences of input data. They are intended for use with data that is comprised of long sequences of data, up to 200 to 400 time steps. They may be a good fit for this problem.

The model can support multiple parallel sequences of input data, such as each axis of the accelerometer and gyroscope data. The model learns to extract features from sequences of observations and how to map the internal features to different activity types.

The benefit of using LSTMs for sequence classification is that they can learn from the raw time series data directly, and in turn do not require domain expertise to manually engineer input features. The model can learn an internal representation of the time series data and ideally achieve comparable performance to models fit on a version of the dataset with engineered features.

This section is divided into four parts; they are:

  1. Load Data
  2. Fit and Evaluate Model
  3. Summarize Results
  4. Complete Example

Load Data

The first step is to load the raw dataset into memory.

There are three main signal types in the raw data: total acceleration, body acceleration, and body gyroscope. Each has 3 axises of data. This means that there are a total of nine variables for each time step.

Further, each series of data has been partitioned into overlapping windows of 2.56 seconds of data, or 128 time steps. These windows of data correspond to the windows of engineered features (rows) in the previous section.

This means that one row of data has (128 * 9), or 1,152 elements. This is a little less than double the size of the 561 element vectors in the previous section and it is likely that there is some redundant data.

The signals are stored in the /Inertial Signals/ directory under the train and test subdirectories. Each axis of each signal is stored in a separate file, meaning that each of the train and test datasets have nine input files to load and one output file to load. We can batch the loading of these files into groups given the consistent directory structures and file naming conventions.

The input data is in CSV format where columns are separated by whitespace. Each of these files can be loaded as a NumPy array. The load_file() function below loads a dataset given the fill path to the file and returns the loaded data as a NumPy array.

We can then load all data for a given group (train or test) into a single three-dimensional NumPy array, where the dimensions of the array are [samples, time steps, features].

To make this clearer, there are 128 time steps and nine features, where the number of samples is the number of rows in any given raw signal data file.

The load_group() function below implements this behavior. The dstack() NumPy function allows us to stack each of the loaded 3D arrays into a single 3D array where the variables are separated on the third dimension (features).

We can use this function to load all input signal data for a given group, such as train or test.

The load_dataset_group() function below loads all input signal data and the output data for a single group using the consistent naming conventions between the directories.

Finally, we can load each of the train and test datasets.

The output data is defined as an integer for the class number. We must one hot encode these class integers so that the data is suitable for fitting a neural network multi-class classification model. We can do this by calling the to_categorical() Keras function.

The load_dataset() function below implements this behavior and returns the train and test X and y elements ready for fitting and evaluating the defined models.

Fit and Evaluate Model

Now that we have the data loaded into memory ready for modeling, we can define, fit, and evaluate an LSTM model.

We can define a function named evaluate_model() that takes the train and test dataset, fits a model on the training dataset, evaluates it on the test dataset, and returns an estimate of the model’s performance.

First, we must define the LSTM model using the Keras deep learning library. The model requires a three-dimensional input with [samples, time steps, features].

This is exactly how we have loaded the data, where one sample is one window of the time series data, each window has 128 time steps, and a time step has nine variables or features.

The output for the model will be a six-element vector containing the probability of a given window belonging to each of the six activity types.

Thees input and output dimensions are required when fitting the model, and we can extract them from the provided training dataset.

The model is defined as a Sequential Keras model, for simplicity.

We will define the model as having a single LSTM hidden layer. This is followed by a dropout layer intended to reduce overfitting of the model to the training data. Finally, a dense fully connected layer is used to interpret the features extracted by the LSTM hidden layer, before a final output layer is used to make predictions.

The efficient Adam version of stochastic gradient descent will be used to optimize the network, and the categorical cross entropy loss function will be used given that we are learning a multi-class classification problem.

The definition of the model is listed below.

The model is fit for a fixed number of epochs, in this case 15, and a batch size of 64 samples will be used, where 64 windows of data will be exposed to the model before the weights of the model are updated.

Once the model is fit, it is evaluated on the test dataset and the accuracy of the fit model on the test dataset is returned.

