Cloud edge collaborative early warning method for early recognition of cerebral apoplexy

Abstract

The invention discloses a cloud edge collaborative early warning method for early recognition of cerebral apoplexy, and the method comprises the steps: collecting the posture data and voice data of an edge end, and carrying out the preprocessing of the data; on the edge end, lightweight model reasoning is performed on the attitude data and the voice data based on a sub-modal processing and feature fusion strategy, and finally the cerebral apoplexy risk probability is output; and cooperative and multi-terminal linkage early warning is carried out based on the cloud and the edge terminal. According to the method, active capture of the cerebral apoplexy risk is realized through a framework of edge-end multi-modal real-time processing, cloud collaborative optimization and multi-end linkage early warning, the mode is converted from passive help calling to active early warning-rapid linkage, and the early warning precision and the response speed are improved.

Description

Technical Field

[0001] This invention belongs to the field of data processing technology, and in particular relates to a cloud-edge collaborative early warning method for early identification of stroke. Background Technology

[0002] Existing early warning technologies for stroke have three main limitations that make it difficult to meet clinical needs:

[0003] 1. Single-modal perception with high false alarm and false negative rates: Existing devices mostly use pure motion sensing (such as relying solely on IMU inertial sensors to monitor posture) or single-speech modality (such as recognizing "slurred speech" through a mobile phone microphone), which cannot simultaneously capture the full range of early stroke signs: face-arm-speech-balance (FAST). For example, pure motion sensing solutions (refer to patent CN118626823A, "A Fall Detection Method Based on Inertial Sensors") completely miss the symptom of "slurred speech," while single-speech solutions (refer to the paper "Research on Stroke Voice Early Warning Based on Mobile Phone Microphones") have a false alarm rate exceeding 20% ​​due to environmental noise and dialect differences.

[0004] 2. Edge-cloud separation, poor real-time performance: Most solutions adopt a separation architecture of "edge-end data collection + cloud-centralized processing". The cross-network transmission delay of raw data is generally >2 seconds (refer to the paper "Edge-cloud delay analysis of wearable stroke monitoring system"). It cannot meet the real-time early warning requirements of the "golden 4.5 hours" and the full upload of data poses a risk of privacy leakage (such as exposure of the target user's medical history and physiological data).

[0005] 3. Lack of end-to-end early warning through cloud-edge collaboration: Existing technologies have not formed a closed loop of "real-time processing at the edge - optimized scheduling in the cloud - multi-terminal linkage," and can only achieve local single early warning (such as terminal beeping alerts), failing to synchronously link with family members, 120 emergency services, and stroke centers, resulting in a break in the link. For example, a certain stroke monitoring bracelet (a publicly available product) can only send vibration alerts to the target user, and family members and hospitals cannot obtain risk information synchronously, delaying decision-making.

[0006] In summary, existing technologies, due to their limited sensing dimensions, lack of edge-cloud collaboration, and insufficient full-chain linkage, are unable to achieve "active capture" and "rapid response" of early stroke risks. There is an urgent need for an early warning method that integrates cloud-edge collaboration and multimodal fusion to fill the clinical gap. Summary of the Invention

[0007] Purpose of the Invention: The purpose of this invention is to provide a cloud-edge collaborative early warning method for early identification of stroke. It addresses the core problems of existing technologies, namely, high false alarm and false negative rates in single-modal approaches, poor real-time performance due to edge-cloud separation, and lack of end-to-end linkage.

[0008] Technical solution: The present invention provides a cloud-edge collaborative early warning method for early identification of stroke, comprising the following steps:

[0009] Step 1: Collect edge pose data and voice data, and preprocess the data;

[0010] Step 2: At the edge, based on the modal processing and feature fusion strategy, perform lightweight model inference on the pose data and speech data respectively, and finally output the stroke risk probability.

[0011] Step 3: Conduct collaborative and multi-terminal linkage early warning based on cloud and edge terminals.

[0012] Furthermore, step 1 specifically includes the following steps:

[0013] Step 1.1, Attitude Data Acquisition: Acquire triaxial acceleration and triaxial angular velocity. Use a low-pass filter to remove sensor thermal noise >100Hz and muscle tremor noise 10-50Hz to obtain filtered attitude data I. t The formula is:

[0014]

[0015] in Here are the original attitude data at time t, and b and a represent the numerator and denominator coefficients of the filter, respectively.

