A state closed-loop recognition method and system based on force feedback and electroencephalogram signals

Abstract

A method and system for state closed-loop recognition based on force feedback and electroencephalogram (EEG) signals are disclosed. This relates to the fields of biomedical signal processing and brain-computer interface (BCI) technology, specifically a method and system for state closed-loop recognition based on force feedback and EEG signals. The method simultaneously acquires raw EEG signal sequences and time-synchronized multi-channel force feedback data sequences, extracts dynamic contact state features, and calculates the contact confidence of corresponding channels. These features are then fused using frequency-sensitive dual-drive features and input into a basic state recognition model to obtain the state recognition result and corresponding confidence level. When the confidence level is lower than a preset dynamic threshold, a channel contact state adjustment command is generated to optimize the input data quality. The newly acquired data is then used to efficiently fine-tune the model parameters online, generating a target personalized model. This solves the problem that BCI systems cannot achieve coordinated and adaptive response across the entire link from physical acquisition to intelligent recognition when signal quality fluctuates.

Description

Technical Field

[0001] This invention relates to the fields of biomedical signal processing and brain-computer interface technology, specifically to a state closed-loop recognition method and system based on force feedback and electroencephalogram (EEG) signals. Background Technology

[0002] In EEG-based state recognition applications (such as driver fatigue monitoring and attention training), signal acquisition quality is a prerequisite for determining the accuracy of the recognition model. Due to significant individual differences in head size and scalp impedance characteristics among different users, and the inevitable slight head movements that users will make during long-term monitoring, the contact pressure between the EEG electrodes and the scalp is constantly changing.

[0003] Fluctuations in contact pressure directly introduce contact impedance artifacts and motion artifacts, severely contaminating EEG signals. To address these issues, existing technologies offer the following solutions: Existing technology 1, Chinese invention patent application "EEG Cap and EEG Signal Acquisition System" (CN115736929A), discloses an EEG cap integrating a pressure sensor and an airbag adjustment mechanism. It uses the pressure sensor to monitor the contact pressure between each electrode and the scalp in real time. When the pressure falls below a preset threshold, it automatically controls the airbag to inflate and adjust the electrode pressure, or triggers an audible and visual alarm to prompt the user to manually adjust. However, this solution has the following drawbacks: the force feedback data is only used for simple threshold alarms or coarse-grained mechanical adjustments in the front-end hardware, failing to deeply participate in the back-end algorithm calculations and thus unable to suppress artifact interference caused by poor contact at the algorithmic level. Furthermore, none of the existing technologies provide a complete physical-algorithm closed-loop simulation verification framework, making it impossible to verify the closed-loop optimization effect without physical hardware.

[0004] Existing technology 2: FAHIMI F, ZHANG Z, GOH WB, et al. Inter-subject transfer learning with an end-to-end deep convolutional neural network for EEG-based BCI[J]. Journal of Neural Engineering, 2019, 16(2): 026007. It proposes a personalized EEG fatigue state recognition model based on cross-subject transfer learning, which completes the personalized adaptation of the model by acquiring a small amount of offline EEG data from the user. However, the drawback of this scheme is that it adopts an "open-loop" mode for offline training, which cannot complete dynamic self-evolution based on the real-time improvement of the physical state of the device; and the existing state recognition models usually assume that the input signal is clean and lack robustness to the deterioration of physical contact state. Once the device is loosened, the model recognition accuracy will drop sharply.

[0005] The existing technology, LOTTE F, BOUGRAIN L, CICHOCKI A, et al. A review of classification algorithms for EEG-based brain–computer interfaces: a 10 year update[J]. Journal of Neural Engineering, 2018, 15(3): 031005, indicates that existing open-loop online learning models assume all new data to be highly reliable data and cannot identify and filter low-quality "dirty data" caused by poor contact, leading to parameter drift, error accumulation, irreversible performance degradation after long-term operation, and a significant increase in the frequency of system recalibration.

[0006] Therefore, it is urgent to solve the problem that existing brain-computer interface systems cannot achieve coordinated and adaptive response across the entire link from physical acquisition to intelligent recognition when signal quality fluctuates because physical perception, signal processing, and model evolution are isolated from each other. Summary of the Invention

[0007] To address the aforementioned problems, this invention proposes a state closed-loop recognition method and system based on force feedback and electroencephalogram (EEG) signals. This solves the problem in existing brain-computer interface systems where physical perception, signal processing, and model evolution are isolated from each other, leading to an inability to achieve coordinated and adaptive responses across the entire chain from physical acquisition to intelligent recognition when signal quality fluctuates.

[0008] The method includes the following steps: S1. Real-time acquisition of the user's original EEG signal sequence, and a multi-channel force feedback data sequence that is time-synchronized with the original EEG signal sequence; S2. Based on the multi-channel force feedback data sequence, extract the dynamic contact state features of each channel and calculate the contact reliability of the corresponding EEG signal. S3. Extract the EEG features of the original EEG signal sequence and perform frequency band-sensitive dual-drive feature fusion with the contact confidence, input it into the pre-trained state recognition basic model for forward inference, and obtain the state recognition result and the confidence corresponding to the result; S4. Determine whether the confidence level is lower than a preset dynamic threshold. If it is lower, generate a channel contact state adjustment command based on the contact confidence level and dynamic contact state characteristics and send it to the data generation module. S5. Obtain the new EEG signal sequence and new force feedback data sequence returned by the data generation module after executing the channel contact state adjustment command, and fine-tune the parameters of the pre-trained state recognition basic model based on the new EEG signal sequence and new force feedback data sequence to generate and store the target personalized state recognition model.

[0009] Furthermore, the dynamic contact state characteristics include: the pressure mean offset of each channel. and high-frequency volatility of pressure , This is the index value of the channel. The index value of the sliding window; The contact reliability The formula for calculation is: ,in, For mask smoothing coefficient, For the first The first channel, in the... The artifact energy index of a sliding window. The artifact tolerance threshold, It is a natural exponential function; The formula for calculation is: ,in, The standard contact pressure baseline, This is a reference value for normal pressure fluctuations. and These are the weighting coefficients. .

[0010] Furthermore, the EEG characteristics of the original EEG signal sequence include: low-frequency band characteristics. and high frequency band characteristics ; The dual-drive feature fusion includes the following steps: S31. Calculate the low-frequency mask based on the dynamic contact state characteristics. and high-frequency mask , , ,in, For Sigmoid-class smoothing mapping functions; S32, will and , respectively applied to and Decoupling and weighting are performed to obtain the fused feature vector. and ; S33 will and Perform normalization; S34. The weighted result from step S32 and the normalized result from step S33 are concatenated to obtain the dual-drive feature fusion vector.

