A noise-resistant biomedical signal processing method and system based on latent topology evolution, storage medium and program product

Abstract

This invention discloses a noise-resistant biomedical signal processing method, system, storage medium, and program product based on latent topological evolution. The invention acquires one-dimensional time-series biomedical signals and static auxiliary features, which are then fused and dimensionality-reduced using a dual-stream encoder to generate a global latent feature vector located in the Lie algebra tangent space. Geometric random perturbations are injected into the tangent space through an evolutionary dynamics system to generate an effective evolutionary velocity vector. A three-dimensional topological state orthogonal matrix is ​​generated using a single-step deterministic analytical mapping that avoids numerical orthogonalization iteration. The physiological signal state prediction results are output after classification and decoding, and control signals are generated. This invention avoids numerical iterative calculations through analytical mapping, significantly reducing computational load while maintaining accuracy. It is suitable for real-time edge monitoring and high-concurrency cloud processing scenarios, and can effectively eliminate sensor displacement artifact interference.

Description

Technical Field

[0001] This invention relates to the field of biomedical signal processing technology, specifically to a noise-resistant biomedical signal processing method, system, storage medium, and program product based on potential topology evolution. Background Technology

[0002] Biomedical signal processing technology has wide applications in wearable devices, telemedicine, and motion monitoring. Common one-dimensional time-series biomedical signals include electrocardiogram (ECG), photoplethysmography (PPG), electroencephalography (EEG), and surface electromyography (sEMG). With the development of edge computing technology, more and more researchers are attempting to directly deploy signal processing and pattern recognition algorithms on the processors (MCUs) of terminal devices such as smartwatches, wristbands, and EEG caps to achieve low-latency, low-power local physiological signal state monitoring.

[0003] Currently, the existing technologies that are closest to this solution can be mainly divided into the following two categories.

[0004] Category 1: Lightweight models based on one-dimensional deep learning These methods employ structures such as one-dimensional convolutional neural networks (1D-CNN), lightweight Transformers, or recurrent neural networks (RNNs). The acquired one-dimensional time-series signals are directly input into the network, and the flattened features are extracted before the classification results are output. For example, researchers have deployed a simplified 1D-CNN on ARM Cortex-M series MCUs for real-time arrhythmia detection of electrocardiogram signals. These methods can achieve a certain level of classification accuracy in static, low-interference environments.

[0005] Category 2: Time series analyzers based on continuous dynamics This type of method treats biomedical signals as observations of a continuous dynamic system, using Neural ODEs or Neural SDEs to model the signal evolution. It simulates the continuous changes in the system state over time using numerical integration methods, and employs Singular Value Decomposition (SVD) for orthogonalization constraints at each integration step to prevent manifold drift. Theoretically, this type of method can effectively characterize the dynamic properties of the signal.

[0006] However, the aforementioned existing technologies still have the following drawbacks when facing real-world edge deployment scenarios: (i) It is sensitive to the physical displacement of the sensor and is prone to false alarms. In the actual use of wearable devices such as smartwatches and EEG caps, users inevitably engage in behaviors such as running, sweating, and head twisting, causing the sensor electrodes to physically slide or rotate on the skin surface. Existing first-class methods (such as 1D-CNN) lack the ability to model changes in the sensor's perspective in three-dimensional physical space, and cannot distinguish between "signal deformation caused by sensor displacement" and "pathological physiological waveform distortion." This makes it easy for devices to mistakenly identify pure motion artifacts as high-risk physiological events (such as arrhythmia), thus frequently triggering invalid warnings and creating a "boy who cried wolf" effect. This not only consumes the limited power and computing power of edge devices but also reduces users' trust in the warnings.

[0007] (ii) The computational complexity is too high, making it difficult to run in real time on a conventional MCU. Existing second-class methods (such as Neural ODE / SDE) require multi-step cumulative evolution with extremely small time steps to achieve numerical stability, and each step necessitates calling Singular Value Decomposition (SVD) with a time complexity as high as O(n³) for orthogonalization correction. For conventional processors (such as Cortex-M series MCUs) powered only by button batteries and lacking a floating-point unit (FPU), this computational overhead leads to memory overflow and rapid battery depletion, making continuous, real-time edge monitoring impossible.

[0008] (iii) Risk of missing rare anomalies in extremely imbalanced data In real-world monitoring scenarios such as ventricular premature beats (VEB) and epileptic spikes, the probability of rare abnormal physiological features occurring is extremely low (often less than 0.1% of all-day data). Existing models mostly use the standard cross-entropy loss function for training, and their gradient updates are easily overwhelmed by massive amounts of normal background samples, causing the model to neglect learning minority abnormal features. In safety-sensitive applications, this deficiency can lead to false negatives (i.e., missed feature detection), potentially affecting the timely response to changes in the user's physiological signal state.

[0009] In summary, how to simultaneously achieve interference resistance to sensor physical displacement, high-sensitivity identification of rare anomalies, and effective suppression of continuous dynamic noise under the limited computing power constraints of edge devices has become a pressing technical problem in the field of biomedical signal processing. Summary of the Invention

[0010] To address the problems existing in the prior art, the present invention aims to provide a noise-resistant biomedical signal processing method, system, storage medium, and program product based on potential topology evolution, which can achieve low power consumption, resistance to motion artifacts, and high sensitivity to rare anomalous features in biomedical signal processing.

