Intelligent safety helmet operation fatigue risk early warning classification algorithm based on fractional order depth extreme learning machine

Abstract

The invention provides an intelligent safety helmet operation fatigue risk early warning classification algorithm based on a fractional order depth extreme learning machine. Belongs to the technical field of intelligent wearable equipment and artificial intelligence. The classification algorithm is executed through the following method, and the method comprises the steps that original data are obtained in real time through multiple sensors integrated on the intelligent safety helmet, and the original data comprise personnel physiological data and environment data; the method comprises the following steps: preprocessing collected original data, performing fractional differential or integral processing on the data by using a fractional calculus theory, and extracting deeper information features; by means of multiple sensors (such as heart rate, brain wave, acceleration and environment sensors) integrated on the intelligent safety helmet, physiological data and environment data of an operator can be obtained in real time.

Description

Technical Field

[0001] The present invention proposes an intelligent safety helmet operation fatigue risk warning classification algorithm based on a fractional-order deep extreme learning machine, belonging to the technical fields of intelligent wearable devices and artificial intelligence. Background Art

[0002] Power operation personnel often face high-risk scenarios such as high-altitude climbing and high-voltage equipment operation, and their fatigue state directly affects operation safety. Traditional methods rely on manual inspections, with problems such as strong experience dependence and lag in identifying potential safety hazards. In the prior art, feature extraction methods based on time-frequency domain analysis (such as Fourier transform) and entropy value theory (such as multi-scale entropy) have limitations in processing non-stationary signals, while deep learning models (such as CNN, RNN), although having high accuracy, have defects such as large computational resource requirements, long training time, and poor real-time performance. In addition, traditional optimization algorithms are prone to falling into local optima and are difficult to meet the real-time requirements of complex industrial scenarios. Therefore, there is an urgent need for an efficient, lightweight, and adaptable fatigue risk warning solution. Summary of the Invention

[0003] The present invention provides an intelligent safety helmet operation fatigue risk warning classification algorithm based on a fractional-order deep extreme learning machine to solve the problems mentioned in the above background art:

[0004] The intelligent safety helmet operation fatigue risk warning classification algorithm based on a fractional-order deep extreme learning machine proposed by the present invention is executed by the following method, and the method includes:

[0005] S1. Real-time acquisition of raw data through multi-sensors integrated on the intelligent safety helmet, where the raw data includes personnel physiological data and environmental data;

[0006] S2. Preprocess the collected raw data, and perform fractional-order differentiation or integration processing on the data using the fractional-order calculus theory to extract deeper information features;

[0007] S3. Construct a fractional-order deep extreme learning machine model, combine the fractional-order calculus theory with the deep extreme learning machine, and use the fractional-order calculus theory to enhance the model's ability to process complex non-linear data;

[0008] S4. Train the fractional-order deep extreme learning machine model with a large amount of historical data, and optimize the model parameters through cross-validation and grid search;

[0009] S5. Real-time input the newly collected data into the trained model to output the fatigue risk level; according to the risk level, the intelligent safety helmet can send warning signals to the operation personnel through multiple means, and at the same time, the warning information can be uploaded to the cloud management platform.

[0010] Beneficial effects of the present invention: Through the multi-sensors integrated on the intelligent safety helmet (such as heart rate, electroencephalogram, acceleration, and environmental sensors), physiological data and environmental data of the operators can be obtained in real time. Using the fractional calculus theory to preprocess the original data can extract information features at a deeper level, enhancing the representativeness and effectiveness of the data; constructing a fractional-order deep extreme learning machine model, combining the fractional calculus theory with the deep extreme learning machine, can effectively process complex non-linear data. This combination not only enhances the processing ability of the model but also improves the accuracy of fatigue risk prediction; training the model with a large amount of historical data and optimizing the model parameters through cross-validation and grid search ensure the generalization ability and prediction performance of the model; according to the fatigue risk level output by the model, the intelligent safety helmet can issue warnings to the operators through various means (such as sound, vibration, optical signals, etc.), and at the same time upload the warning information to the cloud management platform for remote monitoring and analysis; using clustering analysis and association rule mining techniques can discover the potential laws and influencing factors of the operators' fatigue states, further optimizing work arrangements and preventive measures; after real-time data collection, the system will perform data cleaning and verification, and set different levels of fatigue risk warning thresholds according to business requirements and safety standards. Using an adaptive learning algorithm or machine learning technology to dynamically adjust the warning thresholds to adapt to different working environments and individual differences; through the built-in data stream processing framework and asynchronous processing technology, real-time data collection, processing, and input are realized. Introducing a data caching mechanism and a thread pool technology improves the efficiency of data processing and the response speed of the system; during the data stream processing, system metrics are collected in real time through monitoring tools, and optimization operations are performed according to the monitoring results to ensure the stable operation and high efficiency of the system. Brief Description of the Drawings

[0011] Figure 1 It is the algorithm flowchart of the present invention;

[0012] Figure 2 It is the schematic diagram of the layer-by-layer training mechanism of the fractional-order DELM of the present invention;

[0013] Figure 3 It is the algorithm structure diagram of the present invention. Detailed Embodiments

[0014] The following describes the preferred embodiments of the present invention with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention and are not used to limit the present invention.

[0015] An embodiment of the present invention is an intelligent safety helmet operation fatigue risk warning classification algorithm based on a fractional-order deep extreme learning machine. The classification algorithm is executed through the following method, and the method includes:

[0016] S1. Real-time obtain the original data through multiple sensors integrated on the intelligent safety helmet, where the original data includes personnel physiological data and environmental data;

[0017] S2. Preprocess the collected original data, perform fractional-order differentiation or integration on the data using the fractional calculus theory to extract deeper information features, and based on the preprocessed data, extract the feature vectors related to fatigue risk, where the feature vectors include heart rate variability, electroencephalogram power spectral density, and head movement frequency;

[0018] S3. Construct a fractional-order deep extreme learning machine model, combine the fractional calculus theory with the deep extreme learning machine, and use the fractional calculus theory to enhance the model's processing ability for complex non-linear data;

[0019] S4. Train the fractional-order deep extreme learning machine model with a large amount of historical data, and optimize the model parameters through cross-validation and grid search;

[0020] S5. Input the newly collected data into the trained model in real-time and output the fatigue risk level; according to the risk level, the intelligent safety helmet can send early warning signals to the operators through multiple means (such as audible and visual alarms, vibration prompts, etc.), and at the same time, the early warning information can be uploaded to the cloud management platform.

[0021] The working principle of the above technical solution is as follows: A variety of sensors are integrated on the intelligent safety helmet, including a heart rate sensor, an electroencephalogram sensor, an acceleration sensor, and an environmental sensor. These sensors can collect the physiological parameters (such as heart rate and electroencephalogram) of the operator in real time, as well as the head movement state (such as nodding and shaking the head), and data such as the temperature, humidity, and light intensity of the working environment. These data are the basis for subsequent analysis and early warning; the collected raw data needs to go through preprocessing and feature extraction steps. In the preprocessing stage, fractional calculus theory is used to perform fractional differentiation or integration on the data. This processing method can extract deeper information features and is helpful for subsequent classification and early warning. In the feature extraction stage, feature vectors closely related to fatigue risk are extracted from the preprocessed data, such as heart rate variability, electroencephalogram power spectral density, and head movement frequency. These feature vectors can reflect the fatigue state of the operator; a fractional-order deep extreme learning machine model is constructed. This model combines the advantages of fractional calculus theory and deep extreme learning machine and can enhance the processing ability of complex nonlinear data. The input of the model is the feature vector, and the output is the fatigue risk level (such as low, medium, high). In the model training stage, a large amount of historical data is used to train the model, and the model parameters are optimized through methods such as cross-validation and grid search to ensure the accuracy and stability of the model; the trained model can process newly collected data in real time and output the fatigue risk level. According to the risk level, the intelligent safety helmet can send warning signals to the operator in various ways (such as sound and light alarms, vibration prompts, etc.) to remind them to pay attention to their fatigue state and take corresponding measures. At the same time, the warning information can also be uploaded to the cloud management platform for managers to remotely monitor and manage the fatigue state of the operators.

[0022] The effects of the above technical solutions are as follows: The intelligent safety helmet integrates a heart rate sensor, an electroencephalogram sensor, an acceleration sensor, and an environmental sensor, and can comprehensively and real-time obtain the physiological data (such as heart rate, electroencephalogram, and head movement state) of the operator and environmental data (such as temperature and humidity, light intensity). These data provide a rich information basis for subsequent fatigue risk early warning; by performing fractional-order differentiation or integration processing on the collected raw data, deeper information features can be extracted. This processing method can capture the subtle changes and complex non-linear features of the data better than traditional integer-order calculus, thereby improving the depth and accuracy of data processing; combining the fractional-order calculus theory with the deep extreme learning machine to construct a fractional-order deep extreme learning machine model. This model not only inherits the high efficiency and easy expandability of the deep extreme learning machine, but also enhances its processing ability for complex non-linear data through the fractional-order calculus theory. This enables the model to more accurately perform fatigue risk early warning when facing complex data in the actual working environment; using a large amount of historical data to train the model, and optimizing the model parameters through cross-validation and grid search. This step ensures that the model can maintain high accuracy and stability when processing new data, and at the same time improves the generalization performance of the model; inputting the newly collected data into the trained model in real time, and the model can quickly output the fatigue risk level. According to the risk level, the intelligent safety helmet can send warning signals to the operator in various ways such as sound and light alarms and vibration prompts. This real-time early warning mechanism helps the operator to detect the fatigue state in time and take corresponding measures, thereby reducing the operation risk; the warning information can also be uploaded to the cloud management platform for the management personnel to remotely monitor and manage the fatigue state of the operator.

[0023] In one embodiment of the present invention, the S1 includes:

[0024] S11. Obtain the raw data through multiple sensors;

[0025] S12. Synchronize the timestamps of the raw data collected by all sensors through the clock system of the central processing unit;

[0026] S13. Adopt a sliding window filtering algorithm to initially eliminate obvious outliers and record the time points when abnormal data appears;

[0027] S14. Adopt a median filtering or Kalman filtering algorithm to smooth the heart rate data, and adopt an independent component analysis or principal component analysis algorithm to remove noise components such as eye movement artifacts and electromyogram interference in the electroencephalogram signal.

