An AI-powered intelligent control system for treating mental illness based on multimodal microwaves

Abstract

The present invention discloses an AI intelligent control system for treating mental illness based on multimodal microwaves, comprising: a data acquisition module for collecting multimodal microwave data, wherein the multimodal microwave data includes basic microwave data, neural pathway activity state data, and individual physiological data; an optimization and control module for extracting optimal control data from the multimodal microwave data through an improved deep convolutional neural network; a dynamic strategy module for collecting historical microwave stimulation data, and obtaining optimal microwave stimulation parameters for neural activity in specific brain regions by learning the nonlinear relationship between the optimized control data and microwave stimulation parameters through a hybrid transfer learning model; a treatment execution module for controlling the microwave emission equipment based on the optimal microwave stimulation parameters, emitting microwaves to the corresponding brain regions of the patient for treatment, and generating a three-dimensional visualization report of the treatment progress, dynamically adjusting the model parameters, realizing real-time optimization of the treatment parameters, and overcoming the limitations of the static parameters of the traditional method.

Description

Technical Field

[0001] The present invention relates to the field of microwave intelligent control technology, and in particular to an AI intelligent control mental illness treatment system based on multimodal microwaves. Background Art

[0002] In the current application of mental illness treatment, microwave therapy, as a non-invasive neuromodulation technology, has been initially used to intervene in diseases such as depression and anxiety. The occurrence of mental illness is closely related to abnormal activity of brain neural networks (such as dysfunction of the prefrontal lobe and limbic system), but there are significant differences in neural pathway activity and brain functional connectivity among different patients, and personalized stimulation parameters are required to restore neural balance.

[0003] In the existing microwave-controlled mental illness treatment system, although microwave therapy has been applied, there are problems such as single stimulation parameters, lack of personalization and dynamic adjustment mechanism. Traditional microwave neuromodulation is only based on frequency and intensity, and does not consider the microstructure of the brain area, the activity of neural pathways, the patient's physiological rhythm and psychosocial factors, making it difficult to provide accurate treatment. Therefore, an AI intelligent control mental illness treatment system based on multimodal microwaves is proposed here. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art and to achieve the above-mentioned objectives, the present invention proposes the following technical solutions:

[0005] An AI-powered intelligent control system for treating mental illness based on multimodal microwaves, comprising:

[0006] Data acquisition module: collects multimodal microwave data, including basic microwave data, neural pathway activity status data and individual physiological data;

[0007] Optimization and control module: extracts the optimal control data from multimodal microwave data through an improved deep convolutional neural network;

[0008] Among them, the improved deep convolutional neural network is achieved by adding residual connections between two consecutive convolutional layers in the basic deep convolutional neural network and introducing a self-optimized learning scalar in the residual connection process, and using void convolution to replace ordinary convolution.

[0009] Dynamic Strategy Module: This module collects historical microwave stimulation data and uses a hybrid transfer learning model to learn and optimize the nonlinear relationship between control data and microwave stimulation parameters to obtain the optimal microwave stimulation parameters for neural activity in specific brain regions.

[0010] Among them, the hybrid online transfer learning model is implemented based on a dual-channel neural network to increase and improve the online incremental learning process.

[0011] Treatment execution module: Based on the optimal microwave stimulation parameters, it controls the microwave emission equipment, transmits microwaves to the corresponding brain areas of the patient for treatment and generates a three-dimensional visual report of the treatment progress.

[0012] The implementation process of adding residual connection between two consecutive convolutional layers is as follows:

[0013] Taking multimodal microwave data as input data, two layers of convolutional layer mapping output are obtained after two consecutive convolutional layer operations;

[0014] Based on the mapping output of the two convolutional layers, the multimodal microwave data is added to the output mapped by the two convolutional layers and the individual data points and weight sets are fused to realize the residual connection process.

