Sow Estrus Diagnosis from Sound Samples Based on Improved Deep Learning

The inability to diagnose sow estrus at scale has become the burning issues in the livestock farming. An intelligent screening tool would be a game changer. Changing the dependence on the traditional method of artificial detection, this paper proposes, develops and uses an Artificial Intelligence-based screening solution for sow estrus detection. Sow estrus call is one of the significant calls which can reflect the physiological conditions. This makes the recognition of sow estrus call an exceedingly challenging problem. This problem is addressed by investigating the distinctness of Mel spectrum in the sound system generated in estrus when compared to other pig sounds. To overcome the lack of sow estrus sounds training data, transfer learning is exploited. To reduce the risk of misdiagnosis we raise a two-pronged AI architecture, one is collecting and transmitting sound to the detector, the other is identifying calls in sow estrus diagnosis system. Results show the AI architecture can distinguish among sow estrus calls and other non-sow-estrus-calls. The performance is good enough to encourage a large-scale collection of labeled sow estrus call data to improve the generalization capability of the AI architecture.