Brain wave monitoring-based power inspection personnel safe working time limit prediction system

Abstract

The application relates to an electric power inspection personnel safety working time limit prediction system based on brain wave monitoring. The brain wave of electric power inspection personnel can be monitored, the safety working time limit of the electric power inspection personnel can be predicted, a long sequence time prediction model (LSTPModel) is proposed, the collected EEG signals of the electric power inspection personnel are analyzed, the safety working time limit of the electric power inspection personnel is predicted before work, the safety working time limit is predicted and calibrated according to the real-time transmission of the EEG signals during work, and a reference is provided for on-site work command and dispatch.

Description

Technical Field

[0001] This invention relates to a system for predicting the safe working time limits of power line inspectors based on electroencephalogram (EEG) monitoring. Specifically, it relates to a system for predicting the safe working time limits of power line inspectors based on deep learning technology and EEG monitoring technology. Background Technology

[0002] Power line inspection work often involves a large amount of work at height, requiring a safe working time limit prediction system to predict the safe working time limits of inspection personnel in advance. This provides a reference for on-site work scheduling and ensures the safe conduct of inspection work. Power line inspection personnel need to maintain a high level of concentration for extended periods while working at heights, which often leads to fatigue, impairing judgment and reaction capabilities, increasing the probability of dangerous accidents, and affecting the stable operation of the power system. Therefore, it is necessary to adopt a reliable method for predicting safe working time limits. This method can assess the physiological state of inspection personnel based on physiological signal monitoring, predict their safe working time limits in advance, provide a reference for on-site work scheduling, effectively protect the safety of inspection personnel, improve inspection efficiency, and ensure the stable operation of the power system.

[0003] Monitoring fatigue levels is crucial for the safety of power line inspection personnel working at heights. Using electroencephalography (EEG) signals as a physiological assessment tool can provide a more accurate and objective evaluation, allowing for the prediction of safe working hours for inspection personnel. By collecting EEG signals before work and monitoring them in real time during operations, the condition of inspection personnel can be better assessed, helping to predict safe working hours, assisting in on-site work scheduling, and reducing the probability of accidents.

[0004] Using deep learning technology to process EEG signals offers several advantages for predicting the safe working time limits of power line inspection personnel. First, EEG signals can help capture the real-time working status of inspection personnel, such as fatigue or focus, thus more accurately predicting their working time limits. Second, deep learning technology can facilitate complex classification and feature extraction of EEG signals, leading to a better understanding of the work characteristics and status of inspection personnel. This mechanism can help predict factors such as the inspection personnel's fatigue level, mental state, work ability, and reaction time, providing a reference for on-site work scheduling and ensuring safe operation. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a system for predicting the safe working time limit of power inspection personnel based on electroencephalogram (EEG) monitoring.

[0006] The technical solution adopted in this invention is: a power inspection personnel safety work time limit prediction system based on EEG monitoring, including an EEG monitoring power safety helmet (1) and a safety work time limit prediction module (2). The EEG monitoring power safety helmet (1) collects EEG signals from the power inspection personnel's brain in real time and transmits the data to the safety work time limit prediction module (2) through 5G data transmission. The safety work time limit prediction module (2) proposes a long sequence time prediction model (LSTPModel) to process the EEG signals of the power inspection personnel, predict the safety work time limit of the power inspection personnel before the operation, and perform safety work time limit prediction calibration based on the real-time transmitted EEG signals during the operation, providing a reference for on-site work command and dispatch.

