Control method of a ventricular assist device, readable storage medium, and ventricular assist device

Abstract

The application provides a control method of a ventricular assist device, a readable storage medium and the ventricular assist device, the control method of the ventricular assist device comprises the following steps: obtaining voltage driving force excitation data and driving force motion trajectory excitation data; establishing an implicit dynamics model according to the voltage driving force excitation data and the driving force motion trajectory excitation data; and predicting the state of the ventricular assist device based on the implicit dynamics model, and updating the parameters of a control network according to the predicted state to obtain the control rate of the ventricular assist device. In this way, the implicit dynamics model is established by using two-stage excitation data (referring to the voltage driving force excitation data at the front end and the driving force motion trajectory excitation data at the rear end), thereby effectively improving the accuracy of the dynamics model. Furthermore, the parameters of the control network are updated in the mode of predicting the state of the ventricular assist device by using the implicit dynamics model, thereby simplifying the solving process of the control rate.

Description

Technical Field

[0001] This invention relates to the field of medical device technology, and in particular to a control method, a readable storage medium, and a ventricular assist device. Background Technology

[0002] A ventricular assist device is a device that helps the heart work, or temporarily replaces the heart's function when it is unable to work, to assist or replace the heart in pumping blood to other parts of the body.

[0003] The use of magnetically levitated rotors in ventricular assist devices allows rotation without friction or wear, thereby reducing blood stagnation, turbulence, or hemolysis, and reducing or avoiding mechanical failures.

[0004] Because the dynamic equations of a magnetically levitated rotor differ from those of a traditional mechanical bearing, current technologies derive the control law of the magnetically levitated rotor by establishing a dynamic model and performing mathematical derivation. However, due to the dynamic coupling between the x-axis and y-axis of the magnetically levitated rotor, complete decoupling is not possible. Decoupling the voltage inputs of the four coils along the x and y axes (meaning the x-axis coil only outputs control force in the x-direction, and the y-axis coil only outputs control force in the y-direction) and linearization (meaning the current has a linear relationship with the position, velocity, and acceleration of the magnetically levitated rotor) can easily distort the dynamic model, making it difficult to solve for complex control objectives (zero displacement control of the magnetically levitated rotor under rotation and impact) and to obtain a stable control law under complex operating conditions. Summary of the Invention

[0005] The purpose of this invention is to provide a control method, a readable storage medium, and a ventricular assist device to solve the problems of low accuracy of the dynamic model and complex calculation of the control rate in existing ventricular assist devices.

[0006] To solve the above-mentioned technical problems, the present invention provides a control method for a ventricular assist device, comprising:

[0007] Acquire voltage-driven force excitation data and driving force motion trajectory excitation data;

[0008] An implicit dynamic model is established based on the voltage driving force excitation data and the driving force motion trajectory excitation data; and

[0009] The state of the ventricular assist device is predicted based on the implicit dynamics model, and the parameters of the control network are updated according to the predicted state to obtain the control rate of the ventricular assist device.

[0010] Optionally, the step of establishing an implicit dynamic model based on the voltage driving force excitation data and the driving force motion trajectory excitation data includes:

[0011] The encoding part of the differential neural network is trained using voltage-driven force excitation data;

[0012] The decoding part of the differential neural network is trained using excitation data of the driving force motion trajectory; and

[0013] The implicit dynamic model is obtained by aligning the encoded part after training with the decoded part after training.

[0014] Optionally, the encoding part includes encoding the voltage to drive force; the decoding part includes decoding the drive force to dynamic parameters; the step of aligning the trained encoding part and the trained decoding part includes aligning the drive force of the encoding part with the drive force of the decoding part.

[0015] Optionally, the steps for training a differential neural network include:

[0016] Gradient calculation based on augmented state with adjoint sensitivity.

[0017] Optionally, the method for acquiring voltage driving force excitation data and driving force motion trajectory excitation data is based on an excitation strategy that minimizes the condition number.

[0018] Optionally, the incentive strategy based on minimizing the condition number includes:

[0019] Initialize the dynamic excitation trajectory using a finite-term Fourier series;

[0020] Based on the dynamic excitation trajectory, the excitation strategy is obtained through iterative optimization.

[0021] Optionally, the steps of updating the control network parameters based on the predicted state include:

[0022] Based on the predicted state, the parameters of the control network are updated using a maximum entropy reinforcement learning algorithm.

