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