[0016] Step 1.2, Speech Data Acquisition: Acquire the speech signal, enhance the high-frequency components through pre-emphasis processing, and the transfer function is H(z) = 1 - 0.97z. -1 The preprocessed speech data V is obtained. t ;

[0017] Step 1.3, Space-Time Alignment: This is achieved by aligning the clock module for I... t With V t Add a unified timestamp to solve the problem of spatiotemporal asynchrony in multimodal data.

[0018] Furthermore, step 2 specifically includes the following steps:

[0019] Step 2.1: Constructing the posture anomaly recognition model STL-LSTM: The seasonal-trend decomposition (STL) cyclic trend decomposition method is used, and the Io is smoothed by local estimation of LOESS. t Decomposed into trend components Seasonal portion and residual components The decomposition formula is:

[0020]

[0021] right After normalization, the input is a two-layer LSTM network, which outputs the pose anomaly probability P. attitude The probability is mapped to the interval [0,1].

[0022] Step 2.2: Construct an accurate spectrogram and RepViT speech anomaly recognition model:

[0023] A damped forced vibration function set is designed to suppress spectral leakage, and the function formula is:

[0024]

[0025] Where γ = 0.05 is the damping coefficient, ω is the natural frequency of 100-5220Hz, m is the quality parameter, and F is the speech signal V to be analyzed. t x is displacement, t is time, and the damping function is... for The polynomial; this function group generates a 64×32 Mel-precise spectrogram, which is input into the RepViT lightweight convolutional network and outputs the speech anomaly probability P. speech ;

[0026] Step 2.3, Multimodal Fusion and Risk Output: P in the 32-dimensional embedding space attitude With P speech Weighted fusion is performed, and the stroke risk probability P is output through the Sigmoid activation function. stroke The formula is:

[0027] P stroke =Sigmoid(0.6P) attitude +0.4P speech ).

[0028] Furthermore, step 3 specifically includes the following steps:

[0029] Step 3.1, when the edge satisfies P stroke When the value exceeds 0.8 and remains above 3 seconds, the cloud-edge collaboration process is triggered; at the edge, the target user's ID, real-time GPS location, and P are encrypted and uploaded via an NB-IoT module. stroke The values ​​and multimodal feature vectors are sent to the cloud platform, and the warning status and target user information are displayed based on the closed-loop feedback.

[0030] Step 3.2: Extract abnormal facial features of the target user in the cloud, output the probability of facial abnormality based on the facial abnormality recognition model, and fuse the probability of facial abnormality with the cloud multimodal secondary fusion to update the probability of stroke risk.

[0031] Step 3.3, Multi-terminal dispatch: The early warning information is synchronously pushed to three types of terminals, including the family mobile APP, the 120 emergency platform and the stroke center system, via the MQTT protocol;

[0032] Step 3.4: Perform closed-loop feedback.

[0033] Furthermore, step 3.2 specifically includes the following steps:

[0034] Step 3.2.1, Facial Anomaly Recognition: The cloud receives the target user's facial photo uploaded by the rescuer through an instant website, and performs face detection and alignment: A multi-task cascaded convolutional network MTCNN is used to locate 68 facial key points. Affine transformation is used to correct the face to a frontal pose, i.e., the horizontal lines of the eyes are parallel to the horizontal axis of the image, and the image is cropped into alignment region I. align The formula is:

[0035] I align =wrapAffine(I raw ,M,(224,224))

[0036] Where M is the affine matrix calculated based on key points, I raw This is the original photo;

[0037] Illumination normalization: By limiting contrast adaptive histogram equalization (CLAHE) processing, feature blurring caused by backlighting / overexposure is suppressed and facial texture is enhanced;

[0038] Feature standardization: Normalize the image pixel values ​​to the [-1,1] interval to eliminate differences in device shooting and obtain the standardized region I. norm The formula is:

[0039]

[0040] Step 3.2.2, Facial Abnormality Feature Extraction: Based on typical facial symptoms of stroke, including unilateral facial paralysis, ptosis, and facial asymmetry, three core features are extracted:

[0041] Symmetry feature: Calculate the Euclidean distance difference between key points on the left and right sides of the face, including the distance difference D from the left and right corners of the mouth to the midline of the nose. mouth The height difference D between the left and right eyelids eye This forms a 12-dimensional symmetric eigenvector F. sym ;

[0042] Texture variation features: Local Binary Pattern (LBP) is used to extract facial skin texture, generating a 64-dimensional texture feature vector F. lbp ;

[0043] Dynamic deformation features: If multiple photos are uploaded, the displacement of key points in adjacent frames is calculated to form an 8-dimensional dynamic feature vector F. move ;

[0044] Step 3.2.3: Construct a facial anomaly recognition model: Use a lightweight convolutional neural network for inference;

[0045] Auxiliary information fusion and output: Facial abnormality probability P face It will participate in cloud-based multimodal secondary fusion to update the stroke risk probability as follows:

[0046] P stroke-final =Sigmoid(0.7P) stroke +0.3P face ).

[0047] Furthermore, step 3.4 specifically includes:

[0048] Receive response results from 120 / stroke centers in the cloud and transmit them back to edge terminals and mobile devices;

[0049] The warning status is updated via an alert screen at the edge terminal, forming a complete closed loop of monitoring, processing, early warning, scheduling, and feedback.

[0050] Update target user information and provide follow-up suggestions in the mobile app.

[0051] The present invention also discloses a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method of the present invention.

[0052] The present invention also discloses a computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implements the steps of the method of the present invention.

[0053] The present invention also discloses a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of the method of the present invention.

[0054] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages:

[0055] The main objective of this invention is to provide an early warning method for stroke identification based on a cloud-edge collaborative model. This method addresses the core problems of existing technologies, such as high false alarm and false negative rates in single-modal approaches, poor real-time performance due to edge-cloud separation, and lack of full-chain linkage. Through an architecture of "real-time multimodal processing at the edge + cloud-based collaborative optimization + multi-terminal linkage early warning," the invention achieves "active capture" of stroke risk, transforming the mode from "passive emergency call" to "active early warning and rapid linkage," thereby improving early warning accuracy and response speed and seizing the golden 4.5 hours.

[0056] Posture anomaly recognition: An STL-LSTM model was used, and posture data was decomposed into trend, seasonal, and residual components through LOESS local estimation smoothing to monitor slow posture anomalies, gait periodic asymmetry, and sudden acceleration mutations, respectively. Clinical testing showed that the model achieved an accuracy of 92.3% in recognizing hemiplegic gait and a sensitivity of 96.7% in recognizing sudden situations such as imbalance and falls.

[0057] Speech anomaly recognition: Based on a precise spectrogram + RepViT model architecture, a damped forced vibration function group is designed to suppress spectral leakage, generating a 64×32 Mel precise spectrogram. The lightweight RepViT convolutional network (model size <60KB) achieves a 91.5% accuracy rate in recognizing typical stroke symptoms such as "slurred speech," reducing the false alarm rate by 18.2% compared to traditional speech recognition solutions.

[0058] Facial anomaly recognition: The Lightweight FaceNet model employs a 4-layer depthwise separable convolutional architecture with global average pooling, boasting <1.2M parameters, a test set accuracy of ≥94%, and an F1-score of ≥0.92. It specifically utilizes the Focal Loss function to address class imbalance (normal:abnormal = 3:1), improving the recognition rate of difficult-to-classify samples such as unilateral facial paralysis and ptosis by over 15%.

[0059] In a 32-dimensional embedding space, the probabilities of pose abnormalities and speech abnormalities are weighted and fused (with weights of 0.6 and 0.4, respectively, calibrated based on clinical data), and the initial stroke risk probability is output through a sigmoid activation function. When the cloud-edge collaborative process is triggered, the cloud introduces the facial abnormality recognition results for secondary fusion. The updated stroke risk assessment accuracy reaches 96.8%, an average improvement of 23.5% compared to the single-modal solution. This effectively solves the problems of missed detection of "speech ambiguity" in pure motion sensing solutions and high false alarms caused by environmental noise and dialect differences in single-speech solutions. Attached Figure Description

[0060] Figure 1 This is a general overview diagram of the stroke cloud-edge collaborative early warning method of the present invention.

