A mental health monitoring system based on quantum personalized federated learning

Abstract

The application discloses a mental health monitoring system based on quantum personalized federated learning, and the system comprises the following modules: a data acquisition and initialization module, which acquires user voice data and initializes global and local parameters; a global geometric structure guided data enhancement module, which generates enhanced samples by using anisotropic disturbance constructed by global covariance matrix characteristic decomposition to relieve data heterogeneity; a time modulation quantum gate unit module, which dynamically adjusts quantum state evolution by time modulation offset to realize individualized emotion dynamic representation; a parameter optimization module, which updates parameters by using quantum-classical hybrid training mode; a partial parameter uploading module, which uploads only shared Backbone parameters while retaining personalized classification head parameters; and a federated aggregation module, which performs weighted fusion on shared parameters and returns them. The application combines quantum computing and personalized federated learning, significantly improves model recognition accuracy and cross-client adaptability under the premise of protecting data privacy.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and intelligent medical technology, specifically relating to a mental health monitoring system based on quantum personalized federated learning. Background Technology

[0002] With the integrated development of technologies such as 5G, artificial intelligence, and edge computing, mental health monitoring for long-term individual status assessment is gradually becoming an important research direction in the fields of smart healthcare and digital health. In this intelligent health ecosystem, terminal devices not only undertake the function of voice data collection, but are also gradually evolving into key nodes integrating information perception, edge computing, and privacy protection. Compared with traditional methods that rely on questionnaires or intermittent assessments, continuous perception based on voice signals can achieve long-term dynamic monitoring of an individual's emotional state and psychological changes without interfering with the user's daily life.

[0003] Therefore, constructing a high-precision, low-latency, and secure mental health monitoring model has become a key issue in promoting the implementation of intelligent mental health services. In this context, modeling individual multi-source sensory data and their dynamic states as corresponding digital twins can enable continuous characterization and predictive analysis of individual psychological states. Summary of the Invention

[0004] The purpose of this invention is to overcome the performance degradation caused by data heterogeneity and insufficient dynamic modeling capabilities across time scales in existing federated methods for constructing digital twin systems. It provides a mental health monitoring system based on quantum personalized federated learning. This system preserves individual differences while sharing global temporal representations and exhibits strong robustness under various quantum noise environments, offering a novel solution that integrates quantum computing intelligence for federated digital twin systems in distributed healthcare scenarios.

[0005] To address the above technical problems, this invention provides the following technical solution: a mental health monitoring system based on quantum personalized federated learning, comprising:

[0006] The data acquisition and initialization module, configured on each client, is used to collect users' voice data through terminal devices and perform data storage and preprocessing locally; the server is used to initialize global model parameters, including initializing the quantum parameters of the shared feature extractor. and classic parameters And the personalized classifier parameters for each client. and initialize the parameters Distribute it to each client as the initial model;

[0007] A global geometry-guided data augmentation module is configured on each client to construct augmented samples that approximate the global data distribution on local clients using an uncertainty augmentation method guided by the global geometry. The generation of local samples is guided by global geometric priors to alleviate the problem of data distribution heterogeneity in federated learning environments.

[0008] The time-modulated quantum gate unit module is configured on each client and is used to input the enhanced speech features into the time-modulated quantum gate unit model during the local training phase. The input features are first processed by the classical feature mapping layer for representation learning, and then encoded into the variable quantum circuit. The high-dimensional quantum feature representation is obtained through quantum gate operations and quantum measurement.

[0009] The parameter optimization module, configured on each client, calculates the weighted cross-entropy loss function based on the model prediction results and updates the model parameters using the gradient descent algorithm. Specifically, the shared feature extractor parameters and the personalized classifier parameters are optimized using the backpropagation algorithm. , The quantum circuit parameters are updated by calculating the gradient using parameter shift rules. ;

[0010] A partial parameter upload module, configured on each client, is used to upload only the locally shared feature extractor parameters after local training is completed on the client side. Send to the server, while retaining local personalized classifier parameters. It does not participate in aggregation in order to achieve personalized emotion modeling for individual users;

[0011] The federated aggregation module, configured on the server side, receives shared parameters uploaded by each client and uses a federated aggregation strategy to weight and fuse the shared feature extractor parameters to obtain new globally shared parameters. The updated global parameters are then redistributed to each client for the next round of federated training, until the preset number of communication rounds is reached or the model meets the convergence condition.

