Characterizing the diversity within dairy herds with data mining methods

The expectations of consumers and citizens about well-being in animal breeding are increasing strongly. Sensors present in precision breeding can help objectify the states of animals in a herd by assessing the animals' actual stress levels. The thesis will begin with the study of heat stress. Climate change is characterized, among other things, by episodes of greater heat during the summer, with sometimes very high temperatures, which can be harmful to the health and production of dairy animals. This heat stress is also amplified when the humidity is high, such as during stormy episodes. To take into account the combined effects of these two parameters, the THI (Temperature Humidity Index) is generally used, which is calculated according to the formula: THI = T - ((0.55-0.0055 x U)) x ( T-14.5), with T = air temperature (° C) and U, relative humidity in%. Above a value of 70, discomfort may be felt by animals. Beyond 80, heat stress can be considered severe for dairy cows. Based on this definition, it can be estimated that there are frequent days of intense thermal stress for animals present on most farms.

Today, monitoring tools allow the measurement of real-time data, called time series. In particular, it is possible to follow the evolution of the temperature of an animal over long time steps, both during lactation but also over several lactations. These data make it possible, for example, to observe that the temperature response to stress is not the same for each animal (Fig 1).

These differences could be explained by the temperature threshold level of an animal, its adaptation capacities, individual characteristics (ingestion, production of, weight).
Being able to characterize, but also predict, the response of an animal to heat stress is a real challenge for the health of animals and the maintenance of their performance. To characterize cows heat stress response, we will use discriminant pattern extraction methods, capable of extracting groups whose responses to heat stress differ. In particular, we focus on algorithms for finding out subgroups and characterizing for example groups of cows at risk. However, subgroup discovery algorithms are not integrated to handle sets of time series. A first task is therefore the construction of a method for discovering subgroups making it possible to process time series. Another important task is to define more precisely how to characterize these subgroups in order to find characteristics specific to dairy cows.

Josie Signe has been working on this thesis subject since october 2020 for a period of 3 years. She is supervised by Yannick Le Cozler in the Syslait team, Alexandre Termier and Véronique Masson in the Lacodam team (Irisa), Peggy Cellier in the Semlis team (Irisa).

This work, referred as ANR-16-CONV-0004, was supported by the French National Research Agency under the #Investments for the Future Program”.

Contact

Josie Signe : josie.signe[at]inria.fr (PhD student)
Alexandre Termier : alexandre.termier[at]inria.fr (thesis director)
Yannick Le Cozler : yannick.lecozler[at]inrae.fr (thesis co-director)
Peggy Cellier : peggy.cellier[at]irisa.fr (thesis supervisor)
Véronique Masson : veronique.masson[at]irisa.fr (thesis supervisor)