A noise-resistant knowledge distillation and adaptive measurement method for low-power edge blood pressure monitoring devices

Abstract

This invention discloses a noise-resistant knowledge distillation and adaptive measurement method for low-power edge blood pressure monitoring devices. It injects medical physics prior noise into a student model by constructing a structured physiological noise generator (SPNG), trains a minimal parameter student model using asymmetric KL divergence distillation loss (PNIRD), and embeds a domain adversarial adaptive network (DANN) based on gradient inversion layer (GRL) to achieve cross-device domain alignment from arm to wrist. Finally, it optimizes edge MCU deployment through physiologically perceptual asymmetric quantization and operator-folded LUT lookup table. This method effectively suppresses the expansion of prediction standard deviation caused by real noise in the model and supports unlabeled cross-hardware transfer.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence-assisted medical monitoring and edge intelligent deployment technology, specifically involving a method, device and system for compressing, distilling and aligning a highly robust blood pressure estimation model across device domains in a low-computing-power chip environment. Background Technology

[0002] With the synergistic evolution of integrated circuit technology and biosensors, deploying high-precision blood pressure prediction deep models directly from the cloud to portable wearable devices (such as smartwatches, ear-worn monitors, or miniature cuff monitors) has become a revolutionary trend in the field of digital healthcare. However, the computing resources of edge microprocessors (MCU / DSP / NPU) are extremely scarce (typically SRAM < 1MB, computing power < 100 GOPS), which requires the original model to undergo drastic parameter compression and knowledge distillation (KD) to transfer knowledge from a giant "teacher model" to a very small "student model".

[0003] Existing knowledge distillation techniques perform well in general fields such as internet image recognition and voice interaction, but they exhibit a severe "precision brittleness" defect in clinical pressure measurement scenarios, which are extremely sensitive to noise. Large teacher models, with their redundant high-order parameters, can form a strong nonlinear filtering barrier, thereby smoothing out waveform noise caused by sensor baseline drift, environmental electromagnetic interference, or even minute limb tremors. However, due to the extremely limited parameter capacity of student models, traditional "pure distillation" methods cause student models to lose their ability to isolate and defend against these inherent medical morphological artifacts. As a result, when running on real noisy electronic hardware platforms, the predicted standard deviation (SD) index will rise sharply, often failing to cross the error tolerance limit specified by the ISO 81060-2 clinical gold standard.

[0004] Furthermore, blood pressure measurement devices exhibit significant differences in hardware form factors, typically ranging from high-performance arm-cuff sensors (usually used as source domain data for teacher models) to portable wrist sensors (usually used as target domain applications for student models). Due to physical differences between the brachial and radial arteries in anatomical depth, vascular impedance, and propagation and reflection mechanisms, the raw pulse signals acquired by the two arteries show significant physiological characteristic shifts. Traditional knowledge distillation architectures lack mechanisms to handle this "arm-wrist" morphological consistency, resulting in significant systematic bias (Mean Error, ME) when high-precision knowledge learned on large models is transferred to edge hardware with different forms, greatly hindering the generalization and widespread adoption of edge-side pressure measurement algorithms.

[0005] Currently, while there are attempts in the industry to apply pruning or quantization to medical model compression, these methods are essentially "post-hoc remedies" and fail to inject defense against physiological noise into the very essence of distillation—the knowledge transfer path. Furthermore, adaptation for different hardware components often relies on costly and time-consuming re-labeling, lacking an adaptive technical framework that can leverage existing cloud-based knowledge for unsupervised or weakly supervised cross-domain alignment. Summary of the Invention

[0006] This invention provides a noise-resistant knowledge distillation and adaptive measurement method for low-power edge blood pressure monitoring devices, in order to solve the technical challenges of existing technologies, such as the extreme vulnerability of small-scale parameter networks to sensor noise, the loss of accuracy due to domain offset when deployed across hardware parts, and the destruction of physiological inflection point information by unified low-bit quantization.

