CN111768836B - DBS closed-loop learning method in Parkinson's state based on generalized generative adversarial network - Google Patents

Abstract

The present invention provides a closed-loop learning method for DBS in Parkinson's state based on a generalized generative adversarial network. The present invention creatively defines the generative network and the basal ganglia-thalamus-cortex neuron network built by MATLAB as a generalized generative network. The design is based on the generalized generative adversarial network. Closed-loop DBS deep learning algorithm for the network. The basal ganglia-thalamus-cortex neuron network built by a field-programmable gate array obtains the required data set, and conducts repeated adversarial training on the generalized generative network and the discriminant network, so that when random noise is input to the generalized generative network, the generalized generative network It can output a "false and real" data sequence, thereby obtaining a control signal that can effectively suppress the Parkinson's state and achieve effective control of the Parkinson's state. This algorithm adopts the idea of deep learning and uses an adversarial network algorithm based on a generalized generative network to perform closed-loop DBS modulation of the Parkinson's state, making it possible for adaptive DBS technology to improve the Parkinson's state.

Description

DBS closed-loop learning method in Parkinson's state based on generalized generative adversarial network

Technical field

The invention relates to the fields of biomedical engineering technology and deep learning, in particular to a DBS closed-loop deep learning algorithm in Parkinson's disease based on a generalized generative adversarial network.

Background technique

Parkinson's disease (PD) is a degenerative neurological disease caused by the functional degradation of the central nervous system. Patients with Parkinson's disease show symptoms such as resting tremor, muscle stiffness, bradykinesia, abnormal posture and gait. Research shows that Parkinson's disease is related to abnormal electrophysiological activity of neuronal network circuits in the basal ganglia of the brain. The basal ganglia region of the human brain mainly includes three parts: the subthalamic nucleus (STN), the globus pallidus externa (GPe), and the globus pallidus medial (Globus Pallidus, GPi). Cortical neurons can be divided into cortical pyramids. Neuron (Pyramidal neuron, PY) and cortical interneuron (IN), TH (thalamus) represents thalamic neurons.

The Izhikevich neuron model has been favored by more and more researchers because of its computational efficiency, simple structure, and ability to simulate most of the characteristics of actual neurons. The mathematical description of the model is as follows:

Reset equation after discharge:

If V≥30mV, then

The variable V represents the membrane voltage, and u represents the membrane recovery variable. The four dimensionless parameters a, b, c and d can be combined into various different discharge characteristics.

FPGA technology is a semi-customized circuit technology in the field of application-specific integrated circuits. It has important application value in the research of neuron characteristics, synchronization phenomenon mechanisms, bionics, intelligent systems, etc.

At present, the treatments for PD mainly include drug treatment, surgical treatment and DBS (Deep brain stimulation). Compared with drug treatment and surgical treatment, DBS has the advantages of reversibility and adjustability, and is increasingly favored by patients with refractory PD. Clinical studies have shown that DBS has significant effects and has become the preferred treatment option for patients with refractory PD. However, open-loop DBS is currently mainly used in clinical practice. The parameters of DBS are set based on the experience of clinicians. The fixed stimulation pulse sequence is difficult to adapt to the differences between individual patients. In order to solve the problem that open-loop DBS cannot adapt to differences between individual patients, the present invention proposes a closed-loop deep learning method for DBS in Parkinson's disease based on a generalized generative adversarial network. This method can adjust stimulation according to the differences between different patients. parameters, which can effectively overcome the shortcomings of existing open-loop DBS methods.

Currently, the open-loop DBS stimulation treatment method is used clinically for DBS modulation of Parkinson's disease, which can only make targeted parameter settings for the PD state (a set of parameters can only be effective for one Parkinson's patient, and different Parkinson's patients cannot Parameters need to be readjusted again) and cannot effectively adapt to individual differences to provide personalized treatment, which seriously hinders the recovery process of PD patients. GAN has many advantages and has achieved good results in some fields, but traditional GAN is difficult to apply to closed-loop DBS control of PD states.

