Classifying bodily exercise from smartphone information

Introduction

On this submit we’ll describe how one can use smartphone accelerometer and gyroscope information to foretell the bodily actions of the people carrying the telephones. The info used on this submit comes from the Smartphone-Primarily based Recognition of Human Actions and Postural Transitions Knowledge Set distributed by the College of California, Irvine. Thirty people have been tasked with performing varied fundamental actions with an connected smartphone recording motion utilizing an accelerometer and gyroscope.

Earlier than we start, let’s load the varied libraries that we’ll use within the evaluation:

library(keras)     # Neural Networks
library(tidyverse) # Knowledge cleansing / Visualization
library(knitr)     # Desk printing
library(rmarkdown) # Misc. output utilities 
library(ggridges)  # Visualization

Actions dataset

The info used on this submit come from the Smartphone-Primarily based Recognition of Human Actions and Postural Transitions Knowledge Set(Reyes-Ortiz et al. 2016) distributed by the College of California, Irvine.

When downloaded from the hyperlink above, the info incorporates two totally different ‘components.’ One which has been pre-processed utilizing varied function extraction methods similar to fast-fourier remodel, and one other RawData part that merely provides the uncooked X,Y,Z instructions of an accelerometer and gyroscope. None of the usual noise filtering or function extraction utilized in accelerometer information has been utilized. That is the info set we are going to use.

The motivation for working with the uncooked information on this submit is to help the transition of the code/ideas to time sequence information in much less well-characterized domains. Whereas a extra correct mannequin could possibly be made by using the filtered/cleaned information supplied, the filtering and transformation can fluctuate tremendously from activity to activity; requiring numerous handbook effort and area data. One of many stunning issues about deep studying is the function extraction is discovered from the info, not exterior data.

Exercise labels

The info has integer encodings for the actions which, whereas not essential to the mannequin itself, are useful to be used to see. Let’s load them first.

activityLabels % kable(align = c("c", "l"))

Subsequent, we load within the labels key for the RawData. This file is an inventory of all the observations, or particular person exercise recordings, contained within the information set. The important thing for the columns is taken from the info README.txt.

Column 1: experiment quantity ID, 
Column 2: person quantity ID, 
Column 3: exercise quantity ID 
Column 4: Label begin level 
Column 5: Label finish level 

The beginning and finish factors are in variety of sign log samples (recorded at 50hz).

Let’s check out the primary 50 rows:

labels % 
  head(50) %>% 
  paged_table()

File names

Subsequent, let’s take a look at the precise recordsdata of the person information supplied to us in RawData/

dataFiles % head()

[1] "acc_exp01_user01.txt" "acc_exp02_user01.txt"
[3] "acc_exp03_user02.txt" "acc_exp04_user02.txt"
[5] "acc_exp05_user03.txt" "acc_exp06_user03.txt"

There’s a three-part file naming scheme. The primary half is the kind of information the file incorporates: both acc for accelerometer or gyro for gyroscope. Subsequent is the experiment quantity, and final is the person Id for the recording. Let’s load these right into a dataframe for ease of use later.

fileInfo %
  filter(filePath != "labels.txt") %>% 
  separate(filePath, sep = '_', 
           into = c("sort", "experiment", "userId"), 
           take away = FALSE) %>% 
  mutate(
    experiment = str_remove(experiment, "exp"),
    userId = str_remove_all(userId, "person|.txt")
  ) %>% 
  unfold(sort, filePath)

fileInfo %>% head() %>% kable()

Studying and gathering information

Earlier than we will do something with the info supplied we have to get it right into a model-friendly format. This implies we wish to have an inventory of observations, their class (or exercise label), and the info akin to the recording.

To acquire this we are going to scan by every of the recording recordsdata current in dataFiles, lookup what observations are contained within the recording, extract these recordings and return every little thing to a straightforward to mannequin with dataframe.

