A continuous monitoring single-arm multi-modal acquisition system and feature preserving denoising method

Abstract

This invention discloses a continuous monitoring single-arm multimodal acquisition system and a feature-fidelity denoising method. The system includes: an electrocardiogram (ECG) signal acquisition module, a multimodal signal acquisition module, a communication module, a power supply module, and a host computer. The method, applied to the system described above, specifically includes: using a motion signal as a reference input, performing online modeling and adaptive suppression of motion noise in the ECG signal to obtain a pre-denoised signal; inputting the pre-denoised signal and reference noise information into a pre-trained reconstruction model to generate a reconstructed signal. This invention overcomes the technical problem of balancing portability and signal acquisition quality in existing out-of-hospital scenarios. This invention can be widely applied in the field of physiological signal monitoring technology.

Description

Technical Field

[0001] This invention relates to the field of physiological signal monitoring technology, and in particular to a continuous monitoring single-arm multimodal acquisition system and a feature-fidelity denoising method. Background Technology

[0002] With the increasing demand for proactive health management, chronic disease follow-up, home rehabilitation, and outpatient early warning, physiological signal monitoring is gradually shifting from short-term, static, single-parameter detection within hospitals to long-term continuous monitoring in outpatient environments. In these application scenarios, the monitored individuals are typically in daily life, engaging in mild to moderate activity, undergoing rehabilitation training, or experiencing free positional changes. The signal acquisition environment is open, with complex sources of interference and variable wearing conditions. Compared to the relatively controlled testing environment within hospitals, continuous outpatient monitoring not only requires the device to continuously acquire physiological information but also demands good acquisition stability, ease of use, and data availability even under conditions of prolonged wear, frequent changes in movement, and multi-source noise interference. This provides a reliable data foundation for individual health assessment, abnormal event identification, and subsequent intelligent analysis.

[0003] During continuous out-of-hospital monitoring, human physiological states typically do not change in isolation, but rather fluctuate synchronously with multiple types of information, including electrocardiogram, pulse, movement status, and body surface temperature. While a single modal signal can reflect a local physiological process, it often fails to fully characterize the overall state of an individual in a dynamic environment, especially in accurately distinguishing whether signal anomalies originate from genuine physiological changes or from motion interference. Therefore, the centralized acquisition, synchronous organization, and joint analysis of multimodal physiological signals within a single wearable system would significantly improve the completeness, reliability, and interpretability of continuous out-of-hospital monitoring.

[0004] However, existing wearable monitoring technologies still have significant shortcomings in practical out-of-hospital applications. On the one hand, some existing solutions still rely on chest leads, bilateral upper limb leads, adhesive electrodes, or multi-site sensor placement. While these can achieve a certain level of signal acquisition, they generally suffer from problems such as cumbersome deployment, insufficient comfort, limited mobility, and scattered and difficult-to-integrate multimodal data under long-term wear conditions, which are not conducive to continuous use by users in non-clinical environments. On the other hand, concentrating the acquisition area on a single upper limb, while helping to improve integration and reduce the burden of wearing, also introduces new technical challenges. In particular, the ECG signal amplitude acquired under a single upper limb pathway is relatively low, belonging to unconventional weak ECG signals, which are more susceptible to body motion artifacts, electromyographic interference, contact impedance fluctuations, and environmental noise, leading to waveform distortion, feature submersion, or even ineffective use for subsequent interpretation and analysis. Summary of the Invention

[0005] In view of this, in order to solve the technical problem that existing technologies are difficult to balance portability and signal acquisition quality in continuous monitoring scenarios outside the hospital, this invention proposes a continuous monitoring single-arm multimodal acquisition system and a feature-fidelity denoising method.

[0006] In a first aspect, a continuous monitoring single-arm multimodal acquisition system includes an electrocardiogram (ECG) signal acquisition module, a multimodal signal acquisition module, a communication module, a power supply module, and a host computer, wherein: The ECG signal acquisition module includes, in sequence, a front-end buffer unit, a differential amplification unit, a graded filtering unit, and an analog-to-digital conversion unit; The front-end buffer unit is used to improve the load isolation capability of weak ECG signals from the body surface; the differential amplifier unit is used to perform preliminary extraction and enhancement of weak differential signals; the graded filtering unit is used to process the signal after the differential amplifier unit to suppress baseline drift and high-frequency noise; the analog-to-digital conversion unit is used to convert the analog signal after the graded filtering unit into a digital signal.

[0007] The multimodal signal acquisition module includes a pulse signal acquisition unit, a temperature signal acquisition unit, a motion signal acquisition unit, and a signal enhancement unit; The communication module is used to realize data exchange between the ECG signal acquisition module, the multimodal signal acquisition module and the host computer; The power supply module includes analog power and digital power, and adopts a partitioned power supply method for analog and digital power supply; The host computer receives data from the ECG signal acquisition module and the multimodal signal acquisition module, performs feature-fidelity denoising processing on the ECG signal, and outputs the reconstructed signal. Secondly, a feature-fidelity denoising method, applied to the system described above, includes: S1. Using motion signal as reference input, online modeling and adaptive suppression of motion noise in ECG signal are performed to obtain signal after preliminary denoising. Step S1 includes: using the motion signal as a reference input, establishing a noise estimation model using a second-order Wolterra series to obtain noise estimates; and recovering the denoised signal using residuals based on the noise estimates to obtain the preliminarily denoised signal.

[0008] In addition, the filtering quality of the signal after initial denoising is evaluated. If the evaluation result is not satisfactory, secondary filtering and smoothing are performed on the signal after initial denoising.

[0009] S2. Input the pre-denoised signal and reference noise information into the pre-trained reconstruction model to generate the reconstructed signal.

[0010] In step S2, the pre-training step of the reconstruction model specifically includes: constructing a reconstruction model based on a one-dimensional U-Net structure and combining it with preset constraints; using the first preliminary denoised ECG signal and the second noise reference signal as inputs to the reconstruction model, and using the third reference noiseless ECG signal as a label to train the reconstruction model.

