Human body communication method based on vibration linear spread spectrum and deep learning demodulation

Abstract

The invention provides a human body communication method based on vibration linear spread spectrum and deep learning demodulation, and the method comprises the steps: coding to-be-transmitted data into a vibration signal through a linear frequency modulation spread spectrum modulation technology, and injecting the vibration signal into a human body channel through a vibration actuator; receiving the vibration signal from a human body channel through a vibration transducer, and performing time-frequency transformation processing on the vibration signal to generate two-dimensional feature data; and inputting the two-dimensional feature data into a pre-trained deep neural network model, classifying and identifying the two-dimensional feature data by the model, and demodulating data symbols. The system has low power consumption and high reliability, the calculation burden of the convolutional neural network model is controllable, the system adapts to limited batteries and calculation resources of wearable equipment, can cooperatively work with existing wireless communication modes such as Bluetooth and WiFi to form a heterogeneous communication system, and has wide application value in the field of wearable Internet of Things such as medical health and outdoor sports.

Description

Technical Field

[0001] This invention belongs to the field of wireless communication technology, specifically relating to a human body communication method based on vibration linear spread spectrum and deep learning demodulation. Background Technology

[0002] As a crucial component of the next generation of information technology, the Internet of Things (IoT), particularly wearable IoT, has widely penetrated various fields such as healthcare and outdoor sports. Wearable devices, by integrating sensing and wireless communication technologies, provide users with diverse services including health management and daily living assistance, and their popularity is increasing daily. However, wearable devices are generally characterized by their small size, limited power capacity, and insufficient computing power. This makes it difficult for traditional wireless communication protocols such as Bluetooth and WiFi to fully implement effective security mechanisms. Furthermore, the inherently open transmission environment of these communications makes it highly susceptible to the theft or tampering of private data transmitted by the device (such as vital signs and location information), leading to security risks. Simultaneously, the divergence of wireless signals can cause signal diffusion and attenuation, affecting the accurate perception of user behavior characteristics and limiting practical applications. Therefore, achieving low-power, high-security, integrated communication and sensing for wearable IoT devices under the constraints of limited hardware resources has become a critical issue that urgently needs to be addressed.

[0003] Body Area Networks (BANs), wireless communication networks specifically designed for the human body and its surroundings, perfectly suit the characteristics of most wearable devices that come into direct contact with the human body. With advantages such as short range and high reliability, they have become an important research direction for solving the aforementioned problems. Related research is increasingly focusing on using the human body itself as a communication channel to design low-power communication between wearable devices. Currently, existing research in this field mainly follows two technical paths: one is based on electromagnetic wave transmission. Related research has verified the feasibility of the human body as an electromagnetic wave conductor by achieving wireless power transmission and device pairing through human skin. However, such transmission usually requires special hardware support, resulting in high power consumption and susceptibility to external environmental interference. The second is based on vibration wave transmission. Vibration waves themselves have characteristics such as limited propagation range, strong anti-interference ability, and low power consumption. Existing research has confirmed that the human body can be used as a communication channel to propagate vibration wave signals of specific frequencies, providing new possibilities for low-power, high-reliability communication.

[0004] Despite the significant potential of vibration wave transmission, its application in the human body channel still faces numerous technical challenges. The non-uniformity of human tissue and the dynamic activities of the body cause continuous changes in channel characteristics, severely affecting the stable transmission of vibration signals. Simultaneously, vibration signals are easily absorbed by human tissue during transmission, leading to attenuation, decreased signal strength, and increased bit error rate. While existing vibration communication methods have achieved preliminary signal transmission through the human body, they generally suffer from short transmission distances, failing to meet practical application requirements. This has resulted in relatively weak research on wearable devices based on vibration waves in the field of secure communication, and a mature solution has yet to be developed. Summary of the Invention

[0005] To address the shortcomings and deficiencies of existing technologies, this invention provides a human body communication method and system based on vibration linear spread spectrum and deep learning demodulation. It aims to solve the technical challenges of low power consumption and high security communication under the limited hardware resources of wearable devices, while overcoming problems such as transmission instability and high bit error rate caused by dynamic channel changes and signal attenuation when vibration waves are transmitted in the human body channel.

