Liu Tao, Ren Zhong, Wu Junli, Xiong Chengxin, Peng Wenping
Key Laboratory of Optic-Electronic and Communication, Jiangxi Science and Technology Normal University, Nanchang, China.
Key Laboratory of Optic-Electronic Detection and Information Processing of Nanchang City, Jiangxi Science and Technology Normal University, Nanchang, China.
J Biophotonics. 2022 May;15(5):e202100309. doi: 10.1002/jbio.202100309. Epub 2022 Jan 25.
To accurately identify the blood authenticity, a set of photoacoustic detection system was established. In experiments, five kinds of blood in total of 125 groups were used, the time-resolved photoacoustic signals and peak-to-peak spectra were obtained in 700 to 1064 nm. Experimental results showed the accurate identification of blood authenticity was limited due to overlap of signals and spectra. To solve the problem, wavelet neural network (WNN) was employed to supervised train peak-to-peak spectra of 100 samples. The correct rate was 72% for 25 test samples. To improve correct rate, the parameters of WNN were optimized by quantum-behaved particle swarm optimization (QPSO) algorithm. Meanwhile, the effects of neurons number, learning rate factors, iteration times and training times on correct rate were studied and compared with WNN and WNN-PSO algorithms. Results showed the correct rate of WNN-QPSO was increased to 96%. Then, three kinds of dynamic contraction-expansion coefficients were used. Under the optimal dynamic coefficient, the correct rate reached 100%. Moreover, the truncated mean stabilization strategy (TMSS) was coupled to improve the convergent speed. Finally, 10 algorithms were compared. Results demonstrated that photoacoustic spectroscopy combined with WNN-QPSO coupled with TMSS and dynamic contraction-expansion coefficient had an excellent performance in the identification of blood authenticity.
为准确鉴别血液真伪，搭建了一套光声检测系统。实验中，共使用了5种血液，总计125组，在700至1064nm范围内获取了时间分辨光声信号和峰峰值光谱。实验结果表明，由于信号和光谱的重叠，血液真伪的准确鉴别受到限制。为解决该问题，采用小波神经网络（WNN）对100个样本的峰峰值光谱进行监督训练。25个测试样本的正确率为72%。为提高正确率，采用量子行为粒子群优化（QPSO）算法对WNN的参数进行优化。同时，研究了神经元数量、学习率因子、迭代次数和训练次数对正确率的影响，并与WNN和WNN-PSO算法进行比较。结果表明，WNN-QPSO的正确率提高到了96%。然后，使用了三种动态收缩-扩张系数。在最优动态系数下，正确率达到了100%。此外，结合截断均值稳定策略（TMSS）以提高收敛速度。最后，对10种算法进行了比较。结果表明，光声光谱结合WNN-QPSO、TMSS和动态收缩-扩张系数在血液真伪鉴别方面具有优异性能。
