Blood flow pipeline random excitation control method based on physical information neural network

Abstract

The invention relates to the technical field of medical fluid conveying pipeline dynamics and artificial intelligence, and discloses a blood flow pipeline random excitation control method based on a physical information neural network, and the method comprises the steps: constructing a PINN model which comprises a data collection module, a physical information neural network prediction module and a control decision module; a sensing-prediction-regulation closed loop is constructed, a hemodynamic control equation and a pipeline vibration equation are fused in a PINN loss function, dual-drive prediction of data and physics is achieved, the generalization ability of a model is improved, heart rate fluctuation and vasoconstriction signals are processed through wavelet transform, random excitation feature vectors matched with a blood flow scene are constructed, and therefore the heart rate fluctuation and vasoconstriction signals are processed. The problem that excitation types of a traditional model are not matched is solved, blood flow auxiliary equipment is driven to achieve two-stage intelligent response according to a PINN prediction result, and closed-loop safety control for random excitation is formed.

Description

Technical Field

[0001] This invention relates to the fields of medical fluid delivery pipeline dynamics and artificial intelligence technology, specifically a stochastic excitation control method for blood flow pipelines based on physical information neural networks. Background Technology

[0002] In interventional medical treatment and organ transplantation, the dynamic stability of blood flow conduits (such as artificial blood vessels and interventional catheters) is crucial to the safety of treatment. However, the blood flow environment is generally characterized by "random excitations" generated by physiological activities such as heart rate fluctuations and vasospasm, which can easily lead to excessive vibration of the conduit or a sudden increase in local pressure, resulting in serious complications such as leakage at the vascular anastomosis and catheter displacement. Currently, the industry's dynamic analysis of such problems largely draws on industrial flow pipeline technology. However, industrial models are mainly designed for "deterministic excitations" (such as fluid impacts at a fixed frequency), which differ significantly from the characteristics of "random excitations" in blood flow. Directly applying these models can lead to large prediction errors and fail to meet the high accuracy requirements of clinical practice.

[0003] Meanwhile, although existing neural network models have been applied to pipeline parameter inversion, they generally lack the embedding of specific physical laws governing blood flow. The non-Newtonian fluid properties of blood (viscosity varying with shear rate) and the elastic deformation laws of blood vessel walls are not incorporated into model training, resulting in models being unable to reasonably explain complex clinical phenomena such as "sudden changes in pipeline pressure when blood flow velocity increases sharply." Furthermore, when predicted parameters continue to deteriorate and exceed high-risk thresholds, the system will automatically send closed-loop control commands to blood flow assist devices (such as infusion pumps and extracorporeal circulation machines), reducing the infusion pump flow rate by 10%–20% without human intervention. Medical scenarios urgently need to achieve closed-loop linkage between risk prediction and equipment control; that is, when abnormal pressure is detected, the system should be able to automatically adjust the parameters of blood flow assist devices. However, existing technologies mostly focus only on monitoring and early warning, lacking closed-loop control logic for blood flow pipelines, resulting in delayed clinical emergency response and an inability to effectively prevent risks.

[0004] Therefore, there is an urgent need for an innovative method that can accurately characterize the random excitation characteristics of blood flow, deeply integrate the physical laws of hemodynamics, and realize the prediction-control closed-loop linkage, so as to break through the bottlenecks of existing technologies in terms of accuracy, generalization and response efficiency, and meet the high standards of clinical requirements for the safe management of blood flow channels. Summary of the Invention

[0005] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides a stochastic excitation control method for blood flow channels based on physical information neural networks. This method has the advantages of high prediction accuracy, fast response speed, and strong generalization ability, effectively solving the problems of inaccurate prediction and untimely control of blood flow channel instability caused by stochastic excitation.

