Biomedical multi-mode signal anomaly detection and prediction method and system

Abstract

The invention discloses a biomedical multi-mode signal anomaly detection and prediction method and system, and the method comprises the steps: modeling a biomedical multi-mode signal into a time-varying non-local dynamic system, capturing the long-time-history dependence characteristic of the signal through a memory mechanism of a Caputo fractional derivative, and introducing a time-varying input item to process interference; the method comprises the following steps of: carrying out high-precision numerical integration by adopting an Adams-Bashform-Module solver; optimizing model parameters in combination with a local domain normalization pre-training strategy and a fusion loss function; abnormal detection is realized by calculating comparison between signal reconstruction deviation and a self-adaptive threshold value; a future anomaly probability is generated based on the trajectory prediction. According to the method, the problems of insufficient non-local dependence capture, poor time-varying interference robustness, high false positive rate and the like in the prior art are effectively solved, the accuracy and real-time performance of abnormal detection and prediction of biomedical signals such as electroencephalogram and electrocardiogram are remarkably improved, and the method is suitable for wearable medical equipment and clinical monitoring systems.

Description

Technical Field

[0001] This invention belongs to the field of biomedical signal processing technology, specifically relating to a biomedical multimodal signal anomaly detection and prediction method and system. Background Technology

[0002] With the deepening application of artificial intelligence technology in medical diagnosis, real-time and accurate anomaly detection and prediction of biomedical signals has become a key requirement for clinical monitoring and wearable medical devices. In particular, abnormal patterns in multimodal physiological signals, such as electroencephalograms (EEG) and electrocardiograms (ECG), often indicate important pathological conditions, such as epileptic seizures or arrhythmias. These signals exhibit significant non-stationarity, time-varying characteristics, and noise interference, posing a significant challenge to traditional signal processing methods.

[0003] Currently, biomedical signal anomaly detection technologies are mainly divided into two categories: traditional signal processing algorithms and deep learning models. Traditional methods, such as time-frequency analysis and wavelet transform, are mainly used to extract feature patterns from signals. For example, the paper "Deep Learning for EEG-Based Epileptic Seizure Detection" (IEEE Journal of Biomedical and Health Informatics, 2020) proposes a method combining continuous wavelet transform (CWT) and long short-term memory (LSTM) networks. This method uses Morlet wavelets to convert EEG / ECG signals into time-spectrum maps, and then uses an LSTM network (128-256 hidden units, 2-3 layers) to capture temporal dependencies, achieving a detection accuracy of approximately 85%-92%. However, such methods have obvious drawbacks: First, relying on fixed time-frequency decomposition strategies, they cannot effectively capture the non-local dependency characteristics of biological signals, leading to a significant increase in error in long sequence prediction (RMSE>0.05); second, in environments with time-varying interference (such as motion artifacts), the time-spectrum graph is prone to distortion, causing a 10%-15% decrease in detection accuracy; third, the computational complexity is high, and real-time performance is difficult to guarantee when processing multimodal signals, with processing delays often exceeding 0.5 seconds, limiting their application in mobile medical devices.

[0004] On the other hand, with the development of fractional calculus theory, some researchers have attempted to incorporate it into neural network models. For example, Jiang et al. proposed a recurrent neural network (FORNN) based on Caputo fractional ordinary differential equations (CFODEs) in their paper "Fractional-Order RNNs: Universal Approximation Framework for Non-Local Dynamic System Modeling" (IEEE Transactions on Signal Processing, 2025). The model can be expressed as follows: ,in In hidden state, It is of fractional order. It is a time constant. This is the weight matrix. The activation function is . The model uses an Adams-Bashforth-Moulton (ABM) solver for numerical integration and employs Local Domain Normalization (LDN) for parameter pre-training. This method theoretically proves the general approximation ability of fractional-order models for nonlocal dynamic systems. However, this model is only applicable to time-invariant (autonomous) systems and cannot directly handle time-varying interference commonly found in biomedical signals. Furthermore, parameter domain problems (PDPs) and hallucination problems are prominent during training, leading to unstable convergence and detection bias in noisy EEG / ECG sequences. Additionally, the model lacks specific optimization for abnormal event classification, resulting in lower recognition accuracy compared to dedicated clinical algorithms.