[0011] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A noise-resistant biomedical signal processing method based on latent topology evolution includes the following steps: Step 1: Obtain the one-dimensional time-series biomedical signal sequence of the object to be tested and its corresponding static auxiliary features; configure the target manifold according to the physical prior characteristics of the one-dimensional time-series biomedical signal sequence; use an encoder network to extract morphological features from the one-dimensional time-series biomedical signal sequence to obtain a one-dimensional morphological feature vector; encode the static auxiliary features to obtain a static feature vector; fuse the one-dimensional morphological feature vector and the static feature vector, and then perform spatial dimensionality reduction processing to generate a global latent feature vector; the global latent feature vector is located in the Lie algebra tangent space of the target manifold; Step 2: In the Lie algebra tangent space, inject geometric random perturbation into the global latent feature vector through the evolutionary dynamics system to generate an effective evolutionary velocity vector; wherein, the geometric random perturbation includes at least a diffusion term for absorbing continuous physiological background drift and a jump term for capturing transient waveform topological changes; Step 3: Using a single-step deterministic analytical mapping that avoids numerical orthogonalization iteration, the effective evolution velocity vector is directly mapped to the target orthogonal constrained manifold space, and a three-dimensional topological state orthogonal matrix is ​​generated analytically. Step 4: Flatten the three-dimensional topological state orthogonal matrix and input it into the classification decoder to output the prediction result of the target physiological signal state, and generate a control signal based on the prediction result.

[0012] The evolutionary dynamics system comprises a parallel branch network constructed from multiple miniature perceptrons, specifically including: Drift networks are used to extract deterministic physiological dynamics baselines; A diffusion network is used to generate an envelope term for absorbing high-frequency background noise, said envelope term being multiplicatively coupled to an externally introduced continuous Brownian noise variable; and A jump network is used to jointly predict the mutation intensity and the probability of jump occurrence. The jump network is combined with an externally introduced discrete Poisson jump variable to respond to transient waveform topological mutations. The outputs of the drift network, diffusion network, and hopping network are aggregated with a single preset time step in the Lie algebra tangent space and combined to generate the effective evolution velocity vector.

[0013] The classification decoder uses a dynamically decoupled asymmetric classification loss function for global parameter updates during the model training phase. The asymmetric classification loss function is configured with a dynamic adjustment coefficient for cross-entropy and a false negative penalty amplification parameter for minority class abnormal feature samples. Specifically, the false negative penalty amplification parameter is non-linearly positively correlated with the reciprocal of the occurrence probability of the target abnormal physiological feature within the historical all-weather monitoring time window, and is truncated by a preset dynamic threshold to forcibly extract weak manifold jump features corresponding to low-frequency minority class waveform distortions.

[0014] The encoder network is a dual-stream encoder, and its feature extraction and fusion specifically includes: The one-dimensional time-series biomedical signal sequence is input into the time-series convolution branch, and at least one one-dimensional convolution and batch normalization are performed sequentially. The output feature map is flattened and then layer normalized to obtain a one-dimensional morphological feature vector. The static auxiliary features are input into the static fully connected branch and sequentially processed through at least one fully connected layer, an activation function, and layer normalization to output a static feature vector. The one-dimensional morphological feature vector and the static feature vector are concatenated by channels, and then compressed to a dimension that matches the target orthogonal constrained manifold through a fully connected dimensionality reduction layer to generate the global latent feature vector, which is located in the Lie algebra tangent space of the target manifold.

[0015] Step 3 employs a single-step deterministic analytical mapping with non-numerical orthogonalization iteration, specifically including: Extract the magnitude and unit direction vector of the effective evolution velocity vector; An antisymmetric matrix is ​​constructed based on the unit direction vector, and the three-dimensional topological state orthogonal matrix is ​​analytically generated in a single-step calculation using the Rodriguez rotation formula.

[0016] Based on the specific physiological type of the one-dimensional time-series biomedical signal sequence, the evolutionary dynamics system executes the corresponding adaptive perturbation configuration: If the biomedical signal is a photoplethysmography (PPG) signal, the generation scale of the envelope term in the diffusion network is adaptively increased to absorb continuous baseline drift caused by the optical environment. If the biomedical signal is an electroencephalogram (EEG) signal or a surface electromyogram (EMG) signal, the basic bias of the Poisson probability in the jumping network is adaptively increased to capture transient spikes or high-frequency EMG burst jumping features. If the biomedical signal is a pet multi-lead electrocardiogram signal, then the static auxiliary feature is configured to fuse the target pet's species category, weight, and body shape data, and adaptively increase the generation scale of the envelope term in the diffusion network.

[0017] A noise-resistant biomedical signal processing system based on latent topology evolution, comprising: The physiological signal acquisition module is used to acquire one-dimensional time-series biomedical signal sequences of the subject in real time. Memory, used to store computer instructions and pre-trained weights of topology-evolution-based network models; The processor is communicatively connected to the physiological signal acquisition module and the memory, and the processor implements the noise-resistant biomedical signal processing method as described in any one of claims 1 to 6 when executing the computer instructions; And an alarm and communication module, used to perform local intervention or communication alarm based on control signals generated by the processor.

[0018] A computer-readable storage medium having a computer program or instructions stored thereon, which, when executed by a processor, implement the steps of a noise-resistant biomedical signal processing method based on latent topology evolution as described above.

[0019] A computer program product includes a computer program or instructions that, when executed by a processor, implement the steps of a noise-resistant biomedical signal processing method based on latent topology evolution as described above.

[0020] By adopting the above scheme, this invention compresses one-dimensional time-series signals into a low-dimensional Lie algebra tangent space, introduces drift, diffusion, and hopping combinations for modeling in the tangent space, and then uses Rodrigues analytical mapping to generate a three-dimensional topological state matrix that satisfies orthogonality constraints. Compared with methods that rely on multi-step numerical integration or repeated orthogonalization operations, this invention can reduce the computational overhead caused by high-complexity matrix decomposition and iterative calculations, and improve the deployment feasibility in low-power edge devices.

