Intelligent safety helmet operation fatigue risk early warning classification method based on fractional order long short-term memory network

Abstract

The invention provides an intelligent safety helmet operation fatigue risk early warning classification method based on a fractional order long and short term memory network, and relates to the technical field of risk early warning, and the method comprises the steps: collecting fatigue sign data, obtaining fatigue sign collection data, and classifying the fatigue sign collection data to obtain fatigue category data; performing feature extraction on the fatigue category data to obtain fatigue feature extraction data; obtaining three fatigue early warning category labels, respectively inputting the fatigue feature extraction data and the three fatigue early warning category labels into a fractional order long and short term memory network for training, and performing parameter optimization processing on the training process through a novel Genethink shark optimization algorithm to obtain network optimization data; and inputting the network optimization data into a Softmax function, outputting multi-class probability data, and carrying out fatigue early warning, thereby effectively improving the safety and reliability of power production.

Description

Technical Field

[0001] The present invention proposes an intelligent safety helmet operation fatigue risk warning classification method based on a fractional-order long short-term memory network, which relates to the technical field of risk warning, specifically to the technical field of operation fatigue risk warning. Background Art

[0002] Electric power production is a systematic project with a wide range of involvements, numerous on-site personnel, and complex projects. At the electric power production site, multi-professional and different types of staff such as project managers, supervisors, technicians, and engineers need to closely coordinate and cooperate to process a large amount of data information. Once a mistake occurs in electric power production management, it will not only endanger the safety of electric power production, but may also lead to casualties in severe cases. The traditional electric power inspection process is complex, highly dependent on the work experience and professional level of on-site inspectors, and it is easy to cause untimely problem discovery due to human errors, even leading to major accidents, and at the same time, it cannot guarantee the safety of inspection operation personnel.

[0003] The intelligent helmet sensors worn by electric power operation personnel usually collect time-series data such as electroencephalogram, body temperature, heart rate, and blood oxygen saturation. When extracting features from these time-series data, various algorithms and technologies are generally used. These algorithms can be roughly divided into time-domain feature extraction and frequency-domain feature extraction methods. However, time-domain and frequency-domain feature extraction methods each have their own unique advantages and limitations. In particular, the feature extraction of electroencephalogram data, as a key link in electroencephalogram signal processing, is of great importance. It aims to extract useful information from complex electroencephalogram signals for subsequent analysis and classification. Electroencephalogram data is a typical non-stationary and strongly non-linear data, which makes its feature extraction process more challenging.

[0004] The disadvantages of traditional time-domain feature extraction methods and frequency-domain feature extraction methods are as follows:

[0005] Time-domain features (such as mean, variance, peak value, waveform factor, etc.) directly depend on the amplitude of the original signal and are easily affected by noise interference. For example, high-frequency noise will significantly affect features such as peak value and pulse factor.

[0006] Time-domain features only reflect the change of the signal amplitude over time and cannot characterize the frequency components of the signal. For example, the periodic fault features in a vibration signal may be masked in the time domain but are easier to detect in the frequency domain.

[0007] Signals of different categories may have similar statistical characteristics (such as mean and variance) in the time domain, resulting in insufficient feature discrimination and affecting the classification or diagnosis effect.

[0008] Most time-domain features (such as the autocorrelation function) assume that the signal is stationary, while actual signals (such as mechanical vibrations, speech signals) are often non-stationary, resulting in feature failure.

[0009] Loss of time information: Frequency-domain features (such as the spectrum after Fourier transform, power spectral density, etc.) globally decompose the signal into frequency components but cannot locate the position of specific frequency components on the time axis. For example, it is impossible to determine the specific time when a transient shock event (such as the pulse of a bearing fault) occurs.

[0010] The Fourier transform requires the signal to be stationary, but actual signals (such as speech, vibration, electroencephalogram signals) are usually non-stationary. The short-time Fourier transform (STFT) can alleviate this problem to some extent, but it is limited by the fixed time-frequency resolution.

[0011] The resolution of frequency-domain analysis is limited by the signal length (Heisenberg uncertainty principle). For example, a long signal can improve the frequency resolution but reduce the time resolution, and vice versa for a short signal.

[0012] Most frequency-domain features (such as the power spectrum) only utilize the amplitude information and ignore the phase information, which may result in the loss of key signal features (such as the time synchronization of the signal).

[0013] Insensitive to non-linear signals: Frequency-domain analysis is based on linear system theory and is difficult to characterize the characteristics of non-linear systems (such as harmonic distortion, chaotic phenomena).

[0014] Traditional features need to be manually selected or designed (such as which statistics or frequency bands to choose), which requires high domain knowledge and may miss important features.

[0015] Time-domain and frequency-domain features usually need to be combined, resulting in a high feature dimension and possible redundancy (such as the time-domain variance being related to the frequency-domain energy), which requires dimensionality reduction processing but may lose information.

[0016] For multi-component, non-stationary, and non-linear signals, traditional methods are difficult to effectively characterize their complex characteristics.

[0017] Manually designed features may perform inconsistently in different application scenarios and need to be re-adjusted for specific tasks, lacking universality.

[0018] The essence of the intelligent safety helmet operation fatigue risk warning classification algorithm is a multi-classification problem. At present, two main technical means are used in this field to achieve it:

[0019] A common method is to use the weighted average method. By adjusting the weights of the data from each sensor and performing fusion processing, a comprehensive assessment of the health status of the operator is formed. However, this method has two main drawbacks: First, in the actual scenario, the relationship between various features is not always linear, which leads to the inconsistency between the comprehensive assessment model and the real rules, thus seriously affecting the accuracy of the assessment results. Second, the weights belong to hyperparameters and need to be selected manually, and manual selection often has a large error. Although the weights can be solved by hyperparameter optimization methods, this not only takes time and effort but also may affect the timeliness of early warning, which is not worth the candle.

[0020] Another approach is to use deep learning methods to achieve multi-classification. However, in the engineering scenario, the physical device resources are limited and the team lacks professional knowledge related to deep learning algorithms. Using deep learning technology may face the following drawbacks and disadvantages:

[0021] High training cost: Deep learning models (especially large models) require a large amount of computing resources (GPU / TPU). In the case of limited devices, the training time may be too long or even unable to be completed.

[0022] Low inference efficiency: When the hardware performance is insufficient during deployment (such as only having a CPU), the model inference speed may not meet the real-time requirements, affecting the engineering practicality.

[0023] Energy consumption problem: High-load computing may pose higher requirements on infrastructure such as power and heat dissipation.

[0024] Lack of understanding of the algorithm: If the team lacks an understanding of model principles, hyperparameter tuning, regularization methods, etc., it may lead to poor model performance or training failure.

[0025] Difficult to debug: Problems in model training (such as gradient vanishing / exploding, overfitting) require professional knowledge for diagnosis, and an inexperienced team may not be able to solve them effectively.

[0026] Dependence on external support: It may be necessary to hire experts or outsource development, increasing project costs and management complexity.

[0027] Difficult to obtain data: Deep learning depends on a large amount of labeled data. In the engineering scenario, data may be insufficient due to environmental restrictions (such as difficulty in collecting industrial equipment data).

[0028] High cost of data annotation: When there is a lack of automated annotation tools, manual annotation is time-consuming and laborious and requires domain knowledge support.

[0029] Poor data generalization: If the training data is inconsistent with the real scenario distribution (such as equipment aging, environmental changes), the model may perform poorly.

[0030] Insufficient robustness: The model may be sensitive to noise and abnormal inputs, leading to unexpected failures in engineering scenarios.

