A non-contact real-time heart rate detection method and system based on BCG

Abstract

This application belongs to the field of non-contact heart rate detection, specifically disclosing a non-contact real-time heart rate detection method and system based on BCG. It includes: when a user is detected in close contact with an embedded device, in response to a heart rate detection command, acquiring a BCG signal; the embedded device is equipped with a BCG sensor; envelope detection is performed on the BCG signal to extract vital signs signals related to the user's cardiac activity; based on the vital signs signals and a first neural network model, a one-dimensional time-domain first enhanced signal is obtained; based on the vital signs signals and a second neural network model, an enhanced two-dimensional time-frequency graph is generated, and then inverse short-time Fourier transform is used to recover the one-dimensional time-domain second enhanced signal; heart rate is calculated and fused from the two one-dimensional time-domain enhanced signals to obtain the final heart rate estimate. This application enhances the signal from both the time and frequency domains before performing real-time heart rate detection, effectively improving the accuracy of heart rate detection, and offering high flexibility and good real-time performance.

Description

Technical Field

[0001] This application belongs to the field of non-contact heart rate detection, and more specifically, relates to a non-contact real-time heart rate detection method and system based on BCG (ballistocardiography). Background Technology

[0002] Heart rate detection technology has demonstrated broad application prospects and significant unique advantages in fields such as clinical screening, health monitoring, and emotion recognition. Heart rate detection technologies are mainly divided into two categories: contact and non-contact. On the one hand, current mainstream heart rate detection is still primarily based on contact detection using wearable devices, with typical examples including electrocardiogram (ECG) and photoplethysmography (PPG). These technologies have become widely popular due to the convenience of the devices, but their inherent drawbacks are not to be underestimated: firstly, prolonged wear of these devices can easily lead to skin discomfort; secondly, this technology is completely unsuitable for cases with skin damage; and finally, and most importantly, due to interference from factors such as the subject effect, this technology may face the challenge of mixed psychological and physiological signals in certain scenarios. On the other hand, non-contact heart rate detection often employs cutting-edge technologies such as radar, remote photoplethysmography (rPPG), and BCG, achieving non-intrusive detection of vital signs without direct contact with the human body and without interfering with daily life. Specifically, radar detects vital signs such as heartbeat by analyzing the echo signals reflected from the human body; rPPG uses a camera to capture periodic subtle changes in skin color caused by heartbeat and other factors; and BCG relies on thin-film sensor arrays embedded in everyday items such as cushions and mattresses to accurately capture periodic weak mechanical vibrations caused by heartbeat and other factors.

[0003] BCG-based non-contact detection has rapidly emerged as a hot topic in both academic research and industrial applications due to its advantages such as low power consumption, low cost, and high precision. Compared to radar technology, BCG devices offer greater cost-effectiveness and can be seamlessly integrated into everyday items (such as chairs, chair backs, mattresses, pillows, scales, insoles, and toilet seats), facilitating long-term continuous detection. They also exhibit strong resistance to environmental noise (such as interference from other moving objects in the testing environment). Compared to rPPG technology, BCG completely avoids the privacy risks associated with video image leakage and is unaffected by lighting conditions or individual skin color differences.

[0004] However, BCG technology still faces certain challenges: similar to radar and rPPG technologies, BCG is highly susceptible to motion artifacts in the subject (such as limb swings, violent tremors, etc.) and individual differences, leading to a decline in signal quality and thus limiting the real-time performance, robustness, and cross-scene transferability of the algorithm. Therefore, developing a non-contact heart rate detection system that combines high real-time performance, strong robustness, and high accuracy remains a core challenge that urgently needs to be overcome in this field. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the purpose of this application is to provide a non-contact real-time heart rate detection method and system based on BCG, aiming to solve the problems of interference from factors such as subject motion artifacts and individual differences in BCG-based non-contact heart rate detection technology.

[0006] To achieve the above objectives, in a first aspect, this application provides a non-contact real-time heart rate detection method based on BCG, comprising:

[0007] When a user is detected to be in close contact with the embedded device, the device acquires a BCG signal in response to a heart rate detection command. The embedded device is equipped with a BCG sensor.

[0008] Envelope detection of BCG signals is performed to extract vital signs signals related to the user's cardiac activity from the BCG signals;

[0009] Based on vital signs signals and the first neural network model, a one-dimensional time-domain first enhanced signal is obtained;

[0010] Based on vital signs signals and a second neural network model, an enhanced two-dimensional time-frequency graph is generated, and then the inverse short-time Fourier transform is used to restore it to a one-dimensional time-domain second enhanced signal.

[0011] Heart rate is calculated and fused from two one-dimensional time-domain enhanced signals to obtain the final heart rate estimate.

[0012] The first neural network model and the second neural network model are different.

[0013] Preferably, the first neural network model is based on a parallel one-dimensional convolutional neural network and Transformer fusion architecture, and includes three core modules: a feature extraction module, a multi-level CNN-Transformer fusion module, and a signal reconstruction module. Among them, the vital signs signal first passes through the feature extraction module to extract the low-level temporal features and generate a compact feature representation, and then enters the multi-level CNN-Transformer fusion module. In the parallel structure of convolution and self-attention, the local waveform morphology and global rhythm dependence are fused, and the multi-scale fused features are output to the signal reconstruction module. After convolution and mapping, the signal is restored to the temporal domain morphology and the enhanced one-dimensional temporal domain signal is output.

[0014] Preferably, the feature extraction module includes two one-dimensional convolutional layers, each followed by a batch normalization layer and a LeakyReLU activation function, and ends with an average pooling layer; the multi-level CNN-Transformer fusion module contains three CNN-Transformer fusion units connected in series, each consisting of two parallel paths: the convolutional path includes two one-dimensional convolutional layers, each followed by a batch normalization layer and a LeakyReLU activation function, and ends with an average pooling layer; the Transformer path consists of a multi-head self-attention module, a feedforward module, and a convolutional layer; the outputs of the convolutional path and the Transformer path are concatenated along the channel dimension, input to the one-dimensional convolutional layer, and normalized by BN; the signal reconstruction module includes two one-dimensional convolutional layers, each followed by a batch normalization layer and a LeakyReLU activation function, and ends with a fully connected layer, outputting a one-dimensional temporal augmented signal with the same dimension as the length of the vital sign signal.

