A wearable portable device for mental fatigue detection and intervention

Abstract

The application provides a wearable portable hairband for mental fatigue detection and intervention, a flexible conductive polymer dry electrode is embedded in the inner side of the hairband, the dry electrode is attached to the scalp through mechanical pressure, and thus inductive collection is realized without conductive paste; the output end of the dry electrode collection signal is directly connected to a collection front end located at the side edge of the hairband through FPC, the front end is integrated with an amplifier and an analog-to-digital converter, weak electrical signals on the cerebral cortex are amplified and converted into a data format; a SoC chip is selected as a processor to build a full-integrated edge computing platform inside the flexible fabric hairband, and the calculation task of mental fatigue detection is sunk to the hairband end; an ultrathin bone conduction unit is used as an acoustic output component, which is connected through a flexible circuit and embedded in the inner side of the hairband corresponding to the anatomical site of the user's temple, when the mental fatigue detection result triggers the fatigue warning threshold, a binaural rhythm intervention module is activated to perform binaural rhythm intervention. The application can perform mental fatigue detection and intervention.

Description

Technical Field

[0001] This invention belongs to the field of bioelectric signal processing and embedded deep learning technology, and specifically refers to a wearable and portable headband for detecting and intervening in mental fatigue. Background Technology

[0002] Mental fatigue (also known as brain fatigue, central fatigue, or cognitive fatigue) generally refers to a physiological decline in brain function after prolonged, high-intensity cognitive activity, especially under conditions of sustained attention or complex cognitive tasks. Unlike peripheral muscle fatigue, mental fatigue is primarily characterized by a weakening of the cerebral cortex's ability to integrate information. Its core impact lies in significantly inhibiting an individual's ability to maintain attention, sensory thresholds, and executive control functions. When mental fatigue accumulates continuously, the brain's information processing efficiency and reaction speed decrease significantly compared to a normal state, directly impairing the stability of behavioral execution and the accuracy of actions. This negative impact not only affects daily work and life, but also poses extremely high risks in the following scenarios: In driving large vehicles (such as cars, commercial airliners, and spacecraft) or operating high-precision instruments, even minor judgment errors or reaction delays can trigger catastrophic accidents; in extreme challenges such as alpine skiing and free climbing, as well as highly competitive sports like boxing and racing, mental fatigue weakens an individual's sustained attention and proprioception, increasing the error rate and the risk of disability or death; in sports involving complex dynamic balance, such as skateboarding, gymnastics, and trampolines, the depletion of cognitive resources leads to loss of coordination, exceeding physiological safety limits. Therefore, achieving real-time, quantitative detection of mental fatigue is crucial for building early warning mechanisms for high-risk operations, optimizing athlete performance, and ensuring the safety of complex human-computer interaction systems. A wearable, non-invasive, and high-temporal-resolution real-time technology solution is urgently needed to compensate for the lag in traditional subjective assessments and achieve closed-loop risk management from the physiological to the behavioral levels. Summary of the Invention

[0003] In order to solve the technical problems existing in the prior art, the present invention provides a wearable portable headband for mental fatigue detection and intervention. The headband is a flexible fabric headband that integrates a flexible conductive polymer dry electrode (101), an FPC flexible circuit (102), a SoC chip (103) and a bone conduction unit (104). A flexible conductive polymer dry electrode (101) is embedded in the inside of the headband. It is a sensor that collects EEG signals directly by mechanical pressure by attaching the conductive polymer material to the scalp without the need for conductive paste. The sensor achieves non-intrusive collection without the need for conductive paste by attaching the conductive polymer material to the scalp with mechanical pressure. The output of the dry electrode acquisition signal is directly connected to the acquisition front end located on the side of the headband via a miniaturized flexible circuit board FPC (102), a flexible thin film circuit carrier. The acquisition front end integrates a high input impedance instrumentation amplifier In-Amp and an analog-to-digital converter ADC to amplify the weak electrical signals on the cerebral cortex and convert them into a data format between 0 and 1. The processor uses a system-on-a-chip (SoC) chip (103) with operator compression capability to build a fully integrated edge computing platform inside the flexible fabric headband, thus pushing the computing task of mental fatigue detection down to the headband end; An ultra-thin bone conduction unit (104) is used as the acoustic output component. It is connected to and embedded in the inner side of the headband at the anatomical site corresponding to the user's temporal region through a flexible circuit. The processor drives the bone conduction unit through the I2S digital audio interface. When the mental fatigue detection result triggers the fatigue warning threshold, the binaural beat intervention module is activated. The corresponding frequency audio parameters are retrieved from the preset audio library and converted into two phase-synchronized pulse code modulation (PCM) signals in real time and output to the bone conduction unit. A closed-loop triggering mechanism is constructed based on the real-time fatigue discrimination result to perform binaural beat intervention and achieve non-invasive and dynamic adjustment of mental fatigue state.