[0028] The working principle of the above technical solution is: Obtain the raw data through multiple sensors, including:

[0029] Heart rate sensor: It adopts a medical-grade heart rate monitoring module and collects the heart rate data of the operator in real time through a skin-contact sensor or optical sensing technology. The sensor position needs to be correct and fixed with conductive glue or an elastic band to reduce motion artifacts and improve data accuracy. Heart rate data is an important indicator reflecting the heart health status and fatigue level of the operator.

[0030] Brain wave sensor: It adopts non-invasive EEG technology and is equipped with a 32-channel electrode cap that covers key areas of the brain to capture brain wave activities in different frequency bands. These frequency bands (such as α, β, θ, δ) can reflect different states of the brain, including wakefulness, relaxation, fatigue, etc. Conductive paste is used to reduce impedance and ensure signal quality. Brain wave data is an important basis for evaluating the brain fatigue state of the operator.

[0031] Acceleration sensor: It is built-in with a high-precision triaxial accelerometer to detect subtle head movements, such as nodding and shaking the head. Combining algorithm analysis, it can evaluate the attention concentration of the operator. The sensor is fixed inside the safety helmet to ensure stability and reliability. Acceleration data helps to identify the head movement state of the operator, thereby indirectly reflecting their fatigue level.

[0032] Environmental sensor: It integrates high-precision temperature and humidity sensors and light intensity sensors to monitor the temperature, humidity, and light conditions of the working environment in real time. These environmental factors have a significant impact on the fatigue state of the operator. The sensor is placed outside the safety helmet to avoid occlusion and ensure data accuracy.

[0033] Through the clock system of the central processing unit (such as an ARM Cortex-M series microcontroller), the timestamps of the raw data collected by all sensors are synchronized. This step ensures the temporal consistency of the data collected by different sensors and provides a basis for subsequent data analysis and processing. The sliding window filtering algorithm is adopted to initially eliminate obvious outliers. For example, when the heart rate data exceeds the normal range (such as below 40 beats per minute or above 120 beats per minute) or when the brain wave signal is severely interfered (such as the signal-to-noise ratio is below a certain threshold), these abnormal data are eliminated and the time points when the abnormal data appears are recorded. This step helps to reduce data noise and improve the accuracy of subsequent analysis. The median filtering or Kalman filtering algorithm is used to smooth the heart rate data to reduce random fluctuations and noise interference and improve the smoothness and stability of the data; the independent component analysis (ICA) or principal component analysis (PCA) algorithm is used to remove noise components such as eye movement artifacts and electromyographic interference in the brain wave signal.

[0034] The effects of the above technical solution are as follows: By integrating a heart rate sensor, an electroencephalogram sensor, an acceleration sensor, and an environmental sensor, this technical solution can comprehensively collect the physiological data (heart rate, electroencephalogram, head movement state) of the operator and environmental data (temperature and humidity, light intensity). This comprehensive data collection provides a solid foundation for subsequent fatigue risk warning; the selected heart rate sensor, electroencephalogram sensor, and acceleration sensor are all high-precision devices, which can collect data in real time and accurately. At the same time, through reasonable sensor fixing methods and the use of conductive materials, motion artifacts and signal interference are reduced, further improving the accuracy of the data; the raw data collected by all sensors is time-stamped synchronized through the clock system of the central processing unit, ensuring the temporal consistency of data from different sensors. This helps subsequent data analysis and processing, improving the reliability and stability of the data; the sliding window filtering algorithm is used to initially eliminate obvious outliers, and the median filtering or Kalman filtering algorithm is used to smooth the heart rate data. These steps effectively reduce the noise and abnormal fluctuations in the data, improving the smoothness and stability of the data. At the same time, the independent component analysis (ICA) or principal component analysis (PCA) algorithm is used to remove the noise components in the electroencephalogram signal, further improving the signal-to-noise ratio and reliability of the electroencephalogram data; this technical solution can collect and process data in real time, providing timely and accurate information support for fatigue risk warning. This helps operators and managers to detect fatigue status in a timely manner and take corresponding measures to reduce operation risks; based on the processed data, the intelligent safety helmet can send warning signals to the operator through various means such as sound and light alarms and vibration prompts. This multi-means warning method can more effectively attract the attention of the operator and improve the warning effect; the warning information can be uploaded to the cloud management platform, facilitating managers to remotely monitor and manage the fatigue status of the operator.

[0035] In one embodiment of the present invention, the S2 includes:

[0036] S21. Perform fractional-order differential processing on the heart rate variability data, extract the subtle features of heart rate changes, and use the Grünwald-Letnikov fractional-order differential formula to determine the optimal order through simulation tests;

[0037] S22. Analyze the influence of the fractional-order differential order on the feature extraction effect, use the cross-validation method, compare the model performance under different orders, and determine the optimal order;

[0038] S23. Perform fractional-order integral processing on the electroencephalogram signal to smooth the signal, and use the Riemann-Liouville fractional-order integral formula to determine the optimal order through simulation tests;

[0039] S24. Analyze the influence of the fractional integral order on the extraction of electroencephalogram features, use the grid search method to compare the model performance under different orders, and determine the optimal parameters;

[0040] S25. Calculate the heart rate variability index and evaluate the cardiac autonomic regulation function based on the calculation results;

[0041] S26. Use the fast Fourier transform to analyze the power spectral density of each frequency band, and combine the processing results of fractional integral to extract the power spectral density features of the α, β, θ, and δ frequency bands;

[0042] S27. Statistically analyze the frequency and amplitude of the head movement state (such as nodding and shaking the head), combine the acceleration data to construct a head movement feature vector; and use a method combining time-domain analysis and frequency-domain analysis to extract features.

[0043] The working principle of the above technical solution is as follows: The Grünwald-Letnikov fractional differential formula is used to process the heart rate variability data. This formula allows the signal to be differentiated at any order. By adjusting the order, the feature extraction effect can be optimized; through simulation experiments, compare the feature extraction results under different orders to determine the optimal fractional differential order, so as to maximize the extraction of useful information in heart rate variability; use the cross-validation method to divide the data set into a training set and a test set, and evaluate the fractional differential results of different orders. By comparing the performance of the model on the test set, determine the optimal order; use the Riemann-Liouville fractional integral formula to process the electroencephalogram signal. Similar to fractional differentiation, fractional integral also allows the signal to be processed at any order to optimize the feature extraction effect; through simulation experiments, determine the optimal fractional integral order to maximize the extraction of useful information in the electroencephalogram signal; use the grid search method to search within the preset order range and compare the model performance under different orders. By evaluating the performance of the model on the validation set, determine the best parameters; based on the heart rate variability data, use statistical methods to calculate these indicators. These indicators can reflect the activity state and balance of the cardiac autonomic nervous system; perform FFT transformation on the electroencephalogram signal to obtain its representation in the frequency domain. Then, divide each frequency band according to the frequency range and calculate the power spectral density of each frequency band. Combine the processing results of fractional integral to extract the features of these frequency bands; use an acceleration sensor to collect head movement data and use a method combining time-domain analysis (such as mean and variance) and frequency-domain analysis (such as power spectral density) to extract features. These features can reflect the subtle movement state of the head and provide useful information for subsequent fatigue warning and attention monitoring.

[0044] The effects of the above technical solutions are as follows: By performing fractional-order differential processing on heart rate variability data, it is possible to more precisely capture the subtle features of heart rate changes, such as heart rate acceleration and deceleration, thereby more accurately reflecting the changes in the activity of the autonomic nervous system. This processing method can reveal the dynamic characteristics of the data better than traditional integer-order differential; fractional-order integration processing of electroencephalogram signals can smooth the signals and enhance the contrast of the power spectral density in different frequency bands, making the change trend of brain fatigue more obvious. This helps to improve the accuracy and sensitivity of feature extraction, providing stronger support for subsequent fatigue warning and attention monitoring; using cross-validation method and grid search method to determine the optimal fractional-order differential and integration orders can ensure that the model can achieve the best performance when processing different data sets. This method avoids the problems of overfitting and underfitting, and improves the generalization ability of the model; by comparing the model performance under different orders, the optimal parameter settings can be determined, thereby further optimizing the prediction accuracy and stability of the model; calculating heart rate variability indices (such as SDNN, RMSSD, PNN50, etc.) can comprehensively evaluate the cardiac autonomic regulation function. These indices can reflect the activity status and balance of the cardiac autonomic nervous system, which is of great significance for monitoring heart health and preventing cardiovascular diseases; using fast Fourier transform to analyze the power spectral density of each frequency band and combining with the results of fractional-order integration processing can extract the power spectral density features of α, β, θ, δ frequency bands. These features can reflect the brain's activity in different states, providing important basis for applications such as brain fatigue monitoring and attention assessment; statistically analyzing the frequency and amplitude of head movement states (such as nodding, shaking the head) and combining with acceleration data to construct a head movement feature vector can reflect the subtle movement state of the head. This feature vector has practical application value for monitoring the attention concentration and fatigue level of operators; using a method that combines time-domain analysis (such as mean, variance) and frequency-domain analysis (such as power spectral density) to extract features can make full use of the time-frequency characteristics of the data. This method can more comprehensively reflect the dynamic changes of the data and improve the effectiveness and reliability of feature extraction.

[0045] In one embodiment of the present invention, S27 includes:

[0046] S271. Preprocess the head movement data collected by the acceleration sensor, and based on the preprocessed data, extract the features of head movement, including movement frequency, movement amplitude, movement direction, etc.;

[0047] S272. Adopt a method that combines time-domain analysis and frequency-domain analysis to comprehensively describe the state of head movement;

[0048] S273. Construct a head movement state classification model through machine learning algorithms; use the extracted features as input and the head movement state (such as nodding, shaking the head, stationary, etc.) as output for model training;

[0049] S274. Optimize the classification model using cross - validation, and screen for features that have a significant impact on the classification results through feature importance evaluation;

[0050] S275. Based on the results of the classification model, construct a head motion feature vector, fuse the head motion feature vector with other physiological features such as heart rate variability features and electroencephalogram features, and use feature splicing to integrate feature information from different sources to form a more comprehensive and rich feature set.

