A construction worker behavior risk monitoring method based on electroencephalogram signals

Abstract

This disclosure provides a method for monitoring the behavioral risks of construction workers based on electroencephalogram (EEG) signals, belonging to the field of data processing technology. Specifically, it includes: Step 1, collecting multi-channel EEG signals from construction workers; Step 2, preprocessing the multi-channel EEG signals; Step 3, constructing an RCF-InncepLite lightweight EEG recognition model, wherein the RCF-InncepLite lightweight EEG recognition model includes the InncepLite module and the RCF module with a lightweight strategy; Step 4, training the RCF-InncepLite lightweight EEG recognition model using a sample dataset; Step 5, inputting the preprocessed multi-channel EEG signals into the trained RCF-InncepLite lightweight EEG recognition model, and outputting the monitoring results of unsafe behaviors of construction workers. This disclosed solution improves the accuracy and adaptability of monitoring.

Description

Technical Field

[0001] This disclosure relates to the field of data processing technology, and in particular to a method for monitoring the behavioral risks of construction workers based on electroencephalogram (EEG) signals. Background Technology

[0002] Safety management at construction sites is a crucial aspect of ensuring a safe working environment and reducing accidents. In recent years, with increasingly complex working conditions and mounting pressure from tight deadlines, the cognitive load and psychological state of workers have become significant factors affecting safety. Traditional safety management methods are no longer sufficient to meet the needs of real-time perception and intervention of potential cognitive risks among workers. Therefore, the application of electroencephalogram (EEG) signal monitoring technology offers a new direction for safety management at construction sites.

[0003] Currently, electroencephalography (EEG) signals are widely used in high-risk fields such as aviation, transportation, and industrial manufacturing, primarily for assessing cognitive load, monitoring attention, and detecting fatigue. In these fields, specific frequency band characteristics in EEG signals can effectively reflect an individual's cognitive resource mobilization and alertness level, thus becoming an important indicator for monitoring cognitive state. With the interdisciplinary development of neuroscience and human factors engineering, the application of EEG monitoring technology in construction sites has gradually become a research hotspot, providing real-time monitoring of workers' cognitive states and timely early warning and intervention for "high-risk cognitive states."

[0004] However, the high computational resource consumption and inference latency of existing deep learning models make them difficult to apply on lightweight wearable platforms in construction sites. Furthermore, electromagnetic interference, large-scale movements, and unstable task rhythms in construction sites can indeed affect the accurate acquisition and processing of EEG signals. These factors may lead to a decline in the quality of EEG signals, thereby affecting the accuracy of cognitive state monitoring.

[0005] It is evident that there is an urgent need for a high-accuracy and adaptability method for monitoring the behavioral risks of construction workers based on electroencephalogram (EEG) signals. Summary of the Invention

[0006] In view of this, the present disclosure provides a method for monitoring the behavioral risks of construction workers based on electroencephalogram (EEG) signals, which at least partially solves the problems of poor monitoring accuracy and adaptability in the prior art.

[0007] This disclosure provides a method for monitoring the behavioral risks of construction workers based on electroencephalogram (EEG) signals, including:

[0008] Step 1: Collect multi-channel EEG signals from construction workers;

[0009] Step 2: Preprocess the multi-channel EEG signals;

[0010] Step 3: Construct the RCF-IncepLite lightweight EEG recognition model, wherein the RCF-IncepLite lightweight EEG recognition model includes the InncepLite module and the RCF module with a lightweight strategy introduced;

[0011] Step 4: Train the RCF-IncepLite lightweight EEG recognition model using the sample dataset;

[0012] Step 5: Input the preprocessed multi-channel EEG signals into the trained RCF-IncepLite lightweight EEG recognition model and output the risk monitoring results of unsafe behaviors of construction workers.

[0013] According to one specific implementation of this disclosure, the preprocessing includes filtering, artifact removal, and signal normalization.

[0014] According to a specific implementation of an embodiment of this disclosure, step 2 specifically includes:

[0015] Step 2.1: The multi-channel EEG signal is filtered using a bandpass filter to remove low-frequency drift and power frequency noise. The expression for this filtering process is:

[0016]

[0017] in, The transfer function represents the frequency response of the filter to the input signal. Represents a complex variable in the complex frequency domain. Represents the imaginary unit. Represents angular frequency. Indicates the cutoff frequency of the filter;

[0018] Step 2.2: Use the fast ICA algorithm to remove eye movement and heartbeat artifacts from the filtered multi-channel EEG signals;

[0019] Step 2.3: The multi-channel EEG signal after artifact removal is segmented and processed using a sliding window to adapt to the temporal input format of the RCF-IncepLite lightweight EEG recognition model.

[0020] Step 2.4: Standardize the multi-channel EEG signals after artifact removal in each segment to eliminate amplitude differences between different channels.

