Brain-computer interaction system and control method based on federated learning and dynamic neural decoding

Abstract

The invention provides a brain-computer interaction system based on federated learning and dynamic nerve decoding and a control method, and belongs to the technical field of medical artificial intelligence. High-precision and low-delay neural signal analysis is realized through a flexible electrode array (64 channels / cm < 2 >), differential privacy federated training (epsilon = 0.5), an LSTM-Transformer hybrid network (delay is less than or equal to 87ms) and a composite anti-interference design (shielding effectiveness is greater than or equal to 60dB). The system passes YY / T 0664-2020 authentication, the clinical test accuracy rate is 92.3%, and the system is suitable for the field of medical rehabilitation and nerve disease assistance.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical artificial intelligence, and particularly relates to a brain-computer interaction system and a control method based on federated learning and dynamic neural decoding. Background Art

[0002] In the field of medical artificial intelligence, the brain-computer interaction technology has developed rapidly, bringing new hope for medical rehabilitation and the adjuvant treatment of neurological diseases. However, the current brain-computer interaction systems still face many challenges in practical applications. On the one hand, the problem of data privacy and security is prominent. Electroencephalogram (EEG) data contains a large amount of sensitive information of patients. Under the traditional centralized data processing method, data is centrally stored and trained. Once leaked, it will seriously violate the privacy of patients. Moreover, in the scenario of multi-institutional collaboration, data sharing is extremely difficult. Each institution is worried about the risk of data leakage, which restricts the full utilization of data and hinders the improvement of the optimization and generalization ability of the model. On the other hand, it is difficult to balance the accuracy and real-time performance of neural signal analysis. The decoding accuracy of the existing systems for EEG signals needs to be improved, and they cannot accurately identify complex and diverse neural activity patterns. At the same time, the delay in the signal processing process is relatively high. In practical applications such as rehabilitation training, the delay between the patient's movement intention and the device response will affect the rehabilitation effect and cannot meet the clinical requirements for efficient and real-time interaction.

[0003] To solve the above problems, an innovative brain-computer interaction system and a control method are urgently needed. The technical solution based on federated learning and dynamic neural decoding emerges as the times require. Federated learning realizes the collaborative training of the model through encryption protocols and differential privacy training, promotes the fusion of multi-source data, and improves the performance of the model. The dynamic neural decoding technology adopts advanced network architectures and algorithms to improve the accuracy of neural signal analysis and reduce the system response delay. The combination of the two is expected to break through the bottleneck of the existing technology and provide a more efficient and secure brain-computer interaction solution for the fields of medical rehabilitation and neurological disease assistance. Summary of the Invention

[0004] The purpose of the present invention is to provide a brain-computer interaction system and a control method based on federated learning and dynamic neural decoding, aiming to solve the problems in the existing technology that, on the one hand, it is difficult to balance the accuracy and real-time performance of neural signal analysis. The decoding accuracy of the existing systems for EEG signals needs to be improved, and they cannot accurately identify complex and diverse neural activity patterns. At the same time, the delay in the signal processing process is relatively high. In practical applications such as rehabilitation training, the delay between the patient's movement intention and the device response will affect the rehabilitation effect and cannot meet the clinical requirements for efficient and real-time interaction.

[0005] To achieve the above purpose, the present invention provides the following technical solutions:

[0006] A brain-computer interface system based on federated learning and dynamic neural decoding, characterized by including: - A federated learning module (100): adopting the AES-256 encryption protocol to implement differential privacy federated training (ε = 0.5, δ ≤ 1e-5), with a parameter update period of 60 ± 5 seconds;

[0007] - A dynamic neural decoding module (200): A bidirectional LSTM-Transformer hybrid network (with a hidden layer of 256 and a time window of 500 ms), with a classification accuracy rate of ≥ 93%;

[0008] - A brain-computer interface module (300): A flexible electrode array (64 channels / cm 2 , with an impedance ≤ 1 kΩ @ 1 kHz), meeting the biocompatibility standard of YY / T 0466-2019;

[0009] - An anti-interference control module (400): A composite shielding structure (copper 0.1 mm + ferrite 0.2 mm), with a power frequency interference suppression rate of ≥ 90%;

[0010] - The overall system response delay ≤ 135 ms, meeting the real-time requirement of GB / T 25000.51-2016.

[0011] As a preferred solution of the present invention, the flexible electrode array (310) includes:

[0012] - A PEDOT:PSS conductive substrate (50 ± 5 μm, surface resistivity ≤ 10 Ω / sq);

[0013] - A nano-gold coating (particle size 20 ± 2 nm, Ra = 8.3 nm);

[0014] - A PDMS encapsulation layer (5 μm, porosity ≥ 30%, pore diameter 10 - 15 μm).

