An optical fiber sensor vital sign monitoring method, device and system

Abstract

This invention discloses a method, device, and system for monitoring vital signs using fiber optic sensors, comprising the following steps: acquiring fiber optic signals, and generating effective sequences through detrending, filtering, and wavelet denoising; inputting the effective signals into a one-dimensional convolutional network to extract temporal features and superimposing rhythmic phase vectors to form multimodal features; calculating weights through a modal self-balancing layer to output a unified feature sequence; using a PerceiverIO network for latent feature fusion and multi-query decoding to obtain vital sign estimation results; calculating logical scores and confidence corrections through a consistency gating unit to generate early warning levels, and outputting vital sign information reports and early warning information. This invention, by introducing latent feature association and consistency gating based on PerceiverIO, achieves intelligent fusion and adaptive early warning of multiple vital sign signals, significantly improving the accuracy and reliability of fiber optic vital sign monitoring.

Description

Technical Field

[0001] This invention relates to the field of physiological signal monitoring technology, and in particular to a fiber optic sensor method, device and system for monitoring vital signs. Background Technology

[0002] With the development of fiber optic sensing technology and artificial intelligence algorithms, using fiber optic sensors to monitor human vital signs has become an important direction for intelligent healthcare, remote health management, and wearable devices. Fiber Bragg gratings, micro-bend optical fibers, and other sensing elements have advantages such as high sensitivity, strong electromagnetic interference resistance, and good embeddability, enabling the acquisition of respiratory, pulse, and blood flow-related signals without compromising comfort. Current fiber optic vital sign monitoring systems typically obtain light intensity or wavelength variation sequences through demodulation devices, and then use signal processing methods to extract vital sign parameters.

[0003] However, existing solutions generally suffer from two shortcomings: first, phase drift, noise differences, and time asynchrony exist among the optical fiber signals from multiple locations, resulting in insufficient accuracy of multi-channel data fusion; second, the vital sign estimation process often relies on a single statistical model or shallow network, making it difficult to simultaneously characterize the dynamic correlation and temporal features of different modal signals. Threshold-based or fixed-weight fusion methods are less adaptable to individual differences and environmental changes, making it difficult to achieve stable and reliable early warning output.

[0004] Therefore, how to provide a method, device, and system for monitoring vital signs using fiber optic sensors is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a method, device, and system for monitoring vital signs using fiber optic sensors. This invention utilizes a smart mattress to deploy multi-channel fiber optic sensing units at the chest, wrist, and neck locations, and introduces a unified clock synchronization mechanism to achieve time alignment and phase consistency of the fiber optic signals. Preprocessing methods such as detrending, filtering, and wavelet denoising are employed to improve signal stability. A rhythmic phase vector is introduced during the feature extraction stage to preserve the periodic characteristics of respiration and pulse. A modal self-balancing layer generates learnable weights to achieve dynamic balance of multimodal features. An improved PerceiverIO network is used in the fusion structure to establish a unified latent feature space and complete spatiotemporal fusion. A vital sign consistency gating unit is set at the output end to perform logical verification and confidence correction on the vital sign estimation results, generating tiered early warning information. This method enables high-precision fusion of fiber optic signals and intelligent early warning, improving the robustness, real-time performance, and reliability of the vital sign monitoring system.

[0006] A method for monitoring vital signs using a fiber optic sensor according to an embodiment of the present invention includes the following steps: Fiber optic sensing units are deployed on the smart mattress to collect fiber optic signals and output raw data sequences. The original data sequence is sequentially subjected to detrending, filtering and wavelet denoising. The signal quality index is calculated based on waveform stability and peak shape. When the quality index of a certain channel is lower than a set threshold, the channel data is blocked and a valid signal sequence is output. The effective signal sequence is input into a one-dimensional convolutional neural network to extract local temporal features, and the rhythm phase vector is calculated based on the main frequency and harmonic information of each channel. The vector is then superimposed on the corresponding temporal features to form a feature sequence, and a multimodal feature representation is output. The multimodal feature representation is input into the modal self-balancing layer, and a unified feature sequence is output by calculating the energy of each modal feature and introducing learnable weights. The unified feature sequence is taken as input, and the features are mapped to the latent array through the cross attention mechanism of the PerceiverIO network. The latent feature representation is then deeply fused through the self-attention layer. The latent feature representation is input into the multi-query decoding head of the PerceiverIO network to obtain the estimation result, which is then input into the symptom consistency gating unit to calculate the logical consistency score and correct the confidence level. When an abnormal coupling relationship is detected between symptom signs, a corresponding warning level signal is generated. The vital signs estimation results that have passed the consistency verification are fused with the warning level signal to generate a vital signs information report. When the warning level reaches the set threshold, the warning information is output and the original signal and estimation results for the corresponding time period are saved.

[0007] Furthermore, the step of acquiring the fiber optic signal includes: A fiber Bragg grating sensor unit is attached to the chest along the rib arch on the smart mattress to detect chest deformation signals caused by breathing. A fiber Bragg grating or micro-bend fiber optic sensing unit is attached to the wrist artery location on the smart mattress to detect distal arterial pulse signals. Fiber Bragg grating sensing units are deployed on the smart mattress at the neck or proximal end of the upper arm to detect proximal pulse signals. Each fiber optic sensing unit is connected to the input end of the demodulation device via an optical fiber. The demodulation device uses phase demodulation to sample the reflection spectrum in real time. The demodulation device is equipped with a unified clock module to perform time stamping for synchronous acquisition of multi-channel signals, and integrates the synchronized multi-channel data into an original data sequence for output.