[0023] Optionally, the steps for updating the parameters of the control network based on the maximum entropy reinforcement learning algorithm include:

[0024] Obtain the current status of the ventricular assist device;

[0025] The implicit dynamics model is used to predict the multi-step states of the ventricular assist device and calculate the multi-step reward.

[0026] Execute the action based on the voltage output by the current action network;

[0027] The root mean square loss of the multi-step reward and evaluation network output is used to perform gradient backpropagation, combined with the entropy regularization loss of the action state, to update the parameters of the control network.

[0028] To address the aforementioned technical problems, the present invention also provides a readable storage medium having a program stored thereon, which, when executed, implements the steps of the control method for the ventricular assist device as described above.

[0029] To address the aforementioned technical problems, the present invention also provides a ventricular assist device, which includes a control module and a magnetically levitated rotor. The control module drives the magnetically levitated rotor according to the control method for the ventricular assist device described above.

[0030] To address the aforementioned technical problems, the present invention also provides a ventricular assist device, comprising a position sensor, a control module, a dynamics calculation module, an electromagnetic drive unit, and a magnetically levitated rotor. The position sensor is used to feed back the current displacement information of the magnetically levitated rotor to the control module. The control module controls the magnetically levitated rotor through the electromagnetic drive unit based on the current displacement information of the magnetically levitated rotor and the control output related to the control rate calculated by the dynamics calculation module based on the implicit dynamics model.

[0031] In summary, in the control method, readable storage medium, and ventricular assist device provided by the present invention, the control method of the ventricular assist device includes: acquiring voltage driving force excitation data and driving force motion trajectory excitation data; establishing an implicit dynamic model based on the voltage driving force excitation data and the driving force motion trajectory excitation data; predicting the state of the ventricular assist device based on the implicit dynamic model, and updating the parameters of the control network according to the predicted state to obtain the control rate of the ventricular assist device.

[0032] This configuration, using two-stage excitation data (voltage-driven force excitation data at the front end and driving force trajectory excitation data at the back end), establishes an implicit dynamic model, effectively improving the accuracy of the dynamic model. Furthermore, the parameters of the control network are updated by predicting the state of the ventricular assist device using the implicit dynamic model, simplifying the process of solving for the control law. Attached Figure Description

[0033] Those skilled in the art will understand that the accompanying drawings are provided to better understand the invention and do not constitute any limitation on the scope of the invention. Wherein:

[0034] Figure 1 This is a functional block diagram of the ventricular assist device according to an embodiment of the present invention;

[0035] Figure 2 This is an overall control flowchart of the control method for the ventricular assist device according to an embodiment of the present invention;

[0036] Figure 3This is a schematic diagram of the generation process of the incentive strategy based on minimizing the number of conditions according to an embodiment of the present invention;

[0037] Figure 4 This is a schematic diagram of the architecture of the implicit dynamics model in an embodiment of the present invention;

[0038] Figure 5 This is a schematic diagram of the process of updating the parameters of the control network based on the maximum entropy reinforcement learning algorithm according to an embodiment of the present invention;

[0039] Figure 6 This is a schematic diagram of the effect frame stack of the magnetic levitation rotor from startup to reaching the geometric center, according to an embodiment of the present invention.

[0040] In the attached image:

[0041] 01-Position sensor; 02-Control module; 03-Dynamics calculation module; 04-Electromagnetic drive unit; 05-Magnetic levitation rotor. Detailed Implementation

[0042] To make the objectives, advantages, and features of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that the drawings are all in a very simplified form and are not drawn to scale, and are only used to facilitate and clarify the explanation of the embodiments of this invention. Furthermore, the structures shown in the drawings are often part of the actual structures. In particular, different figures may emphasize different aspects and may sometimes use different scales.

[0043] As used in this invention, the singular forms “a,” “an,” “one,” and “the” include plural objects; the term “or” is generally used to mean “and / or”; the term “a number” is generally used to mean “at least one”; and the term “at least two” is generally used to mean “two or more”. Furthermore, the terms “first,” “second,” and “third” are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined with “first,” “second,” or “third” may explicitly or implicitly include one or at least two of that feature.

[0044] The purpose of this invention is to provide a control method, a readable storage medium, and a ventricular assist device, to solve the problems of low accuracy of the dynamic model and complex calculation of the control rate in existing ventricular assist devices. The following description refers to the accompanying drawings.