[0015] Preferably, the first neural network model is trained end-to-end using a loss function with joint constraints in the time and frequency domains, which includes two parts: waveform reconstruction loss and heart rate consistency loss.

[0016] Waveform reconstruction loss, used to constrain the consistency of the enhanced signal and the noise-free BCG reference signal in the time domain amplitude, is defined in the form of mean square error (MSE).

[0017] Heart rate consistency loss This is used to ensure that the dominant frequency position of the enhanced signal and the reference signal is consistent in the frequency domain, thereby maintaining rhythm stability. It is defined as the square error of the dominant frequencies of the two signals.

[0018]

[0019]

[0020]

[0021] in, This refers to the number of samples during batch processing. For the first The main frequency obtained by soft-argmax of the enhanced signal samples For the first The main frequency obtained by soft-argmax of a reference signal sample For the set heart rate range, For frequency index, For the first The power spectral density of the enhanced signal, For the first The power spectral density of the reference signal, This is a temperature parameter used to adjust the smoothness of soft-argmax.

[0022] Preferably, the second neural network model adopts a deep neural network combining a two-dimensional convolutional neural network and a long short-term memory network, and includes three core modules: an encoder module, an LSTM module, and a decoder module symmetrical to the encoder. The encoder module performs a short-time Fourier transform internally, extracts features through four layers of two-dimensional convolution, and outputs a two-dimensional feature map. After the two-dimensional feature map is input into the LSTM module, it first undergoes dimensional reconstruction to obtain a one-dimensional feature vector, then applies two layers of LSTM to capture long-range temporal dependencies, and finally restores the output to a tensor consistent with the encoder output shape through an inverse transform. The decoder module adopts a symmetrical deconvolution structure and fuses the outputs of the corresponding coding layers through skip connections to generate an enhanced two-dimensional time-frequency map. Subsequently, an inverse short-time Fourier transform is performed to restore the enhanced two-dimensional time-frequency map to a one-dimensional time-domain enhanced signal.

[0023] Preferably, the encoder module includes: a time-frequency conversion and bandwidth extraction unit, used to convert the vital signs signal into a two-dimensional time-frequency map using a short-time Fourier transform (STFT) and output it to a multi-layer CNN feature extraction unit; the multi-layer CNN feature extraction unit includes four cascaded two-dimensional CNN encoder modules, each encoder module including a 2D convolutional layer, batch normalization (BN), and an exponential linear activation function, with the number of convolutional kernels increasing layer by layer, using convolutional kernels of uniform size;

[0024] The decoder module adopts a mirror-replicated encoder topology and contains four two-dimensional decoder units. Each of the first three two-dimensional decoder units is configured with a 2D deconvolution layer, BN, and ELU activation function in sequence. The fourth two-dimensional decoder unit is configured with a 2D deconvolution layer, BN, and Softplus. In the channel dimension, the number of the four convolutional kernels decreases layer by layer, forming an inverse symmetric structure with the encoder, while maintaining the same convolutional kernel size and stride as the encoder, forming a rigorous structural correspondence. The decoder module introduces a skip connection enhancement mechanism, in which the output of each encoder module is injected into the input of the corresponding decoder module through a cross-layer direct connection, forming an efficient bidirectional feature path.

[0025] Preferably, the LSTM module includes: a dimension reconstruction unit for expanding and fusing the frequency domain and channel dimensions, and permuting the tensor to align the time axis with the time step sequence of the LSTM; a two-layer LSTM processing unit for sequentially passing the transformed input through two cascaded LSTM layers, both maintaining the same configuration parameters: the input dimension and the hidden state dimension are the same; and an inverse transformation unit for inversely permuting the LSTM output to restore it to the initial feature structure.

[0026] Preferably, a time-domain signal-to-noise ratio loss is used when training the second neural network:

[0027]

[0028] in, This refers to the number of samples during batch processing. For the first One enhanced signal sample, For the first One reference signal sample, This represents the L2 norm of the signal.

[0029] To achieve the above objectives, in a second aspect, this application provides a non-contact real-time heart rate detection system based on BCG, comprising:

[0030] An embedded device, wherein a BCG sensor is provided in the embedded device for acquiring BCG signals in response to a heart rate detection command;

[0031] Memory, used to store computer programs;

[0032] The detection system is used to read and execute the computer program stored in the memory to implement the detection method as described in the first aspect.

[0033] To achieve the above objectives, in a third aspect, this application provides a computer-readable storage medium storing a computer program that, when run on an electronic device, causes the electronic device to perform the detection method as described in the first aspect.

[0034] To achieve the above objectives, in a fourth aspect, this application provides an electronic device, comprising: a processor and a memory; the memory for storing a computer program; and the processor for executing the computer program stored in the memory to cause the electronic device to perform the detection method as described in the first aspect.

[0035] It is understood that the beneficial effects of the second to fourth aspects mentioned above can be found in the relevant descriptions in the first aspect mentioned above, and will not be repeated here.

[0036] Overall, the technical solutions conceived in this application have the following beneficial effects compared with the prior art:

[0037] This application proposes a non-contact real-time heart rate detection method based on BCG, comprising: when a user is detected to be in close contact with an embedded device, in response to a heart rate detection command, acquiring a BCG signal, wherein the embedded device is equipped with a BCG sensor; performing envelope detection on the BCG signal to extract vital sign signals related to the user's cardiac activity from the BCG signal; obtaining a one-dimensional time-domain first enhanced signal based on the vital sign signals and a first neural network model; generating an enhanced two-dimensional time-frequency graph based on the vital sign signals and a second neural network model, and then recovering it into a one-dimensional time-domain second enhanced signal by inverse short-time Fourier transform; calculating and fusing the two one-dimensional time-domain enhanced signals to obtain the final heart rate estimate. This application enhances the signal from both the time and frequency domains before performing real-time heart rate detection, which can effectively improve the accuracy of heart rate detection. This method uses non-contact technology for real-time heart rate detection, which has the advantages of high flexibility and good real-time performance, making up for the deficiencies of existing heart rate detection technologies and improving the robustness and accuracy of detection. Attached Figure Description

[0038] Figure 1 A flowchart of a non-contact real-time heart rate detection method based on BCG provided in an embodiment of this application.

