ezraerb / neuralnetsensorclassifier Goto Github PK

These programs classify data from a gas sensor array. The array was used to
collect time series of sensor data under various conditions. The goal is to 
classify each time series to the correct set of conditions. A classic gradient
descent neural network was used to do the classification. The network was
explictly implemented instead of using a package to learn the algorithm. 

The program input is two files, one of sensor data tine series and one of the 
category each belongs in. The data came from a public data set available here:
https://archive.ics.uci.edu/ml/datasets/Gas+sensors+for+home+activity+monitoring

The time series are first preprocessed by removing background data and then
normalizing to the range 0..1 based on the lowest and higest values seen over
all samples. The samples are then split by category and then further split into
three groups: training, validation, and test. The split by category ensure that
no group has a bias toward a particular category, which can affect the accuracy
of the network.

The network is trained time series by time series, using the classic early
stopping algorithm. The algorighm has a patience period of 3 epochs, meaning it
stops after the epocs of increasing validation error.

The performance of the final network is tested on the remaining samples to 
produce a confusion matrix. The precision, recall, and balance F statistic are
then calculated for each category, and overall, based on the matrix.

Initial results showed that certain sensors produced very noisy data that was
confusing the network. Two data plot utilities, SamplePlotter and
CrossSampleSensorPlotter were written to visualize the output of various
sensors. With these results, sensors were exluded from the netword to 
potentially improve classification results.

Results show excluding noisy sensors did improve classification, but only to a
point. Excluding more than one-third of sensors led to worse classification and
significantly increased training time.

Sample results:
All sensors:
                           predicted
      actual      banana  background        wine
      banana         5.0         2.0         2.0
  background         2.0         5.0         0.0
        wine         3.0         1.0         6.0
Category: banana Precision: 0.5 Recall: 0.555555555556 Balanced F Statistic: 0.5
26315789474
Category: background Precision: 0.625 Recall: 0.714285714286 Balanced F Statisti
c: 0.666666666667
Category: wine Precision: 0.75 Recall: 0.6 Balanced F Statistic: 0.666666666667
Overall: Precision: 0.625 Recall: 0.62328042328 Balanced F Statistic: 0.61988304
0936

Ignore last two:
                           predicted
      actual      banana  background        wine
      banana         4.0         4.0         1.0
  background         1.0         5.0         1.0
        wine         1.0         1.0         8.0
Category: banana Precision: 0.666666666667 Recall: 0.444444444444 Balanced F Sta
tistic: 0.533333333333
Category: background Precision: 0.5 Recall: 0.714285714286 Balanced F Statistic:
 0.588235294118
Category: wine Precision: 0.8 Recall: 0.8 Balanced F Statistic: 0.8
Overall: Precision: 0.655555555556 Recall: 0.65291005291 Balanced F Statistic: 0
.640522875817

Ignore last three:
                           predicted
      actual      banana  background        wine
      banana         5.0         3.0         1.0
  background         0.0         7.0         0.0
        wine         2.0         1.0         7.0
Category: banana Precision: 0.714285714286 Recall: 0.555555555556 Balanced F Sta
tistic: 0.625
Category: background Precision: 0.636363636364 Recall: 1.0 Balanced F Statistic:
 0.777777777778
Category: wine Precision: 0.875 Recall: 0.7 Balanced F Statistic: 0.777777777778

Overall: Precision: 0.741883116883 Recall: 0.751851851852 Balanced F Statistic:
0.726851851852

The code was written on Python 2.7. It requires NumPy and matplotlib.

  Copyright (C) 2016   Ezra Erb

  This program is free software: you can redistribute it and/or modify
  it under the terms of the GNU General Public License version 3 as published
  by the Free Software Foundation.

  This program is distributed in the hope that it will be useful,
  but WITHOUT ANY WARRANTY; without even the implied warranty of
  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  GNU General Public License for more details.

  You should have received a copy of the GNU General Public License
  along with this program.  If not, see <http://www.gnu.org/licenses/>.

  I'd appreciate a note if you find this program useful or make
  updates. Please contact me through LinkedIn or github (my profile also has
  a link to the code depository)