[0007] The aforementioned electroencephalogram (EEG) monitoring electric safety helmet (1) includes: a power supply module (11), EEG acquisition electrodes (12), a bioelectric signal digital-to-analog converter (13), an MCU processor (14), and a 5G communication module (15). The power supply module (11) is powered by a battery and obtains different operating voltages required by different chips on each module through a voltage conversion chip to power each module. The EEG acquisition electrodes (12) are integrated in the shell of the electric safety helmet, with a total of two electrodes distributed at FP1 and FP2 in the 10-20 international standard lead electrode distribution. The electrodes are connected to the bioelectric signal digital-to-analog converter (13) through wires to collect EEG signals from different brain regions. The described bioelectric signal digital-to-analog converter (13) has a built-in programmable gain amplifier (PGA), internal reference, and on-chip oscillator to perform digital-to-analog conversion on EEG signals; the described MCU processor (14) adopts an Xtensa-LX6 core with a working frequency of 300MHz and communicates with the bioelectric signal digital-to-analog converter (12) via the SPI bus. It is responsible for controlling the bioelectric signal digital-to-analog converter (12) to collect EEG signals at a fixed sampling frequency f = 250Hz and send them to the MCU processor (14) via the SPI bus. After the transmission is completed, the MCU processor (14) transmits the EEG signal data to the safe working time limit prediction module (2) via the 5G communication module (15).

[0008] The safety work time limit prediction module (2) proposes a long sequence time prediction model (LSTPModel), which is deployed in the cloud to process the original EEG signal, extract the time dimension sequence characteristics of the EEG signal, and predict the expected safety work time limit of the operator.

[0009] The proposed Long Sequence Temporal Prediction (LSTP) Model comprises a Gram field transform module, an adaptive shape convolution module, and an Informer encoder module connected in series. The Gram field transform module first concatenates the EEG signals from two channels to obtain a long sequence. Then, it performs the Gram field transform on this long sequence, transforming the one-dimensional sequence information into two-dimensional image information. The specific steps of the Gram field transform algorithm are as follows:

[0010] 1) Convert the EEG signal data X = {x1, x2, ..., x...} n Normalized to the interval [-1, 1], the normalized data is denoted as X. ′ ={x ′ 1,x ′ 2,...,x ′ n}

[0011] 2) Convert the normalized data to polar coordinates, including the polar angle σ. i =cos -1 x i ′ and polar axis r i =t i / N,x i ′ t represents the normalized value of this data point. i Represents the position in the sequence.

[0012] 3) Define the inner product between two sequence points as... <x i ,x j >=cos(σ i +σ j The Gram matrix is ​​calculated using the inner product:

[0013]

[0014] This matrix with a custom inner product is the result of the Gram angle field transform algorithm. It is an n×n square matrix that ultimately transforms the n-dimensional frequency domain feature sequence into n×n×1-dimensional image data.

[0015] The adaptive shape convolution module uses adaptive shape convolution to perform convolution operations on image data. The adaptive shape convolution module includes, in sequence:

[0016] 1) A data input layer, the input data being the image data output by the Gram field transform module;

[0017] 2) An adaptive shape convolutional layer consists of 45 adaptive shape convolutional kernels. Each adaptive shape convolutional kernel is composed of N convolution points with their respective offsets β = {β1, β2, ..., β...}.N It consists of offsets of different lengths in different directions, using PReLU as the activation function. Where N, β, and λ are parameters obtained through training. With N and β obtained through training, the adaptive special-shaped convolutional kernel can adaptively adjust the size and shape of the convolutional kernel according to the characteristics of the input image data, and adaptively and efficiently extract the hidden EEG signal time-dependent sequence features in the EEG sequence Gram field image data to obtain the time-series feature map.

[0018] 3) A batch channel normalization layer normalizes the temporal feature map in parallel along the batch dimension and the channel dimension. Then, the batch dimension normalized output and the channel dimension normalized output are weighted and summed with the adaptive parameters α and 1-α, respectively, to obtain the normalized output of the batch channel normalization layer, which accelerates the model training process and reduces the degree of overfitting.

[0019] 4) A Dropout layer randomly selects a neuron in the previous layer with probability p1 so that it does not output anything, where p1 = 0.25;

[0020] 5) A Flatten layer transforms multidimensional data into one-dimensional data, yielding a time-series feature sequence;

[0021] The adaptive shape convolution module extracts the hidden EEG signal time-dependent sequence features in the EEG sequence Gram field image data through adaptive shape convolution, and finally obtains the high-level semantic feature sequence.