[0051] The working principle of the above technical solution is as follows: Clean the original head motion data collected by the acceleration sensor to remove noise and outliers, ensuring the accuracy and reliability of the data. At the same time, standardize the data to make different features comparable; Based on the pre - processed data, extract the features of head motion, including motion frequency, motion amplitude, motion direction, etc. These features can reflect the motion state and motion trend of the head, providing a basis for subsequent classification and feature vector construction; Use statistical indicators such as mean, variance, and standard deviation to perform time - domain analysis on the head motion data to describe the changes in head motion in the time domain; Use methods such as fast Fourier transform and power spectral density to perform frequency - domain analysis on the head motion data to reveal the features of head motion in the frequency domain; Combine the results of time - domain analysis and frequency - domain analysis to comprehensively describe the state of head motion, providing rich feature information for the subsequent construction of machine learning models; According to the complexity of the problem and the characteristics of the data, select an appropriate machine learning algorithm (such as support vector machine, decision tree, random forest, etc.) as the classifier; Use the extracted head motion features as input and the head motion state (such as nodding, shaking the head, stationary, etc.) as output to train the machine learning algorithm. By adjusting the algorithm parameters and optimizing the model structure, improve the classification accuracy and generalization ability of the model; Use the cross - validation method to optimize the classification model, evaluate the performance of the model on different data sets, and ensure the stability and reliability of the model; Through feature importance evaluation methods (such as feature importance ranking based on decision trees, feature importance evaluation based on random forests, etc.), screen for features that have a significant impact on the classification results. This helps to further optimize the feature selection strategy and improve the classification performance of the model; Based on the results of the classification model, construct a head motion feature vector. This feature vector includes key features such as the probability distribution of head motion state, motion frequency, motion amplitude, etc., which can comprehensively reflect the state and trend of head motion; Fuse the head motion feature vector with other physiological features such as heart rate variability features and electroencephalogram features. Integrate feature information from different sources through feature splicing to form a more comprehensive and rich feature set.

[0052] The effects of the above technical solutions are as follows: Clean the original head movement data collected by the acceleration sensor to remove noise and outliers, ensuring the accuracy and reliability of the data. At the same time, standardize the data to make different features comparable; Based on the preprocessed data, extract the features of head movement, including movement frequency, movement amplitude, movement direction, etc. These features can reflect the movement state and movement trend of the head, providing a basis for subsequent classification and feature vector construction; Use statistical indicators such as mean, variance, and standard deviation to perform time-domain analysis on the head movement data to describe the changes of head movement in the time domain; Use methods such as fast Fourier transform and power spectral density to perform frequency-domain analysis on the head movement data to reveal the characteristics of head movement in the frequency domain; Combine the results of time-domain analysis and frequency-domain analysis to comprehensively describe the state of head movement, providing rich feature information for the construction of subsequent machine learning models; According to the complexity of the problem and the characteristics of the data, select appropriate machine learning algorithms (such as support vector machine, decision tree, random forest, etc.) as classifiers; Use the extracted head movement features as input and the head movement state (such as nodding, shaking, stationary, etc.) as output to train the machine learning algorithm. By adjusting the algorithm parameters and optimizing the model structure, improve the classification accuracy and generalization ability of the model; Use the cross-validation method to optimize the classification model, evaluate the performance of the model on different data sets, and ensure the stability and reliability of the model; Through feature importance evaluation methods (such as feature importance ranking based on decision tree, feature importance evaluation based on random forest, etc.), screen the features that have a significant impact on the classification results. This helps to further optimize the feature selection strategy and improve the classification performance of the model; Based on the results of the classification model, construct a head movement feature vector. This feature vector includes key features such as the probability distribution of head movement state, movement frequency, movement amplitude, etc., which can comprehensively reflect the state and trend of head movement; Integrate the head movement feature vector with other physiological features such as heart rate variability features and electroencephalogram features. Integrate the feature information from different sources through feature splicing to form a more comprehensive and rich feature set.

[0053] In one embodiment of the present invention, the S275 includes:

[0054] On the basis of the original probability distribution of head movement state, movement frequency, and movement amplitude, further extract the dynamic features of head movement, such as acceleration change rate, smoothness of movement trajectory, correlation between the speed and acceleration of head movement, etc.;

[0055] Use time series analysis technology to model the head movement data and extract time series features;

[0056] Concatenate the head movement feature vector with other physiological features such as heart rate variability features (e.g., SDNN, RMSSD, PNN50, etc.) and electroencephalogram features (e.g., power spectral density in α, β, θ, δ frequency bands) to form a preliminarily fused feature set;

[0057] Use principal component analysis (PCA) to perform dimensionality reduction on the preliminarily fused feature set, and through feature importance evaluation, such as feature selection methods based on random forest or Lasso regression, further screen features that make significant contributions to the target tasks (e.g., fatigue monitoring, emotion recognition);

[0058] Utilize the constructed feature set to train a machine learning model, validate the model through cross-validation, evaluate its performance on the target tasks, and iteratively optimize the feature set according to the validation results.

[0059] The working principle of the above technical solution is as follows: by calculating the derivative of the acceleration of the head movement, the rate of change of the acceleration is obtained, which can reflect the speed of the acceleration change of the head movement, that is, the degree of the movement; by calculating the curvature of the head movement trajectory or using filtering technology to evaluate the smoothness of the trajectory, this helps to identify whether the head movement is stable or has sudden changes; analyzing the correlation between the head movement speed and acceleration can reveal the dynamic characteristics of the head movement, such as whether there is a trend of acceleration or deceleration; predicting the future movement state based on the historical information of the head movement data, and evaluating the autocorrelation between the data by calculating the autocorrelation coefficient; predicting the future movement state by weighted averaging the past errors of the head movement data, which helps to smooth the random fluctuations in the data; combining the characteristics of AR and MA, considering both historical information and random fluctuations to more accurately model the time series characteristics of the head movement data; splicing the head movement feature vector with other physiological features such as heart rate variability features and brain wave features to form a preliminary fused feature set. This helps to integrate information from different sources and improve the recognition accuracy of the target task; performing dimensionality reduction processing on the preliminary fused feature set, reducing the number of features by retaining the most important principal components, while trying to maintain the information volume of the original data. This helps reduce the computational complexity of the model and improves operational efficiency; use feature selection methods such as random forest or Lasso regression to further screen features that contribute significantly to the target task. This helps remove redundant or irrelevant features and improve the generalization ability of the model; select appropriate machine learning models (such as deep learning networks, support vector machines, etc.) for training based on the characteristics of the target task and the characteristics of the data; verify the model through cross-validation methods to evaluate its performance on the target task. This helps ensure the stability and reliability of the model and avoid overfitting or underfitting problems; iteratively optimize the feature set based on the verification results, adjust the feature selection strategy or model parameters to improve the performance of the model. This helps to continuously improve the recognition accuracy and generalization ability of the target task.

[0060] The effects of the above technical solutions are as follows: By extracting dynamic features such as the acceleration change rate, the smoothness of the motion trajectory, and the correlation between speed and acceleration, the complexity and diversity of head movements can be more comprehensively reflected; these dynamic features help capture the subtle changes in head movements and improve the accuracy of feature extraction; using autoregressive models (AR), moving average models (MA), or autoregressive moving average models (ARMA) to model head movement data can extract time series features such as autocorrelation coefficients and partial autocorrelation coefficients; these time series features can reveal the temporal dependence and change trends in head movement data, providing more valuable information for subsequent machine learning models; splicing the head movement feature vector with other physiological features such as heart rate variability features and electroencephalogram features to form a preliminarily fused feature set; this way of feature fusion can integrate information from different sources and improve the recognition accuracy of the target task; using principal component analysis (PCA) to perform dimensionality reduction on the preliminarily fused feature set can remove redundant and irrelevant features and reduce the computational complexity of the model; the dimensionality-reduced feature set still retains most of the information in the original data, helping subsequent machine learning models to learn and predict more effectively; through feature selection methods based on random forests or Lasso regression, further screen the features that make significant contributions to the target task; this way of feature selection can remove the features that contribute less to the target task and improve the generalization ability of the model; using the constructed feature set to train machine learning models such as deep learning networks and support vector machines; verifying the model through cross-validation and evaluating its performance on the target task; iteratively optimizing the feature set according to the verification results to continuously improve the recognition accuracy and generalization ability of the model; by monitoring the changes in head movement states and other physiological features, the fatigue states of drivers or operators can be detected in a timely manner, improving safety and work efficiency; combining information such as head movement features and electroencephalogram features can achieve accurate recognition and analysis of individual emotions, providing strong support for fields such as mental health assessment and human-computer interaction.

[0061] In one embodiment of the present invention, S3 includes:

[0062] S31. Construct a fractional-order deep extreme learning machine model, combining fractional-order calculus theory and deep neural network structure to enhance the model's ability to process complex non-linear data;

[0063] S32. The input layer receives the preprocessed feature vector, the hidden layer adopts a multi-layer deep neural network structure, and the output layer uses the softmax function to output the probability distribution of the fatigue risk level.

[0064] S33. During the forward propagation of the neural network, introduce fractional calculus operations, adjust the weight update rule, and adopt an adaptive learning rate adjustment strategy to improve the training efficiency;

[0065] S34. Based on the fusion strategy of fractional calculus in the neural network, such as fractional convolution, fractional pooling, etc., through the fractional convolution kernel, integrate fractional differential or integral operations into the convolution operation to extract deeper features;

[0066] S35. Through the fractional pooling strategy, perform fractional integration or differentiation on the feature map; use a large amount of historical data to train the model, adjust the model parameters through the backpropagation algorithm; and adopt the momentum optimization algorithm to accelerate convergence;

[0067] S36. Implement cross-validation to evaluate the generalization ability of the model, and use grid search technology to optimize the model parameters. At the same time, adopt the early stopping method strategy to prevent the model from overfitting on the training set;

[0068] S37. Use multiple metrics such as accuracy, recall rate, and F1 score to evaluate the model performance; analyze the prediction accuracy of the model at different fatigue risk levels, and evaluate the robustness and generalization ability of the model;

[0069] S38. According to the evaluation results, optimize the model, adopt ensemble learning methods (such as Bagging, Boosting, etc.), and combine multiple fractional-order deep extreme learning machine models to improve the overall prediction performance.

