Skip to main content

Predicting Car Battery Failure With R And H2O - Study

# Loading libraries
suppressWarnings( suppressMessages( library( h2o ) ) ) 
suppressWarnings( suppressMessages( library( data.table ) ) )
suppressWarnings( suppressMessages( library( plotly ) ) )
suppressWarnings( suppressMessages( library( DT ) ) )

# Reading data file
# Data from: https://www.kaggle.com/yunlevin/levin-vehicle-telematics
dataFileName = "/Development/Analytics/AnomalyDetection/AutomovileFailurePrediction/v2.csv"
carData = fread( dataFileName, skip=0, header = TRUE )
carBatteryData = data.table( TimeStamp = carData$timeStamp
                             , BatteryVoltage = as.numeric( carData$battery ) 
                            )
rm(carData)

# Data cleaning, filtering and conversion
carBatteryData = na.omit( carBatteryData ) # Keeping just valid Values

# According to this article: 
# https://shop.advanceautoparts.com/r/advice/car-maintenance/car-battery-voltage-range
#
# A perfect voltage ( without any devices or electronic systems plugged in )  
# is between 13.7 and 14.7V. 
# If the battery isn’t fully charged, it will diminish to 12.4V at 75%, 
# 12V when it’s only operating at 25%, and up to 11.9V when it’s completely discharged. 
#
# Battery voltage while a load is connected is much slower
# it should be something between 9.5V and 10.5V 
#
# This value interval ensures that your battery can store and deliver enough 
# current to start your car and power all your electronics and electric devices 
# without any difficulty

carBatteryData = carBatteryData[BatteryVoltage >= 9.5] # Filtering voltages greater or equal to 9.5
carBatteryData$TimeStamp = as.POSIXct( paste0( substr(carBatteryData$TimeStamp,1,17),"00" ) )
carBatteryData = unique(carBatteryData) # Removing duplicate voltage readings
carBatteryData = carBatteryData[order(TimeStamp)]


# spliting all data, using the last date as testing data and the rest for training.
lastDate = max( as.Date( format( carBatteryData$TimeStamp, "%Y-%m-%d" ) ) )
trainingData = carBatteryData[ as.Date( format( carBatteryData$TimeStamp, "%Y-%m-%d" ) ) != lastDate ]
testingData = carBatteryData[ as.Date( format( carBatteryData$TimeStamp, "%Y-%m-%d" ) ) == lastDate ]



################################################################################
# Creating Anomaly Detection Model
################################################################################

  h2o.init( nthreads = -1, max_mem_size = "5G" )
## 
## H2O is not running yet, starting it now...
## 
## Note:  In case of errors look at the following log files:
##     C:\Users\LaranIkal\AppData\Local\Temp\Rtmp6lTw4H/h2o_LaranIkal_started_from_r.out
##     C:\Users\LaranIkal\AppData\Local\Temp\Rtmp6lTw4H/h2o_LaranIkal_started_from_r.err
## 
## 
## Starting H2O JVM and connecting:  Connection successful!
## 
## R is connected to the H2O cluster: 
##     H2O cluster uptime:         1 seconds 899 milliseconds 
##     H2O cluster timezone:       America/Mexico_City 
##     H2O data parsing timezone:  UTC 
##     H2O cluster version:        3.24.0.2 
##     H2O cluster version age:    1 month and 7 days  
##     H2O cluster name:           H2O_started_from_R_LaranIkal_tzd452 
##     H2O cluster total nodes:    1 
##     H2O cluster total memory:   4.44 GB 
##     H2O cluster total cores:    8 
##     H2O cluster allowed cores:  8 
##     H2O cluster healthy:        TRUE 
##     H2O Connection ip:          localhost 
##     H2O Connection port:        54321 
##     H2O Connection proxy:       NA 
##     H2O Internal Security:      FALSE 
##     H2O API Extensions:         Amazon S3, Algos, AutoML, Core V3, Core V4 
##     R Version:                  R version 3.6.0 (2019-04-26)
  h2o.no_progress() # Disable progress bars for Rmd
  h2o.removeAll() # Cleans h2o cluster state.
## [1] 0
  # Convert the training dataset to H2O format.
  trainingData_hex = as.h2o( trainingData[,2], destination_frame = "train_hex" )
  
  # Build an Isolation forest model
  trainingModel = h2o.isolationForest( training_frame = trainingData_hex
                                       , sample_rate = 0.1
                                       , max_depth = 32
                                       , ntrees = 100
                                      )
  
  # According to H2O doc: 
  # http://docs.h2o.ai/h2o/latest-stable/h2o-docs/data-science/if.html
  #
  # Isolation Forest is similar in principle to Random Forest and is built on the basis of decision trees. 
  
