Introduction:

We will identify anomalous products on the production line by using measurements from testing stations and deep learning models. Anomalous products are not failures, these anomalies are products close to the measurement limits, so we can display warnings before the process starts to make failed products and in this way the stations get maintenance.

 Before starting we need the next software installed and working:

– R language installed.
– All the R packages mentioned in the R sources.
– Testing stations data, I suggest to go station by station.
– H2O open source framework.
– Java 8 ( For H2O ). Open JDK: https://github.com/ojdkbuild/contrib_jdk8u-ci/releases
– R studio.


About the data: Since I cannot use my real data, for this article I am using SECOM Data Set from UCI Machine Learning Repository   

h2o.removeAll() ## Removes the data from the h2o cluster in preparation for our final model.

# Reading SECOM data file
allData = read.csv( “secom.data”, sep = ” “, header = FALSE, encoding = “UTF-8” )

# fixing the data set, there are a lot of NaN records
if( dim(na.omit(allData))[1] == 0 ){
  for( colNum in 1:dim( allData )[2] ){

    # Get valid values from the actual column values
    ValidColumnValues = allData[,colNum][!is.nan( allData[, colNum] )]

    # Check each value in the actual active column.
    for( rowNum in 1:dim( allData )[1] ){
    cat( “Processing row:”, rowNum, “, Column:”, colNum, “Data:”, allData[rowNum, colNum], “\n” )
    if( is.nan( allData[rowNum, colNum] ) ) {
      # Assign random valid value to our row,column with NA value
      allData[rowNum, colNum] = getValue
    }
  }
 }
}

# spliting all data, the fiirst 90% for training and the rest 10% for testing our model.
trainingData = allData[1:floor(dim(allData)[1]*.9),]
testingData = allData[(floor(dim(allData)[1]*.9)+1):dim(allData)[1],]
trainingData_hex = as.h2o( trainingData, destination_frame = “train_hex” )

# Set the input variables
featureNames = colnames(trainingData_hex)

# Creating the first model version.
trainingModel = h2o.deeplearning( x = featureNames
                                                         , model_id = “Station1DeepLearningModel”
                                                         , activation = “Tanh”
                                                         , autoencoder = TRUE
                                                         , reproducible = TRUE
                                                         , l1 = 1e-5
                                                         , ignore_const_cols = FALSE
                                                         , seed = 1234
                                                         , hidden = c( 400, 200, 400 ), epochs = 50 )


# Getting the anomalies from training data to set the min MSE( Mean Squared Error )
# value before setting a record as anomally
trainMSE = as.data.frame( h2o.anomaly( trainingModel
                                                                 , trainingData_hex
                                                                 , per_feature = FALSE ) )

# Check the first 30 descendent sorted trainMSE records to see our outliers
head( sort( trainMSE$Reconstruction.MSE , decreasing = TRUE ), 30)
# 0.020288603 0.017976305 0.012772556 0.011556780 0.010143009 0.009524983 0.007363854
# 0.005889714 0.005604329 0.005189614[11] 0.005185285 0.005118595 0.004639442 0.004497609
# 0.004438342 0.004419993 0.004298936 0.003961503 0.003651326 0.003426971 0.003367108
# 0.003169319 0.002901914 0.002852006 0.002772110 0.002765924 0.002754586 0.002748887
# 0.002619872 0.002474702

# Ploting errors of reconstructing our training data, to have a graphical view
# of our data reconstruction errors

plot( sort( trainMSE$Reconstruction.MSE ), main = ‘Reconstruction Error’, ylab = “MSE Value.” )