In recent years, the escalation of globalization has heightened the threat of invasive species to the economies and ecosystems of various countries, emphasizing the importance of live detection in security. Frequency-modulated continuous wave (FMCW) radar technology has gradually matured as a non-contact method for life signal detection, offering solutions in fields such as biology and security. However, there remains a significant research gap in extracting and recognizing life signal characteristics from cold-blooded organisms like insects. Current physical feature extraction algorithms struggle with issues such as insufficient information and low accuracy when dealing with weak life signals, limiting the application of millimeter-wave radar technology. Therefore, we propose a method using the Transformer neural network architecture with embedded threshold shrinkage residual blocks, called the Trans-shrink-Net neural network, for life signal feature extraction and recognition. This approach aims to enhance the accuracy and generalization capability of millimeter-wave radar in detecting weak life signals. Additionally, we introduce the cell averaging-constant false alarm rate (CA-CFAR) algorithm as a preprocessing step to create high-quality standardized datasets, mitigating issues such as dirty data that could affect network performance. The significance of this method is twofold. First, by leveraging deep neural network technology, we can better explore and utilize the latent information in millimeter-wave radar data, improving the efficiency of life signal extraction and recognition. Second, this method addresses the current research gap in life signal detection, providing new avenues for development in related fields. Most importantly, applying this method will elevate the application level of millimeter-wave radar technology in detecting weak life signals, offering reliable technical support for ecological monitoring, disaster relief, and other areas. Overall, our research aims to propose a new method for life signal detection, addressing the shortcomings of existing technologies and improving the accuracy and generalization capability of millimeter-wave radar in detecting weak life signals, thereby providing important theoretical and methodological support for further research and practice in related fields.

We design a method for extracting and recognizing weak life signals from cold-blooded insects. The CA-CFAR detection algorithm is first used for preprocessing data features, establishing a standard dataset of 80 Gbit, containing over 100000 training data entries, to preprocess the weak life signal features for neural network extraction. We then design a Transformer neural network with embedded threshold shrinkage residual blocks, termed the Trans-shrink-Net neural network. This network enhances the generalization capability of the Transformer network by embedding threshold shrinkage blocks, achieving incremental learning of life signal data. The network effectively controls the feature flow in residual connections based on the product of the mask and multilayer perceptron (MLP) output by threshold shrinkage residual blocks. With dual convolution blocks for preprocessing the amplitude and phase of millimeter-wave radar data, the network captures local features and spatial correlations. The attention mechanism and feedforward neural network in the Transformer block address different aspects of positional features and feature transformation, effectively managing the model's depth and complexity. The fixed learnable positional encoding of the Transformer position encoding allows the network to handle sequential data better. The network learns the global distribution characteristics of weak life signals, introduces position sensitivity through threshold residual blocks, and outputs life signal predictions via a multilayer perceptron.

Experimental results indicate that the proposed method effectively extracts features of weak life signals and recognizes the life signals of small cold-blooded insects. The Trans-shrink-Net achieves a recognition accuracy of 96.17% for weak life signals (Fig. 3). It significantly outperforms common binary classification networks in recognizing micro-motion life signals (Fig. 4). Comparative experiments demonstrate that training the network using the standard dataset created with the CA-CFAR algorithm yielded the best results (Fig. 4). Introducing threshold shrinkage residual blocks enhances the feature representation capability of the standard dataset (Table 5).

Our study introduces the Trans-shrink-Net deep neural network for detecting vital signs of cold-blooded small organisms using millimeter-wave radar. This network integrates the CA-CFAR algorithm for preprocessing millimeter-wave radar micro-motion data to construct a comprehensive dataset for deep neural network learning. Through incremental training, the system exhibits high accuracy and robustness, addressing issues of low accuracy and poor generalization in millimeter-wave radar for recognizing weak life signals. Experimental results demonstrate that the Trans-shrink-Net network achieves a recognition accuracy of up to 96.17% for weak life signals. This method has significant potential for applications in various domains, providing a reliable and efficient means for non-contact life signal detection and recognition.