# Learn contents of single file to a dataframe with accelerometer and gyro information.
readInData % 
    filter(userId == as.integer(curUserId), 
           experiment == as.integer(curExperiment))
  

  # extract observations as dataframes and save as a column in dataframe.
  dataLabels %>% 
    mutate(
      information = map2(startPos, endPos, extractObservation)
    ) %>% 
    choose(-startPos, -endPos)
}

# scan by all experiment and userId combos and collect information right into a dataframe. 
allObservations % 
  right_join(activityLabels, by = c("exercise" = "quantity")) %>% 
  rename(activityName = label)

# cache work. 
write_rds(allObservations, "allObservations.rds")
allObservations %>% dim()

Exploring the info

Now that we have now all the info loaded together with the experiment, userId, and exercise labels, we will discover the info set.

Size of recordings

Let’s first take a look at the size of the recordings by exercise.

allObservations %>% 
  mutate(recording_length = map_int(information,nrow)) %>% 
  ggplot(aes(x = recording_length, y = activityName)) +
  geom_density_ridges(alpha = 0.8)

The very fact there’s such a distinction in size of recording between the totally different exercise sorts requires us to be a bit cautious with how we proceed. If we prepare the mannequin on each class without delay we’re going to need to pad all of the observations to the size of the longest, which would depart a big majority of the observations with an enormous proportion of their information being simply padding-zeros. Due to this, we are going to match our mannequin to simply the biggest ‘group’ of observations size actions, these embody STAND_TO_SIT, STAND_TO_LIE, SIT_TO_STAND, SIT_TO_LIE, LIE_TO_STAND, and LIE_TO_SIT.

An attention-grabbing future course could be trying to make use of one other structure similar to an RNN that may deal with variable size inputs and coaching it on all the info. Nonetheless, you’ll run the chance of the mannequin studying merely that if the commentary is lengthy it’s most definitely one of many 4 longest courses which might not generalize to a situation the place you have been working this mannequin on a real-time-stream of knowledge.

Filtering actions

Primarily based on our work from above, let’s subset the info to simply be of the actions of curiosity.

desiredActivities % 
  filter(activityName %in% desiredActivities) %>% 
  mutate(observationId = 1:n())

filteredObservations %>% paged_table()

So after our aggressive pruning of the info we could have a decent quantity of knowledge left upon which our mannequin can study.

Coaching/testing cut up

Earlier than we go any additional into exploring the info for our mannequin, in an try and be as honest as attainable with our efficiency measures, we have to cut up the info right into a prepare and take a look at set. Since every person carried out all actions simply as soon as (excluding one who solely did 10 of the 12 actions) by splitting on userId we are going to make sure that our mannequin sees new folks completely after we take a look at it.

# get all customers
userIds % distinctive()

# randomly select 24 (80% of 30 people) for coaching
set.seed(42) # seed for reproducibility
trainIds % 
  filter(userId %in% trainIds)

testData % 
  filter(userId %in% testIds)

Visualizing actions

Now that we have now trimmed our information by eradicating actions and splitting off a take a look at set, we will really visualize the info for every class to see if there’s any instantly discernible form that our mannequin could possibly choose up on.

First let’s unpack our information from its dataframe of one-row-per-observation to a tidy model of all of the observations.

unpackedObs % 
  map_df(perform(rowNum){
    dataRow % 
      mutate(
        activityName = dataRow$activityName, 
        observationId = dataRow$observationId,
        time = 1:n() )
  }) %>% 
  collect(studying, worth, -time, -activityName, -observationId) %>% 
  separate(studying, into = c("sort", "course"), sep = "_") %>% 
  mutate(sort = ifelse(sort == "a", "acceleration", "gyro"))

Now we have now an unpacked set of our observations, let’s visualize them!

unpackedObs %>% 
  ggplot(aes(x = time, y = worth, colour = course)) +
  geom_line(alpha = 0.2) +
  geom_smooth(se = FALSE, alpha = 0.7, dimension = 0.5) +
  facet_grid(sort ~ activityName, scales = "free_y") +
  theme_minimal() +
  theme( axis.textual content.x = element_blank() )

So at the least within the accelerometer information patterns undoubtedly emerge. One would think about that the mannequin could have bother with the variations between LIE_TO_SIT and LIE_TO_STAND as they’ve an identical profile on common. The identical goes for SIT_TO_STAND and STAND_TO_SIT.