[0011] Furthermore, the state recognition basic model adopts a Transformer-CNN hybrid architecture, which passes through a backbone network layer and a classification adapter layer from input to output. The backbone network layer includes 3 Transformer encoders, 2 one-dimensional convolutions and 1 global average pooling layer. The classification adapter layer includes 2 fully connected layers.

[0012] Furthermore, the generation process of the channel contact state adjustment command includes the following steps: S41. Extract the set of abnormal channels whose single-channel confidence is lower than the preset dynamic threshold; S42. Calculate the local pressure compensation parameters based on the spatial topological position of each channel in the abnormal channel set and the direction of the current pressure deviation from the standard baseline; The compensation parameters include the target channel, adjustment direction, and adjustment step size, wherein the adjustment step size is 1 gram to 5 grams; S43. The local pressure compensation parameters are encapsulated into the channel contact state adjustment command, the command including at least the target adjustment channel identifier, adjustment direction and adjustment step size.

[0013] Furthermore, the data generation module is an external acquisition device.

[0014] Furthermore, the parameter fine-tuning specifically involves: freezing the weight parameters of the backbone network layer in the pre-trained state recognition base model; and performing gradient updates only on the classification adapter layer of the state recognition base model.

[0015] Furthermore, step S5 also includes a confidence gating update method, specifically: when the confidence level is lower than a preset dynamic threshold, if the single-channel contact confidence level is ≥0.7 and the global average contact confidence level is ≥0.8, a channel contact state adjustment command is generated; Conversely, it will not be generated.

[0016] A state closed-loop recognition system based on force feedback and electroencephalogram (EEG) signals, the system being used to implement the above method, the system comprising: a synchronous data acquisition module, a reliability assessment module, a dual-drive fusion module, a closed-loop control module, and a fine-tuning update module; The synchronous data acquisition module is used to acquire the user's original EEG signal sequence in real time, as well as the multi-channel force feedback data sequence that is time-synchronized with the original EEG signal sequence. The credibility assessment module extracts the dynamic contact state features of each channel based on the multi-channel force feedback data sequence and calculates the contact credibility of the corresponding EEG signal. The dual-drive fusion module is used to extract the EEG features of the original EEG signal sequence and perform frequency band-sensitive dual-drive feature fusion with the contact confidence, and input it into the pre-trained state recognition basic model for forward inference to obtain the state recognition result and the confidence level corresponding to the result; The closed-loop control module is used to determine whether the confidence level is lower than a preset dynamic threshold. If it is lower, it generates a channel contact state adjustment command based on the contact confidence level and dynamic contact state characteristics and sends it to the data generation module. The fine-tuning update module is used to obtain the new EEG signal sequence and the new force feedback data sequence returned by the data generation module after executing the channel contact state adjustment command, and to fine-tune the parameters of the pre-trained state recognition basic model based on the new EEG signal sequence and the new force feedback data sequence, thereby generating and storing the target personalized state recognition model.

[0017] The beneficial effects of the method described in this invention are as follows: (1) The present invention does not regard force feedback as an alarm signal independent of EEG, but transforms it into a signal quality assessment index (adaptive weight mask) and performs weight reduction suppression on the severely contaminated EEG channels at the model inference front end.

[0018] (2) This invention uses the designed “artifact energy index formula” and “frequency band sensitive dual-drive fusion mechanism” to accurately convert the physical quantity of force feedback into a soft weighted mask in the frequency domain of EEG, realizing deep synergy from “mechanical vibration → frequency domain artifact mapping → precise stripping and suppression”. While eliminating interference, it preserves the real EEG features to the greatest extent. This is a technical effect that conventional feature splicing or black box networks cannot learn spontaneously.

[0019] (3) When the model determines the recognition result with low confidence and the data quality is poor, the present invention actively issues adjustment instructions to drive the data generation module to optimize the input data quality. At the same time, it uses the high-quality new data returned after optimization to efficiently fine-tune the parameters of the model, realizing the leap from "passively accepting data" to "actively improving data and self-evolving".

[0020] (4) The present invention introduces a credibility gating update mechanism, which blocks the pollution of the model by "dirty data" from the source by allowing only high credibility data to enter the model update process, and significantly improves long-term stability. Attached Figure Description

[0021] Figure 1 This is a flowchart of the method described in this invention; Figure 2 This is a schematic diagram of the artifact suppression process described in this invention; Figure 3 This is a schematic diagram of the data processing logic for dual-drive feature fusion described in this invention; Figure 4 This is a block diagram of the adaptive adjustment control of the channel contact state according to the present invention; Figure 5 This is a schematic diagram of the parameter efficient fine-tuning (PEFT) structure described in this invention; Figure 6 This is a schematic diagram of the system structure described in this invention. Detailed Implementation

[0022] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Example 1 This embodiment provides a state closed-loop recognition method based on force feedback and electroencephalogram (EEG) signals. The flowchart of the method is as follows: Figure 1 As shown, the method includes the following steps: S1. Real-time acquisition of the user's original EEG signal sequence, and a multi-channel force feedback data sequence that is time-synchronized with the original EEG signal sequence; The relevant operations in step S1 will be introduced with specific examples: The user's original EEG signal sequence, and the multi-channel force feedback data sequence that is time-synchronized with the original EEG signal sequence, can be obtained from publicly available datasets, simulation-generated data, or real-world environments. In this embodiment, the user's original EEG signal sequence and the multi-channel force feedback data sequence that is time-synchronized with the original EEG signal sequence are obtained from the SEED-VIG fatigue driving dataset published by the Brain-Computer Interface Laboratory of Shanghai Jiao Tong University. This dataset natively contains the subject's lead EEG signal (200Hz) and the corresponding channel's synchronized force feedback signal (200Hz). The two signal sequences are exactly the same length, the timestamps correspond one-to-one, and there is no time offset. They can be read and used directly without additional synchronization processing.