[0007] To achieve the above objectives, this invention establishes a Physiological Noise Injection Robust Distillation (PNIRD) pipeline. This pipeline not only forces small models to learn the accuracy metrics of large models while simultaneously mastering "immune recognition" of structured noise, but also achieves adaptive mapping across hardware forms through adversarial alignment. The core design logic of this invention is: "Establishing an asymmetric defensive knowledge map transfer at the distillation bottleneck." Specifically, the system utilizes a high-performance teacher model in the cloud as a physiological truth benchmark and actively implements controlled physiological noise injection during the knowledge transfer process to the edge student models. This mechanism forces the micro-student models, under extreme parameter deprivation, not only to imitate the output results of the large model, but also to seize and consolidate the core "skeleton" feature flow with the most essential physiological resistance, thereby fundamentally severing the fragile decision edges that lead to the expansion of prediction dispersion (SD).

[0008] The proposed method for integrating physically robust edge blood pressure distillation and adaptive techniques adopts the following core technical solutions:

[0009] First, an asymmetric distillation training field is constructed. In this training field, the teacher model is input with a clean acquisition signal of high signal-to-noise ratio to output the "theoretically optimal prediction distribution"; while the student model to be trained is input with a noisy signal processed by a structured physiological noise generator (SPNG).

[0010] Secondly, the aforementioned SPNG generator does not randomly generate statistical noise, but rather synthesizes, based on medical physics priors, specific elements including: slow baseline drift caused by breathing or inflation jitter, high-frequency transient tremors generated by muscle contraction, and morphological clipping or flattening masks caused by analog sensor distortion.

[0011] Next, the improved PNIRD loss function is applied. This loss function forces the prediction vector generated by the student model to achieve strong consistency with the reference output of the teacher model in the Kullback-Leibler Divergence dimension. By introducing this adversarial pressure mechanism, the student model is pre-immunized to the noise of the real environment in the software simulation environment before formal deployment, thereby fundamentally stabilizing its regression standard deviation.

[0012] Finally, for cross-device adaptation, the system deploys a feature domain discriminative adversarial network (DANN) based on a gradient inversion layer (GRL) in the middle of the student model. This network utilizes unsupervised hardware domain prediction loss to inversely constrain the output of the main feature extraction layer to produce consistent feature representations independent of specific hardware configurations.

[0013] Meanwhile, a closed-loop hardware-software co-quantization protection system was constructed: combined with subsequent asymmetric bit quantization protection, it ensures that the core blood pressure feature bits are not lost under INT8 inference, achieving efficient, low-power, and medical-grade reliable sustainable and stable operation at the edge.

[0014] The beneficial effects obtained by this invention are:

[0015] (1) Breaking the physical curse of "precision becoming brittle" for small models: Experiments have shown that, through the PNIRD noise injection mechanism, even if the parameter size of the student model is compressed to 0.4MB (only 1 / 25 to 1 / 40 of that of a large model), its standard variance dispersion (SD) on the real noisy test set can still be strictly suppressed within the limits of the ISO criterion, which is more than 30% higher than the precision of traditional distillation prediction.

[0016] (2) Achieved rapid migration of cross-part sensors with "zero labeling cost": By integrating DANN adaptive logic, this invention successfully bridged the gap between "arm-type high-precision data" and "wrist-type high-noise applications". A large amount of prior knowledge pre-trained in the cloud can be directly mapped losslessly to wrist hardware morphology datasets with labels that have never been seen before, which greatly reduces the deployment cycle and calibration cost of terminal devices.

[0017] (3) A closed-loop hardware and software co-quantization protection was constructed: combined with asymmetric bit quantization protection, it was ensured that the core blood pressure feature bits would not be lost under INT8 inference, and the efficient, low-power and medical-grade reliable sustainable and stable operation of the edge side was achieved. Attached Figure Description

[0018] Figure 1 This is a logical diagram of the entire process of PNIRD knowledge distillation and edge deployment provided by an embodiment of the present invention;

[0019] Figure 2 This is a schematic diagram of the noise physical mask synthesis mechanism of the Structured Physiological Noise Generator (SPNG) provided in an embodiment of the present invention;

[0020] Figure 3 This is the calculation topology diagram of the asymmetric noise injection loss function based on KL divergence provided in the embodiments of the present invention;

[0021] Figure 4 This is a block diagram of the structure of the Domain Adversarial Adaptive (DANN) branch used to align the signal manifolds of the arm and wrist, provided in an embodiment of the present invention.