[0031] Difficult to maintain: The model needs to be continuously monitored and updated. However, if the team lacks experience, it is difficult to handle the problem of model performance degradation. Summary of the Invention

[0032] The present invention provides an intelligent safety helmet operation fatigue risk warning classification method based on a fractional-order long short-term memory network to solve the above problems:

[0033] The intelligent safety helmet operation fatigue risk warning classification method based on a fractional-order long short-term memory network proposed by the present invention, the classification method includes:

[0034] S1. Obtain an intelligent safety helmet wearing device, collect fatigue sign data of power workers, obtain fatigue sign collection data, and classify the fatigue sign collection data according to a preset fatigue category to obtain fatigue category data;

[0035] S2. Extract features from the fatigue category data through a fine composite multi-scale fractional-order reverse dispersion entropy to obtain fatigue feature extraction data;

[0036] S3. Obtain three fatigue warning category labels, input the fatigue feature extraction data combined with the three fatigue warning category labels into a fractional-order long short-term memory network for training respectively, and perform parameter optimization processing on the training process through a new type of Genghis shark optimization algorithm to obtain network optimization data;

[0037] S4. Input the network optimization data into a Softmax function, and then output multi-category probability data for fatigue warning.

[0038] Further, the S1 includes:

[0039] Collect fatigue sign data of power workers through an intelligent safety helmet wearing device, preprocess the collected fatigue sign data to obtain fatigue sign collection data;

[0040] Obtain a preset fatigue category, and the preset fatigue category includes electroencephalogram, body temperature, heart rate, and blood oxygen saturation;

[0041] Classify the fatigue sign collection data according to the preset fatigue category to obtain fatigue category data;

[0042] The fatigue category data includes electroencephalogram data, body temperature data, heart rate data, and blood sample saturation data.

[0043] Further, the S2 includes:

[0044] Obtain a time series of a preset length, and calculate the fractional - order RDE by means of the normal distribution function, linear transformation, phase - space reconstruction, calculation of the diffusion pattern and its probability.

[0045] After performing phase - granulation processing on the time series, calculate the RCMFrRDE to obtain calculation data.

[0046] Furthermore, the step S3 includes:

[0047] Obtain the fractional - order Softplus activation function and the fractional - order tanh activation function;

[0048] Set the fractional - order Softplus activation function and the fractional - order tanh activation function for the long short - term memory network to obtain a long short - term memory network with the fractional - order Softplus activation function and the fractional - order tanh activation function.

[0049] Among them, the setting of the fractional - order Softplus activation function and the fractional - order tanh activation function for the long short - term memory network includes:

[0050] Initialize the fractional - order LSTM cell state C and the hidden state h with zero or small random values;

[0051] Apply the fractional - order Softplus activation function to the weighted sum of the current input x t and the previous hidden state h t-1 to calculate the input - gate activation i t ;

[0052] Apply the fractional - order Softplus activation function to the weighted sum of the current input x t and the previous hidden state h t-1 to determine the forget - gate activation f t ;

[0053] Apply the fractional - order tanh activation function to the weighted sum of the current input x t and the previous hidden state h t-1 to obtain the candidate cell state

[0054] Combine the candidate cell state the previous cell state C t-i and the forget - gate f t to update the cell state C t-i ;

[0055] Apply the fractional - order Softplus activation function to the weighted sum of the current input x t and the previous hidden state h t-1 to determine the output - gate activation O t ;

[0056] Apply the fractional tanh activation function to the updated cell state C t Multiply by the output gate O t to obtain the new hidden state h t ; Store the information to be sent to the next time step and used as the final output in this hidden state h t ;

[0057] The hidden state h t is used as the output of the fractional LSTM cell for the current time step;

[0058] Repeat the operation for each time step in the sequence.

[0059] Furthermore, the S3 further includes:

[0060] In the novel Genghis Khan shark optimization algorithm, calculate a random position through the upper and lower bounds of the search space to obtain the best predation position;

[0061] Perform a comprehensive data search in the target area through the spiral search and wandering strategy, and add a crazy operator for depth calculation to obtain the depth calculation result.

[0062] Furthermore, the classification system includes:

[0063] A data acquisition module for obtaining an intelligent safety helmet wearing device, collecting fatigue sign data of power workers to obtain fatigue sign collection data, and classifying the fatigue sign collection data according to preset fatigue categories to obtain fatigue category data;

[0064] A feature extraction module for extracting features from the fatigue category data through the fine composite multi-scale fractional reverse scatter entropy to obtain fatigue feature extraction data;

[0065] A parameter optimization module for obtaining three fatigue warning category labels, inputting the fatigue feature extraction data combined with the three fatigue warning category labels into a fractional long short-term memory network for training, and performing parameter optimization processing on the training process through the novel Genghis Khan shark optimization algorithm to obtain network optimization data;

[0066] A fatigue warning module for inputting the network optimization data into the Softmax function, and then outputting multi-category probability data for fatigue warning.

[0067] Furthermore, the data acquisition module includes:

[0068] An acquisition module, configured to collect fatigue sign data of power workers through an intelligent safety helmet wearing device, preprocess the collected fatigue sign data, and obtain fatigue sign acquisition data;

[0069] Obtain preset fatigue categories, where the preset fatigue categories include electroencephalogram, body temperature, heart rate, and blood oxygen saturation;

[0070] A classification module, configured to classify the fatigue sign acquisition data through the preset fatigue categories to obtain fatigue category data;

[0071] The fatigue category data includes electroencephalogram data, body temperature data, heart rate data, and blood sample saturation data.

[0072] Further, the feature extraction module includes:

[0073] An RDE calculation module, configured to obtain a time series of a preset length, calculate the fractional order RDE by means of a normal distribution function, linear transformation, phase space reconstruction, calculation of the walking pattern and its probability;

[0074] A phase granulation processing module, configured to perform phase granulation processing on the time series and then calculate RCMFrRDE to obtain calculation data.

[0075] Further, the parameter optimization module includes:

[0076] A function acquisition module, configured to acquire a fractional order Softplus activation function and a fractional order tanh activation function;

[0077] A network processing module, configured to set a fractional order Softplus activation function and a fractional order tanh activation function for a long short-term memory network to obtain a long short-term memory network with a fractional order Softplus activation function and a fractional order tanh activation function;

[0078] Wherein, the network processing module includes:

[0079] Initialize the fractional order LSTM cell state C and the hidden state h with zero or small random values;

[0080] Apply the fractional order Softplus activation function to the weighted sum of the current input x t and the previous hidden state h t-1 to calculate the input gate activation i t ;

[0081] Apply the fractional order Softplus activation function to the weighted sum of the current input x t and the previous hidden state h t-1 to determine the forget gate activation f t ;

[0082] Applies the fractional tanh activation function to the current input x t and the previous hidden state h t-1 The weighted sum of the candidate's cell state is obtained

[0083] Set the candidate cell state Previous cell state C t-i and forget gate f t Combine to update the unit state C t-i ;

[0084] Apply the fractional softplus activation function to the current input x t and the previous hidden state h t-1 The weighted sum of determines the output gate activation O t ;

[0085] Apply the fractional tanh activation function to the updated cell state C t Multiply by output gate O t , get the new hidden state h t ; The information sent to the next time step and used as the final output is stored with this hidden state h t middle;

[0086] Hidden state h t Output of the fractional-order LSTM unit for the current time step;

[0087] Repeat for each time step in the sequence.