[0039] Figure 2 A block diagram of a non-contact real-time heart rate detection based on BCG provided in an embodiment of this application.

[0040] Figure 3 This application provides a real-time multi-process detection framework and technology roadmap based on ZeroMQ for embodiments of the present application.

[0041] Figure 4 A flowchart of time-frequency fusion heart rate estimation based on 1D time domain and 2D spectrum provided for embodiments of this application.

[0042] Figure 5 A flowchart illustrating the generation of a noise-free BCG reference signal provided in an embodiment of this application.

[0043] Figure 6 The mounting positions are for the four thin-film sensors and four load sensors used in the embodiments of this application. Detailed Implementation

[0044] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0045] In this application, the term "and / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent three cases: A existing alone, A and B existing simultaneously, and B existing alone. In this application, the symbol " / " indicates that the related objects are in an "or" relationship, for example, A / B means A or B.

[0046] In this application, the terms "first" and "second," etc., are used to distinguish different objects, not to describe a specific order of objects. For example, "first response message" and "second response message," etc., are used to distinguish different response messages, not to describe a specific order of response messages.

[0047] In the embodiments of this application, the terms "exemplary" or "for example" are used to indicate that something is an example, illustration, or description. Any embodiment or design that is described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design. Specifically, the use of the terms "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.

[0048] In the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more, for example, multiple processing units means two or more processing units, multiple elements means two or more elements, etc.

[0049] The embodiments of this application are described below with reference to the accompanying drawings.

[0050] like Figure 1 As shown, this application provides a non-contact real-time heart rate detection method based on BCG, including the following steps:

[0051] When a user is detected to be in close contact with the embedded device, the device acquires a BCG signal in response to a heart rate detection command. The embedded device is equipped with a BCG sensor.

[0052] Envelope detection of BCG signals is performed to extract vital signs signals related to the user's cardiac activity from the BCG signals;

[0053] Based on vital signs signals and the first neural network model, a one-dimensional time-domain first enhanced signal is obtained;

[0054] Based on vital signs signals and a second neural network model, an enhanced two-dimensional time-frequency graph is generated, and then the inverse short-time Fourier transform is used to restore it to a one-dimensional time-domain second enhanced signal.

[0055] Heart rate is calculated and fused from two one-dimensional time-domain enhanced signals to obtain the final heart rate estimate; the first neural network model and the second neural network model are different.

[0056] Specifically, such as Figure 2 As shown, the system proposed in this application is based on the ZeroMQ framework and relies on three processes to carry out real-time BCG data acquisition, BCG envelope detection, and heart rate detection: the first process, the BCG data acquisition module, acquires the BCG signal of the user to be tested via Bluetooth; the second process, the BCG envelope detection module, extracts vital sign signals from the BCG signal; and the third process, the BCG heart rate detection module, detects the heart rate in real time.

[0057] The system framework of this application is as follows: Figure 3 As shown, firstly, the detection system acquires human BCG signals captured by a thin-film sensor via a Bluetooth interface; secondly, the BCG signals are preprocessed and Hilbert transform is used to extract vital sign signals; finally, time-frequency domain information is fused to perform high-precision heart rate estimation.

[0058] Specifically, it includes three basic modules:

[0059] (1) BCG data acquisition module. This module captures and receives BCG data via Bluetooth and transmits the captured data to the host computer.

[0060] (1-1) Bluetooth Capture Unit. First, the Bleak library is used to establish a connection with the target BLE device, and control commands such as channel selection and gain settings are written to the specified GATT characteristics. After successful connection, the device's data characteristics are subscribed to via the GATT Notify interface, enabling new data packets collected by the device to be pushed to the local callback function in real time. In the callback, the binary data frame is parsed according to a predetermined protocol format, extracting the timestamp, channel number, and sampling rate information. The remaining sampled values ​​are then decoded into a sequence in int16 format. Subsequently, the structured result containing the channel identifier and sampled data is encapsulated as a JSON message and broadcast via ZeroMQ to the specified local port for other processes to subscribe to and process.

[0061] (1-2) Sliding Window Unit. During system initialization, when data is less than 30 seconds long, all data (regardless of length) is directly transmitted to subsequent processing modules. However, when data reaches 30 seconds, data processing and transmission are framed using a 30-second queue to ensure continuous operation of the real-time system and efficient data processing. Figure 3 As shown. Once subsequent processing is complete, the system immediately retrieves the latest data from the queue and performs analysis. This seamless loop effectively balances real-time data response speed and processing efficiency, ensuring stable system operation under varying loads.

[0062] (2) BCG envelope detection module, which extracts periodic characteristic components reflecting vital signs (such as heart rate) through preprocessing and Hilbert transform.

[0063] (2-1) Preprocessing unit, which includes a number of traditional preprocessing operations—covering normalization, detrending and smoothing filtering—aimed at improving signal stability.

[0064] (2-2) The Hilbert transform unit is optimized by deep smoothing using a Savitzky-Golay filter, and then the signal envelope is extracted by Hilbert transform to extract vital signs signals related to cardiac activity.

[0065] (3) BCG heart rate detection module, which is responsible for detecting heart rate status. Specifically, an innovative time-frequency fusion scheme is adopted, which combines signal enhancement methods based on 1D time-domain waveform analysis and 2D spectrum enhancement technology to calculate the heart rate of the two enhanced signals and fuse them to obtain the final heart rate estimate.

[0066] This module comprises three core components: a 1D time-domain signal enhancement module, a 2D spectrum signal enhancement module, and a heart rate estimation module.

[0067] (3-1) 1D Temporal Signal Enhancement Module: This module employs a parallel one-dimensional convolutional neural network fusion architecture with Transformer. It extracts local waveform details through multi-scale convolution and captures long-range rhythm dependencies using a self-attention mechanism to achieve one-dimensional signal enhancement. The following sections will provide a detailed introduction to this module from two dimensions: the main network architecture and the key components involved in model training.