[0012] Furthermore, the aforementioned data augmentation module guided by global geometry specifically includes:

[0013] The local statistical calculation unit is used to calculate the local mean vector and local covariance matrix of the samples for each sentiment category c in client k.

[0014] The global statistical aggregation unit is used to upload local category statistics from each client to the server. The server then aggregates the statistics from all clients to obtain the global mean vector for category c. With the global covariance matrix :

[0015] ,

[0016] ,

[0017] in, This represents the number of samples of category c in client k. This represents the total number of samples of category c across all clients;

[0018] Eigenvalue decomposition unit, used for the global covariance matrix Eigenvalue decomposition is performed as follows:

[0019] in, Represents the eigenvalue matrix. This represents the corresponding eigenvector matrix;

[0020] Geometric perturbation building blocks are used to transfer global geometry information on the server. , After being distributed to each client, for any sample x of category c in client k, random coefficients are sampled from the standard normal distribution. And construct the geometric perturbation term based on the eigenvalue decomposition results of the global covariance matrix.

[0021] Furthermore, the aforementioned time-modulated quantum gating unit module specifically includes: the parameter optimization module specifically includes:

[0022] The category weight construction unit is used to input the feature representation output by TM-QMGU into the personalized classifier to obtain the sentiment category prediction result, and to construct the category weight coefficients according to the number of samples in each category. Where N is the total number of training samples and C is the total number of classes. The number of samples in category c;

[0023] The weighted loss calculation unit is used to calculate the error between the model's prediction results and the true labels using the weighted cross-entropy loss function. And construct the optimization objective as follows:

[0024] ;

[0025] The quantum-classical hybrid training unit is used to update model parameters using a quantum-classical hybrid training approach. Classical network parameters are updated through the backpropagation algorithm, while quantum circuit parameters are optimized by calculating gradients using parameter shift rules.

[0026] Dimension alignment unit, containing a first linear transformation layer Second linear transformation layer The first linear transformation layer is used to align the dimension of the input data with the input dimension of the VQC module, and the second linear transformation layer is used to align the output dimension of the VQC module with the input dimension of the subsequent classical operation.

[0027] Furthermore, the aforementioned federated aggregation module specifically includes: a federated average aggregation unit, used by the server to receive shared Backbone parameters uploaded by each client { , The global model parameters are updated using a federated average aggregation method.

[0028] ,

[0029] ,

[0030] Where K represents the number of clients participating in the training;

[0031] The parameter distribution unit is used by the server to send the updated globally shared parameters after aggregation is complete. , The data is redistributed to each client, but the personalized category header parameters are not included in the aggregation, thus satisfying the requirement. ;

[0032] The classification head fast calibration unit is used to freeze the backbone parameters after receiving new global parameters, and only perform fast calibration updates on the classification head parameters, as follows: ,in, Represents the loss function. Indicates the learning rate. Indicates the category header, This represents the gradient of the loss function with respect to the parameters of the k-th client classification head.

