Added markdown files for three livescripts

README.md

@@ -15,23 +15,23 @@ We show examples on how to perform the following parts of the Deep Learning work
 
 This demo is implemented as a MATLAB project and will require you to open the project to run it. The project will manage all paths and shortcuts you need. There is also a significant data copy required the first time you run the project.
 
-## Part 1 - Data Preparation
+## Part 1 - Data Preparation ([View on Browser](./markdown_view/Part01_DataPreparation.md))
 
 This example shows how to extract the set of acoustic features that will be used as inputs to the LSTM Deep Learning network.
 
 To run:
 1. Open MATLAB project Aircompressorclassification.prj
 2. Open and run Part01_DataPreparation.mlx
 
-## Part 2 - Modeling
+## Part 2 - Modeling ([View on Browser](./markdown_view/Part02_Modeling.md))
 
 This example shows how to train LSTM network to classify multiple modes of operation that include healthy and unhealthy signals.
 
 To run:
 1. Open MATLAB project Aircompressorclassification.prj
 2. Open and run Part02_Modeling.mlx
 
-## Part 3 - Deployment
+## Part 3 - Deployment ([View on Browser](./markdown_view/Part03_Deployment.md))
 
 This example shows how to generate optimized c++ code ready for deployment. 
 

markdown_view/Part01_DataPreparation.md

@@ -0,0 +1,256 @@
+# Air Compressor Data Classification
+# Part 1: Data Preparation
+
+
+Copyright 2020 The MathWorks, Inc.
+
+
+
+
+![image_0.png](Part01_DataPreparation_images/image_0.png)
+
+
+# Configuration
+
+
+Click on the checkboxes below to choose options for how to run this script.
+
+
+
+```matlab:Code
+doFeatureExtraction = false; % if unchecked, this will save time by loading previous results
+```
+
+
+
+Make sure we run this as a project.
+
+
+
+```matlab:Code
+try
+    prj = currentProject;
+catch
+    open("Aircompressorclassification.prj");
+    OpenPart1;
+    prj = currentProject;
+end
+```
+
+# Create Datastore
+
+
+The recorded data is sorted by subfolder. Each subfolder contains over 200 recordings of the labeled state. 
+
+
+
+
+We create our datastore by providing the input data folder and specifying that the source of the labels is the name of the subfolders. 
+
+
+
+```matlab:Code
+dataFolder = 'AirCompressorData';
+ads = audioDatastore(dataFolder,'IncludeSubfolders',true,'LabelSource','foldernames');
+```
+
+
+
+We then reset the random number generator (for consistent results) and shuffle the data.
+
+
+
+```matlab:Code
+rng(3);
+ads = shuffle(ads);
+```
+
+# Split Into Training and Validation Sets
+
+
+Split the data into training and validation by doing a 90% training, 10% validation split. The `countEachLabel` command will show us how many samples of data belong to each category in the dataset.
+
+
+
+```matlab:Code
+[adsTrain,adsValidation] = splitEachLabel(ads,0.9,0.1);
+countEachLabel(adsTrain)
+```
+
+| |Label|Count|
+|:--:|:--:|:--:|
+|1|Bearing|203|
+|2|Flywheel|203|
+|3|Healthy|203|
+|4|LIV|203|
+|5|LOV|203|
+|6|NRV|203|
+|7|Piston|203|
+|8|Riderbelt|203|
+
+
+```matlab:Code
+countEachLabel(adsValidation)
+```
+
+| |Label|Count|
+|:--:|:--:|:--:|
+|1|Bearing|22|
+|2|Flywheel|22|
+|3|Healthy|22|
+|4|LIV|22|
+|5|LOV|22|
+|6|NRV|22|
+|7|Piston|22|
+|8|Riderbelt|22|
+
+# Data Preparation
+## Human Insight
+
+
+The data we are working with are time-series recordings of acoustics from different parts of an air compressor. As such, there are strong relationships between samples in time.
+
+
+
+```matlab:Code
+sampleData = read(adsTrain);
+sampleDataCategory = adsTrain.Labels(1);
+plot(1:numel(sampleData), sampleData);