[0015] The specific implementation process of the improved deep convolutional neural network is as follows:

[0016] Preset a learning scalar and assign an initial learning scalar to the preset learning scalar;

[0017] The initial learning scalar is introduced into the residual connection process to obtain the residual connection output of the introduced initial learning scalar;

[0018] Based on the residual connection output of the initial learning scalar, a true output label is preset, and the loss value between the residual connection output of the initial learning scalar and the true output label is obtained through a regression loss function, and the gradient of the loss value with respect to the self-optimized learning scalar is obtained;

[0019] Based on the gradient of the loss value with respect to the self-optimized learning scalar, a Lookahead optimizer is used to update the self-optimized learning scalar to the preset maximum condition;

[0020] A self-optimizing learning scalar is introduced into the residual connection of the basic deep convolutional neural network, and the void convolution is used to replace the ordinary convolution to obtain an improved deep convolutional neural network.

[0021] The historical microwave stimulation data includes microwave parameter frequency, phase, treatment time point and corresponding treatment effect evaluation data.

[0022] The hybrid transfer learning model training process is as follows:

[0023] Preset an initial hybrid transfer learning model, convert the optimal control data and historical microwave stimulation data into a matrix form and then perform dimensionality reduction, and input the reduced dimensionality historical microwave stimulation data matrix and current optimized control data into the preset initial hybrid transfer learning model respectively;

[0024] Based on the initial hybrid transfer learning model, an improved online incremental learning process is added to obtain the final hybrid transfer learning model;

[0025] The final hybrid transfer learning model is trained using the mini-batch gradient descent method, and the loss value is calculated based on the loss function. The gradient of the loss function with respect to the model parameters is calculated through the back propagation algorithm and the parameters are updated until the loss function meets the preset conditions and converges to obtain the trained hybrid transfer learning model.

[0026] The initial hybrid transfer learning model is implemented based on a dual-channel neural network, which includes a hybrid feature extractor, and the hybrid feature extractor includes an online channel and a transfer channel.

[0027] The improved online incremental learning process is implemented as follows:

[0028] The improved online incremental learning is implemented by introducing an optimized feedback gain coefficient based on sliding window.

[0029] The process of obtaining the optimized feedback gain coefficient is as follows:

[0030] An initial feedback gain coefficient is set based on prior knowledge;

[0031] Through a reinforcement learning approach, the model's adaptation process to the newly input optimal control data and historical microwave stimulation data is regarded as a decision-making process, and the initial feedback gain coefficient is taken as the decision-making action. According to the model's processing effect on the newly input optimal control data and historical microwave stimulation data under different initial feedback gain coefficient values, rewards or penalties are given until convergence, and finally an optimal optimized feedback gain coefficient is obtained.

[0032] The present invention has the following beneficial effects:

[0033] In this invention, first, by introducing residual connections between consecutive convolutional layers, the gradient vanishing problem of deep networks is solved. At the same time, the Lookahead optimizer dynamically adjusts the learning scalar, adaptively balances the fusion weights of input data and convolution features, and improves the efficiency of complex feature extraction. Combined with the dilated convolution to expand the receptive field, it captures multi-scale spatial features without increasing parameters, making it more suitable for heterogeneous feature processing of multimodal data.

[0034] Secondly, by introducing an online channel (a time-recursive structure) to capture dynamic changes in data, and a migration channel (domain-invariant projection) to achieve knowledge transfer between historical data and current patient data, we can effectively leverage past treatment experience and reduce the training cost for new patients. Then, based on the sliding window and the instant feedback gain coefficient of reinforcement learning, we can respond to new data in real time (such as changes in neural activity during treatment), dynamically adjust model parameters, and achieve real-time optimization of treatment parameters, thus overcoming the limitations of static parameters in traditional methods.