[0068] (3-1-1) Main network architecture: The proposed network is based on a fusion architecture of a one-dimensional convolutional neural network and a Transformer, and includes three core modules: feature extraction block, multi-level CNN-Transformer fusion module, and signal reconstruction block. For example... Figure 4 As shown in module 2, the vital signs signal related to cardiac activity, after envelope preprocessing, is input into a parallel convolutional-transformer network. First, the feature extraction block extracts low-level temporal features and generates a compact feature representation. Then, it enters a multi-level CNN-Transformer fusion module, where local waveform morphology and global rhythm dependence are fused in a parallel structure of convolution and self-attention, outputting multi-scale fused features. These features are then fed into a signal reconstruction block, where convolution and mapping restore the temporal domain morphology, outputting an enhanced one-dimensional temporal signal.

[0069] (3-1-1-1) Feature Extraction Block: This block receives the enveloped BCG temporal signal and is used to initially encode the underlying temporal features. It contains two one-dimensional convolutional layers (Conv1D), each with a kernel size of 7, a stride of 1, padding of 3, and 16 convolutional channels. Each convolutional layer is followed by batch normalization (BN) and LeakyReLU activation functions to stabilize gradients and enhance nonlinear mapping capabilities. An average pooling layer (AvgPool1d, stride=2) is applied at the end to reduce temporal resolution while maintaining the rhythmic structure, providing a compact representation for subsequent multi-layer feature extraction.

[0070] (3-1-1-2) Multi-level CNN-Transformer fusion module: This module contains three CNN-Transformer fusion units connected in series to extract and fuse local and global features at different scales. Each fusion unit consists of two parallel paths: a convolutional path responsible for capturing local waveform morphology and short-term dynamic information, containing two one-dimensional convolutional layers (kernel size = 7, stride = 1, padding = 3) and one average pooling layer. The number of convolutional channels in the three fusion units are 32, 64, and 128, respectively. The Transformer path is used to model long-range dependencies and rhythmic structures. This path consists of a Multi-Head Attention Block (MHA Block), a FeedForward Block (FFBlock), and one convolutional layer. The MHA Block has 8 attention heads, and its feature dimensions correspond to those of the convolutional path, used to capture global dependencies between different time points; the FF Block uses a channel expansion factor of 1 and a Swish activation function to enhance non-linear expressive power. Both modules employ LayerNorm and residual connection structures to improve gradient propagation stability. The output of FFBlock is passed through a single convolutional layer (kernel size = 7, stride = 2, padding = 3, number of convolutional channels is the same as the convolutional layers in the same level of the convolutional path), Batch Normalization (BN), and LeakyReLU to achieve temporal downsampling and feature compression. The outputs of the convolutional and Transformer paths are concatenated along the channel dimension and then input into a one-dimensional convolutional layer (kernel size = 7, stride = 1, padding = 3, number of convolutional channels is the same as the convolutional layers in the same level of the convolutional path) and normalized by BN to achieve unified encoding of local-global features. Through the progressive three-layer CNN-Transformer fusion unit, the network can simultaneously perceive changes in heart rate rhythm and individual body movement patterns at different time scales, thereby improving generalization robustness against motion artifacts and individual differences.

[0071] (3-1-1-3) Signal reconstruction block is used to further refine and fuse features and enhance waveform continuity. After processing through two one-dimensional convolutional layers (kernel size = 7, stride = 1, padding = 3, number of convolutional channels = 128), the features are mapped back to the time domain through a fully connected layer (FC). The output dimension is consistent with the length of the vital sign signal after envelope preprocessing, resulting in a one-dimensional time-domain enhanced signal.

[0072] (3-1-2) Design of Model Training Related Components

[0073] (3-1-2-1) Reference Signal Generation. During the training phase, the system generates a noise-free BCG reference signal as a supervised target to train the 1D temporal augmentation network and the 2D spectral signal augmentation network in (3-2), outputting heartbeat waveform data consistent with the real heartbeat rhythm; the performance of the augmentation network is evaluated by comparing the model's output with the reference signal, such as... Figure 5 As shown.

[0074] The specific process is as follows: First, in the preprocessing stage, similar to the envelope detection module, normalization, downsampling, detrending, and smoothing filtering operations are sequentially performed on the noisy raw BCG signal; next, Hilbert transform is performed to extract the signal envelope information; finally, bandpass filtering is performed on the signal based on continuous wavelet transform (CWT): the filter uses the heart rate estimate derived from the actual electrocardiogram (ECG). Centered on; bandwidth set to .

[0075] (3-1-2-2) Loss Function Design. This application employs a loss function with joint constraints in the time and frequency domains for end-to-end training, comprising two parts: waveform reconstruction loss and heart rate consistency loss. This dual-constraint mechanism can explicitly maintain the stability of the heart rate peak position and prevent rhythm drift caused by noise. The overall loss function is defined as:

[0076]

[0077] Among them, weight .

[0078] Waveform reconstruction loss Used for constraint enhancement signals With noise-free BCG reference signal Consistency in amplitude over the time domain is defined using the mean squared error (MSE) form:

[0079]

[0080] in, Indicates the first Enhanced BCG waveform of each sample, This is the corresponding noise-free BCG reference signal. For batch size, This represents the number of sampling points for each signal segment.

[0081] Heart rate consistency loss This is used to ensure that the dominant frequency position of the frequency-enhanced signal is consistent with that of the reference signal in the frequency domain, thereby maintaining rhythm stability. The specific calculation steps are as follows:

[0082] First, for the enhanced signal With noise-free BCG reference signal Calculate the power spectral density separately , .

[0083] Secondly, within the set heart rate range The core frequency position is obtained using the soft-argmax method:

[0084]

[0085]

[0086] Finally, the heart rate consistency loss is defined as the squared error of the dominant frequencies of the two:

[0087]

[0088] in, This refers to the number of samples during batch processing. For the first The main frequency obtained by soft-argmax of the enhanced signal samples For the first The main frequency obtained by soft-argmax of a reference signal sample For the set heart rate range, For frequency index, For the first The power spectral density of the enhanced signal, For the first The power spectral density of the reference signal, This is a temperature parameter used to adjust the smoothness of soft-argmax.