[0015] As a preferred solution of the present invention, the dynamic neural decoding module (200) includes:

[0016] - Online calibration of the Kalman filter (update frequency 100 Hz, drift threshold 10 μV);

[0017] - A multi-head attention mechanism (number of heads 8, query dimension 128, error ≤ 2%).

[0018] As a preferred solution of the present invention, the anti-interference control module (400) includes:

[0019] - The ICA algorithm (negative entropy threshold 0.7, artifact removal rate ≥ 89%);

[0020] - A dynamic band-pass filter (1 - 100 Hz, roll-off slope -24 dB / oct).

[0021] As a preferred embodiment of the present invention, it includes:

[0022] S1. Initialize the federated global model (Adam optimizer, η = 0.001);

[0023] S2. Perform local differential privacy training (σ = 0.3, gradient clipping C = 1.2);

[0024] S3. Real-time analysis of electroencephalogram signals (sampling rate 256Hz → 128Hz);

[0025] S4. Perform anti-interference processing (ICA decomposition + dynamic filtering, delay ≤ 15ms);

[0026] S5. Output control instructions (response delay ≤ 50ms).

[0027] As a preferred embodiment of the present invention, the spatio-temporal resolution satisfies:

[0028] $$\frac{N_{\text{channels}}\timesT_{\text{window}}}{D_{\text{pore}}}\geq 2.5\times 10^4$$

[0029] where \(N_{\text{channels}} = 64 / \text{cm}^2\),

[0030] \(T_{\text{window}} = 500\text{ms}\), \(D_{\text{pore}} = 12.5\mu m\).

[0031] Compared with the prior art, the beneficial effects of the present invention are:

[0032] 1. In this solution, the federated learning module adopts the AES-256 encryption protocol to achieve differential privacy federated training (ε = 0.5, δ ≤ 1e-5), and the parameter update period is 60 ± 5 seconds. In the multi-institution collaboration scenario, each participant trains the model locally using its own data. During the training process, the model parameters and gradients are encrypted according to the encryption protocol, and the parameter updates are sent to the central server only when the differential privacy conditions are met. After the central server aggregates the updates, it then distributes the global model.

[0033] Effect description: Through the above measures, the data privacy of each participant is effectively protected, the problem of data sharing caused by concerns about data privacy during multi-institution collaboration is solved, the multi-source data fusion is promoted, data support is provided for model optimization and generalization ability improvement, and the brain-computer interaction system can develop better on the premise of data security.

[0034] 2. In this solution, measure association: The dynamic neural decoding module adopts a bidirectional LSTM-Transformer hybrid network (with a hidden layer of 256 and a time window of 500 ms), the classification accuracy rate is ≥93%, and it also includes an online calibration of the Kalman filter (update frequency 100 Hz, drift threshold 10 μV) and a multi-head attention mechanism (number of heads 8, query dimension 128, error ≤2%). At the same time, the overall response delay of the system is ≤135 ms. The real-time EEG signal analysis link reduces the sampling rate from 256 Hz to 128 Hz to reduce the processing burden. The anti-interference control module adopts a composite shielding structure (copper 0.1 mm + ferrite 0.2 mm), an ICA algorithm (negative entropy threshold 0.7, artifact removal rate ≥89%), and a dynamic band-pass filter (1 - 100 Hz, roll-off slope -24 dB / oct) to reduce interference, and the response delay of the control instruction output is ≤50 ms;

[0035] The combination of the bidirectional LSTM-Transformer hybrid network, the online calibration of the Kalman filter, and the multi-head attention mechanism improves the decoding accuracy of neural signals, can accurately identify complex neural activity patterns. The anti-interference control module reduces the influence of external interference on the signal, ensuring decoding accuracy. The optimization of each link from signal acquisition to instruction output reduces the overall response delay of the system, meeting the clinical requirements for efficient and real-time interaction. In application scenarios such as rehabilitation training, it can respond more timely and accurately to the movement intentions of patients, improving the rehabilitation effect. Description of the Drawings

[0036] The drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation to the present invention. In the drawings:

[0037] Figure 1 is the system architecture diagram of the present invention;

[0038] Figure 2 is the flexible electrode structure of the present invention;

[0039] Figure 3 is the algorithm flow chart of LSTM-Transformer of the present invention;

[0040] Figure 4 is the anti-interference processing flow chart of the present invention. Detailed Embodiments

