Hybrid CNN-LSTM Model with an Attention Mechanism for Early Detection of Lameness in Cattle Based on Sensor Data

This paper presents a hybrid CNN-LSTM model with an attention mechanism for the early detection of lameness in cattle using data from a pedometer-type wearable sensor unit. The input feature set includes accelerometer, gyroscope, and magnetometer measurements, generated statistical activity indicators, and demographic and physiological variables. In the proposed architecture, CNN layers extract local temporal patterns, LSTM layers model longer-term dependencies, and the attention mechanism emphasizes diagnostically informative sequence fragments. In preliminary validation on the available experimental dataset, the model achieved an accuracy of 98.6%. Further evaluation using an independent test set and data collected under real farm conditions is required to confirm the robustness and practical applicability of the approach.