[0004] Optionally, the flexible conductive polymer dry electrode, based on the results of previous multi-channel EEG research, selects three key brain regions that are highly related to mental fatigue: the frontal region (Fp), the temporal region (T), and the occipital region (O). It also precisely extracts four core electrode sites distributed in the frontal region, behind the ears, and behind the occipital region: Fpz, T5, T6, and Oz, and collects raw EEG signals.

[0005] Optionally, the processor executes a real-time artifact removal strategy, specifically including: The processor processes the acquired EEG signals based on a sliding time window and performs bandpass filtering on the EEG signals within a preset frequency range to suppress low-frequency drift and high-frequency noise interference. Based on this, for EEG signals that still contain residual interference after bandpass filtering, the processor introduces wavelet transform to perform multi-scale decomposition of the EEG signals to achieve feature extraction in time and frequency, thereby increasing the information available for analysis. The multi-scale decomposition includes: performing a wavelet transform on the original signal to obtain low-frequency and high-frequency coefficients, performing a wavelet transform on the low-frequency and high-frequency coefficients obtained in the first iteration in the second iteration, and iterating in sequence. After completing the multi-scale decomposition, the processor combines blind source separation technology and uses the Independent Component Analysis (ICA) algorithm to perform independent analysis on the decomposed signal components, separating the source signals from the mixed signals and removing noise and mixed superimposed signals. The ICA algorithm is based on the following assumption: the EEG signals collected from the scalp are mixed signals formed by multiple independent signal sources propagating in space and linearly superimposing, which include both the neural electrical activity of the cerebral cortex and interference signals generated by eye movement, blinking, facial and temporal muscle contraction. By maximizing the statistical independence between different signal components, the mixed signal is decomposed into multiple independent components, each of which corresponds to a potential signal source, and its characteristics in the time domain waveform and spectral features are relatively stable. Based on the decomposition results, combined with the frequency characteristics of independent components and their distribution on the scalp electrodes, independent components related to electrooculography (EOG) and weak electromyography (EMG) are eliminated, while independent components with typical EEG rhythm characteristics are retained. This enables effective differentiation and separation of signals from different sources, removes EOG artifacts and weak EMG artifacts that overlap with EEG components in the frequency domain, and provides EEG input data with higher signal-to-noise ratio and more complete morphological structure for subsequent mental fatigue detection.

[0006] Optionally, the mental fatigue detection uses a pre-trained CTC-Net model combining CNN and Transformer. The CTC-Net model includes: a CNN feature extraction module, a global pooling and embedding module, a Transformer encoding module, and a classification output module. Furthermore, the weights of the CTC-Net model are compressed to less than 100 kB using 8-bit integer quantization technology to achieve low-power detection at the embedded headband end.

[0007] Optionally, the processing procedure of the CNN feature extraction module includes: The input EEG signal is: , where L is the sampling point length; First, a single temporal convolution layer is used to capture local short-term dynamic patterns in the EEG waveform. The calculation formula is as follows: (1) in, It is a temporal convolutional kernel with a large receptive field, used to cover key neural oscillation cycles, with 32 output channels; It is the bias vector; Represents a one-dimensional convolution operation; ReLU is the activation function. Batch normalization (BN) is introduced to accelerate model convergence and suppress covariate bias. The calculation formula is as follows: (2) in, These are the mean and variance of the current batch, respectively. Learnable scale and offset parameters; Then, a feature convolution layer is used to achieve deep fusion of information between channels and enhance the non-linear expressive power of features. The calculation formula is as follows: (3) In equation (3), It uses a 1×16 convolutional kernel with 64 output channels, and further refines high-order semantic features by sliding point by point. For bias; this layer adopts the same padding strategy, that is, in order to avoid information loss caused by the convolution process, the edges of the input data are automatically padded with a certain number of "0"s to ensure that the output features after convolution are aligned with the original input EEG signal L in the temporal dimension, thus avoiding the loss of temporal information.

[0008] Optionally, the processing steps of the global pooling and embedding module include: First, global average pooling is introduced to suppress overfitting and ensure that the extracted features are translation-invariant. This operation compresses the spatial dimension, condensing the global temporal information into a fixed-length feature vector without introducing additional parameters. The calculation formula is as follows: (4) in, ; Then, Z0 is used to increase the signal dimension to 128 dimensions through the linear mapping layer Embedding Layer, so as to map the local features extracted by convolution to a high-dimensional manifold space suitable for Transformer processing; Simultaneously, a random deactivation layer with a Dropout rate of 0.5 is added to enhance the model's generalization ability and training stability. The calculation formula is as follows: (5) in, ; ; Dropout is defined as: (6) in This is for element-wise multiplication.

