| In order to improve the efficiency of process detection and control of solid-state fermentation (SSF), this work attempted to the feasibility and method of the measurement of process parameters of SSF of protein feed by use of near-infrared spectroscopy (NIRS) techniques. In addition, the pattern recognition of process state of SSF was also focused on by use of NIRS, electronic nose, and multi-sensors information fusion technique in this work. The main research points are summarized as follows:(1) The feasibility and method for the measurement of process parameters of SSF were studied by use of NIRS technique. Firstly, the raw spectra of all fermented samples obtained were preprocessed by use of the first derivative. Secondly, several efficient subintervals were selected by use of the synergy interval partial least squares (siPLS) algorithm. Then, some efficient wavelength variables were selected by use of the genetic algorithm (GA) from the subintervals obtained. Lastly, the partial least squares (PLS) model was developed by use of the efficient wavelength variables selected for the measurement of process parameters (i.e. pH and moisture content) of SSF. Experimental results showed as follows:For the parameter of pH, the optimal detection model was achieved when seven principal components (PCs) included based on the45efficient wavelength variables selected. The result of the root mean square error cross-validation (RMSECV) is0.0583and the correlation coefficient (Rc) is0.9878in the training set. When the performance of the best model is evaluated by the independent samples in the validation set, the result of the root mean square error prediction (RMSEP) is0.0779and the correlation coefficient (Rp) is0.9779in the validation set. For the parameter of moisture content, the optimal detection model was achieved when four PCs included based on the53efficient wavelength variables selected. The result of RMSECV is1.3286%w/w and Rc is0.8992in the training set. When the performance of the best model is evaluated by the independent samples in the validation set, the result of RMSEP is1.2668%w/w and Rp is0.8700in the validation set. The overall results demonstrate the potentials of NIRS technique in rapid measurement of process parameters of SSF. Additionally, it is necessary to select characteristic wavelength variables of near-infrared spectra in model calibration. It can effectively reduce the complexity and improve generalization performance of the detection model when NIRS technique is used for on-line detection of the process parameters of SSF.(2) The pattern recognition for the process state of SSF was studied by use of NIRS technique. Firstly, the raw spectra of all fermented samples obtained were preprocessed by use of the discrete wavelet transform (DWT), and the feature vectors were extracted by use of principal component analysis (PCA) from the spectral data preprocessed. Then, the identified model was developed by use of support vector data description (SVDD) algorithm, which is a one-class classification method. Simultaneously, four traditional two-class classification approaches (i.e. linear discriminant analysis, LDA;K-nearest neighbor, KNN; back propagation neural networks, BPNN; and support vector machine, SVM) were comparatively utilized for monitoring time-related changes that occur during SSF. Experimental results showed that as follows:When the number of samples from the target class and those from non-target class in the training set is equal, all identification models can achieve good classification performance in the validation set. However, the SVDD model could reveal its unique superiority in disposing the imbalance training sets. When the ratios achieve one to four and one to eight, the identification rates of SVDD model are95%and90%in the validation set, respectively. Nevertheless, in the same condition, the classification rates of the other four models are all under70%. The overall results demonstrate that SVDD algorithm is a prominent approach in developing one-class classification model with imbalance training set, and NIRS technique combined with SVDD has high potential to monitor the process state of SSF in a no-invasion way.(3) The pattern recognition for the process state of SSF was studied by use of electronic nose technique. The electronic nose technique, with the help of chemometrics analysis, was attempted in this work. Firstly, the feature vectors were extracted by use of PCA, and the number of PCs were optimized by a five-fold cross-validation in model calibration. Then, LDA, KNN, and SVM respectively were used to calibrate identification models in order to evaluate the influences of different linear and non-linear classification algorithms on the classification performance. Experimental results showed as follows: Investigated from PCs scores plot, seven sample groups appeared in cluster trend along two principal component axes, confirming the presence of seven different clusters just associated with their condition of fermentation. Especially, the samples from the initial stage of SSF of protein feed could be separated directly by PCA. In addition, the identification accuracy of SVM model was superior to those of the other two. and the optimal SVM model was obtained when five PCs were included. The identification rates of the SVM model were97.14%and91.43%in the training and validation sets, respectively. The overall results sufficiently demonstrate that the electronic nose technique coupled with an appropriate chemometrics method could be successfully used in identification of process state of SSF.(4) The pattern recognition for the process state of SSF was studied by use of multi-sensors information fusion technique. Firstly, the raw spectra and electronic nose signals were preprocessed, and some initial feature information were extracted by use of traditional methods. Secondly, PC A was implemented on these initial feature information extracted from near-infrared spectra and electronic nose signals preprocessed, and then PCs vectors were extracted and optimized as the inputs of pattern recognition based on the single technique (i.e. NIRS or electronic nose technique) in model calibration. Lastly, the optimal PCs on the single technique model were fused in feature extraction level fusion by use of independent component analysis (ICA), and the best BP_Adaboost model was developed by use of the optimal independent components (ICs). Experimental results showed that the optimal fusion model of BP_Adaboost based on NIRS and electronic nose techniques was obtained when four ICs included. The identified rate equaled to99.05%in the training set, and94.29%in the validation set. Compared with the best BP_Adaboost model based on the single technique, the identified results of fusion model on the two techniques are much better than those of the single technique model both in the training and validation sets, and the complexity of the fusion model was also less than that of the single technique model. The overall results demonstrate that it is feasible to identify the process state of SSF with multi-sensors information fusion technique. The identification accuracy and stability of the recognition model from the multi-sensors information fusion were better than those of the recognition model from the single-sensor information.This study provides a new idea for the process detection and control of SSF. The main aim of improving the accuracy and timeliness for the measurement of process parameters and the monitoring of process state of SSF has been achieved. The results in this work can provide research foundation for developing instruments and equipment for the monitoring of SSF process. |
PDF Full Text Request