[0006] This invention uses the human body as a vibration signal transmission channel, leveraging the limited propagation range of vibration waves to physically isolate external eavesdropping and interference, ensuring data privacy and security. The transmitting end encodes the raw data into vibration signals using linear frequency modulation spread spectrum (CSS) technology. Based on the spreading factor and bandwidth, it defines the encodeable symbol types and designs a dedicated data packet structure including a preamble, a start-of-frame delimiter, and data segments. The preamble is composed of multiple basic sweep signals, and the start-of-frame delimiter is composed of reverse sweep signals, effectively improving signal synchronization stability. The receiving end captures the vibration signal through a transducer and converts it into an electrical signal. After analog-to-digital conversion, it sequentially performs bandpass filtering, outlier replacement, preamble detection, and symbol segmentation preprocessing. Then, it uses short-time Fourier transform to convert the time-domain signal into a time-frequency amplitude spectrum, generating two-dimensional feature data adapted to a deep learning model. A hierarchical convolutional neural network model with integrated global response normalization mechanism (based on ConvNeXt) is employed. The V2 architecture replaces traditional demodulation methods by adaptively extracting deep robust features of signals through core components such as depthwise separable convolution, point convolution, and residual connections, accurately classifying and identifying FM symbols and demodulating data.

[0007] This invention combines low power consumption and high reliability, with a controllable computational burden for the convolutional neural network model. It is adapted to the limited battery and computing resources of wearable devices and can work in conjunction with existing wireless communication methods such as Bluetooth and WiFi to form a heterogeneous communication system. It has broad application value in wearable Internet of Things fields such as medical health and outdoor sports.

[0008] The specific technical solution adopted by this invention to solve its technical problem is as follows:

[0009] A human body communication method based on vibration linear spread spectrum and deep learning demodulation includes:

[0010] The data to be transmitted is encoded into a vibration signal using linear frequency modulation spread spectrum modulation technology, and the vibration signal is injected into the human body channel through a vibration actuator;

[0011] The vibration signal is received from the human body channel through a vibration transducer, and the vibration signal is processed by time-frequency transformation to generate two-dimensional feature data.

[0012] The two-dimensional feature data is input into a pre-trained deep neural network model, which classifies and identifies the two-dimensional feature data and demodulates the data symbols.

[0013] Furthermore, the vibration signal data packet generated by the linear frequency modulation spread spectrum modulation technique includes a preamble, a frame start separator, and a data segment. The preamble is composed of multiple basic frequency sweep signals, the frame start separator is composed of reverse frequency sweep signals, and the data segment is composed of frequency modulation symbols that encode the original data.

[0014] Furthermore, the time-frequency transformation process is a short-time Fourier transform, and the two-dimensional feature data is the signal amplitude spectrum obtained based on this short-time Fourier transform.

[0015] Furthermore, before performing time-frequency conversion processing on the vibration signal, a signal synchronization step is also included: detecting the preamble and frame start separator through a cross-correlation algorithm to determine the starting position of the data segment, and then dividing the data segment into independent frequency modulation symbols according to a preset symbol length.

[0016] Furthermore, the deep neural network model is a convolutional neural network model that integrates a global response normalization mechanism.

[0017] Furthermore, the convolutional neural network model is the ConvNeXtV2 model.

[0018] Furthermore, the processing of the vibration signal also includes a noise suppression step: removing device noise through bandpass filtering and replacing outliers in the signal.

[0019] Furthermore, the linear frequency modulation spread spectrum modulation technique defines symbol types based on the spreading factor SF and the bandwidth BW, wherein the number of coded symbol types is 2 to the power of SF, and the frequency change rate of the vibration signal is the ratio of BW to the signal duration T.

[0020] Furthermore, the vibration actuator is a bone conduction vibration actuator, and the vibration transducer is a transducer used to collect vibrations from the body surface.

[0021] Furthermore, a human body communication system based on vibration linear spread spectrum and deep learning demodulation includes:

[0022] Transmitting module: includes a vibration actuator and a modulation unit. The modulation unit encodes the data to be transmitted into a vibration signal using linear frequency modulation spread spectrum modulation technology. The vibration actuator injects the vibration signal into the human body channel.

[0023] Receiving and preprocessing module: includes a vibration transducer and a preprocessing unit. The vibration transducer receives vibration signals from the human body channel, and the preprocessing unit performs time-frequency transformation processing on the vibration signals to generate two-dimensional feature data.

[0024] Demodulation module: contains a pre-trained deep neural network model, which classifies and identifies the two-dimensional feature data and demodulates the data symbols.