[0070] The working principle of the above technical solution is as follows: Combine the fractional calculus theory with the deep neural network structure to enhance the model's ability to process complex non-linear data. Fractional calculus is an extension of traditional integer-order calculus and can more precisely describe the dynamic behavior of the system. Based on the deep extreme learning machine, introduce fractional calculus operations to construct a fractional deep extreme learning machine model. Receive the preprocessed feature vectors, which contain multi-physiological feature information such as head movement state, heart rate variability, and electroencephalogram features. Adopt a multi-layer deep neural network structure, such as a convolutional neural network (CNN) for feature extraction and a recurrent neural network (RNN) for capturing time series information. Use the softmax function to output the probability distribution of the fatigue risk level for subsequent prediction and classification. During the forward propagation process of the neural network, introduce fractional calculus operations and adjust the weight update rule to adapt to the learning requirements of fractional features. Adopt an adaptive learning rate adjustment strategy to dynamically adjust the learning rate according to the loss change during training to improve the training efficiency. Through the fractional convolution kernel, incorporate fractional differential or integral operations into the convolution operation to extract deeper features. These features can more accurately reflect the relationship between the head movement state and the fatigue risk level. In the convolutional layer, use the fractional convolution kernel for convolution operations to obtain the fractional feature map. Perform fractional integration or differentiation on the feature map to reduce feature redundancy and improve the generalization ability of the model. Use a large amount of historical data to train the model and adjust the model parameters through the backpropagation algorithm. At the same time, adopt momentum optimization algorithms (such as Adam, RMSprop, etc.) to accelerate convergence. Implement cross-validation to evaluate the generalization ability of the model and ensure the stable performance of the model on different datasets. Use grid search technology to optimize model parameters, including the learning rate, the number of neurons in the hidden layer, the fractional order, the size of the convolution kernel, etc. Adopt the early stopping method strategy to prevent the model from overfitting on the training set and improve the generalization performance of the model. Use multiple metrics such as accuracy, recall, and F1-score to evaluate the model performance and comprehensively reflect the prediction accuracy of the model at different fatigue risk levels. Analyze the performance of the model in different scenarios and datasets to evaluate the robustness and generalization ability of the model. According to the evaluation results, optimize the model, such as adjusting the model structure, parameters, etc., to improve the prediction performance. Adopt an ensemble learning method (such as Bagging, Boosting, etc.) to combine multiple fractional deep extreme learning machine models to improve the overall prediction performance. By integrating the results of multiple models, the error of a single model can be reduced and the overall prediction accuracy can be improved.

[0071] The effects of the above technical solution are as follows: By combining the fractional calculus theory with the deep neural network structure, the processing ability of the model for complex non-linear data is significantly enhanced. Fractional calculus can more precisely describe the dynamic behavior of the system, thereby improving the fitting and prediction ability of the model for complex data; the hidden layer adopts a multi-layer deep neural network structure, such as convolutional neural network and recurrent neural network, which can more effectively extract and fuse features and improve the recognition accuracy of the model; the output layer uses the softmax function, which can output the probability distribution of the fatigue risk level, providing strong support for subsequent decision-making and evaluation; fractional calculus operations are introduced in the forward propagation process of the neural network, and the weight update rule is adjusted to adapt to the learning requirements of fractional features. At the same time, an adaptive learning rate adjustment strategy is adopted to effectively improve the training efficiency of the model; through the fractional convolution kernel and fractional pooling strategy, deeper features can be extracted and feature redundancy can be reduced, thereby improving the generalization ability of the model; cross-validation is implemented to evaluate the generalization ability of the model, and grid search technology is used to optimize the model parameters, including learning rate, the number of neurons in the hidden layer, fractional order, convolution kernel size, etc. It helps to find the optimal combination of model parameters and improve the prediction performance of the model; the early stopping method strategy is adopted to effectively prevent the model from overfitting on the training set and further improve the generalization ability of the model; multiple indicators such as accuracy, recall rate, and F1 score are used to evaluate the model performance, which can comprehensively reflect the prediction accuracy of the model at different fatigue risk levels; the model is optimized according to the evaluation results, and an ensemble learning method is used to combine multiple fractional-order deep extreme learning machine models, which can significantly improve the overall prediction performance. It helps to further improve the accuracy and robustness of the model; this technical solution has broad application prospects in the fields of fatigue monitoring, emotion recognition, human-computer interaction, etc. By accurately predicting the fatigue risk level, timely warnings and interventions can be provided for drivers, operators, etc., improving safety and work efficiency. At the same time, this technical solution can also provide useful references for other physiological signal processing and emotion computing tasks.

[0072] In one embodiment of the present invention, step S4 includes:

[0073] S41. Prepare an independent validation data set, including fatigue risk labels under different job types, environmental conditions, job personnel physical constitutions, etc., and ensure that the validation data set is consistent with the training data set in distribution;

[0074] S42. Comprehensively evaluate the model performance using evaluation indicators such as accuracy, recall rate, F1 score, and AUC-ROC curve; compare the prediction accuracy of the model under different thresholds and determine the optimal threshold;

[0075] S43. Analyze the differences in the prediction performance of the model under different job types, environmental conditions, and job personnel physical constitutions, and evaluate the adaptability and robustness of the model;

[0076] S44. Adopt feature importance evaluation methods (such as Lasso regression, random forest feature importance, etc.) to analyze the impact of different features on the model prediction performance; optimize the feature selection strategy according to the feature importance ranking.

[0077] The working principle of the above technical solution is as follows: To evaluate the generalization ability of the model, an independent validation dataset consistent with the distribution of the training dataset needs to be prepared. This dataset should contain fatigue risk labels under various factors such as different job types, environmental conditions, and the physical constitution of workers; collect and preprocess this diverse data to ensure that they can comprehensively reflect various situations in the actual application scenario. At the same time, through means such as data augmentation and data cleaning, ensure the quality of the validation dataset; adopt evaluation metrics such as accuracy, recall rate, F1 score, and AUC-ROC curve to comprehensively evaluate the performance of the model. These metrics can reflect the prediction accuracy, stability, and robustness of the model from different perspectives; by comparing the prediction accuracy of the model at different thresholds, find the threshold that makes the model performance optimal. This threshold will be used in the subsequent prediction and decision-making process; understand the differences in the prediction performance of the model under different job types, environmental conditions, and the physical constitution of workers to evaluate the adaptability and robustness of the model; group the validation dataset according to different factors, and then evaluate the performance of the model in these groups respectively. By comparing the performance metrics of different groups, the performance differences of the model in different scenarios can be found, and thus targeted optimization can be carried out; adopt evaluation methods such as Lasso regression and random forest feature importance to analyze the impact of different features on the model prediction performance. These methods can quantify the contribution degree of each feature to the model prediction result; optimize the feature selection strategy according to the feature importance ranking. Remove those features that contribute less to the model prediction performance, which can reduce the complexity of the model, improve the training efficiency and prediction speed. At the same time, retain those features that have a significant impact on the model performance to ensure the prediction accuracy of the model.

[0078] The effects of the above technical solutions are as follows: By preparing an independent verification dataset containing fatigue risk labels under different job types, environmental conditions, and the physical constitutions of workers, the verification process of the model performance can be ensured to be comprehensive and accurate. This helps to discover potential problems of the model in specific scenarios and provides strong support for subsequent model optimization; By using multiple evaluation metrics such as accuracy, recall rate, F1 score, and AUC-ROC curve, the performance of the model can be comprehensively evaluated. These metrics can reflect the prediction accuracy, stability, and robustness of the model from different perspectives and provide more comprehensive guidance for model optimization; By comparing the prediction performance differences of the model under different job types, environmental conditions, and the physical constitutions of workers, the adaptability and robustness of the model can be evaluated. This helps to discover the performance differences of the model in different scenarios, so as to conduct targeted optimization and improve the generalization ability of the model in practical applications; By using evaluation methods such as Lasso regression and random forest feature importance, the influence of different features on the model prediction performance can be quantified. This helps to identify the key features that contribute greatly to the model prediction results and provides a scientific basis for subsequent feature selection and model optimization; According to the feature importance ranking, optimizing the feature selection strategy can reduce the complexity of the model, improve the training efficiency and prediction speed. At the same time, retaining those features that have a significant impact on the model performance can ensure the prediction accuracy of the model and further improve the practicality and reliability of the model.

[0079] In one embodiment of the present invention, the step S5 includes:

[0080] S51. Input the newly collected data into the trained model in real time and calculate the fatigue risk level;

[0081] S52. Give early warnings to workers through a multi-means early warning mechanism according to the risk level;

[0082] S53. Upload the early warning information to the cloud management platform for remote monitoring and analysis, and combine big data and artificial intelligence algorithms to further explore the data value;

[0083] S54. Adopt clustering analysis and association rule mining to discover the potential laws and influencing factors of workers' fatigue states, and analyze the changing trends of workers' fatigue states through early warning reports and data analysis reports, and optimize the job arrangements and preventive measures.

[0084] The working principle of the above technical solution is as follows: newly collected data (such as physiological signals, behavioral characteristics, etc.) are input into the trained fractional-order deep extreme learning machine model in real time. The model calculates and outputs the corresponding fatigue risk level according to the input data; data are collected through intelligent devices (such as intelligent safety helmets, wearable sensors, etc.) and sent to the computing platform where the model is located through wireless transmission. After receiving the data, the model immediately performs calculations and outputs the results; according to the fatigue risk level output by the model, early warnings are given to the operators through various means such as acoustic and optical alarms, vibration prompts, and visual displays.