[0024] S2. Based on the multi-channel force feedback data sequence, extract the dynamic contact state features of each channel and calculate the contact reliability of the corresponding EEG signal. The relevant operations in step S2 will be introduced with specific examples: A sliding window time-domain analysis was performed on the multi-channel force feedback data sequence to extract the pressure mean offset of each channel. and high-frequency volatility of pressure As a characteristic of dynamic contact state; This is the index value of the channel. The index value is the sliding window index; a new window is generated each time the window is slid. Starting from 1 and increasing sequentially; The specific method for extracting dynamic contact state features is as follows: a 1-second sliding window with a 0.5-second step size (corresponding to a sampling rate of 200Hz, each window contains 200 sampling points, and adjacent windows overlap by 50%) is used to segment the force feedback data. Pressure mean offset : No. The channel is in The average pressure within each sliding window and the standard contact pressure baseline The difference is calculated using the following formula:

[0025] in, It is a function of the arithmetic mean. For the first The channel is in 200 pressure samples within a window, This is the raw force feedback data for the i-th channel.

[0026] High-frequency volatility of pressure : No. The channel is in The standard deviation of pressure within a sliding window is used to characterize the degree of rapid pressure fluctuation, and the calculation formula is as follows:

[0027] in, This is the standard deviation function. The calculation method is as follows: first, calculate the mean of 200 pressure samples within the window; then, calculate the average of the squared differences between each sample and the mean; finally, take the square root to obtain the standard deviation.

[0028] A model for calculating the artifact energy index is constructed based on the physical mechanism of dynamic contact state characteristics. ,in, This represents the absolute value of the pressure mean offset (both excessively high and low pressures will introduce artifacts). The standard contact pressure baseline, This is a reference value for normal pressure fluctuations. and These are the weighting coefficients. , In this embodiment, the standard contact pressure baseline =8.0 grams, reference value for normal pressure fluctuation =0.5 grams: weighting coefficient =0.6、 =0.4.

[0029] In this embodiment, the artifact suppression process is shown in Figure 2.

[0030] The artifact energy index The input is fed into the Sigmoid mapping function to generate an adaptive weight mask corresponding to the EEG features of that channel. As a measure of the contact reliability: ,in, For mask smoothing coefficient, For the first The first channel, in the... The artifact energy index of a window. The artifact tolerance threshold, For the natural exponential function, in this embodiment, the mask smoothing coefficient is... =5.0, so that the Sigmoid curve has a suitable steepness, ensuring that the contact confidence changes rapidly near the artifact energy threshold, while avoiding feature instability caused by abrupt changes. Artifact tolerance threshold =0.5: When the artifact energy is 0.5, the contact confidence level is 0.5. That is, when the artifact energy is lower than 0.5, the signal quality of this channel is judged to be unqualified.

[0031] S3. Extract the EEG features of the original EEG signal sequence and perform frequency band-sensitive dual-drive feature fusion with the contact confidence, input it into the pre-trained state recognition basic model for forward inference, and obtain the state recognition result and the confidence corresponding to the result; The relevant operations in step S3 will be introduced with specific examples: The process of dual-drive feature fusion in this embodiment is as follows: Figure 3 As shown, the EEG features of the original EEG signal sequence are first extracted. Specifically, a Butterworth bandpass filter is used to filter the original EEG signal to extract the low-frequency band features of the original EEG signal sequence. (0.5-4Hz) and high-frequency band characteristics (20-30Hz); The frequency band sensitive mask is calculated based on the above dynamic contact state characteristics, where the low-frequency mask is dominated by the pressure mean offset: ; High-frequency masks are dominated by high-frequency volatility of pressure: ; in, The frequency band-sensitive sigmoid-like smoothing mapping function is calculated as follows:

[0032] In this embodiment, the frequency band mask smoothing coefficient Bandwidth artifact tolerance threshold , for Input.

[0033] The frequency band sensitive mask is applied to the corresponding frequency band features for decoupling and weighting, specifically: and , respectively applied to and Decoupling and weighting are performed to obtain the fused feature vector. and ; , ,in, The characteristic matrix is ​​multiplied element by element; The dynamic contact state characteristics of the force feedback are then normalized, specifically by using min-max normalization to adjust the pressure mean offset. and high-frequency volatility of pressure Mapping to the interval [0,1] respectively, the formula is: ; Includes: normalized pressure mean offset and normalized high-frequency volatility of pressure

[0034] Among them, for the pressure mean offset, gram, Grams (the reasonable pressure offset range for the covering electrode from complete detachment to excessive compression); for high-frequency pressure fluctuations, gram, Grams (covering the range of fluctuations from no movement to violent head movements).

[0035] Weighted low-frequency fusion features High-frequency fusion characteristics offset from the normalized mean pressure Normalized high-frequency volatility of pressure By directly concatenating the feature dimensions, we obtain the first... Dual-drive fusion feature vector of the channel:

[0036] The fusion feature vectors of all channels are concatenated along the channel dimension to form the final dual-drive fusion feature vector, which is then input into the pre-trained state recognition base model to obtain the state recognition result and the corresponding confidence score. In this embodiment, the state recognition result includes the fatigue state binary classification result (awake / fatigued) and the corresponding confidence score value (between 0 and 1).

[0037] This method can be extended to multi-classification tasks, such as attention state (high / medium / low) and emotional state (pleasant / sad / angry / calm).

[0038] The state recognition base model adopts a Transformer-CNN hybrid architecture with an input dimension of 14×250 (14 channels, 1 second of time-series data). From input to output, it passes through the backbone network layer and the pluggable classification adapter layer in sequence. The specific process is as follows: Backbone network layer feature extraction Input: Dual-drive fused feature vector (7028 dimensions), reshaped into a two-dimensional temporal feature matrix with 14 channels × 502 time points; First layer: Stack 3 Transformer encoders, each encoder contains 4 multi-head self-attention layers and feedforward neural network layers, with an output dimension of 14×64; Second layer: Stack two one-dimensional convolutional layers: First convolutional layer: kernel size = 3, stride = 1, output channels = 32; The second convolutional layer has the following parameters: kernel size = 3, stride = 2, and output channels = 16. Global average pooling: Perform global average pooling on the temporal features of the convolution output, compressing them into a 16-dimensional global feature vector.

[0039] Classification adapter layer Input: The 16-dimensional global feature vector output by the backbone network; The first fully connected layer has an input dimension of 16, an output dimension of 16, and uses ReLU as the activation function. The second fully connected layer has an input dimension of 16 and an output dimension of 1 (suitable for binary classification tasks). It has no activation function and outputs logits values.

[0040] State recognition results and confidence calculation Applying the Sigmoid activation function to the logits values ​​yields the predicted probability P: If P ≥ 0.5, the identification result is fatigue, and the confidence level is P; If P < 0.5, the identification result is "conscious" with a confidence level of 1. P.