[0033] Compared with the prior art, the beneficial technical effects of the present invention using the above technical solution are as follows:

[0034] This invention provides a mental health monitoring model based on quantum personalized federated learning (PBA). 1) This model, by integrating geometrically guided data augmentation and partial parameter aggregation strategies, can effectively improve the model's adaptability and generalization performance to heterogeneous data while protecting user privacy. 2) A TM-QMGU model is designed to effectively process digital twin data. Compared with mainstream quantum recurrent neural networks such as QLSTM and QGRU, TM-QMGU significantly reduces the number of trainable parameters while maintaining expressive power through the collaborative design of simplified gating structures and parameterized quantum circuits. Furthermore, by integrating digital twin technology, TM-QMGU can maintain stable classification accuracy and discriminative ability throughout its entire lifecycle. 3) A partial parameter aggregation method is designed. In the federated aggregation process, this method only extracts the backbone parameters of shared temporal features for FedAvg aggregation, while retaining the classification head parameters of each client for local personalized modeling, thereby decoupling shared representation and personalized decision-making at the parameter level. Compared with traditional full-parameter aggregation methods, PBA effectively alleviates the model shift and negative transfer problems caused by data distribution differences and reduces performance degradation caused by cross-client interference. In summary, this invention can preserve individual differences while sharing global temporal representations, and still exhibits strong robustness under various quantum noise environments, providing a new solution that integrates quantum computing intelligence for federated digital twin systems for distributed medical scenarios. Attached Figure Description

[0035] Figure 1 A diagram of a mental health monitoring model based on quantum personalized federated learning;

[0036] Figure 2 Data flow and operation graph in the quantum minimum gate unit;

[0037] Figure 3 A quantum circuit diagram of the quantum minimum gate unit;

[0038] Figure 4 A quantum-classical structure diagram in the quantum minimum gate unit; Detailed Implementation

[0039] To better understand the technical content of the present invention, specific embodiments are described below in conjunction with the accompanying drawings.

[0040] In this invention, various aspects of the invention are described with reference to the accompanying drawings, in which numerous illustrative embodiments are shown. Embodiments of the invention are not limited to those depicted in the drawings. It should be understood that the invention is implemented through any of the various concepts and embodiments described above, as well as the concepts and embodiments described in detail below, because the concepts and embodiments disclosed herein are not limited to any particular implementation. Furthermore, some aspects of the invention disclosed may be used alone or in any suitable combination with other aspects of the invention disclosed.

[0041] An example provides a mental health monitoring system based on quantum personalized federated learning, comprising:

[0042] The data acquisition and initialization module, configured on each client, is used to collect users' voice data through terminal devices and perform local data storage and preprocessing. Different clients collect users' voice data through terminals such as mobile phones and wearable devices, and perform local data storage and preprocessing. The server initializes global model parameters, including shared feature extractor (Backbone) parameters. And personalized classifier parameters for each client. and will share parameters It is distributed to each client as the initial model. For example... Figure 1 The model architecture of this invention is shown, wherein the main body of the model includes two stages: a client-side local training stage and a server aggregation stage. In the client-side local training stage, the TM-QMGU model is trained on the patient's digital twin data, with the number of qubits in the MT-QMGU set to 4, the learning rate to 0.001, the batch size to 128, and the epochs to 200.

[0043] A global geometry-guided data augmentation module is configured on each client to construct augmented samples that approximate the global data distribution on local clients using an uncertainty augmentation method guided by the global geometry. The generation of local samples is guided by global geometric priors to alleviate the problem of data distribution heterogeneity in federated learning environments.

[0044] This module is configured to perform the following steps:

[0045] Step 2.1 Assume there are a total The client, the first The local training data of each client is represented as follows: .in, This represents the input speech temporal feature vector (composed of multiple frames of acoustic features). These are the sentiment category labels. For each sentiment category c in client k, calculate the local mean vector and local covariance matrix of the samples in that category:

[0046]

[0047]

[0048] in, This represents the set of samples belonging to category c in client k;

[0049] Step 2.2 Each client uploads its local category statistics to the server. The server aggregates the statistics from all clients to obtain the global mean vector and global covariance matrix for category c:

[0050]

[0051]

[0052] in, This represents the number of samples of category c in client k. This represents the total number of samples of category c across all clients;

[0053] Step 2.3 For the global covariance matrix Perform eigenvalue decomposition:

[0054]

[0055] in, Represents the eigenvalue matrix. This represents the corresponding eigenvector matrix;

[0056] Step 2.4 The server will transfer the global geometry information { , Distribute to each client;

[0057] Step 2.5 Perform geometry-guided data augmentation on the client's local data based on the global geometry obtained from server-side aggregation. For any sample x of class c in client k, random coefficients are sampled from a standard normal distribution. And construct the geometric perturbation term based on the eigenvalue decomposition results of the global covariance matrix:

[0058]

[0059] in, This represents the m-th feature value corresponding to category c. This represents the corresponding eigenvector, and P represents the number of principal geometric directions selected.

