Fatigue control system based on brain-computer intelligent wearable device

Abstract

The present application relates to the field of intelligent wearable devices, and proposes a fatigue regulation system based on brain-computer intelligent wearable devices, comprising an electroencephalogram signal conditioning and conversion module, a fatigue degree decoding module, and a fatigue feedback regulation module. First, the electroencephalogram signal conditioning and conversion module is used to collect electroencephalogram signals from the user's head. Then, a light-weight feature self-organizing pruning fusion network L-SFC-FNet is proposed, which uses multiple parameter-fixed layers to replace trainable parameter layers, thereby minimizing the training time and space cost of the model, and automatically matching the optimal parameter solution space for each user to perform real-time fatigue decoding on the electroencephalogram signals and calculate the real-time fatigue degree of the user. The fatigue feedback regulation module is used to apply a micro-current stimulus when the user is in a state of fatigue for a long time, prompting the user to concentrate and recover to a normal state, which is beneficial to improving work efficiency, creating more wealth value, and better playing a social value.

Description

Technical Field

[0001] This invention relates to the field of smart wearable devices, and in particular to a fatigue control system based on brain-computer interface smart wearable devices. Background Technology

[0002] The fast pace of modern life, coupled with high work pressure and numerous tasks, easily leads to work fatigue. Work fatigue, also known as work-related fatigue, is a temporary decrease in work capacity after continuous labor. It is a physiological and psychological state, a defense mechanism adopted by the body to adapt to changes in the internal and external environment. However, work fatigue can pose serious hidden dangers to one's work. For example, driving while fatigued greatly increases the likelihood of a major traffic accident, and fatigue among operators in high-risk industries can easily cause significant personal and property losses, leading to incalculable consequences. Therefore, timely monitoring of fatigue and providing early warnings before it occurs can effectively reduce the probability of major accidents and maximize the protection of personal safety, equipment safety, and property safety.

[0003] In addition, intervention and regulation of individuals prone to fatigue can alleviate their fatigue state, extend their safe working hours, create more wealth, and better realize their social value. Transcranial microcurrent stimulation (TCMS) is a form of nerve stimulation that delivers weak electrical currents to the brain via electrodes placed on the scalp to improve various symptoms. The principle behind this therapy is to stimulate neurons in the brain, regulate the brain's electrical activity, thereby affecting the release of neurotransmitters, increasing cerebral blood flow, and enhancing the brain's metabolic function, thus improving fatigue symptoms.

[0004] However, due to the varying scenarios and needs across different industries, it's impossible to use fixed, large-scale equipment for real-time fatigue monitoring. A more wearable, portable, and lightweight device is essential. Furthermore, traditional fatigue monitoring methods, such as blood oxygen saturation, heart rate, and respiration, carry insufficient information to comprehensively reflect an individual's fatigue state. Electroencephalography (EEG), as a macroscopic reflection of the electrophysiological activity of brain nerve cells, contains a wealth of physiological information, reflecting the brain's activity and cognitive state, and is increasingly becoming the preferred method for fatigue monitoring. In addition, EEG signals are easy to acquire and have high temporal resolution, providing a foundation for real-time fatigue monitoring.

[0005] However, EEG signals are characterized by nonlinearity, complex features, and low signal-to-noise ratio, making their identification extremely difficult. Traditional EEG identification methods mainly fall into two categories: manual interpretation, requiring experienced experts, and decoding using machine learning algorithms, which necessitates manual extraction and analysis of EEG features. Both methods require manual verification, which is unsuitable for long-term monitoring and can lead to the omission of important EEG features due to subjective factors. Deep learning, as the most advanced theory in machine learning, possesses significant advantages in handling complex information and massive datasets and has been widely applied in EEG signal research. Deep learning is an end-to-end learning method capable of directly extracting and learning deeper intrinsic representations from input signals for classification. Currently, numerous deep learning architectures have been proposed and applied to the analysis and decoding of EEG signals.

[0006] However, existing deep learning models are overly complex in design and often require graphics processing units (GPUs) for training, which is prohibitively expensive in terms of both time and space for portable devices. Therefore, designing a lightweight model would facilitate deployment on real-world devices and reduce training costs. Summary of the Invention

[0007] The technical problem to be solved by the present invention is to provide a fatigue regulation system based on brain-computer intelligent wearable devices. By collecting the user's electroencephalogram (EEG) signals, the system uses a lightweight deep learning model to monitor the user's fatigue state in real time, and regulates the user's fatigue state through transcranial microcurrent stimulation technology, thus forming a closed loop of fatigue monitoring and fatigue regulation.