[0005] In summary, existing technologies face three core challenges in the detection and prediction of anomalies in biomedical multimodal signals: First, insufficient capture of nonlocal dependencies, making it difficult for traditional methods to model long-term time-dependent relationships of signals; second, poor robustness to time-varying interference, with existing models exhibiting significant performance degradation in dynamic noise environments; and third, low detection accuracy, leading to a high false alarm rate (>15%) and poor prediction reliability. Therefore, there is an urgent need for a new method capable of simultaneously handling nonlocal dependencies and time-varying interference, and possessing high-precision anomaly detection capabilities, to meet the real-time monitoring needs of biomedical signals in clinical and wearable medical devices. This invention presents an innovative solution addressing these technical deficiencies. Summary of the Invention

[0006] This invention addresses the shortcomings of existing technologies by providing a method and system for detecting and predicting anomalies in biomedical multimodal signals.

[0007] To achieve the above-mentioned objectives, the technical solution adopted by the present invention is as follows:

[0008] A biomedical multimodal signal anomaly detection and prediction method includes the following steps:

[0009] Receive biomedical multimodal signal data and perform normalization preprocessing;

[0010] A Caputo fractional recurrent neural network model is constructed to represent the biomedical multimodal signals as a time-varying Caputo fractional ordinary differential equation system. The model equations are as follows:

[0011]

[0012] in, This represents the Caputo fractional derivative operator. It is a fractional parameter with a value range of 0.6-0.95. for 3D hidden state vector, It is 64-256. It is a time constant with a value range of 0.8-1.5. This is the weight matrix. For activation function, The input vector is time-varying;

[0013] The model equations were numerically integrated using the Adams-Bashforth-Moulton solver to obtain the model output. ;

[0014] Model parameters are optimized through local domain normalization pre-training and fusion loss function;

[0015] Based on model output With the original signal Calculate the deviation And through an adaptive threshold mechanism Identify abnormal events;

[0016] Predicting the future based on the current trajectory The probability of an anomaly in the step. Values ​​range from 5 to 20, generating prediction results.

[0017] Furthermore, the numerical integration process of the Adams-Bashforth-Moulton solver includes a prediction step and a correction step:

[0018] Prediction step:

[0019]

[0020] Calibration step:

[0021]

[0022] in, This is the integration step size, and its value ranges from 0.001 to 0.01. For the Gamma function, and For memory weighting coefficients, , , .

[0023] Furthermore, the fusion loss function is:

[0024]

[0025] in, The original signal trajectory, Output for model reconstruction For the current parameter, , , For bias terms, For pre-trained parameters, It is a balance coefficient and its value ranges from 0.85 to 0.95.

[0026] Furthermore, the adaptive threshold Following Caputo fractional differential equations:

[0027]

[0028] Wherein, initial value The value ranges from 0.05 to 0.1 and is updated using the Adams-Bashforth-Moulton solver.

[0029] Furthermore, the biomedical multimodal signals include electroencephalogram (EEG) signals and electrocardiogram (ECG) signals, wherein the number of EEG signal channels is 8-32, the number of ECG signal channels is 1-12, and the signal sampling rate is 250-500Hz.

[0030] This invention also discloses a biomedical multimodal signal anomaly detection and prediction system for implementing the above method, characterized in that it comprises:

[0031] Data input interface for receiving and preprocessing biomedical multimodal signal data;

[0032] The model computation module is used to implement the Caputo fractional recurrent neural network model and the Adams-Bashforth-Moulton solver, and to compute the hidden state and output.

[0033] The optimization module is used to perform local domain normalization pre-training and optimization of the fusion loss function, and to update the model parameters;

[0034] The anomaly detection module is used to calculate signal deviation and adaptive threshold, mark abnormal events, and predict the probability of future anomalies.

[0035] The data output interface is used to generate and transmit anomaly detection results and prediction data;

[0036] The modules are electrically connected via a data bus to form a serial data processing flow: data input interface, model calculation module, optimization module, anomaly detection module, and data output interface.

[0037] Furthermore, the model calculation module includes an integral calculation submodule, used to implement the prediction-correction logic of the Adams-Bashforth-Moulton solver and receive input signals. and time-varying interference Output model reconstruction signal .

[0038] Furthermore, the optimization module includes a gradient calculation submodule and a loss function submodule. The loss function submodule is placed before the gradient calculation submodule and is used to calculate the fusion loss function and backpropagate the gradient to update the model parameters.

[0039] Furthermore, the anomaly detection module includes a deviation comparison submodule and a threshold update submodule. The threshold update submodule runs in parallel with the deviation comparison submodule and is used to implement the fractional threshold equation and update the adaptive threshold.