[0060] Step 2.6 Add the geometric perturbation term β to the original sample x to generate the enhanced sample:

[0061]

[0062] This process mitigates the data heterogeneity problem in federated learning by anisotropically perturbing local samples along the global principal geometry, guiding augmented data to align with the global class geometry while maintaining client distribution characteristics.

[0063] The Time-Modulated Quantum Gated Unit (TM-QMGU) module, configured on each client, is used to input enhanced speech features into the TM-QMGU model during the local training phase. The input features are first processed through a classical feature mapping layer for representation learning, and then encoded into a Variable Quantum Circuit (VQC). High-dimensional quantum feature representations are obtained through quantum gate operations and quantum measurements. Figure 2 As shown, it illustrates the data flow and operations within the quantum minimum gate unit. We integrate time-aware modulation into a parallel variable quantum circuit, allowing the quantum state to dynamically shift over time to improve the model's temporal representation capability. The time-modulated quantum gate unit module is configured to perform the following steps:

[0064] Step 3.1 Input the speech feature sequence enhanced by GGEUR into the Time Modulation Quantum Minimum Gated Unit (TM-QMGU) model, assuming the input sequence is... ,in This represents the input feature vector at time t;

[0065] Step 3.2 Calculate the time modulation offset based on the hidden state of the previous moment. This is then incorporated into the calculation process of the update gate and candidate hidden states:

[0066]

[0067]

[0068]

[0069]

[0070] In the formula, Represents the time modulation coefficient. This indicates the hidden state at the previous moment. This represents the average calculation; This represents the input feature vector at the current time t. This represents the Sigmoid activation function. This indicates an update of the gate-to-strain component quantum circuit. This represents the hyperbolic tangent activation function. This represents Hadamard element-wise multiplication. This represents the quantum circuit for calculating the strain component of candidate hidden states;

[0071] Step 3.3 After the variable quantum circuit completes the quantum state evolution, the quantum characteristic representation is obtained through quantum measurement. The local model based on TM-QMGU can be regarded as a function The final output is the feature representation result of the digital twin model:

[0072]

[0073] Where W is the set of parameters in a classic network. It is the VQC module in TM-QMGU. It is a trainable phase angle parameter in quantum circuits.

[0074] TM-QMGU is parameterized by a single-layer quantum circuit, such as Figure 3 As shown, it illustrates the quantum circuit diagram of the quantum minimum gate unit, taking four qubits as an example. Each layer consists of two RX gates and a set of controlled NOT gates on adjacent qubits. The three VQC modules corresponding to the update gate, reset gate, and candidate hidden state calculation module share the same architecture.

[0075] The parameter optimization module, configured on each client, calculates the weighted cross-entropy loss function based on the model prediction results and updates the model parameters using the gradient descent algorithm. Specifically, the shared feature extractor parameters and the personalized classifier parameters are optimized using the backpropagation algorithm. , The quantum circuit parameters are updated by calculating the gradient using the parameter shift rule. The parameter optimization module is configured to perform the following steps:

[0076] Step 4.1 Input the feature representation output by TM-QMGU into the personalized classifier to obtain the sentiment category prediction result. To alleviate the data class imbalance problem, class weight coefficients are constructed based on the number of samples in each category:

[0077]

[0078] Where N is the total number of training samples and C is the total number of classes. The number of samples in category c;

[0079] Step 4.2 Calculate the error between the model's prediction and the true label using the weighted cross-entropy loss function:

[0080]

[0081] And construct optimization objectives:

[0082]

[0083] in, This represents the optimal solution for classical circuit optimization. This represents the optimal solution for quantum circuit optimization.

