Study On Use Of Fourier Nir Spectrometer For Both Model Transfer And Pharmaceuticals Identification Method

As pharmaceuticals have great implications for public health, there is a strong application need for identifying genuine. and counterfeit pharmaceuticals and classifying pharmaceuticals in pharmaceutical supervision procedures. Fourier NIR spectrometer is a precision measurement device that combines optic, mechanical, and electrical technologies, allowing rapid and non-destructive on-site tests. With the adoption of statistics or stoichiometry methods, it is often used to measure various physical or chemical values, and has become a necessary device for mobile pharmaceutical test vehicles in China in recent years. In a typical pharmaceutical identification application, over a hundred instruments may be used at the same time. Hence, this paper studies factors that may impact inter-instrument agreement and describes relevant model transfer methods, with a focus on the algorithms for identifying two or more types of pharmaceuticals.This paper starts by introducing different types of NIR spectrometers,operation principle of Fourier transform, and the basic process of NIR spectrum analysis and application. These are followed by discussions on wavelet transform spectrum pre-processing method and basic principle of spectral feature extraction methods such as Autoencoder network.This paper then describes the mechanical structure of Michelson interferometer, which is the core of the Fourier NIR spectrometer, and analyzes both mechanical and environmental factors contributing to spectral detection errors. Next, a model transfer method combining wavelet transform spectrum pre-processing method and simple linear regression direct standardization (SLRDS) algorithm is studied. The test results show that introduction of wavelet transform helps eliminate measurement errors caused by such mechanical and environmental factors related to instruments, thus improving model transfer effect.In this paper, a binary classification algorithm called wSDAGSM is proposed for pharmaceutical identification, combining Sparse Denoising Autoencoder and Gaussian process. In this algorithm, continuous one-dimensional wavelet transform is first performed on spectral data,and then followed by a binary classification based on Sparse Denoising Autoencoder and Gaussian process. According to the test results,compared to BP neural network and other algorithms, the modeling method wSDAGSM proposed in this paper exhibits better performance in terms of classification accuracy and stability. The tests also prove that wavelet transform is a better way to eliminate spectral noise.This paper also proposes a binary classification and multi-classification algorithm called wSDAMRBF for pharmaceutical identification, combining a Sparse Denoising Autodecoder and Support Vector Machine (SVM). In this algorithm, continuous one-dimensional wavelet transform is first performed on spectral data, and then followed by a binary classification and multi-classification based on Sparse Denoising Autoencoder and SVM. Comparative tests are performed here on wSDAGSM and wSDAMRBF algorithms. The results show that both algorithms can achieve a good pharmaceutical identification effect,though wSDAMRBF algorithm provides even better classification accuracy and more stable results.