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

Abstract

The application provides an intelligent safety helmet operation fatigue risk early warning classification method based on a fractional order long short-term memory network, relates to the technical field of risk early warning, collects fatigue sign data, obtains fatigue sign collection data, classifies the fatigue sign collection data, and obtains fatigue category data; the fatigue category data is subjected to feature extraction to obtain fatigue feature extraction data; three kinds of fatigue early warning category labels are acquired, the fatigue feature extraction data is combined with the three kinds of fatigue early warning category labels and respectively input into a fractional order long short-term memory network for training, a new Genghis Khan shark optimization algorithm is used for parameter optimization processing on the training process to obtain network optimization data; the network optimization data is input into a Softmax function, and then multi-category probability data is output, and fatigue early warning is performed; and the application effectively improves the safety and reliability of power production.

Description

Technical Field

[0001] This invention proposes a classification method for early warning of work fatigue risk in smart safety helmets based on fractional-order long short-term memory networks, which relates to the field of risk warning technology, specifically the field of work fatigue risk warning technology. Background Technology

[0002] Power generation is a complex and far-reaching system engineering project involving numerous on-site personnel. At power production sites, project managers, supervisors, technicians, engineers, and other professionals of various types need to coordinate closely to process massive amounts of data. Errors in power production management can not only jeopardize safe power production but also, in severe cases, lead to injuries or fatalities. Traditional power inspection processes are complex and heavily reliant on the experience and expertise of on-site inspectors, making them susceptible to human error that can lead to delayed problem detection and even major accidents. Furthermore, they fail to guarantee the safety of inspection personnel.

[0003] The sensors in smart helmets worn by power workers typically collect time-series data such as electroencephalogram (EEG), body temperature, heart rate, and blood oxygen saturation. Various algorithms and techniques are generally employed to extract features from this time-series data. These algorithms can be broadly categorized into time-domain feature extraction and frequency-domain feature extraction methods. However, each of these methods has its unique advantages and limitations. Feature extraction from EEG data, in particular, is a crucial step in EEG signal processing and its importance is self-evident. It aims to extract useful information from complex EEG signals for subsequent analysis and classification. EEG data is a typical non-stationary, strongly nonlinear data type, making its feature extraction process even more challenging.

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

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

[0006] Time-domain characteristics only reflect the change of signal amplitude over time and cannot characterize the frequency components of the signal. For example, periodic fault characteristics in vibration signals 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 properties (such as mean and variance) in the time domain, resulting in insufficient feature discrimination and affecting classification or diagnosis.

[0008] Most time-domain features (such as autocorrelation functions) assume that the signal is stationary, while real signals (such as mechanical vibrations and speech signals) are often non-stationary, causing the features to fail.

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

[0010] Fourier transform requires the signal to be stationary, but real-world signals (such as speech, vibration, and EEG signals) are usually non-stationary. While the Short Time Fourier Transform (STFT) can alleviate this problem, it is limited by a fixed time-frequency resolution.

[0011] The resolution of frequency domain analysis is limited by the signal length (Heisenberg uncertainty principle). For example, long signals can improve frequency resolution but reduce time resolution, while short signals can do the opposite.

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

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

[0014] Traditional features require manual selection or design (such as choosing which statistics or frequency bands to select), which requires a high level of domain knowledge and may miss important features.

[0015] Time-domain and frequency-domain features often need to be used in combination, resulting in high feature dimensionality and potential redundancy (such as the correlation between time-domain variance and frequency-domain energy). Dimensionality reduction is required, but this may result in information loss.

[0016] Traditional methods struggle to effectively characterize the complex properties of multi-component, non-stationary, and nonlinear signals.

[0017] The characteristics of hand-designed features may vary in different application scenarios, requiring readjustment for specific tasks, and lacking universality.

[0018] The fatigue risk warning and classification algorithm for smart safety helmet operations is essentially a multi-classification problem. Currently, this field mainly uses two technical methods to achieve this:

[0019] A common approach is to use a weighted average method, which adjusts the weights of data from various sensors and then fuses them to form a comprehensive assessment of the health status of workers. However, this method has two main drawbacks: First, in real-world scenarios, the relationships between features are not always linear, leading to inconsistencies between the comprehensive assessment model and actual rules, thus severely impacting the accuracy of the assessment results. Second, weights are hyperparameters that require manual selection, which often introduces significant errors. While hyperparameter optimization methods can be used to solve for the weights, this is not only time-consuming and labor-intensive but may also affect the timeliness of early warnings, making it counterproductive.

[0020] Another approach is to use deep learning methods for multi-class classification. However, in engineering scenarios, physical equipment resources are limited and the team lacks expertise in deep learning algorithms. Therefore, using deep learning technology may face the following drawbacks and disadvantages:

[0021] High training costs: Deep learning models (especially large models) require a lot of computing resources (GPU / TPU), and with limited equipment, training time may be too long or even impossible to complete.

[0022] Low inference efficiency: If the hardware performance is insufficient during deployment (e.g., only CPU), the model inference speed may not meet the real-time requirements, affecting the practicality of the project.

[0023] Energy consumption: High-load computing may place higher demands on infrastructure such as power supply and heat dissipation.

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

[0025] Debugging difficulties: Problems in model training (such as vanishing / exploding gradients, overfitting) require expert diagnosis, which inexperienced teams may not be able to solve effectively.

[0026] Relying on external support may require hiring experts or outsourcing development, increasing project costs and management complexity.

[0027] Data acquisition difficulties: Deep learning relies on a large amount of labeled data, but in engineering scenarios, environmental limitations (such as the difficulty in collecting data from industrial equipment) may lead to insufficient data.

[0028] Data annotation is costly: In the absence of automated annotation tools, manual annotation is time-consuming and labor-intensive, and requires domain knowledge support.

[0029] Poor data generalization: If the training data is inconsistent with the distribution of real-world scenarios (such as aging equipment or 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] Maintenance difficulties: The model needs to be continuously monitored and updated, but if the team lacks experience, it is difficult to deal with the problem of model performance degradation. Summary of the Invention

[0032] This invention provides a smart safety helmet operation fatigue risk warning and classification method based on fractional-order long short-term memory networks to solve the above-mentioned problems:

[0033] The present invention proposes a classification method for early warning of fatigue risk in smart safety helmet operations based on fractional-order long short-term memory networks. The classification method includes:

[0034] S1. Obtain the smart safety helmet wearing device, collect fatigue vital signs data of power workers, obtain fatigue vital signs data, classify the fatigue vital signs data according to the preset fatigue category, and obtain fatigue category data.