[0022] Figure 5 This is a hardware and software co-optimization and quantization mapping logic diagram for edge MCU deployment provided in this embodiment of the invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in further detail below with reference to the accompanying drawings. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0024] Example 1: Medical Masking Injection Mechanism of Structured Physiological Noise Generator (SPNG)

[0025] Figure 1 This is a full-process logic diagram of PNIRD knowledge distillation and edge deployment provided in this embodiment of the invention. The dashed box on the left shows the three sets of noise operators (baseline drift, transient jitter, and morphological flattening) of the Structured Physiological Noise Generator (SPNG) after fusion to generate a specific targeted interference mask, which is then forcibly superimposed on the original signal; the upper right shows the high-precision clinical source data from the cloud computing power area being processed by a deep multi-scale convolutional teacher model to generate a teacher soft-logits output set; the middle shows the multimodal robust constraint distillation penalty core, which achieves asymmetric knowledge alignment through a KL divergence function with noise masking weights; the lower side shows the microprocessor edge mapping constraint end, where a lightweight student model is trained under noisy input through inverse gradient pressure to generate a student perturbed feature output set (Student Logits).

[0026] In this embodiment, the SPNG module operates independently of the main neural network training process. For a given input pulse wave signal segment (Window Length = 1024 points), the system invokes a pre-built medical noise generation function library in real time. This function library contains three sets of specific operators.

[0027] One is the baseline drift operator. Based on a first- or second-order random walk model, it simulates the deep breathing and uneven force on the cuff caused by the urgency of measurement, forming a Gaussian walking baseline drift wave characteristic of 0.1~4Hz.

[0028] Specifically, the iterative formula for the first-order random walk model is expressed as:

[0029] .

[0030] Among them, B t Let be the baseline drift value of the sequence at step t. The variance is expressed as Gaussian white noise step size, sampled at To confine the spectrum within the range of physiological characteristics, the above-mentioned wandering sequence is then smoothed by a low-pass filter (cutoff frequency set to 4Hz) before being superimposed onto the original pulse wave.

[0031] The second is the transient dithering operator. A uniformly distributed dithering operator is used in the 8~20Hz range. , from Random sampling is performed, and the amplitude coefficient A is constrained to be the peak amplitude of the original reference signal. This ensures that the interference intensity closely matches real clinical muscle tremors. The Hanning Window envelope function is used, and its specific operation formula is as follows:

[0032] ;

[0033] In the above formula, N is the length of the injected burst noise segment. By superimposing the Hanning window envelope at both ends of the sinusoidal flutter wave, the start and end boundaries of the noise segment can be smoothly transitioned to zero, thereby effectively avoiding high-frequency step abrupt changes and pseudo-frequency leakage caused by direct splicing and hard truncation.

[0034] Thirdly, there is the morphological clipping mask. The peak amplitude is dynamically detected, and soft clipping is applied to the maximum contraction peak within certain periods to simulate the cut-off distortion of the sensor under high dynamic pressure changes. The soft clipping function can be expressed as: ,in For input signal, The set range cutoff dynamic threshold. Let be the hyperbolic tangent function, defined as This nonlinear transformation can smoothly suppress extreme peaks exceeding the threshold, avoiding non-physiological abrupt changes introduced by hard truncation, and truly reproducing the physical saturation characteristics of the sensor.