  # Isolation Forest creates multiple decision trees to isolate observations.
  # 
  # Trees are split randomly, The assumption is that:
  #   
  #   IF ONE UNIT MEASUREMENTS ARE SIMILAR TO OTHERS,
  #   IT WILL TAKE MORE RANDOM SPLITS TO ISOLATE IT.
  # 
  #   The less splits needed, the unit is more likely to be anomalous.
  # 
  # The average number of splits is then used as a score.

  # Calculate score for training dataset
  score <- h2o.predict( trainingModel, trainingData_hex )
  result_pred <- as.vector( score$predict )


################################################################################
# Setting threshold value for anomaly detection.
################################################################################

  # Setting desired threshold percentage.
  threshold = .995 # Let's say we have 99.5% voltage values correct
  
  # Using avobe threshold to get score limit to filter anomalous voltage readings.
  scoreLimit = round( quantile( result_pred, threshold ), 4 )
  

  
################################################################################
# Get anomalous voltage readings from testing data, using model and scoreLimit got using training data.
################################################################################

  # Convert testing data frame to H2O format.
  testingDataH2O = as.h2o( testingData[,2], destination_frame = "testingData_hex" )
  
  # Get score using training model
  testingScore <- h2o.predict( trainingModel, testingDataH2O )

  # Add row score at the beginning of testing dataset
  testingData = cbind( RowScore = round( as.vector( testingScore$predict ), 4 ), testingData )

  # Check if there are anomalous voltage readings from testing data
  anomalies = testingData[ testingData$RowScore > scoreLimit, ]
# Here there is and additional filter to ensure maintenance recommendation
  # If there are more than 3 anomalous voltage readings, display an alert.
  if( dim( anomalies )[1]  > 3 ) { 
    cat( "Show alert on car display: Battery got anomalous voltage readings, it is recommended to take it to service." )
    
    plot_ly( data = anomalies
             , x = ~TimeStamp
             , y = ~BatteryVoltage
             , type = 'scatter'
             , mode = "lines"
             , name = 'Anomalies') %>%
      layout( yaxis = list( title = 'Battery Voltage.' )
              , xaxis = list( categoryorder='trace', title = 'Date - Time.' )
               )
  }
## Show alert on car display: Battery got anomalous voltage readings, it is recommended to take it to service.
if( dim( anomalies )[1]  > 3 ) { 
  DT::datatable(anomalies[,c(2,3)], rownames = FALSE )
}
Using this approach we may prevent failures on cars, not only for batteries but for many cases when sensors are used.
Carlos Kassab
We are using R, more information about R:

Comments

Popular posts from this blog

Using R and H2O to identify product anomalies during the manufacturing process.

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.

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

How many records?:  Training data set - In my real project, I use 100 thousand test passed records, it is around a month of production data. Testing data set - I use …

Production Line Stations Maintenance Prediction - Process Flow.

Steps Needed in a Process to Detect Anomalies And Have a Maintenance Notice Before We Have Scrap Created on The Production Line.

Describing my previous articles( 1, 2 ) process flow:

Get Training Data.At least 2 weeks of passed units measurements.Data Cleaning.Ensure no null values.At least 95% data must have measurement values.Anomalies Detection Model Creation.Deep Learning Autoencoders.orIsolation Forest.Set Yield Threshold Desired, Normally 99%Get Prediction Value Limit by Linking Yield Threshold to Training Data Using The Anomaly Detection Model Created.Get Testing Data.Last 24 Hour Data From Station Measurements, Passed And Failed Units.Testing Data Cleaning.Ensure no null values.Get Anomalies From Testing Data by Using The Model Created And Prediction Limit Found Before.If Anomalies Found, Notify Maintenance to Avoid Scrap.Display Chart Showing Last 24 Hour Anomalies And Failures Found:


As you can see( Anomalies in blue,Failures in orange ), we are detecting anomalies( Units clo…