[0008] Furthermore, the step of generating the effective signal sequence includes: The original data sequence is fitted with a sliding window using a polynomial method and baseline estimation is performed. Detrending is then performed based on the fitting results. High-pass and low-pass filtering steps are performed sequentially on the de-trended signal to confine the signal to a set frequency band. Multi-level discrete wavelet decomposition is performed on the filtered signal. The coefficients of each decomposition layer are calculated and the signal is reconstructed after threshold judgment based on the noise distribution, thus forming time-series data denoised by wavelet. Peak points, trough points, and envelope features are extracted from the denoised time series data, and the peak spacing, peak ratio, and signal stability of the waveform are calculated. The signal quality index is calculated based on the characteristics and compared with each channel. When the signal quality index of any channel is lower than the set threshold, the channel is marked and related data is stopped from participating in subsequent processing. The filtered valid signal sequence is output.

[0009] Furthermore, the generation step of the multimodal feature representation includes: The selected valid signal sequences are segmented and zero-filled according to time order, and a time index corresponding to each channel is established. Each aligned channel segment is input into the convolutional layer and nonlinear activation layer of a one-dimensional convolutional neural network, and layer-by-layer convolution, normalization and downsampling are performed to obtain a temporal feature map containing local morphological and phase changes. Residual connections and channel recalibration operations are performed on the temporal feature map to generate stacked multi-layer feature tensors while maintaining a one-to-one correspondence with the original time index; The characteristic tensor is subjected to dominant frequency analysis and spectral peak retrieval, and the dominant frequency index and harmonic index set of each channel are determined based on the autocorrelation or spectrum estimation results. The rhythm phase vector is calculated based on the set of main frequency index and harmonic index. The phase data is then expanded, normalized and time-aligned to form a phase sequence corresponding to the time step of each channel. The phase sequence is concatenated or element-wise superimposed with the corresponding temporal feature tensor in the time and channel dimensions to obtain a fusion feature sequence containing rhythm priors. The fused feature sequence is dimensionally organized, channels are arranged, and timestamps are verified to output a multimodal feature representation for subsequent modal self-balancing processing.

[0010] Furthermore, the step of generating the unified feature sequence includes: The multimodal feature representations are grouped according to modal categories. Each modal feature sequence is input into the energy analysis unit to calculate the sum of squared amplitudes and the time mean, thereby obtaining the modal energy vector and the modal center vector. The modal energy vector is input into the weight generation network, and a set of modal weights is established through fully connected layers and normalization layers. The modal weights are updated in each training or iteration cycle. The updated modal weight set is input into the linear mapping layer to perform channel-by-channel weighting on each modal feature channel in order to adjust the relative balance of feature distribution among modalities. Perform layer normalization and channel concatenation operations on the weighted feature sequence, align time steps and channel indices, and form a unified feature matrix. The unified feature matrix is ​​input into the residual connection unit for feature fusion and stabilization, while maintaining the same time index and modality mapping relationship as the input multimodal feature representation; The feature matrices after residual concatenation are recombined in chronological order to output a unified feature sequence.

[0011] Furthermore, the step of generating the latent feature representation includes: The unified feature sequence is subjected to dimensional expansion and positional encoding, and a time index vector and a modality identifier vector are added to each time step to form an input feature matrix containing temporal and modal positional information; The input feature matrix is ​​input into the cross attention unit, a set of potential query vectors is set, and attention weights are calculated by the dot product of query, key and value. The input features are weighted and mapped to the latent space to generate an initial array of potential features. The initial array of latent features is input into a multi-layer self-attention stacking module. Attention weight normalization, linear transformation and residual connection are performed in each layer to update the latent features and maintain the data continuity between layers. A feedforward network is inserted between the multi-layer self-attention stacked modules to perform nonlinear mapping and channel recombination on latent features, and layer normalization and time index alignment are performed at the output of each layer. The latent feature array updated through multiple layers is input into the latent aggregation unit, and aggregation and diffusion operations are performed to weighted merge and redistribute the features from different time steps and modal sources to generate a fused latent feature matrix. Feature compression and dimension adjustment operations are performed on the latent feature matrix to unify the shape and index relationship of the latent space, forming a fused latent feature representation.

[0012] Furthermore, the steps of the vital sign estimation and consistency gating include: The fused latent feature representation is input into the multi-query decoding module. Independent query vector sets are set for heart rate, respiratory rate and blood pressure respectively. Each query vector is matched with the latent feature matrix by dot product to generate the corresponding attention weight and the latent features are weighted and summed to obtain the initial vital sign estimation tensor. The initial vital sign estimation tensor is input into independent linear transformation layers and nonlinear activation layers respectively, and the temporal features related to each vital sign are extracted and branch outputs are formed to establish heart rate estimation branch, respiratory rate estimation branch and blood pressure estimation branch. The three branch outputs are aligned according to the time index and then input into the consistency gating unit. Feature splicing and logical vector construction are performed in this unit to form a joint representation vector that reflects the relationship between the features. An association evaluation sublayer is set up within the consistency gating unit to perform linear mapping, weight normalization, and logical operator operations on the joint representation vector to generate a logical consistency score describing the coupling relationship among the three. A confidence adjustment sublayer is set in the consistency gating unit to adjust the confidence weight of each feature estimation output according to the logical consistency score, and form an updated confidence vector. The logical consistency score and the updated confidence vector are input into the warning triggering sub-layer. In this sub-layer, a warning level signal is generated according to the logical judgment rules and a time index is attached. The estimated outputs of heart rate, respiratory rate, and blood pressure, along with the confidence vector and warning level signal, are combined into a unified vital sign estimation result.