[0041] The classification adapter layer is located at the end of the base model and is a pluggable structure. It can be adapted to tasks such as three-class classification and multi-label classification by simply adjusting its output dimension and activation function, without the need to retrain the backbone network.

[0042] S4. Determine whether the confidence level is lower than a preset dynamic threshold. If it is lower, generate a channel contact state adjustment command based on the contact confidence level and dynamic contact state characteristics and send it to the data generation module. The relevant operations in step S4 will be introduced with specific examples: In this embodiment, the process of adaptively adjusting the channel contact state through the channel contact state adjustment command is as follows: Figure 4 As shown, firstly, the confidence level is determined based on a preset dynamic threshold. In this embodiment, the preset dynamic threshold is 0.8. The confidence levels of the recognition results obtained from dual-drive fusion feature inference for each channel are iterated, and it is determined whether the confidence level is lower than the preset dynamic threshold (0.8). If it is lower, a channel contact state adjustment command is generated based on the single-channel contact confidence level and dynamic contact state characteristics and sent to the data generation module. The data generation module is existing technology and is divided into two categories: simulation data generation module and external acquisition device. The specific implementation is as follows: The simulation data generation module, implemented using Python 3.9, NumPy 1.23.5, and SciPy 1.10.0, simulates electrode pressure changes based on adjustment commands and generates EEG signals with corresponding artifact intensities. Its core logic is as follows: upon receiving an adjustment command, it updates the pressure value of the corresponding channel, calculates the artifact energy index based on the new pressure value, and then superimposes artifacts of corresponding intensity and frequency band onto the original clean EEG signal to generate a new synchronous EEG-force feedback data sequence.

[0043] External acquisition equipment: A commercially available EEG acquisition system with force feedback is used, such as the Neuroscan SynAmps2 EEG amplifier paired with a 14-lead electrode cap integrating an FSR thin-film force sensor. This device supports multimodal simultaneous acquisition, can output EEG signals and force feedback data in real time, and receives pressure adjustment commands from a host computer via an RS232 / USB interface to control the miniature electric actuators built into the electrodes to adjust the contact pressure.

[0044] The process of generating the channel contact state adjustment command includes the following steps: S41. Extract the set of abnormal channels whose single-channel confidence is lower than the preset dynamic threshold; S42. Calculate the local pressure compensation parameters based on the spatial topological position of each channel in the abnormal channel set and the direction of the current pressure deviation from the standard baseline; The spatial topological positions of each channel are obtained by pre-storing the international 10-20 system standard three-dimensional coordinates of the EEG electrodes, such as: Fp1 (-3,0,9), Fp2 (3,0,9), Cz (0,0,0), O1 (0,-8,3), O2 (0,8,3), T7 (-7,0,0), T8 (7,0,0), etc. These coordinates are industry-standard and are used for spatial weight calculation during subsequent multi-channel joint adjustment (in this embodiment, single-channel independent adjustment is used, and spatial weight is not used).

[0045] The formula for calculating the direction of the current pressure deviation from the standard baseline is: ,in, For symbolic functions, For the first The average pressure of the current sliding window in the channel. If This indicates that the current pressure is higher than the baseline and the pressure needs to be reduced; if This indicates that the current pressure is below the baseline and the pressure needs to be increased.

[0046] The method for obtaining local pressure compensation parameters is as follows: the adjustment direction is the opposite of the deviation direction. The adjustment step size is adaptively determined based on the pressure deviation magnitude: when When the weight is 1 gram, adjust the step size to 1 gram. When 2 grams When the weight is 4 grams, adjust the step size to 3 grams. when When the weight is 4 grams, adjust the step size to 5 grams; The compensation parameters include the target channel, adjustment direction, and adjustment step size. The adjustment step size is 1 gram to 5 grams, and the pre-compensation scenario uses a small step size of 1 gram to 2 grams. S43. Encapsulate the local pressure compensation parameters into a channel contact state adjustment command. The command uses the industry-standard JSON text format and includes at least the target adjustment channel identifier, adjustment direction, and adjustment step size. It supports direct parsing and execution by the simulation module and external hardware devices. A specific encapsulation example and field descriptions are as follows: { "timestamp":1717234567, "command_type": "channel_pressure_adjust", "version": "1.0", "commands": [ { "channel_id": "O1", "adjust_direction": "increase", "step": 2,"target_pressure": 8.0}, { "channel_id": "T7", "adjust_direction": "decrease", "step": 1, "target_pressure": 8.0} ]} Field description: timestamp: Unix timestamp (unit: seconds), used for time synchronization of instructions and data, ensuring strict alignment between adjustment operations and subsequent data acquisition; command_type: Command type identifier, used to distinguish different commands such as pressure regulation, data acquisition, and equipment calibration; channel_id: The international 10-20 system standard electrode identifier (such as Fp1, Cz, O1, etc.), which uniquely corresponds to one electrode channel; adjust_direction: Adjusts the direction, with values ​​of increase (increase pressure) or decrease (decrease pressure); Step: Adjust the step size (unit: grams), ranging from 1 to 5 grams; target_pressure: Target pressure value (optional), used for closed-loop regulation verification of external equipment.

[0047] Transmission and execution methods: Simulation data generation module: It directly receives JSON-formatted dictionary parameters via Python function calls, parses them, and updates the pressure value of the corresponding channel. External acquisition device: Sends UTF-8 encoded JSON string via USB serial port or TCP / IP protocol. After parsing, the device controls the built-in miniature electric actuator to perform pressure adjustment.

[0048] The formula for calculating the global average contact confidence level for regulation effect verification and anomaly handling is as follows: ; in, This represents the total number of brainwave channels. For the first The passage is Single-channel contact reliability of a sliding window.

[0049] After the adjustment command is issued, wait for 500ms for hardware execution and signal stabilization buffering time, and then recalculate the global average contact reliability. If the global average contact confidence level is ≥0.8, the adjustment is considered successful, and the process proceeds to step S5 to perform model fine-tuning; if 0.6 ≤ global average contact confidence level <0.8, the adjustment is considered partially effective, and steps S41-S43 are repeated for secondary adjustment, with the step size of the secondary adjustment halved. If the global average contact confidence level is still <0.6 after three consecutive adjustments, the adjustment is deemed a failure, triggering a level three anomaly: 1. An audible and visual alarm prompts the user to manually adjust the electrode position; 2. The system automatically switches to a degraded operation mode: only channels with a contact confidence level ≥0.7 are used for inference, and all model updates are paused; 3. If there is no user intervention within 10 minutes, the system automatically stops data acquisition and saves the anomaly log.