[0035] S2. The fatigue category data is subjected to feature extraction by fine composite multi-scale fractional-order back scattering entropy to obtain fatigue feature extraction data;

[0036] S3. Obtain three fatigue warning category labels, and input the fatigue feature extraction data and the three fatigue warning category labels into a fractional long short-term memory network for training. Use the novel Genghis Khan shark optimization algorithm to optimize the parameters of the training process and obtain network optimization data.

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

[0038] Further, S1 includes:

[0039] The fatigue signs data of power workers are collected by using a smart safety helmet wearable device. The collected fatigue signs data are preprocessed to obtain fatigue signs data.

[0040] Obtain preset fatigue categories, which include EEG, body temperature, heart rate, and blood oxygen saturation;

[0041] The fatigue symptoms data are classified by pre-defined fatigue categories to obtain fatigue category data.

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

[0043] Further, S2 includes:

[0044] Obtain a time series of a preset length, and calculate the scattering pattern and its probability through normal distribution function, linear transformation, phase space reconstruction, and then calculate the fractional RDE;

[0045] After granulation of the time series, RCMFrRDE is calculated to obtain the computational data.

[0046] Further, S3 includes:

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

[0048] Fractional-order Softplus activation function and fractional-order tanh activation function are set on the Long Short-Term Memory network to obtain a Long Short-Term Memory network with fractional-order Softplus activation function and fractional-order tanh activation function;

[0049] The setting of fractional-order Softplus activation function and numerical-order tanh activation function for the Long Short-Term Memory network includes:

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

[0051] Apply the fractional Softplus activation function to the current input x. t and the previously hidden state h t-1 The weighted sum is used to calculate the input gate activation i. t ;

[0052] Apply the fractional Softplus activation function to the current input x. t and the previously hidden state h t-1 The weighted sum is used to determine the forget gate activation f. t ;

[0053] Apply the fractional tanh activation function to the current input x. t and the previously hidden state h t-1 The weighted sum is used to obtain the unit state of the candidate.

[0054] Candidate cell status Previous unit state C t-i and the Gate of Oblivion t Combined, to update the cell state C t-i ;

[0055] Apply the fractional Softplus activation function to the current input x. t and the previously hidden state h t-1 The weighted sum determines the output gate activation O. t ;

[0056] Apply the fractional tanh activation function to the updated cell state C t Multiply by output gate O t Obtain the new hidden state h t The information to be sent to the next time step and used as the final output is stored along with the hidden state h. t middle;

[0057] Hidden state h t Used for the output of the fractional-order LSTM unit at the current time step;

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

[0059] Furthermore, S3 also includes:

[0060] In the novel Genghis Khan shark optimization algorithm, a random position is calculated by using the upper and lower bounds of the search space to obtain the optimal predation position;

[0061] A spiral search walking strategy is used to conduct a comprehensive data search of the target area, and a crazy operator is added to perform depth calculation to obtain the depth calculation results.

[0062] Furthermore, the classification system includes:

[0063] The data acquisition module is used to acquire fatigue vital sign data of power workers through the smart safety helmet wearable device, obtain fatigue vital sign data, classify the fatigue vital sign data according to preset fatigue categories, and obtain fatigue category data.

[0064] The feature extraction module is used to extract features from the fatigue category data through fine composite multi-scale fractional-order backscatter entropy to obtain fatigue feature extraction data.

[0065] The parameter optimization module is used to obtain three fatigue warning category labels. The fatigue feature extraction data is combined with the three fatigue warning category labels and input into the fractional long short-term memory network for training. The training process is optimized by a novel Genghis Khan shark optimization algorithm to obtain network optimization data.

[0066] The fatigue warning module is used to input the network optimization data into the Softmax function, and then output multi-class probability data to perform fatigue warning.

[0067] Furthermore, the data acquisition module includes:

[0068] The acquisition module is used to collect fatigue vital signs data of power workers through a smart safety helmet wearable device, preprocess the collected fatigue vital signs data, and obtain fatigue vital signs data.

[0069] Obtain preset fatigue categories, which include EEG, body temperature, heart rate, and blood oxygen saturation;

[0070] The classification module is used to classify the fatigue signs data according to preset fatigue categories to obtain fatigue category data;

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

[0072] Furthermore, the feature extraction module includes:

[0073] The RDE calculation module is used to obtain a time series of a preset length, and then calculate the fractional-order RDE by using the normal distribution function, linear transformation, phase space reconstruction, calculating the scattering pattern and its probability.

[0074] The phase granulation module is used to perform phase granulation on the time series and then calculate RCMFrRDE to obtain the calculated data.

[0075] Furthermore, the parameter optimization module includes:

[0076] The function acquisition module is used to obtain fractional-order Softplus activation functions and alpha-order tanh activation functions;

[0077] The network processing module is used to set fractional-order Softplus activation function and fractional-order tanh activation function for the Long Short-Term Memory network to obtain a Long Short-Term Memory network with fractional-order Softplus activation function and fractional-order tanh activation function;

[0078] The network processing module includes:

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

[0080] Apply the fractional Softplus activation function to the current input x. t and the previously hidden state h t-1 The weighted sum is used to calculate the input gate activation i. t ;

[0081] Apply the fractional Softplus activation function to the current input x. t and the previously hidden state h t-1 The weighted sum is used to determine the forget gate activation f. t ;

[0082] Apply the fractional tanh activation function to the current input x. t and the previously hidden state h t-1 The weighted sum is used to obtain the unit state of the candidate.

[0083] Candidate cell status Previous unit state C t-i and the Gate of Oblivion t Combined, to update the cell state C t-i ;

[0084] Apply the fractional Softplus activation function to the current input x. t and the previously hidden state h t-1 The weighted sum 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 Obtain the new hidden state h t The information to be sent to the next time step and used as the final output is stored along with the hidden state h. t middle;

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

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

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

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

[0090] The depth computing module is used to perform a comprehensive data search of the target area using a spiral search walking strategy, and incorporates a crazy operator to perform depth computing to obtain the depth computing results.

