Research And Development Of Intelligent Feeding Control Device For Fish Based On Sound And Image Analysis

Feeding is one of the important breeding operations in factory recirculating water aquaculture,and it is difficult to automatically adjust the feeding amount according to the feeding demand of fish in the traditional fixed-point,timed and quantitative way.Studies have shown that the feeding behavior of fish can give feedback on the feeding status,therefore,how to automatically adjust the feeding strategy according to the fish behavior and fish area is the key to achieve accurate feeding.This paper takes rainbow trout in factory aquaculture as the research object,combines acoustics and images,proposes a method to quantify fish feeding behavior based on acoustic information and improved Swin Transformer network,researches a method to identify fish spatial distribution based on light-weight image recognition algorithm,and develops a fish intelligent feeding system,which finally realizes accurate feeding of fish,the main research contents of this paper are as follows The main research contents of this paper are as follows(1)proposed a method to quantify fish feeding intensity based on audio spectrum and improved Swin Transformer model,which can classify fish feeding intensity into "strong","medium","weak" and "none"." and "none".Firstly,the audio and acoustic signals are edited using sliding windows,and the signals are transformed into spectrograms;secondly,the perceptual domain of fish feeding sound spectrum recognition is expanded by using Swin Transformer network and its hierarchical structure;and the model’s generalization performance based on small data sets of audio spectrum is further improved by adding a shift patch tokenization module,a local self-attentive module and an enhanced residual connection.Finally,a predictive optimization module was designed to modify the feeding strategy according to four feeding levels.The experimental results show that the accuracy of the improved Swin Transformer network for quantifying fish feeding behavior reaches 96.16%,which can effectively identify the feeding intensity of four types of fish and lay the foundation for developing intelligent baiting devices.(2)The method of fish spatial distribution recognition based on Mobile ViT network was proposed,and four types of fish spatial distribution,"far","medium","near" and "discrete",were realized.(2)We propose a method based on Mobile ViT network to accurately identify the spatial distribution of fish in "far","medium","near" and "discrete" states.The Mobile ViT model,a lightweight deep learning algorithm,is used to identify and classify the fish activity areas in the breeding pond.The fish distribution data of daytime and nighttime are trained under single view to verify the feasibility of the method and the generalization of the model.Comparison experiments are conducted under multi-view and multi-illumination fish distribution data to analyze the accuracy and recall of each distribution case.Finally,continuous video split-frame prediction is performed for each distribution type with false positive analysis.The experimental results show that the accuracy of fish spatial distribution recognition based on Mobile ViT network reaches99.28%.This method can effectively identify fish location distribution information in aquaculture and provide technical support to realize accurate feeding.(3)An intelligent feeding device is developed,which integrates the fish feeding intensity quantification model and fish spatial distribution recognition model,and adjusts the feeding amount and feeding distance in real time according to the fish feeding intensity and spatial distribution information.The specific implementation process is: the identification information of feeding activity of farmed fish is uploaded to the cloud platform,and the cloud sends RS485 commands to NBIOT(Narrow Band Internet of Things),which finally controls the hardware of feeding device.The hardware part uses NBIOT to receive the driver that supports RS485 communication protocol to control the DC motor to achieve feeding distance adjustment,and uses MCU(Microcontroller Unit)to drive the stepper motor to automatically control the feeding amount.The software part realizes the cloud backend and human-machine interaction front-end,retaining the function of manual remote control of feeding device.This intelligent feeding device can be used in real farming environment to accomplish remote automated feeding tasks.