[0050] S5. Obtain the new EEG signal sequence and new force feedback data sequence returned by the data generation module after executing the channel contact state adjustment command, and fine-tune the parameters of the pre-trained state recognition basic model based on the new EEG signal sequence and new force feedback data sequence to generate and store the target personalized state recognition model.

[0051] The relevant operations in step S5 will be introduced with specific examples: In this embodiment, the parameter fine-tuning process is shown in Figure 5. The parameter efficient fine-tuning (PEFT) strategy is adopted. The core is to freeze the backbone network layer of the pre-trained state recognition basic model (which has learned general EEG temporal features) and only perform gradient updates on the terminal classification adapter layer. While ensuring recognition performance, the amount of computation and parameters is reduced by two orders of magnitude, which is suitable for the real-time operation requirements of edge devices.

[0052] The complete fine-tuning process is as follows: After adjustment, the data acquisition module executes the channel contact state adjustment command and then transmits the new EEG signal sequence and new force feedback data sequence after a 500ms delay (reserving buffer time for hardware to perform adjustment and signal stabilization).

[0053] The supervision signal generation process performs a dual-drive feature fusion on the new data, identical to S3, to obtain a 7002-dimensional fused feature vector. Supervision signals are then generated according to the following priority: First priority: User calibration labels (such as "awake" / "fatigued" states manually marked by the user via buttons, with labels encoded as 0 = awake, 1 = fatigued); Second priority: High confidence pseudo-labels. When there are no user-calibrated labels, the prediction results with a confidence level of ≥0.9 output by the model are used as pseudo-labels. Conflict handling: When the user calibration label is inconsistent with the high-confidence pseudo label, the user calibration label shall be used first, and the pseudo label data of the corresponding time period shall be discarded to avoid the accumulation of erroneous signals.

[0054] The loss function is calculated and backpropagation is used to update the parameters of the classification adapter layer. In this embodiment, for a binary classification state recognition task (awake / fatigued), the binary cross-entropy loss function (BCELoss) is used, and the AdamW optimizer is used for gradient update. The hyperparameters are set as follows: learning rate. Weight decay = weight_decay= β1=0.9, β2=0.999. After calculating the gradient of the loss function, only the parameters of the classification adapter layer are updated through backpropagation, while all weights of the backbone network layer remain frozen.

[0055] The number of trainable parameters is ≤10K, and the time for a single round of fine-tuning is ≤5 seconds.

[0056] Step S5 also includes a confidence gating update mechanism, specifically: data for the corresponding time period is allowed to enter the efficient fine-tuning process of model parameters only when the confidence of a single channel contact is ≥0.7 and the global average confidence is ≥0.8; otherwise, efficient fine-tuning of parameters is not performed, thus blocking low-quality data from polluting the model at the source. Exception scenario: If the contact confidence of all channels is <0.7, but the global average contact confidence is ≥0.6 and the model output confidence is ≥0.9, the data can be allowed to enter the lightweight fine-tuning process after manual confirmation. That is, only the parameters of the last fully connected layer in the classification adapter layer are updated, and all other network weights are frozen.

[0057] Exceptional Scenario Operation Details: (1) Triggering Conditions: All single-channel contact confidence scores are <0.7, but the global average contact confidence score is ≥0.6 and the model output confidence score is ≥0.9, and the above conditions are met for 3 consecutive sliding windows (total 1.5 seconds); (2) Manual Confirmation Process: The system pops up a confirmation dialog box, displaying the current contact confidence score heatmap of each channel and the status recognition curve of the last 10 seconds, and the user judges whether the current recognition result is accurate; (3) Judgment Criteria: When the user confirms that the recognition result is consistent with the actual state, the data is allowed to enter the lightweight fine-tuning process; (4) Lightweight Fine-tuning Parameters: Only update the last fully connected layer of the classification adapter layer, and adjust the learning rate to The batch size is halved to 16, the number of iterations is limited to 1, and all other network weights are frozen.

[0058] The method described in this invention also includes a system state machine control mechanism, specifically: establishing a system state machine that includes a normal contact state, a contact degradation state, a dynamic compensation state, an update freeze state, and a self-recovery state; the system dynamically switches operating states based on the global average contact confidence and the model output confidence. (1) Normal contact state: The global average contact confidence is ≥0.8 and the confidence is ≥0.8, and the inference and model update are performed normally; (2) Contact degradation state: 0.6≤global average contact confidence <0.8, and confidence <0.8, trigger abnormal channel detection, only inference, no fine-tuning; (3) Dynamic compensation status: Issue channel contact status adjustment instructions and pause model updates; if the global average contact confidence does not recover to above 0.6 within 5 seconds after adjustment, a data quality anomaly alert will be triggered. (4) Update the frozen state: if the global average contact confidence is <0.6, the model update is completely frozen and only basic inference is performed; (5) Self-recovery state: After adjustment, the global average contact confidence level rises to above 0.8, and the model update gradually recovers; (6) Insufficient model generalization state: global average contact confidence ≥ 0.8 and confidence < 0.8, triggering local fine-tuning of the classification adapter layer, and no data generation adjustment is performed.

[0059] In this embodiment, the state recognition includes at least one of fatigue state recognition, concentration state recognition, emotional state recognition, movement intention recognition, and neurorehabilitation state assessment.

[0060] Example 2 This embodiment is a further limitation of Embodiment 1. Based on the method described in Embodiment 1, this embodiment also includes a contact degradation prediction and early compensation mechanism and an individualized long-term drift compensation mechanism. The contact degradation prediction and early compensation mechanism is as follows: Based on the contact confidence time series of the past 30 seconds (200Hz sampling corresponds to 6000 frames), the ARIMA time series prediction model is used to predict the contact confidence change trend in the next 5 seconds; when the contact confidence is predicted to be lower than 0.6, the pre-compensation channel contact state adjustment is performed in advance, with an adjustment step size of 1 gram to 2 grams.

[0061] The individualized long-term drift compensation mechanism is as follows: Based on the user's usage history data for 7 consecutive days or more, pressure baseline drift curves and EEG characteristic baseline drift curves are constructed for each signal channel; the corresponding drift compensation parameters are automatically loaded when the user uses the device for the first time each day, and baseline correction is performed on the force feedback data and EEG characteristics acquired in real time.