[0089] This loss effectively maintains the periodic structure of the enhanced signal and the reference signal by aligning the dominant frequency position in the frequency domain, ensuring that the network does not destroy the heart rhythm information while denoising.

[0090] (3-2) 2D Spectral Signal Enhancement Module. This module uses a short-time Fourier transform (STFT) to convert the original cardiac impaction signal into a two-dimensional time-frequency graph. Subsequently, a deep neural network combining a two-dimensional convolutional neural network and a long short-term memory (LSTM) network is designed to enhance the spectral features related to cardiac dynamics. The following sections will provide a detailed introduction to this module from two dimensions: the main network architecture and the key components involved in model training.

[0091] (3-2-1) Main Network Architecture. The proposed network is based on a convolutional recurrent neural network structure and includes three core modules: an encoder module, an LSTM module, and a decoder module (symmetric to the encoder). For example... Figure 4 As shown in module 3, the preprocessed vital signs signal related to cardiac activity, after envelope transformation, is input into the encoder module. The encoder module performs a short-time Fourier transform, extracts features through four layers of two-dimensional convolution, and outputs a two-dimensional feature map. The feature map is then input into the LSTM module, where it undergoes dimensionality reconstruction to obtain a one-dimensional feature vector. Next, two layers of LSTM are applied to capture long-range temporal dependencies. Finally, at the output, an inverse transform restores the output to a tensor consistent with the encoder's output shape. The decoder module employs a symmetric deconvolution structure and fuses the outputs of corresponding coding layers through skip connections to generate an enhanced two-dimensional time-frequency map. Subsequently, an inverse short-time Fourier transform is performed to restore the enhanced time-frequency map to a one-dimensional time-domain enhanced signal.

[0092] (3-2-1-1) Encoder Module. First, time-frequency conversion and frequency band truncation. The input one-dimensional raw signal is processed by an STFT with parameters of size=256, window length=256, and step=64 to generate a two-dimensional time-frequency map, retaining only the heartbeat-related spectral components in the 0~5 Hz frequency band. Second, multi-layer CNN feature extraction. The cropped spectrum is processed sequentially by four cascaded two-dimensional CNN encoder modules. Each encoder module contains a 2D convolutional layer, Batch Normalization (BN), and an exponential linear activation function (ELU); the number of convolutional kernels increases layer by layer (16→32→64→128), and a 3×2 kernel is used uniformly. In these two-dimensional CNN encoder modules, the stride and kernel settings of the convolution are further optimized for the structural characteristics of the heartbeat spectrogram: In the frequency dimension, the kernel size is 3, the stride is 2, and the padding is 0, thereby achieving downsampling of the frequency band and aggregating the feature representation of adjacent frequency intervals to enhance the ability to perceive the frequency domain context; In the time dimension, the kernel size is 2, the stride is 1, and the padding is 1, so that the model can aggregate short-term dynamic information of adjacent time segments while maintaining the temporal resolution.

[0093] (3-2-1-2) LSTM Module. To effectively capture long-range temporal dependencies, a systematic transformation is performed on the encoded features. First, dimensionality reconstruction is performed, fusing the frequency domain and channel dimensions and permuting the tensor to align the time axis with the LSTM time step sequence. Next, a two-layer LSTM process is applied, where the transformed input passes through two cascaded LSTM layers, both maintaining the same configuration parameters: input dimension (input_size) and hidden state dimension (hidden_size) are both set to 128. Finally, an inverse transformation operation is performed, inversely permuting the LSTM output to restore the initial feature structure.

[0094] (3-2-1-3) Decoder Module. The decoder is responsible for reconstructing the features processed by LSTM into a two-dimensional time-frequency map. Its design highlights include: First, a symmetrical mirror architecture—mirroring the encoder's topology and building four two-dimensional decoder units. The first three two-dimensional decoder units are configured with 2D deconvolution layers, BN, and ELU activation functions in sequence. The fourth two-dimensional decoder unit is configured with 2D deconvolution layers, BN, and Softplus activation functions. Softplus is used to ensure the non-negativity of the output. In the channel dimension, the number of the four convolutional kernels decreases layer by layer (128→64→32→16), forming an inverse symmetrical structure with the encoder; at the same time, the same convolutional kernel size (3×2) and stride (stride=2) are maintained with the encoder, forming a rigorous structural correspondence. Second, a skip connection enhancement mechanism is introduced—to improve the gradient propagation path and enhance the efficiency of cross-layer information fusion, the output of each encoder module is injected into the input of the corresponding decoder module through a cross-layer direct connection, forming an efficient bidirectional feature path. The enhanced two-dimensional time-frequency graph output by the decoder module is restored to a one-dimensional time-domain enhanced signal by inverse short-time Fourier transform.

[0095] (3-2-2) Loss Function Design. This application designs a loss function. Although the proposed network performs enhancement in the frequency domain, the ultimate goal is to recover the time-domain signal to preserve the cardiac-related periodic structure, thereby enabling accurate HR (Heart Rate) estimation. Therefore, this study employs a time-domain signal-to-noise ratio (SNR) loss function. Its expression is as follows:

[0096]

[0097] in, It is an enhanced time-domain signal. It is the first The corresponding pure signal of each sample, This represents the L2 norm of the signal.

[0098] (3-3) Heart rate estimation module.

[0099] Both the 1D time-domain signal enhancement module and the 2D spectrum signal enhancement module output one-dimensional time-domain enhanced signals. Power spectrum analysis and peak-and-valley finding algorithms are performed on the two enhanced signals respectively to obtain the corresponding heart rate estimates. Then Enter QuantileRegressor to get the final heart rate estimate. .