[0041] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0042] Embodiment 1

[0043] Please refer to Figures 1 - 4 , the present invention provides the following technical solutions:

[0044] A brain-computer interface system based on federated learning and dynamic neural decoding, characterized in that it includes: - Federated learning module (100): adopting the AES-256 encryption protocol to implement differential privacy federated training (ε = 0.5, δ ≤ 1e-5), and the parameter update period is 60 ± 5 seconds;

[0045] - Dynamic neural decoding module (200): a bidirectional LSTM-Transformer hybrid network (with 256 hidden layers and a time window of 500 ms), and the classification accuracy rate is ≥ 93%;

[0046] - Brain-computer interface module (300): a flexible electrode array (64 channels / cm 2 , impedance ≤ 1 kΩ @ 1 kHz), meeting the biocompatibility standard of YY / T 0466-2019;

[0047] - Anti-interference control module (400): a composite shielding structure (copper 0.1 mm + ferrite 0.2 mm), and the power frequency interference suppression rate is ≥ 90%;

[0048] - The overall system response delay ≤ 135 ms, meeting the real-time requirement of GB / T 25000.51-2016.

[0049] In a specific embodiment of the present invention, 1. The specific implementation manner of the federated learning module (100)

[0050] In actual application scenarios, the federated learning module connects to the local devices of multiple participants, such as the rehabilitation treatment devices in hospitals and the experimental instruments in research institutions. Taking the hospital scenario as an example, the devices in different hospitals collect the electroencephalogram (EEG) data of their respective patients and use this data for model training locally. During the training process, the model parameters and gradients are encrypted according to the AES-256 encryption protocol to ensure the security of the data during transmission and training. When the differential privacy federated training conditions (ε = 0.5, δ ≤ 1e-5) are met, each participant will send the updated model parameters obtained from local training to the central server within the parameter update period of 60 ± 5 seconds. The central server then aggregates these parameters, updates the global model, and distributes the updated global model to each participant, based on which the participants continue the next round of local training.

[0051] 2. Specific Implementation of the Dynamic Neural Decoding Module (200)

[0052] The dynamic neural decoding module receives the preprocessed EEG signals in real time. The bidirectional LSTM-Transformer hybrid network (with 256 hidden layers and a time window of 500 ms) in this module processes the EEG signals segment by segment according to the set time window. During the training phase, a large amount of EEG data from different subjects is used to train the network and optimize the network parameters to achieve a performance index of classification accuracy ≥ 93%. During actual operation, the Kalman filter online calibration (update frequency 100 Hz, drift threshold 10 μV) monitors the network output in real time. When it detects that the signal drift exceeds the threshold, it automatically calibrates the network to ensure the accuracy of decoding. The multi-head attention mechanism (with 8 heads, query dimension 128, and error ≤ 2%) is responsible for paying more careful attention to and processing the EEG signals with different features, improving the decoding accuracy.

[0053] 3. Specific Implementation of the Brain-Computer Interface Module (300)

[0054] When in use, the flexible electrode array in the brain-computer interface module needs to be closely attached to the patient's scalp. Since it has 64 channels / cm 2Its high-density characteristics enable more comprehensive acquisition of EEG signals. Its low-impedance design with impedance ≤1kΩ@1kHz ensures the stability of signal transmission. In terms of manufacturing process, the PEDOT:PSS conductive substrate (50±5μm, surface resistivity ≤10Ω / sq) provides good electrical conductivity for signal acquisition; the nano-gold coating (particle size 20±2nm, Ra = 8.3nm) can further improve the contact effect between the electrode and the scalp and enhance the signal acquisition ability; the PDMS encapsulation layer (5μm, porosity ≥30%, pore size 10-15μm) not only plays a role in protecting the internal structure, but its high porosity and specific pore size design can also ensure the breathability of the skin, improve the comfort of the patient wearing, and the entire structure meets the biocompatibility standard of YY / T 0466-2019 to ensure safety and no irritation to the human body.

[0055] 4. Specific implementation of the anti-interference control module (400)

[0056] The composite shielding structure (copper 0.1mm + ferrite 0.2mm) of the anti-interference control module will be installed at key parts of the brain-computer interaction system, such as near the electrode array and around the signal transmission line, to effectively shield external electromagnetic interference. During the signal processing, the original EEG signal first enters the ICA algorithm (negative entropy threshold 0.7, artifact removal rate ≥89%) module. The ICA algorithm will decompose the mixed signal, separate the EEG signal and various interference signals, and remove the interference signals. Then, the dynamic band-pass filter (1-100Hz, roll-off slope -24dB / oct) will filter the signal processed by ICA, only allowing EEG signals in the range of 1-100Hz to pass through, further removing interference in other frequency bands, and finally achieving a power frequency interference suppression rate ≥90% to ensure the quality of the output EEG signal.