[0021]

[0022] in, Indicates the first Standardized multi-channel EEG signals for each window Indicates the first Multichannel EEG signals from one window Indicates channel The mean, Indicates channel The standard deviation.

[0023] According to a specific implementation of an embodiment of this disclosure, step 2.2 specifically includes:

[0024] Step 2.2.1: Construct an electrical signal matrix based on the multi-channel EEG signals after filtering:

[0025]

[0026] in, Represents an electrical signal matrix. Represents the set of real numbers. Representing the time dimension of the signal, Indicates the number of channels;

[0027] Step 2.2.2: Perform fast ICA decomposition on the electrical signal matrix to obtain the mixture matrix and independent component matrix:

[0028]

[0029] in, Represents a mixture matrix. Represents the independent component matrix;

[0030] Step 2.2.3: Identify the independent components containing artifacts by observing the power spectral density, peak characteristics and channel distribution of each independent component in the independent component matrix.

[0031] Step 2.2.4: Set the identified independent components containing artifacts to zero to obtain the artifact-free independent component matrix. Then, reconstruct the electrical signal matrix using the mixing matrix and the artifact-free independent component matrix.

[0032]

[0033] in, This represents the reconstructed electrical signal matrix. This represents the independent component matrix after artifact removal.

[0034] According to a specific implementation of an embodiment of this disclosure, the RCF-IncepLite lightweight EEG recognition model includes, in sequence, an RCF module, an InncepLite module, a fusion layer, a global average pooling layer, and a fully connected classifier. Step 3 specifically includes:

[0035] Step 3.1: Input the multi-channel EEG signal into the RCF module to extract periodic rhythm features;

[0036] Step 3.2: Input multi-channel EEG signals into the InncepLite module with a lightweight strategy to extract local perturbation features;

[0037] Step 3.3: The periodic rhythm features and local perturbation features are concatenated along the channel dimension in the fusion layer to obtain the fused features:

[0038]

[0039] in, Indicates fusion characteristics, Indicating periodic rhythm characteristics, Indicates local disturbance characteristics;

[0040] Step 3.4: Use a global average pooling layer to perform pooling operations on the fused features, compressing the fused features into a fixed-length feature vector along the time dimension.

[0041]

[0042] Where Z represents a fixed-length eigenvector. Indicates global average pooling;

[0043] Step 3.5: Output the fixed-length feature vector to the fully connected classifier, and obtain the probability prediction values ​​for each category by passing the Softmax activation function.

[0044]

[0045] in, This represents the probability prediction value. This represents the Softmax activation function. This represents the weights of the fully connected classifier. This indicates the bias of the fully connected classifier;

[0046] Step 3.6: The category with the highest predicted probability is taken as the risk monitoring result of unsafe behavior of construction workers, and the construction of the RCF-IncepLite lightweight EEG recognition model is completed.

[0047] According to a specific implementation of an embodiment of this disclosure, step 3.1 specifically includes:

[0048] Step 3.1.1, construct a learnable periodic template for accurately anchoring periodic rhythms:

[0049]

[0050] in, This represents a learnable periodic template, where L represents the length of the periodic template and c represents the number of channels.

[0051] Step 3.1.2: Using a circular indexing mechanism, the learnable periodic template is aligned with the preprocessed multi-channel EEG signal to extract the periodic rhythm features.

[0052]

[0053] in, This indicates a periodic rhythmic characteristic, where t represents the time index. This represents the learnable time offset parameter, where L represents the length of the periodic template. This indicates a modulo operation.

[0054] According to a specific implementation of an embodiment of this disclosure, step 3.2 specifically includes:

[0055] Step 3.2.1: Differentiate the preprocessed multichannel EEG signal with the periodic embedding to extract aperiodic perturbation features:

[0056]

[0057] in, This indicates the characteristics of non-periodic disturbances. This represents the preprocessed multichannel EEG signals. Indicates periodic rhythmic characteristics;

[0058] Step 3.2.2, through Bottleneck convolution reduces the dimensionality of the aperiodic perturbation features to obtain dimensionality-reduced aperiodic perturbation features:

[0059]

[0060] in, This represents the characteristics of dimension reduction aperiodic perturbation. express convolution, Indicates bottleneck convolution;

[0061] Step 3.2.3: Input the dimensionality-reduced aperiodic perturbation features into three one-dimensional depthwise separable convolutions with different kernels to extract the first local perturbation features, the second local perturbation features, and the third local perturbation features at different scales.