[0035] Finally, through small-batch gradient descent and back-propagation, the complex mapping between multimodal data and microwave stimulation parameters (frequency, phase, treatment time) is accurately learned, supporting the generation of personalized parameters for specific brain regions. The treatment execution module combines EEG signals, hemoglobin concentration changes and microwave parameters to generate a dynamic three-dimensional visualization report, intuitively presenting the relationship between treatment progress and physiological indicators, assisting doctors in real-time evaluation of treatment efficacy and adjustment of plans. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 This is a system block diagram of the AI ​​intelligent control mental illness treatment system based on multimodal microwaves proposed by the present invention. DETAILED DESCRIPTION

[0037] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0038] Example 1

[0039] like Figure 1 As shown, the present invention proposes an AI intelligent control mental illness treatment system based on multimodal microwaves, comprising:

[0040] Data acquisition module: collects multimodal microwave data, including basic microwave data, neural pathway activity status data and individual physiological data;

[0041] Basic microwave data acquisition:

[0042] Basic microwave data is collected through microwave sensors, such as superconducting nanowire-based single-photon detectors, which use the principle that superconducting nanowires absorb microwave photons and then transition from a superconducting state to a normal state to generate electrical signals, and are expressed as ;

[0043] Neural pathway activity status data collection:

[0044] The neural pathway activity status data is collected through functional near-infrared spectroscopy (fNIRS) equipment. Based on the absorption characteristics of biological tissues for near-infrared light, the oxygenation and deoxygenation status of hemoglobin are detected to reflect neural activity. Near-infrared light emitters and detectors are placed on the patient's scalp. The changes in hemoglobin concentration are calculated by light intensity attenuation, thereby obtaining the neural pathway activity status. The formula for calculating hemoglobin concentration changes by light intensity attenuation is expressed as:

[0045]

[0046] in, is the absorbance, is the incident light intensity, is the outgoing light intensity, is the molar absorptivity, is the optical path, is the change in hemoglobin concentration;

[0047] Calculate the change in hemoglobin concentration by light intensity attenuation, and then obtain the activity status data of the neural pathway ;

[0048] Personalized physiological data collection:

[0049] Personalized physiological data uses a photoelectric heart rate sensor based on photoplethysmography (PPG). When light is irradiated on the skin, the blood volume in the blood vessels changes with the heartbeat, causing changes in light absorption and scattering. The detected pulse wave signal is converted into an electrical signal and expressed as ;

[0050] Multimodal microwave data is represented as ;

[0051] Multimodal microwave data The scales of each feature dimension vary greatly, which affects the basic microwave data collected. , neural pathway activity status data and personal physiological data Normalization is performed to make the data scale uniform.

[0052] Optimization and control module: extracts the optimal control data from multimodal microwave data through an improved deep convolutional neural network;

[0053] The optimal control data acquisition process is:

[0054] Optimize and improve the convolution layer of the basic deep convolutional neural network, which includes an input layer, two consecutive convolution layers, a pooling layer, and an output layer;

[0055] Add residual connections between two consecutive convolutional layers in the basic deep convolutional neural network to transform multimodal microwave data As input data is received through the input layer, it is mapped by the first convolutional layer to obtain , Represents the basic operation of the first convolutional layer, Map the output of a single convolutional layer, and then map it through a second convolutional layer to obtain , is the basic operation of the second convolutional layer, The output is mapped through two convolutional layers;

[0056] Based on the mapping output of two convolutional layers, multimodal microwave data is directly converted to Directly and through convolutional layer mapping The outputs of the multi-modal microwave data are added and fused ( (middle) Individual data points and weights, the output after residual connection is expressed as: ,in, Represents the mapping relationship after two convolutional layers, Multimodal microwave data Individual data points in As weight, multimodal microwave data Directly and through convolutional layer mapping The output of the convolutional layer is added together, so that the network can learn the difference between the input and the output of the convolutional layer, making the network training process easier and extracting the complex spatial features in the multimodal microwave data more deeply;

[0057] Introducing a self-optimizing learning scalar in the residual connection process , the self-optimizing learning scalar optimization process is:

[0058] Preset a learning scalar , giving the scalar Assign an initial learning scalar (e.g. 0.5), for the initial learning scalar Carry out optimization updates;