Note, it is common to not shuffle sequence data when fitting an LSTM. Here we do shuffle the windows of input data during training (the default). In this problem, we are interested in harnessing the LSTMs ability to learn and extract features across the time steps in a window, not across windows.

The complete evaluate_model() function is listed below.

There is nothing special about the network structure or chosen hyperparameters, they are just a starting point for this problem.

Summarize Results

We cannot judge the skill of the model from a single evaluation.

The reason for this is that neural networks are stochastic, meaning that a different specific model will result when training the same model configuration on the same data.

This is a feature of the network in that it gives the model its adaptive ability, but requires a slightly more complicated evaluation of the model.

We will repeat the evaluation of the model multiple times, then summarize the performance of the model across each of those runs. For example, we can call evaluate_model() a total of 10 times. This will result in a population of model evaluation scores that must be summarized.

We can summarize the sample of scores by calculating and reporting the mean and standard deviation of the performance. The mean gives the average accuracy of the model on the dataset, whereas the standard deviation gives the average variance of the accuracy from the mean.

The function summarize_results() below summarizes the results of a run.

We can bundle up the repeated evaluation, gathering of results, and summarization of results into a main function for the experiment, called run_experiment(), listed below.

By default, the model is evaluated 10 times before the performance of the model is reported.

Complete Example

Now that we have all of the pieces, we can tie them together into a worked example.

The complete code listing is provided below.

Running the example first prints the shape of the loaded dataset, then the shape of the train and test sets and the input and output elements. This confirms the number of samples, time steps, and variables, as well as the number of classes.

Next, models are created and evaluated and a debug message is printed for each.

Finally, the sample of scores is printed, followed by the mean and standard deviation. We can see that the model performed well, achieving a classification accuracy of about 89.7% trained on the raw dataset, with a standard deviation of about 1.3.

This is a good result, considering that the original paper published a result of 89%, trained on the dataset with heavy domain-specific feature engineering, not the raw dataset.

Note: given the stochastic nature of the algorithm, your specific results may vary. If so, try running the code a few times.

Now that we have seen how to develop an LSTM model for time series classification, let’s look at how we can develop a more sophisticated CNN LSTM model.

Develop a CNN-LSTM Network Model

The CNN LSTM architecture involves using Convolutional Neural Network (CNN) layers for feature extraction on input data combined with LSTMs to support sequence prediction.

CNN LSTMs were developed for visual time series prediction problems and the application of generating textual descriptions from sequences of images (e.g. videos). Specifically, the problems of:

  • Activity Recognition: Generating a textual description of an activity demonstrated in a sequence of images.
  • Image Description: Generating a textual description of a single image.
  • Video Description: Generating a textual description of a sequence of images.

You can learn more about the CNN LSTM architecture in the post:

To learn more about the consequences of combining these models, see the paper:

The CNN LSTM model will read subsequences of the main sequence in as blocks, extract features from each block, then allow the LSTM to interpret the features extracted from each block.

One approach to implementing this model is to split each window of 128 time steps into subsequences for the CNN model to process. For example, the 128 time steps in each window can be split into four subsequences of 32 time steps.

We can then define a CNN model that expects to read in sequences with a length of 32 time steps and nine features.

The entire CNN model can be wrapped in a TimeDistributed layer to allow the same CNN model to read in each of the four subsequences in the window. The extracted features are then flattened and provided to the LSTM model to read, extracting its own features before a final mapping to an activity is made.

It is common to use two consecutive CNN layers followed by dropout and a max pooling layer, and that is the simple structure used in the CNN LSTM model here.

The updated evaluate_model() is listed below.

We can evaluate this model as we did the straight LSTM model in the previous section.

The complete code listing is provided below.

Running the example summarizes the model performance for each of the 10 runs before a final summary of the models performance on the test set is reported.

We can see that the model achieved a performance of about 90.6% with a standard deviation of about 1%.

Note: given the stochastic nature of the algorithm, your specific results may vary. If so, try running the code a few times.