[0062] Step 3.2.4: Input the dimensionality-reduced aperiodic perturbation features into max pooling plus standard convolution to extract the fourth local perturbation features;

[0063] Step 3.2.5: Concatenate the first, second, third, and fourth local perturbation features along the channel dimension, and process them using batch normalization and the ReLU activation function to obtain the local perturbation features:

[0064]

[0065] in, This indicates the characteristics of local disturbances. Represents the ReLU activation function. This indicates batch normalization processing. This represents the first local disturbance characteristic. This represents the second local disturbance characteristic. This represents the third local disturbance characteristic. This represents the fourth local disturbance characteristic.

[0066] According to a specific implementation of an embodiment of this disclosure, step 4 specifically includes:

[0067] Step 4.1: Obtain multi-channel EEG signals and behavioral data of the sample personnel to form a sample dataset, wherein the behavioral data includes operational error rate and task completion time;

[0068] Step 4.2: Using the sample dataset as feature data, construct training samples;

[0069] Step 4.3: Classify and label the training samples using behavioral indicators, NASA-TLX subjective evaluation, and expert consensus assessment.

[0070] Step 4.4: Based on the labeled training samples, the RCF-IncepLite lightweight EEG recognition model is iteratively trained using the Adam optimizer with the goal of minimizing the cross-entropy loss function value until the model converges, thus completing the training of the RCF-IncepLite lightweight EEG recognition model.

[0071] According to one specific implementation of this disclosure, the unsafe behavior risk monitoring results include low risk level, medium risk level, and high risk level.

[0072] The construction worker behavior risk monitoring scheme based on EEG signals in this embodiment includes: Step 1, collecting multi-channel EEG signals from construction workers; Step 2, preprocessing the multi-channel EEG signals; Step 3, constructing an RCF-IncepLite lightweight EEG recognition model, wherein the RCF-IncepLite lightweight EEG recognition model includes the InncepLite module and the RCF module with a lightweight strategy; Step 4, training the RCF-IncepLite lightweight EEG recognition model using a sample dataset; Step 5, inputting the preprocessed multi-channel EEG signals into the trained RCF-IncepLite lightweight EEG recognition model, and outputting the unsafe behavior risk monitoring results of the construction workers.

[0073] The beneficial effects of this disclosure are as follows: The RCF-InncepLite lightweight EEG recognition model, through the introduction of lightweight InncepLite and RCF modules, is specifically designed to reduce computational resource consumption and inference latency, enabling the model to run efficiently on wearable devices with limited computational resources. This achieves real-time monitoring and cognitive state recognition on a lightweight wearable platform at construction sites. Simultaneously, preprocessing of multi-channel EEG signals effectively removes noise caused by electromagnetic interference, large-amplitude movements, and unstable task rhythms, improving signal quality and ensuring accurate reflection of the cognitive state of construction workers even in complex construction environments, thereby enhancing the accuracy and reliability of monitoring results. Attached Figure Description

[0074] To more clearly illustrate the technical solutions of the embodiments of this disclosure, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0075] Figure 1 A flowchart illustrating a method for monitoring the behavioral risks of construction workers based on electroencephalogram (EEG) signals, provided as an embodiment of this disclosure;

[0076] Figure 2 A flowchart illustrating a method for monitoring the behavioral risks of construction workers based on electroencephalogram (EEG) signals, provided in this embodiment of the disclosure;

[0077] Figure 3 A schematic diagram of the structure of an RCF-InncepLite lightweight deep learning model provided in this embodiment of the disclosure;

[0078] Figure 4 This is a schematic diagram of the structure of an InncepLite module provided in an embodiment of this disclosure. Detailed Implementation

[0079] The embodiments of this disclosure will now be described in detail with reference to the accompanying drawings.

[0080] The following specific examples illustrate the implementation of this disclosure. Those skilled in the art can easily understand other advantages and effects of this disclosure from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of this disclosure, and not all of them. This disclosure can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this disclosure. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. Based on the embodiments in this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0081] It should be noted that various aspects of embodiments within the scope of the appended claims are described below. It will be apparent that the aspects described herein can be embodied in a wide variety of forms, and any particular structure and / or function described herein is merely illustrative. Based on this disclosure, those skilled in the art will understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects set forth herein can be used to implement the device and / or practice the method. Additionally, this device and / or method can be implemented using structures and / or functionalities other than one or more of the aspects set forth herein.

[0082] It should also be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of this disclosure. The illustrations only show the components related to this disclosure and are not drawn according to the number, shape and size of the components in actual implementation. In actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0083] Furthermore, specific details are provided in the following description to facilitate a thorough understanding of the examples. However, those skilled in the art will understand that the described aspects can be practiced without these specific details.

[0084] This disclosure provides a method for monitoring the behavioral risks of construction workers based on electroencephalogram (EEG) signals. This method can be applied to the monitoring of personnel behavior in scenarios such as aviation, transportation, and industrial manufacturing.