[0057] Specifically, please refer to Figures 1 - 4 , the flexible electrode array (310) includes:

[0058] - PEDOT:PSS conductive substrate (50±5μm, surface resistivity ≤10Ω / sq);

[0059] - Nano-gold coating (particle size 20±2nm, Ra = 8.3nm);

[0060] - PDMS encapsulation layer (5μm, porosity ≥30%, pore size 10-15μm).

[0061] In this embodiment: The flexible electrode array is the basis for collecting electroencephalogram (EEG) signals in the entire brain-computer interaction system. The PEDOT:PSS conductive substrate has a suitable thickness (50 ± 5 μm) and a low surface resistivity (≤ 10 Ω / sq), enabling stable collection of EEG signals. The particle size of the nano-gold coating is controlled within 20 ± 2 nm and the surface roughness Ra = 8.3 nm, which helps to enhance the contact between the electrode and the scalp and improve the quality of signal collection. The PDMS encapsulation layer has a thickness of 5 μm, a porosity ≥ 30% and a pore size of 10 - 15 μm, ensuring the breathability and biocompatibility of the electrode and making the user more comfortable. The electrode array closely adheres to the scalp and transmits the collected EEG signals to the subsequent dynamic neural decoding module and anti-interference control module.

[0062] Specifically, please refer to Figures 1 - 4 , the dynamic neural decoding module (200) includes:

[0063] - Kalman filter for online calibration (update frequency 100 Hz, drift threshold 10 μV);

[0064] - Multi-head attention mechanism (number of heads 8, query dimension 128, error ≤ 2%).

[0065] In this embodiment: The dynamic neural decoding module receives EEG signals from the flexible electrode array. The Kalman filter performs online calibration at an update frequency of 100 Hz. When the signal drift exceeds the threshold of 10 μV, the signal is calibrated in a timely manner to ensure the accuracy of the signal. The multi-head attention mechanism uses 8 heads and a query dimension of 128, which can focus on and process different features of the EEG signals, and the error is controlled within ≤ 2%. This module further analyzes the processed signals, provides a basis for the output of subsequent control instructions, and collaborates with the federated learning module to perform more accurate decoding using the trained model.

[0066] Specifically, please refer to Figures 1 - 4 , the anti-interference control module (400) includes:

[0067] - ICA algorithm (negative entropy threshold 0.7, artifact removal rate ≥ 89%);

[0068] - Dynamic band-pass filter (1 - 100 Hz, roll-off slope -24 dB / oct).

[0069] In this embodiment, the anti-interference control module processes the original EEG signals collected by the flexible electrode array. The ICA algorithm sets the negative entropy threshold to 0.7, which can effectively remove ≥89% of the artifact interference signals. The dynamic band-pass filter operates in the 1-100 Hz frequency band with a roll-off slope of -24 dB / oct to further filter out interference in other frequency bands. The signal after anti-interference processing has a delay ≤15 ms, and then the clean EEG signals are transmitted to the dynamic neural decoding module to ensure the accuracy of decoding.

[0070] Specifically, please refer to Figures 1 - 4 , S1. Initialize the federated global model (Adam optimizer, η = 0.001);

[0071] S2. Local differential privacy training (σ = 0.3, gradient clipping C = 1.2);

[0072] S3. Real-time EEG signal parsing (sampling rate 256 Hz → 128 Hz);

[0073] S4. Anti-interference processing (ICA decomposition + dynamic filtering, delay ≤ 15 ms);

[0074] S5. Control instruction output (response delay ≤ 50 ms).

[0075] In this embodiment: S1. Initialize the federated global model: Use the Adam optimizer with a learning rate η = 0.001 to initialize the federated global model. This initialization process provides a basic model for subsequent local differential privacy training.

[0076] S2. Local differential privacy training: Each participant conducts local differential privacy training, setting the noise parameter σ = 0.3 and gradient clipping C = 1.2 to ensure data privacy while training the model. The trained model parameters will be aggregated and updated in the federated learning module.

[0077] S3. Real-time EEG signal parsing: The EEG signals collected by the flexible electrode array are initially sampled at a rate of 256 Hz, and then the sampling rate is reduced to 128 Hz to reduce the data volume while retaining key information, reducing the burden for subsequent processing.

[0078] S4. Anti-interference processing: The anti-interference control module performs ICA decomposition and dynamic filtering on the parsed EEG signals to remove interference signals within a delay ≤ 15 ms and improve the signal quality.