[0011] This step involves constraints including QRS morphological constraints, noise envelope consistency loss, and frequency band statistical consistency loss.

[0012] Based on the above solution, the beneficial effects of the present invention include: 1. A front-end acquisition link with high input impedance, high common-mode rejection, and low noise was constructed, and combined with graded amplification and bandpass filtering design to improve the discernibility and signal-to-noise ratio of weak physiological signals. Compared with conventional front-end solutions designed for standard intensity signals, this invention can better adapt to the acquisition needs of weak ECG signals in restricted wearing areas, while also ensuring high-quality acquisition of other modal signals such as pulse, motion, and temperature.

[0013] 2. Based on the synchronous acquisition of motion signals by the device, a pre-stage dynamic denoising method based on motion reference input is proposed. This method, through Volterra nonlinear modeling, normalized least mean square (NLMS) online adaptive updating, and post-filtering secondary evaluation processing, can continuously track the changes in motion-related noise in dynamic activity scenarios and perform pre-stage suppression of strong interference components such as body motion artifacts. Compared with traditional fixed-parameter filtering methods or adaptive filtering methods based solely on linear relationship modeling, this invention is more suitable for handling complex interferences with time-varying, non-stationary, and weakly nonlinear coupling characteristics in out-of-hospital environments. It can maintain good denoising stability and consistency under different activity states, thereby improving the usability of weak ECG signals under dynamic monitoring conditions.

[0014] 3. This invention further proposes a post-stage feature-fidelity reconstruction method based on residual noise estimation, building upon the previous stage denoising. This method does not directly output a clean ECG signal end-to-end. Instead, it inputs the initial denoised signal and the noise reference signal into a one-dimensional U-Net model to explicitly estimate the noise components, and then reconstructs the ECG signal using residual methods. Simultaneously, this invention introduces gradient consistency constraints, QRS region masking penalties, noise envelope consistency, and frequency band statistical consistency constraints for QRS morphology preservation, reducing the risk of over-smoothing, peak attenuation, and accidental clipping of key details during denoising. Compared to existing denoising methods that focus more on overall error reduction or waveform smoothing, this invention can better preserve the P wave, QRS complex, T wave, and key morphological features related to the identification of atrial fibrillation, premature beats, and other abnormal rhythms while further reducing residual noise, thereby improving the pathological expression ability and subsequent intelligent analysis value of the processed signal. Attached Figure Description

[0015] Figure 1 This is a structural block diagram of a continuous monitoring single-arm multimodal acquisition system according to the present invention; Figure 2 This is a physical schematic diagram of a continuous monitoring single-arm multimodal acquisition system according to the present invention; Figure 3 This is a schematic diagram of the host computer transmission process of a continuous monitoring single-arm multimodal acquisition system according to the present invention; Figure 4 This is a flowchart illustrating a feature-fidelity denoising method according to the present invention; Figure 5 This is a schematic diagram of the pre-stage dynamic noise suppression process based on Volterra-NLMS in this invention; Figure 6 This is a schematic diagram of the training logic of the reconstruction model of the present invention; Figure 7 This is a schematic diagram showing the comparison of signal quality evaluation after using the present invention. Detailed Implementation

[0016] In addition to the issues mentioned in the background, noise in out-of-hospital dynamic environments also exhibits significant time-varying characteristics, multi-source coupling, and individual variability. While traditional fixed-parameter filtering, static denoising, or simple smoothing methods can reduce noise to some extent, they often struggle to simultaneously achieve noise suppression and preserve key waveform features, particularly weakening P waves, QRS complexes, T waves, and local morphological information related to rhythm abnormalities in ECG signals. For the identification of atrial fibrillation, premature beats, and other abnormal events, excessive smoothing, peak attenuation, positional shifts, or loss of local details during denoising will directly affect the accuracy and reliability of subsequent identification results. Furthermore, if multimodal signals lack synchronized organization under a unified time base, and auxiliary information such as motion is difficult to effectively participate in ECG noise modeling and state discrimination, the advantages of multimodal collaborative monitoring cannot be truly realized.

[0017] This invention aims to achieve centralized acquisition and synchronous organization of multimodal physiological signals such as electrocardiogram, pulse, movement status and body surface temperature in a single wearing area of ​​one upper limb. In addition, it constructs a hierarchical dynamic denoising mechanism that takes into account both motion noise suppression and key morphological preservation for weak physiological signals that are easily interfered with, thereby improving the stability, reliability and practicality of continuous monitoring in out-of-hospital environments.

[0018] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0019] It should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings. Unless otherwise specified, the embodiments and features described in this application can be combined with each other.

[0020] It should be understood that the terms "system," "apparatus," "unit," and / or "module" used in this application are a method of distinguishing different components, elements, parts, sections, or assemblies at different levels. However, if other terms can achieve the same purpose, they may be replaced by other expressions.

[0021] Furthermore, flowcharts are used in this application to illustrate the operations performed by the system according to embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed precisely in sequence. Instead, the steps can be processed in reverse order or simultaneously. Additionally, other operations can be added to these processes, or one or more steps can be removed from them.

[0022] Reference Figure 1 The diagram shows a structural block diagram of an optional example of the continuous monitoring single-arm multimodal acquisition system proposed in this invention. The system proposed in this embodiment may include, but is not limited to, the following parts: The electrocardiogram signal acquisition module includes a front-end buffer unit, a differential amplification unit, a graded filtering unit, and an analog-to-digital conversion unit; The multimodal signal acquisition module includes a pulse signal acquisition unit, a temperature signal acquisition unit, a motion signal acquisition unit, and a signal enhancement unit; The communication module is used to realize data exchange between the ECG signal acquisition module, the multimodal signal acquisition module, and the host computer; The host computer receives data from the ECG signal acquisition module and the multimodal signal acquisition module, performs feature-fidelity denoising processing on the ECG signal, and outputs the reconstructed signal. The power supply module provides power to all modules of the system, including digital and analog power supplies.