[0035] The above three sets of operators are fused to generate a structured noisy augmented signal (PNI Augmented), which is then fed into the student model for training. The specific weighted fusion rule is as follows: The outputs of the above three sets of operators are dynamically weighted and fused with the original pulse wave signal to generate the structured noisy augmented signal. The specific weighted fusion rule can be expressed as follows: ,in, This is the final synthesized noisy enhanced signal. B is the main signal after shape-flattening masking. t This is a baseline drift sequence. This is a transient jitter sequence; λ1, λ2, and λ3 are random weighting coefficients dynamically sampled from a preset uniform distribution interval, used to control the energy proportion of each disturbance; I is a Bernoulli event trigger indicator variable, used in... Transient tremors are randomly injected within a specific time window with a preset probability (e.g., p=0.2), thereby highly replicating the random and sporadic muscle interference distribution characteristics during clinical testing.

[0036] like Figure 2 As shown, after the raw pulse wave signal (Raw PPG) enters the SPNG, the three sets of operators mentioned above are called in parallel by the Physiological Physical Interference Operator Library: baseline drift operator (based on a random walk model), transient fissure operator (for micro-fissures in the 8~20Hz range), and morphological flattening mask (soft truncation / range saturation simulation). The outputs of the three sets of operators are dynamically weighted and fused by the simulated interference to generate a structured noisy augmented signal (PNI Augmented), which is then fed into the student model for training.

[0037] Example 2: Convergence Logic Control of Asymmetric PNIRD Loss Function

[0038] During the model distillation process, the system drives a teacher model T (with fixed weights and no updates) running on the cloud computing side and a student model S (dynamically updated) running on the edge side in parallel.

[0039] Distillation loss The core issue is: forcing the student model to output a probability distribution when faced with contamination. The probability distribution output by the teacher model when faced with a pure signal To achieve maximum overlap.

[0040] Specifically, by introducing temperature parameters Smoothed KL divergence loss term Supervision loss in conjunction with the original label This forms a composite cost function. This asymmetric alignment method forces the student model to abandon overfitting to the fragmented noise features of the signal. Since its parameter count is insufficient to characterize random noise, guided by KL divergence, it focuses all its learning capacity on the core low-frequency structural framework for blood pressure inference. α is the distillation loss balance coefficient, with a value range of (0,1). In this invention, the default value of α is 0.5 to ensure that the model does not lose its ability to regress the absolute value of blood pressure while aligning with soft labels to resist noise.

[0041] like Figure 3 As shown, the asymmetric distillation training field forms an asymmetric input structure between the cloud and the edge. The high-performance reference field (cloud) for the teacher model on the left inputs the high SNR oscilloscope-based raw signal (pure), and the expert-level teacher model, after parameter freezing, outputs a high-precision set of blood pressure probability reference distributions (Soft Targets). The edge deployment training field (noisy) for the student model on the right inputs a noisy physiological noise-contaminated signal processed by SPNG, and the lightweight student model outputs a calibrated prediction distribution (Student Pred). The asymmetric robust knowledge alignment field in the middle achieves forced alignment between the teacher's soft labels and the student's prediction distribution through a KL divergence function with noise masking weights, and updates the student model parameters through reverse gradient pressure.

[0042] Example 3: Dynamic Gradient Reversal and MMD Admission Mechanism of Cross-Domain Adaptive Alignment (DANN)

[0043] For cross-hardware adaptation scenarios (such as arm-to-wrist transfer), this invention introduces a domain adversarial learning mechanism. The student model is designed to include a shared feature extractor. Subsequent task classifier Domain discriminator During backpropagation, a gradient inversion layer (GRL) forces the feature extractor to learn how to fool the discriminator. To prevent inaccurate gradients generated by the discriminator in the early stages of training from damaging the feature extractor, the gradient inversion coefficients of the GRL are... Designed to dynamically increase with the training progress p (gradually changing from 0 to 1), its annealing strategy is expressed as:

[0044] .

[0045] Here, γ is a hyperparameter controlling the growth rate (e.g., set to 10). This mechanism allows the generated intermediate physiological features to gradually become completely "blurred" between the arm and wrist signals, thereby eliminating systematic biases introduced by physical differences in measurement points.