[0062] The individualized long-term drift compensation mechanism is implemented as follows: - Data acquisition: Record the initial pressure baseline and EEG characteristic baseline (average power of Delta, Theta, Alpha, and Beta waves) of each channel during each user session; Drift curve fitting: When the user's cumulative usage time is ≥7 days and the effective data volume is ≥10 hours, a cubic polynomial is used to fit the pressure baseline drift curve of each channel. EEG baseline drift curve ,in For the number of days used; Baseline correction: Upon first use each day, the corresponding drift curve is automatically loaded to correct the real-time force feedback data and EEG characteristics; Update frequency: The drift curve is updated every 7 days, and a sliding window method is used to retain the data of the most recent 30 days for fitting.

[0063] Example 3: This example is a further limitation of Examples 1 and 2. This example provides a state closed-loop recognition system based on force feedback and EEG signals.

[0064] like Figure 6 As shown, the system described in this embodiment includes: a synchronous data acquisition module, a reliability assessment module, a dual-drive fusion module, a closed-loop control module, a fine-tuning update module, a gating control module, a state machine management module, a prediction compensation module, and a drift correction module; The synchronous data acquisition module is used to acquire the user's original EEG signal sequence in real time, as well as the multi-channel force feedback data sequence that is time-synchronized with the original EEG signal sequence. The credibility assessment module extracts the dynamic contact state features of each channel based on the multi-channel force feedback data sequence and calculates the contact credibility of the corresponding EEG signal. The dual-drive fusion module is used to extract the EEG features of the original EEG signal sequence and perform frequency band-sensitive dual-drive feature fusion with the contact confidence, and input it into the pre-trained state recognition basic model for forward inference to obtain the state recognition result and the confidence level corresponding to the result; The closed-loop control module is used to determine whether the confidence level is lower than a preset dynamic threshold. If it is lower, it generates a channel contact state adjustment command based on the contact confidence level and dynamic contact state characteristics and sends it to the data generation module. The fine-tuning update module is used to obtain the new EEG signal sequence and the new force feedback data sequence returned by the data generation module after executing the channel contact state adjustment command, and to fine-tune the parameters of the pre-trained state recognition basic model based on the new EEG signal sequence and the new force feedback data sequence, thereby generating and storing the target personalized state recognition model.

[0065] The gating control module is used to execute the trust gating update mechanism; The state machine management module is used to execute the system state machine control mechanism; The prediction and compensation module is used to perform contact degradation prediction and early compensation mechanisms; The drift correction module is used to implement an individualized long-term drift compensation mechanism.

[0066] Example 4 This embodiment further defines Embodiment 1, and provides performance simulation verification of the closed-loop method based on a public dataset.

[0067] This embodiment is the core verification embodiment of the present invention. Under two typical abnormal channel contact state scenarios—single-channel continuous loosening and multi-channel random transient interference—performance comparison tests of three schemes were completed. The core test results are shown in Tables 1 and 2. To verify the technical effect of the method described in this invention and its advantages over existing technologies, this embodiment uses a combination of software simulation and publicly available standard datasets to fully reproduce the method described in Embodiment 1. The present invention can be fully implemented and achieve the expected technical effect without relying on physical hardware.

[0068] (1) Simulation environment and data foundation This simulation experiment was conducted on a computer configured with an Intel Core i7-12700H processor and 32GB of memory. The software environment was Python 3.9, and the algorithm model and simulation process were built based on the PyTorch 1.13.0, SciPy 1.10.0 and NumPy 1.23.5 libraries.

[0069] The basic EEG data used was the publicly available fatigue driving dataset SEED-VIG, which is recognized in the field of brain-computer interfaces. This dataset contains multi-channel EEG signals and corresponding fatigue state labels (awake / fatigued) from 23 subjects, with a signal sampling rate of 200Hz. In this embodiment, complete test data from 10 subjects were selected for simulation verification, covering the full range of states from awake to mild fatigue to severe fatigue.

[0070] (2) Simulation data generation and comparison scheme settings To simulate artifact interference caused by abnormal channel contact states in real-world scenarios, clean EEG signals from the SEED-VIG dataset were used as a baseline. Simulated force feedback data synchronized with the EEG signals was generated procedurally, and artifact contamination was applied to the EEG signals based on abnormal force feedback states. The specific rules are as follows: Normal contact phase: Generate stable pressure values ​​(mean 8 g, variance 0.5 g) that conform to a Gaussian distribution for all EEG channels, corresponding to a good contact state of the corresponding channels, without adding extra artifacts to the EEG signals; Poor Contact Phase: 20% of the total test duration was randomly selected, along with two typical channels, O1 and T7, which are susceptible to head movement. Two abnormal modes were set: ① Pressure value dropped to 3–4 grams, simulating channel contact loosening; ② Force feedback data was superimposed with 20Hz–30Hz high-frequency fluctuations, simulating channel contact high-frequency interference. Based on the abnormal force feedback mode, matching simulation artifacts were superimposed on the clean EEG signals of the corresponding channels: low-frequency baseline drift below 0.5–4Hz was added when pressure was too low, and a sweeping sinusoidal motion artifact of the same frequency band was added when there were high-frequency fluctuations.

[0071] This embodiment sets up three control schemes to compare the performance of the method of the present invention. The three schemes completely address the defects of the prior art in the background art: Comparison with Scheme A (traditional model, corresponding to existing technology 2): The above-mentioned contaminated EEG signals are directly input into the pre-trained fatigue recognition model. The model uses the Transformer-CNN hybrid architecture that is completely consistent with the present invention to complete the pre-training. The inference process does not utilize any force feedback data, has no artifact suppression mechanism, no closed-loop adjustment, and no model adaptive update. Comparison scheme B (simple hardware threshold strategy, corresponding to existing technology 1): when the simulated force feedback value is lower than the preset threshold of 5 grams, the EEG signal of the corresponding channel is directly set to zero and removed. The rest of the process is the same as that of comparison scheme A. It only realizes the simple threshold alarm and channel removal of force feedback, without algorithm-level artifact suppression or closed-loop adaptive evolution. The present invention implements the method described in Example 1 in its entirety, namely, extracting dynamic contact state features based on simulated force feedback data, generating signal quality assessment indicators in the form of adaptive weight masks, completing the dual-drive feature fusion of EEG features and quality indicators, and inputting them into the pre-trained basic model to obtain recognition results and confidence scores; when the confidence score is lower than the preset dynamic threshold of 0.8, the closed-loop adjustment logic is triggered (in the simulation, after triggering, the simulation data generation adjustment is completed, and the force feedback and EEG data of the corresponding channel are restored to the normal contact state), and the high-quality new data obtained after adjustment is used to perform efficient parameter fine-tuning (PEFT) on the classification adapter layer at the end of the model.