[0022] The Informer encoder module consists of three Informer encoders and a linear regression layer connected in series. The Informer encoder receives high-level semantic feature sequences, processes the temporal feature sequences using a probabilistic sparse self-attention mechanism, and uses self-attention distillation to filter the dominant attention weights, reducing the model size. It then performs further feature extraction on the temporal feature sequences to obtain the correlation between different channels in the temporal feature sequences. The output of the Informer encoder is input into a linear regression layer, and regression yields the prediction results of the safe working time limit for power inspection personnel.

[0023] The long sequence time prediction model (LSTPModel) is pre-trained based on a pre-built database of safe working time for power inspectors. Then, it is federated training based on ten minutes of EEG signal data collected by power inspectors wearing EEG monitoring power safety helmets (1) before high-altitude operations. The model is then personalized and fine-tuned to obtain preliminary prediction results of the safe working time limit for power inspectors, providing a preliminary reference for the command and dispatch of on-site inspection work.

[0024] During their work, power inspectors wear EEG monitoring safety helmets (1) to collect EEG signals in real time and transmit the data to the safe work time limit prediction module (2) via 5G data transmission. The safe work time limit prediction module (2) analyzes the EEG signals in real time and corrects the preliminary prediction results of the safe work time limit of the power inspectors in real time, providing reference and safety guarantee for the command and dispatch of on-site work. Attached Figure Description

[0025] Figure 1 This is a system block diagram of the power inspection personnel safety work time limit prediction system based on electroencephalogram monitoring of the present invention;

[0026] Figure 2 This is a schematic diagram of the internal structure of an electroencephalogram (EEG) monitoring power safety helmet;

[0027] Figure 3 This is a schematic diagram of the electrode distribution locations for 10-20 international standard leads;

[0028] Figure 4 This is a schematic diagram of the structure of the Long Sequence Time Prediction Model (LSTP).

[0029] Figure 5 This is a schematic diagram of the Gram angle field transformation algorithm;

[0030] Figure 6 This is a schematic diagram of an adaptive special-shaped convolutional kernel structure;

[0031] Figure 7 This is a schematic diagram of the Informer encoder structure; Detailed Implementation

[0032] The following detailed description, in conjunction with embodiments and accompanying drawings, illustrates the design method and application of the power inspection personnel safety work time limit prediction system based on electroencephalogram (EEG) monitoring of the present invention.

[0033] The present invention provides a power line inspection personnel safety work time limit prediction system based on electroencephalogram (EEG) monitoring, such as... Figure 1 As shown, the device includes an EEG monitoring power safety helmet 1 and a safe working time limit prediction module 2. The EEG monitoring power safety helmet 1 collects EEG signals from the brains of power inspectors in real time and transmits the data to the safe working time limit prediction module 2 via 5G data transmission. The safe working time limit prediction module 2 proposes a long sequence time prediction model (LSTPModel) to process the EEG signals of power inspectors, predict the safe working time limit of power inspectors before work begins, and perform safe working time limit prediction calibration based on the real-time transmitted EEG signals during work, providing a reference for on-site work command and dispatch.

[0034] EEG monitoring electric safety helmet 1, such as Figure 2As shown, the system includes: a power supply module 11, EEG acquisition electrodes 12, a bioelectric signal digital-to-analog converter 13, an MCU processor 14, and a 5G communication module 15. The power supply module 11 is battery-powered and uses a voltage conversion chip to obtain the different operating voltages required by different chips on each module, thus powering each module. The EEG acquisition electrodes 12 are integrated into the shell of the electric safety helmet; there are two in total, and their positions are as shown in the diagram. Figure 3 The electrodes FP1 and FP2 in the 10-20 international standard lead electrode distribution shown are connected to the bioelectric signal digital-to-analog converter 13 via wires to collect EEG signals from different brain regions. The bioelectric signal digital-to-analog converter 13 has a built-in programmable gain amplifier (PGA), internal reference, and on-chip oscillator to perform digital-to-analog conversion on the EEG signals. The MCU processor 14 uses an Xtensa-LX6 core with a working frequency of 300MHz and communicates with the bioelectric signal digital-to-analog converter 12 via the SPI bus. It is responsible for controlling the bioelectric signal digital-to-analog converter 12 to collect EEG signals at a fixed sampling frequency f = 250Hz and send them to the MCU processor 14 via the SPI bus. After transmission, the MCU processor 14 transmits the EEG signal data to the safe working time limit prediction module 2 via the 5G communication module 15.