[0100] Specifically, power spectral density analysis was first performed on the enhanced one-dimensional time-domain signal to locate the peak heart rate frequency within the range of 0.5–2.0 Hz. The peak frequency is used to calculate the peak distance constraint threshold. ,in, This represents the signal sampling rate. Under this constraint, peak and trough detection is performed on the time-domain signal. The time interval between adjacent peaks (or troughs) is calculated, and the resulting intervals are sorted and truncated to remove abnormal intervals caused by motion artifacts. Finally, the filtered intervals are averaged and converted into heart rate values ​​to obtain the heart rate estimation result. The heart rate estimates obtained from the two branch networks are then compared. The data is input into the QuantileRegressor model, where linear regression is used to generate the final heart rate estimate. The combination of power spectrum analysis and peak-valley finding algorithms is one example of a heart rate estimation algorithm; other methods can also be used for heart rate estimation.

[0101] Example

[0102] To comprehensively evaluate the reliability of the non-contact real-time pressure detection method and system proposed in this application, this embodiment conducted systematic experimental verification based on the open-source dataset Kansas. This authoritative dataset was collected and publicly released to the academic community in 2020 by Professor Charles Carlson's research team at Kansas State University, and includes synchronously acquired BCG, ECG, PPG, and blood pressure signal records. Data acquisition was completed using a custom-designed bed system with four large-area electromechanical thin-film sensors placed under the mattress to capture BCG signals; simultaneously, a GE Datex Ohmeda CardioCap 5 vital signs monitor was used to record ECG and other physiological signals. All signals were synchronously sampled at a frequency of 1 kHz. Data was collected from 40 subjects, aged 18 to 65 years, with a body mass index ranging from 18 to 48, exhibiting demographic diversity. The preprocessing steps implemented for this dataset included: bandpass filtering of BCG signals from 1 to 10 Hz and bandpass filtering of ECG signals from 1 to 40 Hz, aiming to eliminate baseline drift and high-frequency noise while retaining key physiological information.

[0103] In the preprocessing stage, all raw signals were divided into non-overlapping 30-second segments and uniformly downsampled to 30Hz to reduce computational load while preserving key signal features. The experimental results of this embodiment are shown in Tables 1 to 3.

[0104] (4-1) Experimental results.

[0105] To verify the effectiveness of the proposed heart rate estimation method, this application conducted evaluations based on the open-source dataset Kansas and carried out systematic comparative experiments with current mainstream state-of-the-art algorithms. The experimental results are detailed in Table 1. Four core evaluation metrics were used in the evaluation process—including mean absolute error (MAE), Pearson correlation coefficient (PCC), root mean square error (RMSE), and standard deviation (SD)—to comprehensively examine the algorithm's performance in terms of accuracy, stability, and robustness from multiple dimensions.

[0106] As shown in Table 1, the proposed method outperforms existing technologies in most evaluation metrics. Specifically, the MAE of this application reaches 0.458 bpm (beats per minute), compared to... The previously reported optimal MAE value of 0.653 bpm is improved by 29.86%, significantly enhancing heart rate estimation accuracy. Simultaneously, a high PCC value of 0.996 is maintained, indicating a strong linear correlation between the estimated value and the reference heart rate. RMSE and SD are reduced to 1.007 bpm and 0.898 bpm, respectively, effectively suppressing large biases and enhancing the stability of heart rate estimation. In summary, the experimental results confirm that the method proposed in this application can efficiently extract signal characteristics and significantly improve overall heart rate estimation performance.

[0107] Table 1 Comparison of results from different heart rate estimation methods (Kansas open-source dataset)

[0108]

[0109] (4-2) Ablation experiment.

[0110] To further verify the effectiveness of the proposed heart rate estimation method, this application conducts a systematic evaluation based on the open-source dataset Kansas and carries out ablation experiments based on the modular framework design.

[0111] To evaluate the contribution of each module design to the overall algorithm performance, this application set up the following four sets of comparative experiments, and the experimental results are shown in Table 2:

[0112] Module 1: Heart rate estimation using only the raw BCG signal reconstructed from envelope detection;

[0113] Module 1+2: First, perform envelope detection, then perform heart rate estimation based on the 1D time-domain enhanced signal waveform;

[0114] Module 1+3: First, perform envelope detection, then perform heart rate estimation based on the 2D spectrum enhancement signal;

[0115] Module 1+2+3+4 (i.e., the method proposed in this application): First, envelope detection is performed, then 1D time-domain signal enhancement and 2D spectral signal enhancement are performed respectively, and finally the heart rate estimation module generates the final heart rate estimation result.

[0116] Table 2 Ablation Experiments: Comparison of Results for Different Module Architectures (Kansas Open Source Dataset)

[0117]

[0118] As shown in Table 2, each module has a positive impact on the final algorithm performance. Specifically:

[0119] On the open-source Kansas dataset, module 1, which only includes envelope detection preprocessing, achieves a mean absolute error (MAE) of 2.644 bpm. Building on this, on the one hand, adding module 2, which includes time-domain waveform estimation, reduces the MAE to 0.574 bpm, validating the effectiveness of module 2 in extracting time-domain information from the waveform; on the other hand, adding module 3, which includes spectral enhancement estimation, reduces the MAE to 0.514 bpm, validating that module 3 effectively enhances the frequency domain characteristics of the signal.

[0120] When the results of modules 2 and 3 are integrated (i.e., using the method proposed in this application—modules 1+2+3+4), the performance is further improved: the MAE decreases to 0.458 bpm, the PCC reaches 0.996, and the standard deviation SDAE and RMSE also decrease accordingly. This indicates that the method proposed in this application can significantly improve the accuracy and stability of heart rate estimation.

[0121] In summary, the ablation experiments clearly demonstrate the contribution of each module to the model performance, further verifying the rationality and effectiveness of the method proposed in this application.

[0122] (4-3) Sensor position influence.