[0085] See Figure 1This is a flowchart illustrating a method for monitoring the behavioral risks of construction workers based on electroencephalogram (EEG) signals, as provided in this embodiment of the disclosure. Figure 1 and Figure 2 As shown, the method mainly includes the following steps:

[0086] Step 1: Collect multi-channel EEG signals from construction workers;

[0087] In practice, multi-channel EEG signals can be collected from construction workers. Electroencephalography (EEG) records changes in the electrical potential of brain nerve activity by placing electrodes on the scalp. Multi-channel EEG signals refer to EEG signals collected simultaneously at multiple electrode locations, and are typically used to study the activity and interactions between different regions of the brain.

[0088] Step 2: Preprocess the multi-channel EEG signals;

[0089] In practice, preprocessing includes filtering, artifact removal, and signal normalization. The specific preprocessing process can be summarized as follows:

[0090] S201: A bandpass filter is used to filter multi-channel EEG signals, removing low-frequency drift and power frequency noise.

[0091]

[0092] in, Let represent the transfer function of the filter in response to the frequency response of the input signal, and s represent the complex variable in the complex frequency domain. Represents the imaginary unit. Represents angular frequency. This indicates the cutoff frequency of the filter.

[0093] A band-pass filter (BPF) is a commonly used filter in signal processing. It allows signals within a specified frequency range to pass through while suppressing signals below and above that frequency range. In other words, a band-pass filter allows signals within a specific frequency range to pass through while blocking signals of other frequencies.

[0094] Optionally, a second-order Butterworth filter is selected as the bandpass filter, with the passband frequency set to [1 Hz, 45 Hz].

[0095] In this embodiment of the invention, the meaningful frequency bands in the EEG signals are concentrated between 1Hz and 45Hz, such as alpha waves (8-12Hz) and beta waves (13-30Hz). These frequency bands are closely related to cognitive states. The function of the bandpass filter is to retain the signal in the frequency range of interest and filter out frequencies below and above this band, thus providing a basis for accurate monitoring below.

[0096] S202: Employs a fast ICA algorithm to remove eye movement and heartbeat artifacts from filtered multi-channel EEG signals.

[0097] Independent Component Analysis (ICA) is a statistical method for signal separation, designed to extract mutually independent components from multidimensional observed signals.

[0098] Specifically, S202 includes:

[0099] S2021: Construct an electrical signal matrix based on the multi-channel EEG signals after filtering:

[0100]

[0101] in, Represents an electrical signal matrix. Let T represent the set of real numbers, T represent the time dimension of the signal, and C represent the number of channels.

[0102] S2022: Perform fast ICA decomposition on the electrical signal matrix to obtain the mixture matrix and independent component matrix:

[0103]

[0104] Where A represents the mixture matrix and S represents the independent component matrix.

[0105] It should be noted that the mixing matrix is ​​a matrix that describes how the source signals are linearly combined into the observed signal. It defines the transformation process of the signal from the source space to the observation space. In the blind source separation problem, it is usually assumed that we have multiple source signals, which undergo some kind of linear mixing process to generate multiple observation signals.

[0106] The Independent Components (ICM) matrix describes the distribution of the source signals themselves; it is a matrix containing the independent source signals. Using the ICA algorithm, our goal is to recover these source signals from the observed signal, thereby obtaining the ICM matrix.

[0107] S2023: By observing the power spectral density, peak characteristics, and channel distribution of each independent component in the independent component matrix, the independent components containing artifacts are identified.

[0108] S2024 sets the identified independent components containing artifacts to zero, obtaining the artifact-free independent component matrix, and then reconstructs the electrical signal matrix using the mixing matrix and the artifact-free independent component matrix:

[0109]

[0110] in, This represents the reconstructed electrical signal matrix. This represents the independent component matrix after artifact removal.

[0111] Optionally, a sliding window segmentation process is performed on the cleaned multi-channel EEG data to adapt it to the temporal input format of the deep learning model. The window length L is set to 500 time points (corresponding to 2 seconds), and the window sliding step size is 125 (corresponding to 0.5 seconds, 50% overlap).

[0112] In this embodiment of the invention, artifacts such as eye movement, electromyography, and heartbeat are removed by ICA. The cleaned EEG signal can better reflect the real activity of the brain and can effectively improve the accuracy of subsequent analysis.

[0113] S203: The multi-channel EEG signal after artifact removal is segmented and processed using a sliding window to adapt to the temporal input format of the RCF-IncepLite lightweight EEG recognition model.

[0114] It should be noted that long-term EEG signals are difficult to analyze directly. By using the sliding window method, long-term series can be divided into short-term windows, which are suitable for time series analysis.

[0115] S204: Standardize the multi-channel EEG signals after artifact removal in each segment to eliminate amplitude differences between different channels.