[0008] The technical solution adopted in this invention is: a fatigue control system based on a brain-computer interface intelligent wearable device, comprising, in sequence: an EEG signal conditioning and conversion module, a fatigue level decoding module, and a fatigue feedback control module. The EEG signal conditioning and conversion module conditions and converts the simulated EEG signals from the user's head into analog-to-digital signals, which are then transmitted to the fatigue level decoding module. The fatigue level decoding module proposes a lightweight feature self-organizing pruning and fusion network (L-SFC-FNet) to perform real-time fatigue decoding of the EEG signals and calculate the user's real-time fatigue level. The fatigue feedback control module applies microcurrent stimulation to the user based on their real-time fatigue level to improve their concentration and alleviate fatigue. The system has two modes: a calibration mode and a working mode. In calibration mode, the EEG signals are stored and used for personalized model matching; in working mode, the EEG signals are directly used for classifying wakefulness and fatigue states.

[0009] The aforementioned EEG signal conditioning and conversion module includes: EEG electrodes, a multi-channel EEG acquisition module for EEG signal conditioning and analog-to-digital conversion, a microprocessor unit for controlling the multi-channel EEG acquisition module, and a first power supply module connected to the multi-channel EEG acquisition module and the microprocessor unit respectively; wherein,

[0010] The EEG electrodes consist of several low-impedance electrodes coated with a thin layer of Ag / AgCl, featuring low DC offset voltage and high signal-to-noise ratio. They are divided into acquisition electrodes, reference electrodes, and interference-reducing electrodes. The acquisition electrodes are responsible for monitoring the user's simulated EEG signals in four channels: FP1, FP2, F7, and F8. The electrode distribution follows the 10 / 20 international standard lead pattern. The reference electrode is located at the mastoid process behind the left ear, serving as a potential reference point for the acquisition electrodes. The interference-reducing electrode is located at the mastoid process behind the right ear, used to eliminate common-mode interference introduced by the human body and improve the signal-to-noise ratio of the EEG signal.

[0011] The multi-channel EEG acquisition module integrates multiple low-noise programmable gain amplifiers (PGAs) with high common-mode rejection ratios for amplifying analog EEG signals, a high-precision multi-channel analog-to-digital converter (ADC) for converting analog EEG signals into EEG signals, and a right leg drive circuit (RLD) for reducing common-mode interference. The RLD uses operational amplifiers to accumulate the differential input signals of each PGA and compares them with the common-mode voltage reference point. It outputs a reverse voltage signal to the human body through the interference-reducing electrodes to cancel the common-mode signal.

[0012] The microprocessor unit consists of a 32-bit low-power system-on-chip (SoC) chip, featuring GPIO, SPI, and I / O pins. 2 C. The UART interface transmits EEG signals to the fatigue level decoding module via Bluetooth BLE4.2 protocol;

[0013] The first power supply module consists of a single rechargeable lithium battery, a charging management module, and a set of ultra-low dropout linear regulators. Furthermore, the single rechargeable lithium battery has a capacity of no less than 2000mAh and a supply voltage of 3.7–4.2V. The charging management module features constant current / constant voltage charging mode, reverse charging protection, and automatic recharging. The charging voltage is limited to 5V, providing a safe charging guarantee for the single rechargeable lithium battery. The ultra-low dropout linear regulator converts the voltage supplied by the single rechargeable lithium battery to provide a voltage that meets the power supply requirements.

[0014] The fatigue level decoding module proposes a lightweight feature self-organizing pruning and fusion network, L-SFC-FNet, which is optimized for the rich temporal, frequency, and spatial features contained in EEG signals. To improve computational efficiency, the model uses multiple layers with fixed parameters instead of layers with trainable parameters, minimizing training time and space costs and automatically matching the optimal parameter solution space for each user. The model includes the following sequentially connected layers:

[0015] 1) A multi-scale frequency domain feature extraction module: Considering that EEG signals can be divided into different frequency bands, and each frequency band contains frequency domain information of different scales, this model uses a multi-scale frequency domain feature extraction module to extract frequency domain information of different frequency bands. This module includes several wavelet decomposition layers connected in sequence. Each wavelet decomposition layer uses the db4 wavelet to decompose the input signal into two parts: high-frequency signal and low-frequency signal. The high-frequency signal and low-frequency signal are further decomposed. The decomposition level is 3 levels. The decomposed signals of each frequency band are concatenated and spliced ​​as the output of this module.