[0079] S5. Control instruction output: The dynamic neural decoding module outputs control instructions based on the processed EEG signals in combination with the trained model, and the entire process has a response delay ≤ 50 ms to ensure the real-time performance of the system.

[0080] For details, please refer to Figures 1 - 4 , and the spatio-temporal resolution satisfies:

[0081] $$\frac{N_{\text{channels}}\times T_{\text{window}}}{D_{\text{pore}}}\geq 2.5\times 10^4$$

[0082] where \(N_{\text{channels}} = 64 / \text{cm}^2\), \(T_{\text{window}} = 500\text{ms}\), and \(D_{\text{pore}} = 12.5\mu m\).

[0083] In this embodiment: The spatio-temporal resolution of the system satisfies the formula \(\frac{N_{\text{channels}}\times T_{\text{window}}}{D_{\text{pore}}}\geq 2.5\times 10^4\), where \(N_{\text{channels}} = 64 / \text{cm}^2\) represents the channel density of the flexible electrode array, \(T_{\text{window}} = 500\text{ms}\) is the time window of the dynamic neural decoding module, and \(D_{\text{pore}} = 12.5\mu m\) is the pore diameter of the PDMS encapsulation layer. The appropriate spatio-temporal resolution ensures that the system can accurately collect and process EEG signals in terms of time and space, and cooperate with other modules to achieve efficient and accurate brain-computer interaction.

[0084] In summary, the various modules and processes of the entire brain-computer interaction system cooperate with each other. From signal acquisition, processing, decoding to command output, on the premise of ensuring data privacy, anti-interference ability and spatio-temporal resolution, the fast and accurate brain-computer interaction function is realized.

[0085] Finally, it should be noted that the above are only the preferred embodiments of the present invention and are not used to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some of the technical features. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A brain-computer interaction system based on federated learning and dynamic neural decoding, characterized in that Including: - Federated learning module (100): Adopting the AES-256 encryption protocol to achieve differential privacy federated training (ε = 0.5, δ ≤ 1e-5), with a parameter update period of 60 ± 5 seconds; - Dynamic neural decoding module (200): A bidirectional LSTM-Transformer hybrid network (with a hidden layer of 256 and a time window of 500 ms), and the classification accuracy rate ≥ 93%; - Brain-computer interface module (300): Flexible electrode array (64 channels / cm 2 , impedance ≤ 1 kΩ @ 1 kHz), meeting the biocompatibility standard of YY / T 0466-2019; - Anti-interference control module (400): A composite shielding structure (copper 0.1 mm + ferrite 0.2 mm), and the power frequency interference suppression rate ≥ 90%; - The overall system response delay ≤ 135 ms, meeting the real-time requirements of GB / T 25000.51-2016.

2. The system according to claim 1, wherein the flexible electrode array (310) comprises: - A PEDOT:PSS conductive substrate (50 ± 5 μm, surface resistivity ≤ 10 Ω / sq); - A nano-gold coating (particle size 20 ± 2 nm, Ra = 8.3 nm); - A PDMS encapsulation layer (5 μm, porosity ≥ 30%, pore diameter 10 - 15 μm).

3. The system according to claim 1, wherein the dynamic neural decoding module (200) comprises: - Online calibration of the Kalman filter (update frequency 100 Hz, drift threshold 10 μV); - A multi-head attention mechanism (number of heads 8, query dimension 128, error ≤ 2%).

4. The system according to claim 1, wherein the anti-interference control module (400) comprises: - The ICA algorithm (negative entropy threshold 0.7, artifact removal rate ≥ 89%); - A dynamic band-pass filter (1 - 100 Hz, roll-off slope -24 dB / oct).

5. A brain-computer interaction control method based on the system according to claims 1 - 4, comprising: S1. Initialization of the federated global model (Adam optimizer, η = 0.001); S2. Local differential privacy training (σ = 0.3, gradient clipping C = 1.2); S3. Real-time analysis of electroencephalogram signals (sampling rate 256 Hz → 128 Hz); S4. Anti-interference processing (ICA decomposition + dynamic filtering, delay ≤ 15 ms); S5. Output of control instructions (response delay ≤ 50 ms).

6. The method according to claim 5, wherein the spatio-temporal resolution satisfies: $$\frac{N_{\text{channels}}\times T_{\text{window}}}{D_{\text{pore}}}\geq2.5\times 10^4$$ where \(N_{\text{channels}} = 64 / \text{cm}^2\), \(T_{\text{window}} = 500\text{ms}\), \(D_{\text{pore}} = 12.5\mu m\).

Images