[0116]

[0117] in, This represents the standardized multichannel EEG signal of the i-th window. This represents the multichannel EEG signal of the i-th window. This represents the mean c of the channel. This represents the standard deviation of channel c.

[0118] In this embodiment of the invention, EEG signals acquired by different electrodes may differ in amplitude. Standardization processing can eliminate these differences, making the signals from different channels comparable. Simultaneously, by converting the signal of each channel into data with zero mean and unit variance, the computational complexity during model training is reduced.

[0119] Specifically, during the task, multi-channel EEG signals (leads Fp1, Fp2, F3, F4, Fz, etc.) and performance indicators of construction behavior (such as task completion time, operation error rate, etc.) of construction personnel are collected simultaneously. The raw EEG signals are then preprocessed by filtering and independent component analysis to eliminate artifact interference. The processed EEG signals are then segmented by sliding windows of fixed window length to generate standardized data input for direct training and real-time recognition of subsequent deep learning models.

[0120] Step 3: Construct the RCF-IncepLite lightweight EEG recognition model, wherein the RCF-IncepLite lightweight EEG recognition model includes the InncepLite module and the RCF module with a lightweight strategy introduced;

[0121] In practice, the RCF-IncepLite lightweight EEG recognition model is a deep learning model that combines RCF and InncepLite modules. It aims to extract features from multi-channel EEG signals and perform cognitive state monitoring or unsafe behavior risk prediction. This model is specifically designed for low-power and high-performance tasks and can run efficiently on resource-constrained devices such as wearable devices and edge computing devices.

[0122] Specifically, the RCF-IncepLite lightweight EEG recognition model includes, in sequence, an RCF module, an InncepLite module, a fusion layer, a global average pooling layer, and a fully connected classifier. Step 3 mainly includes:

[0123] S301: Based on multi-channel EEG signals, periodic rhythm features are extracted through the RCF module.

[0124] Periodic rhythms typically refer to regular fluctuations in signals that follow certain periodic rules. The period (frequency) of these fluctuations corresponds to different frequency bands of brain electrical activity. The activity in these frequency bands is usually closely related to an individual's cognitive and emotional states; therefore, periodic rhythm characteristics play an important role in many brain electrical signal analysis tasks.

[0125] In one possible implementation, S301 specifically includes:

[0126] S3011: Constructing learnable periodic templates for precise anchoring of periodic rhythms:

[0127]

[0128] in, This represents a learnable periodic template, where L represents the length of the periodic template and c represents the number of channels.

[0129] S3012: Using a cyclic indexing mechanism, the learnable periodic template is aligned with the preprocessed multi-channel EEG signal to extract periodic rhythmic features.

[0130]

[0131] in, This indicates a periodic rhythmic characteristic, where t represents the time index. This represents the learnable time offset parameter, where L represents the length of the periodic template and c represents the channel index. This indicates a modulo operation.

[0132] S302: Based on multi-channel EEG signals, local perturbation features are extracted by introducing the InncepLite module with a lightweight strategy.

[0133] Local disturbance features refer to irregular, short-term fluctuations and changes in a signal. These fluctuations are often related to sudden changes in psychological and physiological states, such as cognitive load, emotional fluctuations, or changes in attention. In electroencephalogram (EEG) signal processing, local disturbances typically reflect non-periodic signal changes, corresponding to periodic rhythms (such as alpha waves, beta waves, etc.).

[0134] Furthermore, S302 specifically includes:

[0135] S3021: The preprocessed multi-channel EEG signal is differentially analyzed with the periodic embedding to extract non-periodic perturbation features:

[0136]

[0137] in, This indicates the characteristics of non-periodic disturbances. This represents the preprocessed multichannel EEG signals. It indicates the characteristics of periodic rhythms.

[0138] S3022: Dimensionality reduction of aperiodic perturbation features is achieved by using 1×1 bottleneck convolution to obtain dimensionality-reduced aperiodic perturbation features:

[0139]

[0140] in, This represents the characteristics of dimension reduction aperiodic perturbation. express convolution, This represents the bottleneck convolution.

[0141] In this embodiment of the invention, by performing dimensionality reduction processing on the local perturbation features, the feature dimensions are reduced, thereby further alleviating the computational burden.

[0142] S3023: The dimensionality-reduced aperiodic perturbation features are input into three one-dimensional depthwise separable convolutions with different kernels to extract the first local perturbation features, the second local perturbation features, and the third local perturbation features at different scales.

[0143] In this embodiment of the invention, by decomposing the standard convolution operation into depthwise convolution and pointwise convolution, the computational load and parameter load are significantly reduced, thereby reducing the computational complexity of the model.

[0144] S3024: Input the dimensionality-reduced aperiodic perturbation features into max pooling plus standard convolution to extract the fourth local perturbation features.