[0016] 2) A spatial domain feature attention layer is implemented using grouped convolution. The number of convolution kernels is (32×m), where m is the number of frequency bands obtained by the multi-scale frequency domain feature extraction module. The length of the convolution kernel is (C×1), where C is the number of EEG signal channels. The convolution kernels are divided into m groups. The k-th group of convolution kernels is responsible for mapping the k-th frequency band of the input to a 32-dimensional high-dimensional feature. Finally, the high-dimensional features mapped by the m groups of convolution kernels are concatenated to obtain an output of (32×m).

[0017] 3) A first batch normalization layer is used to accelerate model training and reduce overfitting;

[0018] 4) A first activation function layer using the GELU activation function, the formula is as follows:

[0019] 5) For a time-domain uncertainty measurement layer, the 2-norm of the feature is calculated along the time dimension, using the following formula:

[0020] 6) A first linear layer with 64 neurons;

[0021] 7) A second activation function layer using the GELU activation function;

[0022] 8) A second linear layer with 2 neurons;

[0023] 9) A third activation function layer using the Softmax activation function, the formula is as follows: Where L = 2;

[0024] The model takes EEG signals as input and the user's current fatigue level (awake / fatigued) as output. The trainable layers include only three layers: the spatial feature attention layer, the first linear layer, and the second linear layer. The parameters of the remaining layers do not need to be updated, which greatly reduces the training cost. The model can not only be deployed on resource-constrained devices, but also iteratively train the model according to the user's personalized information to match the user's current fatigue state.

[0025] The fatigue feedback control module includes: a second power supply module, a microcurrent generation module, an output electrode, and a communication module; wherein,

[0026] The second power supply module consists of a battery pack, an isolation voltage regulator module, and a step-down power supply chip. Further, the battery pack consists of four rechargeable lithium batteries connected in series, providing a power supply voltage of 14.8V to 16.8V. The isolation voltage regulator module uses a UWF1205S-1WR3 module manufactured by MoreSun, which has a wide input voltage and constant output voltage, providing a 5V output voltage. The step-down power supply chip receives the 5V output voltage from the isolation voltage regulator module and provides the 3.3V and 1.8V voltages required for the normal operation of the communication module.

[0027] The microcurrent generating module consists of a boost module and a pulse modulation unit. Furthermore, the boost module uses a B1524S-1WR3 module manufactured by K-Cut to boost the voltage provided by the battery pack to 24V. The pulse modulation unit takes 24V as input and generates a periodic pulse current at the output electrode by controlling the switching of the transistor.

[0028] The output electrodes are connected to the user's earlobes and forehead, and use microcurrent stimulation to relieve fatigue and improve concentration.

[0029] The communication module receives the decoding result from the fatigue level decoding module via the Bluetooth BLE4.2 protocol, and drives the microcurrent generation module to apply microcurrent stimulation to the user based on the decoding result.

[0030] When the device is in calibration mode, the user needs to collect 5 minutes of data each in both awake and fatigued states for calibration. The network parameters are then adaptively adjusted to match the wearer's personalized physiological signal characteristics. The specific steps are as follows:

[0031] 1) User B needs to determine whether they are awake (score ≤ 5 points) or fatigued (score > 5 points) according to the Karolinska Sleepiness Scale. Data should be collected for 5 minutes each in the awake and fatigued states, and recorded as follows: and In the formula, C represents the number of EEG channels, and Γ represents the number of data points;

[0032] 2) and Slicing along the time dimension, with each slice being 1 second long and non-overlapping, yields... and In the formula, S represents the number of samples;

[0033] 3) Order It is a vector of all zeros. A vector consisting entirely of 1s, and and Combined into training set

[0034] 4) Using the training set Ω B A lightweight neural network was trained using the Adam optimizer and backpropagation algorithm, with a learning rate set to 0.001, to obtain the personalized matching model for user B. B .