[0013] Furthermore, the steps for reporting vital signs information and outputting early warnings include: The vital sign estimation results, confidence vector, and warning level signal are input into the report generation module, and the time index alignment and numerical integration of each vital sign parameter are performed to form a structured vital sign data table. In the report generation module, a mapping relationship is established between vital signs parameters, confidence levels, and warning levels. The corresponding fields are normalized, encoded, and tagged to generate vital signs information recording units for storage and display. A trend analysis submodule is set up in the vital signs information recording unit to perform time series combination and trend calculation on the estimated values ​​of vital signs in continuous time segments, and to combine the trend characteristics with the current estimation results to form the report content; In the report generation module, a warning judgment interface is set up to receive the warning level signal and call the alarm triggering module to perform level identification and label generation, and attach the corresponding warning level label to the vital signs information report. When the warning level reaches the set threshold, the data storage module is invoked to save the original signal, vital sign estimation results, confidence vector and warning level signal for the corresponding time period, and the warning data packet containing identification information is output to the external device. After the early warning is issued, the record management unit is invoked to update the index structure of the vital signs information database and the current vital signs information report is archived and stored.

[0014] According to an embodiment of the present invention, a fiber optic sensor vital sign monitoring device includes: a processor and a memory; the memory is used to store a computer program, and the processor calls the computer program stored in the memory to execute a fiber optic sensor vital sign monitoring method.

[0015] A fiber optic sensor-based vital sign monitoring system according to an embodiment of the present invention includes the following units: The signal acquisition unit is used to deploy fiber optic sensing units on the chest, wrist and neck of the human body being tested to acquire respiratory deformation signals, arterial pulse signals and pulse wave conduction time signals, and perform phase demodulation and time synchronization through the demodulation device to output the raw data sequence. The preprocessing unit is used to perform detrending, filtering and wavelet denoising on the original data sequence, calculate the signal quality index and shield low-quality channels, and output an effective signal sequence. The feature extraction unit is used to input the effective signal sequence into a one-dimensional convolutional neural network to extract local temporal features, and calculate the rhythm phase vector based on the main frequency and harmonic information, and superimpose it with the features of each channel to form a multimodal feature representation; The modal self-balancing unit is used to calculate the energy of each modal feature and generate learnable weights, weight normalize and concatenate multimodal features, and output a unified feature sequence. The PerceiverIO fusion unit is used to input the unified feature sequence into the PerceiverIO network, realize the fusion of latent features through cross attention and self attention mechanisms, and output the fused feature representation. The vital signs estimation and consistency gating unit is used to output the estimation results of heart rate, respiratory rate and blood pressure based on the fused feature representation, and to calculate the logical consistency score and confidence correction to generate an early warning level signal; The vital signs reporting and early warning output unit is used to generate a vital signs information report containing vital signs parameters, confidence level and early warning level. When the early warning level reaches the set threshold, it outputs early warning information and stores relevant data.

[0016] The beneficial effects of this invention are: This invention improves the basic accuracy of signal fusion from multiple locations by deploying multi-channel fiber optic sensing units in the chest, wrist, and neck, and introducing a unified clock synchronization mechanism in the demodulation device. This achieves time alignment and phase consistency of the fiber optic signals at the acquisition source. The preprocessing steps of sliding window de-trending, filtering, and multi-layer wavelet denoising effectively suppress baseline drift and environmental noise, ensuring signal stability and reliability in the subsequent feature extraction stage.

[0017] In the feature extraction stage, a rhythmic phase vector is introduced to explicitly embed the dominant frequency and harmonic information into the neural network feature space, enabling the model to recognize the temporal sequence of periodic vital signs and enhancing its perception of the patterns of respiration and pulse changes. By calculating the energy of each channel and generating learnable weights through a modal self-balancing layer, dynamic balance fusion of respiration, heart rate, and blood pressure signals is achieved, avoiding the bias of a single mode on the overall estimation.

[0018] An improved PerceiverIO network is employed in the fusion structure to establish a unified latent space representation through cross-attention and multi-layer self-attention mechanisms, achieving deep correlation and spatiotemporal fusion of multimodal features. A vital sign consistency gating unit is set at the output end, which performs self-verification of the vital sign estimation results through logical operations and confidence correction mechanisms. This enables dynamic identification of abnormal combinations and triggers tiered early warnings, improving the reliability and interpretability of the warnings. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a fiber optic sensor method, device, and system for monitoring vital signs, as proposed in this invention. Figure 2 This is a multimodal feature representation generation diagram for a fiber optic sensor vital sign monitoring method, device, and system proposed in this invention; Figure 3 This is a unified feature sequence generation diagram for a fiber optic sensor vital sign monitoring method, device, and system proposed in this invention. Detailed Implementation