[0145] S3025: The first, second, third, and fourth local perturbation features are concatenated along the channel dimension and processed by batch normalization and ReLU activation function to obtain the local perturbation features:

[0146]

[0147] in, This indicates the characteristics of local disturbances. Represents the ReLU activation function. This indicates batch normalization processing. This represents the first local disturbance characteristic. This represents the second local disturbance characteristic. This represents the third local disturbance characteristic. This represents the fourth local disturbance characteristic.

[0148] S303: Periodic rhythmic features and local perturbation features are concatenated along the channel dimension in the fusion layer to obtain fused features:

[0149]

[0150] in, Indicates fusion characteristics, Indicating periodic rhythm characteristics, This indicates the characteristics of local disturbances.

[0151] In this embodiment of the invention, by splicing together periodic rhythmic features and local perturbation features, the model can simultaneously consider the brain's long-term stable activity and short-term sudden changes, thereby obtaining more comprehensive information and improving the ability to predict cognitive load and unsafe behaviors.

[0152] S304: A global average pooling layer is used to pool the fused features, compressing the fused features into a fixed-length feature vector along the time dimension.

[0153]

[0154] Where Z represents a fixed-length eigenvector. This indicates global average pooling.

[0155] S305: Output the fixed-length feature vector to the fully connected classifier, and obtain the probability prediction values ​​for each class through the Softmax activation function.

[0156]

[0157] in, This represents the probability prediction value. This represents the Softmax activation function. represents the weights of the fully connected classifier, and b represents the bias of the fully connected classifier.

[0158] S306: Use the category with the highest predicted probability as the result of unsafe behavior risk monitoring for construction workers, and complete the RCF-IncepLite lightweight EEG recognition model.

[0159] Specifically, the constructed RCF-IncepLite network adopts a "dual-path + fusion" structure, specifically targeting the periodic rhythmic structure and local non-periodic perturbation characteristics of EEG signals, and performing targeted modeling for each. The input EEG data is first divided into several sliding windows (2 seconds in length and 0.5 seconds in step). The signal from each window enters two modules: one path inputs to the periodic modeling module RCF Block to learn the rhythmic structure in the signal; the other path inputs to the multi-scale perturbation modeling module InncepLite Block to extract non-periodic perturbation patterns; finally, features are concatenated in the fusion layer, and three cognitive states are output by a fully connected classifier.

[0160] In this embodiment of the invention, the model adopts a highly lightweight architecture, with the overall parameter count controlled within 0.3M, effectively reducing memory and computing power consumption and meeting the resource constraints and power consumption limitations of wearable devices. Simultaneously, by introducing an RCF module with rhythmic structure learning capabilities, the robustness of modeling unstable EEG temporal rhythms is significantly improved, enhancing the model's adaptability to changes in β-wave structure under different individual and task conditions.

[0161] Furthermore, the model focuses on neural oscillation features closely related to cognitive load, such as the β band (13–30 Hz), which can capture cognitive state changes induced by task pressure with high sensitivity and achieve efficient identification of increased load states.

[0162] Step 4: Train the RCF-IncepLite lightweight EEG recognition model using the sample dataset;

[0163] In practice, the process of training the RCF-IncepLite lightweight EEG recognition model using a sample dataset is as follows:

[0164] S401: Acquire multi-channel EEG signals and behavioral data of construction workers, including operational error rate and task completion time.

[0165] In this embodiment of the invention, by collecting multi-channel EEG signals and behavioral data, it is possible to simultaneously monitor an individual's EEG activity and actual behavioral performance, providing more comprehensive information. Behavioral data (such as operational error rate and task completion time) can provide the model with important clues about cognitive load and the risk of unsafe behaviors.

[0166] S402: Use multi-channel EEG signals and behavioral data as feature data to construct training samples.

[0167] S403: Classify and label training samples using behavioral indicators, NASA-TLX subjective evaluation, and expert consensus assessment.

[0168] Behavioral indicators are quantitative metrics used to assess an individual's cognitive state or task performance by observing their behavioral behavior. Behavioral indicators typically reflect an individual's psychological and physiological state when performing a task and are widely used in cognitive load monitoring and unsafe behavior prediction.

[0169] NASA-TLX (Task Load Index) is a widely used subjective assessment tool used to measure an individual's cognitive load when completing a task. NASA-TLX quantifies task load through six dimensions, covering multiple psychological and physiological factors in the task execution process.

[0170] Among them, expert consensus assessment is an assessment method based on expert experience and judgment. In this method, domain experts judge the cognitive load and behavioral risk of individuals based on their understanding, experience and observation of the task, and label the training samples.

