DEVELOPING A BIOSENSOR WITH APPLYING KALMAN FILTER AND NEURAL NETWORK TO ANALYZE DATA FOR FUSARIUM DETECTION
Date
2020-03-25
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
ORCID
0000-0003-4365-4306
Type
Thesis
Degree Level
Doctoral
Abstract
Early detection of Fusarium is arduous and highly desired as the detection assists in protecting crops from the harmful potential of plant pathogens which affect the quality and quantity of agriculture products. The thesis work concentrates on searching an approach for Fusarium spore detection and developing portable, reliable and affordable Fusarium detection device. The system can also promptly and continuously sample and sense the presence of the fungus spores in the air. From the investigation of the Fusarium oxysporum Chlamydospores by ATR-FTIR spectroscopy, a distinct infrared spectrum of the Chlamydospore was collected. There are two typical infrared wavelengths can be used for Fusarium detection: 1054cm-1 (9.48µm) and 1642cm-1 (6.08µm). Infrared (IR) light is a form of electromagnetic wave which its wavelength ranges from around 0.75µm to 1000µm and it is invisible to human eyes. To be familiar with the light concepts and quantities, it is necessary to start working with the visible light which is also a form of electromagnetic wave with the wavelength range from about 0.3µm to 0.75µm. A visible spectrometer, which automatically corrects data error caused by unstable light, was built. By using the Kalman filter algorithm, Matlab simulations and training program, the appropriate coefficients to apply to the Kalman corrector were found. The experiments proved that the corrector in the visible spectrometer can reduce the data error in the spectra at the order of 10 times.
From the knowledge of working with the visible spectrometer and visible light, the task of searching the detecting approach and building the device in the IR spectrum was reconsidered. The Fusarium detection device was successfully built. Among other components to build the device, there are two essential thermopiles and one infrared light source. The infrared light source emits an IR spectrum from 2µm to 22µm. The two thermopiles working on the IR wavelengths of λ1=6.09±0.06µm and λ2=9.49±0.44µm are used for Fusarium spore detection analysis. The Beer-Lambert assists in quantifying the number of spores in the sample. The group distinction coefficient supports in distinguishing the Fusarium spore from other particulates in the experiments (pollen, turmeric, and starch). Pollen was chosen as it is often present in crop fields, and the other two samples were chosen as they help to verify the work of the system. The group distinction coefficients of Fusarium (1.14±0.15), pollen (0.13±0.11), turmeric (0.79±0.07) and starch (0.94±0.07) are distinct from each other.
The size of Fusarium spore is from about 10µm to 70µm. To mitigate the influence of the other particulates, such as pollens or dust which their sizes are not in the above range, a bandpass particle filter consists of a cyclone separator and a high voltage trap were designed and built. The particles with the sizes not in the interested range are eliminated by the filter. From simulations by the COMSOL Multiphysics and experiments, the particle filter proves that it works well with the assigned particle size range. The filter is useful as it helps to sample a certain size range which contains the interested bio objects.
As other electronic devices, the Fusarium detection device encountered several common types of noises (thermal noise, burst noise, and background noise). These noises along with the thermopile signals are amplified by the amplifiers. These amplifiers have high gain coefficients to amplify weak signals in nV to µV in magnitude. These noises depend on the operating conditions such as power supplies or environment temperature. If the operating conditions can be monitored, the information of the conditions can be used to correct the error data. To perform the correction task, the neural network was selected. To make a NN working, it requires sufficient data to train. In this research, the training data were collected in one week to record as much as possible working conditions. In addition to the thermopile data, the training data also included the environment temperature and the 5V and 9V voltage-regulator data. Then, the trained NN was applied to fix error data. The contribution of this NN method is the use of operating conditions to fix error data.
Although the errors in the data can be corrected well by the trained NN, several other problems still exist. In the samples of Fusarium, starch, pollen, and turmeric, the group-distinction coefficients of Fusarium and starch are very similar. To distinguish better the samples with similar group distinction coefficients, the existing Fusarium detection device was upgraded with a broadband thermopile. The extra thermopile was used along with λ1 and λ2 thermopiles to analyze the reflecting IR light of the samples. To pre-process the thermal noise and burst noise, an adaptive and cognitive Kalman algorithm was proposed. Burst noise is expressed in the form of outliers in the thermopile data. To detect these outliers, a mechanism of using first-order and second-order discrete differentiation of the data and correcting the burst noise and thermal noise was introduced. To study the effectiveness of this pre-processing, the pre-processed data and raw data were applied in the NN training. The main stopping parameters in the training are the number of epochs, absolute mean error, and entropy. The pre-processed data and the trained NN were used for distinguishing samples. The three-thermopile Fusarium detection device led to a use of a validation area to distinguish the samples with similar group-distinction coefficients. The results prove that the use of three thermopiles works very well.
The research provides a comprehensive approach of designing system, particulate sampling, particulate filtering, signal processing, and sample distinguishing. The results from the experiments prove that the proposed approach can detect not only Fusarium but also many other different bio-objects. For further work from this research, the Fusarium detection apparatus should be tested in the crop fields infected by Fusarium spores. The outcomes of the research can be applied in other areas such as food safety and human living or hospital environment to detect not only Fusarium spores but the other pathogens, spores, and molds
Description
Keywords
spectrometer, spectrum, Kalman, filter, noise reduction, corrector, biosensor, Beer-Lambert law, group distinction coefficient, Fusarium spore, thermopile, detection method, air sampler, particle trap, cyclone, high voltage, simulation, neural network, data correction, operating conditions, background noise, burst noise, outlier
Citation
Degree
Doctor of Philosophy (Ph.D.)
Department
Electrical and Computer Engineering
Program
Electrical Engineering