[0085] Acoustic and optical alarm: The intelligent safety helmet emits sound and light signals of different frequencies to intuitively remind the operator of the fatigue risk. The alarm mode can be set according to actual needs, such as continuous alarm, intermittent alarm, etc.

[0086] Vibration prompt: Tactile feedback is provided through the vibration motor inside the safety helmet to enhance the early warning effect. Different levels of fatigue risk correspond to different frequencies and intensities of vibration.

[0087] Visual display: The fatigue risk level and brief suggestions are displayed on the display screen of the safety helmet, facilitating the operator to immediately understand their own status. At the same time, a historical data query function is provided to help the operator analyze the fatigue trend.

[0088] Early warning devices such as intelligent safety helmets are built with sensors and controllers, which trigger corresponding early warning mechanisms according to the received fatigue risk level information. The early warning information is uploaded to the cloud management platform for remote monitoring and analysis. The data value is further mined through big data and artificial intelligence algorithms to provide a scientific basis for optimizing work arrangements and preventive measures; the early warning device sends the early warning information to the cloud management platform through wireless transmission. After receiving the information, the platform stores, analyzes, and visualizes it. At the same time, the data is deeply mined and analyzed using big data and artificial intelligence algorithms; data mining techniques such as cluster analysis and association rule mining are used to discover the potential laws and influencing factors of the operator's fatigue state. Through forms such as early warning reports and data analysis reports, the change trend of the operator's fatigue state is analyzed, and the work arrangement and preventive measures are optimized; data mining tools and algorithms are used to deeply analyze and mine the data stored in the cloud management platform. Early warning reports and data analysis reports are generated according to the analysis results to provide decision-making support for managers.

[0089] The effects of the above technical solutions are as follows: By collecting data in real time and inputting it into the trained model, the fatigue risk level of the operator can be quickly calculated, providing a basis for immediate warning. Combining various means such as audible and visual alarms, vibration prompts, and visual displays can intuitively and effectively send warning signals to the operator, reminding them to pay attention to the fatigue risk and take timely measures to adjust or rest, thus avoiding potential safety hazards. Designing various alarm modes such as continuous alarm and intermittent alarm can meet the requirements in different scenarios, improving the pertinence and effectiveness of the warning. By combining vibration prompts and visual displays, multiple sensory feedback is provided to enhance the warning effect, enabling the operator to more keenly perceive the fatigue risk. Providing a historical data query function helps the operator analyze their fatigue trend and understand the changing pattern of their own state, so as to better manage their health and working hours. Uploading the warning information to the cloud management platform for remote monitoring and analysis can achieve comprehensive and continuous monitoring of the operator, promptly discovering and handling potential problems. Combining big data and artificial intelligence algorithms to further explore the data value can discover the potential patterns and influencing factors of the operator's fatigue state, providing a scientific basis for optimizing work arrangements and preventive measures. Through technical means such as cluster analysis and association rule mining, the internal rules of the operator's fatigue state can be deeply explored, providing strong support for formulating more scientific and reasonable work arrangements and preventive measures. Generating warning reports and data analysis reports can intuitively display the changing trend of the operator's fatigue state and potential risk points, providing a decision-making basis for managers and helping them promptly adjust work plans, optimize work processes, and strengthen preventive measures, etc.

[0090] In one embodiment of the present invention, the S51 includes:

[0091] S511. After collecting data in real time, perform data cleaning to remove noise, outliers, and incomplete data, and verify the data; format the cleaned data according to the model input requirements and standardize the data.

[0092] S512. Based on the data stream processing technology, collect, process, and input the data in real time through multi-threading.

[0093] S513. Before starting the real-time warning system, load the trained fatigue risk prediction model and perform initialization operations.

[0094] S514. Input the preprocessed data into the model in real time for inference, calculate the fatigue risk level, and output the inference result in real time.

[0095] S515. Set fatigue risk warning thresholds at different levels according to business requirements and safety standards. And use an adaptive learning algorithm or machine learning technology to dynamically adjust the warning thresholds; trigger corresponding multi - means warning mechanisms such as audible and visual alarms, vibration prompts, and visual displays according to the calculated fatigue risk level and the set warning thresholds.

[0096] The working principle of the above - mentioned technical solution is as follows: After real - time data collection, first perform data cleaning work. This step aims to remove noise, outliers, and incomplete data to ensure the quality of the data input into the model. At the same time, verify the data to validate its integrity and accuracy; the cleaned data needs to be formatted according to the model input requirements. For example, convert time - series data into window data of a fixed length to meet the model's input requirements. In addition, standardize the data to eliminate the dimensional differences between different features and ensure that the model can fairly process each feature; adopt data stream processing technology to achieve real - time data collection, processing, and input through multi - threading. This technology can efficiently process a large amount of real - time data and ensure that the data can be input into the model in a timely and accurate manner; before the real - time warning system is started, load the trained fatigue risk prediction model. This model is trained based on historical data and can accurately predict the fatigue risk level of operators; perform an initialization operation on the loaded model to ensure that the model can run normally and is ready to accept the input of real - time data; input the pre - processed data into the model in real - time for inference calculation. The model will calculate the corresponding fatigue risk level according to the input data; output the inferred result in real - time for subsequent warning mechanisms to use; set fatigue risk warning thresholds at different levels according to business requirements and safety standards. These thresholds are used to determine whether an operator is in a dangerous state and trigger corresponding warning mechanisms; considering factors such as individual differences of operators and changes in the working environment, use an adaptive learning algorithm or machine learning technology to dynamically adjust the warning thresholds. This adjustment can ensure that the warning mechanism remains accurate and effective in different situations; trigger corresponding multi - means warning mechanisms according to the calculated fatigue risk level and the set warning thresholds. These mechanisms include audible and visual alarms, vibration prompts, and visual displays, etc., which can intuitively remind operators to pay attention to fatigue risks and take corresponding measures for adjustment or rest.

[0097] The effects of the above technical solutions are as follows: By removing noise, outliers, and incomplete data and validating the data, the quality of the data input into the model is ensured. This helps improve the accuracy and stability of model predictions and reduce false alarms or missed alarms caused by data problems; The cleaned data is formatted according to the model input requirements and standardized to eliminate the dimensional differences between different features. This helps the model better understand and process the data and improves the accuracy and generalization ability of predictions; Based on data stream processing technology, real-time data collection, processing, and input are performed on the data through multi-threading, realizing an efficient data processing flow. This ensures that the data can be input into the model in real time and provides a basis for real-time warning; The preprocessed data is input into the model for inference in real time, the fatigue risk level is calculated, and the inference result is output in real time. This realizes the instant monitoring and warning of the fatigue state of operators and helps to detect potential safety hazards in a timely manner; According to business requirements and safety standards, fatigue risk warning thresholds at different levels are set. This helps achieve accurate warning for different risk levels and improves the pertinence and effectiveness of warnings; Considering factors such as individual differences of operators and changes in the working environment, an adaptive learning algorithm or machine learning technology is used to dynamically adjust the warning threshold. This ensures that the warning mechanism can maintain accuracy and effectiveness in different situations and improves the flexibility and adaptability of warnings; According to the calculated fatigue risk level and the set warning threshold, a corresponding multi-means warning mechanism is triggered, such as audible and visual alarms, vibration prompts, and visual displays. This realizes all-round and multi-sensory warnings for operators and helps to attract their attention and take measures in a timely manner; Through real-time data processing and an efficient warning mechanism, the fatigue state of operators can be detected in a timely manner and corresponding measures can be taken to avoid accidents and injuries caused by fatigue and improve work safety; Through continuous monitoring and analysis of the fatigue state of operators, the working hours and rest hours can be arranged more scientifically and reasonably, the operation process can be optimized, and work efficiency can be improved.

[0098] In one embodiment of the present invention, the S512 includes:

[0099] Based on the built-in data stream processing framework, a data stream processing process is designed; Each link in the data stream processing process is divided into multiple independent tasks, and an independent thread is assigned to each task;

[0100] Asynchronous processing technology is adopted, and non-blocking processing of data is carried out through asynchronous calls and callback mechanisms; And through a thread pool, thread reuse and dynamic adjustment are performed;

[0101] During the data stream processing, a data caching mechanism is introduced. Through caching, the real-time collected data is temporarily stored, and a certain-scale data batch is formed before processing;

[0102] Through a monitoring tool, collect in real time metrics such as the CPU usage rate, memory occupancy rate, and data processing latency of the system, and perform optimization operations based on the monitoring results.

[0103] The working principle of the above technical solution is as follows: First, it is necessary to access real-time data sources, such as sensor data, log data, etc. These data sources will serve as the starting point for data stream processing; before the data enters model inference, preprocessing operations need to be performed, such as data cleaning, formatting, standardization, etc. These operations aim to improve data quality and ensure that the model can accurately understand and process the data; to balance real-time performance and processing efficiency, a data caching mechanism is introduced. The real-time collected data will be temporarily stored in the cache and processed after forming a certain scale of data batches. This helps reduce data processing latency and improve system throughput; the preprocessed data is input into the trained model for inference calculation of the fatigue risk level. This is the core link in the data stream processing process; the inference results are output in real time for subsequent warning mechanisms to use. The output form can be data stream, file, database record, etc.; each link in the data stream processing process is divided into multiple independent tasks, and an independent thread is assigned to each task. This helps achieve parallel processing of data and improve system processing efficiency; in data preprocessing, model inference and other links, asynchronous processing technology is adopted. Through asynchronous calls and callback mechanisms, non-blocking processing of data is achieved. This helps reduce the waiting time during data processing and improve system response speed; a thread pool is used for thread reuse and dynamic adjustment. The thread pool can manage a certain number of threads and dynamically adjust the number of threads according to the system load. This helps avoid resource competition caused by too many threads and processing latency caused by too few threads; a data caching mechanism is introduced to temporarily store the real-time collected data. The cache can be a memory data structure, such as a queue, heap, etc., or a temporary file on disk; when the data in the cache reaches a certain scale, a batch processing operation is triggered. The data in the cache is taken out to form a data batch, and then subsequent processing and inference are performed. This helps reduce data processing latency and improve system throughput; through a monitoring tool, collect in real time metrics such as the CPU usage rate, memory occupancy rate, and data processing latency of the system. These metrics can reflect the operating status and performance bottlenecks of the system; perform corresponding optimization operations based on the monitoring results. For example, adjust the thread pool size, optimize the data processing algorithm, improve the cache strategy, etc. These operations aim to improve the processing efficiency and stability of the system.