Preprocessing

Earlier than we will prepare the neural community, we have to take a few steps to preprocess the info.

Padding observations

First we are going to determine what size to pad (and truncate) our sequences to by discovering what the 98th percentile size is. By not utilizing the very longest commentary size this may assist us keep away from extra-long outlier recordings messing up the padding.

padSize % 
  map_int(nrow) %>% 
  quantile(p = 0.98) %>% 
  ceiling()
padSize

98% 
334 

Now we merely have to convert our record of observations to matrices, then use the tremendous useful pad_sequences() perform in Keras to pad all observations and switch them right into a 3D tensor for us.

convertToTensor % 
  map(as.matrix) %>% 
  pad_sequences(maxlen = padSize)

trainObs % convertToTensor()
testObs % convertToTensor()
  
dim(trainObs)

[1] 286 334   6

Fantastic, we now have our information in a pleasant neural-network-friendly format of a 3D tensor with dimensions (, , ).

One-hot encoding

There’s one last item we have to do earlier than we will prepare our mannequin, and that’s flip our commentary courses from integers into one-hot, or dummy encoded, vectors. Fortunately, once more Keras has equipped us with a really useful perform to do exactly this.

oneHotClasses % 
  {. - 7} %>%        # deliver integers all the way down to 0-6 from 7-12
  to_categorical() # One-hot encode

trainY % oneHotClasses()
testY % oneHotClasses()

Modeling

Structure

Since we have now temporally dense time-series information we are going to make use of 1D convolutional layers. With temporally-dense information, an RNN has to study very lengthy dependencies as a way to choose up on patterns, CNNs can merely stack just a few convolutional layers to construct sample representations of considerable size. Since we’re additionally merely searching for a single classification of exercise for every commentary, we will simply use pooling to ‘summarize’ the CNNs view of the info right into a dense layer.

Along with stacking two layer_conv_1d() layers, we are going to use batch norm and dropout (the spatial variant(Tompson et al. 2014) on the convolutional layers and customary on the dense) to regularize the community.

input_shape % 
  layer_conv_1d(
    filters = filters,
    kernel_size = kernel_size, 
    input_shape = input_shape,
    padding = "legitimate", 
    activation = "relu"
  ) %>%
  layer_batch_normalization() %>%
  layer_spatial_dropout_1d(0.15) %>% 
  layer_conv_1d(
    filters = filters/2,
    kernel_size = kernel_size,
    activation = "relu",
  ) %>%
  # Apply common pooling:
  layer_global_average_pooling_1d() %>% 
  layer_batch_normalization() %>%
  layer_dropout(0.2) %>% 
  layer_dense(
    dense_size,
    activation = "relu"
  ) %>% 
  layer_batch_normalization() %>%
  layer_dropout(0.25) %>% 
  layer_dense(
    num_classes, 
    activation = "softmax",
    title = "dense_output"
  ) 

abstract(mannequin)

______________________________________________________________________
Layer (sort)                   Output Form                Param #    
======================================================================
conv1d_1 (Conv1D)              (None, 327, 24)             1176       
______________________________________________________________________
batch_normalization_1 (BatchNo (None, 327, 24)             96         
______________________________________________________________________
spatial_dropout1d_1 (SpatialDr (None, 327, 24)             0          
______________________________________________________________________
conv1d_2 (Conv1D)              (None, 320, 12)             2316       
______________________________________________________________________
global_average_pooling1d_1 (Gl (None, 12)                  0          
______________________________________________________________________
batch_normalization_2 (BatchNo (None, 12)                  48         
______________________________________________________________________
dropout_1 (Dropout)            (None, 12)                  0          
______________________________________________________________________
dense_1 (Dense)                (None, 48)                  624        
______________________________________________________________________
batch_normalization_3 (BatchNo (None, 48)                  192        
______________________________________________________________________
dropout_2 (Dropout)            (None, 48)                  0          
______________________________________________________________________
dense_output (Dense)           (None, 6)                   294        
======================================================================
Whole params: 4,746
Trainable params: 4,578
Non-trainable params: 168
______________________________________________________________________

Coaching

Now we will prepare the mannequin utilizing our take a look at and coaching information. Notice that we use callback_model_checkpoint() to make sure that we save solely one of the best variation of the mannequin (fascinating since sooner or later in coaching the mannequin could start to overfit or in any other case cease enhancing).