[0061] Figure 2 This is a flowchart of the present invention. Detailed Implementation

[0062] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0063] This invention is based on a cloud-edge collaborative architecture of "real-time edge processing - cloud-based collaborative optimization - multi-terminal linkage early warning". Through multimodal data fusion and hierarchical computing, it achieves proactive capture and full-chain early warning of early stroke risk. The specific process is as follows:

[0064] 1. Edge data acquisition and preprocessing

[0065] A wearable magnetic badge-style terminal (edge ​​device) based on the domestically produced ESP32-S3 chip is used. It employs a QMI8658C six-axis inertial sensor to collect attitude data and a digital microphone to collect voice data, performing the following operations:

[0066] (1) Attitude data acquisition: Triaxial acceleration (acc_x / acc_y / acc_z) and triaxial angular velocity (gyr_x / gyr_y / gyr_z) were acquired at a sampling rate of 60Hz. Low-pass filtering (numerator coefficient b = 0.1, denominator coefficient a = 0.9) was used to filter out sensor thermal noise >100Hz and human muscle tremor noise in the range of 10-50Hz, resulting in filtered attitude data I. t The formula is:

[0067]

[0068] in This is the original attitude data at time t;

[0069] (2) Speech data acquisition: Speech signals were acquired at a sampling rate of 16kHz and pre-emphasis was applied (transfer function H(z) = 1 - 0.97z). -1 Enhance high-frequency components (such as transient features like unclear speech) to obtain preprocessed speech data V. t ;

[0070] (3) Time and space alignment: I is aligned using the PCF85063AT clock module. t With V t Add a unified timestamp (accuracy ±1ms) to solve the problem of spatiotemporal asynchrony of multimodal data.

[0071] 2. Multimodal feature processing and risk calculation at the edge

[0072] At the edge, a "modal processing + feature fusion" strategy is adopted to perform lightweight model inference on pose and speech data respectively, and finally output the stroke risk probability:

[0073] (1) Anomaly detection (STL-LSTM model):

[0074] The STL (Seasonal-Trend Decomposition) cyclic trend decomposition method is used, and I is smoothed using LOESS (Local Estimation Smoothing). t Decomposed into trend components (Monitoring slow postural abnormalities, such as hemiplegic gait), seasonal component (Capturing gait periodic asymmetry), residual components (Identify sudden acceleration mutations, such as imbalance and falls), and the decomposition formula is:

[0075]

[0076] For After normalization (eliminating dimensional differences), it is input into a two-layer LSTM network (with 16 hidden neurons, and 30% of the neurons in the Dropout layer are randomly discarded to prevent overfitting), and the output is the probability of abnormal posture P attitude (Mapped to the [0, 1] interval);

[0077] (2) Voice anomaly recognition (accurate spectrogram + RepViT model):

[0078] Design a damped forced vibration function group to suppress spectral leakage, and the function formula is:

[0079]

[0080] Where F is the voice signal V t , γ = 0.05 is the damping coefficient, ω is the natural frequency from 100 - 5220 Hz (covering the key frequency band of voice), and m is the mass parameter; through this function group, a 64×32 Mel accurate spectrogram is generated and input into the RepViT lightweight convolutional network (model volume < 60KB), and the output is the voice anomaly probability P speech ;

[0081] (3) Multimodal fusion and risk output:

[0082] In the 32-dimensional embedding space, P attitude And P speech Are weighted and fused (weights are 0.6 and 0.4 respectively, calibrated based on clinical data), and the stroke risk probability P stroke Is output through the Sigmoid activation function, and the formula is:

[0083] P stroke = Sigmoid(0.6P attitude + 0.4P speech )

[0084] 3. Cloud-edge collaboration and multi-terminal linkage warning

[0085] When the edge device satisfies "P stroke > 0.8 and lasts for 3 seconds", it triggers the cloud-edge collaboration process:

[0086] (1) Edge device processing:

[0087] ① Data encryption and upload: Upload the target user ID, real-time GPS location, and P strokeValues ​​and multimodal feature vectors are sent to the cloud platform, and only abnormal data is uploaded (not the full data), reducing power consumption and transmission latency;

[0088] ② Target user information display: The closed-loop feedback content displays the necessary early warning status and target user information.