[0045] Please refer to Figure 1The diagram illustrates a functional block diagram of a ventricular assist device. This ventricular assist device could be, for example, a percutaneous ventricular assist device (pVad), a short-term device typically used in emergency situations, such as during and after percutaneous coronary intervention (PCI) in high-risk patients and in patients with cardiogenic shock, for short-term cardiac support. In some embodiments, the pVad is inserted percutaneously into the body via a catheter and connected to the heart, powered by an externally located pump to return blood to the systemic circulation, thereby assisting or taking over the heart's pumping function. The use of a magnetically levitated rotor in the pump increases the blood flow gap, reduces blood trauma and hemolysis, and reduces blood stasis, offering numerous beneficial effects. It should be noted that... Figure 1 The percutaneous ventricular assist device (pVad) shown is merely an example of the ventricular assist device provided in this embodiment and is not a limitation thereof. In other types of ventricular assist devices, when a pump body using a magnetically levitated rotor is included, the control method of the ventricular assist device provided in this embodiment can be applied, and the present invention is not limited thereto.

[0046] Furthermore, the ventricular assist device includes a position sensor 01, a control module 02, a dynamics calculation module 03, an electromagnetic drive unit 04, and a magnetically levitated rotor 05. The position sensor 01 is used to feed back the current displacement information of the magnetically levitated rotor 05 to the control module 02. The control module 02 controls and drives the magnetically levitated rotor 05 through the electromagnetic drive unit 04 based on the current displacement information of the magnetically levitated rotor 05 and the control output (control voltage) calculated by the dynamics calculation module 03.

[0047] To address the issues of low accuracy in the dynamic models and complex control rate calculations in existing ventricular assist devices, this invention provides a control method for such devices. It should be noted that the control rate referred to herein refers to the mathematical expression related to the control of the magnetically levitated rotor 05. The control rate obtained based on this invention can maintain the stability of the magnetically levitated rotor 05 under complex operating conditions. The control module 02 is the embodiment of the control rate, which can provide a reasonable control output based on the current state. The control method for the ventricular assist device includes:

[0048] Step S1: Acquire voltage driving force excitation data and driving force motion trajectory excitation data;

[0049] Step S2: Establish an implicit dynamic model based on the voltage driving force excitation data and the driving force motion trajectory excitation data; and

[0050] Step S3: Predict the state of the ventricular assist device based on the implicit dynamic model, and update the parameters of the control network according to the predicted state to obtain the control rate of the ventricular assist device.

[0051] For the control method of the ventricular assist device provided in the embodiments, please refer to... Figure 2 It shows the overall control flow based on this control method.

[0052] The control method for the ventricular assist device provided in this embodiment mainly establishes an implicit dynamic model through two-stage excitation data (referring to the voltage driving force excitation data at the front end and the driving force motion trajectory excitation data at the back end), which effectively improves the accuracy of the dynamic model. Furthermore, the parameters of the control network are updated by predicting the state of the ventricular assist device through the implicit dynamic model, simplifying the process of solving for the control law.

[0053] In step S1, in an alternative example, to obtain excitation data, an independently controlled servo platform can be built. This servo platform is an external auxiliary measurement device and does not participate in the normal operation of the ventricular assist device. The specific construction of the servo platform can refer to existing technologies, and will not be elaborated in this embodiment. The desired excitation trajectory is provided to the servo platform, which then executes it automatically. At this time, the servo platform can acquire driving force (i.e., electromagnetic force) data through a force sensor, and obtain the voltage and dynamic parameters (such as displacement, velocity, or acceleration) output by the servo platform to obtain voltage driving force excitation data and driving force motion trajectory excitation data. The voltage driving force excitation data is the conversion data from voltage to driving force, and the driving force motion trajectory excitation data is the conversion data from driving force to dynamic parameters. Because the process from voltage to driving force does not have the abnormal problems similar to the process from driving force to displacement, the voltage driving force excitation data can be generated using conventional linear scanning. It should be noted that obtaining excitation data by building a servo platform is only an example of an excitation data acquisition method and not a limitation. Other embodiments are not limited to using a servo platform to obtain excitation data.

[0054] Optionally, the method for obtaining voltage driving force excitation data and driving force motion trajectory excitation data in step S1 is based on an excitation strategy that minimizes the condition number.