[0035] The Safe Working Time Limit Prediction Module 2 proposes a Long Sequence Time Prediction Model (LSTPModel), which is deployed in the cloud to process the original EEG signal, extract the temporal sequential features of the EEG signal, and predict the expected safe working time limit for the workers.

[0036] The structure of the Long Sequence Time Prediction Model (LSTP) is as follows: Figure 4 As shown, the system consists of a Gram field transform module, an adaptive shape convolution module, and an Informer encoder module connected in series. The Gram field transform module first concatenates the EEG signals from two channels to obtain a long sequence. Then, the Gram field transform module performs a Gram field transform on this long sequence, transforming the one-dimensional sequence information into two-dimensional image information, such as... Figure 5 As shown, the specific steps of the Gram angle field transformation algorithm are as follows:

[0037] 1) Convert the EEG signal X = {x1, x2, ..., x} n Normalized to the interval [-1, 1], the normalized data is denoted as X. ′ ={x ′ 1,x ′ 2,...,x ′ n}

[0038] 2) Convert the normalized data to polar coordinates, including the polar angle σ. i =cos -1 xi ′ and polar axis r i =t i / N,x i ′ t represents the normalized value of this data point. i Represents the position in the sequence.

[0039] 3) Define the inner product between two sequence points as... <x i ,x j >=cos(σ i +σ j The Gram matrix is ​​calculated using the inner product:

[0040]

[0041] This matrix with a custom inner product is the result of the Gram angle field transform algorithm. It is an n×n square matrix that ultimately transforms the n-dimensional frequency domain feature sequence into n×n×1-dimensional image data.

[0042] The adaptive shape convolution module uses adaptive shape convolution to perform convolution operations on image data. The adaptive shape convolution module includes the following components:

[0043] 1) A data input layer, the input data being the image data output by the Gram field transform module;

[0044] 2) An adaptive shape convolutional layer consists of 45 adaptive shape convolutional kernels. Each adaptive shape convolutional kernel is composed of N convolution points with their respective offsets β = {β1, β2, ..., β...}. N It consists of offsets of different lengths in different directions, to form... Figure 6 Taking the convolutional kernel structure as an example, PReLU is used as the activation function. Where N, β, and λ are parameters obtained through training, and N and β are obtained through training, with... Figure 6 Taking the adaptive shape convolution kernel with N=5 and β={1,2,1,1,1} as an example, the adaptive shape convolution kernel can adaptively adjust the size and shape of the convolution kernel according to the characteristics of the input image data, and adaptively and efficiently extract the hidden EEG signal time-dependent sequence features in the EEG sequence Gram field image data to obtain the time-series feature map.

[0045] 3) A batch channel normalization layer normalizes the temporal feature map in parallel along the batch dimension and the channel dimension. Then, the batch dimension normalized output and the channel dimension normalized output are weighted and summed with the adaptive parameters α and 1-α, respectively, to obtain the normalized output of the batch channel normalization layer, which accelerates the model training process and reduces the degree of overfitting.

[0046] 4) A Dropout layer randomly selects a neuron in the previous layer with probability p1 so that it does not output anything, where p1 = 0.25;

[0047] 5) A Flatten layer transforms multidimensional data into one dimension, yielding a time-series feature sequence.

[0048] The adaptive shape convolution module extracts the hidden EEG signal time-dependent sequence features in the EEG sequence Gram field image data through adaptive shape convolution, and finally obtains the high-level semantic feature sequence.

