Research On Open-set Recognition Of Organic Pollutants In Water Based On 3D Fluorescence Spectroscopy

Water quality safety is closely related to the national economy and people’s livelihood.With the rapid advancement of industrialization,water pollution by organic matter occurs from time to time.As a result,it is important to establish a fingerprint library of organic pollutants,monitor the drinking water supply system,and identify pollutants in time.Three-dimensional(3D)fluorescence spectroscopy has attracted more and more attention in the field of water quality detection because of its advantages such as high sensitivity,excellent selectivity,fast response,and low detection limit.However,in the current qualitative research on 3D fluorescence spectroscopy,there are still some problems.Unknown pollutants are easy to be misjudged as pollutants in the library.Pollutants with weak fluorescent signals are difficult to detect.Multiple pollutants are difficult to distinguish due to intertwined and overlapping fluorescence spectroscopy.In response to the above problems,the research on the open-set recognition method of organic pollutants in water based 3D fluorescence spectroscopy is carried out in this thesis,and independent feature extraction methods and pollutant discrimination methods are proposed,which realized the goal of quickly and accurately recognizing pollutants in the library and reject pollutants outside the library.The main work and innovation points of this thesis are as follows:(1)The features obtained by some feature extraction methods are not independent,and the closed-set classification method would misjudge the unknown class as the known class.Besides,the fluorescence signal of organic matter at low concentration is weak and difficult to be detected,affecting the recognition rate.Given the problems above,an open-set recognition model of organic matter based on 3D fluorescence spectroscopy is proposed.According to the representative characteristics of the fluorescence spectroscopy such as the position,shape and texture,a network called CoordConv consisting of the position-coding module and the convolution-scanning module is established,which extracts the key independence features and effective high-level abstract information.The key features are combined with the cosine similarity measurement method to realize the open set discrimination.The experimental results show that this method achieves the open-set recognition target of recognizing pollutants in the library and reject pollutants outside the library,and improves the detection ability of low-concentration pollutants.(2)In order to solve the problem that the spectra of multi pollutions are intertwined and overlapped,which may affect the recognition effect,an optimization model for open-set recognition of organic pollutants based on attention mechanism and the extreme value machine is proposed.By introducing the CBAM attention mechanism to the CoordConv network,the fine-grained features are extracted from the perspectives of channel attention and spatial attention.So the model can learn more features with distinguishing ability.Meanwhile,the extreme value machine method is introduced to realize the open-set determination of pollutant categories from the perspective of the sample probability distribution,so as to avoid the problem that the threshold value of the cosine similarity measurement method is sensitive and the model performance is easily affected.The experimental results show that this method has a great ability to recognize the intertwined and overlapped spectra,reduce samples and is insensitive to parameters.In addition,for the substance to be recognized,the model can give the probability of each type of pollutant and unknown pollutant in the library,so as to avoid a single decision and provide a reference for subsequent emergency treatment.(3)There are differences in the spectral distribution of samples collected from tap water in different time,regions and water bodies.In this thesis,the adaptability of the open-set recognition model is investigated by designing experiments.The results show that the model has a certain degree of adaptability,but the false alarm rate is still too high in different water bodies.In response to this problem,a new idea of transfer learning is proposed.The model transfer method and Bottleneck feature transfer strategy were adopted to transfer the model trained on tap water pollutant samples to the application of river pollutant recognition.In this thesis,a fingerprint library of 18 types of organic pollutants based on 3D fluorescence spectroscopy is established.An open-set recognition model of organic pollutants which is adaptable within a certain application range is proposed.The model realizes the extraction of key features in low-concentration samples and intertwined overlapping spectra,having the ability to recognize pollutants in the library and reject pollutants outside the library.The research results of this thesis will provide solutions and technical support for the promotion of environmental intelligent perception and the protection of water health.At the same time,it also provides some reference for the open-set recognition in other fields.