[0025] And a computer device including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the method described above.

[0026] A non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described above.

[0027] Compared with the prior art, the present invention and its preferred embodiments have at least the following beneficial effects:

[0028] In terms of communication security, this invention innovatively uses the human body as the transmission channel for vibration signals. By taking advantage of the naturally limited propagation range of vibration waves, it physically isolates external eavesdropping and wireless interference, effectively avoiding the privacy leakage risks of traditional open wireless communication protocols, and providing reliable security for sensitive data such as vital signs and location transmitted between wearable devices.

[0029] Regarding transmission reliability, this invention addresses the issues of variable channel characteristics, signal attenuation, and distortion caused by the non-uniformity of human tissue and dynamic activities. It endows the signal with strong noise resistance through linear frequency modulation spread spectrum modulation technology, enhances signal synchronization stability with a dedicated data packet structure, and combines a hierarchical deep learning model with an integrated global response normalization mechanism to adaptively extract deep robust features of the signal. This significantly improves the demodulation accuracy and adaptability under complex dynamic channels, effectively reduces the bit error rate, and overcomes the limitations of unstable and distance-limited transmission in existing vibration communication.

[0030] In terms of energy efficiency adaptability, vibration communication itself has low power consumption characteristics, and the convolutional neural network model it uses has an efficient architecture design, with controllable computational burden and no need for additional special hardware support. It is well adapted to the hardware constraints of wearable devices, such as small size, limited battery capacity and insufficient computing resources, thus extending the device's battery life.

[0031] In terms of system compatibility, the solution of this invention can work in conjunction with existing wireless communication methods such as Bluetooth and WiFi to form a heterogeneous communication system that balances high security and flexible connectivity. It can expand application scenarios without making major modifications to existing devices, thereby improving the feasibility and practicality of technology implementation and providing strong support for the in-depth application of wearable IoT in fields such as medical health and outdoor sports. Attached Figure Description

[0032] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0033] Figure 1 This is an overall framework diagram of the communication system according to an embodiment of the present invention;

[0034] Figure 2 This is a flowchart of the ConvNeXt V2 model according to an embodiment of the present invention;

[0035] Figure 3 This is a schematic diagram of the transmitting and receiving devices according to an embodiment of the present invention;

[0036] Figure 4 This is a schematic diagram of a signal transmission scenario according to an embodiment of the present invention;

[0037] Figure 5 This is a comparison chart of symbol error rates under different transmission powers according to an embodiment of the present invention;

[0038] Figure 6 This is a comparison chart of symbol error rates under different spreading factors in embodiments of the present invention. Detailed Implementation

[0039] To make the features and advantages of the present invention more apparent and understandable, specific embodiments are described below in detail:

[0040] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used in this specification have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0041] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0042] 1. Problem Model

[0043] This invention aims to provide a method for private human communication based on vibration signals. Utilizing the human body as a communication channel, a vibration communication system adapted to the characteristics of this channel is constructed. The system fully leverages core components such as vibration motors, receivers, and signal processing modules in wearable devices, and designs a communication mode consisting of three parts: a transmitter, a human body channel, and a receiver.

[0044] To address the complexity of the human body channel, this invention utilizes the mechanism of the linear spread spectrum modulation scheme in LoRa technology to design corresponding linear frequency modulation symbols and data packet structures. The linear frequency modulation signal features linear frequency sweep over time, exhibiting extremely strong noise immunity within a preset bandwidth, effectively overcoming multipath attenuation and environmental interference in the human body channel.

[0045] In the embodiments of this invention, a complete data packet structure comprises three parts: a preamble, an SFD (Start Frame Delimiter), and a data portion. The preamble contains 6 basic upchirps, followed by 2.25 basic downchirps, which serve as the SFD indicating the start of the data segment. The final data portion contains several datachirps that encode the data.