[0009] Optionally, the processing procedure of the Transformer encoding module includes: Using the 128-dimensional feature ztrans as the starting point of the Transformer sequence input, and introducing positional encoding to compensate for the lack of temporal order perception in the self-attention mechanism, the calculation formula is as follows: (7) The position coding uses a fixed sine / cosine coding function to capture position correlations at different frequencies. The calculation formula is as follows: (8) Where p is the sequence position index, and d=128; The Transformer encoding module consists of a two-layer architecture, each of which comprises three parts: Multi-head Self-attention (MHSA), Feedforward Neural Network (FFN), Residual Connection, and Layer Normalization (LN). Each layer of the architecture captures long-range dependencies between features through a multi-head self-attention mechanism and performs spatial transformation through a feedforward neural network. Residual connections and layer normalization are introduced after each sublayer to ensure stable gradient propagation in deep networks. The MHSA calculation is as follows: (9) Where Q, K, and V represent the query, key, and value matrices, respectively. The number of heads, h, is set to 4, and the scaling factor, dk, is set to 32 to prevent the dot product result from becoming too large and entering the softmax gradient saturation region. Its definition is: (10) Linear projection after multi-head stitching ensures the aggregation of multi-dimensional information: (11) in, ; The feedforward network (FFN) achieves nonlinear recombination of features in a high-dimensional space through two layers of linear mapping and the ReLU activation function. (12) in, , ; , After completing the nonlinear combination of the 512-dimensional hidden layers, this layer is back-projected to 128 dimensions. The residual join and LN process are represented as follows: (13) (14) The final output features of the Transformer encoding module after residual and normalization processing are as follows: (15).

[0010] Optionally, the classification output module uses a simplified fully connected layer as the classification head, and maps it to specific fatigue / awake state probabilities through a Softmax layer: (16) in, , Output probability .

[0011] Optionally, the binaural beat intervention module includes a preset audio library that stores monotone signals of specific frequencies, specifically: 240Hz for the left ear and 250Hz for the right ear; The fatigue / awake state probability P output by the classification output module is monitored in real time. When the value of P exceeds the preset fatigue threshold, the subject is determined to be in a "brain fatigue warning state". At this time, the binaural beat intervention module is immediately activated. The brain's current fatigue state is intervened by the 10Hz difference frequency acoustic signal between the left and right ears of the preset audio. The intervention is to regulate the brain rhythm by using the neural entrainment effect. When the two ears receive pure tone signals with a frequency difference of 10Hz, the upper olive nucleus of the brainstem integrates the phase information and generates a perceptual beat equivalent to 10Hz. The beat frequency is in the core range of the brain's alpha wave band, which guides the cortical neurons to generate phase lock, thereby relieving mental fatigue and accelerating the recovery of nerve function. This acoustic induction method does not require the subject to interrupt the current task and realizes real-time regulation of brain function.

[0012] Optionally, the binaural beat intervention module integrates user autonomy control, allowing subjects to manually activate the intervention mode based on their perceived attentional drift, slow reaction, or mental discomfort, without triggering the fatigue threshold. The binaural beat intervention module employs a dynamic volume mapping mechanism. The volume of the intervention signal is finely adjusted according to the fluctuation of the probability value P. Monitoring shows that the fatigue trend continues to increase, which will linearly increase the audio amplitude intensity.

[0013] The beneficial effects of the technical solution provided by this invention include at least the following: (1) Addressing the issues of cumbersome wearing and the need for conductive gel in existing laboratory-grade systems (such as Neuroscan), and the limited accuracy and single channel of consumer-grade devices, this invention precisely extracts core sites such as the forehead (Fp), temporal region (T), and occipital region (O) through preliminary research, reducing the number of electrodes to four, and integrating flexible conductive polymer dry electrodes into a fabric headband. This design achieves comprehensive coverage of key brain regions characterized by mental fatigue without the need for conductive gel, significantly reducing the difficulty of wearing the device and social pressure, making it possible to move high-precision mental fatigue monitoring from the laboratory to everyday work scenarios.

[0014] (2) In view of the problems of cumbersome processing flow, poor timeliness and weak algorithm generalization ability of existing equipment, this invention introduces the CTC-Net model that integrates CNN (local feature extraction) and Transformer (global temporal modeling). By compressing weights, the model is implemented on the device. This edge computing architecture does not rely on mobile phones or cloud transmission, effectively avoids the risk of data transmission delay and privacy leakage, and ensures early warning in high-time-demand scenarios such as driving and high-altitude operations.