[0021] This invention introduces diffusion and jumping terms into the tangent space, enabling unified modeling of continuous background drift, high-frequency noise, and transient waveform changes. This improves the model's robustness to sensor displacement, changes in wearing status, and environmental disturbances, and reduces false alarms caused by changes in non-target signals.

[0022] For target waveform features with a low sample ratio, this invention introduces asymmetric classification loss during the training phase, assigning higher weights to minority class target features and false negative cases, thereby increasing the model's attention to low-frequency target states and reducing the risk of target features being overwhelmed by a large number of background samples.

[0023] Therefore, this invention is applicable to one-dimensional time-series signal processing, status prompts, data recording, and communication alarm scenarios in smart wearable devices, pet monitoring devices, motion monitoring devices, or other edge physiological signal monitoring devices. Attached Figure Description

[0024] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a block diagram illustrating the single-step jump diffusion evolution principle of the present invention; Figure 3 This is a schematic diagram of the processing system of the present invention. Detailed Implementation

[0025] like Figure 1 As shown, this invention discloses a noise-resistant biomedical signal processing method based on latent topology evolution, which includes the following steps: Step 1: Acquisition of multi-source data and global feature encoding.

[0026] Step 1.1: Data Acquisition and Preprocessing.

[0027] The edge device acquires a one-dimensional time-series biomedical signal sequence of the subject in real time through a physiological signal acquisition module. This one-dimensional time-series biomedical signal can be an electrocardiogram (ECG), electroencephalogram (EEG), photoplethysmography (PPG), or surface electromyography (SEMG), etc. The acquired raw signal undergoes necessary preprocessing, including but not limited to baseline drift removal, power frequency notch filtering, and bandpass filtering, and then signal segments are truncated to a preset time window length. Simultaneously, the edge device also acquires static auxiliary features corresponding to the subject. These static auxiliary features include, but are not limited to, time-domain interval statistics (such as RR intervals), user age, weight, body fat percentage, skin impedance index, species classification, and other static or quasi-static parameters.

[0028] Step 1.2, Morphological feature encoding.

[0029] The extracted one-dimensional temporal signal segment is input into the temporal convolutional branch of the dual-stream encoder. This temporal convolutional branch includes at least one one-dimensional convolutional layer, each of which is optionally followed by a batch normalization layer and an activation function layer. The kernel size, stride, number of output channels, and number of layers of the convolutional layers can be configured according to the characteristics of the signal and the computing power of the edge device. After several convolutions and downsampling, the output feature map is flattened into a one-dimensional vector and then subjected to layer normalization to obtain a one-dimensional morphological feature vector.

[0030] Step 1.3: Static feature encoding and fusion.

[0031] The static auxiliary features are input into the static fully connected branch of the two-stream encoder. This static fully connected branch includes at least one fully connected layer, followed by an activation function and layer normalization, outputting a static feature vector. Subsequently, the one-dimensional morphological feature vector and the static feature vector are concatenated channel by channel to obtain a fused feature vector.

[0032] Step 1.4: Global Dimensionality Reduction and Manifold Basis Generation The fused feature vector is input into a fully connected dimensionality reduction layer, the output dimension of which is equal to the dimension of the tangent space of the target manifold (e.g., when the target manifold is a three-dimensional special orthogonal group SO(3), the output dimension is 3). The output vector is denoted as the global latent feature vector. This vector serves as the initial state basis for subsequent evolution processes and resides in the tangent space (Lie algebra space) corresponding to the target manifold.

[0033] The output vector is denoted as the global latent feature vector. The global latent feature vector Located in the Lie algebraic tangent space corresponding to the target manifold (e.g., when the target manifold is SO(3), d=3; in a preferred embodiment, when the target manifold is configured as a three-dimensional special orthogonal group SO(3), the dimension of the global latent eigenvector is configured as 3, corresponding to the Lie algebraic tangent space of the rotational perturbation; in another optional extended embodiment, when the application scenario requires simultaneous characterization of rotational and translational perturbations, the target manifold can be configured as a special Euclidean group SE(3), and the dimension of the global latent eigenvector is correspondingly configured as 6, including 3-dimensional rotational components and 3-dimensional translational components.), serving as the initial state basis for the subsequent evolutionary dynamics system. This Lie algebraic representation is the core premise for the present invention to achieve single-step, iterative-free manifold mapping—it compresses the complex three-dimensional rotational evolution problem into a low-dimensional linear space for computation.

[0034] Step 2, Single-step jump in tangent space - diffusion physical perturbation.

[0035] Step 2.1, Configuration of the evolutionary dynamics network.

[0036] The global latent feature vector generated in step 1.4 The input is fed into an evolutionary dynamics system consisting of multiple miniature multilayer perceptrons (MLPs). For example... Figure 2 As shown, this power system comprises the following three sub-networks in parallel: Drift Network: Used to extract deterministic physiological dynamics baselines. This drift network accepts a global latent feature vector and outputs a drift vector with the same dimension as the tangent space.

[0037] Diffusion network: Used to generate continuous envelope terms to absorb high-frequency background noise and continuous baseline drift. The output of this diffusion network is a diffusion envelope vector with the same dimension as the tangent space.

[0038] Jump Network: Used to capture transient waveform topological changes. This jump network further includes two sub-branches: a jump amplitude sub-network, which outputs the change intensity vector; and a jump intensity sub-network, which outputs the probability of a jump occurring. (The value range is [0,1]).