[0089] (2) Cloud processing:

[0090] ① Supplementing facial image data of target users: Generate an instant website for rescuers to take photos and upload facial features of target users. Rescuers can access the page by scanning a code or touching it with NFC, thereby uploading photos or supplementing notes;

[0091] ② Facial anomaly recognition:

[0092] 1) Facial Image Preprocessing: The cloud receives the target user's facial photo (JPG / PNG format, resolution ≥640×480) uploaded by the rescuer via a real-time website, and performs the following preprocessing:

[0093] a. Face detection and alignment: MTCNN (Multi-Task Cascaded Convolutional Network) was used to locate 68 facial key points (such as the corners of the eyes, mouth, and nose). The face was corrected to a frontal pose (the horizontal lines of the eyes are parallel to the horizontal axis of the image) through affine transformation, and then cropped into a 224×224 pixel alignment region I. align The formula is:

[0094] I align =wrapAffine(I raw ,M,(224,224))

[0095] Where M is the affine matrix calculated based on key points, I raw This is the original photo;

[0096] b. Illumination Normalization: Through CLAHE (Limited Contrast Adaptive Histogram Equalization), feature blurring caused by backlighting / overexposure is suppressed, and facial texture (such as the difference in brightness in sagging areas of the face) is enhanced. Parameter settings: clipLimit = 2.0, mesh size 8×8.

[0097] c. Feature Standardization: Normalize the image pixel values ​​to the [-1, 1] interval to eliminate differences in device shooting and obtain the standardized region I. norm The formula is:

[0098]

[0099] 2) Facial Abnormality Feature Extraction: Based on typical facial symptoms of stroke (such as unilateral facial paralysis, ptosis, and facial asymmetry), three core features are extracted:

[0100] a. Symmetry features: Calculate the Euclidean distance difference between key points on the left and right sides of the face, such as the distance difference D from the left and right corners of the mouth to the midline of the nose. mouth The height difference D between the left and right eyelids eye This forms a 12-dimensional symmetric eigenvector F. sym ;

[0101] b. Texture variation features: Local Binary Pattern (LBP) is used to extract facial skin texture (such as texture changes caused by muscle stiffness on the side of facial paralysis), generating a 64-dimensional texture feature vector F. lbp ;

[0102] c. Dynamic Deformation Features: If multiple photos are uploaded (e.g., a sequence of images of the target user attempting to smile or close their eyes), the displacement of key points in adjacent frames (e.g., the difference in the degree of upward slant of the corners of the mouth) is calculated to form an 8-dimensional dynamic feature vector F. move .

[0103] 3) Facial Anomaly Detection Model (Lightweight FaceNet): Inference can be performed using a lightweight convolutional neural network. The model structure is as follows:

[0104] a. Network architecture: 4 layers of depthwise separable convolutions (output channels 32→64→128→256) + global average pooling layer + 2 layers of fully connected layers (64 hidden neurons), model parameters <1.2M;

[0105] b. Input / Output: Input is the preprocessed 224×224 facial image I. nrom The output is the probability P of facial abnormality. face (Mapped to the [0,1] interval, the higher the value, the more pronounced the symptoms such as facial paralysis);

[0106] c. Training strategy: The model was trained on a clinical stroke facial abnormality dataset (containing 5000+ labeled samples). Focal Loss (a focus loss function that forces the model to focus on difficult-to-distinguish minority class samples by "dynamically adjusting sample weights") was used to solve class imbalance (normal:abnormal = 3:1). The test set accuracy was ≥94%, and the F1-score was ≥0.92.