[0084] Step 4.3 updates the model parameters using a quantum-classical hybrid training approach. Classical network parameters are updated via backpropagation, while quantum circuit parameters are optimized by calculating gradients using parameter shift rules. For example... Figure 4 As shown, it illustrates the quantum-classical structure in TM-QMGU, with the presence of a linear transformation layer. This aligns the dimensions of the input data with the input dimensions of the VQC. Following the quantum circuit, there is a linear transformation layer. This aligns the output dimension of VQC with the input dimension of subsequent classic operations.

[0085] A partial parameter upload module, configured on each client, is used to upload only the locally shared feature extractor parameters after local training is completed on the client side. Send to the server, while retaining local personalized classifier parameters. It does not participate in aggregation in order to achieve personalized emotion modeling for individual users.

[0086] In federated learning's non-IID scenario, the partial parameter upload module often exhibits significant data distribution differences between different clients, particularly in class imbalance and inconsistent decision boundaries. The traditional FedAvg method assumes that each client's data follows an independent and identically distributed (IID) distribution, thus applying a uniform weighted average to all model parameters. However, under heterogeneous data conditions, this "full parameter sharing" strategy leads to two key problems: firstly, the shared classification decision boundary is affected by the mutual interference of different client label distributions, resulting in biases or even conflicts; secondly, client-specific discriminative information is continuously smoothed during multiple rounds of aggregation, ultimately weakening the model's personalized expressive ability and manifesting as insufficient fitting to local data. To alleviate these problems, this invention proposes a personalized aggregation method based on partial parameter decoupling. Specifically, let the complete model parameters be... ,in, and These are the quantum and classical parameters in the backbone (TM-QMGU temporal modeling layer), respectively, used to learn a high-level semantic representation shared across clients; The classification header parameters (hidden2tag linear layer) are used to characterize the client-specific class discrimination boundaries.

[0087] In the At the start of round-robin communication, the server broadcasts global backbone parameters to all clients. The client k retains its own personalized category header parameters. and in local data Execution Round-robin updates, gradient updates take the form of:

[0088]

[0089]

[0090]

[0091] In the formula, This represents the complete model parameters of client k in the r-th round of communication and the e-th local iteration. Indicates the learning rate. , , Let represent the quantum parameters, classical parameters, and personalized classification head parameters of the TM-QMGU time series modeling layer in the r-th round of communication and the e-th local iteration, respectively, and L represent the loss function.

[0092] The federated aggregation module, configured on the server side, receives shared parameters uploaded by each client and uses a federated aggregation strategy to weight and fuse the shared feature extractor parameters to obtain new globally shared parameters. The updated global parameters are then redistributed to each client for the next round of federated training, until the preset number of communication rounds is reached or the model meets the convergence criteria. The federated aggregation module is configured to perform the following steps:

[0093] Step 6.1 The server receives the shared Backbone parameters uploaded by each client. The global model parameters are updated using a federated average aggregation method.

[0094]

[0095]

[0096] Where K represents the number of clients participating in the training;

[0097] Step 6.2 After aggregation is complete, the server will update the globally shared parameters { , The data is redistributed to each client, but the personalized category header parameters are not included in the aggregation, satisfying the following condition:

[0098]

[0099] This design structurally avoids direct interference from differences in client category distributions to the decision layer, enabling each client to maintain a discrimination boundary consistent with its local data. By sharing the backbone, the model can still learn stable, generalized feature representations across clients, thus ensuring overall generalization ability.

