为了提高鱼粉品质检测装置的鉴别能力，利用研制的鱼粉品质检测装置，提取鱼粉样本的响应特征信息（10×6个）构成原始特征矩阵，以多层感知器神经网络的鉴别正确率为评价指标，对其传感器阵列进行多特征数据融合优化。首先，通过不同的归一化处理，得到了最佳的归一化处理方法;其次，通过因子载荷分析结果计算获得1 770个特征距离值，按从小到大的顺序对1 770个距离进行排序，并依据特征值距离原点的欧式距离，剔除欧氏距离较小的19个特征值，获得最高的鉴别正确率；最后，对经过载荷分析优化后的原始特征值进行相关性分析，按相关系数绝对值累加和大小进行排序，当剔除掉相关系数绝对值累加和大于37.2时的8个特征值时，此时鉴别正确率为98.3%，特征子集也更紧凑。研究结果表明：特征优化前后的传感器信号的表征特征发生了明显的变化，33个特征值被用来表征鱼粉样本的传感器特征信号。同时，采用马氏距离解释了MLP神经网络鉴别结果的可信性，进一步说明了特征优化方法的合理性。
In order to improve the identification ability of the fish meal quality detection device,the response characteristic information (10×6) of the fish meal sample is extracted to form the original feature matrix,and the multifeature data fusion optimization of the sensor array is carried out by taking the identification accuracy of the multilayer perceptron neural network as the evaluation index. Firstly,through different normalization processing,the best normalization processing method is obtained. Secondly,1 770 characteristic distances are calculated by factor load analysis results,and 1 770 distances are sorted according to the order from small to large. According to the Euclidean distance between the eigenvalues and the origin,19 eigenvalues with smaller Euclidean distance are eliminated to obtain the highest discriminant accuracy. The original eigenvalues optimized by load analysis are correlated and sorted according to the absolute sum and size of correlation coefficients. When 8 eigenvalues are removed when the absolute sum of correlation coefficients is greater than 37.2,the recognition accuracy is 98.3%,and the feature subset is more compact. The results showed that the characterization characteristics of the sensor signals changed obviously before and after feature optimization. 33 eigenvalues were used to characterize the sensor characteristic signals of fish meal samples. At the same time,the reliability of MLP neural network identification results is explained by Mahalanobis distance,which further explains the rationality of feature optimization method.