[0091] The beneficial effects of this invention are as follows: This invention utilizes sensors in a real-time fatigue monitoring helmet to collect fatigue status data such as electroencephalogram (EEG), body temperature, heart rate, and blood oxygen saturation of workers, and inputs this data into an intelligent classification method. By providing real-time feedback classification labels, i.e., status levels, fatigue status warnings are issued to workers, thereby effectively improving the safety and reliability of power production. Attached Figure Description

[0092] Figure 1Flowchart of a smart safety helmet operation fatigue risk warning and classification algorithm based on fractional-order long short-term memory network;

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

[0094] Figure 3 This is a diagram illustrating the fatigue warning algorithm structure for a smart safety helmet wearable device for power workers. Detailed Implementation

[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 for illustration and explanation only and are not intended to limit the present invention.

[0096] In one embodiment of the present invention, a fatigue risk warning and classification method for smart safety helmet operations based on fractional-order long short-term memory networks is proposed. The classification method includes:

[0097] S1. Obtain the smart safety helmet wearing device, collect fatigue vital signs data of power workers, obtain fatigue vital signs data, classify the fatigue vital signs data according to the preset fatigue category, and obtain fatigue category data.

[0098] S2. The fatigue category data is subjected to feature extraction by fine composite multi-scale fractional-order back scattering entropy to obtain fatigue feature extraction data;

[0099] S3. Obtain three fatigue warning category labels, and input the fatigue feature extraction data and the three fatigue warning category labels into a fractional long short-term memory network for training. Use the novel Genghis Khan shark optimization algorithm to optimize the parameters of the training process and obtain network optimization data.

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

[0101] The working principle of the above technical solution is as follows: Safe production is the most important prerequisite for power operations. The State Grid's production departments operate in complex environments, including working at heights and proximity to high-voltage equipment, posing significant safety risks. Accidents can lead to serious injuries or even death. Therefore, monitoring the health status of workers is crucial in power operations. Due to environmental and terrain limitations at work sites, communication signals are restricted, precise positioning is impossible, and medical emergency response is slow. Improving the safety of employees working in hazardous environments has become an urgent need for the power industry. During power operations, workers can obtain detailed information about their current status through smart safety helmets, enabling rapid analysis and early warning of their physical characteristics. This facilitates work tasks and ensures the reliability and efficiency of safety and dispatch command.

[0102] The main idea of ​​this invention is to comprehensively extract features from field data using refined composite multi-scale fractional-order backscatter entropy, then input the obtained features into a fractional-order long short-term memory network for multi-class classification of health warnings, and finally use a novel Genghis Khan shark optimization algorithm to optimize the network parameters, such as... Figure 1 As shown.

[0103] Feature extraction, a core technology in smart helmet data processing, aims to sift through massive amounts of raw monitoring data to provide a reliable basis for subsequent health status classification. Currently, mainstream feature extraction methods can be categorized into the following three dimensions:

[0104] Time-domain analysis, based on statistical principles, visually reflects the distribution and dynamic characteristics of signal amplitude by calculating statistical measures such as mean, variance, skewness, kurtosis, and interquartile range. This method boasts advantages such as high computational efficiency and strong real-time performance, making it particularly suitable for industrial scenarios requiring rapid response. However, due to the inherent limitations of time-domain analysis, it struggles to effectively capture implicit features within the signal.

[0105] Using Fourier transform and its derivative algorithms (such as short-time Fourier transform) to convert time-domain signals into frequency-domain representations can clearly reveal the frequency composition and energy distribution characteristics of the signal. However, while emphasizing frequency-domain characteristics, these methods suffer from the loss of time-domain information and have limited analytical capabilities for non-stationary signals.

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

[0107] Work fatigue risk warning, as a core function of intelligent safety helmet systems, is essentially a typical multi-classification problem. Current mainstream industry solutions have significant technical limitations, specifically as follows:

[0108] The current weighted average method fuses data by manually setting feature weights. Its main technical shortcomings are: First, the linear assumption fails, as complex nonlinear relationships exist between features in real-world scenarios, making it difficult for linear weighted models to accurately map the true decision boundary. Second, there is the parameter optimization paradox: weights, as hyperparameters, must be determined using time-consuming methods such as grid search, fundamentally conflicting with real-time early warning requirements. Third, dynamic adaptation is lacking; a fixed-weight system cannot respond to changes in the operating environment, leading to a continuous decline in model robustness.

[0109] While deep neural networks possess powerful representational capabilities in theory, they face multiple implementation obstacles in industrial scenarios: computational resource constraints, requiring GPU clusters for training; CPU inference latency exceeding 500ms during inference, failing to meet sub-second response standards; high energy consumption and limited device battery life; severe data sparsity issues in data acquisition; annotation costs: time-consuming single-sample annotation with insufficient domain expert participation; feature distribution mismatch between measured data and training set; poor model interpretability, with the black-box nature of the decision-making process leading to decreased security verification pass rates; high maintenance complexity and strong reliance on talent, requiring algorithm teams to possess knowledge in both signal processing and deep learning.

[0110] Existing technologies have limitations: Traditional time-frequency domain analysis methods (such as Fourier transform and wavelet analysis) are significantly insufficient in capturing deep, implicit features of data, and their linear processing paradigms are unable to effectively characterize nonlinear features under complex operating conditions. Although entropy theory (sample entropy, fuzzy entropy, etc.) has made breakthroughs in feature quantification, existing entropy analysis methods have two key shortcomings: first, they lack the ability to quantify the dynamic evolution features of fractional-order time series; second, fixed-scale calculation models cannot adapt to the non-stationary characteristics of equipment operating conditions.

[0111] Conventional weighted average methods, due to their pre-defined fixed weight allocation mechanism, are prone to information loss during feature fusion. While deep learning models possess powerful feature learning capabilities, they face three major application obstacles: generally high model parameter count leading to high energy consumption, reliance on large-scale labeled datasets, and complex model structures making real-time inference difficult on embedded devices.

[0112] Parameter optimization of neural networks is essentially a problem of finding the optimal solution in a high-dimensional non-convex space, and its loss function surface contains a large number of local extrema. Gradient descent-based optimization algorithms have two inherent drawbacks: first, the learning rate decay strategy easily causes the parameters to get stuck in a flat region, leading to premature convergence; second, the ill-conditioned number of the Hessian matrix severely restricts the practical application of second-order optimization methods.