[0107] ③ Auxiliary information fusion and output: Facial abnormality probability P face It will participate in cloud-based multimodal secondary fusion (supplementing the original solution's posture and speech fusion), and update the stroke risk probability as follows:

[0108] P stroke-final =Sigmoid(0.7P) stroke +0.3P face )

[0109] ④ Data storage: Establish a dedicated database for target users to store historical monitoring data (posture / voice characteristics, early warning records);

[0110] ⑤ Model optimization:

[0111] 1) Based on massive multimodal data (such as more than 500 hours of pose / speech monitoring data and 100,000 annotated facial images) aggregated from multiple edge devices, federated learning (such as the FedAvg algorithm, with a learning rate of 0.001 and 50 iterations) is used to collaboratively update the parameters of STL-LSTM (pose anomaly recognition), RepViT (speech anomaly recognition) and Lightweight FaceNet (facial anomaly recognition) models.

[0112] 2) To address the privacy sensitivity and non-independent identical distribution characteristics of facial image data (such as individual skin color and age differences), optimize client-side screening strategies (e.g., a single client must have ≥50 facial images with an annotation accuracy of ≥90%) and local training parameters (e.g., use Focal Loss to solve class imbalance, and use Batch Size = 16 to adapt to edge computing power). Through a closed loop of "local training - encrypted parameter upload - cloud aggregation - global parameter distribution", simultaneously improve the cross-individual generalization ability of the three types of models, thereby improving the overall multimodal early warning system's recognition accuracy for different groups of people by 3%-5%.

[0113] (3) Multi-terminal scheduling: Early warning information is synchronously pushed to three types of terminals via the MQTT protocol:

[0114] ① Family member mobile app: Displays the target user's real-time location, P stroke Value, risk level (e.g., "high risk", "normal");

[0115] ②120 emergency platform: Generates the optimal emergency route (based on Baidu Maps API), and marks the target user's location and key medical history information (such as "previous ischemic stroke");

[0116] ③ Stroke Center System: Pushes real-time multimodal feature vectors and historical medical history of target users to assist doctors in formulating treatment plans;

[0117] (4) Closed-loop feedback:

[0118] ①Cloud: Receive response results from 120 / stroke center (e.g., "Ambulance has been dispatched and is expected to arrive in 8 minutes") and transmit them back to edge terminals and mobile devices;

[0119] ② Edge terminal: The warning status is updated through the warning screen, forming a closed loop of "monitoring-processing-early warning-scheduling-feedback";

[0120] Mobile App: Update target user information and provide follow-up suggestions.

[0121] Real-time performance of cloud-edge collaboration: The cloud-edge collaboration architecture requires efficient processing in three aspects: edge devices, transmission methods, and cloud response to overcome the bottleneck of the current separation between edge and cloud.

[0122] High-efficiency edge processing: Hardware selection: A wearable magnetic badge-style terminal using the domestically produced ESP32-S3 chip, integrating a six-axis inertial sensor and a digital microphone. In the alternative solution, the STM32H743 chip has an inference latency of <50ms and power consumption of <150mW, meeting the requirements for lightweight real-time processing.

[0123] Algorithm optimization: Attitude data is filtered by low-pass filtering (numerator coefficient b = 0.1, denominator coefficient a = 0.9) to remove noise; speech data is enhanced by pre-emphasis processing to improve high-frequency characteristics; the PCF85063AT clock module achieves spatiotemporal alignment with an accuracy of ±1ms to ensure the synchronization of multimodal data.

[0124] Transmission delay control: Selective upload strategy: Abnormal data (target user ID, real-time GPS, risk value and multimodal feature vector) is encrypted and uploaded only when the risk probability is judged to exceed the threshold for 3 consecutive seconds at the edge, instead of being uploaded in full, reducing the amount of data transmission by more than 70%.

[0125] Communication protocol adaptation: The communication method is dynamically selected according to the scenario. LoRa modules (transmission distance <10km) are used in areas with weak NB-IoT signals; Cat-M1 is selected when large amounts of data need to be transmitted; and eMTC is used to achieve low-power bidirectional interaction in wide-area scenarios such as urban-rural fringe areas. The optimized data transmission latency is controlled within 300ms, which is more than 85% better than the traditional full-data upload solution (latency is generally >2 seconds).

[0126] Rapid cloud response: The cloud adopts a lightweight processing workflow, with facial image preprocessing (MTCNN key point localization, CLAHE illumination normalization, feature standardization) taking less than 200ms, multimodal secondary fusion calculation taking less than 100ms, and the overall cloud processing latency controlled within 500ms, meeting the real-time early warning requirements.