[0023] The hardware system of this invention is used to achieve stable, high-quality acquisition and synchronous transmission of multimodal physiological signals, especially weak electrocardiogram (ECG) signals, within the limited wearing area of ​​a single upper limb. Overall, the hardware system includes an ECG signal front-end acquisition section, a multimodal synchronous acquisition section, a transmission section, and a low-noise power supply and board-level anti-interference section. Each module is collaboratively designed around the core objective of acquiring high signal-to-noise ratio weak ECG signals.

[0024] In some feasible embodiments, the present invention uniformly limits the acquisition area of ​​multimodal physiological signals to the range of one upper limb.

[0025] Preferably, the electrodes and sensors are located in at least a portion of the upper arm, from the mid-span to the posterior segment. Unlike traditional methods that rely on separately placing electrodes and sensors on the chest, both upper limbs, or multiple body surface areas, this invention concentrates the acquisition units for ECG, pulse, motion status, and body surface temperature in the same upper limb area. This allows the system to collaboratively acquire multimodal signals within a single wearing area, thereby reducing the deployment complexity and daily usage burden caused by wearing electrodes and sensors on multiple sites. This makes it more suitable for continuous wearing and repeated use in outpatient settings.

[0026] The selection of the unilateral upper limb, especially the middle and posterior segment of the upper arm, is primarily based on the following considerations: Compared to highly mobile areas such as the wrist, this region experiences less local tissue deformation, making it easier to maintain stable contact with the skin during wear; compared to the chest region, it is easier to construct an integrated wearable device in daily life and out-of-hospital monitoring scenarios, and it also helps reduce user resistance to adhesive electrodes, exposed wires, or multiple placement locations. Simultaneously, this region can spatially accommodate the arrangement of multiple types of sensing units, allowing different modal signals to be functionally arranged according to their respective formation mechanisms, while also forming a relatively stable temporal and spatial correspondence on a unified device, creating conditions for subsequent synchronous acquisition and collaborative processing.

[0027] In terms of specific layout, the ECG signal acquisition module adopts a three-electrode single-lead configuration on the inner side of the arm ring. Two ECG signal electrodes are located at opposite positions around the upper limb, establishing a stable potential difference acquisition path within a single arm. The reference driving electrode is preferably positioned between or near the two signal electrodes to form a common-mode driving path and improve the system's ability to suppress external common-mode interference. The pulse signal acquisition unit is preferably located in the area of ​​the inner upper arm where superficial blood flow signals are more pronounced, improving the acquireability of optical pulse signals. The motion signal acquisition unit is preferably located adjacent to the main control area to reduce additional wiring and transmission delays. The temperature signal acquisition unit is preferably located independently of the main control heating area, with the sensitive element as close to the skin surface as possible to reduce thermal coupling errors.

[0028] In some feasible embodiments, the present invention preferably constructs an acquisition link with high input impedance, low noise, and high common-mode rejection for weak ECG signals. Correspondingly, the ECG signal acquisition module may include a front-end buffer unit, a differential amplification unit, a graded filtering unit, and a high-resolution analog-to-digital converter unit. The front-end buffer unit is used to improve the load isolation capability for weak ECG signals from the body surface and reduce the impact of electrode contact impedance fluctuations; the differential amplification unit is used for preliminary extraction and enhancement of weak differential signals; the graded filtering unit is used to suppress baseline drift and high-frequency noise while preserving the effective ECG frequency band; and the high-resolution analog-to-digital converter unit is used to convert the conditioned analog signal into a digital signal with high precision, thereby providing a stable input for subsequent synchronous acquisition and noise reduction processing.

[0029] Specifically, the acquisition link can be composed of a high input impedance buffer, a graded instrumentation amplifier and a 24-bit differential analog-to-digital converter, and combined with a multi-stage bandpass / lowpass filter structure to achieve stable extraction of weak ECG signals from a single arm.

[0030] An Ad8244 gain buffer is introduced as a front-end buffer module for the circuit. In addition, the Ad8422 instrumentation amplifier and its peripheral circuitry are used for signal amplification, while the Ad8244 gain buffer and its peripheral circuitry constitute a Butterworth second-order filter for signal filtering.

[0031] In some feasible embodiments, the pulse signal acquisition unit preferably uses a multi-wavelength photoplethysmography (PPG) sensor to acquire optical signals; the motion signal acquisition unit preferably uses an inertial measurement unit to acquire at least three-axis acceleration information, and may further include gyroscope information; the temperature signal acquisition unit is used to detect body surface temperature. The sampling rate of each modal link can be configured according to actual monitoring needs, thereby ensuring real-time acquisition while taking into account system power consumption and processing resource usage.

[0032] The signal enhancement unit can be implemented using the Ad8422 instrumentation amplifier.

[0033] In some feasible embodiments, to achieve unified organization and synchronous acquisition of data from different modalities, this invention introduces a multimodal synchronous acquisition design at the hardware architecture level. Preferably, a unified time base is established by the main control module, and timers, data ready signals, or interrupt mechanisms are used to uniformly schedule the acquisition links for different modalities. Specifically, the ECG signal acquisition module preferably uses a dedicated SPI interface for high-speed and stable sampling; the pulse signal acquisition unit preferably occupies a separate I²C bus; and the motion signal acquisition unit and temperature signal acquisition unit preferably share another I²C bus. By adopting a parallel communication structure of dedicated SPI plus dual I²C buses, congestion and timing conflicts caused by multiple sensors sharing a bus can be reduced, enabling multimodal signals to participate in buffering, integration, and packetization within a unified time window, providing a foundation for subsequent real-time transmission and upper-computer reconstruction.

[0034] In some feasible embodiments, a preferred approach for power supply and interference suppression design is to use a partitioned analog and digital power supply. The digital processing and wireless communication sections are powered by a single regulated power supply, while the analog acquisition link uses a separate regulated power supply branch. The analog front-end further generates a bipolar low-noise power supply to meet the symmetrical power supply requirements of instrumentation amplification and analog filtering links. By separating the analog / digital power supply, the coupling effect of digital switching noise and wireless communication noise on the weak ECG acquisition link can be reduced.