[0039] Step 2.2, Single-step random perturbation calculation.

[0040] Set the preset time step parameters for single-step evolution of the evolutionary dynamics system. In the tangent space, an external random variable is introduced: the increment of continuous Brownian noise. and based on Discrete Poisson jump variables. Multiply the drift network output. The effective evolution velocity vector is obtained by element-wise multiplying the diffusion network output with the Brownian noise increment and then multiplying by the global diffusion scale constant, and by element-wise multiplying the jump amplitude network output with the discrete Poisson jump variable. ,Right now ,in, For the deterministic physiological dynamic limit of the drift network output, The envelope term output by the diffusion network is used to absorb high-frequency noise background. The mutation strength output by the jump magnitude network is the effective evolution velocity vector. Located in tangent space, it represents the direction and magnitude of the system's single-step evolution after physical geometric perturbation.

[0041] The basic discrete Poisson jump noise variable introduced from the outside; For effective Poisson jump increments, it acts as a gating variable for topological jumps. This is achieved by using the basic discrete Poisson jump noise variable... With the probability of the jump occurring Perform probability condition analysis to generate the aforementioned (For example, when the generated random variable) Falling into probability When determining the activation interval, (It is set to the active state; otherwise, it is inactive). Through this mechanism, the present invention achieves the technical effect of forcing external disordered random noise to be controlled by the target physiological characteristic dynamics prior, thereby accurately inducing manifold jumps corresponding to rare pathological aberrations in the tangent space.

[0042] Step 3: Perform a single-step deterministic constraint manifold mapping and generate a three-dimensional topological state orthogonal matrix.

[0043] Perform Rodriguez exponential mapping: map the effective evolutionary velocity vector generated in step 2.2. Map to the target special orthogonal group manifold and set the zero-truncation threshold parameter.

[0044] Obtain the effective evolution velocity vector Magnitude and unit direction vector Construct the antisymmetric matrix corresponding to the unit direction vector. The rotation matrix H is calculated using the single-step Rodriguez rotation formula.

[0045] Step 3.2: Generation of the three-dimensional topological state orthogonal matrix.

[0046] The rotation matrix H calculated by the above formula is a square matrix (e.g., 3×3), and its product with its transpose is the identity matrix. , which is a three-dimensional topological state orthogonal matrix. The physical meaning of this three-dimensional topological state orthogonal matrix corresponds to the rotational attitude of the sensor in three-dimensional space, thus possessing mathematical immunity to viewpoint rotation artifacts caused by physical displacement of the electrodes. The entire calculation process involves only basic algebraic operations and trigonometric functions, without requiring high-time-complexity numerical orthogonalization iterations such as singular value decomposition (SVD).

[0047] Step 4: Classification of physiological signal states and updating of higher-order parameters under asymmetric conditions.

[0048] Step 4.1, Classification Decoding.

[0049] The three-dimensional topological state orthogonal matrix generated in step 3 is flattened into a one-dimensional vector and input into the classification decoder. The classification decoder includes at least one fully connected layer followed by a Softmax activation function, and outputs the predicted probability of the target physiological signal state (e.g., the probability of a normal state, the probability of a suspected abnormal state, etc.).

[0050] Step 4.2: Hardware control instruction generation.

[0051] The processor compares the classification results with preset thresholds. When it determines that there is a change in the target's physiological signal state or that the warning conditions are met, it generates a control signal. The control signal is used to trigger edge devices to perform local intervention or communication alarms, including but not limited to: starting the vibration motor, illuminating the LED indicator, displaying prompt information on the screen, sending data packets to peripheral devices via Bluetooth Low Energy (BLE), activating the buzzer, etc.

[0052] Step 4.3, Asymmetric constraints during model training (offline).

[0053] During the offline training phase of the model, a dynamically decoupled asymmetric classification loss function is used for global parameter updates. This classification loss function is configured with a dynamically adjusted cross-entropy coefficient and a false negative penalty amplification parameter for high-risk minority class anomalous feature samples. Simultaneously, dynamic weights are assigned to each classification category based on logarithmic smoothing, and these weights are limited to preset upper and lower bound cutoff intervals to prevent gradient explosion and feature collapse caused by extremely imbalanced physiological data.

[0054] This invention compresses high-dimensional time-series signals into a low-dimensional Lie algebraic tangent space, and then, after employing single-step jump-diffusion perturbation in the tangent space, directly generates a three-dimensional topological state orthogonal matrix using the analytical solution of the Rodrigues rotation formula. This avoids the high-time-complexity computations such as singular value decomposition (SVD) required in existing technologies (such as NeuralODE / SDE). The core system of this invention has extremely few parameters and very low computational cost. Under extremely demanding conditions of significantly reduced computational power consumption and a substantial reduction in the number of parameters, this invention achieves a counter-trend improvement in overall signal classification accuracy. This allows the method to be directly programmed into a conventional processor (MCU) powered only by a button cell or milliampere-level battery and without a floating-point acceleration unit, eliminating the need to package and send data to the cloud for processing. This, to a certain extent, avoids the power consumption and network transmission latency of high-power wireless RF communication modules.

[0055] This invention introduces geometric random perturbations, including diffusion and jump terms, into the Lie algebra tangent space and combines them with a single-step deterministic manifold mapping to construct a dynamic geometric compensation system for edge hardware acquisition perspectives. This system can decouple non-stationary noise caused by vigorous user movement, electrode physical slippage, or skin friction in real-world wearable scenarios in real time. In long-range, non-intrusive monitoring scenarios with significant sensor displacement, this invention can distinguish between "sensor physical displacement artifacts" and "pathophysiological topological distortions" from a mathematical and geometric perspective, significantly reducing the frequency of invalid warnings from edge monitoring devices in complex dynamic environments and eliminating the "boy who cried wolf" effect.

