Detecting lameness in sows from ear tag-sampled acceleration data using wavelets

The objective of this study was to develop an automated monitoring system to detect lameness in group-housed sows early and reliably on the basis of acceleration data sampled from ear tags. To this end, acceleration data from ear tags were acquired from an experimental system deployed at the Futterkamp Agriculture Research Farm from May 2012 until November 2013. The developed method performs a wavelet transform for each individual sow's time series of total acceleration. Feature series are then computed by locally estimating the energy, variation and variance in a small moving window. These feature series are then further decomposed into uniform level sets. From these series of level sets, the highest and lowest levels are monitored for lameness detection. To that end, they are split into a past record to serve as reference data representing a sow's expected behaviour. The deviations between the reference and the remaining detection part of current data, termed feature activated, were then utilised to possibly indicate a lameness condition. The method was applied to a sample of 14 sows, seven of which were diagnosed as lame by a veterinarian on the last day of the sampling period of 14 days each. A prediction part of 3 days was set. Feature activated were clearly separable for the lame and healthy group with means of 8.8 and 0.8, respectively. The day-wise means were 1.93, 9.47 and 15.16 for the lame group and 0.02, 1.13 and 1.44 for the healthy group. A threshold could be set to completely avoid false positives while successfully classifying six lame sows on at least one of the 2 last days. The accuracy values for this threshold were 0.57, 0.71 and 0.78 when restricting to data from the particular day. A threshold that maximised the accuracy achieved values of 0.57, 0.79 and 0.93. Thus, the method presented seems powerful enough to suggest that an individual classification from ear tag-sampled acceleration data into lame and healthy is feasible without previous knowledge of the health status, but has to be validated by using a larger data set.