[0123] To further investigate the impact of sensor deployment location on heart rate estimation accuracy, this application conducted experimental research based on the open-source Kansas dataset. Figure 6 As shown, the experimental setup included: (1) four thin-film sensors arranged vertically along the head-to-tail axis, placed under the mattress; and (2) four load sensors (LC) mounted under the four corner support points of the bed frame. Our proposed algorithm calculates heart rate estimates from three types of data sources: single sensor (eight in total), sensor subgroups (thin-film array and LC array), and full sensor fusion. Comprehensive performance evaluation was conducted using multiple metrics. To ensure consistency across studies, we adopted the experimental setup of our previous study CWT-Gaus2, which reported multi-channel results for the thin-film array and LC array on MAE, SD, and RMSE metrics. The complete experimental results and comparative analysis are shown in Table 3.

[0124] For thin-film sensors, the position effect is quite pronounced: the farther away from the heart, the lower the estimation accuracy. For example, taking MAE as an example, Film 0, closest to the heart, has an accuracy of 0.458 bpm, while Film 3, the farthest, deteriorates to 2.937 bpm; correspondingly, PCC decreases from 0.996 to 0.822, RMSE increases from 1.007 bpm to 6.768 bpm, and SD shows a similar downward trend. This pattern aligns with physiological expectations, namely that signal energy and fidelity decrease with increasing distance from the heart source.

[0125] For LC sensors, position has a relatively weak effect. From LC 1 to LC 3, MAE remained between 0.572 bpm and 0.769 bpm, PCC between 0.984 and 0.990, and RMSE between 1.681 bpm and 2.135 bpm. LC 0 was an exception, with significant performance deviations due to poor bed frame contact in two subjects; after removing these two subjects, the MAE of LC 0 improved from 1.015 bpm to 0.941 bpm (corresponding to LC 0). a However, this performance is still lower than that of LC 1 to LC 3. This indicates that for LC sensors, contact quality is the dominant factor, while for thin-film sensors, geometric distance is the main influencing factor.

[0126] Compared to previous CWT-Gaus2 results based on thin films and LC, our method significantly reduces errors across most metrics. For example, in the four-thin film configuration, MAE decreases from 2.15 bpm to 0.455 bpm, SD from 2.33 bpm to 0.891 bpm, and RMSE from 2.62 bpm to 0.999 bpm, while maintaining the highest PCC. In the four-LC configuration, MAE decreases from 1.63 bpm to 0.544 bpm, SD from 1.40 bpm to 1.094 bpm, and RMSE from 2.14 bpm to 1.660 bpm. These results demonstrate that our design is highly effective in fully utilizing the thin film and load signals, and its overall performance significantly outperforms existing methods.

[0127] When all eight sensors (Film 0~3 + LC 0~3) are fused, the MAE is 0.468 bpm, which is better than the results of single-channel and multi-channel configurations on LC, but worse than the results of single-channel Film0 and multi-channel configurations on Film.

[0128] Table 3 Comparison of results at different sensor locations (Kansas open-source dataset)

[0129]

[0130] The results clearly reveal the dominant role of sensor location and contact quality in heart rate estimation accuracy: the performance of thin-film sensors decreases significantly with increasing distance from the heart, a trend highly consistent with the fundamental mechanism of physiological signal energy attenuation; while for LC sensors, the performance difference is mainly limited by contact stability, rather than geometric distance. This finding strongly suggests that thin-film sensors should be deployed closer to the heart region, while LC sensors require stable contact with the bed frame. Further multi-channel fusion experiments demonstrate that multi-sensor fusion is valuable in improving generalization robustness, but simply having more channels does not guarantee better performance; signal correlation between channels and weight optimization strategies must still be considered.

[0131] It is understood that the detailed functional implementation of each of the above units / modules can be found in the description in the aforementioned method embodiments, and will not be repeated here.

[0132] It should be understood that the above-described device is used to execute the methods in the above embodiments. The implementation principle and technical effect of the corresponding program modules in the device are similar to those described in the above methods. The working process of the device can be referred to the corresponding process in the above methods, and will not be repeated here.

[0133] Based on the methods in the above embodiments, this application provides an electronic device that may include a processor, a communications interface, a memory, and a communication bus, wherein the processor, communications interface, and memory communicate with each other via the communication bus. The processor may invoke logical instructions stored in the memory to execute the methods in the above embodiments.

[0134] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application.

[0135] Based on the methods in the above embodiments, this application provides a computer-readable storage medium storing a computer program that, when run on a processor, causes the processor to execute the methods in the above embodiments.

[0136] Based on the methods in the above embodiments, this application provides a computer program product that, when run on a processor, causes the processor to execute the methods in the above embodiments.

[0137] It is understood that the processor in the embodiments of this application can 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, transistor logic devices, hardware components, or any combination thereof. A general-purpose processor can be a microprocessor or any conventional processor.

[0138] The method steps in this application embodiment can be implemented in hardware or by a processor executing software instructions. The software instructions can consist of corresponding software modules, which can be stored in random access memory (RAM), flash memory, read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), registers, hard disks, portable hard disks, CD-ROMs, or any other form of storage medium known in the art. An exemplary storage medium is coupled to the processor, enabling the processor to read information from and write information to the storage medium. Of course, the storage medium can also be a component of the processor. The processor and the storage medium can reside in an ASIC.

[0139] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted through the computer-readable storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state disk (SSD)).

[0140] It is understood that the various numerical designations used in the embodiments of this application are merely for the convenience of description and are not intended to limit the scope of the embodiments of this application.

[0141] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A non-contact real-time heart rate detection method based on BCG, characterized in that, include: When a user is detected to be in close contact with the embedded device, the device acquires a BCG signal in response to a heart rate detection command. The embedded device is equipped with a BCG sensor. Envelope detection of BCG signals is performed to extract vital signs signals related to the user's cardiac activity from the BCG signals; Based on vital signs signals and the first neural network model, a one-dimensional time-domain first enhanced signal is obtained; Based on vital signs signals and a second neural network model, an enhanced two-dimensional time-frequency graph is generated, and then the inverse short-time Fourier transform is used to restore it to a one-dimensional time-domain second enhanced signal. Heart rate is calculated and fused from two one-dimensional time-domain enhanced signals to obtain the final heart rate estimate. The first neural network model and the second neural network model are different; The second neural network model employs a deep neural network combining a two-dimensional convolutional neural network and a long short-term memory network. It includes three core modules: an encoder module, an LSTM module, and a decoder module symmetrical to the encoder. The encoder module performs a short-time Fourier transform internally, extracts features through four layers of two-dimensional convolution, and outputs a two-dimensional feature map. After the two-dimensional feature map is input into the LSTM module, it first undergoes dimensional reconstruction to obtain a one-dimensional feature vector. Then, two layers of LSTM are applied to capture long-range temporal dependencies. Finally, at the output end, an inverse transform is performed to restore the tensor to the same form as the encoder output. The decoder module adopts a symmetrical deconvolution structure and fuses the outputs of the corresponding coding layers through skip connections to generate an enhanced two-dimensional time-frequency map. Subsequently, an inverse short-time Fourier transform is performed to restore the enhanced two-dimensional time-frequency map to a one-dimensional time-domain enhanced signal.