[0049] The Informer encoder module consists of three such... Figure 7 The diagram shows an Informer encoder and a linear regression layer connected in series. The Informer encoder receives a sequence of high-level semantic features and uses a probabilistic sparse self-attention mechanism to process the temporal feature sequence. Figure 7 The trapezoidal shape represents self-attention distillation, which aims to filter the dominant attention weights, reduce the model size, further extract features from the time-series feature sequence, and obtain the correlation between different channels in the time-series feature sequence. The output of the Informer encoder is input into a linear regression layer, and the regression yields the prediction results of the safe working time limit for power inspection personnel.

[0050] The Long Sequence Time Prediction Model (LSTPModel) is pre-trained based on a pre-built database of safe working time for power inspection personnel. Subsequently, it is federated training based on ten minutes of EEG signal data collected by power inspection personnel wearing EEG monitoring helmets before high-altitude operations. This completes personalized fine-tuning and yields preliminary prediction results of the safe working time limits for power inspection personnel, providing a preliminary reference for the command and dispatch of on-site inspection work.

[0051] During their work, power inspection personnel wear EEG monitoring safety helmets 1 to collect EEG signals in real time and transmit the data via 5G to a safe work time limit prediction module 2. The safe work time limit prediction module 2 analyzes the EEG signals in real time and corrects the previously predicted safe work time limits for the power inspection personnel, providing reference and safety assurance for on-site command and dispatch.

[0052] The present invention has been described in detail above with reference to the accompanying drawings, but is not intended to be limiting. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention, and all such changes should be covered within the scope of the claims of the present invention.

Claims

1. A power inspection personnel safety work time limit prediction system based on electroencephalogram (EEG) monitoring, comprising an EEG monitoring power safety helmet (1) and a safety work time limit prediction module (2), characterized in that, The aforementioned EEG monitoring power safety helmet (1) collects EEG signals from the brains of power inspectors in real time and transmits the data to the safe work time limit prediction module (2) via 5G data transmission. The safe work time limit prediction module (2) proposes a long sequence time prediction model (LSTPModel) to process the EEG signals of power inspectors, predict the safe work time limit of power inspectors before the operation, and perform safe work time limit prediction calibration based on the real-time transmitted EEG signals during the operation, providing a reference for on-site work command and dispatch. The long sequence time prediction model (LSTPModel) proposed is composed of a Gram field transform module, an adaptive shape convolution module, and an Informer encoder module connected in series. The Gram field transform module first concatenates the EEG signals from two channels to obtain a long sequence. Then, the Gram field transform module is applied to this long sequence to transform the one-dimensional sequence information into two-dimensional image information. The specific steps of the Gram field transform algorithm are as follows: 1) EEG signal Normalized to the interval [-1, 1], the normalized data is denoted as ; 2) Convert the normalized data to polar coordinates, including polar angles. and polar axis , This represents the normalized value of the data point. Represents the position in the sequence; 3) Define the inner product between two sequence points as... The Gram matrix is ​​obtained by calculating the inner product: This matrix with a custom inner product is the result of the Gram angle field transformation algorithm, which is an n An n-dimensional square matrix is ​​used to transform the n-dimensional frequency domain feature sequence into an n-dimensional matrix. n 1D EEG sequence Gram field image data; The adaptive shape convolution module uses adaptive shape convolution to perform convolution operations on image data. The adaptive shape convolution module includes, in sequence: 1) A data input layer, the input data being the EEG sequence Gram field image data output by the Gram field transformation module; 2) An adaptive shape-shifting convolutional layer consists of 45 adaptive shape-shifting convolutional kernels. Each adaptive shape-shifting convolutional kernel is composed of N convolution points with their respective offsets. It consists of offsets of different lengths in different directions, using PReLU as the activation function. N, , The parameters are obtained through training, and N and N are obtained through training. The adaptive shape convolution kernel can adaptively adjust the size and shape of the convolution kernel according to the characteristics of the input image data, and adaptively and efficiently extract the hidden EEG signal time-dependent sequence features in EEG sequence Gram field image data to obtain a time-series feature map. 3) A batch-channel normalization layer normalizes the temporal feature map in parallel along both the batch and channel dimensions. The batch-dimensional normalized output and the channel-dimensional normalized output are then compared with adaptive parameters. and Weighted summation yields the normalized output of the batch channel normalization layer, accelerating the model training process and reducing overfitting. 4) A Dropout layer, with probability Randomly select a neuron in the previous layer and prevent it from outputting. = 0.25; 5) A Flatten layer transforms multidimensional data into one-dimensional data, yielding a time-series feature sequence; The adaptive shape convolution module extracts the hidden EEG signal time-dependent sequence features in the EEG sequence Gram field image data through adaptive shape convolution, and finally obtains the high-level semantic feature sequence. The Informer encoder module consists of three Informer encoders and a linear regression layer connected in series. The Informer encoder receives a high-level semantic feature sequence, processes the high-level semantic feature sequence using a probabilistic sparse self-attention mechanism, and uses self-attention distillation to filter the dominant attention weights, thereby reducing the model size. It further extracts features from the time-series feature sequence to obtain the correlation between different channels in the high-level semantic feature sequence. The output of the Informer encoder is input into a linear regression layer, and regression yields the prediction result of the safe working time limit for power inspection personnel.