[0015] (3) In view of the shortcomings of existing technologies such as emphasizing monitoring, neglecting intervention, and the single and blind intervention methods, this invention systematically integrates the binaural beat technology based on the neural entrainment theory into the detection device for binaural beat intervention for the first time. Based on the mental fatigue detection results, a 10Hz difference frequency signal is triggered, and the α wave phase lock is directionally induced through the bone conduction unit. This non-invasive intervention does not require the subject to interrupt the current task and can be actively intervened in the early stage of fatigue.

[0016] (4) In view of the problems of existing headband structures being heavy, oppressive and having high power consumption, the present invention uses an FPC flexible circuit board to connect an ultra-thin bone conduction unit, and combines it with a fabric headband to achieve personalized head circumference adaptation, which greatly improves the comfort of wearing for a long time. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of a wearable and portable headband for detecting and intervening in mental fatigue, provided by an embodiment of the present invention; Figure 2 This is a block diagram of the CTC-Net model structure provided in an embodiment of the present invention; Figure 3 This is a block diagram of the CNN feature extraction module provided in an embodiment of the present invention; Figure 4 This is a structural block diagram of the Transformer encoding module provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of binaural beat intervention provided in an embodiment of the present invention. Detailed Implementation

[0019] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0020] like Figure 1 As shown, this embodiment of the invention provides a wearable portable headband for detecting and intervening in mental fatigue. The headband is a flexible fabric headband that integrates a flexible conductive polymer dry electrode (101), an FPC flexible circuit (102), a SoC chip (103), and a bone conduction unit (104). A flexible conductive polymer dry electrode (101) is embedded in the inside of the headband. It is a sensor that collects EEG signals directly by mechanical pressure by attaching the conductive polymer material to the scalp without the need for conductive paste. The sensor achieves non-intrusive collection without the need for conductive paste by attaching the conductive polymer material to the scalp with mechanical pressure. The output of the dry electrode acquisition signal is directly connected to the acquisition front end located on the side of the headband via a miniaturized flexible circuit board FPC (102), a flexible thin film circuit carrier. The acquisition front end integrates a high input impedance instrumentation amplifier In-Amp and an analog-to-digital converter ADC to amplify the weak electrical signals on the cerebral cortex and convert them into a data format between 0 and 1. The processor uses a system-on-a-chip (SoC) chip (103) with operator compression capability to build a fully integrated edge computing platform inside the flexible fabric headband, thus pushing the computing task of mental fatigue detection down to the headband end; An ultra-thin bone conduction unit (104) is used as the acoustic output component. It is connected to and embedded in the inner side of the headband at the anatomical site corresponding to the user's temporal region through a flexible circuit. The processor drives the bone conduction unit through the I2S digital audio interface. When the mental fatigue detection result triggers the fatigue warning threshold, the binaural beat intervention module is activated. The corresponding frequency audio parameters are retrieved from the preset audio library and converted into two phase-synchronized pulse code modulation (PCM) signals in real time and output to the bone conduction unit. A closed-loop triggering mechanism is constructed based on the real-time fatigue discrimination result to perform binaural beat intervention and achieve non-invasive and dynamic adjustment of mental fatigue state.

[0021] Optionally, the flexible conductive polymer dry electrode, based on the results of previous multi-channel EEG research, selects three key brain regions that are highly related to mental fatigue: the frontal region (Fp), the temporal region (T), and the occipital region (O). It also precisely extracts four core electrode sites distributed in the frontal region, behind the ears, and behind the occipital region: Fpz, T5, T6, and Oz, and collects raw EEG signals.

[0022] Currently, wearable EEG testing devices on the market primarily use a single electrode in the frontal region, mainly for meditation training, and are not suitable for accurate assessment of mental fatigue. Clinical-level mental fatigue assessments mostly rely on laboratory-grade multi-channel EEG systems, which are technically demanding and inconvenient, making them unsuitable for widespread use.

[0023] This invention is based on previous laboratory research on multi-channel EEG for identifying mental fatigue. Through scientific comparison and analysis, it was found that in addition to the frontal region (Fp), the temporal region (T) and occipital region (Oz) have significant value in identifying brain fatigue and emotion in visual-related tasks (the frontal region Fp and temporal region T are currently the main brain regions used by portable wearable devices to detect emotion and fatigue; however, this invention found that the occipital region Oz, which is closely related to vision, has very important value in identifying brain fatigue in visual-related tasks). Therefore, the number of measuring electrodes can be greatly reduced to four, significantly simplifying the device components. This solves the problems of complexity and high technical requirements in existing EEG monitoring technologies, and can be used in wearable EEG devices, suitable for the scientific assessment of mental fatigue in a wide range of populations.