[0046] Furthermore, as a quantitative assessment of alignment quality, before feeding the solution data to the deeper classification layers of the network, the relative expected feature cluster center of the extracted subdomain distribution of the new target feature group is calculated, along with the mean distance between the cluster regions of the baseline of the clean, healthy samples pre-embedded in the original training data. The maximum mean discrepancy (MMD) tug-of-war loss constraint unit is then invoked to limit the distribution discrepancy. When the high-dimensional latent space projection divergence of the current wrist pulse center anchor point is forcibly contracted to a safe threshold ring boundary not exceeding twice the variance of the feature cluster radius statistically obtained from the offline arm girder source domain standard library, then the fully connected laminar flow for subsequent pressure measurement results is enabled. This eliminates the systematic bias introduced by physical differences in the measurement points.

[0047] like Figure 4 As shown, the Domain Discrimination Feature Distribution Offset Adversarial Field (DANN) receives a standard arm band sample set (large number - labeled) and wrist target domain migration samples (small number - unlabeled). After the shared event feature extractor extracts common features, the domain-independent feature extractor backpropagates gradients through a gradient inversion layer (GRL), forcing the shared feature output to be indistinguishable between the standard arm domain and the wrist offset domain by a specific domain morphology classifier, thereby eliminating the systematic bias introduced by physical differences in measurement points. The domain-aligned feature stream is then fed into the subsequent blood pressure estimation task network, outputting the wrist blood pressure estimation task results (SBP / DBP).

[0048] Example 4: Asymmetric Quantization Protection and Operator Lookup Table Optimization for Edge MCUs

[0049] For deployment environments with limited resources at the edge, this invention implements a physical awareness-based asymmetric quantization strategy.

[0050] First, for the high-pass sensing feature hierarchy group that captures the slope of the small abrupt change in the descending tidal wave containing information on vascular compliance poles, since the relative amplitude of its differential features is attenuated to less than 15% of the baseband peak, this invention implements a calibration ladder preservation layer, freezing and defining its network weights and activation state format to a numerical precision width of at least 16 bits, thus avoiding the smooth erasure of physiological inflection point information due to tail quantization.

[0051] Secondly, for the low-frequency global large-scale average pooling and convolutional layer group responsible for tracking and sensing the long-term envelope evolution of macroscopic gas sleeve venting, a forced low-bit integer truncation operation is performed to load the relevant fully connected kernel weight polarization compression amount into a discrete variable set of no more than 8 bits.

[0052] Finally, a mapping transformation is performed within the microprocessor (MCU) memory: For the truncated payload and on-chip compiled convolutional layer (Conv Layer) that generates the deployment target graph using continuous linear dense multiply-accumulate operations, and the subsequent batch-Normalization layer (Batch-Normalization Layer) that performs serial operations cascaded on the time logic line, a software-executed pre-analyzed mathematical identity folding operation is implemented. The specific operator folding formula is expressed as follows:

[0053] ,

[0054] .

[0055] in, and These represent the weights and biases of the original convolution operator; μ and γ and β are the moving mean and moving variance obtained during training of subsequent BN layers, respectively; γ and β are the scaling and translation parameters learned by BN, respectively; ε is a minimal constant to prevent division by zero.

[0056] The merged normalization parameters are transformed into a continuous and independent integer-mapped discrete set lattice, and all of them are pre-stored in the edge processor ROM query memory dictionary matrix (LUT). By calling the table through the address bus index offset, the table is instantly replaced and the original power-consuming multiply-accumulate matrix solution is skipped, realizing near real-time inference in a resource-constrained internal SRAM environment.

[0057] like Figure 5 As shown, the deployment phase for edge MCUs is divided into an algorithm compression and bit-width adaptive optimization layer, a static weight rearrangement layer, and an edge hardware resource mapping layer. After being processed by an asymmetric bit-width allocation strategy, the feature backbone layer, together with the higher-order redundant prediction branches, is written into instruction-level compact memory (TCM / SRAM) through static weight rearrangement. Low-bit inference is then performed by the vector processing unit (VPU / NPU Accelerator) in the ARM Cortex-M / DSP architecture, ultimately achieving ISO-level precision preservation while ensuring an extreme compression ratio and sub-millisecond real-time inference latency.