[0035] Furthermore, it is preferable to adopt design principles such as separating analog and digital grounds, isolating critical links, avoiding critical signals, and layering high-speed signals and sensitive analog signals in PCB or flexible circuit board designs to improve the overall anti-interference capability and long-term stability of the system.

[0036] In some feasible embodiments, to achieve stable wearing and reusability of the single-arm upper limb multimodal acquisition system, this invention employs an arm-ring or semi-ring-type wearing carrier as the structural basis of the device. This carrier is positioned around at least a portion of the area from the mid-section to the posterior section of one upper limb, with the inner side for contact with the skin and housing various acquisition units, and the outer side for supporting the main control, power supply, communication, and structural adjustment components. Compared to the innovation of the mechanical structure itself, this part focuses more on providing a stable spatial carrier for the aforementioned multimodal acquisition and high-quality acquisition of weak ECG data, enabling the arrangement of sensor units, electrode bonding, and module partitioning to be implemented within a single wearable structure.

[0037] To accommodate the differences in upper limb sizes among different users and improve the fit stability under dynamic wearing conditions, this invention preferably incorporates a connection structure within the armband that combines adjustability and elastic cushioning. This structure may include an elastic connector arranged circumferentially around the armband and a corresponding adjustment mechanism. The adjustment mechanism can adjust the tightness by changing the effective length of the elastic connector, enabling the armband to provide stable coverage while adapting to different body types, and reducing slippage and contact instability during slight movement or local tissue changes through elastic recovery.

[0038] Furthermore, to reduce the electrical, thermal, and mechanical coupling between system modules and adjustment mechanisms, this invention preferably adopts a partitioned modular arrangement. The carrier can be composed of multiple circumferentially distributed structural units, each accommodating different functional modules, and the overall assembly is achieved through flexible or elastic connection structures. This partitioned arrangement allows the ECG electrodes, pulse acquisition unit, and temperature unit to be closer to their respective suitable skin areas. It also helps to separate modules that may generate heat or electromagnetic interference, such as the main control and power supply, from the sensitive acquisition areas, thereby improving the overall system stability and wearing comfort.

[0039] For specific equipment details and wearing instructions, please refer to... Figure 2 ,in, Figure 2 (a) and (c) are practical equipment diagrams from different perspectives. Figure 2 (b) and (d) are schematic diagrams of device wearing from different perspectives.

[0040] In some feasible embodiments, a supporting software system is also included to realize unified scheduling, buffer management, real-time packetization, wireless transmission and synchronous restoration of multimodal signals to the host computer.

[0041] In the embedded system, a multi-task acquisition framework is preferably built based on a real-time operating system. The acquisition processes for different modalities, such as ECG, pulse, motion, and temperature, are organized into independent tasks that are scheduled under a unified time base. Different tasks execute according to their respective sampling frequencies and data readiness mechanisms, and write the sampling results to the buffer within a fixed time window. The main control module performs time organization and synchronization management of the data from each channel under a unified time base, ensuring a clear correspondence between different modalities during the output stage. The embedded software flow can be found in [reference needed]. Figure 3 .

[0042] In terms of data organization and transmission, this invention preferably adopts a unified data frame structure to encapsulate and output multimodal data within the same time window. This data frame may include a frame identifier, frame sequence number, timestamp, length information for each modality of data, and sequentially arranged data payloads. Different modal sampling results participate in unified organization during the encapsulation stage according to their corresponding sampling rate ratios, thereby ensuring that the host computer can perform real-time unpacking, reassembly, and time alignment based on the frame header information at the receiving end. The embedded terminal preferably continuously transmits the encapsulated data to the host computer via a wireless communication link. The communication method can be TCP transmission under Wi-Fi to achieve real-time monitoring and data interaction in off-site usage scenarios. On the host computer, the software system receives the data packets transmitted by the embedded terminal and performs real-time parsing, modal data restoration, unified timeline synchronous display, storage, and further processing according to a preset protocol. Through this software system, multimodal signals can not only be presented in a unified time dimension but also provide structured input data for subsequent front-end dynamic denoising and back-end feature-fidelity reconstruction.

[0043] After completing the above-mentioned device system design, this invention can simultaneously acquire multimodal physiological signals such as electrocardiogram, pulse, motion status, and body surface temperature within a single wearing area of ​​a single upper limb. Among them, the motion status signal is not only used to characterize the subject's activity level in the out-of-hospital environment, but also provides a reference basis for the subsequent identification and suppression of motion-related interference in weak electrocardiogram signals.

[0044] Reference Figure 4 This invention proposes a feature-fidelity denoising method, which can be applied to the aforementioned system, and specifically includes the following steps: Step S1: Using the motion signal as a reference input, perform online modeling and adaptive suppression of motion noise in the electrocardiogram signal to obtain the signal after preliminary denoising; Step S2: Input the pre-denoised signal and reference noise information into the pre-trained reconstruction model to generate the reconstruction signal.

[0045] Because the amplitude of ECG signals acquired via a single upper limb pathway is relatively low, they are more susceptible to motion artifacts, contact impedance fluctuations, and other motion-related noise during daily activities, rehabilitation training, or changes in body position. Therefore, before subsequent feature-fidelity reconstruction, it is preferable to preprocess motion interference with larger amplitudes and stronger correlations using a pre-stage dynamic denoising method to improve the initial quality of the input signal and reduce the burden of subsequent reconstruction. Based on this, this invention proposes a motion-reference-based pre-stage dynamic denoising method. This method uses motion signals synchronously acquired by the device as reference input to perform online modeling and adaptive suppression of motion-related noise in weak ECGs, thereby obtaining preliminary denoising results. Unlike methods that only use fixed filters or static denoising models, the pre-processing of this invention emphasizes the ability to continuously track time-varying interference in dynamic activity environments, enabling it to maintain good adaptability and stability under different activity intensities and motion states.

[0046] In some feasible embodiments, step S1 is a pre-stage denoising process based on Volterra-NLMS, which specifically includes: This method directly uses the synchronously acquired triaxial acceleration signal as the motion reference input, and no longer performs motion state characterization and channel fusion separately. Instead, it uses the Volterra-NLMS filtering framework to model and adaptively suppress motion noise online, and further performs secondary evaluation, denoising and smoothing after filtering, thereby obtaining stable and high-quality physiological signal output.