[0088] Furthermore, the parameter optimization module further includes:

[0089] The position acquisition module is used to calculate a random position by the upper and lower bounds of the search space in the new Genghis Khan shark optimization algorithm to obtain the best predation position;

[0090] The depth calculation module is used to perform comprehensive data search in the target area through a spiral search and wandering strategy, add a crazy operator to perform depth calculation, and obtain the depth calculation results.

[0091] The present invention utilizes sensors in real-time fatigue monitoring helmets to collect fatigue data, including EEG, body temperature, heart rate, and blood oxygen saturation, and then feeds this data into an intelligent classification method. By providing real-time feedback of classification labels (i.e., status levels), fatigue warnings are provided to workers, effectively improving the safety and reliability of power generation. BRIEF DESCRIPTION OF THE DRAWINGS

[0092] Figure 1It is a flow chart of an intelligent safety helmet operation fatigue risk warning classification algorithm based on a fractional-order long short-term memory network;

[0093] Figure 2 It is a schematic diagram of the fractional-order LSTM neuron structure;

[0094] Figure 3 It is a structural diagram of the fatigue warning algorithm for the intelligent safety helmet wearing device of electric power operation personnel. Specific implementation manners

[0095] The preferred embodiments of the present invention will be described below 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.

[0096] In one embodiment of the present invention, an intelligent safety helmet operation fatigue risk warning classification method based on a fractional-order long short-term memory network is proposed. The classification method includes:

[0097] S1. Obtain an intelligent safety helmet wearing device, collect fatigue sign data of electric power workers to obtain fatigue sign collection data, and classify the fatigue sign collection data according to a preset fatigue category to obtain fatigue category data;

[0098] S2. Extract features from the fatigue category data through a fine composite multi-scale fractional-order reverse dispersion entropy to obtain fatigue feature extraction data;

[0099] S3. Obtain three fatigue warning category labels, input the fatigue feature extraction data combined with the three fatigue warning category labels into a fractional-order long short-term memory network for training respectively, and perform parameter optimization processing on the training process through a new type of Genghis Khan shark optimization algorithm to obtain network optimization data;

[0100] S4. Input the network optimization data into a Softmax function, and then output multi-category probability data for fatigue warning.

[0101] The working principle of the above technical solution is as follows: Safety production is the most important prerequisite for power operations. The working environment of the production department of the State Grid is complex, with high-altitude operations, proximity to high-voltage equipment, etc., and the safety risks are relatively large. Safety accidents may cause serious injuries or even life-threatening. Therefore, in power operations, it is particularly important to monitor the physical health status of operators. Due to environmental and topographical limitations at the operation site, communication signals are restricted, accurate positioning is impossible, and the medical emergency response is slow. Improving the safety of employees working in dangerous environments has become an urgent need in the power industry. During the process of power operation by operators, detailed information about their current status can be obtained through intelligent safety helmet wearing devices, and rapid analysis and early warning of their own physical signs can be carried out in advance, so as to facilitate the operation tasks of the staff and ensure the reliability and efficiency of safety and dispatching command.

[0102] The main idea of the present invention is to use the fine composite multi-scale fractional-order reverse dispersion entropy to comprehensively extract the features of on-site data, then input the obtained features into the fractional-order long short-term memory network for multi-classification processing of health early warning, and finally use the new Genghis shark optimization algorithm to optimize the network parameters, as Figure 1 shown.

[0103] Feature extraction is the core technical link in the data processing of intelligent safety helmets. Its core task is to screen effective information from a large amount of original monitoring data and provide a reliable basis for subsequent health status classification. The current mainstream feature extraction methods can be divided into the following three dimensions:

[0104] The time-domain analysis method based on statistical principles intuitively reflects the distribution law and dynamic characteristics of signal amplitude by calculating statistics such as mean, variance, skewness, kurtosis, and interquartile range. This method has the advantages of high calculation efficiency and strong real-time performance, and is particularly suitable for industrial scenarios that require rapid response. However, due to the inherent characteristics of time-domain analysis, it is difficult to effectively capture the hidden features in the signal.

[0105] Using the Fourier transform and its derivative algorithms (such as the short-time Fourier transform) to convert the time-domain signal into a frequency-domain representation can clearly show the frequency composition and energy distribution characteristics of the signal. However, while emphasizing the frequency-domain characteristics, these methods have the defect of losing time-domain information and have limited analytical ability for non-stationary signals.

[0106] Aiming at the limitations of traditional methods in fatigue state recognition, the analysis method based on entropy theory shows unique advantages. This method can directly quantify the complexity of the monitoring signal without data decomposition, and effectively distinguish the system response differences corresponding to different fatigue levels by calculating characteristic parameters such as approximate entropy and sample entropy.

[0107] Job fatigue risk warning, as the core function of the intelligent safety helmet system, essentially belongs to a typical multi-classification problem. The current mainstream solutions in the industry have significant technical limitations, which are specifically manifested as follows:

[0108] The current weighted average method performs data fusion by manually setting feature weights. Its technical defects are mainly reflected in: the linear hypothesis fails, there are complex non-linear associations between features in the actual scenario, and the linear weighted model is difficult to accurately map the real decision boundary. The parameter optimization paradox, where weights, as hyperparameters, need to be determined by time-consuming methods such as grid search, fundamentally conflicts with the real-time warning requirement. The lack of dynamic adaptation, as the fixed weight system cannot respond to changes in the working environment, resulting in a continuous decline in the model's robustness.

[0109] Although deep neural networks have powerful representation capabilities at the theoretical level, they face multiple implementation obstacles in industrial scenarios: computational resource constraints, requiring GPU cluster support during the training phase. During the inference phase, the inference latency of CPU devices is greater than 500ms, unable to meet the sub-second response standard. The energy consumption cost is relatively high, and the device's battery life is weak. There are serious data sparsity problems in data collection. Annotation cost: single-sample annotation is time-consuming, and the participation rate of domain experts is insufficient. There is a distribution shift in the feature matching between the measured data and the training set. The model interpretability is poor, and the black-box nature of the decision-making process leads to a decrease in the safety verification passing rate. The maintenance complexity is relatively high, with a strong dependence on talents. The algorithm team needs to possess knowledge in both signal processing and deep learning fields simultaneously.

[0110] The defects of existing technologies: Traditional time-frequency domain analysis methods (such as Fourier transform, wavelet analysis) have significant deficiencies in capturing the deep hidden features of data. Their linear processing paradigm is difficult to effectively represent non-linear features under complex working conditions. Although the entropy value theory (sample entropy, fuzzy entropy, etc.) has made breakthroughs in feature quantification, there are two key defects in existing entropy analysis methods: First, there is a lack of quantification ability for the dynamic evolution characteristics of fractional-order time series; second, the fixed-scale calculation mode cannot adapt to the non-stationary characteristics of the equipment operating state.

[0111] Due to the preset fixed weight allocation mechanism, the conventional weighted average method is prone to information loss during the feature fusion process. Although deep learning models have powerful feature learning capabilities, they face three application obstacles: generally high model parameter quantities leading to high energy consumption, dependence on large-scale labeled data sets, and complex model structures making it difficult to achieve real-time inference on embedded devices.

[0112] The parameter optimization of neural networks is essentially a problem of solving the optimal solution in a high-dimensional non-convex space, and there are a large number of local extreme points on the loss function surface. The optimization algorithm based on gradient descent has two inherent defects: First, the learning rate decay strategy is likely to make the parameters fall into a flat area, resulting in premature convergence; second, the ill-conditioned number of the Hessian matrix severely restricts the practical application of second-order optimization methods.