[0035] Real-time control of fatigue status when the equipment is in working mode includes the following steps:

[0036] 1) Initialize fatigue level to 0;

[0037] 2) Every 5 seconds, the EEG signal conditioning and conversion module 1 collects 1 second of the user B's EEG signal and sends it to the fatigue level decoding module 2;

[0038] 3) The fatigue level decoding module 2 uses the proposed L-SFC-FNet model to classify the 1-second EEG signal as either "awake" or "fatigued".

[0039] 4) If the classification result is "fatigue state", then the fatigue level increases by 1; otherwise, the fatigue level remains unchanged.

[0040] 5) If the fatigue level is ≥100, the fatigue feedback control module 3 generates microcurrent stimulation for 1 minute, and at the same time, the acquisition process of the EEG signal conditioning and conversion module 1 is paused, the fatigue level is set to 0, and the process returns to step 2) until user B stops using the system; if the fatigue level is <100, the process returns directly to step 2) until user B stops using the system. Attached Figure Description

[0041] Figure 1 This is a block diagram of the fatigue control system based on a brain-computer interface intelligent wearable device according to the present invention;

[0042] Figure 2 This is a schematic diagram of the fatigue control system based on a brain-computer interface intelligent wearable device according to the present invention.

[0043] Figure 3 This is a block diagram of the EEG signal conditioning and conversion module of the present invention;

[0044] Figure 4 This is a block diagram of the lightweight feature self-organizing pruning and fusion network L-SFC-FNet proposed in this invention;

[0045] Figure 5 This is a block diagram of the fatigue feedback control module of the present invention;

[0046] Figure 6 This is a flowchart of the calibration steps for the lightweight feature self-organizing pruning and fusion network L-SFC-FNet proposed in this invention.

[0047] Figure 7 This is a flowchart of the real-time control steps for fatigue state according to the present invention. Detailed Implementation

[0048] The fatigue control system based on brain-computer interface intelligent wearable device of the present invention will be described in detail below with reference to the embodiments and accompanying drawings.

[0049] like Figure 1 As shown, the fatigue control system based on a brain-computer interface smart wearable device includes, in sequence: an EEG signal conditioning and conversion module 1, a fatigue level decoding module 2, and a fatigue feedback control module 3. The EEG signal conditioning and conversion module 1 conditions and performs analog-to-digital conversion on the simulated EEG signals from the user's head, converting them into EEG signals and transmitting them to the fatigue level decoding module 2. The fatigue level decoding module 2 proposes a lightweight feature self-organizing pruning and fusion network L-SFC-FNet to perform real-time fatigue decoding on the EEG signals and calculate the user's real-time fatigue level. The fatigue feedback control module 3 applies microcurrent stimulation to the user based on the user's real-time fatigue level to improve the user's concentration and alleviate fatigue. The system has two modes: a calibration mode and a working mode. In calibration mode, the EEG signals are stored and used for personalized model matching. In working mode, the EEG signals are directly used for classifying wakefulness and fatigue states.

[0050] like Figure 2 As shown in the figure, in a specific embodiment, the fatigue control system based on the brain-computer intelligent wearable device has an EEG signal conditioning and conversion module 1 and a fatigue feedback control module 3 placed inside the safety helmet. It communicates with the fatigue level decoding module 2 via Bluetooth BLE4.2 protocol. The EEG electrodes 11 and output electrodes 33 are placed near the brim inside the safety helmet and are connected to the EEG signal conditioning and conversion module 1 and the fatigue feedback control module 3 via FPC flexible flat cables. The tightness of the safety helmet can be adjusted to maintain good contact with the forehead.

[0051] like Figure 3 As shown, the EEG signal conditioning and conversion module 1 includes: EEG electrodes 11, a multi-channel EEG acquisition module 12 for EEG signal conditioning and analog-to-digital conversion, a microprocessor unit 13 for controlling the multi-channel EEG acquisition module 12, and a first power supply module 14 connected to the multi-channel EEG acquisition module 12 and the microprocessor unit 13 respectively; wherein,

[0052] The EEG electrode 11 consists of several low-impedance electrodes coated with a thin layer of Ag / AgCl, featuring low DC offset voltage and high signal-to-noise ratio. It is divided into acquisition electrodes, reference electrodes, and interference removal electrodes. The acquisition electrodes are responsible for monitoring the simulated EEG signals of the user's four channels: FP1, FP2, F7, and F8. The distribution of the electrodes follows the 10 / 20 international standard leads. The reference electrode is located at the mastoid process behind the left ear, serving as the potential reference point for the acquisition electrodes. The interference removal electrode is located at the mastoid process behind the right ear, used to eliminate common-mode interference introduced by the human body and improve the signal-to-noise ratio of the EEG signal.