[0127] Privacy and security:

[0128] Data transmission security: Data uploads are encrypted using the AES-128 algorithm at the edge. In resource-constrained scenarios, the lightweight Simon encryption algorithm can be used. Hardware serial / parallel pipelined processing improves encryption and decryption efficiency. This encrypted transmission solution meets medical data security standards and has a low risk of data leakage.

[0129] Federated learning privacy protection: Based on massive multimodal data (over 500 hours of pose / speech monitoring data and 100,000 annotated facial images) aggregated from multiple edge devices, the FedAvg algorithm (learning rate 0.001, 50 iterations) is used to collaboratively update model parameters. Through a closed loop of "local training - encrypted parameter upload - cloud aggregation - global parameter distribution," the raw data is always stored locally, and only the model parameters are shared. To address the privacy sensitivity of facial images, the client-side screening strategy is optimized (≥50 facial images per client with annotation accuracy ≥90%), ensuring that the model optimization process complies with privacy regulations such as HIPAA and GDPR.

[0130] End-to-end collaborative response:

[0131] Multi-terminal linkage mechanism: Warning information is synchronously pushed to three types of terminals via the MQTT protocol: the family's mobile app displays real-time location and risk level; the 120 emergency platform generates the optimal emergency route based on the Baidu Maps API; and the stroke center system receives real-time multimodal characteristics and medical history data. The WebSocket protocol is used to achieve real-time two-way interaction between the cloud and the terminal, improving response speed by 25% compared to traditional one-way push.

[0132] Optimized response time: Test data shows that the cloud-edge collaborative early warning method controls the entire response time from anomaly detection to emergency response dispatch to within 8 minutes, which is 46.7% shorter than the traditional "passive emergency call" mode (average delay of over 15 minutes), thus seizing a valuable golden window for stroke target users. The closed-loop feedback mechanism ensures that the edge terminal updates the early warning status in real time, forming a complete closed loop of "monitoring-processing-early warning-dispatch-feedback".

Claims

1. A cloud-edge collaborative early warning method for early stroke identification, characterized in that, Includes the following steps: Step 1: Collect edge pose data and voice data, and preprocess the data; Step 2: At the edge, based on the modal processing and feature fusion strategy, perform lightweight model inference on the pose data and speech data respectively, and finally output the stroke risk probability. Step 3: Conduct collaborative and multi-terminal linkage early warning based on cloud and edge terminals.

2. The cloud-edge collaborative early warning method for early stroke identification according to claim 1, characterized in that, Step 1 specifically includes the following steps: Step 1.1, Attitude Data Acquisition: Acquire triaxial acceleration and triaxial angular velocity. Use a low-pass filter to remove sensor thermal noise >100Hz and muscle tremor noise 10-50Hz to obtain filtered attitude data I. t The formula is: in Here are the original attitude data at time t, and b and a represent the numerator and denominator coefficients of the filter, respectively. Step 1.2, Speech Data Acquisition: Acquire the speech signal, enhance the high-frequency components through pre-emphasis processing, and the transfer function is H(z) = 1 - 0.97z. -1 The preprocessed speech data V is obtained. t ; step 1.3 Spatiotemporal Alignment: This is achieved through the clock module for I... t With V t Add a unified timestamp to solve the problem of spatiotemporal asynchrony in multimodal data.