[0020] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0021] refer to Figure 1 - Figure 3 A method for monitoring vital signs using a fiber optic sensor includes the following steps: Fiber optic sensing units are deployed on the chest, wrist, and neck of the human body being tested on the smart mattress to collect fiber optic signals, including respiratory deformation signals, arterial pulse signals, and pulse wave propagation time signals. The fiber optic signals are demodulated by a demodulation device to form a multi-channel time series, and time synchronization is achieved through a unified clock to output the original data sequence. The original data sequence is sequentially subjected to detrending, filtering and wavelet denoising. The signal quality index is calculated based on waveform stability and peak shape. When the quality index of a certain channel is lower than a set threshold, the data of that channel is blocked and a valid signal sequence is output. The effective signal sequence is input into a one-dimensional convolutional neural network to extract local temporal features, and the rhythm phase vector is calculated based on the main frequency and harmonic information of each channel. The rhythm phase vector is superimposed on the corresponding temporal features to form a feature sequence with physiological rhythm priors, and the multimodal feature representation embedded by the rhythm is output. The multimodal feature representation is input into the modal self-balancing layer, and a unified feature sequence after amplitude balancing is output by calculating the energy of each modal feature and introducing learnable weights. The unified feature sequence processed by modal self-balancing is taken as input, and the features are mapped to the latent array through the cross attention mechanism of the PerceiverIO network. The latent feature representation is then deeply fused through the self-attention layer and output as a fused latent feature representation. The latent feature representation is input into the multi-query decoder head of the PerceiverIO network to obtain the estimation results of heart rate, respiratory rate and blood pressure respectively. The three estimation results are input into the vital sign consistency gating unit to calculate the logical consistency score and correct the confidence. When an abnormal coupling relationship is detected between vital signs, a corresponding warning level signal is generated. The vital signs estimation results that have passed the consistency verification are fused with the warning level signal to generate a vital signs information report. The report includes heart rate, respiratory rate, blood pressure, trend information and confidence level. When the warning level reaches a set threshold, the warning information is output and the original signal and estimation results for the corresponding time period are saved.

[0022] In this embodiment, the step of acquiring fiber optic signals includes: A fiber Bragg grating sensor unit is attached to the chest along the rib arch on the smart mattress to detect chest deformation signals caused by breathing. A fiber Bragg grating or micro-bend fiber optic sensing unit is attached to the wrist artery location on the smart mattress to detect distal arterial pulse signals. Fiber Bragg grating sensing units are deployed on the smart mattress at the proximal end of the neck or upper arm to detect proximal pulse signals. Each fiber optic sensing unit is connected to the input end of the demodulation device via an optical fiber. The demodulation device uses phase demodulation to sample the reflection spectrum in real time. The demodulation device is equipped with a unified clock module to perform time stamping for synchronous acquisition of multi-channel signals, and integrates the synchronized multi-channel data into an original data sequence for output.

[0023] In this embodiment, the step of generating the effective signal sequence includes: The original data sequence is subjected to polynomial fitting and baseline estimation using a sliding window method. The sliding window is shifted segment by segment on the original fiber time series data at a fixed length. Low-order polynomial fitting is performed on the data points in each window, and the fitted curve is used as the local baseline of the window. Detrending operation is performed based on the fitting result. Specifically, the baseline is smoothed by overlapping intervals between windows. Then, the original signal of the corresponding segment is subtracted point by point using the local baseline to complete the baseline correction and detrending process. High-pass and low-pass filtering steps are performed sequentially on the de-trended signal to confine the signal to a set frequency band. Multi-level discrete wavelet decomposition is performed on the filtered signal. The coefficients of each decomposition layer are calculated and the signal is reconstructed after threshold judgment based on the noise distribution, thus forming time-series data denoised by wavelet. Peak points, trough points, and envelope features are extracted from the denoised time series data, and the peak spacing, peak ratio, and signal stability of the waveform are calculated. The signal quality index is calculated based on the characteristics and compared with each channel. When the signal quality index of any channel is lower than the set threshold, the channel is marked and related data is stopped from participating in subsequent processing. The filtered valid signal sequence is output.

[0024] In this embodiment, the step of generating the multimodal feature representation includes: The selected valid signal sequences are segmented and zero-filled in chronological order. Segmentation is achieved by dividing continuous signals into fixed-length windows. Zero-filling is used to fill in the length differences of different channels at the segment boundaries and to establish a time index corresponding to each channel. Each aligned channel segment is input into the convolutional layer and nonlinear activation layer of a one-dimensional convolutional neural network, and layer-by-layer convolution, normalization and downsampling are performed to obtain a temporal feature map containing local morphological and phase changes. Residual connections and channel recalibration operations are performed on the temporal feature map to generate stacked multi-layer feature tensors while maintaining a one-to-one correspondence with the original time index; The feature tensor is subjected to dominant frequency analysis and spectral peak retrieval. The dominant frequency analysis is completed by detecting the periodic changes of the feature tensor, and the dominant frequency index and harmonic index set of each channel are determined based on the autocorrelation or spectrum estimation results. The rhythmic phase vector is calculated based on the set of the main frequency index and harmonic index. The rhythmic phase vector is calculated from the position of the main frequency and its harmonic spacing. The phase data is expanded, normalized and time-aligned to form a phase sequence corresponding to the time step of each channel. The phase sequence is concatenated or element-wise superimposed with the corresponding temporal feature tensor in the time and channel dimensions to obtain a fusion feature sequence containing rhythm priors. The fused feature sequence is dimensionally organized, channels are arranged, and timestamps are verified to output a multimodal feature representation for subsequent modal self-balancing processing.

