CalvingDetection

This github repository keeps the code and data for calving detection.

---

Requirements

• Python 3.11.12
• Libraries: os, pandas, numpy, scikit-learn, random, matplotlib, joblib, seaborn, datetime, dateutil, pytz, torch

---

This project focuses on detecting calving behavior in cattle using sensor data. It is divided into two main stages:

Stage 1: Anomaly Detection

First change all the file paths marked with # Change Required #

The paths in the code point to folders in this Github Repo just to indicate which data is being used. When you actually run it you may want to download the data and change the corresponding path to your local path.

Workflow

Training autoencoder

1. Load and parse CSV files (Processed_data) from the Processed_data folder (one file per cow).
2. Compute features:
   • KNN_Sum: sum of KNN2–KNN5
   • ts: derived behavioral index using ActivityCount_Difference, Distance_moved, and KNN_Sum
   • sin_hour: $np.sin(2 * pi * hours / 24)$
   • cos_hour: $np.cos(2 * pi * hours / 24)$
3. Drop null entries and set Rounded_Time as the index.
4. Remove outliers using IQR filtering on each feature.
5. Normalize the data using MinMaxScaler (individually per cow).
6. Save the fitted scaler for future use.
7. Train the autoencoder model.

Applying the trained autoencoder

8. Process test data.
   • Compute features needed
   • Drop null entries
   • Normalize the test data using saved MinMaxScaler (individually per cow).
9. Apply the trained AE model to detect anomalies.

Combined with the random forest model obtained in stage 2

Output

• Individual Min-Max scaler files for each cow
• Trained autoencoder model
• Anomalies

---

Stage 2: Classification

First change all the file paths marked with # Change Required #

Workflow

1. Load detected anomaly CSVs for each cow.
2. Add sliding features including (MA, STD, Diff1, and Diff3) for each feature.
   • Moving average (MA) and standard deviation (STD) over 6 and 12 hours
   • Difference from previous hours (1 and 3 hours)
   • Applied on features: KNN2–KNN5, Distance_moved, ActivityCount_Difference
3. Label anomalies with boolean value to mark if it's a calving event.
4. Train Random Forest model

Output

• Trained Random Forest model

You can then go back to the Stage1 code and turn USE_CLF to be True in order to plot the results with the classification results.

---