[0024] Traditional electrode caps, due to their need to cover the entire brain with electrodes, lack aesthetic appeal and are difficult to wear independently. Currently, most meditation devices on the market are headbands, but because the electrodes are primarily located in the forehead area, their accuracy in assessing mental fatigue is insufficient, and their aesthetics are also limited. This invention's EEG electrode headband product, based on previous fatigue monitoring research, accurately extracts four core electrode sites distributed in the forehead, temporal region, and occipital region. These electrodes are embedded in a high-elasticity fabric headband, solving the problems of insufficient aesthetics, comfort, and style of current wearable EEG monitoring headbands on the market. It reduces the "laboratory look" of traditional headbands, lowers the social pressure on users in daily use, and, most importantly, adds occipital region electrodes that reflect visual task-related mental fatigue. The headband material is a fabric with a certain degree of elasticity, and different colors can be used to match different needs. At the same time, its flexibility provides users with both aesthetic appeal and significant comfort.

[0025] Optionally, the processor executes a real-time artifact removal strategy, specifically including: The processor processes the acquired EEG signals based on a sliding time window and performs bandpass filtering on the EEG signals within a preset frequency range to suppress low-frequency drift and high-frequency noise interference. Based on this, for EEG signals that still contain residual interference after bandpass filtering, the processor introduces wavelet transform to perform multi-scale decomposition of the EEG signals to achieve feature extraction in time and frequency, thereby increasing the information available for analysis. The multi-scale decomposition includes: performing a wavelet transform on the original signal to obtain low-frequency and high-frequency coefficients, performing a wavelet transform on the low-frequency and high-frequency coefficients obtained in the first iteration in the second iteration, and iterating in sequence. After completing the multi-scale decomposition, the processor combines blind source separation technology and uses the Independent Component Analysis (ICA) algorithm to perform independent analysis on the decomposed signal components, separating the source signals from the mixed signals and removing noise and mixed superimposed signals. The ICA algorithm is based on the following assumption: the EEG signals collected from the scalp are mixed signals formed by multiple independent signal sources propagating in space and linearly superimposing, which include both the neural electrical activity of the cerebral cortex and interference signals generated by eye movement, blinking, facial and temporal muscle contraction. By maximizing the statistical independence between different signal components, the mixed signal is decomposed into multiple independent components, each of which corresponds to a potential signal source, and its characteristics in the time domain waveform and spectral features are relatively stable. Based on the decomposition results, combined with the frequency characteristics of independent components and their distribution on the scalp electrodes, independent components related to electrooculography (EOG) and weak electromyography (EMG) are eliminated, while independent components with typical EEG rhythm characteristics are retained. This enables effective differentiation and separation of signals from different sources, removes EOG artifacts and weak EMG artifacts that overlap with EEG components in the frequency domain, and provides EEG input data with higher signal-to-noise ratio and more complete morphological structure for subsequent mental fatigue detection.

[0026] Optionally, such as Figure 2 As shown, the mental fatigue detection uses a pre-trained CTC-Net model that combines CNN and Transformer. The CTC-Net model includes: a CNN feature extraction module, a global pooling and embedding module, a Transformer encoding module, and a classification output module. Furthermore, it employs 8-bit integer quantization technology (compressing the high-precision floating-point parameters of the AI ​​model into 8-bit integers to reduce storage pressure) to compress the weights of the CTC-Net model to within 100 kB, thereby achieving low-power detection at the embedded headband end.

[0027] Optionally, such as Figure 3 As shown, the processing procedure of the CNN feature extraction module includes: The input EEG signal is: , where L is the sampling point length; First, a single temporal convolution layer is used to capture local short-term dynamic patterns in the EEG waveform (such as transient discharges or rhythm changes in specific frequency bands). The calculation formula is as follows: (1) in, It is a temporal convolutional kernel with a large receptive field, used to cover key neural oscillation cycles, with 32 output channels; It is the bias vector; Represents a one-dimensional convolution operation; ReLU is the activation function. Batch normalization (BN) is introduced to accelerate model convergence and suppress covariate bias. The calculation formula is as follows: (2) in, These are the mean and variance of the current batch, respectively. Learnable scale and offset parameters; Then, a feature convolution layer is used to achieve deep fusion of information between channels and enhance the non-linear expressive power of features. The calculation formula is as follows: (3) In equation (3), It uses a 1×16 convolutional kernel with 64 output channels, and further refines high-order semantic features by sliding point by point. For bias; this layer adopts the same padding strategy, that is, in order to avoid information loss caused by the convolution process, the edges of the input data are automatically padded with a certain number of "0"s to ensure that the output features after convolution are aligned with the original input EEG signal L in the temporal dimension, thus avoiding the loss of temporal information.