[0113] The technical advantages of the above solution are as follows: Addressing the shortcomings of existing traditional time-frequency domain analysis methods and traditional entropy theory, this invention employs refined composite multi-scale fractional-order backscattered entropy for feature extraction from time-series data such as EEG, body temperature, heart rate, and blood oxygen saturation. Through multi-scale decomposition, it can capture the signal complexity characteristics at different time scales, thus more comprehensively reflecting the dynamic changes of the signal. The introduction of fractional order increases the model's flexibility and adaptability, enabling better handling of nonlinear and non-stationary signals. Backscattered entropy itself has good noise resistance, effectively suppressing the impact of noise on signal complexity measurement. The refined composite multi-scale structure further improves the stability of complexity measurement, maintaining good performance under different signal-to-noise ratios.

[0114] This invention proposes a fractional LSTM algorithm for multi-classification. It replaces the traditional LSTM algorithm's Sigmoid and tanh activation functions with fractional Softplus and fractional tanh activation functions.

[0115] The fractional-order Softplus activation function inherits the smoothness of the Softplus function, being continuously differentiable across its entire domain. This smoothness allows neural networks to avoid gradient vanishing or exploding during backpropagation, thus improving training stability and convergence speed. The fractional-order Softplus activation function does not output zero in negative input regions, preventing "dead neurons." This allows the network to remain more active during training, preventing some neurons from stopping learning due to zero gradients. Furthermore, by introducing the concept of fractional calculus, it can better adapt to different data distributions and network structures. This adaptability makes it excellent at handling complex nonlinear problems, especially in scenarios requiring flexible adjustment of the activation function shape. Computational efficiency can be improved through polynomial approximation. This approximation method not only reduces computational complexity but also improves the model's generalization ability while maintaining the nonlinearity of the activation function. Its design is based on the theoretical framework of fractional calculus, better satisfying the conditions of the general approximation theorem. This means it can approximate complex functions more efficiently, thus performing well in deep learning tasks.

[0116] The fractional tanh activation function inherits the zero-centeredness property of the traditional tanh function, meaning the output values ​​are symmetrically distributed around the zero point. This property maintains gradient balance during training, avoiding training instability caused by output offset. The traditional tanh function faces the vanishing gradient problem when the input values ​​are large or small, but the fractional tanh activation function, through the introduction of fractional calculus, can alleviate this problem to some extent, thereby improving the training efficiency of the network. This activation function maintains the symmetry of the tanh function, effectively handling positive and negative input values, while introducing nonlinearity, enabling the network to learn complex patterns and relationships. The introduction of fractional calculus makes it better suited to complex input data distributions, especially performing well when dealing with data with nonlinear characteristics. Due to its zero-centeredness and symmetry, the fractional tanh activation function is particularly suitable for recurrent neural networks (such as LSTM), providing balanced input for subsequent time steps and improving the learning of long-term dependencies. These improvements give the fractional tanh activation function a greater advantage in handling complex tasks.

[0117] Fractional activation functions can capture more complex nonlinear relationships than traditional integer activation functions, thus exhibiting higher accuracy when processing complex dynamic systems and time series data. Fractional activation functions can achieve faster convergence speeds and shorter training times. By introducing fractional parameters, fractional activation functions increase the model's flexibility and adaptability, resulting in better generalization ability when handling complex datasets. Fractional calculus has memory capabilities, enabling it to better simulate dynamic systems with memory characteristics. Introducing fractional activation functions into LSTM models allows them to more effectively capture long-term dependencies, further enhancing their ability to model time series data.

[0118] The parameter optimization process of the fractional-order LSTM algorithm is equivalent to solving a strongly nonlinear, nonconvex optimization problem. Traditional machine learning optimization methods that rely on gradient information often fall into the trap of local optima, failing to achieve global optimization. This invention, based on the Genghis Khan Shark Optimization Algorithm, enhances its global optimization capability by introducing a spiral search walk strategy and a crazy operator.

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

[0120] The fatigue signs data of power workers are collected by using a smart safety helmet wearable device. The collected fatigue signs data are preprocessed to obtain fatigue signs data.

[0121] Obtain preset fatigue categories, which include EEG, body temperature, heart rate, and blood oxygen saturation;

[0122] The fatigue symptoms data are classified by pre-defined fatigue categories to obtain fatigue category data.

[0123] The fatigue category data includes EEG data, body temperature data, heart rate data, and blood 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 assessment. As a novel nonlinear dynamic analysis method, RDE is based on Permutation Entropy (PE) and combines the advantages of Dispersion Entropy (DE) and Reverse Permutation Entropy (RPE), exhibiting good stability and discriminative power.

[0125] For a time series X = {x(1), x(2), ..., x(T)} of length T, the steps for calculating the fractional 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 Y = {y(i), i = 1, 2, ..., T}, where y(i) ranges from 0 to 1, and μ is the expectation and σ is the mean. 2 Let Variance be the variance.

[0129] (2) Through linear transformation:

[0130]

[0131] Map the new sequence Y = {y(i), i = 1, 2, ..., T} to the range Z = {z(i), i = 1, 2, ..., T}, where int is the floor 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 time delay d and embedding dimension 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 to a scattering pattern π. The total number of possible scattering patterns is c. m There are several types, 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 ) represents the ratio of the number of scattering patterns to the number of embedding vectors; Number{π i} refers to the number of embedding vectors mapped to the i-th scattering pattern.

[0142] (6) Calculate the fractional RDE:

[0143]

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

[0145] Fine-grained composite multi-scale fractional-order inverse scattering entropy:

[0146] FrRDE has significant 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, multi-scale analysis methods have emerged. However, multi-scale calculations based on coarse-grained processes depend on the length of the time series and do not consider the relationships between the newly calculated coarse-grained sequences, leading to information loss. To address these issues, RCMFrRDE is proposed, combining a refined composite multi-scale method. The calculation steps are as follows:

[0147] (1) Coarsening treatment:

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

[0149]

[0150] (2) Calculate RCMFrRDE:

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

[0152]

[0153] in Let be the i-th scattering pattern under the scaling factor τ.