[0058] Verification of technical effectiveness:

[0059] This invention embodiment performed four ablation experiments (B1-B4) on the technical solution of Patent 2:

[0060] B1 (Baseline): Train student models directly on a standard edge hardware architecture (approximately 0.4MB).

[0061] B2 (Standard KD): Employs conventional probabilistic distillation (noiseless injection).

[0062] B3(PNI-RD): (Core of this scheme) Incorporate physiological noise injection for robust distillation.

[0063] B4 (Full Stack): This builds upon B3 by adding a DANN domain alignment module and forcibly switching to the wrist-based target domain for cross-domain migration testing. It evaluates the migration limits of the full-stack solution under an extreme capacity constraint of 0.4MB. The purpose of this setup is not to pursue optimal accuracy within the same domain, but rather to verify the feasibility and capacity sensitivity of unsupervised domain migration paths.

[0064] The performance of each experimental group on the validation set (covering a hybrid domain of PrivateB and Carditech) with high environmental interference and different pressure sensor specifications is shown in the table below:

[0065]

[0066] Comparing B2 and B3, it is evident that after injecting simulated noise using the PNI-RD module proposed in this invention, the MAE on the validation set was optimized from 9.29 mmHg to 7.38 mmHg. This quantitative data demonstrates that instilling anti-interference knowledge into the model during the distillation stage is more efficient than simply increasing parameters. This metric successfully approaches the performance upper limit of the teacher model (7.33 mmHg), proving the inventiveness of this invention in achieving extremely high measurement stability on low-computing-power devices.

[0067] By observing the performance of group B4, although the introduction of the DANN domain discrimination mechanism leads to a certain sacrifice in MAE accuracy on a very small-scale (0.4MB) edge model due to capacity bottlenecks, its core value lies in preliminarily verifying the feasibility of the unsupervised domain transfer path. For subsequent commercial deployment, this conclusion guides developers to achieve accuracy transfer support from standard hardware to customized hardware without large-scale clinical annotation, by appropriately increasing the feature extractor capacity while maintaining the core logic of the PNI-RD invention.

Claims

1. A noise-resistant knowledge distillation adaptive measurement method for low-power edge blood pressure monitoring devices, characterized in that, Includes the following steps: S1. Obtain cardiovascular pulse wave flow data and corresponding reference blood pressure labels from cloud nodes, and train a high-precision blood pressure prediction model containing multi-scale convolutional layers as a teacher model; S2. Construct a structured physiological noise generator (SPNG) to generate structured physiological noise masks based on medical physics priors. The noise masks include: baseline drift waves, transient tremor waves, and morphological flattening masks. S3. The noise mask is dynamically weighted and fused with the original pulse wave signal to generate a structured noisy enhancement signal; S4. Construct an asymmetric distillation training field: Input the high signal-to-noise ratio pure acquisition signal into the teacher model to output the theoretically best predicted distribution, and input the structured noisy enhancement signal into the student model to be trained. By introducing the KL divergence loss term after temperature parameter smoothing, combined with the supervision loss of the original blood pressure label, a composite PNIRD loss function is formed to train the student model. S5. Deploy a Domain Adversarial Neural Network (DANN) based on a gradient inversion layer (GRL) in the middle of the student model. The DANN includes a shared feature extractor, a task classifier, and a domain discriminator. By backpropagating the gradient of the feature extractor through the gradient inversion layer during backpropagation, the loss is predicted using unsupervised hardware domains. The output of the main feature extraction layer is constrained to be a consistent feature representation that does not depend on the specific hardware form, thus aligning the feature distribution of the arm-type source domain and the wrist-type target domain. S6. Implement asymmetric quantization for different levels of the student model: For the high-pass sensing feature level group that captures the slope of the small abrupt change in the descending tidal wave containing information on vascular compliance poles, implement a calibration ladder preservation layer, freezing and defining its network weights and activation state format to a numerical precision width of at least 16 bits; For the low-frequency global large-scale average pooling and convolutional level group responsible for tracking and sensing the evolution of the long-term envelope of macroscopic gas sleeve deflation, perform forced low-bit integer truncation operation, loading the relevant fully connected kernel weight polarization compression amount into a discrete variable set of no more than 8 bits; S7. For the truncated payload and the continuous linear dense multiply-accumulate convolution operator that generates the deployment target map on-chip, and the batch processing regular nonlinear feature normalization mapping function operation layer that is subsequently executed serially on the time logic line, a software-executed pre-analyzed mathematical identity folding operation is implemented to transform the merged normalization parameters into a continuous and independent integer mapping discrete set lattice, and all of them are pre-stored in the edge processor ROM query storage dictionary matrix. The table is then instantly replaced by calling the table through the address bus index offset, and the original power-consuming multiply-accumulate matrix solution stage is skipped.