[0104] The effects of the above technical solutions are as follows: By designing a clear data flow processing process and dividing each link into multiple independent tasks, the modularization and parallelization of data processing are achieved. This not only improves the efficiency of data processing but also makes the system easier to maintain and expand; Independent threads are assigned to each task, and asynchronous processing technology is adopted to achieve non-blocking data processing. This greatly reduces the waiting time during data processing and improves the response speed and throughput of the system; Through the thread pool for thread reuse and dynamic adjustment, the resource consumption caused by the frequent creation and destruction of threads is effectively avoided. At the same time, the number of threads is dynamically adjusted according to the system load to ensure the reasonable utilization of resources and the stability of the system; A data caching mechanism is introduced to temporarily store the real-time collected data and process it after forming a certain scale of data batches. This helps to reduce the frequency of data processing, reduce the processing pressure of the system, and improve the efficiency of data processing; Through monitoring tools, indicators such as the CPU usage rate, memory occupancy rate, and data processing delay of the system are collected in real time, providing strong data support for system performance evaluation; Corresponding optimization operations are performed according to the monitoring results, such as adjusting the thread pool size and optimizing the data processing algorithm. This helps to timely discover and solve system performance bottlenecks and improve the overall performance and stability of the system; Through the data flow processing framework and multi-threaded technology, the real-time collection and processing of data are achieved. This ensures that data can be input into the model in a timely and accurate manner, providing timely and accurate data support for the subsequent warning mechanism; The preprocessed data is input into the trained model for inference calculation of the fatigue risk level. Since efficient processing technologies are adopted in both the data preprocessing and model inference links, the accuracy and reliability of the inference results can be ensured.

[0105] In one embodiment of the present invention, the composite multi-scale fractional-order attention entropy is used to comprehensively extract feature data, and then the obtained features are input to the fractional-order DELM for multi-class fatigue warning processing. Finally, a new escape optimization algorithm is used to optimize the network parameters, such as Figure 1 shown. And, the specific method includes:

[0106] 1. Feature extraction

[0107] 1.1. Fractional-order attention entropy

[0108] Compared with traditional entropy, attention entropy has the advantages of strong robustness to the time series length and no need to set hyperparameters. Calculating the time series attention entropy can be summarized into the following steps:

[0109] (1) If each point in the time series is regarded as a system, the change of its state can be regarded as the system's adjustment to the environment. The peak point can effectively characterize the change of the upper and lower bounds of the local state, so the local peak point is defined as the key point.

[0110] (2) Given a finite sequence X, first define the key examples Ω. Secondly, calculate the subsequences u of any given interval X i , u k and u j , which satisfy that u i and u j match in the examples Ω, but for any i < k < j, u k does not match in Ω.

[0111] If a point satisfies one of the following conditions, we define it as a peak point, including local maximum and local minimum:

[0112] x i-1 < x i and x i > x i+1 (defined as local maximum)

[0113] x i < x i-1 and x i < x i+1 (defined as local minimum)

[0114] If each point in the time series is regarded as a state of the system, then the change of state can be regarded as the system's adjustment to the environment. A complex system is expected to have a complex state change process when adapting to the environment. Peak points represent the local upper and lower limits of state changes, making them potential key examples. Then, the time series can be represented by the peak point sequence. Then we calculate the interval between two consecutive peak points.

[0115] Set the intervals of local maximum to local maximum (Max - Max), local minimum to local minimum (Min - Min), local maximum to local minimum (Max - Min), and local minimum to local maximum (Min - Max) as key points according to four different strategies, calculate the number of interval points between adjacent key points, and calculate the frequency of all intervals.

[0116] (3) Calculate the Shannon entropy of the interval frequency between adjacent key points. The following is the fractional - order Shannon entropy formula for the j - th strategy:

[0117]

[0118] In the formula: p(x) is the probability of x occurring; b is the number of types of interval points, where Γ(·) is the gamma function, ψ(·) is the digamma function, and α is the fractional - order.

[0119] (4) Define the mean value of the fractional - order Shannon entropy calculated by the four different strategies as the fractional - order attention entropy, that is:

[0120]

[0121] 1.2. Composite Multiscale Fractional-Order Attention Entropy

[0122] FrATE has great advantages in analyzing the randomness and dynamic mutation behavior of time series at a single scale. To measure the complexity of time series at different scales, multiscale analysis methods have emerged. However, the multiscale calculation defined based on the coarse-graining process depends on the length of the time series and does not consider the relationship between the newly calculated coarse-grained sequences, resulting in information loss. To address the above problems, combined with the composite multiscale method, CMFrATE is proposed, and the calculation steps are as follows:

[0123] (1) Coarse-graining processing

[0124] For the original time series X = {x1, x2, …, x T}, its k-th coarse-grained sequence at the scale factor τ is given by the following formula:

[0125]

[0126] where: 1 ≤ k ≤ τ.

[0127] (2) Calculate CMFrATE

[0128] The CMFrATE at each scale factor τ is defined as:

[0129]

[0130] According to the CMFrATE algorithm, at the scale factor τ, the original sequence is continuously segmented into coarse-grained sequences of length τ with initial points [1, τ] respectively, then the average value of each small segment of the sequence is calculated, and then τ new coarse-grained sequences are obtained by arranging them in order. After calculating the CMFrATE values of the new sequences, the average of all entropy values is taken. This not only reduces the fluctuation of the entropy value with the increase of the scale factor, but also has better stability than the traditional coarse-grained multiscale process.

[0131] 2. Fractional-Order Deep Extreme Learning Machine

[0132] 2.1. Fractional-Order ELU (FracELU) Activation Function

[0133] FracELU is a fractional-order variant of ELU. It is defined using fractional calculus as:

[0134]

[0135] where a i ∈(0, 1) and b are fixed parameters. Then, the fractional derivative is calculated as follows:

[0136]

[0137] The FracELU activation function is an improved activation function. Compared with traditional activation functions, its core advantage lies in introducing fractional hyperparameters to adaptively balance the linear and nonlinear response characteristics, thereby enhancing the model's expressive power while optimizing the gradient dynamics. Specifically, this function maintains linear characteristics in the positive value region to avoid gradient saturation, and adopts a continuously differentiable mapping with a fractional exponent in the negative value region, which not only alleviates the neuron death problem of ReLU-like functions but also has a more flexible nonlinear adjustment ability compared to the ELU function. Its main advantages are as follows:

[0138] Enhanced nonlinear ability: By introducing a fractional exponential function, FracELU further enhances the nonlinear ability of the activation function. This enables the neural network to more effectively capture complex data patterns, thereby improving the model's expressive power and generalization performance.

[0139] Alleviating the neuron death problem. Compared with ReLU, FracELU introduces non-zero outputs in the negative input region, avoiding the common "dead neuron" problem in ReLU. This design ensures that there is still gradient flow in the negative input region of the network, thereby improving the network's stability and training efficiency.

[0140] Output mean close to zero: The output mean of FracELU is closer to zero, which helps to reduce the mean shift problem during network training. This characteristic can accelerate the convergence speed of the neural network and improve training efficiency.

[0141] Smooth function characteristics: The output of FracELU in the negative input region is continuous and smooth, which helps to reduce the risk of gradient vanishing and improve the model's stability and convergence speed.

[0142] Flexibility and adjustability: By introducing adjustable parameters (such as fractional parameters), FracELU provides higher flexibility for the model. These parameters can be adjusted according to different tasks and datasets to optimize the model performance.

[0143] Applicable to multiple tasks: FracELU performs well in various deep learning tasks, including image classification, object detection, time series prediction, etc. Its enhanced nonlinear ability and stability enable it to effectively handle complex features and dynamic changes.

[0144] 2.2. Fractional Extreme Learning Machine

[0145] When training a traditional feedforward neural network using the gradient descent algorithm, the parameters gradually approach the optimal solution with each update and computational iteration. Although this method has advantages, traditional feedforward neural networks have limitations such as slow learning speed and complex network topologies. Therefore, the main goal of ELM is to solve these problems. The fractional-order ELM method is an improved version of the traditional feedforward perceptron neural network, with advantages such as fast learning speed, excellent generalization ability, and the generation of a single optimal solution.

[0146] The fractional-order ELM model consists of an input layer, a hidden layer, and an output layer. This algorithm is used to train a set of S arbitrary samples (x i , y i ), where x i represents the input variable and y i represents the corresponding desired output. Equation (7) shows the mathematical structure of the fractional-order ELM for S different samples:

[0147]

[0148] where α j is the weight vector connecting the i-th hidden node and the output node, ω j represents the weight vector connecting the i-th hidden node and the input node, and ψ j represents the bias vector, represents the FrELU activation function. Here, m represents the total number of hidden neurons. If m ≤ S, then the S samples can be accurately approximated without error. Otherwise, y j is approximated as the training data target t i . Therefore, there exist certain values of α j , ω j , x j and ψ j such that for i = 1,..., S and j = 1,..., S, Equation (8) is satisfied:

[0149]

[0150] In Equation (7), t i represents the element of the training data target matrix T. ω j , ψ j and the order α i of the FrELU activation function are randomly generated numbers for the network and are all determined numbers. The weights of the hidden nodes and the output nodes are unknowns. The S equations in Equation (7) can be concisely expressed as:

[0151] H = αT(9)

[0152] In Equation (8),

[0153]

[0154] represents the output matrix generated by the hidden layer,

[0155]

[0156] represents the target matrix of the training data,

[0157]

[0158] Calculate the "Moore-Penrose inverse matrix" of H and record it as H * After calculating, ELM continues to learn the value of α using the following equation (9):

[0159] α=H * T. (13)

[0160] The output matrix Y is described as:

[0161] Y=Hα (14)

[0162] 2.3. Multi-layer fractional extreme learning machine

[0163] The fractional-order DELM algorithm constructs a deep network by combining hierarchical feature extraction with fast parameter analysis. Its working principle can be divided into three core stages:

[0164] (1) Hierarchical random feature mapping: A multi-layer cascade structure is adopted, and each layer is composed of a randomly initialized fractional-order ELM autoencoder. The input data is transformed nonlinearly layer by layer, and the present invention uses the fractional exponential linear unit (FrELU) to generate sparse feature representation in the latent space.