[0154] For each scale factor τ, the RCMFrRDE is defined as follows:

[0155]

[0156] in

[0157] The technical effects of the above solutions are as follows: FRDE combines the advantages of scatter entropy and reverse permutation entropy, exhibiting good stability and discriminative power, thus more accurately assessing signal complexity. This is particularly important for applications such as fault diagnosis, as it can help identify subtle signal changes. Through phase space reconstruction and scatter pattern calculation, FRDE can capture the nonlinear dynamic characteristics in time series. This is significant for understanding the dynamic behavior of systems, predicting future system states, and identifying potential system faults. RCMFrRDE, by combining a refined composite multi-scale method, overcomes the shortcomings of traditional multi-scale analysis methods in terms of information loss. It can measure the complexity of time series at different scales, thus providing a more comprehensive understanding of the dynamic behavior of systems. This is particularly useful for analyzing complex systems with multi-scale characteristics. Both FRDE and RCMFrRDE employ preprocessing steps such as normal distribution functions and linear transformations, which enhance the robustness of the algorithms, enabling them to handle time series of different types and characteristics. Furthermore, these algorithms are applicable to various practical applications, such as biomedical signal processing and mechanical fault diagnosis. FRDE and RCMFrRDE provide new and effective tools for nonlinear dynamic analysis. These methods can reveal hidden information and dynamic characteristics in time series, providing researchers with deeper understanding and analytical tools. This is of great significance for promoting the development of the field of nonlinear dynamics. This method has demonstrated significant technical effects in signal processing and nonlinear dynamics analysis, making important contributions to improving the accuracy of signal complexity assessment, enhancing the ability to capture the dynamic characteristics of time series, providing multi-scale analysis capabilities, improving the robustness and applicability of the algorithm, and providing new tools for nonlinear dynamics analysis.

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

[0159] Obtain a time series of a preset length, and calculate the scattering pattern and its probability through normal distribution function, linear transformation, phase space reconstruction, and then calculate the fractional RDE;

[0160] After granulation of the time series, RCMFrRDE is calculated to obtain the computational data.

[0161] The working principle of the above technical solution is as follows: The fractional-order Softplus activation function includes the fractional-order Softplus (FracSoftplus). The sigmoid function is generalized using the softplus function through the fractional derivative, and it is defined as follows:

[0162]

[0163] Then the expression is calculated as follows:

[0164]

[0165] Where a i These are trainable parameters.

[0166] Fractional tanh activation function:

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

[0168]

[0169] Then the expression is calculated as follows:

[0170]

[0171] Where a i These are trainable parameters.

[0172] The technical effects of the above-mentioned solution are as follows: The fractional-order Softplus activation function is an activation function based on fractional calculus. It inherits the smoothness of the Softplus function and is continuously differentiable throughout its domain. This smoothness allows the neural network to avoid the problems of vanishing or exploding gradients during backpropagation, thereby improving training stability and convergence speed. By introducing the concept of fractional calculus, fractional-order Softplus can better adapt to different data distributions and network structures. This adaptability makes it perform well in handling complex nonlinear problems, especially in scenarios where the shape of the activation function needs to be flexibly adjusted. In some cases, fractional-order Softplus can improve computational efficiency through polynomial approximation. This approximation method not only reduces computational complexity but also improves the model's generalization ability while maintaining the nonlinearity of the activation function. The design of fractional-order Softplus is based on the theoretical framework of fractional calculus, which can better satisfy the conditions of the general approximation theorem. The gradient of fractional-order Softplus is smoother, avoiding the problem of excessive gradient changes at certain points in Softplus, resulting in a more stable gradient flow. Fractional-order Softplus introduces richer nonlinear characteristics through fractional derivatives, enabling the model to learn more complex patterns. Overall, it not only improves computational efficiency and numerical stability but also enhances gradient and nonlinear characteristics, resulting in superior model performance. These improvements give Fractional Softplus a greater advantage in handling complex tasks and deep networks. The fractional tanh activation function inherits the zero-centeredness property of the traditional tanh function, meaning the output values ​​are symmetrically distributed around the zero point. This property maintains gradient balance during training, avoiding training instability caused by output offset. The traditional tanh function faces the problem of gradient vanishing when the input values ​​are large or small, while the fractional tanh activation function, through the introduction of fractional calculus, can alleviate this problem to some extent, thereby improving the training efficiency of the network. This activation function maintains the symmetry of the tanh function, effectively handling positive and negative input values, while introducing nonlinear characteristics, enabling the network to learn complex patterns and relationships. The introduction of fractional calculus makes it better suited to complex input data distributions, especially performing well when dealing with data with nonlinear features. 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 input for subsequent time steps and improving the learning of long-term dependencies. These improvements give the fractional tanh activation function a greater advantage when handling complex tasks.

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

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

[0175] Fractional-order Softplus activation function and fractional-order tanh activation function are set on the Long Short-Term Memory network to obtain a Long Short-Term Memory network with fractional-order Softplus activation function and fractional-order tanh activation function;

[0176] The setting of fractional-order Softplus activation function and numerical-order tanh activation function for the Long Short-Term Memory network includes:

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

[0178] The fractional-order Softplus activation function is applied to the weighted sum of the current input xt and the previous hidden state ht-1 to compute the input gate activation it;

[0179] The fractional-order Softplus activation function is applied to the weighted sum of the current input xt and the previous hidden state ht-1 to 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] The fractional-order tanh activation function is applied to the weighted sum of the current input xt and the previous hidden state ht-1 to obtain the cell state of the candidate. At this stage, new potential values ​​for cell states are calculated;

[0181] Candidate cell status The previous cell state Ct-i is combined with the forget gate ft to update the cell state Ct-i; at this stage, the data to be retained in the cell state is determined.

[0182] The fractional-order Softplus activation function is applied 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-order tanh activation function to the updated cell state Ct and multiply it by the output gate Ot to obtain a new hidden state ht; store the information sent to the next time step and used as the final output in this hidden state ht.

[0184] The hidden state ht is used for the output of the fractional-order LSTM unit 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 high-level description of the fractional 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 using zero or small random values.

[0188] 2. By applying the fractional-order Softplus activation function to the current input x t and the previously hidden state h t-1 The weighted sum is used to calculate the input gate activation i. 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 current input x. t and the previously hidden state h t-1 The weighted sum is used to determine the forget gate activation f. t Cell states C that should be forgotten or maintained t-1 The information 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 tanh activation function to the current input x. t and the previously hidden state h t-1 The weighted sum is used to obtain the unit state of the candidate. At this stage, new potential values ​​for cell states are calculated.

[0193]

[0194] 5. Candidate cell status Previous unit state C t-1 and the Gate of Oblivion t They are combined to update the cell state C t-1 At this stage, the data that should be retained in the cell state is determined.