[0100] Step 6.3 Since the backbone parameters are updated after each round of communication, the feature representation space may experience slight drift, leading to a mismatch between the original classification head and the new feature space. To mitigate this issue, a lightweight head warm-up mechanism is introduced to freeze the backbone parameters before model inference or the next training round. Only for the classification header parameters Quick updates:

[0101]

[0102] This mechanism can be viewed as a localized, rapid adaptation process, similar to recalibrating classifier parameters on a fixed feature space. Compared to retraining the entire model, this approach has extremely low computational cost but effectively restores the consistency between the classification boundary and the feature representation, thereby significantly improving the model's stability in dynamic federated environments.

[0103] While the present invention has been described above with reference to preferred embodiments, it is not intended to limit the invention. Those skilled in the art can make various modifications and refinements without departing from the spirit and scope of the invention. Therefore, the scope of protection of the present invention shall be determined by the claims.

Claims

1. A mental health monitoring system based on quantum personalized federated learning, characterized in that, include: The data acquisition and initialization module is configured on each client and is used to collect users' voice data through terminal devices and complete data storage and preprocessing locally. The server is used to initialize global model parameters, including initializing the quantum parameters of the shared feature extractor. and classic parameters And the personalized classifier parameters for each client. and initialize the parameters Distribute it to each client as the initial model; A global geometry-guided data augmentation module is configured on each client to construct augmented samples that approximate the global data distribution on local clients using an uncertainty augmentation method guided by the global geometry. The generation of local samples is guided by global geometric priors to alleviate the problem of data distribution heterogeneity in federated learning environments. The time-modulated quantum gate unit module is configured on each client and is used to input the enhanced speech features into the time-modulated quantum gate unit model during the local training phase. The input features are first processed by the classical feature mapping layer for representation learning, and then encoded into the variable quantum circuit. The high-dimensional quantum feature representation is obtained through quantum gate operations and quantum measurement. The parameter optimization module, configured on each client, calculates the weighted cross-entropy loss function based on the model prediction results and updates the model parameters using the gradient descent algorithm. Specifically, the shared feature extractor parameters and the personalized classifier parameters are optimized using the backpropagation algorithm. , The quantum circuit parameters are updated by calculating the gradient using parameter shift rules. ; A partial parameter upload module, configured on each client, is used to upload only the locally shared feature extractor parameters after local training is completed on the client side. Send to the server, while retaining local personalized classifier parameters. It does not participate in aggregation in order to achieve personalized emotion modeling for individual users; The federated aggregation module, configured on the server side, receives shared parameters uploaded by each client and uses a federated aggregation strategy to weight and fuse the shared feature extractor parameters to obtain new globally shared parameters. The updated global parameters are then redistributed to each client to begin the next round of federated training, until the preset number of communication rounds is reached or the model meets the convergence condition.

2. The mental health monitoring system based on quantum personalized federated learning according to claim 1, characterized in that, The data augmentation module guided by global geometry specifically includes: The local statistical calculation unit is used to calculate the local mean vector and local covariance matrix of the samples for each sentiment category c in client k. The global statistical aggregation unit is used to upload local category statistics from each client to the server. The server then aggregates the statistics from all clients to obtain the global mean vector for category c. With the global covariance matrix : ， ， in, This represents the number of samples of category c in client k. This represents the total number of samples of category c across all clients; Eigenvalue decomposition unit, used for the global covariance matrix Eigenvalue decomposition is performed as follows: in, Represents the eigenvalue matrix. This represents the corresponding eigenvector matrix; Geometric perturbation building blocks are used to transfer global geometry information on the server. , After being distributed to each client, for any sample x of category c in client k, random coefficients are sampled from the standard normal distribution. And construct the geometric perturbation term based on the eigenvalue decomposition results of the global covariance matrix.