[0171] Specifically, the labeling process employs a standardized approach that integrates behavioral indicators, NASA-TLX subjective evaluation, and expert consensus assessment to generate unified and objective cognitive status labels, categorized into low-risk, medium-risk, and high-risk levels.

[0172] In this embodiment of the invention, by combining three annotation methods (behavioral metrics, NASA-TLX, and expert evaluation), uniform and objective labels can be generated for training samples, ensuring high quality of model training data and thereby improving the model's generalization ability in different environments and scenarios.

[0173] S404: Based on the labeled training samples, the RCF-IncepLite lightweight EEG recognition model is iteratively trained using the Adam optimizer with the goal of minimizing the cross-entropy loss function, until the model converges, thus completing the training of the RCF-IncepLite lightweight EEG recognition model.

[0174]

[0175] in, Let N represent the cross-entropy loss function, and C represent the total number of training samples and the total number of classes. This indicates the category to which the true label of the i-th sample belongs. This represents the predicted category of the i-th sample. Represents a logarithmic function.

[0176] Step 5: Input the preprocessed multi-channel EEG signals into the trained RCF-IncepLite lightweight EEG recognition model and output the risk monitoring results of unsafe behaviors of construction workers.

[0177] In practice, the trained RCF-IncepLite lightweight EEG recognition model can promptly identify and predict the cognitive load and behavioral risks of construction workers. For example, the monitoring results of unsafe behavior risks include low-risk, medium-risk, and high-risk levels. Simultaneously, this model helps construction workers identify cognitive load and behavioral risks in a timely manner, enabling them to react quickly, reduce the occurrence of unsafe behaviors, and improve overall safety management efficiency.

[0178] The construction worker behavior risk monitoring method provided in this embodiment utilizes the RCF-IncepLite lightweight EEG recognition model. By incorporating lightweight InncepLite and RCF modules, it is specifically designed to reduce computational resource consumption and inference latency, enabling the model to run efficiently on resource-constrained wearable devices. This achieves real-time monitoring and cognitive state recognition on a lightweight wearable platform at the construction site. Furthermore, preprocessing of multi-channel EEG signals effectively removes noise from electromagnetic interference, large-amplitude movements, and unstable task rhythms, improving signal quality and ensuring accurate reflection of the cognitive state of construction workers even in complex construction environments, thereby enhancing the accuracy and reliability of monitoring results.

[0179] It should be understood that the various parts of this disclosure can be implemented in hardware, software, firmware, or a combination thereof.

[0180] The above description is merely a specific embodiment of this disclosure, but the scope of protection of this disclosure is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this disclosure should be included within the scope of protection of this disclosure. Therefore, the scope of protection of this disclosure should be determined by the scope of the claims.

Claims

1. A method for monitoring behavioral risks of construction workers based on electroencephalogram (EEG) signals, characterized in that, include: Step 1: Collect multi-channel EEG signals from construction workers; Step 2: Preprocess the multi-channel EEG signals; Step 3: Construct the RCF-IncepLite lightweight EEG recognition model, wherein the RCF-IncepLite lightweight EEG recognition model includes a ncepLite module and an RCF module that incorporate a lightweight strategy; The RCF-IncepLite lightweight EEG recognition model includes, in sequence, an RCF module, an InncepLite module, a fusion layer, a global average pooling layer, and a fully connected classifier. Step 3 specifically includes: Step 3.1: Input the multi-channel EEG signal into the RCF module to extract periodic rhythm features; Step 3.2: Input multi-channel EEG signals into the InncepLite module with a lightweight strategy to extract local perturbation features; Step 3.3: The periodic rhythm features and local perturbation features are concatenated along the channel dimension in the fusion layer to obtain the fused features: in, Indicates fusion characteristics, Indicating periodic rhythm characteristics, Indicates local disturbance characteristics; Step 3.4: Use a global average pooling layer to perform pooling operations on the fused features, compressing the fused features into a fixed-length feature vector along the time dimension. Where Z represents a fixed-length eigenvector. Indicates global average pooling; Step 3.5: Output the fixed-length feature vector to the fully connected classifier, and obtain the probability prediction values ​​for each category by passing the Softmax activation function. in, This represents the probability prediction value. This represents the Softmax activation function. This represents the weights of the fully connected classifier. This indicates the bias of the fully connected classifier; Step 3.6: Take the category with the highest predicted probability as the risk monitoring result of unsafe behavior of construction workers, and complete the construction of the RCF-IncepLite lightweight EEG recognition model; Step 3.1 specifically includes: Step 3.1.1, construct a learnable periodic template for accurately anchoring periodic rhythms: in, This represents a learnable periodic template. Let L represent the set of real numbers, and L represent the length of the periodic template. Indicates the number of channels; Step 3.1.2: Using a circular indexing mechanism, the learnable periodic template is aligned with the preprocessed multi-channel EEG signal to extract the periodic rhythm features. in, This indicates a periodic rhythmic characteristic, where t represents the time index. This represents the learnable time offset parameter, where L represents the length of the periodic template. Represents modulo operation; Step 3.2 specifically includes: Step 3.2.1: Differentiate the preprocessed multichannel EEG signal with the periodic embedding to extract aperiodic perturbation features: in, This indicates the characteristics of non-periodic disturbances. This represents the preprocessed multichannel EEG signals. Indicates periodic rhythmic characteristics; Step 3.2.2, through Bottleneck convolution reduces the dimensionality of the aperiodic perturbation features to obtain dimensionality-reduced aperiodic perturbation features: in, This represents the characteristics of dimension reduction aperiodic perturbation. express convolution, Indicates bottleneck convolution; Step 3.2.3: Input the dimensionality-reduced aperiodic perturbation features into three one-dimensional depthwise separable convolutions with different kernels to extract the first local perturbation features, the second local perturbation features, and the third local perturbation features at different scales. Step 3.2.4: Input the dimensionality-reduced aperiodic perturbation features into max pooling plus standard convolution to extract the fourth local perturbation features; Step 3.2.5: Concatenate the first, second, third, and fourth local perturbation features along the channel dimension, and process them using batch normalization and the ReLU activation function to obtain the local perturbation features: in, This indicates the characteristics of local disturbances. Represents the ReLU activation function. This indicates batch normalization processing. This represents the first local disturbance characteristic. This represents the second local disturbance characteristic. This represents the third local disturbance characteristic. This indicates the fourth local disturbance characteristic; Step 4: Train the RCF-IncepLite lightweight EEG recognition model using the sample dataset; Step 5: Input the preprocessed multi-channel EEG signals into the trained RCF-IncepLite lightweight EEG recognition model and output the risk monitoring results of unsafe behaviors of construction workers.