Contents of the invention

The present invention aims at the problem that the open-loop DBS treatment method has serious limitations, and the traditional generative adversarial network is difficult to apply to the closed-loop DBS control problem of Parkinson's disease. In order to solve the current difficulties faced by the clinical analysis and treatment of neuroscience diseases, the inventiveness of the present invention proposed the concept of generalized generative network (5), defined the generative network (6) and MATLAB basal ganglia-thalamus-cortex neuron network (7) as generalized generative network (5), and designed and developed a generalized generative adversarial network based on The DBS closed-loop deep learning method in Parkinson's disease calculates and analyzes the local field potential signal of PD state through neural network. According to the changes of the real-time local field potential signal, the DBS is adaptively modulated through the generalized generative adversarial network, and the closed-loop control of the Parkinson's state based on the generalized generative adversarial network is realized, which solves the problem of the control of the Parkinson's state by the traditional generative adversarial network. It also provides new ideas for closed-loop control of Parkinson's disease, making personalized closed-loop DBS modulation possible.

The technical solution adopted by the present invention is to provide a DBS closed-loop learning method in Parkinson's state based on a generalized generative adversarial network, which is characterized in that the method combines the generative network (6) and the MATLAB basal ganglia-thalamus-cortex neuron network ( 7) Defined as a generalized generative network (5), through repeated confrontation training of the generalized generative network and the discriminant network, when random noise is input to the generalized generative network, the generalized generative network can output a data sequence that is "false and real", thus obtaining Effectively suppresses the control signals of Parkinson's state and achieves effective control of Parkinson's state.

The method includes the following steps:

The first step is to obtain the data set: adjust the FPGA basal ganglia-thalamic neuron network (1) to a "normal person" state, imitate the local field potential signals of the normal people's basal ganglia-thalamic neuron network, and collect multiple "normal The local field potential signals of the PY nuclei of "people" over a period of time are expanded through data processing to obtain n sets of data pairs as training data sets;

The second step is to build the model:

First, the MATLAB basal ganglia-thalamus-cortex neuron network (7) is adjusted to the Parkinson's state. The generalized generative network (5) is composed of the generative network (6) and the MATLAB basal ganglia-thalamus-cortex neuron network (7). Generate The network (6) generates a corresponding output by inputting random noise (4), and the control signal output by the generated network (6) is input into the MATLAB basal nucleus-thalamus-cortex neuron network (7), and the MATLAB basal nucleus-thalamus-cortex The local field potential signal of the PY nucleus output by the neuron network (7) is input into the discriminant network (9) for discrimination.

Among them, the generation network (6) and the discriminant network (9) both adopt error back propagation neural network;

If the local field potential signal generated by the generalized generation network (5) is judged to meet the consistency requirements, the control signal generated by the generalized generation network is the optimal control signal, and the optimal control signal directly passes through the DBS pulse generator (10) to generate an electrical stimulation signal Add it to the FPGA Basal Nucleus-Thalamus Neuron Network (1); if the consistency requirements are not met, the error back propagation algorithm must be used to correct the weights of the generated network (6), and at the same time, the MATLAB Basal Nucleus-Thalamus-Cortex The local field potential signal generated by the neuron network (7) is set to label 0, and the weight of the discriminant network (9) is updated;

The third step, model training:

Set the label of the "normal person" training data set (8) obtained in the first step to 1, and input the training data set (8) into the discriminant network (9) for training and adjust the network parameters of the discriminant network; after training the discriminant network In the process of (9), first fix the generalized generating network (5), update the network weights of the discriminating network (9), then fix the discriminating network (9), and update the generating network (6) according to the discriminating result of the discriminating network (9). The weights are trained repeatedly and iteratively until both sides reach a dynamic balance;

