基于短时动态卷积的吞咽动作识别研究

Studies on Swallowing Action Recognition Using Short-Time Dynamic Convolution

摘要: 吞咽困难或吞咽功能障碍严重影响人类健康、尤其影响老年人群体，而目前现有的吞咽识别模型存在精度不足、模型参数冗余等问题。针对上述问题，本研究提出采用自主研发的柔性压电聚乳酸传感器作为数据采集终端，通过一种短时动态卷积(STDyConv)方法，用于分析受试者进行吞咽动作时采集的生理信号。该方法是由一个短时频率自适应卷积模块和一个全局传递卷积核组成，其中短时自适应卷积模块的参数存储在两个张量中，分别代表频率特征的实部和虚部。通过基于输入生理信号的参数微调和优化，卷积核能够精确响应输入信号的多维频率特征。为了进一步增强模型的场景适应能力，在每个卷积核中引入了频率加权因子，以实现对频率特定特征的差异化关注。全局传递卷积核仅需存储一个初始卷积核，其余卷积核均基于该初始卷积核生成，有效抑制了动态卷积模型参数的冗余增长。实验结果表明，STDyConv与传统的卷积神经网络(CNN)架构具有良好的兼容性，并且与传统方法相比，吞咽动作识别的平均准确率提高了20%。

Abstract: Dysphagia or swallowing dysfunction seriously affects human health, especially for the elderly. At present, the existing swallowing recognition models have problems such as insufficient accuracy and redundant model parameters. To address the aforementioned issues, this study proposes using a self-developed flexible piezoelectric polylactic acid sensor as the data acquisition terminal and employing a Short-Time Dynamic Convolution (STDyConv) method to analyze physiological signals collected during swallowing. This method consists of a short-time frequency adaptive convolution module and a globally propagating convolution kernel. The parameters of the short-time adaptive convolution module are stored in two tensors, representing the real and imaginary parts of the frequency features, respectively. Through parameter fine-tuning and optimization based on the input physiological signals, the convolution kernel can accurately respond to the multidimensional frequency features of the input signal. To further enhance the model’s scene adaptability, a frequency weighting factor is introduced into each convolution kernel to achieve differentiated attention to frequency-specific features. The globally propagating convolution kernel only needs to store one initial convolution kernel, and all other convolution kernels are generated based on this initial kernel, effectively suppressing the redundant growth of parameters in the dynamic convolution model. Experimental results show that STDyConv has good compatibility with traditional convolutional neural network (CNN) architectures, and compared with traditional methods, the average accuracy of swallowing action recognition is improved by 20%.