[0053] The multi-channel EEG acquisition module 12 integrates multiple low-noise programmable gain amplifiers (PGAs) with high common-mode rejection ratios for amplifying analog EEG signals, a high-precision multi-channel analog-to-digital converter (ADC) for converting analog EEG signals into EEG signals, and a right leg drive circuit (RLD) for reducing common-mode interference. The RLD uses operational amplifiers to accumulate the differential input signals of each PGA and compares them with the common-mode voltage reference point. It outputs a reverse voltage signal to the human body through the interference-reducing electrode to cancel the common-mode signal.

[0054] The microprocessor unit 13 consists of a 32-bit low-power system-on-chip (SoC) chip, and has GPIO, SPI, and I / O pins. 2 C. The UART interface transmits EEG signals to the fatigue level decoding module 2 via Bluetooth BLE4.2 protocol;

[0055] The first power supply module 14 consists of a single rechargeable lithium battery 141, a charging management module 142, and a set of ultra-low dropout linear regulators 143. Further, the single rechargeable lithium battery 141 has a capacity of not less than 2000mAh and a supply voltage of 3.7–4.2V. The charging management module 142 has constant current / constant voltage charging mode, reverse charging protection, and automatic recharging functions. The charging voltage is limited to 5V, providing a safe charging guarantee for the single rechargeable lithium battery 141. The ultra-low dropout linear regulators 143 convert the voltage supplied by the single rechargeable lithium battery 141 to provide a voltage that meets the power supply requirements.

[0056] like Figure 4As shown, the fatigue level decoding module 2 proposes a lightweight feature self-organizing pruning and fusion network L-SFC-FNet. This model is optimized for the rich temporal, frequency, and spatial features contained in EEG signals. Simultaneously, to improve computational efficiency, the model uses multiple layers with fixed parameters instead of layers with trainable parameters, minimizing the training time and space costs of the model and automatically matching the optimal parameter solution space for each user. The model includes the following sequentially connected layers:

[0057] 1) A multi-scale frequency domain feature extraction module, which includes several wavelet decomposition layers connected in sequence. Each wavelet decomposition layer uses the db4 wavelet to decompose the input signal into two parts: high-frequency signal and low-frequency signal. The high-frequency signal and low-frequency signal are further decomposed. The decomposition level is 3. The decomposed frequency band signals are concatenated and spliced ​​as the output of the module.

[0058] 2) A spatial domain feature attention layer is implemented using grouped convolution. The number of convolution kernels is (32×m), where m is the number of frequency bands obtained by the multi-scale frequency domain feature extraction module. The length of the convolution kernel is (C×1), where C is the number of EEG signal channels. The convolution kernels are divided into m groups. The k-th group of convolution kernels is responsible for mapping the k-th frequency band of the input to a 32-dimensional high-dimensional feature. Finally, the high-dimensional features mapped by the m groups of convolution kernels are concatenated to obtain an output of (32×m).

[0059] 3) A first batch normalization layer is used to accelerate model training and reduce overfitting;

[0060] 4) A first activation function layer using the GELU activation function, the formula is as follows:

[0061] 5) For a time-domain uncertainty measurement layer, the 2-norm of the feature is calculated along the time dimension, using the following formula:

[0062] 6) A first linear layer with 64 neurons;

[0063] 7) A second activation function layer using the GELU activation function;

[0064] 8) A second linear layer with 2 neurons;

[0065] 9) A third activation function layer using the Softmax activation function, the formula is as follows: Where L = 2;

[0066] The model takes EEG signals as input and the user's current fatigue level (awake / fatigued) as output. The trainable layers include only three layers: the spatial feature attention layer, the first linear layer, and the second linear layer. The parameters of the remaining layers do not need to be updated, which greatly reduces the training cost. The model can not only be deployed on resource-constrained devices, but also iteratively train the model according to the user's personalized information to match the user's current fatigue state.