[0025] In this embodiment, the step of generating the unified feature sequence includes: The multimodal feature representations are grouped according to modal categories, which are divided into chest, wrist, and neck channels. Each modal feature sequence is input into the energy analysis unit to calculate the sum of squared amplitudes and the time mean, thereby obtaining the modal energy vector and the modal center vector. The modal energy vector is input into the weight generation network. A set of modal weights is established through fully connected layers and normalization layers. The weight generation network receives the modal energy vector as input, maps it through the fully connected layer to generate preliminary weights, and then adjusts the weights proportionally through the normalization layer. The modal weights are updated in each training or iteration cycle. The updated modal weight set is input into the linear mapping layer to perform channel-by-channel weighting on each modal feature channel in order to adjust the relative balance of feature distribution among modalities. Perform layer normalization and channel concatenation operations on the weighted feature sequence, align time steps and channel indices, and form a unified feature matrix. The unified feature matrix is ​​input into the residual connection unit for feature fusion and stabilization. The residual connection unit performs a transformation on the unified feature matrix and superimposes the result with the original input matrix, while maintaining the same time index and modality mapping relationship as the input multimodal feature representation. The feature matrices after residual concatenation are recombined in chronological order to output a unified feature sequence.

[0026] In this embodiment, the step of generating the latent feature representation includes: The unified feature sequence is subjected to dimensional expansion and positional encoding. Dimensional expansion is achieved by rearranging the unified feature sequence according to the channel dimension. Positional encoding inserts the encoding vector calculated from the time position at each time step. A time index vector and a modality identifier vector are added to each time step to form an input feature matrix containing temporal and modal positional information. The input feature matrix is ​​input into the cross attention unit, a set of potential query vectors is set, and attention weights are calculated by the dot product of query, key and value. The input features are weighted and mapped to the latent space to generate an initial array of potential features. The initial array of latent features is input into a multi-layer self-attention stacking module. Attention weight normalization, linear transformation and residual connection are performed in each layer to update the latent features and maintain the data continuity between layers. A feedforward network is inserted between the multi-layer self-attention stacked modules to perform nonlinear mapping and channel recombination on latent features, and layer normalization and time index alignment are performed at the output of each layer. The latent feature array updated through multiple layers is input into the latent aggregation unit, and aggregation and diffusion operations are performed to weighted merge and redistribute the features from different time steps and modal sources to generate a fused latent feature matrix. Feature compression and dimension adjustment operations are performed on the latent feature matrix to unify the shape and index relationship of the latent space, forming a fused latent feature representation.

[0027] In this embodiment, the steps of vital sign estimation and consistency gating include: The fused latent feature representation is input into the multi-query decoding module, which is implemented based on the decoding structure of PerceiverIO. Independent query vector sets are set for heart rate, respiratory rate and blood pressure respectively. Each query vector is matched with the latent feature matrix by dot product to generate the corresponding attention weight and the latent features are weighted and summed to obtain the initial vital sign estimation tensor. The initial vital sign estimation tensor is input into independent linear transformation layers and nonlinear activation layers respectively, and the temporal features related to each vital sign are extracted and branch outputs are formed to establish heart rate estimation branch, respiratory rate estimation branch and blood pressure estimation branch. The three branch outputs are aligned according to the time index and then input into the consistency gating unit. Feature splicing and logical vector construction are performed in this unit to form a joint representation vector that reflects the relationship between the features. An association evaluation sublayer is set up within the consistency gating unit to perform linear mapping, weight normalization, and logical operator operations on the joint representation vector to generate a logical consistency score describing the coupling relationship among the three. A confidence adjustment sublayer is set in the consistency gating unit to adjust the confidence weight of each feature estimation output according to the logical consistency score, and form an updated confidence vector. The logical consistency score and the updated confidence vector are input into the warning triggering sub-layer. In this sub-layer, a warning level signal is generated according to the logical judgment rules and a time index is attached. The estimated outputs of heart rate, respiratory rate, and blood pressure, along with the confidence vector and warning level signal, are combined into a unified vital sign estimation result.

[0028] In this embodiment, the steps of reporting vital signs information and outputting early warning include: The vital sign estimation results, confidence vector, and warning level signal are input into the report generation module, and the time index alignment and numerical integration of each vital sign parameter are performed to form a structured vital sign data table. In the report generation module, a mapping relationship is established between vital signs parameters, confidence level, and warning level. The mapping relationship is realized through preset field templates. The corresponding fields are normalized, encoded, and tagged to generate vital signs information recording units for storage and display. A trend analysis submodule is set up in the vital signs information recording unit to perform time series combination and trend calculation on the estimated values ​​of vital signs in continuous time segments, and to combine the trend characteristics with the current estimation results to form the report content; In the report generation module, an early warning determination interface is set up to receive the early warning level signal and call the alarm triggering module to perform level identification and identifier generation. The early warning determination interface parses the early warning level at each time step. The alarm triggering module generates an identifier field corresponding to the level according to preset rules and appends it to the corresponding position in the report data structure. The corresponding early warning level identifier is then appended to the vital signs information report. When the warning level reaches the set threshold, the data storage module is invoked to save the original signal, vital sign estimation results, confidence vector and warning level signal for the corresponding time period, and the warning data packet containing identification information is output to the external device. After the early warning is issued, the record management unit is invoked to update the index structure of the vital signs information database and the current vital signs information report is archived and stored.

[0029] A fiber optic sensor vital sign monitoring device includes a processor and a memory; the memory stores a computer program, and the processor calls the computer program stored in the memory to execute a fiber optic sensor vital sign monitoring method.