2. The method of claim 1, wherein, The structured physiological noise generator SPNG mentioned in step S2 includes: The baseline drift operator, based on a first-order random walk model, simulates the deep breathing and uneven cuff stress caused by measurement urgency, forming the 0.1~4Hz Gaussian walk baseline drift wave characteristics; the iterative formula of the first-order random walk model can be expressed as: ,in, Let be the baseline drift value of the sequence at step t; The variance is Gaussian white noise step size, sampled at To confine the spectrum within the range of physiological characteristics, the above-mentioned wandering sequence is then smoothed by a low-pass filter and finally superimposed onto the original pulse wave. The transient dithering operator employs a uniform distribution within the 8–20 Hz range. exist Random sampling is performed; the envelope window function uses the Hanning window to smooth the boundaries and avoid abrupt pseudo-frequency changes; the range of the injected amplitude coefficient A is constrained to 0.05~0.2 times the amplitude of the original reference signal to ensure that the interference intensity closely matches the real clinical muscle tremor; A shape-shaving mask operator dynamically detects peak amplitude and selectively performs soft shaving on the maximum contraction peak within certain periods to simulate the range cutoff distortion of the sensor under high dynamic pressure changes; the soft shaving function can be expressed as: ,in For input signal, The set range cutoff dynamic threshold, Let be the hyperbolic tangent function, defined as This nonlinear transformation can smoothly suppress extreme peaks that exceed the threshold, avoid non-physiological abrupt changes introduced by hard truncation, and truly reproduce the physical saturation characteristics of the sensor. And a simulated interference dynamic weighted fusion module, used to perform the dynamic weighted fusion of the above three sets of operator outputs with the original pulse wave signal to generate the structured noisy enhanced signal; the weighted fusion rule is expressed as: ,in, This is the final synthesized noisy enhanced signal. B is the main signal after shape-flattening masking. t For baseline drift sequences, This is a transient jitter sequence; λ1, λ2, and λ3 are random weighting coefficients dynamically sampled from a preset uniform distribution interval, used to control the energy proportion of each disturbance; I is a Bernoulli time trigger indicator variable, sampled at... Transient tremors are randomly injected within a specific time window with a preset probability, thereby highly replicating the distribution characteristics of random muscle interference during clinical testing.

3. The method of claim 1, wherein, Step S5, which involves aligning the feature distributions of the arm-type source domain and the wrist-type target domain, further includes: Before feeding the solution data to the deeper classification end of the network, calculate the relative expected feature cluster center of the new target feature group extracted subdomain distribution, the distance between the mean of the clustering region of the pure and healthy sample baseline embedded in the original training data, and call the maximum mean difference MMD saw-off loss constraint unit to restrict the distribution dispersion. When the high-dimensional latent space projection divergence of the current wrist pulse center anchor point is forcibly contracted to a safe threshold ring boundary that is no higher than twice the variance of the feature clustering radius obtained by statistical calculation of the offline armband source domain standard library, then the fully connected laminar flow release for subsequent pressure measurement results is initiated.

4. The method of claim 1, wherein, The mathematical identity folding operation in step S7 merges the weights and biases of the convolutional layer with the scaling factor, offset, mean, and variance of the batch normalization layer into equivalent integer mapping parameters. The entire discrete set of the integer mapping is pre-stored in the edge processor ROM query storage dictionary matrix, and the table is called by the address bus index offset, skipping the original power-consuming multiplication and addition matrix solution stage.

