中文摘要：

为明确自行设计的滚筒式红茶发酵机性能参数,以无量纲化的综合评分为发酵品质评价指标,采用响应面法和基于改进型神经网络的遗传算法（BP-AdaBoost-GA）对影响发酵品质的3个因素（发酵温度、发酵时间、翻拌间隔）进行优化,并对2种方法的优化效果进行比较。结果表明,各因素对发酵品质的影响重要性顺序为：发酵温度、翻拌间隔、发酵时间;采用响应面法优化,当发酵温度、发酵时间、翻拌间隔分别为25℃、150 min、20 min时,综合评分预测值和实际值分别为0.863和0.856,相对误差为0.8%;而采用BP-AdaBoost-GA优化,当发酵温度、发酵时间、翻拌间隔分别为27℃、170 min、25 min时,预测值和实际值分别为0.871和0.868,相对误差为0.3%;BPAdaBoost预测模型的决定系数和相对分析误差分别为0.994和18.456,高于响应面法的0.988和9.577,且预测均方根误差较低,为0.017。在红茶发酵工艺的参数优化中,采用BP-AdaBoost-GA方法能比响应面法更好地拟合模型,以及在全局变量范围内推导最优发酵条件。

英文摘要：

Fermentation is the key procedure in processing of congou black tea, which directly decides the quality and flavor of tea products. Fermentation experiments were conducted on a novel drum-type fermentation machine as the platform, the performance parameters of fermentation machine were clarified. Methodologically, with dimensionless comprehensive scores as a measure of fermentation quality, response surface methodology （RSM） and back-propagation adaptive boosting based genetic algorithm （BP- AdaBoost- GA） were used separately to optimize three parameters （fermentation temperature x1, fermentation time x：, rotational interval x3 ） that affect fermentation quality. Also the optimizing effects of RSM and BP- AdaBoost- GA were compared. Results showed that the importance degrees of the three parameters ranked as x1 〉x3 〉x2. With RSM at xt =25℃ , x2 = 150 min and x3 =20 min, the predicted and actual values of comprehensive scores were 0. 863 and 0. 856, respectively, showing relative error of 0. 8%. With BP-AdaBoost-GA at xI =27℃ , x2 = 170 min and x3 =25 min, the predicted and actual values of comprehensive scores were 0. 871 and 0. 868, respectively, showing relative error of 0. 3%. When the BP- AdaBoost had seven nodes in the hidden layer and a prediction error threshold of 0.25, its determination coefficient was greater than that of RSM （0. 994 vs 0. 988） , and it had lower root mean square error of prediction （RMSEP） of 0. 017 and residual predictive deviation （RPD） equaled to 18. 456. Both RSM and BP- However, the fitting ability of AdaBoost - GA were feasible for optimization of fermentation parameters. RSM was limited because it was based on quadratic polynomial regression,while the fitting ability over experimental data was limited. The algorithm combining improved neural network and GA had higher global extremum prediction ability and higher accuracy. Thus, it can be concluded that even though RSM was the most widely used method for fermentation parameter optimization, BP- A