[0030] A fiber optic sensor-based vital sign monitoring system includes the following units: The signal acquisition unit is used to deploy fiber optic sensing units on the chest, wrist and neck of the human body being tested to acquire respiratory deformation signals, arterial pulse signals and pulse wave conduction time signals, and perform phase demodulation and time synchronization through the demodulation device to output the raw data sequence. The preprocessing unit is used to perform detrending, filtering and wavelet denoising on the original data sequence, calculate the signal quality index and shield low-quality channels, and output an effective signal sequence. The feature extraction unit is used to input the effective signal sequence into a one-dimensional convolutional neural network to extract local temporal features, and calculate the rhythm phase vector based on the main frequency and harmonic information, and superimpose it with the features of each channel to form a multimodal feature representation; The modal self-balancing unit is used to calculate the energy of each modal feature and generate learnable weights, weight normalize and concatenate multimodal features, and output a unified feature sequence. The PerceiverIO fusion unit is used to input the unified feature sequence into the PerceiverIO network, realize the fusion of latent features through cross attention and self attention mechanisms, and output the fused feature representation. The vital signs estimation and consistency gating unit is used to output the estimation results of heart rate, respiratory rate and blood pressure based on the fused feature representation, and to calculate the logical consistency score and confidence correction to generate an early warning level signal; The vital signs reporting and early warning output unit is used to generate a vital signs information report containing vital signs parameters, confidence level and early warning level. When the early warning level reaches the set threshold, it outputs early warning information and stores relevant data.

[0031] Example 1: To verify the feasibility of this invention in practice, it was applied to a smart home-based elderly care monitoring project. Three adult subjects were selected as subjects, and multi-channel vital signs signals were collected by deploying fiber Bragg grating sensing units on the chest, wrist, and neck. Based on the signal preprocessing, modal fusion, and PerceiverIO vital sign estimation process proposed in this invention, the estimation and early warning judgment of heart rate, respiratory rate, and blood pressure were completed.

[0032] The experiment used a multi-channel fiber optic demodulator as the data acquisition terminal, with a uniform sampling frequency set to 500Hz. Subjects maintained a natural sitting posture, with the chest sensor attached along the costal arch, the wrist sensor attached to the surface of the radial artery, and the neck sensor attached to the area adjacent to the carotid artery. All three sensors were connected to the demodulator's input via fiber optic cables, and a unified clock module was used to synchronize multi-channel sampling.

[0033] The acquired raw fiber optic signals sequentially enter the preprocessing module, which includes polynomial fitting detrending, double filtering, discrete wavelet denoising, and peak feature extraction. In this stage, the invention generates a signal quality index to filter out channels whose quality is degraded due to body movement, poor fit, or noise disturbances.

[0034] The signal quality index statistics for each channel were obtained by processing the continuous 2-minute fiber optic signals of the three subjects, as shown in the table below.

[0035] Table 1 Statistical results of signal quality index for different channels

[0036] As shown in Table 1, significant quality degradation was observed in the wrist or neck channels in subjects S2 and S3, with quality indices all below the preset threshold. The method of this invention automatically removes these low-quality channels during the preprocessing stage while maintaining the integrity of the timeline, thereby avoiding structural inconsistencies due to data gaps in the subsequent neural network input stage.

[0037] After effective channel data is segmented, aligned, and convolutionally feature-extracted, a phase sequence containing periodic structures is generated through dominant frequency analysis and rhythmic phase calculation. This invention uses a modal self-balancing layer to calculate the energy ratio of chest, wrist, and neck features, and dynamically adjusts the contribution of different modalities to the final fused features through learnable weights, giving the model better cross-scene adaptability.

[0038] The fused unified feature sequence is input into the PerceiverIO network, where cross-attention maps high-dimensional features to a latent array. The final latent features are then obtained through multi-layer self-attention stacking and a feedforward network. A multi-query decoder generates estimates of heart rate, respiratory rate, and blood pressure, respectively, and a vital sign consistency gating unit verifies the logical coupling relationship between the three vital signs.

[0039] Table 2. Statistics of Model Vital Signs Estimation Errors

[0040] As shown in Table 2, the average error of heart rate and respiratory rate remained within 2 units, indicating that the rhythm phase vector enhancement mechanism of the present invention effectively improved the ability to identify periodic signs.

[0041] The blood pressure estimation error is mainly concentrated in the range of 4–7 mmHg, which is within the reasonable range for wearable cuffless estimation, reflecting that PerceiverIO latent space fusion has good generalization ability in multimodal association learning.

[0042] The error was stable among the three subjects, demonstrating that the modal self-balancing layer has an inhibitory effect on modal shifts caused by different body surface structures and changes in attachment position.

[0043] In a test conducted on S3 subjects, deep breathing and brief changes in systolic blood pressure were intentionally introduced to verify the judgment capability of the consistency gating unit. The model identified an increase in heart rate and periodic instability, while the blood pressure characteristics were inconsistent with these. The gating unit generated a medium-level warning and recorded the original signal for that period. This process demonstrates that the present invention can promptly label abnormal coupling relationships in real-world scenarios.

