Quantity Prediction Of Mixed Freshwater Fish Based On Passive Underwater Acoustic Signal

Freshwater fish farming is an important industry in China,and the output of freshwater fish is increasing year by year.One of the most important problems in fishery farming is the estimation of fish quantity.Traditional estimation methods are mainly applicable to oceans and large lakes,but lack of means for estimating fish quantity in general freshwater farming.In this paper,two freshwater fishes,breams and crucians,were taken as the research objects.The sound signals of the breams and crucians were collected at 7 proportions and 15 quantitative gradients.After normalization and different filtering methods,Proportional recognition model and quantitative prediction model of breams and crucians were established respectively.The main conclusions were as follows:1.Sound signals of breams and crucians were collected at the proportions of 1:1,1:2,1:3,1:4,2:1,3:1 and 4:1 respectively,and at the proportion of 1:1,the quantity increased from 2,4,6 to 30,respectively,and were processed by Butterworth lowpass filter,equalripple filter,wavelet filter and wiener filter.Three features of sound signals of breams and crucians in different proportions and numbers were extracted: frequency band energy based on wavelet packet decomposition,average Mel cepstral coefficient,and main peak frequency and main peak value based on the power spectrum.2.Mixed proportional recognition model of breams and crucians was established.When establishing the proportional recognition model,two classifiers,probabilistic neural network and support vector machine,were used to optimize the hyper-parameters of the classifier by using quadratic grid search.Finally,the average classification accuracy after 10 fold cross-validation was used to evaluate the classification effect.The results showed that in the model established by probabilistic neural network,when the equalripple filter was used,the smoothing factor was 0.09,the model performance was the best,and the recognition rate reached 0.9067.In the model established by the support vector machine,when the equalripple filter was adopted,the penalty coefficient was 90.5097,and the Gaussian kernel function parameter was 2.000,the model performance was the best,and the recognition rate reached 0.9583.3.Mixed quantitative prediction model of breams and crucians was established.When establishing quantitative prediction model,Rank-RS method was used to divide the sample set into training set and verification set.Multivariate linear regression based on ordinary least squares method was used to establish the model.The results showed that:(1)In the multivariate linear regression model constructed,the performance of the model built by using butterworth lowpass filtering data was the best.The determining coefficient of training set was 0.9522,the adjusted determining coefficient was 0.9514,the calibration mean square error was 0.9443,the determining coefficient of verification set was 0.9555,the verification mean square error was 0.9108,and the relative analysis error was 4.7571.(2)When t-test was used to judge the significant effect of each characteristic value on the numerical gradient in the above model,there were 12 features that had significant effect on the numerical gradient(P < 0.05).According to the absolute value of the normalized partial regression coefficient,the most significant effected on the numerical gradient were the 22 th,23th,4th and 20 th features,which represented respectively the 6th and 7th average Mel cepstrum coefficients,the band energy within 375-500 Hz and the 4th average Mel cepstrum coefficients.