2. The method according to claim 1, characterized in that, The preprocessing includes filtering, artifact removal, and signal normalization.

3. The method according to claim 1, characterized in that, Step 2 specifically includes: Step 2.1: The multi-channel EEG signal is filtered using a bandpass filter to remove low-frequency drift and power frequency noise. The expression for this filtering process is: in, The transfer function represents the frequency response of the filter to the input signal. Represents a complex variable in the complex frequency domain. Represents the imaginary unit. Represents angular frequency. Indicates the cutoff frequency of the filter; Step 2.2: Use the fast ICA algorithm to remove eye movement and heartbeat artifacts from the filtered multi-channel EEG signals; Step 2.3: The multi-channel EEG signal after artifact removal is segmented and processed using a sliding window to adapt to the temporal input format of the RCF-IncepLite lightweight EEG recognition model. Step 2.4: Standardize the multi-channel EEG signals after artifact removal in each segment to eliminate amplitude differences between different channels. in, Indicates the first Standardized multi-channel EEG signals for each window Indicates the first Multichannel EEG signals from one window Indicates channel The mean, Indicates channel The standard deviation.

4. The method according to claim 3, characterized in that, Step 2.2 specifically includes: Step 2.2.1: Construct an electrical signal matrix based on the multi-channel EEG signals after filtering: in, Represents an electrical signal matrix. Represents the set of real numbers. Representing the time dimension of the signal, Indicates the number of channels; Step 2.2.2: Perform fast ICA decomposition on the electrical signal matrix to obtain the mixture matrix and independent component matrix: in, Represents a mixture matrix. Represents the independent component matrix; Step 2.2.3: Identify the independent components containing artifacts by observing the power spectral density, peak characteristics and channel distribution of each independent component in the independent component matrix. Step 2.2.4: Set the identified independent components containing artifacts to zero to obtain the artifact-free independent component matrix. Then, reconstruct the electrical signal matrix using the mixing matrix and the artifact-free independent component matrix. in, This represents the reconstructed electrical signal matrix. This represents the independent component matrix after artifact removal.

5. The method according to claim 4, characterized in that, Step 4 specifically includes: Step 4.1: Obtain multi-channel EEG signals and behavioral data of the sample personnel to form a sample dataset, wherein the behavioral data includes operational error rate and task completion time; Step 4.2: Using the sample dataset as feature data, construct training samples; Step 4.3: Classify and label the training samples using behavioral indicators, NASA-TLX subjective evaluation, and expert consensus assessment. Step 4.4: Based on the labeled training samples, the RCF-IncepLite lightweight EEG recognition model is iteratively trained using the Adam optimizer with the goal of minimizing the cross-entropy loss function value until the model converges, thus completing the training of the RCF-IncepLite lightweight EEG recognition model.

6. The method according to claim 5, characterized in that, The results of the unsafe behavior risk monitoring include low-risk, medium-risk, and high-risk levels.

Images