[0046] 2. Method Description

[0047] 2.1 Overall Overview

[0048] System overview as follows Figure 1As shown, a complete closed-loop communication system was constructed, fully covering the entire process from raw data to demodulation results. First, the system constructs the basic signal by generating a chirp array. The raw data to be transmitted is encoded into a linear frequency modulated signal with noise resistance robustness using CSS modulation technology based on the set spreading factor and bandwidth. Subsequently, the modulated signal is applied to actuators such as bone conduction vibration motors through the drive circuit, converted into mechanical vibration, and injected into the human body channel, using the human body as a dedicated transmission medium to complete the signal transmission. After the receiving end captures the vibration analog signal through the transducer and converts it into a digital signal stream, it performs receiving preprocessing operations in sequence: First, it achieves signal purification through bandpass filtering and equipment noise removal. Preferably, it can further optimize signal quality by replacing outliers with LPC prediction values ​​(e.g., using the Linear Predictive Coding (LPC) algorithm to perform forward prediction on the digital signal stream, calculating the deviation between the predicted value and the actual sampled value of the signal, identifying sampling points with deviations exceeding a preset range as outliers, and replacing the outliers with the corresponding LPC prediction values ​​to optimize signal quality). Finally, it completes signal synchronization based on the cross-correlation detection of the preamble and SFD. The preprocessed signal is converted into a time-frequency amplitude spectrum through symbol segmentation and short-time Fourier transform, and is fed as input into a deep learning model based on the ConvNeXt V2 architecture. The model completes feature extraction and classification recognition through core mechanisms such as hierarchical convolutional blocks and global response normalization, and finally outputs the demodulated symbol sequence, which is reconstructed to restore the complete original application data, completing a safe and reliable communication process.

[0049] 2.2 CSS Modulation of Signals

[0050] A linear frequency modulated (LFM) signal is a signal whose frequency changes linearly with time. It sweeps its frequency linearly over a preset bandwidth, exhibiting extremely strong noise immunity. This embodiment achieves this by shifting the frequency of the essentially swept frequency signal (Upchirp). By embedding data, the resulting linear frequency modulated signal can be expressed as:

[0051] (1)

[0052] in This basically represents a frequency sweep signal. To correspond to the frequency offset of the data symbol, each data symbol to be transmitted is mapped to a unique frequency offset value that is discretely distributed within the bandwidth. . The offset component is used to shift the starting frequency of the essentially swept signal to... . The rate of change of frequency, For bandwidth, This represents the duration of the linear frequency modulated signal.

[0053] Similar to LoRa, this embodiment uses a spreading factor. Define an encodeable symbol type, which is defined as follows:

[0054] (2)

[0055] Given bandwidth , The larger the length of each chirp signal, the greater the length of the chirp signal. The longer, The number of possible starting frequencies for each signal (i.e., the types of symbols that can be encoded).

[0056] 2.3 Signal Preprocessing

[0057] Receiver signal preprocessing includes the following key steps:

[0058] 1) Preamble Detection: To detect input data packets, this embodiment determines the start position of the data packet based on the known structure of the preamble (containing 6 basically swept frequency signals) and the start of frame delimiter. During detection, a sliding window with a length twice the total length of the preamble and SFD is first applied to the received signal stream to extract signal segments; then, the cross-correlation between each segment and the preamble is calculated, and the segment with the highest correlation is taken as the preliminary detection result. Based on this, intervals of 5 signal lengths before and after this result are taken for further cross-correlation calculation with the SFD. If the highest correlation value exceeds a set threshold, the data packet is determined to exist, and the peak position is determined as the start position of the data portion.

[0059] 2) Symbol Segmentation: Starting from a defined data start position, the data portion is segmented into independent signals based on the symbol length. For example, when the bandwidth is 2kHz and the spreading factor is 4, the symbol length T is calculated to be 8 milliseconds. This process outputs separated linear frequency modulated symbols for subsequent symbol conversion.

[0060] 3) Symbol Conversion: In the symbol conversion stage, each separated linear frequency modulated signal is first converted into a time-frequency representation through a short-time Fourier transform, and the amplitude spectrum is calculated based on the real and imaginary parts of its complex values. Subsequently, this time-frequency amplitude information is encoded into a three-channel color spectrogram, converting the one-dimensional time-domain signal into a two-dimensional time-frequency image for processing by the deep learning model. This process simplifies the input data structure, adapts to real-valued neural networks, and fully preserves the key sweeping mode features required for accurate decoding.

[0061] Preferably, outlier replacement may also be included: that is, using the linear predictive coding (LPC) algorithm to perform forward prediction on the digital signal stream, calculating the deviation between the predicted value and the actual sampled value of the signal, identifying the sampled points whose deviation exceeds a preset range as outliers, and replacing the outliers with the corresponding LPC predicted values ​​to optimize signal quality.