[0055] Please refer to Figure 3 It exemplifies the generation process of an incentive strategy based on minimizing the condition number, which includes:

[0056] Step S11: Use a finite-term Fourier series as the initialization expression for the dynamic excitation trajectory.

[0057]

[0058] In the formula an b n q0 is the parameter to be optimized, ω f To excite the fundamental frequency, N is the number of terms in the Fourier series, q, These represent displacement, velocity, and acceleration, respectively, with t representing time.

[0059] The dynamic excitation trajectory is the desired trajectory obtained by optimizing a finite-term Fourier series, and the excitation trajectory is determined by the optimization parameter a. n b n q0 is uniquely determined, substituting it into a n b n q0 can be used to obtain q. The expected value matrix of (position, velocity, acceleration). Fourier series can approximate any function, offering high degrees of freedom. In practice, three to five terms are typically used in the Fourier series; increasing the number of terms yields diminishing returns, so only the most influential terms are considered. The fundamental excitation frequency refers to the operating frequency of the system to be excited, chosen independently for different systems. Displacement, velocity, and acceleration q... Let be the dynamic parameters of the magnetically levitated rotor. The condition number is equal to the norm of the matrix multiplied by the norm of the matrix inverse. The condition number is an indicator of the ill-conditioned nature of a matrix. Minimizing the condition number can reduce the impact of sensor sampling data noise on the establishment of the dynamic model.

[0060] Step S12: Use the interior point method of the obstacle function to iteratively optimize and obtain the optimal excitation trajectory:

[0061]

[0062] Where t0 and t f These are the start time and the end time, respectively. The optimization objective is to minimize the displacement, velocity, and acceleration of the condition number trajectory. The obstacle function interior point method can be solved using mathematical tools such as MATLAB. This embodiment will not elaborate on this.

[0063] Optionally, in step S2, a two-stage differential neural network can be used to establish an implicit dynamic model. The implicit dynamic model established in this way is a nonlinear model in which the magnetic levitation rotor and the electromagnetic multiphysics field are fully coupled, effectively improving accuracy. Step S2 preferably includes:

[0064] Step S21: Train the encoding part of the differential neural network using voltage-driven force excitation data;

[0065] Step S22: Train the decoding part of the differential neural network using the driving force motion trajectory excitation data; and

[0066] Step S23: Align the trained encoding part with the trained decoding part to obtain the implicit dynamic model.

[0067] Please refer to Figure 4 It illustrates the architecture of the implicit dynamics model and expresses the continuous flow layer of the differential neural network. Among them, Figure 4 The upper part represents the encoding part (i.e., the implicit dynamics front end), and the lower part represents the decoding part (i.e., the implicit dynamics back end). Optionally, the encoding part includes encoding the voltage to drive force; the decoding part includes decoding the drive force to dynamic parameters; step S23 preferably includes aligning the drive force of the encoding part with the drive force of the decoding part, thus obtaining the implicit dynamics model based on the training of the differential neural network.

[0068] The following is an illustrative explanation of the training process for a differential neural network:

[0069] Step S41: Input dynamic parameters θ, start time t0, end time t1, final state z(t1), and loss gradient.

[0070] Step S42: Initialize augmented state

[0071] Step S43: Define the augmented state dynamics behavior aug_dynamics([z(t), a(t)], t, θ);

[0072] Step S44: Calculate the Jacobian vector

[0073] Step S45: Solve the differential equation

[0074]

[0075] Step S46: Return the gradient

[0076] Wherein, the dynamic parameter θ is the neuron parameter in the neural differential equation, the start time t0 is the initial time of the differential equation calculation, the end time t1 is the termination time of the differential equation calculation, the final state z(t1) is the dynamic state obtained from the solution, and the loss gradient is... The gradient for parameter updates during backpropagation of the neural network; a(t) is the adjoint sensitivity; T is the matrix transpose, which represents the neural network equation, the final state, and the partial derivatives of the neurons, respectively; ODESolve is the differential equation solver, and aug_dynamics is the augmented dynamic state.

[0077] Preferably, the steps of training a differential neural network include: calculating the gradient based on the augmented state with adjoint sensitivity. Since the overhead of using a differential equation solver to calculate the gradient is huge, it is preferable to introduce the adjoint sensitivity method and use the augmented state to calculate the gradient.