[0113] The technical effects of the above technical solution are as follows: Aiming at the defects of existing traditional time-frequency domain analysis methods and traditional entropy value theories, the present invention will use the refined composite multi-scale fractional-order reverse dispersion entropy for feature extraction of time-series data such as electroencephalogram, body temperature, heart rate, and blood oxygen saturation. Through multi-scale decomposition, the signal complexity characteristics at different time scales can be captured, so as to more comprehensively reflect the dynamic changes of the signal. The introduction of the fractional order increases the flexibility and adaptability of the model and can better process non-linear and non-stationary signals. The reverse dispersion entropy itself has good anti-noise performance and can effectively suppress the influence of noise on the signal complexity measurement. The refined composite multi-scale structure further improves the stability of the complexity measurement and enables it to maintain good performance under different signal-to-noise ratios.

[0114] The present invention proposes a fractional-order LSTM algorithm as a multi-classifier. By using the fractional-order Softplus activation function and the fractional-order tanh activation function to replace the Sigmoid activation function and the tanh activation function of the traditional LSTM algorithm.

[0115] The fractional-order Softplus activation function inherits the smoothness of the Softplus function and is continuously differentiable throughout its domain. This smoothness enables the neural network to avoid the problems of gradient vanishing or explosion during the backpropagation process, thereby improving the stability and convergence speed of training. The fractional-order Softplus activation function does not output zero in the negative input region, so the situation of "dead neurons" will not occur. This enables the network to maintain a more active state during training and avoid some neurons from stopping learning due to a zero gradient. And by introducing the concept of fractional-order calculus, it can better adapt to different data distributions and network structures. This self-adaptability makes it perform excellently in dealing with complex non-linear problems, especially in scenarios where the shape of the activation function needs to be flexibly adjusted. The computational efficiency can be improved through polynomial approximation. This approximation method not only reduces the computational complexity but also improves the generalization ability of the model while maintaining the non-linearity of the activation function. Its design is based on the theoretical framework of fractional-order calculus and can better meet the conditions of the universal approximation theorem. This means that it can approximate complex functions in a more efficient way and thus perform excellently in deep learning tasks.

[0116] The fractional tanh activation function inherits the zero-centered property of the traditional tanh function, meaning that the output values are symmetrically distributed around zero. This property maintains gradient balance during training, preventing training instability caused by output drift. The traditional tanh function suffers from the vanishing gradient problem when the input values are large or small. By incorporating fractional calculus, the fractional tanh activation function alleviates this problem to a certain extent, thereby improving network training efficiency. This activation function maintains the symmetry of the tanh function, effectively handling positive and negative input values, while also incorporating nonlinear properties, enabling the network to learn complex patterns and relationships. The introduction of fractional calculus allows it to better adapt to complex input data distributions, particularly when processing data with nonlinear characteristics. Due to its zero-centeredness and symmetry, the fractional tanh activation function is particularly well-suited for recurrent neural networks (such as LSTMs), providing balanced inputs for subsequent time steps and improving the learning of long-term dependencies. These improvements give the fractional tanh activation function a significant advantage in handling complex tasks.

[0117] Fractional activation functions can capture more complex nonlinear relationships than traditional integer-order activation functions, resulting in higher accuracy when processing complex dynamic systems and time series data. They can achieve faster convergence and shorter training times. By introducing fractional-order parameters, fractional activation functions increase model flexibility and adaptability, resulting in better generalization when processing complex datasets. Fractional calculus has a memory function, enabling better simulation of dynamic systems with memory properties. By introducing fractional activation functions into LSTM, the model can more effectively capture long-term dependencies, further improving its ability to model time series data.

[0118] The parameter optimization process for the fractional-order LSTM algorithm is equivalent to solving a strongly nonlinear, non-convex optimization problem. Traditional machine learning optimization methods that rely on gradient information often fall into the trap of local minima, which can be solved by global optimization. This invention, based on the Genghis Khan Shark optimization algorithm, enhances its global optimization capabilities by introducing a spiral search walk strategy and a crazy operator.

[0119] In one embodiment of the present invention, the S1 includes:

[0120] The fatigue sign data of the power workers are collected through the intelligent helmet wearing device, and the collected fatigue sign data are pre-processed to obtain fatigue sign collection data;

[0121] Obtaining preset fatigue categories, the preset fatigue categories including EEG, body temperature, heart rate, and blood oxygen saturation;

[0122] Classify the fatigue sign acquisition data through preset fatigue categories to obtain fatigue category data;

[0123] The fatigue category data includes electroencephalogram data, body temperature data, heart rate data, and blood oxygen saturation data.

[0124] The working principle of the above technical solution is as follows: Reverse Dispersion Entropy (RDE) is a concept used in signal processing and nonlinear dynamics analysis, especially in fault diagnosis and signal complexity evaluation. As a new nonlinear dynamic analysis method, RDE is based on Permutation Entropy (PE) and combines the advantages of Dispersion Entropy (DE) and Reverse Permutation Entropy (RPE), showing good performance in stability and discrimination.

[0125] For a time series X = {x(1), x(2), …, x(T)} of length T, the calculation steps of fractional-order RDE are as follows:

[0126] (1) Through the normal distribution function:

[0127]

[0128] Map the time series X = {x(i), i = 1, 2, …, T} to the range of Y = {y(i), i = 1, 2, …, T}, where the range of y(i) is between 0 and 1, and μ is the expectation and σ 2 is the variance.

[0129] (2) Through linear transformation:

[0130]

[0131] Map the new sequence Y = {y(i), i = 1, 2, …, T} to the range of Z = {z(i), i = 1, 2, …, T}, where int is the integer function, c is the number of categories, and z(i) is a positive integer from 1 to c.

[0132] (3) Phase space reconstruction:

[0133] Reconstruct Z into L embedding vectors with a time delay of d and an embedding dimension of m. The matrix composed of all embedding vectors is

[0134]

[0135] In the formula, the number of embedding vectors L is T - (m - 1)d.

[0136] (4) Calculate the scattering pattern:

[0137] Each embedding vector can be mapped into a scattering pattern π. The total number of all possible scattering patterns is c m kinds, because each embedding vector has m components, and each component can be an integer between 1 and c.

[0138] (5) Calculate the probability of each scattering pattern:

[0139] The relative probability of the i-th scattering pattern can be expressed as

[0140]

[0141] In the formula: P(π i ) is the ratio of the number of the i-th scattering pattern to the number of embedding vectors; Number{π i} refers to the number of embedding vectors mapped into the i-th scattering pattern.

[0142] (6) Calculate the fractional-order RDE:

[0143]

[0144] Among them, Γ(·) is the gamma function, ψ(·) is the digamma function, and α is the fractional order.

[0145] Refined composite multi-scale fractional-order reverse scattering entropy:

[0146] FrRDE has great advantages in analyzing the randomness and dynamic mutation behavior of time series at a single scale. In order to measure the complexity of time series at different scales, multi-scale analysis methods have emerged. However, the multi-scale calculation defined based on the coarse-graining process depends on the length of the time series and does not consider the relationship between the new coarse-grained sequences after calculation, resulting in information loss. To address the above problems, combined with the refined composite multi-scale method, RCMFrRDE is proposed, and the calculation steps are as follows:

[0147] (1) Coarse-graining processing:

[0148] For the original time series X = {x1, x2,..., x T}, its k-th coarse-grained sequence at the scale factor of τ is given by the following formula:

[0149]

[0150] (2) Calculate RCMFrRDE:

[0151] At each scale factor τ, combining formula (6) with formulas (1) to (4) can obtain:

[0152]

[0153] where is the \(i\)-th spreading pattern under the scale factor \(\tau\).