[0059] The initial learning scalar In the process of introducing residual connection, for the input multimodal microwave data , after the mapping relationship of two convolutional layers Get the convolution layer output, at the same time, the input multimodal microwave data Together with the convolutional layer output, it participates in the residual connection calculation. At this time, the residual connection output is:

[0060]

[0061] To introduce the residual connection output of the initial learning scalar;

[0062] The output value will continue to propagate forward in the network, participating in the calculation of subsequent layers and generating the final output result;

[0063] Based on the residual connection output of the initial learning scalar , preset a true output label, and obtain the residual connection output of the initial learning scalar through a regression loss function The loss value between the actual output label , loss value Backward propagation is performed in the network through the backpropagation algorithm. During the backpropagation process, the loss value is calculated according to the chain derivative rule. On Self-Optimizing Learning Scalars The gradient of , the gradient value reflects The direction and extent of the impact of changes in the loss value;

[0064] Specifically, the true output label is obtained by analyzing historical data. A large amount of historical data related to multimodal microwave therapy is collected, including microwave stimulation parameters and patient feedback data before and after treatment. Statistical analysis methods are used to find parameter combinations that are highly correlated with treatment effects as the true output label.

[0065] According to the calculated gradient , use a Lookahead optimizer to update, and its update formula is:

[0066]

[0067] in Lookahead optimizer learning rate, used to update parameters The value of Before the update Value, updated The value is recorded as , used for the next round of forward propagation and subsequent calculations

[0068] As the training process progresses, the forward propagation, loss calculation, backpropagation, and parameter update steps are repeated until the preset maximum condition is reached, for example:

[0069] Set a loss threshold and the number of consecutive iterations , when continuous In the iterations, the decrease in the loss value L is less than At this time, the obtained That is the optimal learning scalar;

[0070] Specific, different Affects the effect of residual connection. Approaching 0, middle, Output The contribution of becomes smaller, which means that the network relies more on the mapping of the convolutional layer itself , focusing on direct transformation and feature extraction of input data, when When it approaches 1, the residual connection is close to the traditional method, emphasizing the input of multimodal microwave data With the convolutional layer output Difference learning between

[0071] At the same time, in the convolution layer, the hole convolution is used to replace the ordinary convolution. In the hole convolution, for the input data , convolution kernel , the void rate is , the i-th element of the output is expressed as:

[0072]

[0073] in, represents the jth convolution kernel, is the maximum number of convolution kernels;

[0074] The preprocessed multimodal microwave data is input into the optimized deep neural network. The improved convolution layer adaptively optimizes the spatial features of different scales. Then, the output of the improved convolution layer is downsampled through the pooling layer to retain the optimal features. The final formula is expressed as:

[0075]

[0076] Finally, a series of feature data is output through the output layer That is, the optimal control data;

[0077] Specifically, we introduce the self-optimizing learning scalar Finally, the deep neural network can automatically adjust the residual connection according to the characteristics of the input data and the needs of the current task. For example, for some multimodal microwave data with complex features that require deep transformation by convolutional layers, the deep neural network can adjust the residual connection automatically. By making it close to 0, the interference information entering the network through the residual connection in the complex features is also reduced, making the features output by the convolution layer purer. For some features similar to historical data, the deep neural network adjustment When the value approaches 1 and further learns the difference information, when analyzing the changes in the activity state data of the neural pathway before and after microwave stimulation, by emphasizing this difference learning, the network can more keenly capture the changes in the activity state of the neural pathway caused by microwave stimulation, thereby extracting features that are closely related to the treatment effect.

[0078] Dynamic Strategy Module: This module collects historical microwave stimulation data and uses a hybrid transfer learning model to learn and optimize the nonlinear relationship between control data and microwave stimulation parameters to obtain the optimal microwave stimulation parameters for neural activity in specific brain regions.