[0044] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for monitoring vital signs using an optical fiber sensor, characterized in that, Includes the following steps: Fiber optic sensing units are deployed on the smart mattress to collect fiber optic signals and output raw data sequences. The original data sequence is sequentially subjected to detrending, filtering and wavelet denoising. The signal quality index is calculated based on waveform stability and peak shape. When the quality index of a certain channel is lower than a set threshold, the channel data is blocked and a valid signal sequence is output. The effective signal sequence is input into a one-dimensional convolutional neural network to extract local temporal features, and the rhythm phase vector is calculated based on the main frequency and harmonic information of each channel. The vector is then superimposed on the corresponding temporal features to form a feature sequence, and a multimodal feature representation is output. The multimodal feature representation is input into the modal self-balancing layer, and a unified feature sequence is output by calculating the energy of each modal feature and introducing learnable weights. The unified feature sequence is taken as input, and the features are mapped to the latent array through the cross attention mechanism of the PerceiverIO network. The latent feature representation is then deeply fused through the self-attention layer. The latent feature representation is input into the multi-query decoding head of the PerceiverIO network to obtain the estimation result, which is then input into the symptom consistency gating unit to calculate the logical consistency score and correct the confidence level. When an abnormal coupling relationship is detected between symptom signs, a corresponding warning level signal is generated. The vital signs estimation results that have passed the consistency verification are fused with the warning level signal to generate a vital signs information report. When the warning level reaches the set threshold, the warning information is output and the original signal and estimation results for the corresponding time period are saved.

2. The fiber optic sensor method for monitoring vital signs according to claim 1, characterized in that, The steps for acquiring fiber optic signals include: A fiber Bragg grating sensor unit is attached to the chest along the rib arch on the smart mattress to detect chest deformation signals caused by breathing. A fiber Bragg grating or micro-bend fiber optic sensing unit is attached to the wrist artery location on the smart mattress to detect distal arterial pulse signals. Fiber Bragg grating sensing units are deployed on the smart mattress at the neck or proximal end of the upper arm to detect proximal pulse signals. Each fiber optic sensing unit is connected to the input end of the demodulation device via an optical fiber. The demodulation device uses phase demodulation to sample the reflection spectrum in real time. The demodulation device is equipped with a unified clock module to perform time stamping for synchronous acquisition of multi-channel signals, and integrates the synchronized multi-channel data into an original data sequence for output.

3. The fiber optic sensor method for monitoring vital signs according to claim 1, characterized in that, The steps for generating the effective signal sequence include: The original data sequence is fitted with a sliding window using a polynomial method and baseline estimation is performed. Detrending is then performed based on the fitting results. High-pass and low-pass filtering steps are performed sequentially on the de-trended signal to confine the signal to a set frequency band. Multi-level discrete wavelet decomposition is performed on the filtered signal. The coefficients of each decomposition layer are calculated and the signal is reconstructed after threshold judgment based on the noise distribution, thus forming time-series data denoised by wavelet. Peak points, trough points, and envelope features are extracted from the denoised time series data, and the peak spacing, peak ratio, and signal stability of the waveform are calculated. The signal quality index is calculated based on the characteristics and compared with each channel. When the signal quality index of any channel is lower than the set threshold, the channel is marked and related data is stopped from participating in subsequent processing. The filtered valid signal sequence is output.

4. The fiber optic sensor method for monitoring vital signs according to claim 1, characterized in that, The steps for generating the multimodal feature representation include: The selected valid signal sequences are segmented and zero-filled according to time order, and a time index corresponding to each channel is established. Each aligned channel segment is input into the convolutional layer and nonlinear activation layer of a one-dimensional convolutional neural network, and layer-by-layer convolution, normalization and downsampling are performed to obtain a temporal feature map containing local morphological and phase changes. Residual connections and channel recalibration operations are performed on the temporal feature map to generate stacked multi-layer feature tensors while maintaining a one-to-one correspondence with the original time index; The characteristic tensor is subjected to dominant frequency analysis and spectral peak retrieval, and the dominant frequency index and harmonic index set of each channel are determined based on the autocorrelation or spectrum estimation results. The rhythm phase vector is calculated based on the set of main frequency index and harmonic index. The phase data is then expanded, normalized and time-aligned to form a phase sequence corresponding to the time step of each channel. The phase sequence is concatenated or element-wise superimposed with the corresponding temporal feature tensor in the time and channel dimensions to obtain a fusion feature sequence containing rhythm priors. The fused feature sequence is dimensionally organized, channels are arranged, and timestamps are verified to output a multimodal feature representation for subsequent modal self-balancing processing.

5. The fiber optic sensor method for monitoring vital signs according to claim 1, characterized in that, The steps for generating the unified feature sequence include: The multimodal feature representations are grouped according to modal categories. Each modal feature sequence is input into the energy analysis unit to calculate the sum of squared amplitudes and the time mean, thereby obtaining the modal energy vector and the modal center vector. The modal energy vector is input into the weight generation network, and a set of modal weights is established through fully connected layers and normalization layers. The modal weights are updated in each training or iteration cycle. The updated modal weight set is input into the linear mapping layer to perform channel-by-channel weighting on each modal feature channel in order to adjust the relative balance of feature distribution among modalities. Perform layer normalization and channel concatenation operations on the weighted feature sequence, align time steps and channel indices, and form a unified feature matrix. The unified feature matrix is ​​input into the residual connection unit for feature fusion and stabilization, while maintaining the same time index and modality mapping relationship as the input multimodal feature representation; The feature matrices after residual concatenation are recombined in chronological order to output a unified feature sequence.

