循环水养殖大口黑鲈摄食颗粒饲料的声学特征

为突破智能投饲系统的技术瓶颈，近年来采用被动声学技术开展鱼虾摄食行为研究成为热点之一。该研究主要采用被动声学技术获取单体大口黑鲈（Micropterus salmoides）摄食声信号，从混合信号中提取完善的摄食信号，筛选可作为衡量大口黑鲈摄食活跃度的声学特征参数，以期对摄食活跃度进行量化。根据大口黑鲈喂食期间的同步音频与视频记录，确定信号类别并进行标记，主要提取每次吞食饲料的时域与频域特征，对比各参数与吞食次序之间的相关度。研究结果表明，摄食声信号能量主要集中于4.2～7.4 kHz，且大口黑鲈吞食间隔与吞食次序呈正相关，稳定性较强；而时域特征中的波形振幅极差与频域特征的功率积分值均与吞食次序呈负相关。吞食间隔、振幅极差及功率积分值均可以作为衡量摄食活跃度的量化指标，而共振峰与平均梅尔倒谱系数可作为摄食声识别参数，研究结果可为今后养殖鱼类被动声学智能投饲系统研发提供理论基础。

Acoustic characteristics of the feeding pellets for Micropterus salmoides in circulating aquaculture

Abstract: A passive acoustic technology has been widely used to monitor the behavior of aquatic organisms for the intelligent feeding system in recent years. Taking six sizes of Micropterus salmoides as research objects, this study aims to acquire the acoustic signals of pellet feeding in circulating aquaculture using passive acoustic techniques. The signals were first identified to classify for the post-processing using simultaneous audio and video recordings during feeding. The feeding activity was then quantified to extract and screen the characteristic parameters from the acoustic signals. Six kinds of pre-processing were utilized for the feeding sound signals, including A/D conversion, denoise, pre-emphasis, windowed framing, and endpoint detection. A Fast Fourier transform, real-time, and Mel frequency cepstrum methods were also used to extract the time- and frequency-domain features of each swallowing signal in the complete feeding acoustic signal, in order to obtain the correlation between each acoustic feature parameter and the swallowing order. Specifically, the swallowing interval, the peak-to-peak value of voltage, and the maximum amplitude were extracted from the time-domain features. It was found that the swallowing interval was positively correlated with the order of swallowing (R2>0.68), whereas, the maximum and range amplitude was negatively correlated with the order of swallowing (R2 <-0.61), but there was no significant difference between the correlation coefficient of three time-domain characteristic parameters. Furthermore, the maximum sum of power intensity and integral value was extracted from the power spectrum of each swallowing signal. Among them, P=0.05 was assumed as the basis to evaluate the integral value of power, where a more stable and reliable measurement was achieved for the characteristic parameters of feeding activity. In addition, the formant frequency and the average Mel cepstrum coefficient (AMFCC) were extracted to find each acoustic signal of feeding mainly in 4-7 kHz. More importantly, the third coefficient in AMFCC presented an outstanding and stable peak. Particularly, the feeding activity decreased significantly, as the feeding sequence increased. The extraction of power integral parameters depended significantly on subjective factors, although both time domain and frequency domain parameters presented an excellent correlation with the order of swallowing. The feature parameters of the time domain also behaved more reliable stability. Subsequently, the feature parameters for the activity of eating were screened out, according to the correlation between the acoustic feature parameters of ingestion and the order of swallowing. Correspondingly, the feature parameters of multi-feature fusion can be expected to better quantify the feeding activity, indicating the best choice for the swallowing interval and peak-to-peak value of voltage. The finding can also provide theoretical support to identify the sound signal of farmed fish in the intelligent feeding system.