[0154] For each scale factor \(\tau\), the RCMFrRDE is defined as:

[0155]

[0156] where

[0157] The technical effects of the above technical solution are as follows: FRDE combines the advantages of spreading entropy and reverse permutation entropy, and can perform well in terms of stability and distinguishability, so as to more accurately evaluate the complexity of signals. This is particularly important for applications such as fault diagnosis, because it can help identify subtle signal changes. Through phase space reconstruction and the calculation of spreading patterns, FRDE can capture the non-linear dynamic characteristics in time series. This is of great significance for understanding the dynamic behavior of systems, predicting the future state of systems, and identifying potential system faults. By combining the fine composite multi-scale method, RCMFrRDE overcomes the deficiencies of traditional multi-scale analysis methods in terms of information loss. It can measure the complexity of time series at different scales, so as to more comprehensively understand the dynamic behavior of systems. This is particularly useful for analyzing complex systems with multi-scale characteristics. Both FRDE and RCMFrRDE adopt preprocessing steps such as normal distribution functions and linear transformations, which can enhance the robustness of the algorithms and enable them to process time series of different types and characteristics. In addition, these algorithms are also applicable to various practical application scenarios, such as biomedical signal processing, mechanical fault diagnosis, etc. FRDE and RCMFrRDE provide new effective tools for non-linear dynamics analysis. They can reveal the hidden information and dynamic characteristics in time series, providing researchers with a deeper understanding and analysis means. This is of great significance for promoting the development of the field of non-linear dynamics. This method has shown remarkable technical effects in signal processing and non-linear dynamics analysis, and has made important contributions to improving the accuracy of signal complexity evaluation, enhancing the ability to capture the dynamic characteristics of time series, providing the ability of multi-scale analysis, improving the robustness and applicability of algorithms, and providing new tools for non-linear dynamics analysis.

[0158] In one embodiment of the present invention, the S2 includes:

[0159] Obtain a time series of a preset length, and calculate the fractional order RDE by means of a normal distribution function, linear transformation, phase space reconstruction, calculation of the walking pattern and its probability;

[0160] After performing phase granulation processing on the time series, calculate the RCMFrRDE to obtain the calculation data.

[0161] The working principle of the above technical solution is as follows: The fractional Softplus activation function includes the fractional Softplus (FracSoftplus), which uses the softplus function to generalize the Sigmoid function through fractional derivatives and is defined as:

[0162]

[0163] Then the calculation expression is:

[0164]

[0165] where a i is a trainable parameter.

[0166] Fractional tanh activation function:

[0167] Fractional tanh (FracTanh) is a generalization of the fractional tanh and is defined as:

[0168]

[0169] Then the calculation expression is:

[0170]

[0171] where a i is a trainable parameter.

[0172] The technical effects of the above technical solution are as follows: The fractional Softplus activation function is an activation function based on fractional calculus. It inherits the smoothness of the Softplus function and is continuously differentiable throughout the domain. This smoothness enables the neural network to avoid the problems of gradient vanishing or explosion during backpropagation, thereby improving the stability and convergence speed of training. By introducing the concept of fractional calculus, the fractional Softplus can better adapt to different data distributions and network structures. This adaptability makes it perform excellently in dealing with complex non-linear problems, especially in scenarios where the shape of the activation function needs to be flexibly adjusted. In some cases, the fractional Softplus can improve the computational efficiency through polynomial approximation. This approximation method not only reduces the computational complexity but also enhances the generalization ability of the model while maintaining the non-linearity of the activation function. The design of the fractional Softplus is based on the theoretical framework of fractional calculus and can better meet the conditions of the universal approximation theorem. The gradient of the fractional Softplus is smoother, avoiding the problem of excessive gradient change at certain points of the Softplus, and enabling a more stable gradient flow. The fractional Softplus introduces richer non-linear characteristics through fractional derivatives, allowing the model to learn more complex patterns. Generally speaking, it not only improves the computational efficiency and numerical stability but also enhances the gradient and non-linear characteristics, thus performing better in terms of model performance. These improvements give the Fractional Softplus greater advantages in dealing with complex tasks and deep networks. The fractional tanh activation function inherits the zero-centered property of the traditional tanh function, that is, the output values are symmetrically distributed around zero. This property can maintain the balance of the gradient during training and avoid training instability caused by output offset. The traditional tanh function faces the problem of gradient vanishing when the input value is large or small, while the fractional tanh activation function can alleviate this problem to a certain extent through the introduction of fractional calculus, thereby improving the training efficiency of the network. This activation function maintains the symmetry of the tanh function, can effectively handle positive and negative input values, and at the same time introduces non-linear characteristics, enabling the network to learn complex patterns and relationships. The introduction of fractional calculus enables it to better adapt to complex input data distributions, especially performing excellently when dealing with data with non-linear characteristics. Due to its zero-centered and symmetric properties, the fractional tanh activation function is particularly suitable for recurrent neural networks (such as LSTM), can provide balanced input for subsequent time steps, and can improve the learning of long-term dependencies. These improvements give the fractional tanh activation function greater advantages in dealing with complex tasks.

[0173] In one embodiment of the present invention, S3 includes:

[0174] Obtain the fractional Softplus activation function and the fractional tanh activation function;

[0175] Set the fractional Softplus activation function and the fractional tanh activation function for the long short-term memory network to obtain a long short-term memory network with the fractional Softplus activation function and the fractional tanh activation function;

[0176] Among them, the setting of the fractional Softplus activation function and the fractional tanh activation function for the long short-term memory network includes:

[0177] Initialize the fractional LSTM cell state C and the hidden state h with zero or small random values;

[0178] Apply the fractional Softplus activation function to the weighted sum of the current input xt and the previous hidden state ht-1 to calculate the input gate activation it;

[0179] Apply the fractional Softplus activation function to the weighted sum of the current input xt and the previous hidden state ht-1 to further determine the forget gate activation ft; The information of the cell state Ct-i that should be forgotten or maintained is determined by this gate;

[0180] Apply the fractional tanh activation function to the weighted sum of the current input xt and the previous hidden state ht-1 to obtain the candidate cell state Calculate the new potential value of the cell state at this stage;

[0181] The candidate cell state Combine the previous cell state Ct-i and the forget gate ft to update the cell state Ct-i; At this stage, the data that should be retained in the cell state is determined;

[0182] Apply the fractional Softplus activation function to the weighted sum of the current input xt and the previous hidden state ht-1 to determine the output gate activation Ot; The cell state information that should be generated as the hidden state ht is determined by this gate;

[0183] Apply the fractional tanh activation function to the updated cell state Ct multiplied by the output gate Ot to obtain the new hidden state ht; The information to be sent to the next time step and used as the final output is stored in this hidden state ht;

[0184] The hidden state ht is used for the output of the fractional LSTM cell at the current time step;

[0185] Repeat the operation for each time step in the sequence.

[0186] The working principle of the above technical solution is as follows: The advanced description of the fractional-order LSTM algorithm in terms of mathematical operations and data flow is as follows:

[0187] 1. Initialize the fractional-order LSTM cell state C and hidden state h with zero or small random values.

[0188] 2. Calculate the input gate activation i by applying the fractional-order Softplus activation function to the weighted sum of the current input x t and the previous hidden state h t-1 . t

[0189] i t = σ FS ([h t-1 , x t *w i + b i ) (13)

[0190] 3. Apply the fractional-order Softplus activation function to the weighted sum of the current input x t and the previous hidden state h t-1 to determine the forget gate activation f t . The information of the cell state C t-1 that should be forgotten or maintained is determined by this gate.

