smartdolphin / collider-prediction

Collider detection challenge using acceleration sensor (by Korea Atomic Energy Research Institute)

collider prediction using acceleration sensor

Collider detection using acceleration sensor task was to predict collider 2D position(x, y), mass and velocity.

Summary of approach

• LB Score: Entered top 7% (Public 0.0063, Private: 0.00662)

Data

• Data visualization
  
  
  
  
  

Method

• Residual CNN for time-series (M)
• Multi-layer perceptron (X, Y, V)
• Model average ensemble
• Permutation feature selection

Feature engineering

• Distance Feature using signal arrival time delta
• Original Signal
  
  
• Fast Fourier Transform (FFT)
  
  
• Power density estimation (PSD)
  
  
• Auto correlation
  
  
• Spectrogram
  
  
• Mel spectrogram (except)
  
  
• Statistical information
• Peak distance mean/std