[0040] Furthermore, the system is deployed on a general-purpose computer device, including a processor, memory, and input / output devices. The processor is a CPU or GPU used to perform model calculations and parameter optimization. The memory is used to store model parameters, training data, and program code. The signal processing latency is controlled within 0.5 seconds, the anomaly detection accuracy is greater than 92%, and the false positive rate is less than 5%.

[0041] Compared with the prior art, the advantages of the present invention are as follows:

[0042] 1. This invention utilizes the nonlocal memory property of fractional derivatives to effectively capture long-term dependencies in biomedical signals, overcoming the problem of prediction error accumulation in traditional methods when processing long-sequence signals, and significantly improving the accuracy of signal modeling.

[0043] 2. By introducing time-varying input terms into the model, this invention can effectively adapt to and resist various time-varying interferences (such as motion artifacts, environmental noise, etc.), maintain stable detection performance in complex clinical environments, and solve the defect of significant performance degradation of existing technologies in interference environments.

[0044] 3. The Adams-Bashforth-Moulton solver is used for numerical integration, and the optimized calculation process significantly reduces the computational complexity of fractional-order models, making real-time processing of multimodal signals possible and meeting the low-latency processing requirements of clinical monitoring and wearable medical devices.

[0045] 4. This invention expands the applicability of fractional recurrent neural networks, enabling them to handle time-varying nonlinear systems. It overcomes the limitation of traditional fractional models being applicable only to time-invariant systems and better adapts to the dynamic variability of biomedical signals.

[0046] 5. By using a local domain normalization pre-training strategy and a collaborative optimization mechanism of the fusion loss function, the parameter domain problem and training illusion are effectively alleviated, and the convergence stability and generalization ability of the model in high-noise environments are improved.

[0047] 6. The adaptive threshold mechanism designed in this invention can dynamically adjust the anomaly judgment criteria according to signal characteristics. Combined with the deviation calculation method, it significantly improves the accuracy and specificity of anomaly event classification and reduces the false alarm rate.

[0048] 7. The overall system architecture adopts a closed-loop feedback design, which organically integrates signal modeling, parameter optimization and anomaly detection to form a complete processing flow. This not only improves the ability to capture non-local dependencies, but also enhances the robustness to time-varying interference, making the prediction results more reliable. Attached Figure Description

[0049] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0050] Figure 1 This is a schematic diagram of the structure of the biomedical multimodal signal anomaly detection and prediction system in this embodiment of the invention;

[0051] Figure 2 This is a flowchart of the biomedical multimodal signal anomaly detection and prediction method in this embodiment of the invention;

[0052] Figure 3 This is a schematic diagram of the model in an embodiment of the present invention;

[0053] Figure 4 This is a time series diagram of the method in the embodiments of the present invention. Detailed Implementation

[0054] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0055] I. System Implementation Environment

[0056] The system implementation environment of this invention is a general-purpose computer device, including a central processing unit (CPU) or graphics processing unit (GPU), random access memory (RAM), read-only memory (ROM), input / output interfaces, and a bus system. In specific implementations, the processor can be an Intel Xeon series or an NVIDIA Tesla series, the memory capacity is not less than 16GB, the operating system can be Linux or Windows, the programming language is Python 3.8 or above, and the deep learning framework is PyTorch 1.10 or TensorFlow 2.6.

[0057] II. System Module Implementation

[0058] like Figure 1 As shown, the system of the present invention includes the following core modules:

[0059] 1. Data Input Interface Module: Implemented as a software API interface, this module receives biomedical multimodal signal data. It includes a data buffer and parsing functions, capable of processing biomedical signals such as EEG (8-32 channels) and ECG (1-12 channels), and supports standard medical data formats such as EDF and MAT. In its code implementation, this module employs a multi-threaded design to ensure that data reading does not block subsequent processing flows.

[0060] 2. Model Computation Module: Implements the Caputo Fractional Recurrent Neural Network (CFORNNs) model and the Adams-Bashforth-Moulton (ABM) solver. The core of this module is the numerical integration algorithm for fractional differential equations, including complete implementations of the prediction and correction steps. In a GPU environment, this module utilizes CUDA parallel computing to accelerate the integration process, significantly improving computational efficiency.

[0061] 3. Optimization Module: Implements Local Domain Normalization (LDN) pre-training strategy and fusion loss function calculation. This module includes gradient calculation submodules and loss function submodules, employs automatic differentiation technology to calculate fractional gradients, supports optimizers such as Adam and SGD, and can dynamically adjust the learning rate and batch size.