[0191] f t = σ FS ([h t-1 , x t *w f + b f ) (14)

[0192] 4. Apply the fractional-order tanh activation function to the weighted sum of the current input x t and the previous hidden state h t-1 to obtain the candidate cell state The new potential value of the cell state is calculated at this stage.

[0193]

[0194] 5. The candidate cell state the previous cell state C t-1 and the forget gate f t are combined to update the cell state C t-1 . At this stage, the data to be retained in the cell state is determined.

[0195]

[0196] ​6. Apply the fractional Softplus activation function to the current input x t and the previous hidden state h t-1 in a weighted sum to determine the output gate activation O t . The cell state information generated by the hidden state h t is determined by this gate.

[0197] O t = σ FS ([h t-1 , x t *w o + b o )(17)

[0198] 7. Apply the fractional tanh activation function to the updated cell state C t multiplied by the output gate O t to obtain the new hidden state h t . The information to be sent to the next time step and possibly used as the final output is included in this hidden state.

[0199] h t = O t * Frtanh(C t )(18)

[0200] 8. The hidden state h t can be used as the output of the fractional LSTM cell at the current time step.

[0201] 9. Repeat the above steps for each time step in the sequence. The fractional LSTM neuron structure is as Figure 2 shown.

[0202] where i t is the input gate activation, σ FS is the fractional Softplus activation function, w i is the weight matrix of the input gate, x t is the current input, h t-1 is the previous hidden state, b i is the bias of the input gate, f t is the forget gate activation, w f is the weight matrix of the forget gate, b f is the bias of the forget gate, is the candidate cell state, b c is the bias of the candidate cell state, Frtanh is the fractional tanh activation function, w c is the weight matrix of the candidate cell state, C t is the updated cell state, C t-1 is the previous cell state, O tThe output gate is activated, w o is the weight matrix of the output gate, b o is the bias of the output gate, h t is the new hidden state.

[0203] The cross-entropy loss function is one of the most commonly used loss functions in multi-classification tasks. It is used to measure the difference between the probability distribution of the model output and the true labels. For multi-classification problems, the categorical cross-entropy loss function is usually used, and its mathematical expression is:

[0204]

[0205] where y true,i is the one-hot encoding of the true label, and ŷ pred,i is the probability distribution predicted by the model. In the fractional-order LSTM model, the cross-entropy loss function can effectively guide the model to learn how to correctly classify the input sequence into different categories.

[0206] The technical effects of the above-mentioned technical solution are as follows: The fractional-order Softplus and fractional-order tanh activation functions introduce the concept of fractional-order calculus, enabling LSTM units to capture more complex nonlinear relationships. This enhances the model's expressive power, enabling it to better adapt to and process a variety of complex time series data. By using fractional-order activation functions, the model can learn richer feature representations during training, thereby improving its generalization ability. This means that the model can demonstrate better performance and stability when processing unseen data. LSTM networks inherently have long-term memory capabilities, and the introduction of fractional-order activation functions can further optimize this memory mechanism. By adjusting the fractional-order parameters, the model's sensitivity to historical information and forgetting rate can be controlled, thereby achieving a better balance between short-term and long-term memory. Fractional-order activation functions have smooth transition characteristics, which can reduce model oscillations and instability during training. Therefore, LSTM networks using fractional-order activation functions are more robust to noise and outliers. Fractional-order parameters provide an additional adjustment dimension for the model, allowing researchers to flexibly adjust the model's complexity and performance based on specific tasks and data characteristics. This allows finding the optimal model configuration for different application scenarios. The introduction of fractional calculus into deep learning provides new insights and methods for related research. This not only promotes the development of recurrent neural networks such as LSTM but also has the potential to positively impact other types of neural networks and machine learning algorithms. Applying the fractional-order Softplus activation function and the fractional-order tanh activation function to long-short-term memory networks can significantly improve the model's nonlinear expressiveness, generalization, memory capacity, and robustness, while also providing flexible model adjustment methods.

[0207] In one embodiment of the present invention, the S3 further includes:

[0208] In the new Genghis Khan shark optimization algorithm, a random position is calculated by the upper and lower bounds of the search space to obtain the best predation position;

[0209] A comprehensive data search of the target area is performed through a spiral search wandering strategy, and a crazy operator is added for deep calculation to obtain deep calculation results.

[0210] The above technical solution works as follows: Genghis Khan sharks are top predators in freshwater rivers. They typically hang out near the bottom, seemingly on patrol. In reality, they do this to ensure that no predators stronger than themselves are in the area. This not only ensures their own safety but also expands their search range across the entire space to locate the best prey. GKS do not attack prey suddenly, but once they locate their prey, they wait for the perfect moment to launch a full-scale attack.

[0211] To simulate the natural behavior of the Genghis Khan shark, in GKSO, a new random position is calculated as the "optimal predation position" by using the upper and lower bounds of the search space. The position update formula of the Genghis Khan shark is as follows:

[0212]

[0213] Among them, X i j (t+1) represents the position of the i-th member in the j-th dimension at time t+1, ub j and lb j They represent the upper and lower bounds of the j-th dimension, r1 is a random number in the interval [0,1], N represents the group size, D represents the problem dimension, it represents the current number of iterations, and T represents the total number of iterations.

[0214] To capture higher-quality prey, the Genghis Khan shark has a certain probability of employing a strategy of continuously approaching the best prey, relying on its keen sense of smell. It also has a certain probability of employing a spiral search around the best prey. The specific mathematical model of this behavior is as follows:

[0215]

[0216] in, represents the value of the best known prey location in the jth dimension at time t, and s represents the olfactory intensity of the Genghis Khan shark when moving towards the best prey, which depends on the concentration of the odor emitted by the prey.

[0217] The attraction model is established as follows:

[0218] s=mI r (twenty three)

[0219] where r is a random number on the interval [0,1], representing the degree to which the search agent absorbs the scent of the prey.

[0220] The spiral search strategy is a search technique that simulates spiral motion in nature. Its core idea is to achieve full coverage of the target area by continuously adjusting the search direction. η is a random number in the interval [0, 1]. By comparing the value with 0.5, one of the two strategies is selected. The parameter k is the coefficient of variation, and the parameter l is a random number that satisfies a uniform distribution in the interval [-1, 1]. T is the current number of iterations, T max is the maximum number of iterations.

[0221] Therefore, the parameter r controls the behavior of the Genghis Khan shark. I is the attribute strength, which depends on the ability of the individuals in the group, that is, the current fitness value of each agent. Another important parameter is m, which is a non-negative constant that needs to be determined based on the characteristics of the optimization problem. In addition, m is a key parameter that affects the convergence speed of the new GKSO. We retain the positions of the first two optimal agents and update the positions of the other agents based on these two optimal positions to make them as close to the optimal position as possible. The position update formula of the Genghis Khan shark is as follows:

[0222]

[0223] During the hunting process, fish express themselves in many unexpected ways or strategies. After approaching the optimal prey, the Genghis Khan shark uses the prey as a reference point, exploits its speed advantage to quickly swim in front of the prey, and delivers a fatal blow to the prey's head with its large mouth. The entire process follows a parabolic trajectory. To overcome premature convergence of the algorithm, the present invention incorporates a crazy operator. The specific formula is as follows:

[0224]

[0225] Where r2 is a random number in the interval [0,1], λ is a random number with a value of 1 or -1, and p is a parameter that controls the step size of the new GKSO. This parameter is a nonlinear convergence factor with perturbations. When the value of p is large, the new GKSO focuses primarily on global exploration; as the value of p decreases, local exploration gradually becomes dominant. The formula for calculating p is as follows:

[0226]

[0227] Among them, |ω(t+1)| is the weight coefficient at time t+1, and the calculation formula is as follows:

[0228] |ω(t+1)|=1-2ω 4 (t) (29)

[0229] The initial value ω(0)=0.1.