[0078] Step S3 applies the implicit dynamics model established in step S2. In an example, to control the magnetically levitated rotor, a control network based on the implicit dynamics model prediction can be used. This control network may include an action network and a Critic network Q. ω1 (s, a), Q ω2 (s, a) and the evaluation network Actor network π θ (s). The final output of the action network is the control rate. ω1, ω2, and θ are the parameters of the control network. Setting these parameters aims to more accurately output the control voltage based on the current position of the magnetic levitation rotor 05. Step S3 introduces an implicit dynamic model predictive control improved maximum entropy reinforcement learning algorithm, which can effectively reduce the control rate when solving complex targets. It is particularly effective for targets like cardiac assist devices, such as magnetic levitation blood pumps, which have strict requirements for various control parameters and complex scenarios.

[0079] Please refer to Figure 5 In one example, step S3, which involves updating the parameters of the control network based on the maximum entropy reinforcement learning algorithm, includes:

[0080] Step S31: Obtain the current state of the ventricular assist device; here, the current state of the ventricular assist device mainly includes the current position information of the magnetic levitation rotor, which can be obtained through a position sensor;

[0081] Step S32: Predict the multi-step states of the ventricular assist device based on the implicit dynamics model, and calculate the multi-step reward;

[0082] Step S33: Execute an action based on the voltage output by the current action network (such as turning the MOSFET of the electromagnetic drive unit 04 on and off).

[0083] Step S34: Based on the root mean square loss output by the multi-step reward and evaluation network, gradient backpropagation is performed in combination with the entropy regularization loss of the action state to update the parameters of the control network.

[0084] The following exemplifies the steps for updating the parameters of a control network based on the maximum entropy reinforcement learning algorithm:

[0085] The Critic network Q is initialized with parameters ω1, ω2, and θ from a random control network. ω1 (s, a), Q ω2 (s, a) and Actor network πθ (s);

[0086] Copy the same parameters Initialize the target network respectively and

[0087] Initialize the experience replay pool R;

[0088] for the sequence e = 1 → Edo;

[0089] Obtain the initial environmental state s1;

[0090] for time step t = 1 → Tdo;

[0091] Select action a based on the current strategy. t =π θ (s t );

[0092] Perform action a t Using an implicit dynamics model to predict n-step states, calculate the reward r. t The environmental state becomes s t+1 ;

[0093] (s) t a t r t s t+1 Store in the return visit pool R;

[0094] For training rounds k = 1, we move to Kdo;

[0095] Sample N tuples {(s) from R t a t r t s t+1 )} i＝1，...，N ;

[0096] Calculate separately

[0097] Minimize the loss function for two Critic networks

[0098] Sample motion using parameterization techniques Then update the current Actor network using the following loss function:

[0099]

[0100] Update the coefficient 'a' of the entropy regularization term;

[0101] Update the target network:

[0102] endfor;

[0103] endfor;

[0104] endfor;

[0105] Where e is the index variable of the action sequence, and E is the length of the action sequence. → represents increment, the initial position of the magnetic levitation rotor in the initial state of the environment; the current policy refers to the execution based on the voltage output by the current action network; the reward is the constraint designed to achieve the goal during reinforcement learning; γmin j＝1，2 The outputs and minimum values ​​of the two evaluation networks are given. It is preferable to use two evaluation networks to prevent bootstrapping. The coefficient 'a' of the entropy regularization term is the proportion of entropy loss. τ is the retention factor in the parameter update strategy.

[0106] Please refer to Figure 6 This illustration demonstrates a control method for the ventricular assist device provided in this embodiment. It shows the image frame stack for each time step (e.g., 0.1 ms) from startup to reaching the geometric center of the magnetic levitation rotor 05. The image frame stack shows that the magnetic levitation rotor 05 smoothly and quickly reaches the geometric center. Specifically, the displacement information of the magnetic levitation rotor 05 is detected by the position sensor 01 and fed back to the control module 02. The control module 02 calculates the control output (e.g., control voltage) based on the displacement information of the magnetic levitation rotor 05 and the dynamics calculation module 03, and controls the electromagnetic drive unit 04, thus changing the position of the magnetic levitation rotor 05. In some embodiments, this loop is performed every 0.1 ms, and the closed-loop iteration allows the magnetic levitation rotor 05 to smoothly and quickly reach the geometric center. The image frame stack for each step in this process is as follows: Figure 6 As shown.