[0062] 4. Anomaly Detection Module: Implements deviation calculation and adaptive thresholding mechanism. This module runs in parallel with two sub-modules: the deviation comparison sub-module is responsible for calculating the Euclidean distance between the original signal and the model output; the threshold update sub-module implements the fractional-order threshold equation and dynamically adjusts the anomaly judgment criteria. The two sub-modules interact through shared memory to ensure the synchronization of threshold updates and deviation calculations.

[0063] 5. Data Output Interface Module: Generates structured anomaly detection results and predicted data, supporting real-time visualization and historical data playback. This module provides a standard API interface, which can be integrated with hospital information systems (HIS) or remote monitoring platforms.

[0064] like Figure 3 As shown, the internal signal flow of this invention presents a closed-loop feedback system structure. This block diagram clearly illustrates the complete process from the input signal, through the integrator to generate the model output, and then through an optimization loop to dynamically adjust the model parameters. In this structure, the anomaly detection part, as a parallel computing branch, calculates the signal reconstruction deviation in real time, forming a feedback mechanism with the main signal flow. Specifically, the input signal... Interference items First, the data is fed into the ABM solver to generate the hidden state. and output Output On the one hand, with the original signal The comparison generates reconstruction errors, while the data is sent to the anomaly detection module to calculate the deviation. Reconstruction errors update model parameters through backpropagation, while biases... With adaptive threshold Anomaly markers are generated after comparison; the adaptive threshold itself is also updated in real time through fractional-order dynamic equations, forming a complete closed-loop control system. This architecture design ensures that the model can simultaneously optimize signal reconstruction accuracy and anomaly detection sensitivity, improving the overall performance and robustness of the system.

[0065] III. Implementation Process of Biomedical Multimodal Signal Anomaly Detection and Prediction Methods

[0066] like Figure 2 As shown, the method of the present invention includes the following specific steps:

[0067] Step 1: Data Preprocessing

[0068] 1.1 Acquire biomedical multimodal signals, including electroencephalogram (EEG) and electrocardiogram (ECG), with the sampling rate set to 250-500Hz to ensure compliance with the Nyquist sampling theorem.

[0069] 1.2 The original signal is normalized, mapping the data of each channel to the interval [-1, 1], using the following formula:

[0070]

[0071] 1.3 To enhance the robustness of the model, simulated time-varying disturbances can be optionally injected. The disturbance signal... Sine waves, Gaussian white noise, or common clinical artifact models can be used, with the amplitude range controlled within... Inside.

[0072] 1.4 The preprocessed signal is divided into fixed-length sequences, each containing 500-2000 sampling points, with 20% overlap between adjacent sequences to avoid boundary effects.

[0073] Step 2: Model Building and Forward Integration

[0074] 2.1 Construct a Caputo fractional recurrent neural network model, mathematically expressed as:

[0075]

[0076] Among them, fractional-order parameters Set to 0.6-0.95, time constant Set to 0.8-1.5 to hide the state dimension. Set to 64-256, weight matrix Initialize to a normal distribution with a mean of 0 and a variance of 0.01, and use the sigmoid function as the activation function σ. It is a time-varying input vector.

[0077] 2.2 Numerical integration was performed using the Adams-Bashforth-Moulton (ABM) solver:

[0078] Initialization: Set the initial state Integral step size Calculate the Gamma function value ;

[0079] Prediction step: For each time point ,calculate

[0080]

[0081] Among them, the memory weight coefficient ;

[0082] Correction step: Calculation

[0083]

[0084] Where the correction coefficient , ;

[0085] Output generated: ,in This is the output weight matrix.

[0086] like Figure 4 As shown, the numerical integration process of this invention exhibits a clear sequential characteristic in the time dimension. From the initial moment... Initially, the hidden state z(0) originates from the compact domain. Obtained by mid-sampling; at each integration step At this point, the system injects time-varying input. Then the state function is calculated. ; Execute the prediction step of the ABM solver to obtain the predicted state. Then, the accurate state is obtained through a calibration step. These state values ​​accumulate along the time axis to the end of the sequence T, ultimately generating the complete output trajectory. The figure particularly emphasizes the non-local dependency property, which is achieved by memorizing the weight coefficients. and The accumulation and manifestation of these coefficients make the current state dependent on all historical states, rather than just the most recent few time points, thus effectively capturing the long-term dependencies of biomedical signals. This time series processing mechanism enables the model to maintain high-precision trajectory approximation capabilities when faced with complex, non-stationary biomedical signals.

