Quantitative Analysis of Artificial Neural Networks for Spectral Measurement

　　The quantitative issue of artificial neural nrtworks ( ANN) had been addressed by using examples of absorption and fluorescent spectra. In this report, we proposed a new ANN architecture which combines nonlinear and linear process elements (neurons) for better quantitative results. We used MATLAB and MATHENADTICA to simulate several important aspedts of spectral measurement, e.g. normalization method, interference from untrained components and noise, and verified the feasibility of this model for quantitative applications, such as interpolation and extraploation. With generalized delta rule and back error propagation, the network can identify the principal components within the training spectra and evolve toward the convergent state with the information of principal components being stored in the weight motrices. The convergent weight matrices can then be used for the prediction of untrained spectra in either interpolating or extrapolating manners with quantitative output. The simulation results indicate that: 1. Euclidcan norm method can be used to identify the principal components. This normalization method will have better replications of input patterns than the infinite norm method . Intuitively, Euclidean norm will allow cach neuron control its total output and let the connection weights grow with maximum variations in the input patterns. 2. The proposed new architecture improves the linearity of the network's output response. 3. The trained netwoek has noise- tolerant capability which has been tested against the interference of extra untdrained component and added random moise. 4. The trained netwoek is capable of doing interpolation and linited range of extrapolation.#0578#