# Compile mannequin
mannequin %>% compile(
  loss = "categorical_crossentropy",
  optimizer = "rmsprop",
  metrics = "accuracy"
)

trainHistory %
  match(
    x = trainObs, y = trainY,
    epochs = 350,
    validation_data = record(testObs, testY),
    callbacks = record(
      callback_model_checkpoint("best_model.h5", 
                                save_best_only = TRUE)
    )
  )

The mannequin is studying one thing! We get a decent 94.4% accuracy on the validation information, not unhealthy with six attainable courses to select from. Let’s look into the validation efficiency just a little deeper to see the place the mannequin is messing up.

Analysis

Now that we have now a skilled mannequin let’s examine the errors that it made on our testing information. We will load one of the best mannequin from coaching based mostly upon validation accuracy after which take a look at every commentary, what the mannequin predicted, how excessive a likelihood it assigned, and the true exercise label.

# dataframe to get labels onto one-hot encoded prediction columns
oneHotToLabel % 
  mutate(quantity = quantity - 7) %>% 
  filter(quantity >= 0) %>% 
  mutate(class = paste0("V",quantity + 1)) %>% 
  choose(-number)

# Load our greatest mannequin checkpoint
bestModel % 
  predict(testObs) %>% 
  as_data_frame() %>% 
  mutate(obs = 1:n()) %>% 
  collect(class, prob, -obs) %>% 
  right_join(oneHotToLabel, by = "class")

predictionPerformance % 
  group_by(obs) %>% 
  summarise(
    highestProb = max(prob),
    predicted = label[prob == highestProb]
  ) %>% 
  mutate(
    fact = testData$activityName,
    right = fact == predicted
  ) 

predictionPerformance %>% paged_table()

First, let’s take a look at how ‘assured’ the mannequin was by if the prediction was right or not.

predictionPerformance %>% 
  mutate(end result = ifelse(right, 'Right', 'Incorrect')) %>% 
  ggplot(aes(highestProb)) +
  geom_histogram(binwidth = 0.01) +
  geom_rug(alpha = 0.5) +
  facet_grid(end result~.) +
  ggtitle("Possibilities related to prediction by correctness")

Reassuringly it appears the mannequin was, on common, much less assured about its classifications for the wrong outcomes than the proper ones. (Though, the pattern dimension is just too small to say something definitively.)

Let’s see what actions the mannequin had the toughest time with utilizing a confusion matrix.

predictionPerformance %>% 
  group_by(fact, predicted) %>% 
  summarise(rely = n()) %>% 
  mutate(good = fact == predicted) %>% 
  ggplot(aes(x = fact,  y = predicted)) +
  geom_point(aes(dimension = rely, colour = good)) +
  geom_text(aes(label = rely), 
            hjust = 0, vjust = 0, 
            nudge_x = 0.1, nudge_y = 0.1) + 
  guides(colour = FALSE, dimension = FALSE) +
  theme_minimal()

We see that, because the preliminary visualization steered, the mannequin had a little bit of bother with distinguishing between LIE_TO_SIT and LIE_TO_STAND courses, together with the SIT_TO_LIE and STAND_TO_LIE, which even have comparable visible profiles.

Future instructions

The obvious future course to take this evaluation could be to aim to make the mannequin extra normal by working with extra of the equipped exercise sorts. One other attention-grabbing course could be to not separate the recordings into distinct ‘observations’ however as a substitute hold them as one streaming set of knowledge, very similar to an actual world deployment of a mannequin would work, and see how nicely a mannequin might classify streaming information and detect modifications in exercise.