Develop a ConvLSTM Network Model

A further extension of the CNN LSTM idea is to perform the convolutions of the CNN (e.g. how the CNN reads the input sequence data) as part of the LSTM.

This combination is called a Convolutional LSTM, or ConvLSTM for short, and like the CNN LSTM is also used for spatio-temporal data.

Unlike an LSTM that reads the data in directly in order to calculate internal state and state transitions, and unlike the CNN LSTM that is interpreting the output from CNN models, the ConvLSTM is using convolutions directly as part of reading input into the LSTM units themselves.

For more information for how the equations for the ConvLSTM are calculated within the LSTM unit, see the paper:

The Keras library provides the ConvLSTM2D class that supports the ConvLSTM model for 2D data. It can be configured for 1D multivariate time series classification.

The ConvLSTM2D class, by default, expects input data to have the shape:

Where each time step of data is defined as an image of (rows * columns) data points.

In the previous section, we divided a given window of data (128 time steps) into four subsequences of 32 time steps. We can use this same subsequence approach in defining the ConvLSTM2D input where the number of time steps is the number of subsequences in the window, the number of rows is 1 as we are working with one-dimensional data, and the number of columns represents the number of time steps in the subsequence, in this case 32.

For this chosen framing of the problem, the input for the ConvLSTM2D would therefore be:

  • Samples: n, for the number of windows in the dataset.
  • Time: 4, for the four subsequences that we split a window of 128 time steps into.
  • Rows: 1, for the one-dimensional shape of each subsequence.
  • Columns: 32, for the 32 time steps in an input subsequence.
  • Channels: 9, for the nine input variables.

We can now prepare the data for the ConvLSTM2D model.

The ConvLSTM2D class requires configuration both in terms of the CNN and the LSTM. This includes specifying the number of filters (e.g. 64), the two-dimensional kernel size, in this case (1 row and 3 columns of the subsequence time steps), and the activation function, in this case rectified linear.

As with a CNN or LSTM model, the output must be flattened into one long vector before it can be interpreted by a dense layer.

We can then evaluate the model as we did the LSTM and CNN LSTM models before it.

The complete example is listed below.

As with the prior experiments, running the model prints the performance of the model each time it is fit and evaluated. A summary of the final model performance is presented at the end of the run.

We can see that the model does consistently perform well on the problem achieving an accuracy of about 90%, perhaps with fewer resources than the larger CNN LSTM model.

Note: given the stochastic nature of the algorithm, your specific results may vary. If so, try running the code a few times.

Extensions

This section lists some ideas for extending the tutorial that you may wish to explore.

  • Data Preparation. Consider exploring whether simple data scaling schemes can further lift model performance, such as normalization, standardization, and power transforms.
  • LSTM Variations. There are variations of the LSTM architecture that may achieve better performance on this problem, such as stacked LSTMs and Bidirectional LSTMs.
  • Hyperparameter Tuning. Consider exploring tuning of model hyperparameters such as the number of units, training epochs, batch size, and more.

If you explore any of these extensions, I’d love to know.

Further Reading

This section provides more resources on the topic if you are looking to go deeper.

Papers

Articles

Summary

In this tutorial, you discovered three recurrent neural network architectures for modeling an activity recognition time series classification problem.

Specifically, you learned:

  • How to develop a Long Short-Term Memory Recurrent Neural Network for human activity recognition.
  • How to develop a one-dimensional Convolutional Neural Network LSTM, or CNN LSTM, model.
  • How to develop a one-dimensional Convolutional LSTM, or ConvLSTM, model for the same problem.

Do you have any questions?
Ask your questions in the comments below and I will do my best to answer.


Develop Deep Learning models for Time Series Today!

Deep Learning for Time Series Forecasting

Develop Your Own Forecasting models in Minutes

…with just a few lines of python code

Discover how in my new Ebook:
Deep Learning for Time Series Forecasting

It provides self-study tutorials on topics like: CNNs, LSTMs,
Multivariate Forecasting, Multi-Step Forecasting and much more…

Finally Bring Deep Learning to your Time Series Forecasting Projects

Skip the Academics. Just Results.