[0062] 2.4 Traditional Demodulation Methods

[0063] Dechirp (despreading) is currently the most widely used linear frequency modulation (LFM) signal demodulation method and is also the benchmark method compared in this embodiment. It involves multiplying the received baseband upchirp signal by the downchirp signal to convert it into a single-frequency signal. This concentrates the chirp signal energy onto a single frequency. The formula is as follows:

[0064] (3)

[0065] in For the complex conjugate of the essentially swept frequency signal, the resulting signal (i.e. By performing an FFT (Fast Fourier Transform), a peak value can be obtained in the frequency domain. The frequency corresponding to this peak value is the signal's starting frequency. The encoded data in the signal can be determined by the index corresponding to the starting frequency, thus obtaining the SF bit data.

[0066] 2.5 Neural Network Model

[0067] Deep learning-based image classification methods demonstrate broad adaptability in signal image classification due to their efficient feature learning capabilities. Convolutional Neural Networks (CNNs), with their local perception and parameter sharing characteristics, can effectively handle strongly structured data such as time-frequency maps, overcoming the feature design limitations of traditional methods. ConvNeXt V2, as a representative of the new generation of convolutional networks, with its pure convolutional architecture and global response normalization mechanism, is well-suited to the demodulation requirements of human vibration signal time-frequency maps; therefore, this invention adopts this model to replace traditional demodulation methods.

[0068] 2.5.1 ConvNeXt V2

[0069] This embodiment uses a hierarchical convolutional model based on ConvNeXt V2 to replace the traditional demodulation method. For example... Figure 2 As shown in the model flowchart, the ConvNeXt V2 model adopts a pure convolutional architecture and integrates a global response normalization mechanism. The model input is a 224×224 pixel RGB image. First, a 4×4 convolutional layer (with a stride of 4) is used for initial feature extraction and spatial downsampling, compressing the input image size from 224×224 to 56×56, while generating an initial feature map with 64 channels. Then, a normalization layer projects the features into the embedding space, laying the foundation for subsequent feature extraction.

[0070] Figure 2In this example, the input data corresponds to a 224×224 pixel RGB image. The depth-separable convolution, GELU activation, residual connection and other modules correspond one-to-one with the basic block components in this example. The downsampling layer corresponds to the 2×2 convolutional layer between stages, which is used to compress the feature map size and multiply the number of channels. The ConvNeXt V2 module covers four stacked stages (Stage1-Stage4), which correspond to the hierarchical convolutional block stacking structure of stages 1, 2, 3 and 4 in this example.

[0071] The selected model comprises four stages, each consisting of multiple stacked ConvNeXt V2 base blocks. Each base block consists sequentially of depthwise separable convolutions, layer normalization, pointwise convolutions, global response normalization layers, and residual connections. The residual connections skip some convolutional operations, effectively mitigating the vanishing gradient problem in deep network training, while simultaneously fusing shallow and deep features to enhance the model's representational capabilities. In stages 1, 2, and 4, the base block is repeated three times; in stage 3, it is repeated nine times, extracting more complex semantic information through multiple feature iterations. Spatial downsampling is performed between stages using 2×2 convolutional layers, further reducing the feature map size to half its original size while doubling the number of channels (from 64 to 128, 256, and 512 respectively), forming a pyramid-shaped feature hierarchy that ensures multi-scale feature representation while progressively increasing the abstraction level of the features. Finally, the feature map is compressed to 1×1 through global average pooling, and then normalized. The image classification prediction result is generated through a linear classification head, thus completing the feature recognition and decoding of the input time-frequency map.

[0072] The core components of each basic block are described below:

[0073] Global Response Normalization Layer: This module enhances the competitive interaction between feature channels by calculating the L2 norm comparison between channels. GRN employs a global feature aggregation and normalization mechanism to effectively prevent feature collapse in self-supervised pre-training and improve the representation of feature diversity.

[0074] Depthwise separable convolution: This module employs depthwise convolution operations with large 7×7 kernels to expand the receptive field and capture richer spatial contextual information. The depthwise separability design significantly reduces computational complexity while maintaining representational power, achieving efficient feature extraction.

[0075] Point-based convolution and non-linear activation: This module expands and compresses the channel dimension through two linear transformation layers, and introduces non-linear transformation capabilities in conjunction with the GELU activation function. This design enhances the model's ability to model complex visual patterns while maintaining parameter efficiency.