[0056] For abnormal physiological waveforms that occur with extremely low probability but are easily ignored by models, leading to missed detection, this invention employs a dynamically decoupled asymmetric classification loss function to independently amplify and penalize false negative samples, effectively resisting the gradient overwhelming effect of massive amounts of normal background electrophysiological data. While maintaining an extremely low parameter count, this invention not only significantly improves the overall diagnostic accuracy but also enhances the sensitivity to subtle abnormal physiological features. This means that in extreme local environments without external cloud computing support, this invention successfully establishes a safety net—when extremely weak and rare physiological topological mutations occur, the edge system can quickly and sensitively complete edge-side identification and directly trigger local hardware-level intervention.

[0057] To provide a detailed explanation of the technical solution of this invention, specific application examples will be given below, combining four different types of biomedical signals. In these examples, the parameters of each step will be adaptively adjusted according to the signal characteristics, but the overall method remains unchanged.

[0058] Implementation 1 This embodiment uses electroencephalography (EEG) signals as an example to illustrate the application of the present invention in a portable EEG cap or head-mounted device. The direct purpose of this embodiment is to extract abnormal rhythmic features of the EEG (such as spikes suspected to be epileptic discharges, abnormal slow waves after stroke, etc.) in real time during the user's daily activities (such as gait rehabilitation training) as intermediate reference data for subsequent manual review or system alarms, without directly outputting disease diagnosis conclusions.

[0059] This embodiment specifically includes the following steps: Step 1: Data Acquisition and Preprocessing.

[0060] Given the high sampling rate of EEG signals (typically above 256Hz), this embodiment uses a sampling rate of 256Hz to acquire EEG signals from a single channel or a small number of leads. The time window length L is set to 512 sampling points (corresponding to 2 seconds) to include multiple EEG rhythm cycles. Static auxiliary features may include the user's age, gender, etc.

[0061] During morphological feature encoding, the dilation rate of the convolutional layers is adaptively adjusted for specific frequency bands in EEG (e.g., Delta waves 0.5-4Hz, Theta waves 4-8Hz, Alpha waves 8-13Hz, Beta waves 13-30Hz, Gamma waves >30Hz): a larger dilation rate (e.g., dilation=4) is used to capture low-frequency slow waves, and a smaller dilation rate (e.g., dilation=1) is used to capture high-frequency spikes. The convolutional layer configuration is a first 1D-CNN layer (kernel size 7, stride 2, output channels 16) and a second 1D-CNN layer (kernel size 5, stride 2, output channels 32), followed by flattening and layer normalization to obtain a one-dimensional morphological feature vector.

[0062] Static auxiliary features are concatenated with morphological feature vectors after passing through a fully connected branch, and then output as a global latent feature vector with dimension d=3 through a fully connected dimensionality reduction layer. , located in the SO(3) Lie algebra tangent space.

[0063] Step 2, Single-step jump in tangent space - diffusion physical perturbation.

[0064] To address the continuous characteristics of electrooculography (EOG) and electromyography (EMG) artifacts in EEG, the output scale of the diffusion network is appropriately increased. The single-step evolution time step is adaptively reduced. In this embodiment, a value of 0.02 can be used to adapt to the high sampling rate. The continuous envelope generated by the diffusion network is specifically designed to dynamically absorb continuous background baseline drift caused by sweating or minor friction; the skip network, by increasing the sensitivity of the Poisson probability λ, is specifically designed to accurately capture extremely brief suspected anomalous spike features in the tangent space. The effective evolution velocity vector is calculated. .

[0065] Step 3: Perform a single-step deterministic constraint manifold mapping and generate a three-dimensional topological state orthogonal matrix.

[0066] Effective evolution velocity vector (Located in the SO(3) tangent space) mapped to the SO(3) manifold. The effective evolutionary velocity vector is obtained. Magnitude and unit direction vector Construct the antisymmetric matrix corresponding to the unit direction vector. The rotation matrix H is calculated using the single-step Rodriguez rotation formula, and H is a three-dimensional topological state orthogonal matrix. This three-dimensional topological state orthogonal matrix effectively avoids the physical displacement of the EEG cap electrodes on the scalp (equivalent to SO(3) rotation), eliminating the need for additional motion sensor correction.

[0067] Step 4: Classification of physiological signal states and output of hardware control commands The three-dimensional topological state orthogonal matrix H is flattened into a 9-dimensional vector and input into a fully connected classification head (9→2). After passing through Softmax, it outputs "probability of suspected abnormal spikes" and "probability of normal background". When the "probability of suspected abnormal spikes" exceeds a preset threshold (e.g., 0.7), the edge device (EBD cap MCU) generates a control signal: triggering local storage (timestamping), sending intermediate data to a mobile app via BLE, or activating an LED indicator. This embodiment does not output a "seizure" diagnosis; it only provides feature extraction and early warning prompts.

[0068] This embodiment applies a general method to EEG signal processing, effectively suppressing artifacts from electrooculography (EOG), electromyography (EMG), and electrode displacement interference. It eliminates the need for additional motion sensors to prevent physical displacement of the EEG cap, while significantly improving the detection sensitivity for rare abnormal features such as epileptic spikes, and achieving ultra-low power edge deployment.