[0079] Collect past microwave stimulation related data, including microwave parameter frequency during each microwave treatment , Phase , treatment time point T, and corresponding treatment effect evaluation data , normalize and store these data to form historical microwave stimulation data ;

[0080] Preset an initial hybrid transfer learning model to optimally control data and historical microwave stimulation data Convert to matrix form and then reduce the dimension, and the reduced dimension historical microwave stimulation data matrix and current optimization and control data Input them into the preset initial hybrid transfer learning model respectively;

[0081] The initial hybrid transfer learning model is implemented based on a dual-channel neural network and is expressed as:

[0082]

[0083] in, is the input result of the initial hybrid transfer learning model for the input data, represents the bias term of the two-channel neural network, represents the weight matrix of the two-channel neural network, Represents a hybrid feature extractor The historical microwave stimulation data matrix after dimensionality reduction and current optimization and control data Perform feature extraction;

[0084] Further, Including online channels and migration channels :

[0085] Online Channel It is a time-recursive mechanism that can continuously capture the dynamic changes of data over time. When recording the effect of microwave stimulation on neural activity over time, it can obtain the fluctuations and trends of the data at each time point.

[0086] Migration Channel To introduce a domain-invariant projection mechanism to enable knowledge transfer between data in different domains;

[0087] Add an improved online incremental learning process based on the initial hybrid transfer learning model to make the model more adaptable to new optimal control data and historical microwave stimulation data , the implementation process of improving the online incremental learning process is:

[0088] Improved online incremental learning is implemented by introducing an optimized feedback gain coefficient based on a sliding window. The process of obtaining the optimized feedback gain coefficient is as follows:

[0089] Setting the initial feedback gain coefficient based on prior knowledge , such as, based on prior knowledge, discovering When the model is in a certain range, it can respond well to new data, and a value can be selected within this range as the initial feedback gain coefficient;

[0090] Through a reinforcement learning idea, the model is trained to respond to new data (optimal control data of new input) and historical microwave stimulation data ) is considered as a decision-making process. The value of is set as the decision action, according to the model in different The processing effect of new data (the degree of improvement in prediction accuracy) is rewarded or punished until convergence. (The convergence condition is that the model gradually finds the most suitable value to achieve the best response to new data), and finally obtain an optimal feedback gain coefficient ;

[0091] Specifically, the basis of the reinforcement learning idea is that reinforcement learning is good at dealing with dynamically changing environments. Integrating reinforcement learning into this specific scenario innovatively, new data (optimal control data) and historical microwave stimulation data ) is continuously input, and the model needs to dynamically adapt to these new data. Reinforcement learning can continuously adjust its decision-making based on the changes in the environment brought about by the new data (changes in the model's processing effect of new data). Values ​​are taken to allow the model to better cope with new data;

[0092] Based on the optimal feedback gain coefficient When new data is input, the process of improving the online incremental learning process is expressed as:

[0093]

[0094] in, represents the learning rate, represents the parameters of the model, represents the new parameters of the model, Represents the loss function of the model About model parameters The gradient, Represents the newly input data (newly input optimal control data and historical microwave stimulation data ), Represents the model prediction results, The feature matrix representing the new input data;

[0095] Learning rate A learning rate decay strategy is used. A relatively large learning rate (0.1) is initially set. During training, the learning rate is reduced by a certain ratio (multiplied by 0.9) after every specific number of rounds (10 rounds). This allows the model to converge quickly in the early stages of training and fine-tune parameters in the later stages.

[0096] Specifically, by introducing a feedback gain coefficient in the sliding window to achieve improved online incremental learning, the model can quickly respond to the difference between new data and predicted values. If the predicted value deviates significantly from the actual value of the new data, the feedback gain coefficient will cause a significant adjustment to the model parameters, allowing the model to quickly adapt to the characteristics of the new data.