[0067] like Figure 5 As shown, the fatigue feedback control module 3 includes: a second power supply module 31, a microcurrent generation module 32, an output electrode 33, and a communication module 34; wherein,

[0068] The second power supply module 31 consists of a battery pack 311, an isolation voltage regulator module 312, and a step-down power supply chip 313. Further, the battery pack 311 consists of four rechargeable lithium batteries connected in series, providing a power supply voltage of 14.8V to 16.8V. The isolation voltage regulator module 312 adopts the UWF1205S-1WR3 module with wide voltage input and constant voltage output manufactured by MoreSun, providing a 5V output voltage. The step-down power supply chip 313 receives the 5V output voltage from the isolation voltage regulator module 312 and provides the 3.3V and 1.8V voltages required for normal operation of the communication module 34.

[0069] The micro-current generating module 32 consists of a boost module 321 and a pulse modulation unit 322. Further, the boost module 321 adopts a B1524S-1WR3 module manufactured by K-Cut Company to boost the voltage provided by the battery pack 311 to 24V; the pulse modulation unit 322 takes 24V voltage as input and generates a periodic pulse current at both ends of the output electrode 33 by controlling the switching of the transistor.

[0070] The output electrode 33 is connected to the user's earlobe and forehead, and uses microcurrent stimulation to relieve fatigue and improve concentration.

[0071] The communication module 34 receives the decoding result from the fatigue level decoding module 2 via the Bluetooth BLE4.2 protocol, and drives the microcurrent generation module 32 to apply microcurrent stimulation to the user based on the decoding result.

[0072] like Figure 6 As shown, when the device is in calibration mode, the user needs to collect 5 minutes of data each in both awake and fatigued states for calibration. The network parameters are then adaptively adjusted to match the wearer's personalized physiological signal characteristics. The specific steps are as follows:

[0073] 1) User B needs to determine whether they are awake (score ≤ 5 points) or fatigued (score > 5 points) according to the Karolinska Sleepiness Scale. Data should be collected for 5 minutes each in the awake and fatigued states, and recorded as follows: and In the formula, C represents the number of EEG channels, and Γ represents the number of data points;

[0074] 2) and Slicing along the time dimension, with each slice being 1 second long and non-overlapping, yields... and In the formula, S represents the number of samples;

[0075] 3) Order It is a vector of all zeros. A vector consisting entirely of 1s, and and Combined into training set

[0076] 4) Using the training set Ω B A lightweight neural network was trained using the Adam optimizer and backpropagation algorithm, with a learning rate set to 0.001, to obtain the personalized matching model for user B. B .

[0077] like Figure 7 As shown, when the equipment is in working mode, the real-time control of fatigue status includes the following steps:

[0078] 1) Initialize fatigue level to 0;

[0079] 2) Every 5 seconds, the EEG signal conditioning and conversion module 1 collects 1 second of the user B's EEG signal and sends it to the fatigue level decoding module 2;

[0080] 3) The fatigue level decoding module 2 uses the proposed L-SFC-FNet model to classify the 1-second EEG signal as either "awake" or "fatigued".

[0081] 4) If the classification result is "fatigue state", then the fatigue level increases by 1; otherwise, the fatigue level remains unchanged.

[0082] 5) If the fatigue level is ≥100, the fatigue feedback control module 3 generates microcurrent stimulation for 1 minute, and at the same time, the acquisition process of the EEG signal conditioning and conversion module 1 is paused, the fatigue level is set to 0, and the process returns to step 2) until user B stops using the system; if the fatigue level is <100, the process returns directly to step 2) until user B stops using the system.