[0069] Example 2 This embodiment uses multi-lead electrocardiogram (ECG) signals from pets (such as cats and dogs) as an example to illustrate the application of this invention in pet collars or medical monitoring suits. The direct purpose of this embodiment is to monitor heart rate changes and rhythm characteristics in real time during pets' daily activities, providing pet owners with non-diagnostic health tips (such as "Pet's heart rate is high; it is recommended to reduce strenuous activity"), without replacing veterinary diagnosis. This embodiment specifically includes the following steps: Step 1: Acquisition of multi-source data and global feature encoding.

[0070] Considering that pets have significantly higher heart rates than humans (normal heart rates in cats and dogs often reach 120-200 bpm) and shorter ECG waveform intervals, this embodiment shortens the time window length L (e.g., L = 90 sampling points, corresponding to 0.36 seconds). Static auxiliary features are forcibly fused with pet species (dog / cat), weight, and size (small / medium / large) data to construct a feature space that conforms to the physiological baseline of the specific species. The kernel size of the first layer of the temporal convolution branch is appropriately reduced (e.g., kernel size = 3) to capture a narrower QRS complex. The remaining encoding process is the same as the aforementioned method, ultimately outputting a global latent feature vector.

[0071] Step 2, Single-step jump in tangent space - diffusion physical perturbation.

[0072] The pet's violent writhing caused extreme physical displacement noise in the electrocardiogram signal; the drift network was responsible for anchoring the normal sinus rhythm baseline. The envelope scale of the diffusion network was significantly enlarged, and Brownian noise was used to forcibly filter out artifacts caused by hair friction and unstable skin impedance. The skip network maintained high sensitivity to skip probability for rare mutations such as ventricular ectopic beats. The effective evolutionary velocity vector was calculated.

[0073] Step 3: Perform a single-step deterministic constraint manifold mapping and generate a three-dimensional topological state orthogonal matrix.

[0074] An orthogonal rotation matrix is ​​generated using the Rodriguez rotation formula. This single-step deterministic mapping eliminates the need for SVD calculations, allowing the pet collar device to maintain high-frequency ECG monitoring while extending battery life. The orthogonal topological matrix directly cancels out visual interference caused by the pet's body rotation.

[0075] Step 4: Classification of physiological signal states and output of hardware control instructions.

[0076] The classification decoder outputs "heart rate rhythm status" (normal / high / low / irregular). When it is determined to be "irregular" and the probability exceeds the threshold, the collar device emits a beep or vibrates to remind the pet owner to pay attention to the pet's condition. This embodiment does not output diagnostic conclusions such as "heart disease," but only provides exercise and health management prompts.

[0077] This embodiment applies a general method to pet ECG monitoring, adapting to high heart rate and intense exercise noise, effectively filtering burr artifacts caused by hair friction, and significantly reducing the false alarm rate; at the same time, the SVD-free operation significantly extends the battery life of the pet collar device, and the orthogonal topology matrix directly cancels out the visual interference caused by the pet's body rotation.

[0078] Example 3 This embodiment uses photoplethysmography (PPG) signals as an example to illustrate the application of the present invention in all-weather wearable devices such as smartwatches and fitness trackers. The direct purpose of this embodiment is to monitor pulse wave morphology and heart rate variability in real time during users' daily activities (walking, running, sleeping), providing exercise intensity suggestions and daily health management references, and is not intended for diagnosing heart disease.

[0079] Step 1: Acquisition of multi-source data and global feature encoding.

[0080] To address the extremely low frequency of PPG signals (fundamental frequency typically 1-3Hz) and their sensitivity to minute random wrist movements and ambient light scattering, this embodiment employs a one-dimensional convolution with an ultra-high dilation rate in the temporal convolution branch. This achieves a large receptive field covering the entire pulse wave cycle (approximately 0.3-1 seconds) with minimal parameter usage. The static auxiliary feature branch forcibly integrates parameters such as the user's age and skin tone impedance index. The time window length L is appropriately increased (e.g., L = 300 sampling points, corresponding to approximately 2 seconds) to include at least two complete pulse wave cycles. The remaining encoding process is the same as described above, outputting a global latent feature vector.

[0081] Step 2, Single-step jump in tangent space - diffusion physical perturbation.

[0082] Unlike electrophysiological signals, PPG's environmental optical noise exhibits strong continuity. This embodiment adaptively increases the global diffusion noise scale constant. The envelope generated by the diffusion network, combined with enhanced continuous Brownian noise, is specifically designed to dynamically "absorb" continuous baseline drift caused by light refraction, sweating, or minute displacements in the latent space. A skip network is used to capture suspected rhythmic anomalies (such as sudden changes in pulse wave morphology), with the skip probability threshold appropriately increased to avoid overreacting to normal fluctuations. The effective evolutionary velocity vector is calculated.

[0083] Step 3: Perform a single-step deterministic constraint manifold mapping and generate a three-dimensional topological state orthogonal matrix.

[0084] An orthogonal rotation matrix is ​​generated using the Rodriguez rotation formula. The physical sliding of the smartwatch on the wrist is equivalent to the SO(3) rotation of the incident / reflected angle of the optical sensor. This orthogonal topological state matrix effectively avoids this angle rotation offset, avoids SVD operations, and allows the model to run smoothly in the RTOS smartwatch MCU with a milliampere-level battery capacity.

[0085] Step 4: Classification of physiological signal states and output of hardware control instructions.