Then, any given random noise is input into the generalized generating network, and the generalized generating network outputs the local field potential signal to the discriminating network. The discriminating network determines whether the consistency requirement is met. If it is met, electrical stimulation is generated through the DBS pulse generator (10) The signal is added to the FPGA basal ganglia-thalamic neuron network (1); if it is not satisfied, the error back propagation algorithm is used to correct the weight of the generated network (6), and the MATLAB basal ganglia-thalamus-cortical neuron network (6) is also added. The local field potential signal generated by the meta-network (7) is set to label 0, and the label of the "normal person" training data set (8) obtained in the first step is set to 1, and the optimized discriminant network (9) is trained again. Network parameters, thus completing a single alternating iterative training, and training and alternating iterations are repeated until the discriminant network (9) cannot distinguish which data sequence is generated by the generalized generative network;

The fourth step, "Parkinson's state" control:

Adjust the FPGA basal ganglia-thalamic neuron network (1) to the "Parkinson's patient" state and act as a controlled object. The optimal control signal output by the generalized generation network (5) after judgment by the discriminant network is passed through the DBS pulse generator ( 10) Input to the FPGA basal ganglia-thalamic neuron network (1) for "Parkinson's state" modulation.

The number of neurons in the input layer, hidden layer and output layer of the generative network (6) and the discriminant network (9) are 100, 150 and 1 respectively; the learning rates are all 0.1, and the activation functions of the hidden layer neurons are all Use the Sigmoid function.

The parameters of the Izhikevich neuron model of the MATLAB basal ganglia-thalamic-cortical neuron network (7) are:

.

The synaptic delay time is introduced into the MATLAB basal ganglia-thalamus-cortex neuron network (7), and the value of the delay parameter is:

τ _GPe→STN =6ms

τGPe _→GPi =6ms

τGPi _→Th =6ms

_τTh→PY =6ms

τ _PY→STN =6ms

τ _STN→GPe =2ms

τ _STN→GPi =2ms

τGPe _→GPe =2ms

τPY _→IN =2ms

τ _IN→PY =2ms.

Compared with the prior art, the beneficial effects of the present invention are:

The present invention uses the advantages of GAN to creatively propose the concept of generalized generative network, defines the generative network and MATLAB basal ganglia-thalamus-cortex neuron network as generalized generative network, and realizes the use of adversarial network algorithms based on generalized generative network to treat Parkinson's disease. Closed-loop DBS control of the Parkinson's state solves the problem that traditional GAN is difficult to apply in the control of Parkinson's state. It also provides a new idea for the closed-loop control of Parkinson's state and becomes the development direction of the future treatment of Parkinson's disease.

Generative Adversarial Networks (GAN) is a deep learning model. The main structure of GAN includes a generator (Generator, G) and a discriminator (Discriminator, D). GAN produces quite good output through mutual game learning of G and D. Here we take generating a data sequence as an example. Assume there are two networks G and D. G is a network that generates a data sequence. It receives a random noise and generates a data sequence through this random noise. D is a discriminant network that determines whether the data sequence is "real". The input of the discriminant network is the real data sequence and the data sequence generated by the generation network, and the output is the probability of being a "real" data sequence. If it is 1, it means it is a "real" data sequence. If the output is 0, it means it is not a "real" data sequence. real" data sequence. During the training process, the goal of the generation network G is to try to generate "real" data sequences to deceive the discriminant network D. The goal of D is to try to distinguish the data sequence generated by G from the real data sequence. In this way, G and D constitute a dynamic game process. In an ideal state, G can generate a data sequence that is enough to “disguise the real” from the fake. For D, it is difficult to determine whether the data sequence generated by G is real.

The present invention conducts mutual game training through the generalized generative network (5) and the discriminant network (9) in the generalized generative adversarial network, and obtains the optimal control signal to treat the PD state through the DBS pulse generator (10), and through closed-loop control ( Closed-loop control refers to the generalized generative adversarial network to obtain the optimal control signal u through adversarial game training. For different Parkinson's patients, his Parkinson's state is different, that is to say, his MATLAB basal ganglia-thalamus-cortex neuron network ( 7) are different states, so the present invention uses generalized generative adversarial network and discriminant network to compete against each other for game training. No matter what kind of Parkinson's state is targeted, the optimal control signal for this Parkinson's state can be obtained, thereby achieving Personalized treatment), which can treat personalized PD patients and achieve personalized treatment.