[0006] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: a stochastic excitation control method for blood flow channels based on physical information neural networks, comprising the following steps: Step 1: Data Input Design: Design a data acquisition module to collect clinical measured data and finite element simulation data, merge the data to generate sample groups, and preprocess the data by wavelet transform denoising to output a standardized sample set; Step 2, Model Building Stage: Based on the preprocessed data, the PINN model is built, and the network structure of the input layer, hidden layer and output layer is designed. The Navier-Stokes equation and Euler-Bernoulli beam equation are explicitly embedded in the network design as core physical constraints. The model has a special structural optimization feedback mechanism: when the verification shows that the physical consistency does not meet the standard, the reconfiguration and re-optimization of the network structure will be triggered. Step 3, Model Training and Validation: The PINN model is iteratively trained using the gradient descent algorithm. The network parameters are optimized by minimizing the loss function until the displacement prediction error is <5% and the pressure prediction error is <3%. The model prediction results are validated using clinical experimental data. Its accuracy and generalization ability are comprehensively evaluated. If the validation fails, return to Step 2 for re-optimization. Step 4: System Deployment and Module Initialization: Deploy the validated PINN model to the medical edge computing terminal and integrate it into the surgical console or postoperative monitoring system; deploy the data acquisition module at the clinical monitoring front end, and deploy the physical information neural network prediction module and control decision module at the edge computing terminal, completing the initial configuration of each module; Step 5: Real-time data acquisition and prediction: The data acquisition module collects blood flow parameters, duct structure parameters and random excitation characteristics in real time, and transmits them to the physical information neural network prediction module, which outputs the prediction results of vibration displacement and local pressure for the next 5 to 10 seconds. Step Six: Control Decision and Closed-Loop Execution: After receiving the prediction results, the control decision module triggers a graded early warning mechanism when the safety threshold is exceeded, driving the extracorporeal circulation machine to adjust the pressure compensation value or the infusion pump to adjust the flow rate, thereby achieving closed-loop control of blood flow parameters.

[0007] Preferably, in step one, the data acquisition module collects real-time blood flow data through medical sensors, including blood flow velocity, blood viscosity, elastic modulus of artificial blood vessels, and heart rate fluctuation frequency.

[0008] Preferably, in step one, the data preprocessing involves denoising the acquired random excitation signal and extracting the excitation feature vector.

[0009] Preferably, the input layer in step two contains 7 to 8 neurons, which correspond to parameters such as blood flow velocity, blood viscosity, real-time blood pressure, blood flow channel diameter, vascular elastic modulus, heart rate fluctuation frequency, vascular contraction amplitude, and blood flow pulsation cycle.

[0010] Preferably, in step two, the hidden layer contains 3-4 fully connected layers, each with 30-32 neurons, and the activation function is ReLU.

[0011] Preferably, in step two, physical constraints are embedded as follows: a hemodynamic control equation, i.e., a simplified form of the Navier-Stokes equation, is added to the fluid domain loss function; considering the elasticity of blood vessels, the Euler-Bernoulli beam equation is used to constrain the vibration of the vessel wall in the solid domain loss function. The combination of the two function formulas makes the model training process follow physical laws, and the simplified expression of the total function is: total loss function = data fitting loss + physical law loss.

[0012] Preferably, the fluid domain loss function uses the Navier-Stokes equations to constrain the fluid dynamics behavior, and its calculation formula is as follows: ; In the formula, This represents the total loss function in the fluid domain. This represents the data loss term based on the velocity field, reflecting the state of blood flow. This represents the loss term in the momentum equation. This represents the loss term in the continuity equation. Represents the boundary condition loss term. This represents the data loss weighting coefficient, which is adjusted based on the reliability and density of the measurement data. When the data quality is high, a larger weight is used to enhance the data fit. The physical equation loss weighting coefficients are adjusted based on the importance of the Navier-Stokes equations and are used to balance the contribution of physical constraints to the total loss. The boundary condition loss weight coefficient is used to adjust the penalty intensity of the boundary condition terms. When the boundary conditions are accurate, a large weight is used to force the boundary constraints to be met.