[0086] The classification decoder outputs "pulse wave status" (normal / suspected irregular). When the probability of suspected irregularity exceeds a threshold (e.g., 0.7), the watch device triggers a vibration alert and displays non-diagnostic text such as "Abnormal heart rate variability, please remain still" on the screen. Simultaneously, the data can be recorded locally for user daily review or sharing with doctors (not for diagnostic purposes).

[0087] This embodiment applies a general method to photoplethysmography (PPG) monitoring. Through ultra-large receptive field encoding and continuous baseline drift dynamic absorption, it effectively resists interference from wrist movement, sweat refraction, and ambient light scattering. The orthogonal topology matrix effectively avoids the rotation of the viewing angle when the watch is sliding, realizing ultra-low power continuous monitoring on the smartwatch MCU.

[0088] Example 4 This embodiment uses surface electromyography (sEMG) signals as an example to illustrate the application of the present invention in extreme sports monitoring or rehabilitation training. The direct purpose of this embodiment is to assess muscle fatigue, explosive power level, and abnormal muscle tremors in real time during high-intensity training for athletes or rehabilitation for patients, provide suggestions for adjusting exercise intensity, and prevent sports injuries. It is not intended for the diagnosis of neuromuscular diseases.

[0089] This embodiment specifically includes the following steps: Step 1: Acquisition of multi-source data and global feature encoding.

[0090] To address the characteristics of sEMG signals—extremely high sampling rates (often exceeding 1000Hz) and the dramatic topological jumps exhibited by sudden, explosive muscle contractions—this embodiment adaptively increases the convolution stride and pooling window in the temporal convolution branch to compress high-frequency electromyographic pulse sequences while maintaining computational efficiency. The static auxiliary feature branch forcibly incorporates the athlete's biomechanical characteristics, including BMI, body fat percentage, core muscle mass, or parameters related to the patient's rehabilitation stage. The time window length L is set to a relatively short window (e.g., L=256, corresponding to 0.256 seconds) to capture rapid transient changes in electromyography. The remaining encoding process is the same as the general method, outputting a global latent feature vector.

[0091] Step 2, Single-step jump in tangent space - diffusion physical perturbation.

[0092] A drift network extracts a stable baseline during sustained muscle contraction; a diffusion network absorbs continuous frictional noise caused by minute movements of the electrode patches; and a jump network is emphasized: adaptively increasing the base bias of the Poisson probability in the jump intensity subnetwork and amplifying the perception threshold of abrupt changes. Discrete Poisson jump variables are combined with this network to precisely anchor and amplify high-frequency jump features representing muscle fatigue limits, weak electromyographic signals during nerve repair, or minute tears within the tangent space perturbation. An effective evolutionary velocity vector is then calculated.

[0093] Step 3: Perform a single-step deterministic constraint manifold mapping and generate a three-dimensional topological state orthogonal matrix.

[0094] An orthogonal rotation matrix is ​​generated using the Rodriguez rotation formula. During intense running or physical contact, athletes experience extreme stretching and torsional deformation between the electromyography (EMG) patch and the skin. This deterministic SO(3) geometric constraint directly counteracts this complex physical three-dimensional deformation noise. The SVD-free design results in millisecond-level extremely low inference latency, meeting the business requirements of providing "zero-latency" real-time feedback on muscle status in sports competitions.

[0095] Step 4: Classification of physiological signal states and output of hardware control instructions.

[0096] The classification decoder outputs parameters such as "probability of muscle fatigue" and "probability of abnormal tremors." When the probability of muscle fatigue exceeds a threshold, the system provides a voice prompt to the athlete via Bluetooth headphones (e.g., "Muscle fatigue is too high, it is recommended to reduce the intensity"), or provides a tactile warning via a vibrating waist belt. In rehabilitation scenarios, this output can be used to trigger parameter adjustments for functional electrical stimulation devices, forming a closed-loop feedback loop.

[0097] This embodiment applies a general method to surface electromyography (EMG) signal monitoring. By using a skip network to enhance the capture of muscle fatigue precursors and abnormal tremor characteristics, the detection latency is extremely short. Deterministic geometric constraints directly neutralize the physical deformation noise of the EMG patch, and the millisecond-level inference latency meets the real-time feedback requirements, making it suitable for closed-loop control of extreme sports protection and rehabilitation training.

[0098] Because the method of this invention ensures high accuracy while having low computational load and low power consumption, it can be deployed in edge devices or cloud servers. Especially in scenarios involving real-time processing of massive concurrent data streams, it can improve cloud throughput and reduce concurrent response latency, thus providing solid algorithmic support for large-scale medical IoT.

[0099] like Figure 3 As shown, based on the same inventive concept, this invention also discloses a noise-resistant biomedical signal processing system based on latent topological evolution, which includes: The physiological signal acquisition module is used to acquire one-dimensional time-series biomedical signal sequences of the subject in real time. Memory, used to store computer instructions and pre-trained weights of micromanifold intelligent network models based on topology evolution; The processor is communicatively connected to the physiological signal acquisition module and the memory. When executing the computer instructions, the processor implements the noise-resistant biomedical signal processing method described above. And an alarm and communication module, connected to the processor, for performing local intervention or communication alarms based on control signals generated by the processor.

