肠鸣音是人体肠道的生理信号之一，具有重要的医学潜力，通过监测肠鸣音的数量可以辅助检测肠道疾病。当前临床的肠鸣音监测方法大多为人工听诊，医疗成本较高且对医师专业性有要求，而现有的肠鸣音识别方法存在准确率太低或计算量过高的问题，无法进行长时间的有效肠鸣音监测，结合课题组现有的可穿戴式体音记录仪，设计出一套高准确率、低开销的边缘端肠鸣音硬件识别电路具有重要意义。本文首先提出了基于LS-SVM分类器的肠鸣音信号识别算法，算法采用勒让德多项式平滑频谱和LS-SVM分类器结合的方法，肠鸣音识别准确率达90.1%，然后基于课题组已有的时域特征提取技术和专用二值卷积神经网络结合的肠鸣音识别算法，提出了整体的边缘端肠鸣音识别硬件架构并完成设计，同时针对低功耗需求，提出关键模块的低开销、低计算量优化技术，完成了整套识别电路的模块确定与设计、仿真综合、FPGA验证等流程，本文设计的肠鸣音识别电路可将1秒钟的输入声音片段分类为肠鸣音、说话声或高斯噪声片段。本文提出的硬件架构分为三部分：全局控制模块、数据存储模块和数据处理单元。针对边缘端设备低功耗的主要设计需求，本文提出了关键的数据处理单元和存储模块优化技术。通过对数据集特征的分析，优化了特征提取和浮点运算模块，并对高位宽浮点数进行低比特量化处理，有效降低识别流程中的运算次数。在计算密集型的卷积运算模块中，利用位运算替代乘法运算、popcount运算替代加法运算等技术，在保持网络推理精度的基础上，大幅度降低卷积运算功耗。除此之外，考虑到硬件面积和资源使用率，本文对运算模块与存储模块进行复用设计，在保证声音片段的识别实时性的同时，有效降低硬件开销。本文基于FPGA完成硬件验证，在电路工作频率不高于30kHz时即可具备实时识别功能，同时消耗的FPGA片上资源不超过15%，具备低频低功耗的特性。利用测试集进行验证，本文工作的输入声音片段识别准确率高达99.33%，同时肠鸣音虚警率几乎为0， 硬件模块的网络参数存储空间、操作数与核心电路等效门数仅为0.47KB、4.89K和27.3K，远优于相关工作。本文提出的肠鸣音识别硬件工作在保留高准确率的同时，具有最低的硬件开销，满足边缘端肠鸣音识别的需求。

Bowel sounds is one of the physiological signals of human intestinal tract, which has important medical potential. Monitoring the number of bowel sounds can help to detect intestinal diseases. At present, most of the clinical monitoring methods of bowel sounds are manual auscultation, which has high medical cost and professional requirements for doctors. At the same time, the existing recognition methods of bowel sounds have the problems of low accuracy or high calculation, which can not effectively monitor bowel sounds for a long time. Combined with the existing wearable body sound recorder of the research group, it is of great significance to design a set of hardware recognition circuits of edge end bowel sounds with high accuracy and low cost.In this paper, a recognition algorithm of bowel sounds based LS-SVM classifier is proposed, and the accuracy reaches 90.1%. Then, based on the time-domain feature extraction technology and special binary convolutional neural network algorithm, this paper proposes the overall hardware architecture of bowel sound recognition and completes the design. At the same time, aiming at the low power consumption requirements, it proposes the low cost and low computational complexity optimization technology of key modules, and completes the module identification and design, simulation synthesis, FPGA verification of the whole recognition circuits. The circuits can classify the one-second input sound segments into bowel sound, speaking sound or Gaussian noise. The hardware architecture is divided into three parts: global control module, data storage module and data processing unit. For the main demand of low power consumption of edge devices, this paper optimizes the key data processing unit and storage module. By analyzing the characteristics of the data set, the feature extraction and floating-point operation modules are optimized, and the 32-bit floating-point number is quantized with 16-bit, which effectively reduces the number of operations in the recognition process. In the calculation intensive convolution module, bit operation is used to replace multiplication operation, and popcount operation is used to replace addition operation, which greatly reduces the power consumption of convolution operation on the basis of maintaining the accuracy of network reasoning. In addition, considering the hardware area and resource utilization, the operation modules and storage modules are reused to ensure the real-time recognition of sound clips and effectively reduce the hardware overhead. In our work, FPGA is used to complete the hardware verification, the circuit working frequency is not higher than 30kHz can have real-time recognition function, the consumption percentage of on-chip resources is not more than 15%. It has the characteristics of low frequency and low power consumption. The test results show that the recognition accuracy of the input voice segment is 99.33%, and the false alarm rate of bowel sounds is almost zero. The network parameter storage space, the number of operands and the equivalent gates of the core circuit of the hardware module are only 0.47KB, 4.89K and 27.3K, which are far better than the related work. The work of this paper has the lowest hardware overhead while retaining high accuracy, which can meet the needs of edge end bowel sound recognition.