[0013] Preferably, the Euler-Bernoulli beam equation is used to constrain the pipe wall vibration in the solid domain loss function, and its calculation formula is as follows: ; In the formula, This represents the total loss function in the solid domain, used to measure the degree of agreement between the pipe wall vibration predicted by the neural network and the physical laws. This represents the data loss term based on the displacement field, reflecting the deformation state of the pipe wall. This represents the residual loss of the physical equations, penalizing the degree of violation of the Euler-Bernoulli beam equations. This represents the initial condition loss weighting coefficient, used to adjust the contribution of initial displacement and initial velocity conditions to the total loss in the solid domain. This represents the initial condition loss.

[0014] Preferably, in step five, the physical information neural network prediction module receives the collected data in real time and outputs the pipeline vibration displacement and local pressure within the next 5 to 10 seconds.

[0015] Preferably, in step six, the control decision module sets up a two-level response mechanism, including: (1) Level 1 warning: When the predicted parameter reaches the primary warning threshold, the system automatically sends a visual and audio-visual prompt signal to the medical terminal and clearly marks the abnormal parameter item, prompting medical staff to immediately conduct manual review and clinical assessment; (2) Level 2 warning: When the predicted parameters continue to deteriorate and exceed the high-risk threshold, the system will automatically send a closed-loop control command to the blood flow assist device, reduce the infusion pump flow rate by 10% to 20% under unattended conditions, or dynamically adjust the pressure compensation parameters of the extracorporeal circulation machine until the predicted value falls back to the preset safe range.

[0016] Compared with existing technologies, this invention provides a stochastic excitation control method for blood flow channels based on physical information neural networks, which has the following beneficial effects: 1. This invention embeds the Navier-Stokes equation and the Euler-Bernoulli beam equation as physical constraints into the neural network loss function, achieving a dual drive from data-driven and physical laws. This results in improved prediction accuracy and model generalization ability, ensuring that even in regions with sparse clinical data, the model's prediction results strictly conform to the basic laws of hemodynamics and structural mechanics. This effectively avoids abnormal predictions that violate physical common sense, such as "increased flow velocity but decreased pressure." Ultimately, the displacement and pressure prediction errors are reduced to within 5% and 3%, respectively.

[0017] 2. This invention constructs a closed-loop prediction and control system that links the PINN prediction module and the control decision module, and designs a graded early warning mechanism to achieve the beneficial effects of significantly shortening emergency response time and realizing proactive risk prevention and control. When the predicted parameters continue to deteriorate and exceed the high-risk threshold, the system will automatically send closed-loop control commands to blood flow assist devices (such as infusion pumps and extracorporeal circulation machines) to reduce the infusion pump flow rate by 10% to 20% without human intervention. This method realizes the upgrade from a delayed response that relies on manual interpretation and intervention to an integrated intelligent closed loop of perception, decision-making, and execution, thereby taking intervention measures before or in the early stage of the risk of pipeline vibration or pressure exceeding the threshold, and ultimately effectively avoiding complications such as vascular leakage and catheter displacement.

[0018] 3. This invention uses wavelet transform to denoise and extract features from signals such as heart rate fluctuations and vasoconstriction, constructing a feature vector that accurately describes the random excitation of blood flow as model input. Ultimately, this invention fundamentally solves the problem of excitation type mismatch in traditional industrial models, enabling the model to accurately learn and respond to the random dynamic characteristics of blood flow. It overcomes the huge prediction bias caused by applying industrial deterministic excitation models, providing a reliable analysis tool that is truly applicable to the physiological environment for clinical use. Attached Figure Description

[0019] Figure 1 This is a diagram illustrating the steps of the method of the present invention; Figure 2 This is a flowchart of the module execution of the present invention. Detailed Implementation

[0020] 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 embodiments of the present invention, and not all embodiments. 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.