Description of drawings

Figure 1 is a schematic diagram of the system structure of the present invention;

Figure 2 is a control flow chart of the present invention;

In the picture:

1. FPGA basal ganglia-thalamic neuron network, 2. Local field potential calculation module, 3. Adversarial network algorithm based on generalized generative network, 4. Random noise, 5. Generalized generative network, 6. Generative network, 7. MATLAB base Nuclear-thalamic-cortical neuron network, 8. Training data set, 9. Discriminative network, 10. DBS pulse generator, 11. Local field potential signal.

Detailed ways

The DBS closed-loop deep learning algorithm in Parkinson's state based on the generalized generative adversarial network of the present invention will be further elaborated below with reference to the accompanying drawings.

The steps of the DBS closed-loop learning method in Parkinson's state based on the generalized generative adversarial network of the present invention are:

The first step is to build the MATLAB basal ganglia-thalamus-cortex neuron network (7): Based on the basal ganglia-thalamus-cortex neuron network model proposed by Lu et al. (Lu ML, Wei XL and Loparo KA (2017) InvestigatingSynchronous Oscillation and Deep Brain Stimulation Treatment in A Model ofCortico-Basal Ganglia Network.IEEE Trans.Neural Syst.Rehabil.Eng., vol.25, pp.1950-1958.), introduced synaptic delay time, and modified the Izhikevich neuron model parameters to simulate normal and parkinsonian states, making it closer to the real brain basal ganglia-thalamo-cortical neuron network.

The second step is to obtain the data set: deep brain stimulation of basal ganglia-thalamic neuron network (1) in FPGA (Deng Bin, Zhang Maohua, Wang Xiaojun, Wei Xile, Li Huiyan, Yu Haitao, Wang Jiang, Parkinson's disease basal ganglia-thalamic network) Experimental platform: China, CN103691058A[P].2014-04-02.) is adjusted to a "normal person" state and imitates normal people through the local field potential calculation module 2 (the specific form of the local field potential calculation module is implemented based on existing technology) The local field potential signals of the basal ganglia-thalamus-cortex neuron network are collected from the local field potential signals of the PY nuclei of 10 "normal people" over a period of time, and then expanded through data processing such as denoising, normalization, and grouping. Obtain 100 sets of data pairs to obtain the required training data set (8);

The third step is to build the model: In this invention, both the generation network (6) and the discriminant network (9) adopt error back propagation (BP) neural network. The generation network (6) and the discriminant network (9) are two completely independent neural network models. The training method is individual alternating iterative training. The numbers of neurons in the input layer, hidden layer and output layer of the generative network (6) and the discriminative network (9) are 100, 150 and 1 respectively. The learning rates are all 0.1, and the activation functions of hidden layer neurons all use the Sigmoid function; the generative network (6) and the MATLAB basal ganglia-thalamus-cortex neuron network (7) are defined as generalized generative networks (5). Generalized generative network Network (5) and discriminant network (9) constitute a generalized generative adversarial network, and adversarial game training is performed through generalized generative network (5) and discriminant network (9).

The fourth step, model training: Set the label of the “normal person” training data set (8) obtained in the first step to 1, and input the training data set (8) into the discriminant network (9) for training and adjustment of network parameters. In the process of training the discriminant network (9), first fix the generalized generative network (5), update the network weights of the discriminant network (9), then fix the discriminant network (9), and update the generative network according to the discrimination results of the discriminant network (9) The weights of (6) are repeated in this way, alternately and iteratively trained, until both parties reach a dynamic balance. During the training process, the MATLAB basal ganglia-thalamus-cortex neuron network (7) is adjusted to the parameters of different "Parkinson's states" so that the model can identify different "Parkinson's states";

The fifth step, "Parkinson's state" control: adjust the FPGA basal ganglia-thalamus neuron network (1) to a certain "Parkinson's patient" state, act as a controlled object, and adjust the MATLAB basal ganglia-thalamus-cortex accordingly The neuron network (7) is adjusted to the parameters corresponding to the state of "Parkinson's patient"; after the local potential signal output by the generalized generation network (5) is judged by the discriminant network and confirmed as the optimal control signal u, it passes through the DBS pulse generator (10) is input into the FPGA basal ganglia-thalamic neuron network (1) for "Parkinson's state" modulation.

