Research On The Prediction Model Of Brown Tide Algae Concentration Combined With Deep Learning In The Improved Gramian Angle Field

With the rapid development of the global economy,improper discharge of industrial and agricultural wastewater,increasingly serious seawater pollution and the frequent occurrence of algal blooms,the aquatic economy and the human living environment have caused enormous harm.Therefore,accurate and rapid detection of algal blooms is particularly important.Traditional detection techniques are complex and time-consuming to operate,and fluorescence spectroscopy has the characteristics of accuracy,speed,sensitivity,simplicity,and pollution-free.Therefore,laser-induced fluorescence spectroscopy was used to collect the spectrum of the algal of brown tides.The characteristic bands of spectral information were selected for information conversion,The main work of this study is as follows:Firstly,based on the principle of fluorescence generation and the corresponding characteristic fluorescence bands of different algae fluorescent pigments,the relationship between fluorescence intensity and the concentration of brown tide algae was established according to Lambert-Beer’s law.It was found that when the concentration was low,the fluorescence intensity had a linear relationship with the concentration of algae cells,and when the concentration exceeded a certain range or the intensity of excitation was too strong,both would exhibit a non-linear correlation;A collection scheme for culture spectral information and concentration data of aureococcus anophagefferens was designed,and preprocessing operations were performed on the spectral.The correlation trends of the characteristic bands and spectral peaks of the algal fluorescence spectra with concentration changes were analyzed.It lays a theoretical foundation for building a concentration prediction model of brown tide algae and provides corresponding data preparation.In order to combine 1D spectral sequences with visual neural networks,Gramian angular field(GAF)was used to convert 1D spectral sequences into 2D images,and the improvements were made to address the shortcomings in the processing spectral data at different concentrations.Firstly,the internal normalization method of GAF was improved from local normalization to global one.Secondly,a new conversion method that adds weighting factors to control the main and sub diagonal features is constructed.The improved Gramian angle field(IGAF)can effectively convert the spectra of different concentrations of the algae into 2D image matrix,and its spectral information can effectively correspond to the characteristics of the image changes.The depthwise separable convolutional neural network(DSC)is used to extract features from converted images.This network separates spatial feature learning and channel feature learning,which can effectively reduce the parameters and improve the efficiency of the network.Finally,combining the DSC and long short-term memory network(LSTM)to realize high-precision regression of the algae concentration,R~2 was 97.8%,with the lowest MSE and MAE.Compared with traditional methods,the proposed method has faster convergence and higher prediction accuracy.Aiming at the problem that convolutional neural network(CNN)shares weights in image feature extraction,and feature maps of different channels are assigned the same weight,which makes it impossible to distinguish the contribution of feature maps of each channel to the model.A efficient channel attention module(ECANet)is proposed to adaptively allocate attention weights to the feature maps.Feature maps that contribute more to the model will be assigned higher weight values.The model is relatively lightweight,reducing network complexity while improving the efficiency of the model.Then,the feature was input into the recurrent neural network(RNN)to encoding and then decoding by the attention mechanism to prediction the concentration.This method can integrate RNN with attention mechanism prediction results to improve the accuracy of prediction,R~2 was 98.5%.This method provides a reliable technical reference for online monitoring of brown tide algae concentration,and can be applied to quantitative analysis of substances in other fields.