[0021] Please see Figure 1 - Figure 2 A stochastic excitation control method for blood flow channels based on physical information neural networks includes the following steps: Step 1: Data Input Design: Design a data acquisition module to collect 100 clinical experimental data and finite element simulation data, merge the data to generate 5000 sets of samples, and use wavelet transform for noise reduction preprocessing to output a standardized sample set; Step 2, Model Building Stage: Based on the preprocessed data, the PINN model is built, and the network structure of the input layer (8 neurons), hidden layer (3×32 neurons), and output layer (2 neurons) is designed. The Navier-Stokes equation and Euler-Bernoulli beam equation are explicitly embedded in the network design as core physical constraints. The model has a special structural optimization feedback mechanism: when the verification shows that the physical consistency does not meet the standard, the reconfiguration and re-optimization of the network structure will be triggered, thereby ensuring the reliability of the solution and improving the convergence efficiency. Step 3, Model Training and Validation: The PINN model is iteratively trained using the gradient descent algorithm. The network parameters are optimized by minimizing the loss function until the displacement prediction error is less than 5% and the pressure prediction error is less than 3%. The model prediction results are validated using clinical experimental data. Its accuracy and generalization ability are comprehensively evaluated. If the validation fails, return to Step 2 for re-optimization. Step 4: System Deployment and Module Initialization: Deploy the validated PINN model to the medical edge computing terminal and integrate it into the surgical console or postoperative monitoring system; deploy the data acquisition module at the clinical monitoring front end, and deploy the physical information neural network prediction module and control decision module at the edge computing terminal, completing the initial configuration of each module; Step 5: Real-time data acquisition and prediction: The data acquisition module collects blood flow parameters (flow velocity, pressure, viscosity), duct structure parameters (inner diameter, elastic modulus) and random excitation characteristics (heart rate fluctuation frequency, vasoconstriction amplitude) in real time, and transmits them to the physical information neural network prediction module, which outputs the prediction results of vibration displacement and local pressure for the next 5 to 10 seconds. Step Six: Control Decision and Closed-Loop Execution: After receiving the prediction results, the control decision module will trigger a graded early warning mechanism when the safety thresholds (vibration ≤ 0.2 mm, pressure ≤ 120 mmHg) are exceeded. The first-level early warning sends a prompt message to the medical terminal, and the second-level early warning sends a control command to the blood flow assist device to drive the extracorporeal circulation machine to adjust the pressure compensation value or the infusion pump to adjust the flow rate, thereby realizing closed-loop control of blood flow parameters.

[0022] This method constructs a PINN model, which comprises three parts: a data acquisition module, a physical information neural network prediction module, and a control decision module. The data acquisition module collects blood flow parameters, pipeline structure parameters, and random excitation characteristics in real time, and transmits them to the physical information neural network prediction module after preprocessing. The physical information neural network prediction module outputs the predicted results of pipeline vibration displacement and local pressure. When a safety threshold is exceeded, a control signal is triggered and transmitted to the control decision module. The control decision module connects to blood flow assist devices (such as infusion pumps or extracorporeal circulation machines) and sends parameter adjustment commands (such as reducing flow rate or adjusting pressure compensation values). The three modules work together to form a complete technical closed loop from real-time perception and accurate prediction to intelligent control. By integrating the hemodynamic control equation and the pipeline vibration equation into the PINN loss function, dual-driven prediction based on data and physics is achieved, improving the model's generalization ability. Wavelet transform is used to process heart rate fluctuations and vasoconstriction signals to construct a random excitation feature vector adapted to the blood flow scenario, solving the problem of excitation type mismatch in traditional models. By linking the PINN prediction results with blood flow auxiliary equipment, a two-level response mechanism is designed to achieve real-time safety control under random excitation, breaking through the limitation of existing technologies that "emphasize prediction but neglect control".

[0023] Specifically, in step one, the data acquisition module collects real-time blood flow data through medical sensors (such as intravascular pressure sensors and ultrasonic flow meters), including blood flow velocity (0.5-1.5 m / s), blood viscosity (3.5-5.5 mPa·s), elastic modulus of artificial blood vessels (10-30 MPa), and heart rate fluctuation frequency (0.8-1.2 Hz).