Example 1

This embodiment is based on the DBS closed-loop deep learning method in Parkinson's state based on generalized generative adversarial network.

The first step is to build a MATLAB basal ganglia-thalamus-cortex neuron network (7).

Based on the cortical-basal ganglia-thalamic neuron network model proposed by Lu et al., synaptic delay time was introduced, and the parameters of the Izhikevich neuron model were modified to simulate normal and Parkinson's states. The specific modifications of the parameters of the Izhikevich neuron model are detailed in Table 1.

Table 1 Parameter modification of Izhikevich neuron model

Table 1 shows the parameter changes of Lu's model and modified model from normal state to Parkinson's state. The value before "→" represents the normal state, and the value after "→" represents the Parkinson's state value.

Synaptic current from j-th neuron to i-th neuron It is defined as shown in equation (4), which shows the synaptic current calculation formula after taking into account the synaptic delay time.

Among them, g _ij describes the synaptic coupling strength, E _syn is the reversal potential, and τ is the delay time. See Table 2 for specific values. Based on these modifications, it is closer to the real brain basal ganglia-thalamo-cortical neuron network.

Table 2 Synaptic delay time parameters

The second step is to obtain the data set.

First, adjust the FPGA basal ganglia-thalamus neuron network (1) to a "normal person" state, imitate the local field potential signals of the normal people's basal ganglia-thalamus-cortex neuron network, and collect 10 "normal people" over a period of time. Due to the small amount of data, the local field potential signal of the PY core group is used in the present invention to expand the data using data enhancement methods. According to the characteristics of large data interference, this experiment uses overlapping sliding windows, adding noise and other methods to appropriately expand the data. The network training is more complete and the effect is better. The required training data set is obtained through data preprocessing methods such as data enhancement, filtering, and normalization (8).

The third step is to build the model.

The present invention designs and develops a closed-loop deep learning method for DBS in Parkinson's state based on a generalized generative adversarial network. Through mutual game training of the generalized generative network (5) and the discriminant network (9), the DBS is adaptively modulated to realize the PD state. of personalized treatment.

(1) Build a generalized generative network model

First, the MATLAB basal ganglia-thalamus-cortex neuron network (7) is adjusted to the Parkinson's state. The generalized generative network (5) is composed of the generative network (6) and the MATLAB basal ganglia-thalamus-cortex neuron network (7). Generate The network (6) generates a corresponding output by inputting random noise (4), and the output of the generating network (6) is input into the MATLAB basal ganglia-thalamic-cortical neuron network (7), and the MATLAB basal ganglia-thalamic-cortical neuron network The local field potential signal of the PY nuclei output by the network (7) is input into the discriminating network (9) for discrimination. In the present invention, the generating network (6) adopts the BP neural network. The number of neurons in its input layer, hidden layer and output layer are 100, 150 and 1 respectively. The learning rate is 0.1, and the activation function of the hidden layer neurons adopts the Sigmoid function.

(2) Build a discriminant network model

The discriminant network (9) determines whether the input data sequence is true or false, and the generalized generative network (5) and the discriminant network (9) are trained by playing games with each other, thereby making the data sequence generated by the generalized generative network (5) more realistic. The discriminant network (9) in the present invention also adopts BP neural network. The number of neurons in its input layer, hidden layer and output layer are 100, 150 and 1 respectively. The learning rate is 0.1, and the activation function of the hidden layer neurons adopts the Sigmoid function.

The purpose of the generalized generative network is to generate data that is the same as the normal state, and the purpose of the discriminant network is to distinguish between true and false, that is, to judge that the output of the generalized generative network 5 is false data. The goal of the generalized generative network is to generate data that is as real as possible to fool the discriminant network. The purpose of the discriminant network is to identify that the output of the generalized generative network 5 is false data, so that they form a process of confrontational game.