[0195]

[0196] 6. Apply the fractional-order Softplus activation function to the current input x. t and the previously hidden state h t-1 The weighted sum is used to determine the output gate activation O. t It should be treated as a hidden state h. t The generated cell state information 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 Multiply by output gate O t To obtain the new hidden state h t Information that will be sent to the next time step and may be used as the final output is contained in this hidden state.

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

[0200] 8. Hidden state h t It can be used as the output of a fractional-order LSTM unit at the current time step.

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

[0202] Among them, i t It is input gate activation, σ FS It is a fractional-order Softplus activation function, w i It is the weight matrix of the input gate, x t This is the current input, h t-1 It is the previous hidden state, b i It is the bias of the input gate, f t It's the Forgotten Gate activated, w f This is the weight matrix of the forget gate, b f It's a bias in the forgetting gate. It is the candidate unit state, b c It is the bias of the candidate unit state, Frtanh is the fractional tanh activation function, w c C is the weight matrix of the candidate unit states. t It is the updated unit state, C t-1 It is the previous unit state, O tIt's output gate activation, w o It is the weight matrix of the output gate, b o It is the output gate bias, h t It is a new hidden state.

[0203] Cross-entropy loss is one of the most commonly used loss functions in multi-class classification tasks. It measures the difference between the probability distribution of the model's output and the true labels. For multi-class problems, the categorical cross-entropy loss function is typically used, and its mathematical expression is:

[0204]

[0205] Among them, y true,i It is the One-Hot encoding of the real label, y pred,i This represents the probability distribution predicted by the model. In fractional LSTM models, the cross-entropy loss function effectively guides the model to learn how to correctly classify input sequences into different categories.

[0206] The technical effects of the above solutions are as follows: Fractional-order Softplus and fractional-order tanh activation functions introduce the concept of fractional calculus, enabling LSTM units to capture more complex nonlinear relationships. This enhances the model's expressive power, allowing it to better adapt to and process various 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 exhibit better performance and stability when processing unseen data. LSTM networks inherently possess long-term memory capabilities, and the introduction of fractional-order activation functions can further optimize its memory mechanism. By adjusting the fractional-order parameters, the model's sensitivity to historical information and the rate of forgetting can be controlled, thus better balancing 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 exhibit stronger robustness when dealing with 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 according to specific tasks and data characteristics. This allows for finding the optimal model configuration in different application scenarios. Introducing fractional calculus into deep learning provides new ideas and methods for related research. This can not only promote the development of recurrent neural networks such as LSTM, but also potentially have a positive impact on other types of neural networks and machine learning algorithms. Applying fractional Softplus activation functions and fractional tanh activation functions to Long Short-Term Memory networks can significantly improve the model's nonlinear expressive power, generalization ability, memory capacity, robustness, and provide flexible model adjustment methods.

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

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

[0209] A spiral search walking strategy is used to conduct a comprehensive data search of the target area, and a crazy operator is added to perform depth calculation to obtain the depth calculation results.

[0210] The working principle of the above technical solution is as follows: Genghis Khan sharks are apex predators in freshwater river areas. They typically roam near the bottom, seemingly on patrol. In reality, they do this to ensure that there are no predators stronger than themselves in the area. In this way, Genghis Khan sharks can not only ensure their own safety but also expand their search range to find the best prey location. GKS do not suddenly attack their prey, but once they lock onto it, they wait for the optimal moment to launch a full-scale attack.

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

[0212]

[0213] Among them, X i j (t+1) represents the value of the position of the i-th member in the j-th dimension at time t+1, ub j and lb j Let r1 and r2 represent the upper and lower bounds on the j-th dimension, respectively. r1 is a random number in the interval [0,1]. N represents the population size, D represents the problem dimension, it represents the current iteration number, and T represents the total number of iterations.

[0214] To capture higher-quality prey, the Genghis Khan shark employs a strategy with a certain probability: relying on its keen sense of smell to continuously approach the optimal prey. It also employs a strategy of spiraling and searching around the optimal prey. The specific mathematical model for this behavior is as follows:

[0215]

[0216] in, Let represent the value of the best predation position known at time t in the j-th dimension, and s represent the olfactory intensity of the Genghis Khan shark as it moves toward 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 in the interval [0,1], representing the degree to which the search agent absorbs the scent of its prey.

[0220] The spiral search walking strategy is a search technique that simulates the spiral motion in nature. Its core idea is to achieve comprehensive coverage of the target area by continuously adjusting the search direction. η is a random number in the interval [0,1], and is selected from two strategies by comparing it with 0.5. Parameter k is the coefficient of variation, and parameter l is a random number uniformly distributed in the interval [-1,1]. T is the current iteration number. max This represents the maximum number of iterations.

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

[0222]

[0223] During hunting, fish employ many unexpected methods and strategies to express themselves. The Genghis Khan shark, after approaching its optimal prey, uses the prey as a reference point, leveraging its speed advantage to quickly swim in front of the prey and deliver a fatal blow to the prey's head with its massive jaws; the entire process follows a parabolic trajectory. To overcome premature convergence in algorithms, this invention incorporates a crazy operator. The specific formula is given below:

[0224]

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

[0226]

[0227] Where |ω(t+1)| is the weighting 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] Where ξ is a random number in the interval [0,1], Θ is the given probability of insanity, and κ is a very small constant.

[0231] Over the long course of evolution in their natural environment, freshwater sharks of the same genus have developed certain reflexive behaviors to adapt to their surroundings. Color change is usually influenced by external factors. During foraging, Genghis Khan sharks frequently encounter predators that threaten their safety or competitors for prey. Therefore, to escape these predators, Genghis Khan sharks possess a color-changing mechanism similar to that of squid. When startled, their tails and bodies lighten in color, thus frightening away predators and causing them to flee quickly.