[0028] Optionally, the processing steps of the global pooling and embedding module include: First, global average pooling is introduced to suppress overfitting and ensure that the extracted features are translation-invariant. This operation compresses the spatial dimension, condensing the global temporal information into a fixed-length feature vector without introducing additional parameters. The calculation formula is as follows: (4) in, ; Then, Z0 is used to increase the signal dimension to 128 dimensions through the linear mapping layer Embedding Layer, so as to map the local features extracted by convolution to a high-dimensional manifold space suitable for Transformer processing; Simultaneously, a random deactivation layer with a Dropout rate of 0.5 is added to enhance the model's generalization ability and training stability. The calculation formula is as follows: (5) in, ; ; Dropout is defined as: (6) in This is for element-wise multiplication.

[0029] Optionally, such as Figure 4 As shown, the processing procedure of the Transformer encoding module includes: Using the 128-dimensional feature ztrans as the starting point of the Transformer sequence input, and introducing positional encoding to compensate for the lack of temporal order perception in the self-attention mechanism, the calculation formula is as follows: (7) The position coding uses a fixed sine / cosine coding function to capture position correlations at different frequencies. The calculation formula is as follows: (8) Where p is the sequence position index, and d=128; The Transformer encoding module consists of a two-layer architecture, each of which comprises three parts: Multi-head Self-attention (MHSA), Feedforward Neural Network (FFN), Residual Connection, and Layer Normalization (LN). Each layer of the architecture captures long-range dependencies between features through a multi-head self-attention mechanism and performs spatial transformation through a feedforward neural network. Residual connections and layer normalization are introduced after each sublayer to ensure stable gradient propagation in deep networks. The MHSA calculation is as follows: (9) Where Q, K, and V represent the query, key, and value matrices, respectively. The number of heads, h, is set to 4, and the scaling factor, dk, is set to 32 to prevent the dot product result from becoming too large and entering the softmax gradient saturation region. Its definition is: (10) Linear projection after multi-head stitching ensures the aggregation of multi-dimensional information: (11) in, ; The feedforward network FFN achieves nonlinear recombination of features in a high-dimensional space (512-dimensional hidden layers) through two layers of linear mapping and the ReLU activation function. (12) in, , ; , After completing the nonlinear combination of the 512-dimensional hidden layers, this layer is back-projected to 128 dimensions. The residual join and LN process are represented as follows: (13) (14) The final output features of the Transformer encoding module after residual and normalization processing are as follows: (15).

[0030] Optionally, the classification output module uses a simplified fully connected layer as the classification head, and maps it to specific fatigue / awake state probabilities through a Softmax layer: (16) in, , Output probability .

[0031] This invention uses a flexible conductive polymer dry electrode to collect raw EEG signals, which are then processed by a deep learning model for hierarchical feature extraction, fusion, and classification output. Through the synergistic effect of the above-mentioned functional modules, stable identification and real-time output of mental fatigue state are achieved, thus providing reliable technical support for mental fatigue monitoring and application in low-power and portable scenarios.

[0032] Optionally, the binaural beat intervention module includes a preset audio library storing monotone signals of specific frequencies, specifically: 240Hz for the left ear and 250Hz for the right ear, such as... Figure 5 As shown; The fatigue / awake state probability P output by the classification output module is monitored in real time. When the value of P exceeds the preset fatigue threshold (e.g., 0.5), the subject is determined to be in a "brain fatigue warning state". At this time, the binaural beat intervention module is immediately activated. It intervenes in the current fatigue state of the brain by using the 10Hz difference frequency acoustic signal between the left and right ears of the preset audio. The intervention is to regulate the brain rhythm by using the neural entrainment effect. When the two ears receive pure tone signals with a frequency difference of 10Hz, the upper olive nucleus of the brainstem integrates the phase information and generates a perceptual beat equivalent to 10Hz. The beat frequency is in the core range of the brain's alpha wave band (8-13Hz), which guides the cortical neurons to generate phase lock, thereby relieving mental fatigue and accelerating the recovery of neural function. This acoustic induction method does not require the subject to interrupt the current task and realizes real-time regulation of brain function.

[0033] Optionally, the binaural beat intervention module integrates user autonomy control, allowing subjects to manually activate the intervention mode based on their perceived attentional drift, slow reaction, or mental discomfort, without triggering the fatigue threshold. The binaural beat intervention module employs a dynamic volume mapping mechanism. The volume of the intervention signal is finely adjusted according to the fluctuation of the probability value P. Monitoring shows that the fatigue trend continues to increase, which will linearly increase the audio amplitude intensity.