[0165] (2) Layered pseudo-inverse analysis: The optimal output weights are directly solved for each layer through the Moore-Penrose generalized inverse matrix. This process avoids the iterative calculation of back propagation and greatly reduces the time complexity of single-layer training.

[0166] If the matrix product H T If H is non-singular, the pseudo-inverse matrix (Moore-Penrose inverse matrix) of H is expressed as:

[0167] H * =(H T H) -1HT (15)

[0168] Similarly, if HH T is non-singular, then the pseudo-inverse matrix (Moore-Penrose inverse matrix) of H is expressed as:

[0169] H * =H T (HH T ) -1 (16)

[0170] By using the ridge regression theory, the diagonal elements of the matrix HH T need to contain the value where λ is the regularization parameter.

[0171] The single-layer learning equation of the fractional-order DELM is transformed into:

[0172]

[0173] The input to the k-th hidden layer is described as:

[0174] For n external inputs, and k ranging from 1 to J,

[0175] In this framework, the transformation matrix of the k-th hidden layer is represented as:

[0176]

[0177] where, is the transformation vector that controls the representation learning, and j ranges from 1 to m. In this way, the input is represented as:

[0178] X k =H k α k . (19)

[0179] Learn the value of α k by using Equation (15) or Equation (16):

[0180]

[0181] (3) Hierarchical cascade fine-tuning optimization: After completing the unsupervised layer-by-layer pre-training, improve the representation ability through global supervised fine-tuning. Adopt a new escape algorithm to minimize the loss function.

[0182] For the J-layer, we can describe the output matrix of the last layer as H J , and use this output matrix to learn the weight matrix α J of the last layer:

[0183]

[0184] That is, this equation is equivalent to solving the following least squares problem,

[0185]

[0186] Its training mechanism is as follows Figure 2 as shown

[0187] 3. Novel Escape Optimization Algorithm

[0188] The novel escape algorithm (Escape algorithm, ESC) is a meta - heuristic algorithm inspired by crowd evacuation behavior. This algorithm is inspired by crowd evacuation behavior and is used to solve real - world cases and benchmark problems. The novel ESC algorithm simulates the behavior of people during the evacuation process. During the exploration stage, the crowd is divided into three groups: calm, aggregated, and panicked, reflecting different levels of decision - making and emotional states. Calm individuals guide the crowd towards safety, aggregated individuals imitate others in less safe areas, and panicked individuals make unstable decisions in the most dangerous areas. These models allow the novel ESC algorithm to simulate various adaptive strategies used by people in life - threatening situations and transform them into a computational framework that can more effectively solve complex optimization problems.

[0189] (1) Algorithm and Population

[0190] The novel ESC algorithm aims to simulate the behavior of people during an emergency evacuation, where individuals must move towards an exit in a dynamic and uncertain environment. The novel ESC algorithm introduces the concept of an elite pool, which represents the best - performing individuals and symbolizes the potential exits identified by the crowd. This mechanism enhances the algorithm's ability to explore the solution space deeply, considering multiple directions and avoiding local optima.

[0191] The novel ESC algorithm first initializes a population, and each population is described by an expression of a D - dimensional vector: each is described by a D - dimensional vector \(x\) i =(x i,1 ,x i,2 …,x i,D ). The value of the \(i\) - th individual in the \(j\) - th dimension is given by:

[0192] x i,j =lb j +r i,j ×(ub j - lb j ), r i,j ~U(0,1) (23)

[0193] where lb j and ub j represent the lower and upper bounds of the \(j\) - th dimension respectively, ensuring that the initial positions of each individual are randomly distributed within the feasible space. The random variable \(r\) i,j is uniformly distributed between 0 and 1, reflecting the randomness of the initial decision - making process during the evacuation.

[0194] After initializing the population, the fitness function \(f\) is usedi = f(x i ) evaluates the fitness of each individual. Then, the population is sorted in ascending order according to the fitness, and the top individuals are stored in the elite pool E, which represents the number of potential safe exits found by the population:

[0195] E = {x (1) , x (2) ,..., x (exist)} (24)

[0196] (2) Panic Index and Iteration Process

[0197] The new ESC algorithm models the iteration process to reflect the evolution of crowd behavior during evacuation. According to different behavioral responses during evacuation, the actions of individuals are classified into three categories: calm, aggregation, or panic.

[0198] At the beginning of each iteration t, the panic index P(t) is calculated as follows:

[0199]

[0200] More chaotic behavior. As t ranges from 0 to the number of iterations T, this index decreases over time, simulating the adaptation of the crowd to the evacuation environment.

[0201] (3) Exploration

[0202] In the exploration stage, when t ≤ T / 2 (T is the maximum number of iterations of the algorithm, and t is the current number of iterations), the population is divided into a calm group, an aggregation group, and a panic group according to the fitness level of the population. Specifically, the population is sorted in ascending order of fitness, and the individuals are divided into three groups in proportion: the calm group c = 0.15, the aggregation group h = 0.35, and the panic group p = 0.5. This stratification reflects the different responses of individuals in the crowd during evacuation, where some people remain calm, some follow group behavior, and some panic. To extend the scope of exploration and improve exploration efficiency, the present invention adds a Gaussian random walk strategy to the mathematical expressions of the three groups.

[0203] (4) Update of the Calm Group

[0204] Individuals in the calm group behave rationally and move towards the central position C j , which represents the collective decision of the group as:

[0205]

[0206]

[0207] where C jis the center of the calm group in the j - dimension, calculated as the average of all calm individuals in this dimension. is to generate a Gaussian distribution random number with the current optimal position and the constant σ as the mean and variance. The vector v c,j is defined by the following formula:

[0208]

[0209] where,

[0210]

[0211] is a position randomly generated within the boundary of the calm group. and represent the minimum and maximum values of all individuals in the calm group in the j - dimension respectively, represents a small adjustment of the individual's movement. The binary variable m1 is determined by the Bernoulli distribution, allowing partial updates, which simulates the part of the dimension that is not updated due to crowding. Specifically, when generated, it takes the value 0 or 1 with the same probability. w1 is an adaptive Levy weight, using the Levy distribution to simulate the step size in the exploration stage.

[0212] (5) Aggregation group update

[0213] Aggregation group individuals follow the behaviors of calm and panic groups. Their positions are updated according to the following two aspects of influence:

[0214]

[0215] In this equation, x p,j is an individual randomly selected from the panic group, representing the potential direction of panic - driven movement. w2 is another adaptive Levy weight for the corresponding group. The vector v h,j is defined by the following formula:

[0216]

[0217] where,

[0218]

[0219] is a position randomly generated within the boundary of the driving group. and represent the minimum and maximum values of all individuals in the driving group in the j - dimension respectively. m2 is a binary variable, generated by the same mechanism as m1.

[0220] (6) Panic group update

[0221] Affected by potential exits and other randomly indicated individuals, panic-driven individuals explore the solution space more irregularly:

[0222]

[0223] Among them, E j is an individual randomly selected from the elite pool, representing a possible exit that a panic-driven individual might move towards. x rand,j represents an individual randomly selected from the population, introducing an element of randomness into the panic-driven movement. The vector v h,j is defined by

[0224]

[0225] where,

[0226]

[0227] is a position randomly generated within the boundaries of the panic group. and represent the minimum and maximum values of all panic group individuals in the j-th dimension respectively.

[0228] (7) Development stage

[0229] When the iteration exceeds T / 2, the algorithm transitions to the development stage, where all individuals are considered calm. The focus shifts to fine-tuning the positions based on the best solutions determined so far. In this stage, individuals optimize their positions by approaching the members of the elite pool, which represents the possible safe exits and the best solutions determined in previous iterations, as well as individuals randomly selected from the group. This process simulates the crowd gradually converging towards the determined best exits. A triangular walk strategy will be introduced in the present invention. The triangular walk strategy uses a triangular path to walk around the optimal position, which can increase randomness, avoid the algorithm falling into local optimal solutions, and at the same time have a greater chance of exploring the search space, thus finding better solutions. The position update in this stage is given by:

[0230]

[0231] where,

[0232]

[0233] L2 = λ1·L1 (38)

[0234] Θ = L1 2 +L2 2 -2·L1·L2·cos(2πλ2) (39)

[0235] In the formula, x i,j Represents the position of the i-th individual in the j-dimension. j is the position of the elite pool member, symbolizing a possible safe exit and one of the best solutions identified so far. rand,j is the position of an individual randomly selected from the population. This allows individuals to optimize their positions by moving closer to elite pool members and randomly selected individuals, simulating the population gradually converging to a determined optimal exit. λ1,λ2,λ3 are random numbers in the interval [0, 1].

[0236] 4. Intelligent safety helmet operation fatigue risk warning classification algorithm based on fractional-order deep extreme learning machine

[0237] The present invention uses composite multi-scale fractional-order attention entropy to perform comprehensive feature extraction on various time series data collected by sensors of smart helmet wearable devices for power workers, and then inputs the obtained feature data into a fractional-order deep extreme learning machine for fatigue warning multi-classification processing. The optimizer of this network is a new escape optimization algorithm.