[0232] To mathematically simulate this behavior of the Genghis Khan shark, we conducted this process in the fourth stage of the novel GKSO, and 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, calculated using the following formula:

[0235]

[0236] Where rand is a random number in the interval [0,1], and l1 is a binary parameter that takes the value 0 or 1. Furthermore, ρ is an adaptive coefficient, calculated as follows:

[0237]

[0238] Where, β min and β max We take values ​​of 0.2 and 1.2 respectively. Additionally, a randomly generated solution... and The formulas for calculating t at two times are as follows:

[0239]

[0240] Calculate the solution The formula is as follows:

[0241]

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

[0243] The technical effects of the above solution are as follows: By simulating the behavior of a Genghis Khan shark swimming near the bottom of the water to find the best prey location, the algorithm can explore extensively within the search space, effectively avoiding premature convergence and improving global search capabilities. The algorithm introduces two strategies: relying on smell to approach the best prey and spiral search walking. These two strategies can be flexibly adjusted according to the current search state, making the search process more efficient and enabling rapid location near the optimal solution. The introduction of parabolic foraging behavior allows the algorithm to utilize its speed advantage for fine-grained searching when approaching the optimal solution, thereby improving local exploitation capabilities and finding higher-quality optimal solutions. By adding a crazy operator, the algorithm can maintain a certain degree of perturbation during the search process, effectively avoiding getting trapped in local optima, further improving the algorithm's search efficiency and robustness. Several parameters in the algorithm (such as olfactory intensity, coefficient of variation, and activity step size) can be adjusted according to the characteristics of the optimization problem, enabling the algorithm to adapt to different types of optimization problems. The introduction of a self-protection mechanism allows the algorithm to simulate the color-changing behavior of a Genghis Khan shark during the search process, thereby avoiding interference from other search agents to a certain extent and improving the algorithm's stability. The algorithm's mathematical model is clear and straightforward, facilitating implementation and expansion. Furthermore, each step and parameter in the algorithm has a clear physical meaning and mathematical expression, aiding in theoretical analysis and experimental verification. The novel swarm intelligence optimization algorithm (GKSO) based on the natural behavior of the Genghis Khan shark demonstrates significant technical advantages in global search capability, search efficiency, local exploitation capability, overcoming premature convergence, adaptability, stability, and ease of implementation. These advantages give GKSO considerable potential and advantages in solving complex optimization problems.

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

[0245] The data acquisition module is used to acquire fatigue vital sign data of power workers through the smart safety helmet wearable device, obtain fatigue vital sign data, classify the fatigue vital sign data according to preset fatigue categories, and obtain fatigue category data.

[0246] The feature extraction module is used to extract features from the fatigue category data through fine composite multi-scale fractional-order backscatter entropy to obtain fatigue feature extraction data.

[0247] The parameter optimization module is used to obtain three fatigue warning category labels. The fatigue feature extraction data is combined with the three fatigue warning category labels and input into the fractional long short-term memory network for training. The training process is optimized by a novel 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 to perform fatigue warning.

[0249] The working principle of the above technical solution is as follows: Safe production is the most important prerequisite for power operations. The State Grid's production departments operate in complex environments, including working at heights and proximity to high-voltage equipment, posing significant safety risks. Accidents can lead to serious injuries or even death. Therefore, monitoring the health status of workers is crucial in power operations. Due to environmental and terrain limitations at work sites, communication signals are restricted, precise positioning is impossible, and medical emergency response is slow. Improving the safety of employees working in hazardous environments has become an urgent need for the power industry. During power operations, workers can obtain detailed information about their current status through smart safety helmets, enabling rapid analysis and early warning of their physical characteristics. This facilitates work tasks and ensures the reliability and efficiency of safety and dispatch command.

[0250] The main idea of ​​this invention is to comprehensively extract features from field data using refined composite multi-scale fractional-order backscatter entropy, then input the obtained features into a fractional-order long short-term memory network for multi-class classification of health warnings, and finally use a novel Genghis Khan shark optimization algorithm to optimize the network parameters, such as... Figure 1 As shown.

[0251] The fatigue warning algorithm for smart safety helmets worn by power workers should train a fractional-order long short-term memory network with four feature data extracted using a refined composite multi-scale fractional-order backscatter entropy method (EEG, body temperature, heart rate, and blood oxygen saturation) and three fatigue warning category labels (normal, fatigued, and high-risk). This network consists of eight cascaded fractional-order LSTM neurons and is connected to a Softmax function to output multi-class probabilities (predicted labels). Its network structure is as follows: Figure 3 As shown.

[0252] The technical advantages of the above solution are as follows: Addressing the shortcomings of existing traditional time-frequency domain analysis methods and traditional entropy theory, this invention employs refined composite multi-scale fractional-order backscattered entropy for feature extraction from time-series data such as EEG, body temperature, heart rate, and blood oxygen saturation. Through multi-scale decomposition, it can capture the signal complexity characteristics at different time scales, thus more comprehensively reflecting the dynamic changes of the signal. The introduction of fractional order increases the model's flexibility and adaptability, enabling better handling of nonlinear and non-stationary signals. Backscattered entropy itself has good noise resistance, effectively suppressing the impact of noise on signal complexity measurement. The refined composite multi-scale structure further improves the stability of complexity measurement, maintaining good performance under different signal-to-noise ratios.

[0253] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A classification method for early warning of fatigue risk in smart safety helmet operations based on fractional-order long short-term memory networks, characterized in that, The classification method includes: S1. Obtain the smart safety helmet wearing device, collect fatigue vital signs data of power workers, obtain fatigue vital signs data, classify the fatigue vital signs data according to the preset fatigue category, and obtain fatigue category data. S2. The fatigue category data is subjected to feature extraction by fine composite multi-scale fractional-order back scattering entropy to obtain fatigue feature extraction data; S3. Obtain three fatigue warning category labels, and input the fatigue feature extraction data and the three fatigue warning category labels into a fractional long short-term memory network for training. Use the novel Genghis Khan shark optimization algorithm to optimize the parameters of the training process and obtain network optimization data. S4. Input the network optimization data into the Softmax function, and then output multi-class probability data for fatigue warning; The fractional-order backscattering entropy of fine composite multiscale is obtained by combining FrRDE with the fine composite multiscale method; The calculation steps for the fine composite multi-scale fractional-order backscatter entropy are as follows: Coarse-graining treatment: For the original time series Its scale factor is The first time coarse-grained sequences It is given by the following formula: ； Calculate RCMFrRDE: ； in scale factor Under the conditions A distribution pattern; For each scale factor The definition of RCMFrRDE is: ； in ; in, Let m be the number of categories, d be the number of time delays (L), and α be the fractional order. Wherein, S3 includes: Obtain the fractional-order Softplus activation function and the alpha-order tanh activation function; Fractional-order Softplus activation function and fractional-order tanh activation function are applied to the Long Short-Term Memory (LSTM) network to obtain an LSM network with fractional-order Softplus activation function and fractional-order tanh activation function; The setting of fractional-order Softplus activation function and numerical-order tanh activation function for the Long Short-Term Memory network includes: Initialize the fractional-order LSTM cell state C and hidden state h with zero or small random values; Apply the fractional Softplus activation function to the current input. and the previously hidden state The weighted sum is used to calculate the input gate activation. ; Apply the fractional Softplus activation function to the current input. and the previously hidden state The weighted sum is used to determine the activation of the forget gate. ; Apply the fractional tanh activation function to the current input. and the previously hidden state The weighted sum is used to obtain the unit state of the candidate. ; Candidate cell status Previous unit state And the Gate of Oblivion Combine them to update the cell state. ; Apply the fractional Softplus activation function to the current input. and the previously hidden state The weighted sum determines the output gate activation. ; Apply the fractional tanh activation function to the updated cell states. Multiply by the output gate Gain a new hidden state ; The information stored in the hidden state will be sent to the next time step and used as the final output. middle; Hidden state Used for the output of the fractional-order LSTM unit at the current time step; Repeat the operation for each time step in the sequence.