6. The fiber optic sensor method for monitoring vital signs according to claim 1, characterized in that, The steps for generating the latent feature representation include: The unified feature sequence is subjected to dimensional expansion and positional encoding, and a time index vector and a modality identifier vector are added to each time step to form an input feature matrix containing temporal and modal positional information; The input feature matrix is ​​input into the cross attention unit, a set of potential query vectors is set, and attention weights are calculated by the dot product of query, key and value. The input features are weighted and mapped to the latent space to generate an initial array of potential features. The initial array of latent features is input into a multi-layer self-attention stacking module. Attention weight normalization, linear transformation and residual connection are performed in each layer to update the latent features and maintain the data continuity between layers. A feedforward network is inserted between the multi-layer self-attention stacked modules to perform nonlinear mapping and channel recombination on latent features, and layer normalization and time index alignment are performed at the output of each layer. The latent feature array updated through multiple layers is input into the latent aggregation unit, and aggregation and diffusion operations are performed to weighted merge and redistribute the features from different time steps and modal sources to generate a fused latent feature matrix. Feature compression and dimension adjustment operations are performed on the latent feature matrix to unify the shape and index relationship of the latent space, forming a fused latent feature representation.

7. The fiber optic sensor method for monitoring vital signs according to claim 1, characterized in that, The steps of the vital sign estimation and consistency gating include: The fused latent feature representation is input into the multi-query decoding module. Independent query vector sets are set for heart rate, respiratory rate and blood pressure respectively. Each query vector is matched with the latent feature matrix by dot product to generate the corresponding attention weight and the latent features are weighted and summed to obtain the initial vital sign estimation tensor. The initial vital sign estimation tensor is input into independent linear transformation layers and nonlinear activation layers respectively, and the temporal features related to each vital sign are extracted and branch outputs are formed to establish heart rate estimation branch, respiratory rate estimation branch and blood pressure estimation branch. The three branch outputs are aligned according to the time index and then input into the consistency gating unit. Feature splicing and logical vector construction are performed in this unit to form a joint representation vector that reflects the relationship between the features. An association evaluation sublayer is set up within the consistency gating unit to perform linear mapping, weight normalization, and logical operator operations on the joint representation vector to generate a logical consistency score describing the coupling relationship among the three. A confidence adjustment sublayer is set in the consistency gating unit to adjust the confidence weight of each feature estimation output according to the logical consistency score, and form an updated confidence vector. The logical consistency score and the updated confidence vector are input into the warning triggering sub-layer. In this sub-layer, a warning level signal is generated according to the logical judgment rules and a time index is attached. The estimated outputs of heart rate, respiratory rate, and blood pressure, along with the confidence vector and warning level signal, are combined into a unified vital sign estimation result.

8. The fiber optic sensor method for monitoring vital signs according to claim 1, characterized in that, The steps for reporting vital signs information and outputting early warnings include: The vital sign estimation results, confidence vector, and warning level signal are input into the report generation module, and the time index alignment and numerical integration of each vital sign parameter are performed to form a structured vital sign data table. In the report generation module, a mapping relationship is established between vital signs parameters, confidence levels, and warning levels. The corresponding fields are normalized, encoded, and tagged to generate vital signs information recording units for storage and display. A trend analysis submodule is set up in the vital signs information recording unit to perform time series combination and trend calculation on the estimated values ​​of vital signs in continuous time segments, and to combine the trend characteristics with the current estimation results to form the report content; In the report generation module, a warning judgment interface is set up to receive the warning level signal and call the alarm triggering module to perform level identification and label generation, and attach the corresponding warning level label to the vital signs information report. When the warning level reaches the set threshold, the data storage module is invoked to save the original signal, vital sign estimation results, confidence vector and warning level signal for the corresponding time period, and the warning data packet containing identification information is output to the external device. After the early warning is issued, the record management unit is invoked to update the index structure of the vital signs information database and the current vital signs information report is archived and stored.

9. A fiber optic sensor-based vital sign monitoring device, characterized in that, include: Processor and memory; The memory is used to store computer programs, and the processor calls the computer programs stored in the memory to execute the fiber optic sensor vital sign monitoring method according to any one of claims 1 to 8.

10. A fiber optic sensor-based vital sign monitoring system, characterized in that, A fiber optic sensor method for monitoring vital signs according to any one of claims 1 to 8 comprises the following units: The signal acquisition unit is used to deploy fiber optic sensing units on the chest, wrist and neck of the human body being tested to acquire respiratory deformation signals, arterial pulse signals and pulse wave conduction time signals, and perform phase demodulation and time synchronization through the demodulation device to output the raw data sequence. The preprocessing unit is used to perform detrending, filtering and wavelet denoising on the original data sequence, calculate the signal quality index and shield low-quality channels, and output an effective signal sequence. The feature extraction unit is used to input the effective signal sequence into a one-dimensional convolutional neural network to extract local temporal features, and calculate the rhythm phase vector based on the main frequency and harmonic information, and superimpose it with the features of each channel to form a multimodal feature representation; The modal self-balancing unit is used to calculate the energy of each modal feature and generate learnable weights, weight normalize and concatenate multimodal features, and output a unified feature sequence. The PerceiverIO fusion unit is used to input the unified feature sequence into the PerceiverIO network, realize the fusion of latent features through cross attention and self attention mechanisms, and output the fused feature representation. The vital signs estimation and consistency gating unit is used to output the estimation results of heart rate, respiratory rate and blood pressure based on the fused feature representation, and to calculate the logical consistency score and confidence correction to generate an early warning level signal; The vital signs reporting and early warning output unit is used to generate a vital signs information report containing vital signs parameters, confidence level and early warning level. When the early warning level reaches the set threshold, it outputs early warning information and stores relevant data.

Images