[0238] The four feature data extracted by the composite multi-scale fractional-order attention entropy method, namely EEG, body temperature, heart rate and blood oxygen saturation, and three fatigue warning category labels, namely normal, fatigue and high-risk, should be input into the fractional-order deep extreme learning machine for training. The fractional-order DELM algorithm constructs a deep network by combining hierarchical feature extraction with fast parameter analysis. Its working principle can be divided into three core stages:

[0239] (1) Hierarchical random feature mapping: A multi-layer cascade structure is adopted, and each layer is composed of a randomly initialized fractional-order ELM autoencoder. The input data is transformed nonlinearly layer by layer, and the present invention uses the fractional exponential linear unit (FrELU) to generate sparse feature representation in the latent space.

[0240] (2) Layered pseudo-inverse analysis: The optimal output weights are directly solved for each layer through the Moore-Penrose generalized inverse matrix. This process avoids the iterative calculation of back propagation and greatly reduces the time complexity of single-layer training.

[0241] (3) Layer-by-layer fine-tuning optimization: After completing unsupervised layer-by-layer pre-training, global supervised fine-tuning is used to improve the representation ability. A new escape optimization algorithm is used to minimize the loss function. Its network structure diagram Figure 3 shown.

[0242] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalent technologies, the present invention is also intended to include these modifications and variations.

Claims

1. An intelligent safety helmet operation fatigue risk warning classification algorithm based on a fractional-order deep extreme learning machine, characterized in that The classification algorithm is executed through the following method, which includes: S1. Real-time acquisition of raw data through multiple sensors integrated on the intelligent safety helmet, where the raw data includes personnel physiological data and environmental data; S2. Preprocess the collected raw data, perform fractional-order differentiation or integration on the data using the fractional calculus theory, and extract deeper information features; S3. Construct a fractional-order deep extreme learning machine model, combine the fractional calculus theory with the deep extreme learning machine, and use the fractional calculus theory to enhance the model's processing ability for complex non-linear data; S4. Train the fractional-order deep extreme learning machine model with a large amount of historical data, and optimize the model parameters through cross-validation and grid search; S5. Real-time input the newly collected data into the trained model to output the fatigue risk level; according to the risk level, the intelligent safety helmet can send warning signals to the operators through multiple means, and at the same time upload the warning information to the cloud management platform.

2. The intelligent safety helmet operation fatigue risk warning classification algorithm based on the fractional-order deep extreme learning machine according to claim 1, characterized in that The above S1 includes: S11. Acquire raw data through multiple sensors, and the sensors include: a heart rate sensor, an electroencephalogram sensor, an acceleration sensor, and an environmental sensor; S12. Synchronize the timestamps of the raw data collected by all sensors through the clock system of the central processing unit; S13. Use the sliding window filtering algorithm to preliminarily eliminate obvious outliers and record the time points when abnormal data appears; S14. Use the median filtering or Kalman filtering algorithm to smooth the heart rate data, and use the independent component analysis or principal component analysis algorithm to remove the noise components in the electroencephalogram signal.

3. The intelligent safety helmet operation fatigue risk warning classification algorithm based on the fractional-order deep extreme learning machine according to claim 1, wherein The above S2 includes: S21. Perform fractional-order differentiation on the heart rate variability data, extract the subtle features of heart rate changes, and use the Grünwald-Letnikov fractional-order differentiation formula to determine the optimal order through simulation tests; S22. Analyze the influence of the fractional-order differentiation order on the feature extraction effect, use the cross-validation method to compare the model performance under different orders, and determine the optimal order; S23. Perform fractional-order integration on the electroencephalogram signal to smooth the signal, and use the Riemann-Liouville fractional-order integration formula to determine the optimal order through simulation tests; S24. Analyze the influence of the fractional-order integration order on the electroencephalogram feature extraction, use the grid search method to compare the model performance under different orders, and determine the best parameters; S25. Calculate the heart rate variability index and evaluate the cardiac autonomic regulation function based on the calculation results; S26. Use the fast Fourier transform to analyze the power spectral density of each frequency band, and combine the fractional-order integration processing results to extract the power spectral density features of the α, β, θ, and δ frequency bands; S27. Statistically analyze the frequency and amplitude of the head movement state, combine the acceleration data to construct a head movement feature vector; and use a method combining time-domain analysis and frequency-domain analysis to extract features.

4. The intelligent safety helmet operation fatigue risk early warning classification algorithm based on the fractional-order deep extreme learning machine according to claim 3, wherein The above S27 includes: S271. Preprocess the head movement data collected by the acceleration sensor, and extract the features of the head movement based on the preprocessed data; S272. Adopt a method combining time-domain analysis and frequency-domain analysis to comprehensively describe the state of head movement; S273. Construct a head movement state classification model through a machine learning algorithm; use the extracted features as input and the head movement state as output for model training; S274. Optimize the classification model using cross-validation, and screen features that have a significant impact on the classification results through feature importance evaluation; S275. Based on the results of the classification model, construct a head movement feature vector, fuse the head movement feature vector with other physiological features, and use feature splicing to integrate feature information from different sources to form a feature set.

5. The intelligent safety helmet operation fatigue risk early warning classification algorithm based on the fractional-order deep extreme learning machine according to claim 4, characterized in that, The above S275 includes: On the basis of the original head movement state probability distribution, movement frequency, and movement amplitude, further extract the dynamic features of head movement; Use time series analysis technology to model the head movement data and extract time series features; Splice the head movement feature vector with other physiological features to form a preliminarily fused feature set; Use principal component analysis to perform dimensionality reduction on the preliminarily fused feature set, and further screen features that have a significant contribution to the target task through feature importance evaluation; Use the constructed feature set to train a machine learning model, verify the model through cross-validation, evaluate its performance on the target task, and iteratively optimize the feature set according to the verification results.

6. The intelligent safety helmet operation fatigue risk warning classification algorithm based on the fractional-order deep extreme learning machine according to claim 1, characterized in that The above S3 includes: S31. Construct a fractional-order deep extreme learning machine model, combine the fractional-order calculus theory with the deep neural network structure to enhance the model's processing ability for complex non-linear data; S32. The input layer receives the preprocessed feature vector, the hidden layer adopts a multi-layer deep neural network structure, and the output layer uses the softmax function to output the probability distribution of the fatigue risk level; S33. During the forward propagation process of the neural network, introduce fractional-order calculus operations, adjust the weight update rule, and adopt an adaptive learning rate adjustment strategy to improve the training efficiency; S34. Based on the fusion strategy of fractional-order calculus in the neural network, through the fractional-order convolution kernel, integrate fractional-order differential or integral operations into the convolution operation to extract deeper features; S35. Through the fractional-order pooling strategy, perform fractional-order integration or differentiation on the feature map, use historical data to train the model, adjust the model parameters through the backpropagation algorithm; and adopt the momentum optimization algorithm to accelerate convergence; S36. Implement cross-validation to evaluate the generalization ability of the model, and use the grid search technology to optimize the model parameters. At the same time, adopt the early stopping method strategy to prevent the model from overfitting on the training set; S37. Use multiple indicators to evaluate the model performance; analyze the prediction accuracy of the model at different fatigue risk levels, and evaluate the robustness and generalization ability of the model; S38. According to the evaluation results, optimize the model, adopt the ensemble learning method, combine multiple fractional-order deep extreme learning machine models to improve the overall prediction performance.

7. The intelligent safety helmet operation fatigue risk warning classification algorithm based on the fractional-order deep extreme learning machine according to claim 1, characterized in that, The above S4 includes: S41. Prepare an independent validation data set and ensure that the validation data set is consistent with the training data set in distribution; S42. Comprehensively evaluate the model performance using evaluation metrics; compare the prediction accuracy of the model at different thresholds to determine the optimal threshold; S43. Analyze the differences in the prediction performance of the model under different operation types, environmental conditions, and operator constitutions, and evaluate the adaptability and robustness of the model; S44. Adopt a feature importance evaluation method to analyze the impact of different features on the model prediction performance; optimize the feature selection strategy according to the feature importance ranking.

8. The intelligent safety helmet operation fatigue risk warning classification algorithm based on the fractional-order deep extreme learning machine according to claim 1, characterized in that The said S5 includes: S51. Input the newly collected data into the trained model in real time to calculate the fatigue risk level; S52. Issue warnings to the operators through a multi-means warning mechanism according to the risk level; S53. Upload the warning information to the cloud management platform for remote monitoring and analysis, and combine big data and artificial intelligence algorithms to further explore the data value; S54. Adopt cluster analysis and association rule mining to discover the potential rules and influencing factors of the operators' fatigue states, and analyze the changing trends of the operators' fatigue states through warning reports and data analysis reports, and optimize the operation arrangements and preventive measures.

9. The intelligent safety helmet operation fatigue risk early warning classification algorithm based on the fractional-order deep extreme learning machine according to claim 8, characterized in that, The said S51 includes: S511. After collecting data in real time, perform data cleaning and data verification; format the cleaned data according to the model input requirements and perform standardization processing on the data; S512. Based on the data stream processing technology, perform real-time data collection, processing, and input on the data through multi-threading; S513. Before starting the real-time warning system, load the trained fatigue risk prediction model and perform initialization operations; S514. Input the preprocessed data into the model for inference in real time, calculate the fatigue risk level, and output the inference result in real time; S515. Set fatigue risk warning thresholds at different levels according to business requirements and safety standards; and adopt an adaptive learning algorithm or machine learning technology to dynamically adjust the warning thresholds; trigger the corresponding multi-means warning mechanism according to the calculated fatigue risk level and the set warning thresholds.

10. The intelligent safety helmet operation fatigue risk warning classification algorithm based on the fractional-order deep extreme learning machine according to claim 9, wherein, The said S512 includes: Based on the built-in data stream processing framework, design the data stream processing process; divide each link in the data stream processing process into multiple independent tasks and allocate an independent thread to each task; Adopt asynchronous processing technology to perform non-blocking processing of the data through asynchronous calls and callback mechanisms; and perform thread reuse and dynamic adjustment through a thread pool; During the data stream processing, introduce a data caching mechanism to temporarily store the real-time collected data through caching and form data batches before processing; Real-time collect the metrics of the system through monitoring tools and perform optimization operations according to the monitoring results.

Images