蜉蝣算法是一种以蜉蝣生物的飞行和社会行为为参照的新型元启发式算法，算法将群体分为雌、雄个体种群，并且采用不同的方法来更新雄性和雌性的速度和位置，对于一些雌性个体，它们的更新状态取决于雄性个体。蜉蝣算法具有可伸缩性、计算效率高、可扩展性强的特点。它的搜索空间更小，可以有效地解决复杂的优化问题，自提出受到广大学者的青睐，并广泛应用于工程问题。但是该算法仍然存在易陷入局部最优、稳定性有待增强等问题，针对存在的不足之处，本文提出了两种改进的蜉蝣算法，具体内容如下：

首先针对蜉蝣算法收敛性能欠佳和稳定性不足，提出一种基于混沌初始化和模拟退火机制改进的蜉蝣算法，该算法使用Logistic混沌映射初始化种群来提高解的质量，并引入模拟退火机制改进个体速度更新方式，以一定的概率接受部分个体的速度更新，增强算法寻优能力和稳定性。通过在7个不同维度测试函数上的仿真实验，验证了改进后算法的稳定性和收敛速率得到提高。其次基于传统蜉蝣算法位置更新缺乏样本多样性支撑，导致算法早熟，陷入局部最优，且随机性较弱，尤其在搜索后期，算法需要一定的变体用于跳出局部最优，并且蜉蝣算法中的多群体机制虽然一定程度上提升了搜索能力，但是复杂的参数和个体会使得搜索和开发失去平衡，降低算法的稳定性，造成个体间信息交流不足。为了解决以上问题，提出了一种黄金模拟交叉变异蜉蝣算法。算法保留已引入的模拟退火机制改进速度更新方法，并使用黄金正弦算法改进位置更新方式，得以缩减搜索空间，平衡算法的全局搜索和局部搜索能力，之后引入不同的自适应交叉变异算子替换固定算子，增加算法在后期的样本多样性和跳出局部最优的能力。通过在12个不同维度和模态的测试函数上进行仿真实验，验证了算法在快速收敛、求解精度和鲁棒性方面有显著优势。

本文基于黄金模拟交叉变异蜉蝣算法优化支持向量机构建新的集成分类模型，选取不同的UCI数据集和性能指标进行验证和评价。基于五位受试者的P300脑电信号数据，采取频段截取、低高通滤波、ICA算法过滤降噪等预处理过程，并提出一种多时频域融合特征提取方法，使用建立的集成分类模型对五位受试者数据进行数值实验，对比结果验证了所建立的模型具有更佳的分类性能。

The mayfly algorithm is a novel metaheuristic algorithm with reference to the flight and social behavior of mayfly organisms. The algorithm divides the population into populations of female and male individuals and uses different methods to update the speed and position of males and females, and for some females, their update status depends on the males. The mayfly algorithm has the characteristics of scalability, high computational efficiency and scalability, and its smaller search space can effectively solve complex optimization problems, and has been favored by the majority of scholars since it was proposed and widely used in engineering problems. However, the algorithm still has problems such as easy to fall into local optimum, stability to be enhanced, etc. For the existing shortcomings, this paper proposes two improved mayfly algorithms as follows:

Firstly, to address the poor convergence performance and stability of the mayfly algorithm, an improved mayfly algorithm based on chaotic initialization and simulated annealing mechanism is proposed, which uses the logistic chaotic mapping to initialize the population to improve the quality of the solution, and introduces a simulated annealing mechanism to improve the individual velocity update method, accepting the velocity update of some individuals with a certain probability to enhance the algorithm's merit-seeking ability and stability. Simulation experiments on 7 different dimensional test functions demonstrate that the improved algorithm's stability and convergence rate is improved. Secondly, based on the traditional mayfly algorithm location update lacks sample diversity support, which leads to premature algorithm and falls into local optimum, and the randomness is weak, especially in the late stage of the search, the algorithm needs certain variants for jumping out of local optimum, and although the multi-population mechanism in the mayfly algorithm improves the search capability to some extent, the complex parameters and individuals will make the search and development lose balance, reduce the stability of the algorithm, and cause insufficient information exchange among individuals. In order to solve the above problems, a golden simulated cross-mutation mayfly algorithm is proposed. The algorithm retains the introduced simulated annealing mechanism to improve the velocity update method and uses the golden sine algorithm to improve the position update method, which is able to reduce the search space and balance the global and local search ability of the algorithm, after which different adaptive cross-mutation operators are introduced to replace the fixed operators to increase the sample diversity and the ability of the algorithm to jump out of the local optimum at a later stage. Through simulation experiments on 12 test functions with different dimensions and modes, the algorithm is demonstrated to have significant advantages in terms of fast convergence, solution accuracy, and robustness.

In this paper, a new integrated classification model is built based on the golden simulated cross-variance mayfly algorithm optimized support vector machines, and different UCI datasets and performance metrics are selected for validation and evaluation. Based on the P300 EEG signal data of five subjects, the pre-processing processes such as frequency band interception, low and high-pass filtering, and noise reduction by ICA algorithm filtering are adopted, and a multi-time-frequency domain fusion feature extraction method is proposed. Numerical experiments are conducted on the data of five subjects using the established integrated classification model, and the comparison results verify that the established model has better classification performance.