Doctor of Philosophy
Sarah C. Rutan
In this research, a hidden node pruning algorithm was developed for an artiﬁcial neural network (ANN) that automatically determined a more efﬁcient size of the hidden layer, caused the ANN to resize itself, and then continued to train using a standard back-propagation algorithm. The hidden-node pruning algorithm was based on determining the number of signiﬁcant eigenvalues present in the matrix of values produced by the hidden layer, starting with an excessive number of hidden nodes.
Eight sulfonylurea herbicides were used as the target analytes in this study. The ability of an ANN to simplify the sample preparation needed for analysis using a liquid chromatograph/particle beam/mass spectrometer (LC/PB/MS) was evaluated.
The results derived from this research demonstrated that ANNs allow the clean-up procedure to be simpliﬁed, while still obtaining reliable identiﬁcation of the sulfonylurea herbicides in complex matrices such as soil. Speciﬁcally, this was accomplished by using retention times from the LC and MS when the herbicides were injected individually in pure forms combined with MS data obtained from extracted samples. This information was used by a trained neural network to identify sulfonylurea herbicides as both individual components and components in a mixture.
Two different neural networks were created. One was trained with a single mass spectrum from each herbicide, resulting in an 8-training-sample network, and one was trained with ﬁve mass spectra of each herbicide, resulting in a 40-training-sample network. Both ANNs had 47 input nodes and eight output nodes. Starting with an excess of 20 hidden nodes, the networks resized themselves to contain 6 hidden nodes for the 8-training-sample network and 7 hidden nodes for the 40-training-sample network. An optimum sum-squared error (SSE) goal was determined to be 0.3 for the 8-training-sample network by using a statistical t-test . to establish the smallest SSE where the standard error of prediction was not signiﬁcantly greater than the standard error of calibration. Results demonstrated that the 8-training-sample ANN performed just as well as the 40-training-sample ANN. When compared to the Hewlett-Packard probability-based matching (HP- PBM) library searching system, both neural networks out-performed the HP-PBM system in the identiﬁcation of unknown mass spectra.
© The Author
Is Part Of
VCU University Archives
Is Part Of
VCU Theses and Dissertations
Date of Submission