[0047] During exercise, the original physiological signals acquired can be represented as the superposition of effective physiological components and motion-related noise, i.e.: in, For a moment Noisy observation signals, For the effective physiological components that are expected to be retained, This refers to the noise component introduced by factors such as body movement, relative displacement of electrodes, and fluctuations in contact impedance.

[0048] Since this type of noise is strongly correlated with the motion state, this invention denotes the synchronously acquired triaxial acceleration signal as... And use it as the reference input for the filter.

[0049] To simultaneously characterize the linear and weakly nonlinear coupling relationships between motion noise and the reference input, this invention employs a second-order Volterra series to establish a noise estimation model, which can be expressed as follows: in, To estimate the obtained motion noise, The channel index of the reference input signal, i.e., the first... One channel, This represents the time delay index in the first-order term. This represents the memory length of a first-order Volterra term. The channel index of the reference input signal, i.e., the first... One channel, This represents the memory length of a second-order Volterra term. This represents the time index of the current sampling point. The first-order term describes the linear mapping relationship between the acceleration reference input and the noise, while the second-order term describes the weak nonlinear coupling relationship between different axes and different delay terms. and These represent the memory lengths for the first-order and second-order terms, respectively. The linear memory length is preferably 8, and the second-order delay is randomly selected from 1 to 8 sampling points. This approach balances the model's expressive power with control over feature dimensionality and online computational complexity. This represents a nonlinear time characteristic term.

[0050] To further improve online computation efficiency, this invention further expands the linear terms, square terms, cross terms, and delay terms in the above Volterra model into an extended feature vector. And the corresponding coefficients are represented as weight vectors. Therefore, the noise estimation process can be written as: in, The transpose of the matrix is ​​used. In this processing method, the linear feature block is composed of historical samples of triaxial acceleration, and the total dimension is preferably 24. The zero-delay second-order feature block is preferably 6-dimensional, and the time-delay second-order feature block is preferably 6-dimensional. Therefore, the total feature dimension after expansion is preferably 36-dimensional. To avoid update imbalance caused by differences in numerical scales among different feature blocks, this invention preferably performs root mean square normalization on the linear feature block, the zero-delay second-order feature block, and the time-delay second-order feature block. To suppress the excessive dominance of nonlinear terms on weight updates, the scaling factor of the zero-delay second-order term is preferably 0.08, and the scaling factor of the time-delay second-order term is preferably 0.04. Through the above normalization and scaling design, the stability of online recursive calculation can be improved while retaining the weak nonlinear representation capability.

[0051] After obtaining the noise estimate, the denoised signal is recovered using the residual method, i.e.: in, This represents the signal after initial noise reduction.

[0052] To enable the model to continuously correct the mapping relationship between the reference input and noise as the motion state changes, this invention employs the NLMS algorithm to update the weights online. The update formula is as follows: in, δ is the step size factor, and δ is the regularization term. Indicates the first Round weights, Indicates the first Weighting of rounds.

[0053] In a preferred embodiment, the base step size is preferably 0.03, and the regularization term is preferably [missing value]. .

[0054] To further improve the update stability under different motion states, the present invention preferably introduces a gated update mechanism based on motion amplitude.

[0055] Specifically, the motion gating threshold can be set according to the statistical distribution of the triaxial acceleration modulus, preferably using the 0.70 quantile of the acceleration modulus of the entire sequence as the gating threshold; when the motion amplitude at the current moment is lower than the threshold, in order to avoid over-updating in the resting or low motion segment, the step size is preferably reduced to 0.10 times the base step size; when the motion amplitude at the current moment is higher than the threshold, the normal step size is used for updating.

[0056] Through the closed-loop processing of "acceleration reference input - Volterra nonlinear modeling - residual output - NLMS adaptive update - gating adjustment", motion artifacts in physiological signals can be continuously tracked and suppressed in dynamic motion scenarios.

[0057] After Volterra-NLMS filtering, the output is further evaluated for denoising and smoothing to suppress residual high-frequency fluctuations and local abnormal interference, thereby improving the smoothness and stability of the output signal. This secondary processing preferably includes median filtering, bandpass smoothing, amplitude constraint, and adaptive smoothing based on QRS event recognition. Specifically, the filtered output is first subjected to a median filter of length 3 to suppress isolated spike interference; then, a fourth-order Butterworth bidirectional bandpass filter is used to smooth the signal's frequency band, preferably with a bandpass range of 0.5–45 Hz; furthermore, amplitude constraints are applied to the filtered signal, preferably limiting abnormal peak values ​​to within six times the signal standard deviation, thereby reducing the impact of local abnormal fluctuations on the overall output.

[0058] Building upon the aforementioned processing, this invention further introduces an adaptive smoothing strategy based on QRS event recognition. Different smoothing intensities are applied to the QRS and non-QRS regions to balance peak shape preservation and overall waveform smoothing. The QRS detection process preferably employs a 5–20Hz bandpass filter, a 150ms moving integration window, and a 250ms minimum RR interval constraint. The pre-guard window for the QRS region is preferably 120ms, the post-guard window is preferably 200ms, and the boundary transition width is preferably 30ms. In weakly smoothed regions, a Savitzky-Golay filter with a window length of 20ms and an order of 3 is preferred; in strongly smoothed regions, a Hampel filter with a window length of 100ms and a threshold coefficient of 3.5 is preferred, further combined with a Savitzky-Golay filter with a window length of 120ms and an order of 3 for smoothing. Through this secondary evaluation denoising and adaptive smoothing process, residual noise in non-QRS regions can be further suppressed while key morphological information such as QRS groups is better preserved, thereby improving the usability of the final output signal and the reliability of subsequent analysis.