[0100] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of hardware embodiments, software embodiments, or embodiments combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0101] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0102] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0103] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

Claims

1. A noise-resistant biomedical signal processing method based on latent topological evolution, characterized in that, Includes the following steps: Step 1: Obtain the one-dimensional time-series biomedical signal sequence of the object to be tested and the corresponding static auxiliary features; configure the target manifold according to the physical prior characteristics of the one-dimensional time-series biomedical signal sequence; use an encoder network to extract morphological features from the one-dimensional time-series biomedical signal sequence to obtain a one-dimensional morphological feature vector; encode the static auxiliary features to obtain a static feature vector; fuse the one-dimensional morphological feature vector and the static feature vector, and then perform spatial dimensionality reduction processing to generate a global latent feature vector; The global latent feature vector is located in the Lie algebraic tangent space of the target manifold; Step 2: In the Lie algebra tangent space, inject geometric random perturbation into the global latent feature vector through the evolutionary dynamics system to generate an effective evolutionary velocity vector; wherein, the geometric random perturbation includes at least a diffusion term for absorbing continuous physiological background drift and a jump term for capturing transient waveform topological changes; Step 3: Using a single-step deterministic analytical mapping that avoids numerical orthogonalization iteration, the effective evolution velocity vector is directly mapped to the target orthogonal constrained manifold space, and a three-dimensional topological state orthogonal matrix is ​​generated analytically. Step 4: Flatten the three-dimensional topological state orthogonal matrix and input it into the classification decoder to output the prediction result of the target physiological signal state, and generate a control signal based on the prediction result.

2. The noise-resistant biomedical signal processing method based on latent topology evolution according to claim 1, characterized in that, The evolutionary dynamics system comprises a parallel branch network constructed from multiple miniature perceptrons, specifically including: Drift networks are used to extract deterministic physiological dynamics baselines; A diffusion network is used to generate an envelope term for absorbing high-frequency background noise, said envelope term being multiplicatively coupled to an externally introduced continuous Brownian noise variable; and A jump network is used to jointly predict the intensity of a mutation and the probability of a jump. The jump network is combined with an externally introduced discrete Poisson jump variable to respond to transient waveform topological mutations. The outputs of the drift network, diffusion network, and hopping network are aggregated with a single preset time step in the Lie algebra tangent space and combined to generate the effective evolution velocity vector.

3. The method according to claim 1 or 2, characterized in that, The classification decoder uses a dynamically decoupled asymmetric classification loss function for global parameter updates during the model training phase. The asymmetric classification loss function is configured with a dynamic adjustment coefficient for cross-entropy and a false negative penalty amplification parameter for minority class abnormal feature samples. Specifically, the false negative penalty amplification parameter is non-linearly positively correlated with the reciprocal of the occurrence probability of the target abnormal physiological feature within the historical all-weather monitoring time window, and is truncated by a preset dynamic threshold to forcibly extract weak manifold jump features corresponding to low-frequency minority class waveform distortions.

4. The noise-resistant biomedical signal processing method based on latent topology evolution according to claim 1, characterized in that, The encoder network is a dual-stream encoder, and its feature extraction and fusion specifically includes: The one-dimensional time-series biomedical signal sequence is input into the time-series convolution branch, and at least one one-dimensional convolution and batch normalization are performed sequentially. The output feature map is flattened and then layer normalized to obtain a one-dimensional morphological feature vector. The static auxiliary features are input into the static fully connected branch and sequentially processed through at least one fully connected layer, an activation function, and layer normalization to output a static feature vector. The one-dimensional morphological feature vector and the static feature vector are concatenated by channels, and then compressed to a dimension that matches the target orthogonal constrained manifold through a fully connected dimensionality reduction layer to generate the global latent feature vector, which is located in the Lie algebra tangent space of the target manifold.

5. The noise-resistant biomedical signal processing method based on latent topology evolution according to claim 1, characterized in that, Step 3 employs a single-step deterministic analytical mapping with non-numerical orthogonalization iteration, specifically including: Extract the magnitude and unit direction vector of the effective evolution velocity vector; An antisymmetric matrix is ​​constructed based on the unit direction vector, and the three-dimensional topological state orthogonal matrix is ​​analytically generated in a single-step calculation using the Rodriguez rotation formula.

6. The noise-resistant biomedical signal processing method based on latent topology evolution according to claim 1, characterized in that, Based on the specific physiological type of the one-dimensional time-series biomedical signal sequence, the evolutionary dynamics system executes the corresponding adaptive perturbation configuration: If the biomedical signal is a photoplethysmography (PPG) signal, the generation scale of the envelope term in the diffusion network is adaptively increased to absorb continuous baseline drift caused by the optical environment. If the biomedical signal is an electroencephalogram (EEG) signal or a surface electromyogram (EMG) signal, the basic bias of the Poisson probability in the jumping network is adaptively increased to capture transient spikes or high-frequency EMG burst jumping features. If the biomedical signal is a pet multi-lead electrocardiogram signal, then the static auxiliary feature is configured to fuse the target pet's species category, weight, and body shape data, and adaptively increase the generation scale of the envelope term in the diffusion network.

7. A noise-resistant biomedical signal processing system based on latent topological evolution, characterized in that, include: The physiological signal acquisition module is used to acquire one-dimensional time-series biomedical signal sequences of the test subject in real time; Memory, used to store computer instructions and pre-trained weights of topology-evolution-based network models; The processor is communicatively connected to the physiological signal acquisition module and the memory, and the processor implements the noise-resistant biomedical signal processing method as described in any one of claims 1 to 6 when executing the computer instructions; And an alarm and communication module for performing local intervention or communication alarms based on control signals generated by the processor.

8. A computer-readable storage medium having a computer program or instructions stored thereon, characterized in that, When the computer program or instructions are executed by the processor, they implement the steps of a noise-resistant biomedical signal processing method based on potential topology evolution as described in any one of claims 1 to 6.

9. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instructions are executed by the processor, they implement the steps of a noise-resistant biomedical signal processing method based on potential topology evolution as described in any one of claims 1 to 6.

Images