5. A model storage medium for a low-power edge blood pressure monitoring device, characterized by, The storage medium stores quantized student model parameters obtained by the method of any one of claims 1-4, the parameters including: The network weights and activation state parameters after asymmetric quantization are as follows: the network weights and activation state format of the high-pass sensing feature layer group that captures the slope of the small abrupt change in the descending tidal wave containing information on vascular compliance poles are frozen and defined with a numerical precision width of at least 16 bits; the fully connected kernel weights of the low-frequency global large-scale average pooling and convolution layer group responsible for tracking and sensing the evolution of the long-term envelope of macroscopic gas sleeve deflation are polarized and compressed into a discrete variable set of no more than 8 bits; and the integer-mapped discrete set group lattice is written into the ROM query storage dictionary matrix after operator folding. The integer-mapped discrete set group lattice is used to call the table lookup through the address bus index offset and skip the original power-consuming multiply-accumulate matrix solution.

6. A noise-resistant knowledge distillation adaptive measurement system for low-power edge blood pressure monitoring devices, characterized in that, include: The teacher model module, configured on a cloud node, is used to train a high-precision blood pressure prediction model with multi-scale convolutional layers based on cardiovascular pulse wave flow data and corresponding reference blood pressure labels. The Structured Physiological Noise Generator (SPNG) is used to generate structured physiological noise masks based on medical physics priors. The noise masks include a 0.1-4 Hz baseline drift wave, an 8-20 Hz transient tremor wave, and a morphological flattening mask. The student model distillation training module is connected to the teacher model module and SPNG, and is used to receive the structured noisy enhancement signal processed by the SPNG, and to perform distillation training through the composite PNIRD loss function under the supervision of the soft label output by the teacher model. A domain adversarial adaptive alignment module, embedded in the middle of the student model, includes a domain adversarial neural network (DANN) based on a gradient inversion layer (GRL) and a domain discriminator, used to align the feature distributions of the arm-type source domain and the wrist-type target domain. The asymmetric quantization module is used to retain at least 16 bits of numerical precision for the high-pass perceptual feature hierarchical group of the student model and to compress the low-frequency global hierarchical group to no more than 8 bits. The model storage module includes an edge processor ROM, which is used to pre-store the integer mapping weight parameters of the student model after operator folding and LUT solidification, as well as the query dictionary matrix.

7. The system of claim 6, wherein, The domain adversarial adaptive alignment module also includes a maximum mean difference (MMD) constraint module, which is used to calculate the relative expected feature cluster center distance between the target domain and the source domain feature distribution before sending the solution data to the deeper classification end of the network, and to enable the full-connected laminar flow to release the subsequent pressure measurement results when the distance is not higher than twice the variance of the feature cluster radius obtained by statistical calculation of the offline armband source domain standard library.

8. A noise-resistant knowledge distillation adaptive measurement device for low-power edge blood pressure monitoring devices, characterized in that, include: At least one edge baseband digital processing chip (MCU) with embedded-level standby wake-up capability; The ROM physical storage module is fixedly connected to the MCU. The ROM physical storage module is pre-loaded with model weight parameters and lookup table data quantized and fixed according to the method of any one of claims 1-4. The model weight parameters include student model weights after asymmetric quantization processing. The lookup table data includes integer mapping discrete set lattice that is written into the ROM lookup storage dictionary matrix after operator folding. When the MCU is running, it calls the integer-mapped discrete set group grid and model weight parameters in the ROM according to the address bus index offset in a timing sequence, processes the arterial pressure sign signal output by the physical sensor, and outputs the blood pressure value.

9. A low power edge blood pressure monitoring device, characterized by, The device includes a physical sensor, an MCU as described in claim 8, and a ROM physical storage module, wherein the ROM physical storage module is pre-loaded with model weight parameters and LUT lookup data quantized and solidified based on the method described in any one of claims 1-4.

Images