[0059] In summary, this invention combines lightweight Volterra feature expansion, NLMS online updating, motion gating adjustment, and post-filtering evaluation smoothing to achieve pre-stage suppression of motion artifacts in physiological signals during motion scenes and to filter out significant motion interference caused by motion noise, thus providing a stable input signal for further model reconstruction. The specific process is as follows: Figure 5 As shown.

[0060] After the initial dynamic denoising process, the large-amplitude motion-related interference in the original weak ECG signal has been initially suppressed. However, in continuous out-of-hospital monitoring scenarios, residual mixed noise, local waveform distortion, and masking of low-amplitude features may still exist in the signal. Especially for weak ECG signals, even if motion artifacts are partially removed, if subsequent processing cannot further balance noise suppression and key morphological restoration, it may still affect the integrity of the expression of P waves, QRS complexes, T waves, and rhythm abnormality-related structures. Therefore, after the initial dynamic denoising, this invention further sets up a post-stage feature-fidelity reconstruction method to refine and restore the output results of the initial stage, so that the system can improve the overall signal quality while maintaining as many key physiological features as possible related to subsequent abnormality identification and clinical interpretation. Unlike the initial processing, which mainly focuses on online suppression of strong motion-related noise, the post-stage processing focuses more on further correcting residual noise and local distortion, and by introducing morphological protection constraints, it reduces the risk of excessive smoothing or erroneous clipping of key waveform structures by the deep model during the denoising process.

[0061] In some feasible embodiments, step S2 is a feature-fidelity-oriented post-stage denoising and reconstruction. It employs a residual reconstruction approach centered on noise estimation, inputting the preliminary denoising result and reference noise information into the model, outputting the noise estimation result, and then reconstructing the ECG using a residual method. This approach balances denoising effectiveness and interpretability. Specifically, it includes: A one-dimensional U-Net-based network framework is introduced to refine and restore the pre-processed ECG signal in the subsequent stage. Unlike the pre-stage dynamic denoising, which mainly focuses on online suppression of strong motion interference, this stage emphasizes further attenuation of residual mixed noise and the recovery and preservation of key morphological details of weak ECG signals. By inputting the pre-stage dynamic denoising results and auxiliary reference information into the reconstruction model, the model's ability to distinguish between effective ECG structures and residual noise components can be further improved, building upon the reduction in noise amplitude and interference complexity in the previous stage. This enhances the quality and interpretability of the final output signal.

[0062] In one embodiment, the present invention segments the original sequence using a fixed-length window. Each sample is saved as an npy file, with a fixed length of 2500 sampling points per sample. The sampling frequency is adjustable and preferably set to 250 Hz, thus each sample corresponds to a 10-second time window. The sample data is stored in a three-column structure, including one preliminary denoised ECG signal, one noise reference signal, and one noise-free reference ECG signal. Specifically, the first noisy ECG signal is used to characterize the original noisy ECG obtained from one upper limb under actual dynamic acquisition conditions; the second noise reference signal is derived from motion signal mapping results and is used to characterize noise reference information related to motion artifacts; the third noise-free reference ECG signal is derived from reference lead ECGs synchronously acquired and aligned by other devices and is used as the target reference signal during model training.

[0063] In terms of training data organization, this invention uses the first pre-denoised ECG signal and the second noisy reference signal as model input, and the third reference noise-free ECG signal as the supervision target output. Therefore, the model input tensor can be represented as a two-dimensional time series combination with the following shape: ,in The shape of the corresponding supervised target tensor is In actual training, the training is further organized into batches. and The tensor form, in which For batch size.

[0064] The core network of this invention adopts a one-dimensional U-Net structure. The network has 2 input channels and 1 output channel. The output is not a direct denoised ECG, but rather an estimation of the noise components. The basic network channel count is set to 64, forming a progressively expanding encoder-decoder structure. The encoder includes a four-level convolutional feature extraction module and a one-level intermediate bottleneck module, with channel counts of 64, 128, 256, 512, and 1024 respectively. The decoder includes a four-level upsampling recovery module, which is concatenated with the features of the corresponding layers in the encoder via skip connections. Each convolutional feature extraction module uses a combination of two layers of one-dimensional convolution, batch normalization, and ReLU activation function, with a kernel size of 3, a stride of 1, and padding of 1, to extract local temporal features while maintaining the time series length. The downsampling module uses one-dimensional max pooling with a kernel size of 2; the upsampling module uses a combination of linear interpolation and convolution to restore temporal resolution and reduce artifacts that may be caused by direct transposed convolution.

[0065] To ensure the reconstruction of various ECG features, this invention introduces an attention module in the shallow feature extraction part of the network to enhance the model's focus on key time-domain regions and important channel features. The attention module can adopt a one-dimensional CBAM structure, including a channel attention submodule and a temporal attention submodule. Channel attention generates channel descriptions through average pooling and max pooling, and obtains channel weights through compression-recovery mapping; temporal attention generates a temporal attention map through average and max projection along the channel dimension, thereby strengthening the time-domain regions related to P waves, QRS complexes, and rhythm features. This attention module can be enabled or disabled according to task requirements. In the final implementation, this invention considers it as part of the model composition.

[0066] Regarding signal reconstruction, let the noisy input ECG be... The noise reference input is Reference noise-free ECG is Network mapping is denoted as Then the model output noise estimate It can be represented as: Furthermore, the reconstructed electrocardiogram signal It can be represented as: Therefore, the model's learning objective shifts from "directly generating a clean ECG of the target" to "estimating and removing noise components," thereby improving the interpretability of the noise decomposition process and the controllability of preserving pathological features. To ensure that the reconstructed signal approximates the reference noise-free ECG as closely as possible in terms of overall waveform, while avoiding excessive perturbation of training by occasional large error points, this invention employs Charbonnier loss as the basic reconstruction loss. Its form is as follows: in, For the smoothing term, the value is... This loss is more robust to outliers than ordinary mean squared error, and has better differentiability and training smoothness than simple absolute error. To prevent the model from mistakenly treating the rapidly changing QRS structure as noise during denoising, this invention further constructs gradient consistency constraints for QRS shape preservation. The first-order difference operator is defined as: The QRS gradient preservation loss based on the reconstructed signal and the reference signal can then be expressed as: This loss enhances the protection of the fast edge structure of QRS by comparing the changing trends of the reconstructed signal and the reference signal at local rising and falling edges, and suppresses the excessive peak clipping phenomenon caused by the network's pursuit of smoothness. In this embodiment, the weighting coefficient of this loss is set to 0.5.