The fourth step is model training.

First, random noise (4) is input to the generating network (6) to generate an output. The output of the generating network (6) is input to the local field generated in the MATLAB basal ganglia-thalamic-cortical neuron network (7) through the DBS pulse generator. The potential signal is input to the discriminant network (9) for discrimination. If the consistency condition is met, it means that the control signal generated by the generalized generation network (5) meets the requirements, and the electrical stimulation signal can be generated directly through another DBS pulse generator (10). Added to the FPGA Basal Nucleus-Thalamus Neuron Network (1), if the consistent conditions are not met, the error back propagation algorithm must be used to correct the weights of the generated network (6), and at the same time, the MATLAB Basal Nucleus-Thalamus-Cortex The local field potential signal generated by the neuron network (7) is set to label 0, the label of the "normal person" training data set (8) obtained in the first step is set to 1, and the optimized discrimination network (9) is trained again. Network parameters, update the weights of the discriminant network (9), thus completing a single alternating iterative training. Repeat the training and alternate iterations until the discriminant network (9) cannot distinguish which data sequence is generated by the generalized generative network.

The fifth step is "Parkinson's state" control.

Adjust the FPGA basal ganglia-thalamus neuron network (1) to the "Parkinson's patient" state to act as a controlled object, and accordingly adjust the MATLAB basal ganglia-thalamus-cortex neuron network (7) to correspond to the "Parkinson's patient" state. "Parameters in the state; the optimal control signal after the generalized generative network (5) is determined to be true by the discriminant network is input into the FPGA basal ganglia-thalamic neuron network (1) through the second DBS pulse generator (10) Perform "Parkinsonian state" modulation.

Noise is randomly given and input into the generalized generative network 5. After game training of the generalized generative network and the adversarial network, it is given how to modulate the current patient. Instead, based on the current patient status, the generalized generative adversarial network performs modulation based on the current patient status. Adversarial game training gives how the current patient should be modulated, thus achieving personalized treatment.

According to the characteristics of the neuron network in the basal ganglia area of the brain, the present invention uses MATLAB to reasonably build a basal ganglia-thalamus-cortex neuron network. The built MATLAB basal ganglia-thalamus-cortex neuron network (7) is achieved by adjusting MATLAB basal ganglia-thalamus. -The parameters of the cortical neuron network (7) can correctly simulate the "normal person" state and the "Parkinson's state". There is no need to adjust the parameters of the MATLAB basal ganglia-thalamus-cortical neuron network (7) for a Parkinson's state. If To change the Parkinson's state, you need to adjust it flexibly according to the actual Parkinson's state. For example, adjust the a, b, c, d parameters of the Izhikevich neuron model and the prominent coupling strength g _ij and other parameters according to different Parkinson's states to simulate different Parkinson's state), the present invention also uses FPGA to build a basal ganglia-thalamic neuron network, which can also simulate the "normal person" state and "Parkinson's state". The FPGA basal ganglia-thalamic neuron network is used to obtain training data Set and serve as a controlled object to simulate the human brain.

The parts not described in the present invention are applicable to the existing technology.

Claims (4)