2. The detection method as described in claim 1, characterized in that, The first neural network model is based on a parallel one-dimensional convolutional neural network and Transformer fusion architecture, which includes three core modules: a feature extraction module, a multi-level CNN-Transformer fusion module, and a signal reconstruction module. Among them, the vital signs signal first passes through the feature extraction module to extract the low-level temporal features and generate a compact feature representation. Then it enters the multi-level CNN-Transformer fusion module, which fuses the local waveform morphology and global rhythm dependence in the parallel structure of convolution and self-attention, and outputs multi-scale fused features to the signal reconstruction module. After convolution and mapping, it is restored to the temporal domain morphology and outputs the enhanced one-dimensional temporal domain signal.

3. The detection method as described in claim 2, characterized in that, The feature extraction module contains two one-dimensional convolutional layers, each followed by a batch normalization layer and a LeakyReLU activation function, and ends with an average pooling layer. The multi-level CNN-Transformer fusion module contains three CNN-Transformer fusion units connected in series. Each CNN-Transformer fusion unit consists of two parallel paths: the convolutional path contains two one-dimensional convolutional layers, each followed by a batch normalization layer and a LeakyReLU activation function, and ends with an average pooling layer; the Transformer path consists of a multi-head self-attention module, a feedforward module, and a single convolutional layer. The outputs of the convolutional path and the Transformer path are concatenated along the channel dimension, input to the one-dimensional convolutional layer, and normalized by BN. The signal reconstruction module contains two one-dimensional convolutional layers, each followed by a batch normalization layer and a LeakyReLU activation function, and ends with a fully connected layer, outputting a one-dimensional temporal augmented signal with the same dimension as the length of the vital sign signal.

4. The detection method as described in claim 2, characterized in that, The first neural network model is trained end-to-end using a loss function that is jointly constrained by the time domain and the frequency domain, and includes two parts: waveform reconstruction loss and heart rate consistency loss. Waveform reconstruction loss, used to constrain the consistency of the enhanced signal and the noise-free BCG reference signal in the time domain amplitude, is defined in the form of mean square error (MSE). Heart rate consistency loss This is used to ensure that the dominant frequency position of the enhanced signal and the reference signal is consistent in the frequency domain, thereby maintaining rhythm stability. It is defined as the square error of the dominant frequencies of the two signals. in, This refers to the number of samples during batch processing. For the first The main frequency obtained by soft-argmax of the enhanced signal samples For the first The main frequency obtained by soft-argmax of a reference signal sample For the set heart rate range, For frequency index, For the first The power spectral density of the enhanced signal, For the first The power spectral density of the reference signal, This is a temperature parameter used to adjust the smoothness of soft-argmax.

5. The detection method as described in claim 1, characterized in that, The encoder module includes: a time-frequency conversion and bandwidth extraction unit, used to convert vital sign signals into a two-dimensional time-frequency map using short-time Fourier transform (STFT) and output it to a multi-layer CNN feature extraction unit; the multi-layer CNN feature extraction unit includes four cascaded two-dimensional CNN encoder modules, each encoder module includes a 2D convolutional layer, batch normalization (BN), and exponential linear activation function, with the number of convolutional kernels increasing layer by layer, using convolutional kernels of uniform size; The decoder module adopts a mirror-replicated encoder topology and contains four two-dimensional decoder units. Each of the first three two-dimensional decoder units is configured with a 2D deconvolution layer, BN, and ELU activation function in sequence. The fourth two-dimensional decoder unit is configured with a 2D deconvolution layer, BN, and Softplus. In the channel dimension, the number of the four convolutional kernels decreases layer by layer, forming an inverse symmetric structure with the encoder, while maintaining the same convolutional kernel size and stride as the encoder, forming a rigorous structural correspondence. The decoder module introduces a skip connection enhancement mechanism, in which the output of each encoder module is injected into the input of the corresponding decoder module through a cross-layer direct connection, forming an efficient bidirectional feature path.

6. The detection method as described in claim 1, characterized in that, The LSTM module includes: a dimension reconstruction unit, used to expand and fuse the frequency domain and channel dimensions, and to permutate the tensor to align the time axis with the time step sequence of the LSTM; a two-layer LSTM processing unit, used to pass the transformed input sequentially through two cascaded LSTM layers, both of which maintain the same configuration parameters: the input dimension and the hidden state dimension are the same; and an inverse transformation unit, used to perform inverse permutation on the LSTM output to restore it to the initial feature structure.

7. The detection method as described in claim 1, characterized in that, Temporal signal-to-noise ratio loss is used when training the second neural network. : in, This refers to the number of samples during batch processing. For the first One enhanced signal sample, For the first One reference signal sample, This represents the L2 norm of the signal.

8. A non-contact real-time heart rate detection system based on BCG, characterized in that, include: An embedded device, wherein a BCG sensor is provided in the embedded device for acquiring BCG signals in response to a heart rate detection command; Memory, used to store computer programs; The detection system is used to read and execute the computer program stored in the memory to implement the detection method as described in any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when run on an electronic device, causes the electronic device to perform the detection method as described in any one of claims 1 to 7.

10. An electronic device, characterized in that, include: Processor and memory; The memory is used to store computer programs; The processor is configured to execute a computer program stored in the memory to cause the electronic device to perform the detection method as described in any one of claims 1 to 7.

Images