[0087] 2.3 Parallel processing of multi-channel signals, performing the integration process independently for each channel while retaining the correlation information between channels.

[0088] Step 3: Parameter Optimization

[0089] 3.1 Local Domain Normalization (LDN) Pre-training:

[0090] In the tight area Inside( )sampling One training point;

[0091] Construct the objective function; ;

[0092] Using the Adam optimizer (initial learning rate 2×10⁻⁶) −3 -5×10 −5 Iterate 300-600 times;

[0093] Save pre-training parameters .

[0094] 3.2 Formal Training:

[0095] Define the fusion loss function: ;

[0096] Adjust the balance coefficient according to the signal type. (EEG signal set to 0.92, ECG signal set to 0.88);

[0097] Backpropagation is achieved by calculating the fractional gradient using the ABM solver.

[0098] Update parameters using batch processing (batch size 32-128);

[0099] Implement an early stopping mechanism: stop training when the loss decreases by less than 0.001 for 10 consecutive rounds or the total number of iterations reaches 600.

[0100] Step 4: Calculation of Abnormal Deviations

[0101] 4.1 Calculate the deviation vector: ,in The original signal, Output for the model.

[0102] 4.2 Implementing an adaptive threshold mechanism:

[0103] Initialize threshold ;

[0104] Threshold update equation: ;

[0105] The threshold is updated using the ABM solver, with an update frequency set to once every 10-50 sampling points;

[0106] Applying moving average filtering: .

[0107] 4.3 Anomaly Detection: If If the time interval is t, then time t is marked as abnormal; otherwise, it is normal.

[0108] Step 5: Predict the output

[0109] 5.1 Based on the current hidden state Use a trained model to predict the future step( The signal value of ).

[0110] 5.2 Calculate the probability of anomalies: ,in This is the sigmoid function.

[0111] 5.3 Generate structured output:

[0112] Exception label: A binary flag (0 / 1) indicates whether an exception has occurred;

[0113] Anomaly probability: A continuous value between 0 and 1, representing the likelihood of an anomaly;

[0114] Abnormal type: classified according to deviation characteristics (such as epileptic seizures, arrhythmias, etc.);

[0115] Prediction time window: Identifies the time range within which an anomaly is expected to occur.

[0116] 5.4 Visualization Results: Generates time-frequency graphs, deviation curves, and predicted probability curves to assist in clinical diagnosis.

[0117] III. Examples

[0118] To verify the effectiveness of the present invention, two specific embodiments are provided below:

[0119] Example 1: Epilepsy Seizure Prediction

[0120] Dataset: The publicly available CHB-MITEEG database contains electroencephalogram (EEG) recordings from 23 patients with epilepsy.

[0121] Configuration parameters: , , , The sequence length T = 1000;

[0122] Preprocessing: 23-channel EEG signal, sampling rate 256Hz, normalized to [-1,1];

[0123] Training: LDN pre-training with 1000 points, formal training with 500 rounds, batch size 64;

[0124] Results: A 30-second warning system was developed to detect epileptic seizures, with a sensitivity of 94.2% and a specificity of 91.8%.

[0125] Example 2: Cardiac Arrhythmia Detection

[0126] Dataset: The dataset uses the MIT-BIH ECG database and contains 48 ECG records;

[0127] Configuration parameters: , , , Sequence length ;

[0128] Preprocessing: 2-channel ECG signal (MLII and V1), sampling rate 360Hz, normalized;

[0129] Training: LDN pre-training with 1200 points, formal training with 400 rounds, batch size 32;

[0130] Results: The accuracy rate of arrhythmia classification was 96.5%, and the F1 value for atrial fibrillation detection was 0.94.

[0131] IV. System Deployment and Implementation

[0132] This invention can be deployed and implemented on multiple platforms:

[0133] 1. Clinical Monitoring System: Integrated into the hospital's central monitoring station, this system processes data from multiple beds in real time, providing abnormal alerts and historical data analysis. The system employs a distributed architecture, with the main server handling model inference and edge devices responsible for data acquisition and preprocessing.

[0134] 2. Wearable medical devices: Lightweight deployment in embedded systems (such as ARM Cortex-A series processors), reducing computational complexity through model pruning and quantization techniques, enabling 24-hour continuous monitoring, and battery life of up to 48 hours or more.