[0067] In addition to directly constraining the QRS morphology of the reconstructed signal, this invention further constrains the noise estimation components of the model output to prevent the model from "extracting" the ECG spike structures that should be preserved. To this end, this invention generates a soft peak map based on the noisy ECG signal and further constructs a QRS region mask. Specifically, a peak response map is first constructed through local average smoothing and threshold response, and then a QRS region mask is obtained through power-law enhancement and smoothing normalization. Subsequently, the absolute value of the first-order gradient of the noise estimation components is calculated within the corresponding region of the mask and added as a penalty term to the total loss to suppress excessively strong noise spike structures output by the model in the QRS region. This loss helps enhance the model's ability to protect the main morphology of the QRS.

[0068] This invention further incorporates noise envelope consistency loss and frequency band statistical consistency loss. The noise envelope consistency loss calculates the absolute value moving average envelope for both the noise estimate and the mapped noise reference, comparing their differences to ensure the model's output noise estimate matches the overall energy variation trend with the noise reference. The frequency band statistical consistency loss decomposes the signal into low-frequency smoothing components and high-frequency residual components, calculating their root mean square values ​​separately. It then compares the consistency between the noise estimate and the noise reference in low-frequency and high-frequency energy distributions, thereby enhancing the statistical rationality of the model's output noise.

[0069] Furthermore, this invention can introduce other additional constraints according to specific application needs, including but not limited to peak correlation constraints, atrioventricular fidelity constraints, attention alignment constraints, rhythm consistency constraints, soft peak constraints, and gradient preservation and anti-mining penalty constraints. Among these, peak correlation constraints are used to enhance the consistency between the reconstructed signal and the reference signal at the rhythm peak positions; atrioventricular fidelity constraints are used to enhance the model's ability to preserve P-waves and atrial-related regions; and attention alignment constraints are used to ensure that the shallow attention map is consistent with the atrial window distribution, thereby improving the model's ability to focus on anomalous rhythm-related time-domain segments. All of the above modules are considered part of the overall method of this invention in the final implementation.

[0070] In summary, the total loss function of this invention can be expressed as: in, Based on the loss of reconstruction, To reconstruct the signal gradient consistency loss, The gradient penalty term for the QRS region in the noise estimation. For noise envelope consistency loss, For the statistical consistency loss of the noise frequency band, For other optional additional constraints, These are the corresponding weighting coefficients. In the current embodiment of this invention, the basic reconstruction loss, QRS gradient preservation loss, noise envelope consistency loss, noise band statistical consistency loss, noise QRS region gradient penalty term, and attention-related and peak-plot-related constraints are all considered as part of the overall framework.

[0071] In terms of training strategy, this invention adopts a supervised learning approach, using Adam as the training optimizer and setting the initial learning rate to [value missing]. The batch size was set to 16, and the number of training epochs was set to 100. To improve convergence stability during training, an adaptive learning rate decay strategy based on the change in validation set loss was further adopted. When the test loss did not decrease significantly over several consecutive epochs, the learning rate was automatically reduced. During training, the model parameters and the parameters of the noisy reference mapper were updated synchronously. After each training epoch, the total loss was calculated on the test set, and the model parameters with the best performance on the test set were saved as the final model.

[0072] By combining the aforementioned data organization methods, network structure design, residual noise estimation relationships, key morphological protection losses, and multiple statistical consistency constraints, this invention can effectively reconstruct noisy electrocardiogram (ECG) signals in a real-world dynamic monitoring scenario of a single upper limb. While suppressing motion artifacts, electromyographic interference, baseline drift, and other multi-source mixed noise, it preserves as much as possible the P wave, QRS complex, T wave, and key pathological features related to the identification of atrial fibrillation, premature beats, and other abnormal events. This provides higher-quality input signals for subsequent abnormality detection, rhythm analysis, and clinical interpretation. The algorithm training logic can be found in [reference needed]. Figure 6 .

[0073] In some feasible embodiments, step S2, the reasoning stage, specifically includes: The denoised ECG signal and the reference noise information are aligned in the same time window and used as dual-channel time series data input into the pre-trained reconstruction model. Within the model: ECG main features and noise-related features are jointly extracted through the encoding structure, and the temporal resolution is restored through the decoding structure, finally outputting the residual noise estimation result.

[0074] The residual noise estimation result is subtracted from the signal after initial denoising to obtain the final reconstructed ECG signal.

[0075] The QRS morphology preservation constraint, noise envelope consistency constraint, and frequency band statistical consistency constraint set during the training phase are not calculated separately during inference. Instead, they are implemented through the pre-trained model parameters, which make the model tend to remove interference components that are consistent with the reference noise during denoising, while preserving key ECG morphologies such as P waves, QRS complexes, and T waves.

[0076] In some feasible embodiments, considering that the second noise reference signal originates from a motion signal mapping and may differ from the true noise components in scale, bias, and phase, this invention does not require the noise estimation to be consistent with the noise reference point-by-point. Instead, it introduces a learnable one-dimensional linear mapper to adaptively map the noise reference signal and then maintain statistical consistency with the model output noise. The mapper employs... A one-dimensional convolutional structure is used to learn the magnitude scaling and bias correction relationships between the noise reference and the noise estimate.

[0077] In some feasible embodiments, to reduce the impact of amplitude differences between different devices, individuals, and measurement ranges on training stability, this invention performs standardization processing on the input and reference signals during the training and inference phases. The standardization method employs robust z-score, which calculates the median, first quartile, and third quartile for each signal and normalizes them based on the interquartile range. Compared to ordinary mean-standard deviation standardization, this method is less sensitive to abnormal peaks and outliers, making it more suitable for preprocessing noisy ECGs in dynamic acquisition scenarios.