Click to learn more.


16 Responses to How to Develop RNN Models for Human Activity Recognition Time Series Classification

  1. Tin September 28, 2018 at 9:20 am #

    Hi Jason,

    So enjoy reading with your stuff, very helpful. As we can use CNN+LSTM to predict the spatial-temporal data, can we reverse the architecture as LSTM+CNN to do the same job? Any examples for LSTM + CNN?

    • Jason Brownlee September 28, 2018 at 2:59 pm #

      Not that I have seen. What application did you have in mind exactly? Sequence to image?

  2. Akilesh October 29, 2018 at 12:40 pm #

    Hi Jason,
    Can you please explain the choice of parameters for the LSTM network?
    Especially the LSTM layer and the dense layer?
    What does the value 100 signify?

  3. Amir October 31, 2018 at 8:13 pm #

    Thanks for such a comprehensive tutorial
    I try it and it worked as expressed in the tutorial.

    Now, I’m going to try it with my data that comes from a single axis accelerometer. It means I have only one feature so I don’t need a 3D array but a 2D.
    You mentioned “The model (RNN LSTM) requires a three-dimensional input with [samples, time steps, features]. This is exactly how we have loaded the data.”
    Then, it means that it won’t work with a 2D array? or I can consider a 3D array but the 3rd dimension has only one member?

    and my second question is:
    I need the model for a real-time classification, so I need to train once and then save the model and use it in my web application.
    how can I save the model after training and use it?

    • Jason Brownlee November 1, 2018 at 6:05 am #

      Yes, even if the 3rd dimension has 1 parameter, it is still a 3D array.

      You can call model.save() to save the model.

  4. Daniel Aguilera Garcia November 21, 2018 at 10:24 am #

    Hello Jason.

    In my case I have 2 time series from EGG and I have to diseign a model that classify in two types de signal. I dont understand exactly how should i reshape the data.

    The freq is 256 values per second so i could divide in windows like you did before. The problem is that i dont know how to put the 3rd dimension of caracteristics. From each window I have 7 caracteristics not from each moment (max, min, std, fft bandwidths, fft centroids, Arima 1, Arima 2)

    Please, how could I do what you mean [samples, time steps, features] in my case??

    • Jason Brownlee November 21, 2018 at 2:10 pm #

      Perhaps those 7 characteristics are 7 features.

      If you have 256 samples per section, you can choose how many samples/seconds to use as time steps and perhaps change the resolution via averaging or removing samples.

      • Daniel Aguilera Garcia November 21, 2018 at 9:16 pm #

        Let’s see im going to make an easy example:

        channel1 (values in one second)=2,5,6,8,54,2,8,4,7,8,…,5,7,8 (in total 256 values per second)

        channel2 (values in one second)=2,5,6,8,54,2,8,4,7,8,…,5,7,8 (in total 256 values per second)

        7 diferent features

        [samples,timesteps,features]=[2, 256, 7]?

        and another questions, for example the mean feature:

        chanel 1:

        feat_mean[0]=2
        feat_mean[1]=(2+5)/2=3.5
        feat_mean[2]=(2+5+6)/3=4.33
        etc…

        is it correct? what I understood is that I have to substract features for each each moment?

  5. coolyj November 22, 2018 at 5:21 pm #

    where does the “561 element vector of features” apply to?

    • Jason Brownlee November 23, 2018 at 7:44 am #

      That vector was the pre-processed data, prepared by the authors of the study that we do not use in this tutorial.

  6. Daniel Aguilera Garcia November 26, 2018 at 8:29 pm #

    hello jason!

    should we normalize each feature??

    • Jason Brownlee November 27, 2018 at 6:33 am #

      Perhaps try it and evaluate how it impacts model performance.

  7. Leon December 3, 2018 at 7:05 am #

    Hi Jason, thanks for the great article.
    I notice in your load data section, you probably mean 2.56 seconds instead of 2.65, since 128(time step) * 1/50(record sampling rate) = 2.56.

Leave a Reply