[0076] Residual connections: By directly adding the input features of the base block to the output features after a series of convolution and normalization operations, features are transferred across layers. This preserves the original feature information, alleviates the training difficulty of deep networks, and improves the convergence speed and generalization performance of the model.

[0077] Overall, the ConvNeXt V2 model, through its carefully designed pure convolutional architecture combined with global response normalization and residual connections, significantly improves the quality of feature representations while maintaining computational efficiency. It not only boasts the advantage of short training time but also meets the low-power requirements of the communication system proposed in this embodiment.

[0078] 2.5.2 Implementation of Key Technologies

[0079] The following describes the key technologies for adapting deep learning models to the human vibration communication scenario in this embodiment, focusing on the accuracy of signal demodulation and the low power consumption requirements of the device:

[0080] Global response normalization

[0081] As the core feature adjustment mechanism of the ConvNeXt V2 model in this embodiment, it aims to solve the channel activation redundancy problem that easily occurs in the extraction of human vibration signal time-frequency map features, and improve feature discriminability to adapt to the complex and ever-changing human channel. This mechanism is implemented in three steps: first, the spatial dimension L2 norm of the time-frequency feature map of each channel is calculated to obtain the global response intensity; then, the relative importance of each channel is calibrated through cross-channel competitive normalization; finally, learnable scaling and translation parameters and residual connections are introduced to ensure stable training while adaptively adjusting the normalization intensity. This design forces global comparison and competition between feature channels, effectively preserves the key sweep frequency mode features of the vibration signal, suppresses redundant information caused by environmental interference and channel changes, and provides support for demodulation accuracy.

[0082] Hierarchical convolutional block design

[0083] The hierarchical convolutional blocks customized for the model in this embodiment are the core building blocks for achieving a balance between low power consumption and efficient feature extraction. They are specifically adapted to the resource constraints of wearable devices and the spatial contextual feature requirements of vibration signals. Each convolutional block starts with a 7×7 large-kernel depthwise separable convolution, expanding the receptive field while controlling the number of parameters to accurately capture deep frequency sweep features in the time-frequency map. Subsequently, the training dynamics are stabilized through layer normalization, and the nonlinear modeling capability is enhanced through an inverse residual structure of "channel expansion-compression". At the same time, a global response normalization mechanism is embedded to further optimize the feature quality. Finally, feature transfer across layers is achieved through residual connections, alleviating the training difficulty of deep networks, and random depth regularization is used to improve the model's generalization performance. This modular design can flexibly adapt to the demodulation requirements of vibration signals by adjusting the number of channels and stacked layers. While ensuring computational efficiency, it fully explores the time-frequency features of vibration signals, meeting the low power consumption and high reliability communication requirements of wearable devices.

[0084] 3. Experimental Results and Analysis

[0085] 3.1 Experimental Setup

[0086] 1) Hardware implementation scheme: such as Figure 3 As shown, the vibration-based human body local area network communication prototype designed in this embodiment uses a bone conduction vibration motor as the transmitter, which is driven by a PAM8406 audio driver chip to receive CSS modulated signals. Its power consumption is monitored by a Juwei A3 power meter. At the receiving end, an Adeline AD-35 transducer is used to collect vibration signals, which are then sampled and processed by a Raspberry Pi Zero development board in conjunction with an IQaudio codec. The entire system is powered by a portable power supply, facilitating flexible deployment. The vibration motor and transducer are fixed to two different parts of the human body to complete the transmission and reception of communication data.

[0087] 2) Data Acquisition Settings: In the CSS modulation configuration, BW=2kHz and SF=4 are set as the default settings to achieve a balance between data rate and demodulation accuracy. To evaluate the system's performance in a real-world scenario, this embodiment conducted data acquisition on different individuals with fixed transmitter and receiver locations. A total of 70,000 symbols were acquired.

[0088] 3) Training and Testing Methods: This embodiment uses the PyTorch framework to implement the model and employs the Adam optimizer for training. The initial learning rate is set to 0.00002, and the batch size is set to 16. During training, the cross-entropy loss function is used to measure the difference between the model's predicted values ​​and the true labels, and the model parameters are optimized through backpropagation. Simultaneously, a stochastic deep regularization strategy is applied to the base blocks of the ConvNeXt V2 model, with a base block dropout rate of 0.1, to improve the model's generalization performance and avoid overfitting. Each training iteration consists of 50 epochs, a setting sufficient for model convergence. The best-performing model from these 50 epochs is selected as the final output training model. All experiments were performed on an NVIDIA GeForce RTX 3090 GPU equipped with 24 GB of VRAM. The dataset is divided into training and test sets in an 8:2 ratio.