[0107] Based on the control method for the ventricular assist device described above, embodiments of the present invention also provide a readable storage medium storing a program thereon. When the program is executed, it implements the steps of the control method for the ventricular assist device described above. This readable storage medium can be set independently or attached to the control module of the ventricular assist device; the present invention is not limited in this regard. Furthermore, embodiments of the present invention also provide a ventricular assist device, which includes a control module and a magnetically levitated rotor. The control module drives the magnetically levitated rotor according to the control method for the ventricular assist device described above.

[0108] In summary, the control method, readable storage medium, and ventricular assist device provided by this invention include the following steps: acquiring voltage driving force excitation data and driving force motion trajectory excitation data; establishing an implicit dynamic model based on the voltage driving force excitation data and the driving force motion trajectory excitation data; predicting the state of the ventricular assist device based on the implicit dynamic model; and updating the parameters of the control network according to the predicted state to obtain the control rate of the ventricular assist device. This configuration, using two-stage excitation data (referring to the front-end voltage driving force excitation data and the back-end driving force motion trajectory excitation data) to establish the implicit dynamic model, effectively improves the accuracy of the dynamic model. Furthermore, updating the parameters of the control network by predicting the state of the ventricular assist device using the implicit dynamic model simplifies the process of solving for the control rate.

[0109] It should be noted that the above embodiments can be combined with each other. The above description is only a description of preferred embodiments of the present invention and is not intended to limit the scope of the present invention in any way. Any changes or modifications made by those skilled in the art based on the above disclosure shall fall within the protection scope of the present invention.

Claims

1. A control method for a ventricular assist device, characterized in that, include: Acquire voltage-driven force excitation data and driving force motion trajectory excitation data; An implicit dynamic model is established based on the voltage driving force excitation data and the driving force motion trajectory excitation data; as well as The state of the ventricular assist device is predicted based on the implicit dynamics model, and the parameters of the control network are updated according to the predicted state to obtain the control rate of the ventricular assist device. The steps for establishing an implicit dynamic model based on the voltage driving force excitation data and the driving force motion trajectory excitation data include: The encoding part of the differential neural network is trained using voltage-driven force excitation data; The decoding part of the differential neural network is trained using excitation data of the driving force motion trajectory; and Aligning the trained encoding part with the trained decoding part yields the implicit dynamics model; wherein the encoding part includes encoding of voltage conversion to driving force; the decoding part includes decoding of driving force conversion to dynamic parameters; the step of aligning the trained encoding part with the trained decoding part includes aligning the driving force of the encoded part with the driving force of the decoded part; the step of updating the parameters of the control network according to the predicted state includes updating the parameters of the control network based on the predicted state using a maximum entropy reinforcement learning algorithm.

2. The control method for the ventricular assist device according to claim 1, characterized in that, The steps for training a differential neural network include: Gradient calculation based on augmented state with adjoint sensitivity.

3. The control method for the ventricular assist device according to claim 1, characterized in that, The method for acquiring voltage driving force excitation data and driving force motion trajectory excitation data is based on an excitation strategy that minimizes the condition number.

4. The control method for the ventricular assist device according to claim 3, characterized in that, The incentive strategy based on minimizing the condition number includes: Initialize the dynamic excitation trajectory using a finite-term Fourier series; Based on the dynamic excitation trajectory, the excitation strategy is obtained through iterative optimization.

5. The control method for the ventricular assist device according to claim 1, characterized in that, The steps for updating the parameters of the control network based on the maximum entropy reinforcement learning algorithm include: Obtain the current status of the ventricular assist device; The implicit dynamics model is used to predict the multi-step states of the ventricular assist device and calculate the multi-step reward. Execute the action based on the voltage output by the current action network; The root mean square loss of the multi-step reward and evaluation network output is used to perform gradient backpropagation, combined with the entropy regularization loss of the action state, to update the parameters of the control network.

6. A readable storage medium having a program stored thereon, characterized in that, When the program is executed, it implements the steps of the control method for the ventricular assist device as described in any one of claims 1 to 5.

7. A ventricular assist device, characterized in that, The device includes a control module and a magnetically levitated rotor, wherein the control module drives the magnetically levitated rotor according to the control method of any one of claims 1 to 5 for a ventricular assist device.

Images