[0230] Among them, ξ is a random number in the interval [0,1], Θ is a given crazy probability, and κ is a constant with an extremely small value.

[0231] Over the course of long-term evolution in their natural environment, freshwater sharks of the same genus have developed a number of reflexes to adapt to their surroundings. Color change is often influenced by external factors. During foraging, Genghis Khan sharks frequently encounter predators or competitors for prey. Therefore, to escape these predators, Genghis Khan sharks have developed a color-changing mechanism similar to that of squids. When startled, the color of their tail and body lightens, deterring predators and allowing them to quickly flee.

[0232] In order to mathematically simulate the behavior of the Genghis Khan shark, we carried out this process in the fourth stage of the new GKSO. The specific mathematical model is as follows:

[0233]

[0234] Where k1 is a random number uniformly distributed in the interval [-1, 1], and k2 is a normally distributed random number with a mean of 0 and a standard deviation of 1. a1, a2, and a3 are three random numbers, and the calculation formula is as follows:

[0235]

[0236] Here, rand is a random number in the interval [0,1], and l1 is a binary parameter with a value of 0 or 1. In addition, ρ is an adaptive coefficient, which is calculated as follows:

[0237]

[0238] Among them, β min and β max Take 0.2 and 1.2 respectively. In addition, the randomly generated solution and The calculation formula at two moments t is as follows:

[0239]

[0240] Calculation solution The formula is as follows:

[0241]

[0242] in, represents a set of solutions generated by random initialization at time t, is a randomly selected solution at time t. The value of l2 is determined in the same way as l1.

[0243] The technical advantages of the above-mentioned technical solution are as follows: By simulating the behavior of the Genghis Khan shark, which wanders near the bottom of the water in search of optimal prey, the algorithm is able to explore extensively within the search space, effectively avoiding premature convergence and improving global search capabilities. The algorithm incorporates two strategies: relying on olfaction to approach optimal prey and spiral search wandering. These strategies can be flexibly adjusted based on the current search state, making the search process more efficient and enabling rapid localization near the optimal solution. The introduction of parabolic foraging behavior enables the algorithm to leverage its speed advantage to conduct a refined search near the optimal solution, thereby improving local exploration capabilities and finding higher-quality optimal solutions. The inclusion of a crazy operator allows the algorithm to maintain a certain level of perturbation during the search process, effectively avoiding being trapped in local optima, further improving the algorithm's search efficiency and robustness. Multiple parameters in the algorithm (such as olfactory intensity, coefficient of variation, and active step size) can be adjusted based on the characteristics of the optimization problem, making the algorithm adaptable to different types of optimization problems. The introduction of a self-protection mechanism allows the algorithm to simulate the color-changing behavior of the Genghis Khan shark during the search process, thereby preventing interference from other search agents to a certain extent and improving the algorithm's stability. The algorithm's mathematical model is clear and straightforward, making it easy to implement and expand. Furthermore, each step and parameter in the algorithm has clear physical meaning and mathematical expressions, facilitating theoretical analysis and experimental verification. The novel swarm intelligence optimization algorithm (GKSO), based on the natural behavior of Genghis Khan sharks, demonstrates significant technical results in terms of global search capability, search efficiency, local exploitation, overcoming premature convergence, adaptability, stability, and ease of implementation. These technical results give GKSO significant potential and advantages in solving complex optimization problems.

[0244] In one embodiment of the present invention, the classification system includes:

[0245] A data acquisition module is used to acquire the intelligent helmet wearing device, collect fatigue sign data of the power workers, obtain the fatigue sign collected data, classify the fatigue sign collected data according to preset fatigue categories, and obtain fatigue category data;

[0246] A feature extraction module is used to extract features from the fatigue category data by using fine composite multi-scale fractional order inverse scatter entropy to obtain fatigue feature extraction data;

[0247] A parameter optimization module is used to obtain three fatigue warning category labels, input the fatigue feature extraction data combined with the three fatigue warning category labels into the fractional-order long short-term memory network for training, and perform parameter optimization processing on the training process using the new Genghis Khan shark optimization algorithm to obtain network optimization data;

[0248] The fatigue warning module is used to input the network optimization data into the Softmax function, and then output multi-class probability data for fatigue warning.

[0249] The working principle of the above technical solution is as follows: Safety production is the most important prerequisite for power operations. The working environment of the production department of the State Grid is complex, with high-altitude operations, proximity to high-voltage equipment, etc., and the safety risks are relatively large. Safety accidents may cause serious injuries or even life-threatening. Therefore, in power operations, it is particularly important to monitor the physical health status of operators. Due to environmental and topographical limitations at the operation site, communication signals are restricted, accurate positioning is impossible, and the medical emergency response is slow. Improving the safety of employees working in dangerous environments has become an urgent need in the power industry. During the power operation of the operator, detailed information about the current self-status can be obtained through the intelligent safety helmet wearing device, and a quick analysis and warning of the self-physiological signs can be carried out in advance, so as to facilitate the operator's task, and ensure the reliability and efficiency of safety and dispatching command.

[0250] The main idea of the present invention is to use the fine composite multi-scale fractional-order reverse dispersion entropy to comprehensively extract the features of on-site data, then input the obtained features into the fractional-order long short-term memory network for multi-classification processing of health warning, and finally use the new Genghis Khan shark optimization algorithm to optimize the network parameters, as Figure 1 shown.

[0251] The fatigue warning algorithm for the intelligent safety helmet wearing device of power operation personnel should train the four types of feature data extracted by the fine composite multi-scale fractional-order reverse dispersion entropy method, namely electroencephalogram, body temperature, heart rate, and blood oxygen saturation, and three fatigue warning category labels, namely normal, fatigue, and high-risk, by inputting them into the fractional-order long short-term memory network. This network consists of 8 series-connected fractional-order LSTM neurons and is connected to the Softmax function to output multi-class probabilities (predicted labels). Its network structure is as Figure 3 shown.

[0252] The technical effects of the above technical solution are as follows: Aiming at the defects of the existing traditional time-frequency domain analysis method and traditional entropy theory, the present invention will use the fine composite multi-scale fractional-order reverse dispersion entropy to extract the features of time series data such as electroencephalogram, body temperature, heart rate, and blood oxygen saturation. Through multi-scale decomposition, the signal complexity features at different time scales can be captured, so as to more comprehensively reflect the dynamic changes of the signal. The introduction of fractional order increases the flexibility and adaptability of the model, and can better process non-linear and non-stationary signals. The reverse dispersion entropy itself has good anti-noise performance and can effectively suppress the influence of noise on the signal complexity measurement. The fine composite multi-scale structure further improves the stability of the complexity measurement, so that it can maintain good performance under different signal-to-noise ratios.

[0253] Obviously, those skilled in the art can make various modifications and variations 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 also intends to include these modifications and variations.