3. The cloud-edge collaborative early warning method for early stroke identification according to claim 1, characterized in that, Step 2 specifically includes the following steps: Step 2.1: Constructing the posture anomaly recognition model STL-LSTM: The seasonal-trend decomposition (STL) cyclic trend decomposition method is used, and the Io is smoothed by local estimation of LOESS. t Decomposed into trend components Seasonal portion and residual components The decomposition formula is: right After normalization, the input is a two-layer LSTM network, which outputs the pose anomaly probability P. attitude The probability is mapped to the interval [0,1]. Step 2.2: Construct an accurate spectrogram and RepViT speech anomaly recognition model: A damped forced vibration function set is designed to suppress spectral leakage, and the function formula is: Where F is the speech signal V t γ = 0.05 is the damping coefficient, ω is the natural frequency of 100-5220Hz, m is the quality parameter, and F is the speech signal V to be analyzed. t x represents displacement, and t represents time; this function group generates a 64×32 Mel-precise spectrogram, which is then input into the RepViT lightweight convolutional network, and outputs the speech anomaly probability P. speech ; Step 2.3, Multimodal Fusion and Risk Output: P in the 32-dimensional embedding space attitude With P speech Weighted fusion is performed, and the stroke risk probability P is output through the Sigmoid activation function. stroke The formula is: P stroke =Sigmoid(0.6P attitude +0.4P speech )。 4. The cloud-edge collaborative early warning method for early stroke identification according to claim 3, characterized in that, Step 3 specifically includes the following steps: Step 3.1, when the edge satisfies P stroke When the value is >0.8 and remains above 3 seconds, the cloud-edge collaboration process is triggered. At the edge, the target user's ID, real-time GPS location, and P are encrypted and uploaded via an NB-IoT module. stroke The values ​​and multimodal feature vectors are sent to the cloud platform, and the warning status and target user information are displayed based on the closed-loop feedback. Step 3.2: Extract abnormal facial features of the target user in the cloud, output the probability of facial abnormality based on the facial abnormality recognition model, and fuse the probability of facial abnormality with the cloud multimodal secondary fusion to update the probability of stroke risk. Step 3.3, Multi-terminal dispatch: The early warning information is synchronously pushed to three types of terminals, including the family mobile APP, the 120 emergency platform and the stroke center system, via the MQTT protocol; Step 3.4: Perform closed-loop feedback.

5. A cloud-edge collaborative early warning method for early stroke identification according to claim 4, characterized in that, Step 3.2 specifically includes the following steps: Step 3.2.1, Facial Anomaly Recognition: The cloud receives the target user's facial photo uploaded by the rescuer through an instant website, and performs face detection and alignment: A multi-task cascaded convolutional network MTCNN is used to locate 68 facial key points. Affine transformation is used to correct the face to a frontal pose, i.e., the horizontal lines of the eyes are parallel to the horizontal axis of the image, and the image is cropped into alignment region I. align The formula is: I align =wrapAffine(I raw ,M,(224,224)) Where M is the affine matrix calculated based on key points, I raw This is the original photo; Illumination normalization: By limiting contrast adaptive histogram equalization (CLAHE) processing, feature blurring caused by backlighting / overexposure is suppressed and facial texture is enhanced; Feature standardization: Normalize the image pixel values ​​to the [-1,1] interval to eliminate differences in device shooting and obtain the standardized region I. norm The formula is: Step 3.2.2, Facial Abnormality Feature Extraction: Based on typical facial symptoms of stroke, including unilateral facial paralysis, ptosis, and facial asymmetry, three core features are extracted: Symmetry feature: Calculate the Euclidean distance difference between key points on the left and right sides of the face, including the distance difference D from the left and right corners of the mouth to the midline of the nose. mouth The height difference D between the left and right eyelids eye This forms a 12-dimensional symmetric eigenvector F. sym ; Texture variation features: Local Binary Pattern (LBP) is used to extract facial skin texture, generating a 64-dimensional texture feature vector F. lbp ; Dynamic deformation features: If multiple photos are uploaded, the displacement of key points in adjacent frames is calculated to form an 8-dimensional dynamic feature vector F. move ; Step 3.2.3: Construct a facial anomaly recognition model: Use a lightweight convolutional neural network for inference; Auxiliary information fusion and output: Facial abnormality probability P face It will participate in cloud-based multimodal secondary fusion to update the stroke risk probability as follows: P stroke-final =Sigmoid(0.7P stroke +0.3P face )。 6. The cloud-edge collaborative early warning method for early stroke identification according to claim 4, characterized in that, Step 3.4 specifically involves: Receive response results from 120 / stroke centers in the cloud and transmit them back to edge terminals and mobile devices; The warning status is updated via an alert screen at the edge terminal, forming a complete closed loop of monitoring, processing, early warning, scheduling, and feedback. Update target user information and provide follow-up suggestions in the mobile app.

7. A computer device comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method of claim 1.

8. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method of claim 1.

9. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method of claim 1.

Images