[0072] The simulation data generation rules described above are based on bioelectrical mechanisms: a decrease in electrode contact pressure leads to an increase in contact impedance, which in turn introduces low-frequency baseline drift artifacts; high-frequency pressure fluctuations cause rapid changes in contact impedance, which in turn introduces high-frequency motion artifacts in the same frequency band. This invention establishes a quantitative mapping relationship between "pressure change → artifact characteristics" to ensure that the simulation data can realistically simulate signal contamination in actual acquisition scenarios.

[0073] (3) Simulation results and quantitative analysis Performance comparison tests of three schemes were conducted under two typical abnormal channel contact state scenarios: continuous loosening of a single channel and random instantaneous interference of multiple channels. The core test results are shown in Tables 1 and 2.

[0074] Table 1 Performance comparison of different solutions under abnormal channel contact conditions.

[0075] Table 2 Comparison of computational power and efficiency of model fine-tuning strategies

[0076] (4) Simulation conclusions The simulation experiments based on publicly available standard datasets quantify and verify the core beneficial effects of this invention: Significantly improved artifact suppression and recognition performance: The physical-algorithm dual-driven feature fusion method proposed in this invention can make full use of force feedback information to actively suppress artifact interference introduced by poor contact at the algorithm level. Under typical interference scenarios, the recognition accuracy is improved by more than 27 percentage points compared with the traditional solution of the existing technology, which solves the core defect of poor robustness of the model to the deterioration of the contact state in the existing technology. The closed-loop adaptive evolution mechanism is effective and feasible: Based on the model recognition confidence trigger, the closed-loop logic of "anomaly judgment - contact state adjustment instruction generation - data quality optimization - model fine-tuning" can complete the acquisition environment optimization and model self-evolution in a very short time. After adjustment, the model's recognition accuracy in similar interference scenarios is stable at around 90%, realizing a technological leap from passive reception to active optimization. The feasibility of edge computing has been fully verified: The parameter-efficient fine-tuning strategy adopted in this invention only requires updating less than 1% of the model parameters (approximately 9.6K parameters) to achieve recognition performance comparable to full fine-tuning. The required data volume and computation time are only 25% and 20% of those for full fine-tuning, respectively. On typical ARM Cortex-A series edge processors, matrix operations of this magnitude typically take hundreds of milliseconds to seconds, far below the time window requirements for real-time EEG analysis. This theoretically confirms the feasibility of this method running in real-time on portable edge EEG devices with limited computing power.

[0077] This simulation embodiment rigorously reproduces all the core technical steps described in the claims at the algorithm level and completes comparative testing with existing technologies on a recognized public dataset, fully demonstrating the advanced principle and predictable stable technical effect of the method of the present invention compared with existing technologies.

[0078] Example 5 This embodiment is a further limitation of Embodiment 1.

[0079] The pure software closed-loop generation method of this invention has strong scenario versatility and can be adapted to all EEG state recognition scenarios described in Example 1. All scenarios are implemented based on the pure software simulation framework of the core Example 1, requiring no dedicated hardware equipment. Only the corresponding public dataset and scenario-based parameters need to be replaced to complete the closed-loop generation and verification of personalized models. The following are the complete implementation processes for two typical application examples: (1) Vehicle fatigue state recognition scenario This scenario addresses the need for real-time monitoring of driver fatigue during long-term driving. It constructs a simulation environment based on the publicly recognized fatigue driving dataset SEED-VIG in the field of brain-computer interfaces, and generates a personalized fatigue recognition model for the driver group.

[0080] Dataset and Model Foundation: The dataset uses 14-lead EEG signals from 23 subjects in the SEED-VIG dataset and corresponding awake / fatigue binary classification labels. The pre-trained base model adopts a Transformer-CNN hybrid architecture, focusing on the power ratio feature of Delta waves (0.5–4Hz) and Alpha waves (8–13Hz).

[0081] Scenario-based parameter adaptation: Set the EEG sampling rate to 250Hz, and the force feedback sampling rate to be synchronized with the EEG; the artifact tolerance threshold τ=0.5, the confidence dynamic threshold 0.8; the standard contact pressure baseline is 8 grams, and the channel contact state adjustment step size is 1 gram to 5 grams.

[0082] Simulation verification process and results: Following the steps of core embodiment 1, typical interference scenarios such as road bumps and channel contact degradation caused by prolonged driving during vehicle operation are simulated. For example, at the 30-minute mark of the simulation, a sudden pressure fluctuation (change rate > 2 g / s) occurs in channel O1. The method of this invention automatically suppresses the adaptive weight mask of this channel to below 0.15, and simultaneously generates a channel contact state adjustment command to drive the simulation data generation module to restore a high-quality signal. After efficient fine-tuning of parameters based on the adjusted new data, the model's fatigue state recognition accuracy under subsequent similar interference scenarios remains stable at over 88%, an improvement of 27.2 percentage points compared to the traditional closed-loop model.

[0083] (2) Online Education Attention Status Recognition Scenarios This scenario addresses the need for real-time assessment and training of students' attention in online education. It constructs a simulation environment based on the publicly available EEG attention dataset SEED and generates a personalized attention recognition model for the student population.

[0084] Dataset and Model Foundation: Using 8-lead EEG signals from 15 subjects in the SEED dataset and corresponding high / medium / low attention three-classification labels, the pre-trained basic model focuses on the power ratio characteristics of Beta waves (13–30Hz) and Theta waves (4–7Hz) in the prefrontal cortex.

[0085] Scenario-based parameter adaptation: Parameters are optimized for younger users: standard contact pressure baseline 4g-8g, channel contact state adjustment step 1g-3g, and confidence dynamic threshold 0.75; in the fine-tuning stage, in addition to high-confidence pseudo-labels (confidence ≥0.9), simulated student classroom answer accuracy data can also be used as user calibration labels. When the calibration label is inconsistent with the pseudo-label, the calibration label is used first for model update.

[0086] Simulation verification process and results: Following the steps of core embodiment 1, typical interference scenarios such as student head movement and changes in channel contact state were simulated. Simulation verification showed that the method of this invention achieved a focus recognition accuracy of over 84% under the aforementioned interference scenarios, an improvement of over 25 percentage points compared to traditional models. Simultaneously, through an individualized long-term drift compensation mechanism, the cross-day recognition accuracy fluctuation can be controlled within 5%, meeting the needs of long-term continuous use in online education.