3. The mental health monitoring system based on quantum personalized federated learning according to claim 1, characterized in that, The time-modulated quantum gating unit module specifically includes: The time modulation offset calculation unit is used to calculate the time modulation offset based on the hidden state of the previous time step. In the formula, Represents the time modulation coefficient. This indicates the hidden state at the previous moment. This represents the average calculation; The update gate calculation unit is used to incorporate the time modulation offset into the update gate calculation process: In the formula, This represents the input feature vector at the current time t. This represents the Sigmoid activation function. This indicates the variable quantum circuit corresponding to the update gate. The candidate hidden state calculation unit is used to incorporate the time modulation offset into the candidate hidden state calculation process. In the formula, This represents the hyperbolic tangent activation function. This represents Hadamard element-wise multiplication. This represents the variable quantum circuit corresponding to the candidate hidden state calculation module. Hidden state update unit, used to calculate the hidden state at the current time: ， The quantum feature output unit is used to obtain the quantum feature representation through quantum measurement after the quantum state evolution is completed in the variable quantum circuit, and output the feature representation result of the digital twin model: ， Where W is the set of parameters in a classic network. It is the VQC module in TM-QMGU. It is a trainable phase angle parameter in a quantum circuit; the TM-QMGU is parameterized as a l-layer quantum circuit, each layer consisting of two RX gates and a set of controlled NOT gates on adjacent qubits, and the three VQC modules corresponding to the update gate, reset gate and candidate hidden state calculation module share the same architecture.

4. The mental health monitoring system based on quantum personalized federated learning according to claim 1, characterized in that, The time-modulated quantum gating unit module specifically includes: The parameter optimization module specifically includes: The category weight construction unit is used to input the feature representation output by TM-QMGU into the personalized classifier to obtain the sentiment category prediction result, and to construct the category weight coefficients according to the number of samples in each category. Where N is the total number of training samples and C is the total number of classes. The number of samples in category c; The weighted loss calculation unit is used to calculate the error between the model's prediction results and the true labels using the weighted cross-entropy loss function. And construct the optimization objective as follows: ； The quantum-classical hybrid training unit is used to update model parameters using a quantum-classical hybrid training approach. Classical network parameters are updated through the backpropagation algorithm, while quantum circuit parameters are optimized by calculating gradients using parameter shift rules. Dimension alignment unit, containing a first linear transformation layer Second linear transformation layer The first linear transformation layer is used to align the dimension of the input data with the input dimension of the VQC module, and the second linear transformation layer is used to align the output dimension of the VQC module with the input dimension of the subsequent classical operation.

5. The mental health monitoring system based on quantum personalized federated learning according to claim 1, characterized in that, In the partial parameter upload module, the complete model parameters are represented as follows: ,in and These represent the quantum parameters and classical parameters shared in the backbone, respectively. This represents the personalized classification header parameters; after the client completes local training, it only uploads the shared Backbone parameters. Federated aggregation is performed on the server while retaining personalized category header parameters. Locally.

6. The mental health monitoring system based on quantum personalized federated learning according to claim 5, characterized in that, The federated aggregation module specifically includes: The federated average aggregation unit is used by the server to receive shared Backbone parameters uploaded by each client. , The global model parameters are updated using a federated average aggregation method. ， ， Where K represents the number of clients participating in the training; The parameter distribution unit is used by the server to send the updated globally shared parameters after aggregation is complete. , The data is redistributed to each client, but the personalized category header parameters are not included in the aggregation, thus satisfying the requirement. ; The classification head fast calibration unit is used to freeze the backbone parameters after receiving new global parameters, and only perform fast calibration updates on the classification head parameters, as follows: ,in, Represents the loss function. Indicates the learning rate. Indicates the category header, This represents the gradient of the loss function with respect to the parameters of the k-th client classification head.

7. A mental health monitoring system based on quantum personalized federated learning according to claim 1, characterized in that, In the data acquisition and initialization module, the number of qubits in the TM-QMGU is set to 4, the learning rate is 0.001, the batch size is 128, and the epoch is 200.

8. A mental health monitoring system based on quantum personalized federated learning, characterized in that, include: One or more processors; Storage device for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to perform the functions of all or part of the modules of the system as described in any one of claims 1 to 7.

Images