[0097] Then, the final hybrid transfer learning model is expressed as:

[0098]

[0099] in, The final hybrid transfer learning model is combined with the output of improved online incremental learning;

[0100] Finally, the final hybrid transfer learning model is trained using the mini-batch gradient descent method. Each time, a mini-batch of data is taken from the training set and input into the model to calculate the output. , and then calculate the loss value according to the loss function L, and calculate the loss function about the model parameters through the back propagation algorithm Gradient , and according to the gradient update rule The parameters are updated and this process is repeated until the loss function meets the preset conditions (the decrease in the loss value is less than a certain threshold after several consecutive rounds of training) and converges. After the training is completed, the final hybrid transfer learning model outputs a series of data, which is the optimal microwave stimulation parameters for neural activity in a specific brain area obtained by learning the nonlinear relationship between the optimized control data and the microwave stimulation parameters;

[0101] Specifically, the final hybrid transfer learning model output represents the output value of the microwave stimulation parameters given by the model after comprehensively considering historical microwave stimulation data and current optimized control data. This value is a quantitative representation of the optimal microwave stimulation parameters for neural activity in a specific brain region, where:

[0102] Nonlinear relationship expression: through hybrid feature extractor , which includes online channels and migration channels The online channel captures the dynamic changes of data over time, and the migration channel realizes the knowledge migration of data in different domains. The combination of the two can extract multiple features from complex data, and then characterize the complex nonlinear relationship between the optimization control data and microwave stimulation parameters, rather than a simple linear mapping;

[0103] Obtaining optimal parameters: The model is based on input data and uses a dual-channel neural network structure for calculation. Then, through a specific improved online incremental learning process, the model parameters are continuously adjusted according to the difference between the new data and the predicted value. (Updated to ), after multiple iterative learning, the model output continuously approaches the optimal microwave stimulation parameters for neural activity in a specific brain area, and finally outputs That is, the optimal microwave stimulation parameters.

[0104] Treatment execution module: Based on the optimal microwave stimulation parameters, it controls the microwave emission equipment, transmits microwaves to the corresponding brain areas of the patient for treatment, and generates a three-dimensional visualization report of the treatment progress;

[0105] The relevant data of the new patient (multimodal microwave data) is input into the trained model to obtain the optimal microwave stimulation parameters. The optimal microwave stimulation parameters include the frequency, phase frequency and treatment time that best suit the patient's data. For example:

[0106] frequency (denormalized value), phase =180°, treatment time minute;

[0107] Treatment operation: The treatment execution module uses Python's Socket library to establish a wired network connection with the microwave transmitting device (assuming the device supports TCP / IP protocol), sends control instructions to the device, and sets its output frequency to , the phase is 180°. During the treatment process, physiological indicators such as the patient's EEG signals and microwave treatment equipment parameters (such as frequency, phase, power, etc.) are collected in real time through physiological monitoring equipment such as EEG monitors. The Python Plotly visualization library is used to construct a three-dimensional coordinate system with the time axis as one dimension and the hemoglobin concentration change value and microwave treatment equipment parameters as the other two dimensions. The data at each time point are plotted as data points in the coordinate system, and the Plotly animation function is used to display the dynamic changes of data points driven by time, generate a three-dimensional visualization report of the treatment progress, and intuitively present the parameter change trends and their relationships.

[0108] In the application, several formulas involved are calculated by taking their numerical values ​​after removing the dimensions, and the formulas are established by collecting a large amount of data and performing software simulation to obtain a formula for the most recent real situation. Some coefficients or weights in the formulas are set by technical personnel in this field according to actual conditions, so they will not be elaborated here.

[0109] The above embodiments can be implemented in whole or in part by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. Those skilled in the art will appreciate that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution.