2. The power line inspection personnel safety work time limit prediction system based on electroencephalogram (EEG) monitoring according to claim 1, characterized in that, The electroencephalogram (EEG) monitoring electric safety helmet (1) includes: a power supply module (11), an EEG acquisition electrode (12), a bioelectric signal digital-to-analog converter (13), an MCU processor (14), and a 5G communication module (15). The power supply module (11) is powered by a battery and obtains the different working voltages required by different chips on each module through a voltage conversion chip to power each module. The EEG acquisition electrodes (12) are integrated in the shell of the electric safety helmet. There are two electrodes in total, which are distributed in the FP1 and FP2 positions of the 10-20 international standard lead electrode distribution. The electrodes are connected to the bioelectric signal digital-to-analog converter (13) through wires to acquire EEG signals from different brain regions. The aforementioned bioelectric signal digital-to-analog converter (13) has a built-in programmable gain amplifier (PGA), internal reference, and on-chip oscillator to perform digital-to-analog conversion on EEG signals; The MCU processor (14) adopts the Xtensa-LX6 core, operates at a frequency of 300MHz, and communicates with the bioelectric signal digital-to-analog converter (13) via the SPI bus. It is responsible for controlling the bioelectric signal digital-to-analog converter (13) to collect EEG signals at a fixed sampling frequency of f = 250Hz and send them to the MCU processor (14) via the SPI bus. After the transmission is completed, the MCU processor (14) transmits the EEG signal data to the safe working time limit prediction module (2) via the 5G communication module (15).

3. The power line inspection personnel safety work time limit prediction system based on electroencephalogram (EEG) monitoring according to claim 1, characterized in that, The safety work time limit prediction module (2) proposes a long sequence time prediction model LSTPModel, which is deployed in the cloud to process the original EEG signal, extract the time dimension sequence characteristics of the EEG signal, and predict the expected safety work time limit of the operator.

4. The power line inspection personnel safety work time limit prediction system based on electroencephalogram (EEG) monitoring according to claim 1, characterized in that, The long-sequence time-series prediction model LSTPModel is pre-trained based on a pre-built database of safe working time of power inspection personnel. Then, it is federated training based on ten minutes of EEG signal data collected by power inspection personnel wearing EEG monitoring power safety helmets (1) before high-altitude operations. The model is then personalized and fine-tuned to obtain preliminary prediction results of the safe working time limit of power inspection personnel, providing preliminary reference for the command and dispatch of on-site inspection work.

5. The power line inspection personnel safety work time limit prediction system based on electroencephalogram (EEG) monitoring according to claim 1, characterized in that, During their work, power inspectors wear EEG monitoring safety helmets (1) to collect EEG signals in real time and transmit the data to the safe work time limit prediction module (2) via 5G data transmission. The safe work time limit prediction module (2) analyzes the EEG signals in real time and corrects the preliminary prediction results of the safe work time limit of the power inspectors in real time, providing reference and safety guarantee for the command and dispatch of on-site work.

Images