Successfully reported this slideshow.
Your SlideShare is downloading. ×

[DSC Europe 22] Anomaly detection within a hydroelectric power plan - Cyrille Feudjio

Ad
Ad
Ad
Ad
Ad
Ad
Ad
Ad
Ad
Ad
Ad

Check these out next

1 of 15 Ad

[DSC Europe 22] Anomaly detection within a hydroelectric power plan - Cyrille Feudjio

Download to read offline

The goal of the project is to develop an unsupervised machine learning model to predict when parts are likely to fail within a hydroelectric power plant. The value of such a model is gained by applying its predictions to improve the maintenance and planning schedule of the power plant in order to improve its safety and overall efficiency in its operations. The data set is collected from sensors in the power plant at intervals. These sensors captured different data types such as temperature and pressure.

The goal of the project is to develop an unsupervised machine learning model to predict when parts are likely to fail within a hydroelectric power plant. The value of such a model is gained by applying its predictions to improve the maintenance and planning schedule of the power plant in order to improve its safety and overall efficiency in its operations. The data set is collected from sensors in the power plant at intervals. These sensors captured different data types such as temperature and pressure.

Advertisement
Advertisement

More Related Content

More from DataScienceConferenc1 (20)

Recently uploaded (20)

Advertisement

[DSC Europe 22] Anomaly detection within a hydroelectric power plan - Cyrille Feudjio

  1. 1. Cyrille and Faith Anomaly Detection within a Hydroelectric Power Plant Electricity everywhere and for everyone! 1
  2. 2. 2 About me Cyrille Feudjio Data Scientist & Industrial Engineer ● MSc degree in Industrial Mathematics AIMS, Cameroon. MSc degree in Industrial Engineering NHPSD Doula, Cameroon. I like learning new things and helping others
  3. 3. 3 Context We need Electricity
  4. 4. 4 Context Business Losses Loss of communication Lack of security Unable to perform daily operations
  5. 5. 5 Problem Overview Better schedule of maintenance Detect, predict and explain failures
  6. 6. ETL 2 ETL1 Data Preparation 6 LANDING ZONE RAW ZONE TRUSTED ZONE Great Expectation
  7. 7. 7 Solution 1: Model Isolation Forest Compute all the paths to reach the node with one point and choose the shortest ones based on the contamination.
  8. 8. 8 Solution 1: Results anomalies normal
  9. 9. Solution 2: Model 9 Signature matrices Generation Encoding spatial information. Encoding temporal information. Decoding previous information Reconstructed signature matrices Residual signature matrices - Multi-Scale Convolutional Recurrent Encoder-Decoder Model (MSCRED) WHY MSCRED? ● Detect anomaly events at certain time steps. ● Identify abnormal time series that are most likely to be the causes of each anomaly.
  10. 10. Solution 2: Anomalies 10 All the time step with anomaly score above the red line are anomalies
  11. 11. Solution 2: Root causes 11 55 12 50 19 53 Root causes sensors
  12. 12. 12 Solution 3: Model LSTM (16,16) LSTM (16,8) LSTM (8,16) LSTM (16,16)
  13. 13. 13 Solution 3: Data augmentation  Slicing: The general concept behind slicing is that the data is augmented by slicing time steps off the ends of the pattern Research paper: “An empirical survey of data augmentation for time series classification with neural networks” by Brian Kenji Iwana ,Seiichi Uchida published on july 15, 2021  Window warping is a popular method of time warping. It takes a random window of the time series and stretches it by 2 or contracts it by 1/2.
  14. 14. 14 Solution 3: Results ( in progress)
  15. 15. 15 Cyrille Feudjio cyrille@ishango.ai

×