[0034] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A wearable and portable headband for detecting and intervening in mental fatigue, characterized in that, The headband is a flexible fabric headband that integrates a flexible conductive polymer dry electrode (101), an FPC flexible circuit (102), a SoC chip (103), and a bone conduction unit (104). A flexible conductive polymer dry electrode (101) is embedded in the inside of the headband. It is a sensor that collects EEG signals directly by mechanical pressure by attaching the conductive polymer material to the scalp without the need for conductive paste. The sensor achieves non-intrusive collection without the need for conductive paste by attaching the conductive polymer material to the scalp with mechanical pressure. The output of the dry electrode acquisition signal is directly connected to the acquisition front end located on the side of the headband via a miniaturized flexible circuit board FPC (102), a flexible thin film circuit carrier. The acquisition front end integrates a high input impedance instrumentation amplifier In-Amp and an analog-to-digital converter ADC to amplify the weak electrical signals on the cerebral cortex and convert them into a data format between 0 and 1. The processor uses a system-on-a-chip (SoC) chip (103) with operator compression capability to build a fully integrated edge computing platform inside the flexible fabric headband, thus pushing the computing task of mental fatigue detection down to the headband end; An ultra-thin bone conduction unit (104) is used as the acoustic output component. It is connected to and embedded in the inner side of the headband at the anatomical site corresponding to the user's temporal region through a flexible circuit. The processor drives the bone conduction unit through the I2S digital audio interface. When the mental fatigue detection result triggers the fatigue warning threshold, the binaural beat intervention module is activated. The corresponding frequency audio parameters are retrieved from the preset audio library and converted into two phase-synchronized pulse code modulation (PCM) signals in real time and output to the bone conduction unit. A closed-loop triggering mechanism is constructed based on the real-time fatigue discrimination result to perform binaural beat intervention and achieve non-invasive and dynamic adjustment of mental fatigue state.

2. The headband according to claim 1, characterized in that, Based on previous multichannel EEG research results, the flexible conductive polymer dry electrode selects three key brain regions that are highly related to mental fatigue: the frontal region (Fp), the temporal region (T), and the occipital region (O). It also precisely extracts four core electrode sites distributed in the frontal region, behind the ears, and behind the occipital region: Fpz, T5, T6, and Oz, and collects raw EEG signals.

3. The headband according to claim 1, characterized in that, The processor executes a real-time artifact removal strategy, specifically including: The processor processes the acquired EEG signals based on a sliding time window and performs bandpass filtering on the EEG signals within a preset frequency range to suppress low-frequency drift and high-frequency noise interference. Based on this, for EEG signals that still contain residual interference after bandpass filtering, the processor introduces wavelet transform to perform multi-scale decomposition of the EEG signals to achieve feature extraction in time and frequency, thereby increasing the information available for analysis. The multi-scale decomposition includes: performing a wavelet transform on the original signal to obtain low-frequency and high-frequency coefficients, performing a wavelet transform on the low-frequency and high-frequency coefficients obtained in the first iteration in the second iteration, and iterating in sequence. After completing the multi-scale decomposition, the processor combines blind source separation technology and uses the Independent Component Analysis (ICA) algorithm to perform independent analysis on the decomposed signal components, separating the source signals from the mixed signals and removing noise and mixed superimposed signals. The ICA algorithm is based on the following assumption: the EEG signals collected from the scalp are mixed signals formed by multiple independent signal sources propagating in space and linearly superimposing, which include both the neural electrical activity of the cerebral cortex and interference signals generated by eye movement, blinking, facial and temporal muscle contraction. By maximizing the statistical independence between different signal components, the mixed signal is decomposed into multiple independent components, each of which corresponds to a potential signal source, and its characteristics in the time domain waveform and spectral features are relatively stable. Based on the decomposition results, combined with the frequency characteristics of independent components and their distribution on the scalp electrodes, independent components related to electrooculography (EOG) and weak electromyography (EMG) are eliminated, while independent components with typical EEG rhythm characteristics are retained. This enables effective differentiation and separation of signals from different sources, removes EOG artifacts and weak EMG artifacts that overlap with EEG components in the frequency domain, and provides EEG input data with higher signal-to-noise ratio and more complete morphological structure for subsequent mental fatigue detection.

4. The headband according to claim 1, characterized in that, The mental fatigue detection uses a pre-trained CTC-Net model that combines CNN and Transformer. The CTC-Net model includes a CNN feature extraction module, a global pooling and embedding module, a Transformer encoding module, and a classification output module. Furthermore, 8-bit integer quantization technology is used to compress the weights of the CTC-Net model to less than 100 kB to achieve low-power detection at the embedded headband end.

5. The headband according to claim 4, characterized in that, The processing steps of the CNN feature extraction module include: The input EEG signal is: , where L is the sampling point length; First, a single temporal convolution layer is used to capture local short-term dynamic patterns in the EEG waveform. The calculation formula is as follows: （1） in, It is a temporal convolutional kernel with a large receptive field, used to cover key neural oscillation cycles, with 32 output channels; It is the bias vector; Represents a one-dimensional convolution operation; ReLU is the activation function. Batch normalization (BN) is introduced to accelerate model convergence and suppress covariate bias. The calculation formula is as follows: （2） in, These are the mean and variance of the current batch, respectively. Learnable scale and offset parameters; Then, a feature convolution layer is used to achieve deep fusion of information between channels and enhance the non-linear expressive power of features. The calculation formula is as follows: （3） In equation (3), It uses a 1×16 convolutional kernel with 64 output channels, and further refines high-order semantic features by sliding point by point. For bias; this layer adopts the same padding strategy, that is, in order to avoid information loss caused by the convolution process, the edges of the input data are automatically padded with a certain number of "0"s to ensure that the output features after convolution are aligned with the original input EEG signal L in the temporal dimension, thus avoiding the loss of temporal information.