[0089] 3.2 Performance Evaluation Results

[0090] To evaluate the robustness of deep learning methods, this embodiment utilizes Figure 3 The experimental setup shown was used for actual communication testing. The transmitter and receiver were secured to the wrist and elbow respectively with tape, as detailed below. Figure 4 As shown.

[0091] 3.2.1 Power and SER

[0092] With a fixed communication configuration of SF=4 and BW=2000, this embodiment collected the communication performance of vibration signals on the human body under five different power levels. The results are as follows: Figure 5 As shown.

[0093] Figure 5 The diagram illustrates the SER performance of the proposed method and the baseline method at various power levels. At all power levels, the proposed method consistently outperforms the baseline linear frequency modulator. Notably, the SER curve of the deep learning method is consistently significantly below the limit; at a transmit power of 0.12W, the SER of the deep learning method can be reduced by up to 8.33% for the same power setting.

[0094] 3.2.2 SER performance under different SFs

[0095] With BW=2000 and transmit power fixed at 0.12W, this embodiment tested the SER cases with communication bandwidth SF set to 4, 5 and 6 respectively.

[0096] Figure 6The results show the symbol error rate (SER) at different spreading factors (4, 5, and 6) with a fixed bandwidth of 2 kHz. As the SF decreases, the SER gradually increases, which is consistent with expectations: lower SF values ​​correspond to shorter linear frequency modulated (LFM) symbols, making them more susceptible to noise interference. However, at lower SF values, the signal transmission bit rate is higher.

[0097] Compared to existing technologies, the present invention offers several advantages. First, in terms of security, it innovatively uses the human body itself as a communication channel to transmit vibration signals. The naturally limited propagation range effectively isolates the device from external eavesdropping and wireless interference at the physical level, providing extremely high privacy and security for data transmission between wearable devices. Second, regarding reliability, to address the signal attenuation and distortion issues caused by the complex and variable nature of the human body channel, a hierarchical deep learning model based on ConvNeXt V2 is used to replace the traditional physical layer demodulation method. This model can adaptively learn and extract deep robust features from the signal time-frequency diagram, significantly improving the demodulation accuracy of linear frequency modulated signals and its adaptability in dynamic channels, thus achieving more stable and reliable communication. Third, in terms of energy efficiency, vibration communication itself has low power consumption characteristics, and the adopted convolutional neural network model structure is efficiently designed with controllable computational burden. The combination of these two factors makes this solution particularly suitable for wearable devices with limited battery capacity and computing resources. Furthermore, this solution has good complementarity and can work in conjunction with existing wireless communication methods such as Bluetooth and WiFi to form a heterogeneous communication system that balances high security and flexible connectivity.

[0098] Based on such Figure 1 The system framework shown in this invention, and the reference implementation process provided in the final embodiment of this invention, specifically includes the following steps:

[0099] Step 1, Signal Modulation and Transmission: The transmitting end encodes the raw data to be transmitted into a corresponding linear frequency modulation vibration signal using CSS modulation technology according to the set spreading factor and bandwidth; subsequently, the electrical signal is driven by the drive circuit to act on actuators such as bone conduction vibration motors, converting it into mechanical vibration and transmitting it to the human body channel.

[0100] Step 2, Signal Reception and Preprocessing: The transducer at the receiving end captures the analog signal of vibration from the human body surface, and obtains a digital signal stream after analog-to-digital conversion; then, the receiving system performs a preprocessing process, including preamble detection through cross-correlation algorithm to locate the start of the data packet, segmentation according to symbol length, and conversion of each time domain signal into a time-frequency amplitude spectrum through short-time Fourier transform.

[0101] Step 3, Deep Learning Demodulation: The preprocessed time-spectrum graph is used as input and fed into a pre-trained deep learning model based on the ConvNeXt V2 architecture. The model extracts features and classifies the spectrogram graph through its hierarchical convolutional blocks and global response normalization mechanisms, and finally outputs the decoded data symbol corresponding to each signal.

[0102] Step four, data output: The symbol sequence demodulated from the model is reassembled to restore the complete original application data, completing one communication cycle. This process is repeated cyclically to achieve continuous, safe, and reliable human body vibration data communication.

