随着光谱分析及荧光检测技术的快速发展, 单色荧光标记已无法对细胞样本进行精准判断, 必须采用双染色或多色荧光标记来分析细胞内部结构。 然而, 使用光谱测量方法进行多色荧光分析时, 由于通常使用多种标记物同时对待测细胞进行标记, 发射光谱会产生部分光谱重叠, 为了准确对其进行分析, 需将重叠峰分解为独立谱峰。 针对光谱重叠现象, 提出了遗传算法优化BP神经网络(GA_BP)的重叠峰解析算法。 首先确定了BP神经网络具体结构, 并对重叠峰信号进行二次微分预处理, 确定重叠峰中单峰个数及单峰位置, 将其作为重叠峰信号的特征值送入BP神经网络的输入层； 其次将BP神经网络权值及阈值初始化, 利用遗传算法全局搜索的优势, 进行算法初始种群及种群规模等最优参数的选取, 通过选择、 交叉、 变异等一系列遗传进化操作进行寻优计算, 得到包含BP神经网络最优权值和阈值的个体； 然后确定网络最优参数并进行相应网络训练, 使优化后的BP神经网络可从输出节点处获得独立单峰的峰宽及强度； 最后结合二次微分处理得到的重叠峰特征值, 即可分离出单个谱峰。 以随机生成的多组高斯重叠峰数学模型作为实验数据进行仿真实验, 结果表明该方法具有较高的精确度。 其中, 双峰重叠峰及三峰重叠峰分解后峰强度及峰宽的最大相对误差分别为0.30%, 3.57%和0.64%, 3.83%； 同时也可对四峰重叠峰进行较为准确的分解。 此外, 将GA_BP网络模型与未经优化的BP神经网络模型作对比, 结果表明GA_BP网络运行5步后即可达到预设的误差值, 而未经优化的网络模型则需19步方可达到, 进一步证明GA_BP网络模型收敛更快且误差较低。 由此可见, GA_BP算法在重叠光谱分析中有较好的效果, 并可应用于其他能谱重叠峰的分解, 与传统方法相比具有明显的优势, 具有一定的实用价值。

With the rapid development of spectroscopy and fluorescence detection technology, monochrome fluorescence labeling is unable to analyze cell samples accurately and has been gradually replaced by two-color or multi-color fluorescence labeling. In the multicolor fluorescence analysis, since the cells were labeled with a variety of fluorescein usually, partial spectral overlap will occur in the emission spectrum, which need to be decomposed into an independent spectral peak to analyze accurately. Aiming at it, optimized BP neural network based on genetic algorithm (GA_BP) were used for overlapping spectral peak analysis. Firstly, the concrete structure of BP neural network was determined, and the overlapping peak was pre-processed by quadratic differential to find out the number and positions of single peaks as the characteristic value of overlapping peaks to be the input layer of BP neural network; in addition, weights and thresholds of BP neural network were Initialized, and optimal parameters like initial population and population size of the genetic algorithm were selected by using the advantage of global search; after a series of genetic evolution operations like selecting, crossing and mutating, the individuals containing the optimal weights and thresholds of BP neural network were obtained; and then the optimal parameters of the network were selected to carry out network training, which the width and intensity of the independent peak can be calculated from the output node of the optimized BP neural network; finally, combined with the eigenvalues of overlapping peak identified by quadratic differential, independent spectral peak can be separated. The randomly generated Gaussian overlapping peaks model was used as experimental simulation data, and the decomposition experiments showed high precision of the peak intensity and peak width. Wherein, the maximum relative error of decomposition of two overlapping peaks was 0.30% and 3.57%, and which of the three overlapping peaks was 0.64% and 3.83%. It can also be decomposed when the four overlapping peaks. Moreover, compared the GA_BP network model with the unoptimized BP neural network model, the results showed that the GA_BP network could reach the preset error value after five steps, while the unoptimized network model takes 19 steps. This further proves that the GA_BP network model converged faster with a fairly high precision that can be widely used for the decomposition of spectral and other overlapping peaks, which has a certain practical value compared with traditional methods.