[0024] Specifically, in step one, data preprocessing involves denoising the acquired random excitation signal (using wavelet transform algorithm) and extracting the excitation feature vector (frequency distribution, maximum amplitude) as model input.

[0025] By combining wavelet transform denoising and excitation feature extraction in step one above, high-frequency noise interference in clinical monitoring can be effectively filtered out while retaining the key dynamic features of random excitation. This significantly improves the quality and signal-to-noise ratio of the input data, thereby enhancing the stability and convergence speed of model training.

[0026] Specifically, in step two, the input layer contains eight neurons, which correspond to eight key parameters: blood flow velocity, blood viscosity, real-time blood pressure, blood flow channel diameter, vascular elastic modulus, heart rate fluctuation frequency, vascular contraction amplitude, and blood flow pulsation cycle.

[0027] Specifically, in step two, the hidden layer contains 3-4 (4 layers) fully connected layers, each with 30-32 (32) neurons, and the activation function is ReLU.

[0028] By combining the multi-parameter input layer and the deep fully connected hidden layer in step two above, the 8-dimensional physical parameters achieve comprehensive feature representation, and the 4-layer × 32-neuron structure provides sufficient nonlinear fitting ability, ultimately achieving the beneficial effect of improving the model's accuracy in mapping complex hemodynamics and its anti-overfitting ability.

[0029] Step two involves embedding physical constraints: Hemodynamic control equations (simplified forms of the Navier-Stokes equations, adapted to the non-Newtonian characteristics of blood flow) are added to the fluid domain loss function (describing blood flow motion); considering vascular elasticity, the Euler-Bernoulli beam equations are used to constrain pipe wall vibration in the solid domain loss function (describing pipe structure response). The combination of these two function formulas ensures that the model training process follows physical laws. The simplified expression of the total function is: Total Loss Function = Data Fitting Loss + Physical Law Loss. The data fitting loss ensures that the model's predictions match limited experimental or clinical measurement data, while the physical law loss ensures that the model's predictions comply with the fundamental physical laws of hemodynamics (Navier-Stokes equations) and the fundamental physical laws of pipe vibration (Euler-Bernoulli beam equations) everywhere (including areas without measurement data).

[0030] (1) Fluid domain loss function (describing blood flow motion), using the Navier-Stokes equations to constrain fluid dynamics behavior, is calculated as follows: ; In the formula, This represents the total loss function in the fluid domain, used to measure the degree of agreement between the neural network's predictions and physical laws; the data loss term... Based on the velocity field (vector), reflecting the state of blood flow, , It is the velocity field (vector) predicted by the neural network. It is the measured velocity field (vector). It is the number of measurement points. These are the spatial coordinates and time of the measurement point; the loss term in the momentum equation. = , ρ is blood density, u is the velocity field (predicted by a neural network), and p is the pressure field (predicted by a neural network). It is the deviatoric stress tensor, which is related to the strain rate tensor; f is the body force. It is the number of sampling points within the computational domain; the loss term in the continuity equation. Boundary condition loss term , It is the number of boundary condition points. It is the velocity given at the boundary point; This represents the data loss weighting coefficient, which is adjusted based on the reliability and density of the measurement data. When the data quality is high, a larger weight is used to enhance the data fit. The physical equation loss weighting coefficients are adjusted based on the importance of the Navier-Stokes equations and are used to balance the contribution of physical constraints to the total loss. The boundary condition loss weight coefficient is used to adjust the penalty intensity of the boundary condition terms. When the boundary conditions are accurate, a large weight is used to force the boundary constraints to be met.

[0031] By employing the Navier-Stokes equations to constrain fluid dynamics, and introducing a dual physical mechanism of momentum and mass conservation, solutions that conform to hemodynamic laws are automatically generated in sparse data regions. This effectively suppresses prediction jumps caused by measurement noise, thereby improving the model's extrapolation ability and robustness.