[0103] Based on the same inventive concept, this invention also provides a computer device, comprising: one or more processors, and a memory for storing one or more computer programs; the programs include program instructions, and the processor executes the program instructions stored in the memory. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, used to implement one or more instructions, specifically for loading and executing one or more instructions stored in a computer storage medium to implement the above-described method.

[0104] It should be further explained that, based on the same inventive concept, the present invention also provides a computer storage medium storing a computer program, which, when executed by a processor, performs the above-described method. This storage medium can be any combination of one or more computer-readable media. A computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. A computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In the present invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0105] It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0106] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

[0107] This invention is not limited to the preferred embodiment described above. Anyone inspired by this invention can derive other forms of human body communication methods based on vibration linear spread spectrum and deep learning demodulation. All equivalent variations and modifications made within the scope of the claims of this invention shall fall within the scope of this invention.

Claims

1. A human body communication method based on vibration linear spread spectrum and deep learning demodulation, characterized in that, include: The data to be transmitted is encoded into a vibration signal using linear frequency modulation spread spectrum modulation technology, and the vibration signal is injected into the human body channel through a vibration actuator; The vibration signal is received from the human body channel through a vibration transducer, and the vibration signal is processed by time-frequency transformation to generate two-dimensional feature data. The two-dimensional feature data is input into a pre-trained deep neural network model, which classifies and identifies the two-dimensional feature data and demodulates the data symbols.

2. The human body communication method based on vibration linear spread spectrum and deep learning demodulation according to claim 1, characterized in that: The vibration signal data packet generated by the linear frequency modulation spread spectrum modulation technique includes a preamble, a frame start separator, and a data segment. The preamble is composed of multiple basic frequency sweep signals, the frame start separator is composed of reverse frequency sweep signals, and the data segment is composed of frequency modulation symbols that encode the original data.

3. The human body communication method based on vibration linear spread spectrum and deep learning demodulation according to claim 1, characterized in that: The time-frequency transformation process is a short-time Fourier transform, and the two-dimensional feature data is the signal amplitude spectrum obtained based on this short-time Fourier transform.

4. The human body communication method based on vibration linear spread spectrum and deep learning demodulation according to claim 2, characterized in that: Before performing time-frequency conversion processing on the vibration signal, a signal synchronization step is also included: detecting the preamble and frame start separator through a cross-correlation algorithm to determine the starting position of the data segment, and then dividing the data segment into independent frequency modulation symbols according to the preset symbol length.

5. The human body communication method based on vibration linear spread spectrum and deep learning demodulation according to claim 1, characterized in that: The deep neural network model is a convolutional neural network model that integrates a global response normalization mechanism.

6. The human body communication method based on vibration linear spread spectrum and deep learning demodulation according to claim 5, characterized in that: The convolutional neural network model is the ConvNeXtV2 model.

7. The human body communication method based on vibration linear spread spectrum and deep learning demodulation according to claim 1, characterized in that: The processing of the vibration signal also includes a noise suppression step: removing equipment noise through bandpass filtering and replacing outliers in the signal.

8. The human body communication method based on vibration linear spread spectrum and deep learning demodulation according to claim 1, characterized in that: The linear frequency modulation spread spectrum modulation technique defines symbol types based on the spreading factor SF and the bandwidth BW, wherein the number of coded symbol types is 2 to the power of SF, and the frequency change rate of the vibration signal is the ratio of BW to the signal duration T.

9. The human body communication method based on vibration linear spread spectrum and deep learning demodulation according to claim 1, characterized in that: The vibration actuator is a bone conduction vibration actuator, and the vibration transducer is a transducer used to collect vibrations from the body surface.

10. A human body communication system based on vibration linear spread spectrum and deep learning demodulation, characterized in that, include: Transmitting module: includes a vibration actuator and a modulation unit. The modulation unit encodes the data to be transmitted into a vibration signal using linear frequency modulation spread spectrum modulation technology. The vibration actuator injects the vibration signal into the human body channel. Receiving and preprocessing module: includes a vibration transducer and a preprocessing unit. The vibration transducer receives vibration signals from the human body channel, and the preprocessing unit performs time-frequency transformation processing on the vibration signals to generate two-dimensional feature data. Demodulation module: contains a pre-trained deep neural network model, which classifies and identifies the two-dimensional feature data and demodulates the data symbols.

Images