6. The headband according to claim 5, characterized in that, The processing steps of the global pooling and embedding module include: First, global average pooling is introduced to suppress overfitting and ensure that the extracted features are translation-invariant. This operation compresses the spatial dimension, condensing the global temporal information into a fixed-length feature vector without introducing additional parameters. The calculation formula is as follows: （4） in, ; Then, Z0 is used to increase the signal dimension to 128 dimensions through the linear mapping layer Embedding Layer, so as to map the local features extracted by convolution to a high-dimensional manifold space suitable for Transformer processing; Simultaneously, a random deactivation layer with a Dropout rate of 0.5 is added to enhance the model's generalization ability and training stability. The calculation formula is as follows: （5） in, ; ; Dropout is defined as: （6） in This is for element-wise multiplication.

7. The headband according to claim 6, characterized in that, The processing steps of the Transformer encoding module include: Using the 128-dimensional feature ztrans as the starting point of the Transformer sequence input, and introducing positional encoding to compensate for the lack of temporal order perception in the self-attention mechanism, the calculation formula is as follows: （7） The position coding uses a fixed sine / cosine coding function to capture position correlations at different frequencies. The calculation formula is as follows: （8） Where p is the sequence position index, and d=128; The Transformer encoding module consists of a two-layer architecture, each of which comprises three parts: Multi-head Self-attention (MHSA), Feedforward Neural Network (FFN), Residual Connection, and Layer Normalization (LN). Each layer of the architecture captures long-range dependencies between features through a multi-head self-attention mechanism and performs spatial transformation through a feedforward neural network. Residual connections and layer normalization are introduced after each sublayer to ensure stable gradient propagation in deep networks. The MHSA calculation is as follows: （9） Where Q, K, and V represent the query, key, and value matrices, respectively. The number of heads, h, is set to 4, and the scaling factor, dk, is set to 32 to prevent the dot product result from becoming too large and entering the softmax gradient saturation region. Its definition is: （10） Linear projection after multi-head stitching ensures the aggregation of multi-dimensional information: （11） in, ; The feedforward network (FFN) achieves nonlinear recombination of features in a high-dimensional space through two layers of linear mapping and the ReLU activation function. （12） in, , ; , After completing the nonlinear combination of the 512-dimensional hidden layers, this layer is back-projected to 128 dimensions. The residual join and LN process are represented as follows: （13） （14） The final output features of the Transformer encoding module after residual and normalization processing are as follows: （15）。 8. The headband according to claim 7, characterized in that, The classification output module uses a simplified fully connected layer as the classification head, and maps it to specific fatigue / awake state probabilities through a Softmax layer: （16） in, , Output probability .

9. The headband according to claim 8, characterized in that, The binaural beat intervention module includes a preset audio library that stores monotone signals of specific frequencies, specifically: 240Hz for the left ear and 250Hz for the right ear. The fatigue / awake state probability P output by the classification output module is monitored in real time. When the value of P exceeds the preset fatigue threshold, the subject is determined to be in a "brain fatigue warning state". At this time, the binaural beat intervention module is immediately activated. The brain's current fatigue state is intervened by the 10Hz difference frequency acoustic signal between the left and right ears of the preset audio. The intervention is to regulate the brain rhythm by using the neural entrainment effect. When the two ears receive pure tone signals with a frequency difference of 10Hz, the upper olive nucleus of the brainstem integrates the phase information and generates a perceptual beat equivalent to 10Hz. The beat frequency is in the core range of the brain's alpha wave band, which guides the cortical neurons to generate phase lock, thereby relieving mental fatigue and accelerating the recovery of nerve function. This acoustic induction method does not require the subject to interrupt the current task and realizes real-time regulation of brain function.

10. The headband according to claim 9, characterized in that, The binaural beat intervention module integrates user autonomy and allows subjects to manually activate the intervention mode based on their own perceived attentional drift, slow reaction, or mental discomfort, without triggering the fatigue threshold. The binaural beat intervention module employs a dynamic volume mapping mechanism. The volume of the intervention signal is finely adjusted according to the fluctuation of the probability value P. Monitoring shows that the fatigue trend continues to increase, which will linearly increase the audio amplitude intensity.

Images