+xlabel("Sample number");
+ylabel("Amplitude");
+title("Class: " + string(sampleDataCategory));
+```
+
+
+![figure_0.png](Part01_DataPreparation_images/figure_0.png)
+
+
+
+Listen to a sample of the audio if desired.
+
+
+
+```matlab:Code
+% sound(sampleData,16000);
+```
+
+
+
+Because of this, we should be able to use a type of recurrent neural network (RNN) as a model for the data. The type of RNN we will eventually select is a bi-directional long short term memory (LSTM) network. However, before we can get to modeling, it's important to prepare the data adequately. Oftentimes, it is best to transform or extract features from 1-dimensional signal data in order to increase a model's representative power, as we see in the diagram below:
+
+
+
+
+![image_1.png](Part01_DataPreparation_images/image_1.png)
+
+
+
+
+In this part of the workflow, we will focus on how to engineer features from the original data that will aid in the model's ability to classify the inputs.
+
+
+# Generate Training Features
+
+
+The next step is to extract the set of acoustic features that will be used as inputs to the network. 
+
+
+
+
+The Audio Toolbox provides a set of Spectral Descriptor features that are commonly used as inputs to deep learning networks.
+
+
+
+
+We can extract the features with individual functions, or we can simplify the workflow and use a single object called audioFeatureExtractor to do it all at once. 
+
+
+
+```matlab:Code
+trainingFeatures = cell(1,numel(adsTrain.Files));
+windowLength = 512;
+overlapLength = 0;
+
+aFE = audioFeatureExtractor('SampleRate',16e3, ...
+    'Window',hamming(windowLength,'periodic'),...
+    'OverlapLength',overlapLength,...
+    'spectralCentroid',true, ...
+    'spectralCrest',true, ...
+    'spectralDecrease',true, ...
+    'spectralEntropy',true,...
+    'spectralFlatness',true,...
+    'spectralFlux',false,...                
+    'spectralKurtosis',true,...
+    'spectralRolloffPoint',true,...
+    'spectralSkewness',true,...
+    'spectralSlope',true,...
+    'spectralSpread',true);
+
+if doFeatureExtraction
+    reset(adsTrain);
+    index = 1;
+    tic;
+    while hasdata(adsTrain)
+        data = read(adsTrain);
+        trainingFeatures{index} = extract(aFE,data);
+        index = index + 1;
+    end
+    fprintf('Extraction took %f seconds.\n',toc);
+else
+    load("TrainingFeatures.mat");
+    disp("Training data features loaded.")
+end
+```
+
+
+```text:Output
+Extraction took 31.829615 seconds.
+```
+
+# Normalize Training Features
+
+
+Networks will often train better when normalized. Calculate the mean and standard deviation and normalize each element of the training feature set. 
+
+
+
+```matlab:Code
+allTrainingFeatures = cat(1,trainingFeatures{:});
+M = mean(allTrainingFeatures);
+S = std(allTrainingFeatures);
+
+for index = 1:numel(adsTrain.Files)
+   trainingFeatures{index} =  ((trainingFeatures{index} - M)./S).';
+end
+```
+
+# Generate and Normalize Validation Features
+
+
+Repeat the feature extraction for the validation features. Perform the normalization inside the loop.
+
+
+
+```matlab:Code
+validationFeatures = cell(1,numel(adsValidation.Files));
+
+if doFeatureExtraction
+    index = 1;
+    tic;
+    while hasdata(adsValidation)
+       data = read(adsValidation);
+       validationFeatures{index} = extract(aFE,data);
+       validationFeatures{index} = ((validationFeatures{index}  - M) ./ S).';
+       index = index + 1;
+    end
+    fprintf('Validation Extraction took %f seconds.\n',toc);
+else
+    load("ValidationFeatures.mat");
+end
+```
+
+
+```text:Output
+Validation Extraction took 3.328731 seconds.
+```
+