Lameness Detection in Cows Using Hierarchical Deep Learning and Synchrosqueezed Wavelet Transform

Objectives: Identification of cow lameness is important to farmers improve and manage cattle health welfare. No validated tools exist for automatic detection. In this research, we aim early detect the by identifying instantaneous fundamental gait harmonics from low frequency (16Hz) acceleration signals recorded using leg-worn sensors. Methods: A triaxial accelerometer has been worn on each leg. Synchrosqueezed wavelet transform (SSWT) applied generate initial time-frequency spectrum related gait. This given as an input a designed deep neural network including based long short-term memory (LSTM) estimate frequencies at time point. An inverse SSWT (ISSWT) then used recover harmonic enhanced spectrum. Results: Validation provided leg (combined 23 cows) time-series cross validator across three folds are provided. The average mean squared errors in 3 obtained 0.036, 0.033, 0.044 0.042 left-front, right-front, right-back left-back legs, respectively. Conclusion: Estimation proved useful identification phases, detection, accurate estimation speed, coherency movement among legs non-gait episodes. Moreover, proposed method can be new ridge exploiting many other applications.