[0078] Signal quality assessment has been conducted for the dynamic denoising method proposed in this invention. In the assessment, the original noisy ECG signal was used as input, and dynamic denoising based on motion reference input was applied. The signal quality before and after denoising was then compared and analyzed. In the table, "RAW" represents the original signal, and "DEN" represents the signal result processed by the denoising method of this invention.

[0079] This evaluation employed multiple signal quality metrics, including coverage, mean absolute error (MAE), normalized mean square error (NMSE), percentage root-mean-square difference (PRD), and signal-to-noise ratio improvement (SNR improvement), to measure the changes in signal quality before and after denoising from different perspectives. Figure 7 The results shown include, Figure 7 (a) is a comparison diagram of Coverage. Figure 7 (b) is a diagram showing the comparison of NMSE. Figure 7 (c) is a comparison diagram of PRDs. Figure 7(d) is a schematic diagram of SNR improvement comparison. In 22 sets of test data, after processing by the method of this invention, multiple evaluation indicators showed a significant improvement trend. The average results show that: Coverage increased from 0.81 of the original signal to 0.88, indicating that the identifiability of the effective signal has been improved; MAE decreased from 24.70 to 2.60, indicating that the error between the denoised signal and the target signal has been significantly reduced; NMSE decreased from 0.265 to 0.012, indicating that the normalization error has been significantly reduced; PRD decreased from 35.14 to 11.57, indicating that the distortion level has been significantly reduced; at the same time, the average SNR-improvement reached 12.93dB, indicating that the method of this invention can effectively improve the signal-to-noise ratio.

[0080] As can be further seen from Table 1, the method of this invention can achieve relatively stable signal quality improvement on most test subjects, indicating that the proposed denoising method has a good suppression effect on motion-related noise. The above results show that the dynamic denoising method proposed in this invention can reduce motion artifacts and interference components in the original signal while maintaining the main morphological characteristics of the effective electrocardiogram signal, thereby improving the signal stability, recognizability, and reliability of subsequent analysis.

[0081] Table 1 Noise Reduction Data Test A storage medium storing processor-executable instructions, which, when executed by a processor, are used to implement a feature-fidelity denoising method as described above.

[0082] The content of the above method embodiments is applicable to this storage medium embodiment. The specific functions implemented in this storage medium embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0083] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. A continuous monitoring single-arm multimodal acquisition system, characterized in that, It includes an ECG signal acquisition module, a multimodal signal acquisition module, a communication module, and a host computer, among which: The electrocardiogram signal acquisition module includes, in sequence, a front-end buffer unit, a differential amplification unit, a graded filtering unit, and an analog-to-digital conversion unit; The multimodal signal acquisition module includes a pulse signal acquisition unit, a temperature signal acquisition unit, a motion signal acquisition unit, and a signal enhancement unit, wherein the pulse signal acquisition unit, the temperature signal acquisition unit, and the motion signal acquisition unit are connected to the signal enhancement unit; The communication module is used to realize data exchange between the electrocardiogram signal acquisition module, the multimodal signal acquisition module and the host computer; The host computer is used to receive data from the ECG signal acquisition module and the multimodal signal acquisition module, perform feature-fidelity denoising processing on the ECG signal, and output the reconstructed signal.

2. The continuous monitoring single-arm multimodal acquisition system according to claim 1, characterized in that: The differential amplifier unit is used to initially extract and enhance the electrocardiogram signal to obtain the enhanced signal; The graded filtering unit is used to process the enhanced signal, suppress baseline drift and high-frequency noise, and obtain a conditioned signal; The analog-to-digital converter is used to convert the conditioned signal into a digital signal.

3. The continuous monitoring single-arm multimodal acquisition system according to claim 1, characterized in that, Also includes: The power supply module includes a digital power supply and an analog power supply, wherein the digital power supply and the analog power supply provide power in separate zones.

4. A feature-fidelity denoising method, characterized in that, The system is applied to a continuous monitoring single-arm multimodal acquisition system as described in claim 1, comprising: Using motion signals as reference input, motion noise in electrocardiogram signals is modeled and adaptively suppressed online to obtain a preliminarily denoised signal. The pre-denoised signal and reference noise information are input into the pre-trained reconstruction model to generate the reconstructed signal.

5. The feature-fidelity denoising method according to claim 4, characterized in that, The step of using motion signals as reference input to perform online modeling and adaptive suppression of motion noise in electrocardiogram signals to obtain a preliminarily denoised signal specifically includes: Using motion signals as reference input, a noise estimation model is established using a second-order Wolterra series to obtain noise estimates. Based on the noise estimate, the denoised signal is recovered using a residual method to obtain the preliminarily denoised signal.

6. The feature-fidelity denoising method according to claim 5, characterized in that, The weights in the noise estimation model are updated using the normalized least mean square algorithm, and the update formula is as follows: in, Indicates the first Round weights, Indicates the first Round weights, Indicates the step size factor. This represents the signal after initial noise reduction. This represents the extended eigenvector of the noise estimation model. This indicates a regular term.

7. The feature-fidelity denoising method according to claim 5, characterized in that, Also includes: The filtering quality of the initially denoised signal is evaluated. If the evaluation result is not satisfactory, the initially denoised signal is subjected to secondary filtering and smoothing.

8. The feature-fidelity denoising method according to claim 4, characterized in that, The pre-training steps of the reconstruction model specifically include: Based on the one-dimensional U-Net structure, a reconstruction model is constructed by combining preset constraints; The first pre-denoised ECG signal and the second noise reference signal are used together as inputs to the reconstruction model, and the third reference noise-free ECG signal is used as a label to train the reconstruction model.

9. The feature-fidelity denoising method according to claim 8, characterized in that, During training, the loss function of the reconstruction model is: in, Represents the total loss function. Based on the loss of reconstruction, To reconstruct the signal gradient consistency loss, The gradient penalty term for the QRS region in the noise estimation. For noise envelope consistency loss, For the statistical consistency loss of the noise frequency band, For other additional constraints, These are the corresponding weighting coefficients.

Images