1. A closed-loop learning method for DBS in Parkinson’s state based on generalized generative adversarial network, which is characterized in that the method includes the following steps: The first step is to obtain the data set: adjust the FPGA basal ganglia-thalamic neuron network (1) to a "normal person" state, imitate the local field potential signals of the normal people's basal ganglia-thalamic neuron network, and collect multiple "normal The local field potential signals of the PY nuclei of "people" over a period of time are expanded through data processing to obtain n sets of data pairs as training data sets; The second step is to build the model: First, the MATLAB basal ganglia-thalamus-cortex neuron network (7) is adjusted to the Parkinson's state. The generalized generative network (5) is composed of the generative network (6) and the MATLAB basal ganglia-thalamus-cortex neuron network (7). Generate The network (6) generates a corresponding output by inputting random noise (4), and the control signal output by the generated network (6) is input into the MATLAB basal nucleus-thalamus-cortex neuron network (7), and the MATLAB basal nucleus-thalamus-cortex The local field potential signal of the PY nucleus output by the neuron network (7) is input into the discriminant network (9) for discrimination. Among them, the generation network (6) and the discriminant network (9) both adopt error back propagation neural network; If the local field potential signal generated by the generalized generation network (5) is judged to meet the consistency requirements, the control signal generated by the generalized generation network is the optimal control signal, and the optimal control signal directly passes through the DBS pulse generator (10) to generate an electrical stimulation signal Add it to the FPGA Basal Nucleus-Thalamus Neuron Network (1); if the consistency requirements are not met, the error back propagation algorithm must be used to correct the weights of the generated network (6), and at the same time, the MATLAB Basal Nucleus-Thalamus-Cortex The local field potential signal generated by the neuron network (7) is set to label 0, and the weight of the discriminant network (9) is updated; The third step, model training: Set the label of the "normal person" training data set (8) obtained in the first step to 1, and input the training data set (8) into the discriminant network (9) for training and adjust the network parameters of the discriminant network; after training the discriminant network In the process of (9), first fix the generalized generating network (5), update the network weights of the discriminating network (9), then fix the discriminating network (9), and update the generating network (6) according to the discriminating result of the discriminating network (9). The weights are trained repeatedly and iteratively until both sides reach a dynamic balance; Then, any given random noise is input into the generalized generating network, and the generalized generating network outputs the local field potential signal to the discriminating network. The discriminating network determines whether the consistency requirement is met. If it is met, electrical stimulation is generated through the DBS pulse generator (10) The signal is added to the FPGA basal ganglia-thalamic neuron network (1); if it is not satisfied, the error back propagation algorithm is used to correct the weight of the generated network (6), and the MATLAB basal ganglia-thalamus-cortical neuron network (6) is also added. The local field potential signal generated by the meta-network (7) is set to label 0, and the label of the "normal person" training data set (8) obtained in the first step is set to 1, and the optimized discriminant network (9) is trained again. Network parameters, thus completing a single alternating iterative training, and training and alternating iterations are repeated until the discriminant network (9) cannot distinguish which data sequence is generated by the generalized generative network; The fourth step, "Parkinson's state" control: Adjust the FPGA basal ganglia-thalamic neuron network (1) to the "Parkinson's patient" state and act as a controlled object. The optimal control signal output by the generalized generation network (5) after judgment by the discriminant network is passed through the DBS pulse generator ( 10) Input to the FPGA basal ganglia-thalamic neuron network (1) for "Parkinson's state" modulation. 2. The learning method according to claim 1, characterized in that the number of neurons in the input layer, hidden layer and output layer of the generating network (6) and the discriminating network (9) are 100, 150 and 1 respectively. ; The learning rates are all 0.1, and the activation functions of hidden layer neurons all use Sigmoid functions. 3. The learning method according to claim 1, characterized in that the parameters of the Izhikevich neuron model of the MATLAB basal ganglia-thalamus-cortex neuron network (7) are: a _GPe value is 0.025, a _GPi value is 0.025, a _PY has a value of 0.001, b _PY has a value of 0.1, b _IN has a value of 0.2, c _GPi has a value of -65, d _STN has a value of 15→10, d _GPe has a value of 12→12, d The value of _GPi is 12→4, the value of d _PY is 4, and the value of d _IN is 8. 4. The learning method according to claim 3, characterized in that synaptic delay time is introduced in the MATLAB basal ganglia-thalamus-cortex neuron network (7), and the delay parameter value is: τ _GPe→STN =6ms τGPe _→GPi =6ms τGPi _→Th =6ms _τTh→PY =6ms τ _PY→STN =6ms τ _STN→GPe =2ms τ _STN→GPi =2ms τGPe _→GPe =2ms τPY _→IN =2ms τ _IN→PY =2ms.