Animal blood and semen analysis plays a significant role
in national
biological resource management, wildlife conservation, and customs
security quarantine. Traditional blood analysis methods have disadvantages,
such as complex sample preparation, time consumption, and false positives.
Therefore, proposing a rapid and highly accurate analysis method is
highly valuable. Raman spectroscopy has been widely used in blood
analysis, and efficient and accurate analysis results can be obtained
through the machine learning algorithm feature extraction. Recently,
the transformer network structure was applied to Raman spectroscopy
recognition. However, the multihead self-attention mechanism does
not perform well in extracting local feature peaks, although it obtains
global feature relations. This paper proposes a neural network based
on the combination of one-dimensional convolution and multihead self-attention
mechanism (Raman ConvMSANet) to identify 52 species of blood and semen
Raman spectra. The network can achieve reliable identification effects
in multiclassification and sample imbalance situations, and the average
identification accuracy of blood and semen can reach more than 98.5%.
The proposed network model can be applied not only to blood and semen
identification but also to other biological fields.