[0032] (2) The Euler-Bernoulli beam equation is used to constrain pipe wall vibration in the solid domain loss function (describing the pipe structure response). The calculation formula is as follows:

[0033] In the formula, This represents the total loss function in the solid domain, used to measure the degree of agreement between the pipe wall vibration predicted by the neural network and the physical laws. The data loss term is also included. Based on the displacement field (scalar or vector), it reflects the deformation state of the pipe wall. There are N sensors measuring data. It refers to the spatial coordinates and time of the measurement point. These are measured values. It is the displacement predicted by the neural network; M points are randomly selected within the computational domain. Calculate the residuals of the Euler-Bernoulli beam equation = The physical loss is ; Boundary condition loss ; , , These are hyperparameters used to balance various losses; is the density of the beam; A is the cross-sectional area of ​​the beam; c is the damping coefficient; f(x,t) is the distributed external force. This represents the initial condition loss weighting coefficient, used to adjust the contribution of initial displacement and initial velocity conditions to the total loss in the solid domain. It is dynamically adjusted based on the reliability and importance of the initial conditions; initial condition loss. Representing the initial displacement and initial velocity conditions ; By using the Euler-Bernoulli beam equation to constrain the pipe wall vibration, an explicit mechanical relationship between fluid pressure load and structural displacement response is established. Under limited measurement point conditions, the higher-order modal vibration characteristics of the blood vessel wall are accurately captured. Furthermore, by explicitly constraining the initial conditions, the divergence problem of transient prediction is avoided, ultimately achieving the beneficial effect of improving the accuracy and stability of fluid-structure interaction prediction.

[0034] The training data mentioned above are all derived from clinical experimental data (dynamic data of blood flow channels from 100 patients) and finite element simulation data. They were obtained by constructing 5000 training samples. The training objective is to ensure that the vibration displacement error predicted by the model is less than 5% and the pressure error is less than 3%. The pressure field output by the fluid domain is applied as a distributed external force to the beam equation of the solid domain. The pipe wall displacement predicted by the solid domain reverses the calculation boundary of the fluid domain. The two domains achieve bidirectional information transmission by sharing the boundary condition loss term at the coupling interface.

[0035] Specifically, in step five, the physical information neural network prediction module receives the collected data in real time and outputs the pipeline vibration displacement (safety threshold: ≤0.2mm) and local pressure (safety threshold: ≤120mmHg) within the next 5 to 10 seconds.

[0036] By combining multi-source heterogeneous data fusion with physical constraint prediction in step five above, the low-latency architecture based on edge computing ensures the real-time performance of the prediction, and the dual physical field coupling mechanism guarantees the physiological rationality of the prediction results. This achieves the beneficial effect of providing early warning of abnormal blood flow during surgery 5 to 10 seconds in advance and reducing the risk of delayed clinical intervention.

[0037] Specifically, in step six, the control decision module sets up a two-level response mechanism, including: Level 1 warning (predicted value close to 80% of the threshold): When the predicted parameter reaches the primary warning threshold, the system automatically sends a visual and audio-visual prompt signal to the medical staff terminal, and clearly marks the abnormal parameter item, prompting medical staff to immediately conduct manual review and clinical assessment; Level 2 warning (predicted value exceeds threshold): Automatically sends instructions to blood flow assist devices, such as reducing the infusion pump flow rate by 10% to 20%, or adjusting the pressure compensation value of the extracorporeal circulation machine, until the predicted value falls back to a safe range; By combining the graded early warning mechanism with automated closed-loop control in step six above, the feedforward early warning with an 80% threshold retains a window for manual intervention by medical staff, and the automatic adjustment of the secondary early warning can avoid human response delay. The response time of both levels is controlled within 5 seconds, achieving the beneficial effect of minimizing blood flow fluctuation complications while ensuring medical safety.

[0038] By combining "data-driven" and "physical laws" through the PINN model, we can improve prediction accuracy by using clinical data and avoid "abnormal results that do not conform to hemodynamics" through physical constraints (such as predicting "increased flow rate but decreased pressure"). At the same time, we form a closed-loop design for prediction and control to shorten emergency response time from the traditional "manual discovery - manual adjustment" (30-60 seconds) to "automatic prediction - automatic control" (less than 5 seconds).