[0135] 3. Mobile Health Applications: Through API interfaces, these applications integrate with smartphone apps to enable patients to monitor their health at home. Abnormal events are automatically pushed to doctors, supporting remote diagnosis and intervention.

[0136] Through the above specific embodiments, those skilled in the art can realize the biomedical multimodal signal anomaly detection and prediction method and system described in this invention, achieving the technical effects of improving detection accuracy, enhancing robustness to time-varying interference, and reducing false alarm rate. It should be noted that the above embodiments are merely illustrative examples, and the scope of protection of this invention is not limited thereto. Any modifications, substitutions, or improvements that do not depart from the spirit and scope of this invention should be included within the scope of protection of this invention.

[0137] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0138] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0139] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for detecting and predicting anomalies in biomedical multimodal signals, characterized in that, Includes the following steps: Receive biomedical multimodal signal data and perform normalization preprocessing; A Caputo fractional recurrent neural network model is constructed to represent the biomedical multimodal signals as a time-varying Caputo fractional ordinary differential equation system. The model equations are as follows: ； in, This represents the Caputo fractional derivative operator. It is a fractional parameter with a value range of 0.6-0.

95. for 3D hidden state vector, It is 64-256. It is a time constant with a value range of 0.8-1.

5. This is the weight matrix. For activation function, The input vector is time-varying; The model equations were numerically integrated using the Adams-Bashforth-Moulton solver to obtain the model output. ; Model parameters are optimized through local domain normalization pre-training and fusion loss function; Based on model output With the original signal Calculate the deviation And through an adaptive threshold mechanism Identify abnormal events; Predicting the future based on the current trajectory The probability of an anomaly in the step. Values ​​range from 5 to 20, generating prediction results.

2. The method according to claim 1, characterized in that, The numerical integration process of the Adams-Bashforth-Moulton solver includes a prediction step and a correction step: Prediction step: ； Calibration step: ； in, This is the integration step size, and its value ranges from 0.001 to 0.

01. For the Gamma function, and For memory weighting coefficients, , , .

3. The method according to claim 1 or 2, characterized in that, The fusion loss function is: ； in, The original signal trajectory, Output for model reconstruction For the current parameter, , , For bias terms, For pre-trained parameters, It is a balance coefficient and its value ranges from 0.85 to 0.

95.

4. The method according to claim 1, characterized in that, The adaptive threshold Following Caputo fractional differential equations: ； Wherein, initial value The value ranges from 0.05 to 0.1 and is updated using the Adams-Bashforth-Moulton solver.

5. The method according to claim 1, characterized in that, The biomedical multimodal signals include electroencephalogram (EEG) signals and electrocardiogram (ECG) signals, wherein the number of EEG signal channels is 8-32, the number of ECG signal channels is 1-12, and the signal sampling rate is 250-500Hz.

6. A biomedical multimodal signal anomaly detection and prediction system implementing the method of any one of claims 1-5, characterized in that, include: Data input interface for receiving and preprocessing biomedical multimodal signal data; The model computation module is used to implement the Caputo fractional recurrent neural network model and the Adams-Bashforth-Moulton solver, and to compute the hidden state and output. The optimization module is used to perform local domain normalization pre-training and optimization of the fusion loss function, and to update the model parameters; The anomaly detection module is used to calculate signal deviation and adaptive threshold, mark abnormal events, and predict the probability of future anomalies. The data output interface is used to generate and transmit anomaly detection results and prediction data; The modules are electrically connected via a data bus to form a serial data processing flow: data input interface, model calculation module, optimization module, anomaly detection module, and data output interface.

7. The system according to claim 6, characterized in that, The model calculation module includes an integral calculation submodule, which implements the prediction-correction logic of the Adams-Bashforth-Moulton solver and receives input signals. and time-varying interference Output model reconstruction signal .

8. The system according to claim 6, characterized in that, The optimization module includes a gradient calculation submodule and a loss function submodule. The loss function submodule is placed before the gradient calculation submodule and is used to calculate the fusion loss function and backpropagate the gradient to update the model parameters.

9. The system according to claim 6, characterized in that, The anomaly detection module includes a deviation comparison submodule and a threshold update submodule. The threshold update submodule runs in parallel with the deviation comparison submodule and is used to implement the fractional threshold equation and update the adaptive threshold.

10. The system according to claim 6, characterized in that, The system is deployed on a general-purpose computer device and includes a processor, memory, and input / output devices. The processor is a CPU or GPU and is used to perform model calculations and parameter optimization. The memory is used to store model parameters, training data, and program code.

Images