[0083] Those skilled in the art will readily understand that the above description is merely the preferred embodiment of the present invention and is not intended to limit the invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A fatigue regulation system based on a brain-computer interface intelligent wearable device, comprising, in sequence: an EEG signal conditioning and conversion module (1), a fatigue level decoding module (2), and a fatigue feedback regulation module (3), characterized in that, The EEG signal conditioning and conversion module (1) conditions and converts the simulated EEG signal from the user's head into an analog-to-digital signal, which is then transmitted to the fatigue level decoding module (2). The fatigue level decoding module (2) proposes a lightweight feature self-organizing pruning and fusion network L-SFC-FNet to perform real-time fatigue decoding on the EEG signal and calculate the user's real-time fatigue level. The fatigue feedback control module (3) applies microcurrent stimulation to the user based on the user's real-time fatigue level to improve the user's concentration and alleviate fatigue. The system has two modes: calibration mode and working mode. When the system is in calibration mode, the EEG signals will be stored and used for personalized model matching. When the system is in working mode, the EEG signals will be directly used for classification of awake and fatigued states. Among them, the fatigue level decoding module (2) proposes a lightweight feature self-organizing pruning and fusion network L-SFC-FNet. This model is optimized for the rich temporal, frequency, and spatial features contained in EEG signals. At the same time, in order to improve computational efficiency, the model uses multiple layers with fixed parameters instead of layers with trainable parameters, which minimizes the training time and space cost of the model and can automatically match the optimal parameter solution space for each user. The model includes the following sequentially connected layers: 1) A multi-scale frequency domain feature extraction module, including several wavelet decomposition layers connected in sequence. Each wavelet decomposition layer uses the db4 wavelet to decompose the input signal into two parts: high-frequency signal and low-frequency signal. The high-frequency signal and low-frequency signal are further decomposed. The decomposition level is 3 levels. The decomposed frequency band signals are concatenated and spliced ​​as the output of the module. 2) A spatial feature attention layer, implemented using grouped convolutions, with the number of convolutional kernels being... , The number of frequency bands obtained by the multi-scale frequency domain feature extraction module; the length of the convolution kernel is , The number of EEG signal channels; the convolution kernel is divided into Group, No. The group convolution kernel is responsible for processing the input into the first convolution kernel. Each frequency band is mapped to a 32-dimensional high-dimensional feature, and finally... The high-dimensional features obtained by mapping the convolutional kernels are concatenated to obtain... The output; 3) A first batch normalization layer is used to accelerate model training and reduce overfitting; 4) A first activation function layer using the GELU activation function, the formula is as follows: ; 5) For a time-domain uncertainty measurement layer, the 2-norm of the feature is calculated along the time dimension, using the following formula: ; 6) A first linear layer with 64 neurons; 7) A second activation function layer using the GELU activation function; 8) A second linear layer with 2 neurons; 9) A third activation function layer, using the Softmax activation function, with the following formula: ,in ; The model takes EEG signals as input and outputs the user's current level of fatigue, which is divided into alertness or fatigue. The trainable layers include only three layers: a spatial feature attention layer, a first linear layer, and a second linear layer. The parameters of the remaining layers do not need to be updated, which greatly reduces the training cost. The model can not only be deployed on resource-constrained devices, but also iteratively train the model according to the user's personalized information to match the user's current fatigue state.

2. The fatigue control system based on a brain-computer interface intelligent wearable device according to claim 1, characterized in that, The EEG signal conditioning and conversion module (1) includes: EEG electrodes (11), a multi-channel EEG acquisition module (12) for EEG signal conditioning and analog-to-digital conversion, a microprocessor unit (13) for controlling the multi-channel EEG acquisition module (12), and a first power supply module (14) connected to the multi-channel EEG acquisition module (12) and the microprocessor unit (13) respectively; wherein, The EEG electrode (11) consists of several low-impedance electrodes covered with a thin layer of Ag / AgCl, which have the characteristics of low DC offset voltage and high signal-to-noise ratio. It is divided into acquisition electrode, reference electrode and interference removal electrode. The acquisition electrode is responsible for monitoring the analog EEG signals of the user's four channels FP1, FP2, F7 and F8. The distribution of the electrodes follows the 10 / 20 international standard leads. The reference electrode is located at the mastoid process behind the left ear as the potential reference point of the acquisition electrode. The interference removal electrode is located at the mastoid process behind the right ear to eliminate common-mode interference introduced by the human body and improve the signal-to-noise ratio of the EEG signal. The multi-channel EEG acquisition module (12) integrates multiple low-noise programmable gain amplifiers (PGAs) with high common-mode rejection ratios for amplifying analog EEG signals, a high-precision multi-channel analog-to-digital converter (ADC) for converting analog EEG signals into EEG signals, and a right leg drive circuit (RLD) for reducing common-mode interference. The right leg drive circuit uses an operational amplifier to accumulate the differential input signals of each low-noise programmable gain amplifier and compares them with the common-mode voltage reference point. It outputs a reverse voltage signal to the human body through the interference removal electrode to cancel the common-mode signal. The microprocessor unit (13) consists of a 32-bit low-power system-on-chip (SoC) chip, and has GPIO, SPI, and I / O pins. 2 C. UART interface, transmits EEG signals to fatigue level decoding module (2) via Bluetooth BLE4.2 protocol; The first power supply module (14) consists of a single rechargeable lithium battery (141), a charging management module (142), and a set of ultra-low dropout linear regulators (143). Further, the single rechargeable lithium battery (141) has a capacity of not less than 2000mAh and a power supply voltage of 3.7~4.2V. The charging management module (142) has constant current / constant voltage charging mode, anti-reverse charging, and automatic recharging functions. The charging voltage is limited to 5V, providing a safe charging guarantee for the single rechargeable lithium battery (141). The ultra-low dropout linear regulator (143) converts the voltage provided by the single rechargeable lithium battery (141) to provide a voltage that meets the power supply requirements.