[0039] Example Using the method of this invention, the data acquisition module connects an intravascular pressure sensor (model: St. Jude Medical PressureWire X) and an ultrasonic flow meter (model: Philips EPIQ 7), and deploys the PINN model on a medical edge computing terminal (computing power: 20 TOPS), controlling the extracorporeal circulation machine (model: Terumo System 1). During implementation, the postoperative patient's heart rate fluctuated at a frequency of 1.0–1.1 Hz, blood viscosity was 4.2 mPa·s, and the elastic modulus of the artificial blood vessel was 25 MPa. The system collects data in real time, and the PINN model predicts that the local pressure in the pipeline will rise to 125 mmHg within the next 5 seconds (exceeding the safe threshold of 120 mmHg), triggering a level-two warning. Results: The extracorporeal circulation machine automatically reduced the pressure compensation value by 15%, and the pipeline pressure dropped back to 112 mmHg after 10 seconds. No leakage occurred at the vascular anastomosis site. The patient recovered well after the operation and there were no related complications.

[0040] In summary, compared with traditional industrial pipeline models, the method of this invention reduces the vibration displacement prediction error from over 30% to within 5% and the pressure prediction error from over 25% to within 3%, meeting the high-precision requirements of medical scenarios. The "prediction-control" closed loop shortens the emergency response time to within 5 seconds, effectively avoiding complications such as vascular leakage and catheter displacement caused by pipeline vibration and sudden pressure increases, reducing clinical risks by more than 60%. The PINN model adapts to the blood flow parameters (blood pressure, blood viscosity differences) of different patients through physical constraints, eliminating the need to retrain the model for each patient and reducing clinical application costs by 40%.

[0041] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A blood flow conduit random excitation control method based on a physical information neural network, characterized by, The method comprises the following steps: Step one, data input design: design a data collection module to collect clinical measured data and finite element simulation data, merge the data to generate a sample set, and use wavelet transform denoising for preprocessing to output a standardized sample set; Step two, model building stage: build a PINN model based on the preprocessed data, design the network structure of the input layer, hidden layer and output layer, explicitly embed the Navier-Stokes equation and Euler-Bernoulli beam equation as the core physical constraints in the network design, and the model is specially provided with a structure optimization feedback mechanism: when the verification shows that the physical consistency does not meet the standard, the network structure will be reconfigured and optimized again; Step three, model training and verification: use the gradient descent algorithm to iteratively train the PINN model, optimize the network parameters by minimizing the loss function, until the displacement prediction error is less than 5% and the pressure prediction error is less than 3%, verify the model prediction results with clinical measured data, and comprehensively evaluate the accuracy and generalization ability of the model. If the verification fails, return to step two for re-optimization; Step four, system deployment and module initialization: deploy the verified PINN model to a medical edge computing terminal, integrate it into a surgical control console or postoperative monitoring system; deploy the data collection module to the clinical monitoring front end, deploy the physical information neural network prediction module and control decision module to the edge computing terminal, and complete the initialization configuration of each module; Step five, real-time data collection and prediction: the data collection module collects blood flow parameters, pipe structure parameters and random excitation characteristics in real time, and transmits them to the physical information neural network prediction module, which outputs the future 5-10 second prediction results of vibration displacement and local pressure; Step six, control decision and closed-loop execution: after receiving the prediction results, the control decision module triggers a hierarchical warning mechanism when the safety threshold is exceeded, drives the extracorporeal circulation machine to adjust the pressure compensation value or the infusion pump to adjust the flow rate, and realizes closed-loop control of blood flow parameters.

2. The physical information neural network-based blood flow conduit stochastic excitation control method of claim 1, wherein: The data collection module in step one collects real-time blood flow pipe data through medical sensors, including blood flow velocity, blood viscosity, artificial blood vessel elastic modulus, and heart rate fluctuation frequency.