[0087] (3) Other extended scenarios The method of this invention can be further extended to the following application scenarios using the same pure software simulation framework: Emotional state recognition: Based on the DEAP public emotion dataset, generate a personalized emotion (pleasure / sadness / anger / calm) recognition model; Motion Intent Recognition: Based on the publicly available BCICompetition motion imagery dataset, generate an upper / lower limb motion intent recognition model; Neurological rehabilitation status assessment: Based on publicly available EEG datasets of stroke patients, a quantitative assessment model for the rehabilitation status of patients' motor function is generated.

[0088] All extended scenarios do not require modification of the core algorithm logic; only the corresponding dataset and scenario-based parameters need to be replaced to achieve complete closed-loop model generation, fully demonstrating the universality and scalability of this invention.

Claims

1. A state closed-loop recognition method based on force feedback and electroencephalogram (EEG) signals, characterized in that, The method includes the following steps: S1. Real-time acquisition of the user's original EEG signal sequence, and a multi-channel force feedback data sequence that is time-synchronized with the original EEG signal sequence; S2. Based on the multi-channel force feedback data sequence, extract the dynamic contact state features of each channel and calculate the contact reliability of the corresponding EEG signal. S3. Extract the EEG features of the original EEG signal sequence and perform frequency band-sensitive dual-drive feature fusion with the contact confidence, input it into the pre-trained state recognition basic model for forward inference, and obtain the state recognition result and the confidence corresponding to the result; S4. Determine whether the confidence level is lower than a preset dynamic threshold. If it is lower, generate a channel contact state adjustment command based on the contact confidence level and dynamic contact state characteristics and send it to the data generation module. S5. Obtain the new EEG signal sequence and new force feedback data sequence returned by the data generation module after executing the channel contact state adjustment command, and fine-tune the parameters of the pre-trained state recognition basic model based on the new EEG signal sequence and new force feedback data sequence to generate and store the target personalized state recognition model.

2. The state closed-loop recognition method based on force feedback and electroencephalogram (EEG) signals according to claim 1, characterized in that, The dynamic contact state characteristics include: the average pressure offset of each channel. and high-frequency volatility of pressure , This is the index value of the channel. The index value of the sliding window; The contact reliability The formula for calculation is: ,in, For mask smoothing coefficient, For the first The first channel, in the... The artifact energy index of a sliding window. The artifact tolerance threshold. It is a natural exponential function; The formula for calculation is: ,in, The standard contact pressure baseline, This is a reference value for normal pressure fluctuations. and These are the weighting coefficients. .

3. The state closed-loop recognition method based on force feedback and electroencephalogram (EEG) signals according to claim 2, characterized in that, The EEG characteristics of the original EEG signal sequence include: low-frequency band characteristics. and high frequency band characteristics ; The dual-drive feature fusion includes the following steps: S31. Calculate the low-frequency mask based on the dynamic contact state characteristics. and high-frequency mask , , ,in, For Sigmoid-class smoothing mapping functions; S32, will and , respectively applied to and Decoupling and weighting are performed to obtain the fused feature vector. and ; S33 will and Perform normalization; S34. The weighted result from step S32 and the normalized result from step S33 are concatenated to obtain the dual-drive feature fusion vector.

4. The state closed-loop recognition method based on force feedback and electroencephalogram (EEG) signals according to claim 3, characterized in that, The state recognition basic model adopts a Transformer-CNN hybrid architecture, which passes through a backbone network layer and a classification adapter layer from input to output. The backbone network layer includes 3 Transformer encoders, 2 one-dimensional convolutions and 1 global average pooling. The classification adapter layer includes 2 fully connected layers.

5. The state closed-loop recognition method based on force feedback and electroencephalogram (EEG) signals according to claim 4, characterized in that, The process of generating the channel contact state adjustment command includes the following steps: S41. Extract the set of abnormal channels whose single-channel confidence is lower than the preset dynamic threshold; S42. Calculate the local pressure compensation parameters based on the spatial topological position of each channel in the abnormal channel set and the direction of the current pressure deviation from the standard baseline; The compensation parameters include the target channel, adjustment direction, and adjustment step size, wherein the adjustment step size is 1 gram to 5 grams; S43. The local pressure compensation parameters are encapsulated into the channel contact state adjustment command, the command including at least the target adjustment channel identifier, adjustment direction and adjustment step size.

6. The state closed-loop recognition method based on force feedback and electroencephalogram (EEG) signals according to claim 5, characterized in that, The data generation module is an external acquisition device.

7. The state closed-loop recognition method based on force feedback and electroencephalogram (EEG) signals according to claim 6, characterized in that, The parameter fine-tuning specifically involves: freezing the weight parameters of the backbone network layer in the pre-trained state recognition base model; and updating the gradient only for the classification adapter layer of the state recognition base model.

8. The state closed-loop recognition method based on force feedback and electroencephalogram (EEG) signals according to claim 7, characterized in that, Step S5 also includes a confidence gating update method, specifically: when the confidence level is lower than a preset dynamic threshold, if the single-channel contact confidence level is ≥0.7 and the global average contact confidence level is ≥0.8, a channel contact state adjustment command is generated; Conversely, it will not be generated.

9. A state closed-loop recognition system based on force feedback and electroencephalogram (EEG) signals, characterized in that, The system is used to implement the method according to any one of claims 1 to 8, and the system includes: a synchronous data acquisition module, a reliability assessment module, a dual-drive fusion module, a closed-loop control module, and a fine-tuning update module; The synchronous data acquisition module is used to acquire the user's original EEG signal sequence in real time, as well as the multi-channel force feedback data sequence that is time-synchronized with the original EEG signal sequence. The credibility assessment module extracts the dynamic contact state features of each channel based on the multi-channel force feedback data sequence and calculates the contact credibility of the corresponding EEG signal. The dual-drive fusion module is used to extract the EEG features of the original EEG signal sequence and perform frequency band-sensitive dual-drive feature fusion with the contact confidence, and input it into the pre-trained state recognition basic model for forward inference to obtain the state recognition result and the confidence level corresponding to the result; The closed-loop control module is used to determine whether the confidence level is lower than a preset dynamic threshold. If it is lower, it generates a channel contact state adjustment command based on the contact confidence level and dynamic contact state characteristics and sends it to the data generation module. The fine-tuning update module is used to obtain the new EEG signal sequence and the new force feedback data sequence returned by the data generation module after executing the channel contact state adjustment command, and to fine-tune the parameters of the pre-trained state recognition basic model based on the new EEG signal sequence and the new force feedback data sequence, thereby generating and storing the target personalized state recognition model.

Images