3. The fatigue control system based on a brain-computer interface intelligent wearable device according to claim 1, characterized in that, The fatigue feedback control module (3) includes: a second power supply module (31), a microcurrent generation module (32), an output electrode (33), and a communication module (34); wherein, The second power supply module (31) consists of a battery pack (311), an isolation voltage regulator module (312), and a step-down power supply chip (313). Further, the battery pack (311) is composed of rechargeable lithium batteries connected in series. The isolation voltage regulator module (312) converts the voltage of the battery pack (311) into a 5V output voltage. The step-down power supply chip (313) receives the 5V output voltage from the isolation voltage regulator module (312) and provides the 3.3V and 1.8V voltages required for normal operation of the communication module (34). The microcurrent generating module (32) consists of a boost module (321) and a pulse modulation unit (322). Further, the boost module (321) pumps the voltage provided by the battery pack (311) to 24V; the pulse modulation unit (322) takes 24V voltage as input and generates periodic pulse microcurrents at both ends of the output electrode (33) by controlling the switching of the transistor. The output electrode (33) is connected to the user's earlobe and forehead, and uses microcurrent stimulation to relieve fatigue and improve concentration. The communication module (34) receives the decoding result from the fatigue level decoding module (2) via the Bluetooth BLE4.2 protocol, and drives the microcurrent generation module (32) to apply microcurrent stimulation to the user based on the decoding result.

4. The fatigue control system based on a brain-computer interface intelligent wearable device according to claim 1, characterized in that, When the system is in calibration mode, the user needs to collect 5 minutes of data in both awake and fatigued states for calibration. When the system is in working mode, the fatigue level decoding module (2) will classify the awake or fatigued state in real time and send the classification results to the fatigue feedback control module (3) for real-time control of the fatigue state.

5. The fatigue control system based on a brain-computer interface intelligent wearable device according to claim 4, characterized in that, The calibration operation includes the following steps: 1) User B needs to determine whether they are awake or fatigued according to the Karolinska Sleepiness Scale, where an awake score is less than or equal to 5 points and a fatigue score is greater than 5 points; data should be collected for 5 minutes each in the awake and fatigued states, and recorded as follows: and In the formula Represents the number of brainwave channels. Represents the number of data points; 2) and Slicing along the time dimension, with each slice being 1 second long and non-overlapping, yields... and In the formula Represents the number of samples; 3) Order It is a vector of all zeros. A vector consisting entirely of 1s, and and Combined into training set ; 4) Using the training set A lightweight neural network was trained using the Adam optimizer and backpropagation algorithm, with a learning rate set to 0.001, yielding the desired user results. Personalized matching model .

6. The fatigue control system based on a brain-computer interface intelligent wearable device according to claim 5, characterized in that, The real-time control of fatigue state includes the following steps: 1) Initialize fatigue level to 0; 2) Every 5 seconds, the EEG signal conditioning and conversion module (1) collects the 1-second EEG signal of user B and sends it to the fatigue level decoding module (2); 3) The fatigue level decoding module (2) uses the proposed L-SFC-FNet model to classify the 1-second EEG signal as "awake state" or "fatigue state"; 4) If the classification result is "fatigue state", the fatigue level increases by 1; otherwise, the fatigue level remains unchanged. 5) If the fatigue level is ≥100, the fatigue feedback control module (3) generates microcurrent stimulation for 1 minute, and at the same time, the acquisition process of the EEG signal conditioning and conversion module (1) is paused, the fatigue level is set to 0, and the system returns to step 2) until user B stops using the system; if the fatigue level is <100, the system returns directly to step 2) until user B stops using the system.

Images