优化BP神经网络提高高光谱检测调理鸡肉菌落总数精度

针对调理鸡肉菌落总数在贮藏期间易受到外界因素影响,提出了一种优化反向传播(back propagation,BP)神经网络的调理鸡肉菌落总数预测方法。以贮藏在4℃条件下的调理鸡肉为研究对象,采集其表面400~1 000 nm高光谱信息共计419个波段作为全波段,并利用竞争性自适应重加权(competitive adaptive reweighted sampling,CARS)算法筛选出34个特征波段,分别以全波段和特征波段对应的光谱值作为BP神经网络输入,采用鸟群算法(bird swarm algorithm,BSA)和免疫算法(immune algorithm,IA)优化BP神经网络的初始权重和阈值,建立调理鸡肉菌落总数的BP、BSA-BP、IA-BP、BSA-IA-BP预测模型。试验结果表明:经过CARS筛选特征波长的BSA-IA-BP模型预测效果最佳,预测集相关系数RP、均方根误差、剩余预测偏差分别为0.93、0.31lg(CFU/g)、2.68,且模型稳定性最好。该研究为基于BP神经网络实现调理鸡肉菌落总数快速无损检测提供了算法支撑和理论基础。

Improving hyperspectral detection accuracy of total bacteria in prepared chicken using optimized BP neural network

Meat spoilage is a relatively complicated process, in which microorganisms increase nonlinearly. As a non-linear model, BP neural network has strong generalization ability and fitting ability, but there are some shortcomings in the application, such as slow convergence speed, easy to fall into local minima and overfitting. Thus an optimized BP neural network was proposed. Prepared chicken was stored in a refrigerator at 4 ℃, and 240 samples were collected. After obtaining 400-1000 nm hyperspectral images of each prepared chicken sample, sub-samples were randomly selected from each homogenized sample to determine the total bacteria. Then, the spectral data was preprocessed by different methods such as differentiation, standard normalized variate, and multiplicative scatter correction. The PLSR model was cross-validated by the leave-one-out method, and the best preprocessing method was determined based on RMSECV(root mean square error of cross validation). After that, based on the pre-processed spectral information, 34 characteristic bands were extracted by CARS(competitive adaptive reweighted sampling) algorithm. Finally, the spectral values corresponding to the full-band and filtered characteristic bands were used as the input of the BP(back propagation) neural network, and the total bacteria was used as the output of the BP neural network. Bird swarm algorithm(BSA) and immune algorithm(IA) optimization were used to optimize the initial weight and threshold of the BP neural network. The prediction models of the total bacteria were established by using BP, BSA-BP, IA-BP, and BSA-IA-BP. The results showed that: 1) by introducing the IA algorithm’s immune operation, after iterative stabilization, the total fitness of BSA-IA-BP was significantly lower than BSA-BP based on training samples. This showed that the search ability of the BSA-IA fusion algorithm was improved, which could effectively prevent the BSA algorithm from falling into a local optimum in the later stage. At the same time, among the four models of BP, IA-BP, BSA-BP, and BSA-IA-BP, the BSA-IA-BP model had the best prediction accuracy and convergence speed. Among them, the BSA-IA-BP model in the characteristic band had the highest prediction accuracy. The RP(the correlation coefficient), RMSEP(the root mean square error) and RPD(the residual predictive deviation) of the prediction set was 0.93, 0.31 lg(CFU/g), 2.68, respectively. 2) By comparing the characteristic band and the full band, the overall prediction effect of the characteristic band was better than the full band, which indicating that the CARS algorithm could effectively delete the wavelengths, reduced redundant information interference, and improved the model prediction efficiency. In general, the use of hyperspectral technology for non-destructive testing of the total bacteria in prepared chicken was feasible, which can provide technical support for the online testing of prepared chicken.