3. The physical information neural network-based blood flow conduit stochastic excitation control method of claim 1, wherein: The data preprocessing in step one: denoising processing is performed on the collected random excitation signal to extract the excitation feature vector.

4. The physical information neural network-based blood flow conduit stochastic excitation control method of claim 1, wherein: The input layer in step two contains 7-8 neurons corresponding to blood flow velocity, blood viscosity, real-time blood pressure, blood flow pipe diameter, blood vessel elastic modulus, heart rate fluctuation frequency, blood vessel contraction amplitude, and blood flow pulsation cycle parameters.

5. The physical information neural network-based blood flow conduit stochastic excitation control method of claim 1, wherein: The hidden layer in step two is provided with 3-4 fully connected layers, each layer having 30-32 neurons, and the activation function uses ReLU.

6. The physical information neural network-based blood flow conduit stochastic excitation control method of claim 1, wherein: The physical constraint embedding in step two: the blood flow dynamics control equation, i.e. the simplified form of Navier-Stokes equation, is added to the fluid domain loss function; considering the blood vessel elasticity, the Euler-Bernoulli beam equation is used to constrain the pipe wall vibration in the solid domain loss function, and the combination of the two function formulas makes the model training process follow the physical law. The total function simplification formula is: total loss function = data fitting loss + physical law loss.

7. The physical information neural network-based blood flow conduit stochastic excitation control method of claim 6, wherein: The fluid domain loss function adopts Navier-Stokes equation to constrain fluid dynamics behavior, and the calculation formula is as follows: ； In the formula, represents the total loss function of the fluid domain, represents the data loss term based on the velocity field, reflecting the blood flow motion state, represents the momentum equation loss term, represents the continuity equation loss term, represents the boundary condition loss term, represents the data loss weight coefficient, which is adjusted according to the reliability and density of the measured data. When the data quality is high, a large weight is used to strengthen the data fitting, represents the physical equation loss weight coefficient: adjusted according to the importance of the Navier-Stokes equation, used to balance the contribution degree of the physical constraint to the total loss, represents the boundary condition loss weight coefficient: used to adjust the punishment strength of the boundary condition term, when the boundary condition is accurate, a large weight is used to force to meet the boundary constraint.

8. The physical information neural network-based blood flow conduit stochastic excitation control method of claim 6, wherein: The solid domain loss function adopts Euler-Bernoulli beam equation to constrain pipe wall vibration, and the calculation formula is as follows: ； In the formula, represents the total loss function of the solid domain, which is used to measure the consistency of the neural network prediction of the pipe wall vibration with the physical law, represents the data loss term based on the displacement field, which reflects the deformation state of the pipe wall, represents the physical equation residual loss, which punishes the violation degree of the Euler-Bernoulli beam equation, represents the initial condition loss weight coefficient, which is used to adjust the contribution degree of the initial displacement and initial velocity conditions to the total loss of the solid domain, represents the initial condition loss.

9. The physical information neural network-based blood flow conduit stochastic excitation control method of claim 1, wherein: The physical information neural network prediction module in the fifth step receives the collected data in real time and outputs the pipe vibration displacement and local pressure in the next 5-10 seconds.

10. The physical information neural network-based blood flow conduit stochastic excitation control method of claim 1, wherein: The control decision module in the sixth step sets a two-stage response mechanism, including: (1) First warning: when the predicted parameter reaches the primary warning threshold, the system automatically sends visual and audible prompt signals to the medical terminal and clearly marks the abnormal parameter item, prompting medical personnel to immediately perform manual review and clinical evaluation; (2) Second warning: when the predicted parameter continues to deteriorate and breaks through the high-risk threshold, the system will automatically send a closed-loop control instruction to the blood flow auxiliary equipment, reducing the infusion pump flow rate by 10%-20% without human intervention, or dynamically adjusting the extracorporeal circulation machine pressure compensation parameters until the predicted value falls within the preset safety interval.

Images