2. The intelligent safety helmet operation fatigue risk early warning and classification method based on fractional-order long short-term memory network according to claim 1, characterized in that, S1 includes: The fatigue signs data of power workers are collected by using a smart safety helmet wearable device. The collected fatigue signs data are preprocessed to obtain fatigue signs data. Obtain preset fatigue categories, which include EEG, body temperature, heart rate, and blood oxygen saturation; The fatigue symptoms data are classified by pre-defined fatigue categories to obtain fatigue category data. The fatigue category data includes EEG data, body temperature data, heart rate data, and blood saturation data.

3. The intelligent safety helmet operation fatigue risk early warning and classification method based on fractional-order long short-term memory network according to claim 1, characterized in that, S2 includes: Obtain a time series of a preset length, and calculate the scattering pattern and its probability through normal distribution function, linear transformation, phase space reconstruction, and then calculate the fractional RDE; After granulation of the time series, RCMFrRDE is calculated to obtain the computational data.

4. The intelligent safety helmet operation fatigue risk early warning and classification method based on fractional-order long short-term memory network according to claim 1, characterized in that, S3 further includes: In the novel Genghis Khan shark optimization algorithm, a random position is calculated by using the upper and lower bounds of the search space to obtain the optimal predation position; A spiral search walking strategy is used to conduct a comprehensive data search of the target area, and a crazy operator is added to perform depth calculation to obtain the depth calculation results.

5. A classification system for a smart safety helmet operation fatigue risk early warning classification method based on fractional-order long short-term memory networks as described in claim 1, characterized in that, The classification system includes: The data acquisition module is used to acquire fatigue vital sign data of power workers through the smart safety helmet wearable device, obtain fatigue vital sign data, classify the fatigue vital sign data according to preset fatigue categories, and obtain fatigue category data. The feature extraction module is used to extract features from the fatigue category data through fine composite multi-scale fractional-order backscatter entropy to obtain fatigue feature extraction data. The parameter optimization module is used to obtain three fatigue warning category labels. The fatigue feature extraction data is combined with the three fatigue warning category labels and input into the fractional long short-term memory network for training. The training process is optimized by a 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-class probability data to perform fatigue warning. RCMFrRDE was obtained by combining FrRDE with a fine composite multiscale method; The fractional-order backscattering entropy of fine composite multiscale is obtained by combining FrRDE with the fine composite multiscale method; The calculation steps for the fine composite multi-scale fractional-order backscatter entropy are as follows: Coarse-graining treatment: For the original time series Its scale factor is The first time coarse-grained sequences It is given by the following formula: ； Calculate RCMFrRDE: ； in scale factor Under the conditions A distribution pattern; For each scale factor The definition of RCMFrRDE is: ； in ; in, Let m be the number of categories, d be the number of time delays (L), and α be the fractional order. The parameter optimization module includes: The function acquisition module is used to obtain fractional-order Softplus activation functions and alpha-order tanh activation functions; 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, so as to obtain the Long Short-Term Memory network with the fractional-order Softplus activation function and the fractional-order tanh activation function; The network processing module includes: Initialize the fractional-order LSTM cell state C and hidden state h with zero or small random values; Apply the fractional Softplus activation function to the current input. and the previously hidden state The weighted sum is used to calculate the input gate activation. ; Apply the fractional Softplus activation function to the current input. and the previously hidden state The weighted sum is used to determine the activation of the forget gate. ; Apply the fractional tanh activation function to the current input. and the previously hidden state The weighted sum is used to obtain the unit state of the candidate. ; Candidate cell status Previous unit state And the Gate of Oblivion Combine them to update the cell state. ; Apply the fractional Softplus activation function to the current input. and the previously hidden state The weighted sum determines the output gate activation. ; Apply the fractional tanh activation function to the updated cell states. Multiply by the output gate Gain new hidden state ; The information stored in the hidden state will be sent to the next time step and used as the final output. middle; Hidden state Used for the output of the fractional-order LSTM unit at the current time step; Repeat the operation for each time step in the sequence.

6. The classification system according to claim 5, characterized in that, The data acquisition module includes: The acquisition module is used to collect fatigue vital signs data of power workers through a smart safety helmet wearable device, preprocess the collected fatigue vital signs data, and obtain fatigue vital signs data. Obtain preset fatigue categories, which include EEG, body temperature, heart rate, and blood oxygen saturation; The classification module is used to classify the fatigue signs data according to preset fatigue categories to obtain fatigue category data; The fatigue category data includes EEG data, body temperature data, heart rate data, and blood saturation data.

7. The classification system according to claim 5, characterized in that, The feature extraction module includes: The RDE calculation module is used to obtain a time series of a preset length, and then calculate the fractional-order RDE by using the normal distribution function, linear transformation, phase space reconstruction, calculating the scattering pattern and its probability. The phase granulation module is used to perform phase granulation on the time series and then calculate RCMFrRDE to obtain the calculated data.

8. The classification system according to claim 5, characterized in that, The parameter optimization module further includes: The location acquisition module is used in the new Genghis Khan shark optimization algorithm to calculate a random location by the upper and lower bounds of the search space and obtain the best predation location. The depth computing module is used to perform a comprehensive data search of the target area using a spiral search walking strategy, and incorporates a crazy operator to perform depth computing to obtain the depth computing results.

Images