Claims

1. An intelligent safety helmet operation fatigue risk warning classification method based on a fractional-order long short-term memory network, characterized in that The classification method includes: S1. Obtain an intelligent safety helmet wearing device, collect fatigue sign data of power workers to obtain fatigue sign collection data, and classify the fatigue sign collection data according to a preset fatigue category to obtain fatigue category data; S2. Extract features from the fatigue category data through a fine composite multi-scale fractional-order reverse dispersion entropy to obtain fatigue feature extraction data; S3. Obtain three fatigue warning category labels, input the fatigue feature extraction data combined with the three fatigue warning category labels into a fractional-order long short-term memory network for training, and perform parameter optimization processing on the training process through a new Genghis shark optimization algorithm to obtain network optimization data; S4. Input the network optimization data into a Softmax function, and then output multi-category probability data for fatigue warning.

2. The intelligent safety helmet operation fatigue risk warning classification method based on the fractional-order long short-term memory network according to claim 1, wherein The S1 includes: Collect fatigue sign data of power workers through an intelligent safety helmet wearing device, preprocess the collected fatigue sign data to obtain fatigue sign collection data; Obtain a preset fatigue category, where the preset fatigue category includes electroencephalogram, body temperature, heart rate, and blood oxygen saturation; Classify the fatigue sign collection data according to the preset fatigue category to obtain fatigue category data; The fatigue category data includes electroencephalogram data, body temperature data, heart rate data, and blood sample saturation data.

3. The intelligent safety helmet operation fatigue risk early warning classification method based on the fractional-order long short-term memory network according to claim 1, wherein The S2 includes: Obtain a time series of a preset length, calculate the fractional-order RDE by using a normal distribution function, linear transformation, phase space reconstruction, calculating the diffusion pattern and its probability; After performing phase granulation processing on the time series, calculate the RCMFrRDE to obtain calculation data.

4. The intelligent safety helmet operation fatigue risk warning classification method based on the fractional-order long short-term memory network according to claim 1, wherein, The S3 includes: Obtain a fractional-order Softplus activation function and a fractional-order tanh activation function; Set the fractional-order Softplus activation function and the fractional-order tanh activation function for the long short-term memory network to obtain a long short-term memory network with a fractional-order Softplus activation function and a fractional-order tanh activation function; Among them, the setting of the fractional-order Softplus activation function and the fractional-order tanh activation function for the long short-term memory network includes: Initialize the fractional-order LSTM cell state C and the hidden state h with zero or small random values; Apply the fractional Softplus activation function to the current input x t and the previous hidden state h t-1 for the weighted sum to calculate the input gate activation i t ; Apply the fractional Softplus activation function to the current input x t and the previous hidden state h t-1 to determine the forget gate activation f t ; Apply the fractional-order tanh activation function to the current input x t and the previous hidden state h t-1 to obtain the cell state of the candidate as a weighted sum Combine the candidate cell state with the previous cell state C t-i and the forget gate f t to update the cell state C t-i ; Apply the fractional Softplus activation function to the current input x t and the previous hidden state h t-1 to determine the output gate activation O t ; Apply the fractional-order tanh activation function to the updated cell state C t Multiply by the output gate O t , to obtain the new hidden state h t ; Store the information that will be sent to the next time step and used as the final output with this hidden state h t ; Hidden state h t Output for the fractional-order LSTM cell at the current time step; Repeat the operation for each time step in the sequence.

5. The intelligent safety helmet operation fatigue risk warning classification method based on the fractional-order long short-term memory network according to claim 1, characterized in that, The S3 further includes: In the new Genghis shark optimization algorithm, calculate a random position through the upper and lower bounds of the search space to obtain the best predation position; Perform a comprehensive data search in the target area through a spiral search and wandering strategy, and add a crazy operator for in-depth calculation to obtain an in-depth calculation result.

6. A classification system for an intelligent safety helmet operation fatigue risk warning classification method based on a fractional-order long short-term memory network as described in claim 1, characterized in that, The classification system includes: A data collection module, which is used to obtain an intelligent safety helmet wearing device, collect fatigue sign data of power workers to obtain fatigue sign collection data, and classify the fatigue sign collection data according to a preset fatigue category to obtain fatigue category data; A feature extraction module, which is used to extract features from the fatigue category data through a fine composite multi-scale fractional-order reverse dispersion entropy to obtain fatigue feature extraction data; The parameter optimization module is used to obtain three fatigue warning category labels, input the fatigue feature extraction data combined with the three fatigue warning category labels into the fractional-order long short-term memory network for training respectively, and perform parameter optimization processing on the training process through the novel Genghis Khan shark optimization algorithm to obtain network optimization data; The fatigue warning module is used to input the network optimization data into the Softmax function, and then output multi-category probability data for fatigue warning.

7. The intelligent safety helmet operation fatigue risk warning classification method based on the fractional-order long short-term memory network according to claim 6, characterized in that The data acquisition module includes: The acquisition module is used to collect the fatigue sign data of power workers through the intelligent safety helmet wearing device, preprocess the collected fatigue sign data, and obtain the fatigue sign acquisition data; Obtain the preset fatigue categories, and the preset fatigue categories include electroencephalogram, body temperature, heart rate and blood oxygen saturation; The classification module is used to classify the fatigue sign acquisition data through the preset fatigue categories to obtain fatigue category data; The fatigue category data includes electroencephalogram data, body temperature data, heart rate data and blood sample saturation data.

8. The intelligent safety helmet operation fatigue risk warning classification method based on the fractional-order long short-term memory network according to claim 6, characterized in that The feature extraction module includes: The RDE calculation module is used to obtain a time series of a preset length, calculate the fractional-order RDE by means of the normal distribution function, linear transformation, phase space reconstruction, calculation of the walking pattern and its probability; The phase granulation processing module is used to perform phase granulation processing on the time series and then calculate the RCMFrRDE to obtain calculation data.

9. The intelligent safety helmet operation fatigue risk warning classification method based on the fractional-order long short-term memory network according to claim 6, characterized in that, The parameter optimization module includes: The function acquisition module is used to obtain the fractional-order Softplus activation function and the fractional-order tanh activation function; The network processing module is used to set the fractional-order Softplus activation function and the fractional-order tanh activation function for the long short-term memory network to obtain a long short-term memory network with the fractional-order Softplus activation function and the fractional-order tanh activation function; Among them, the network processing module includes: Initialize the fractional-order LSTM cell state C and the hidden state h with zero or small random values; Apply the fractional Softplus activation function to the current input x t and the previous hidden state h t-1 to calculate the input gate activation i t ; Apply the fractional Softplus activation function to the current input x t and the previous hidden state h t-1 to determine the forget gate activation f t ; Apply the fractional-order tanh activation function to the current input x t and the previous hidden state h t-1 to obtain the cell state of the candidate as a weighted sum Combine the candidate cell state with the previous cell state C t-i and the forget gate f t to update the cell state C t-i ; Apply the fractional Softplus activation function to the current input x t and the previous hidden state h t-1 to determine the output gate activation O t ; Apply the fractional-order tanh activation function to the updated cell state C t Multiply by the output gate O t to obtain the new hidden state h t ; Store the information to be sent to the next time step and used as the final output in this hidden state h t ; Hidden state h t Output for the fractional-order LSTM cell at the current time step; Repeat the operation for each time step in the sequence.

10. The intelligent safety helmet operation fatigue risk warning classification method based on the fractional-order long short-term memory network according to claim 6, characterized in that, The parameter optimization module further includes: The position acquisition module is used to calculate a random position through the upper and lower bounds of the search space in the novel Genghis Khan shark optimization algorithm to obtain the best predation position; The depth calculation module is used to perform comprehensive data search in the target area through the spiral search and wandering strategy, add the crazy operator for depth calculation, and obtain the depth calculation result.

Images