[0110] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. An AI intelligent control system for treating mental illness based on multimodal microwaves, characterized in that: include: Data acquisition module: collects multimodal microwave data, including basic microwave data, neural pathway activity status data and individual physiological data; Optimization and control module: extracts the optimal control data from multimodal microwave data through an improved deep convolutional neural network; Among them, the improved deep convolutional neural network is achieved by adding residual connections between two consecutive convolutional layers in the basic deep convolutional neural network and introducing a self-optimized learning scalar in the residual connection process, and using void convolution to replace ordinary convolution. The implementation process of adding residual connection between two consecutive convolutional layers is as follows: Taking multimodal microwave data as input data, two layers of convolutional layer mapping output are obtained after two consecutive convolutional layer operations; Based on the output of the two-layer convolutional layer mapping, the multimodal microwave data is added to the output of the two-layer convolutional layer mapping and the individual data points and weight sets are fused to realize the residual connection process; The specific implementation process of the improved deep convolutional neural network is as follows: Preset a learning scalar and assign an initial learning scalar to the preset learning scalar; The initial learning scalar is introduced into the residual connection process to obtain the residual connection output of the introduced initial learning scalar; Based on the residual connection output of the initial learning scalar, a true output label is preset, and the loss value between the residual connection output of the initial learning scalar and the true output label is obtained through a regression loss function, and the gradient of the loss value with respect to the self-optimized learning scalar is obtained; Based on the gradient of the loss value with respect to the self-optimized learning scalar, a Lookahead optimizer is used to update the self-optimized learning scalar to the preset maximum condition; A self-optimized learning scalar is introduced into the residual connection of the basic deep convolutional neural network, and the hole convolution is used to replace the ordinary convolution to obtain an improved deep convolutional neural network; Dynamic Strategy Module: This module collects historical microwave stimulation data and uses a hybrid transfer learning model to learn and optimize the nonlinear relationship between control data and microwave stimulation parameters to obtain the optimal microwave stimulation parameters for neural activity in specific brain regions. Among them, the hybrid online transfer learning model is implemented based on a dual-channel neural network to increase and improve the online incremental learning process; Treatment execution module: Based on the optimal microwave stimulation parameters, it controls the microwave emission equipment, transmits microwaves to the corresponding brain areas of the patient for treatment and generates a three-dimensional visual report of the treatment progress.

2. The AI ​​intelligent control mental illness treatment system based on multimodal microwaves according to claim 1 is characterized in that: The historical microwave stimulation data includes microwave parameter frequency, phase, treatment time point and corresponding treatment effect evaluation data.

3. The AI ​​intelligent control mental illness treatment system based on multimodal microwaves according to claim 1 is characterized in that: The hybrid transfer learning model training process is as follows: Preset an initial hybrid transfer learning model, convert the optimal control data and historical microwave stimulation data into a matrix form and then perform dimensionality reduction, and input the reduced dimensionality historical microwave stimulation data matrix and current optimized control data into the preset initial hybrid transfer learning model respectively; Based on the initial hybrid transfer learning model, an improved online incremental learning process is added to obtain the final hybrid transfer learning model; The final hybrid transfer learning model is trained using the mini-batch gradient descent method, and the loss value is calculated based on the loss function. The gradient of the loss function with respect to the model parameters is calculated through the back propagation algorithm and the parameters are updated until the loss function meets the preset conditions and converges to obtain the trained hybrid transfer learning model.

4. The AI ​​intelligent control mental illness treatment system based on multimodal microwaves according to claim 3 is characterized in that: The initial hybrid transfer learning model is implemented based on a dual-channel neural network, which includes a hybrid feature extractor, and the hybrid feature extractor includes an online channel and a transfer channel.

5. The AI ​​intelligent control mental illness treatment system based on multimodal microwaves according to claim 1 is characterized in that: The improved online incremental learning process is implemented as follows: The improved online incremental learning is implemented by introducing an optimized feedback gain coefficient based on sliding window.

6. The AI ​​intelligent control mental illness treatment system based on multimodal microwaves according to claim 5 is characterized in that: The process of obtaining the optimized feedback gain coefficient is as follows: An initial feedback gain coefficient is set based on prior knowledge; Through a reinforcement learning approach, the model's adaptation process to the newly input optimal control data and historical microwave stimulation data is regarded as a decision-making process, and the initial feedback gain coefficient is taken as the decision-making action. According to the model's processing effect on the newly input optimal control data and historical microwave stimulation data under different initial feedback gain coefficient values, rewards or penalties are given until convergence, and finally an optimal optimized feedback gain coefficient is obtained.

Images