Health early warning method, system and equipment for operating personnel and medium

Abstract

The invention discloses a health early warning method, system, equipment and medium for an operator, and relates to the technical field of health monitoring, and the method comprises the steps: collecting various data of wearable equipment of the operator, and extracting the features of the various data; inputting the features into a t distribution probability neural network, performing activation function operation through a t distribution probability density function represented by Euclidean distance, adding operation results according to classes, judging the added classes, and taking the class with the maximum probability value in judgment results as an output result; and outputting the output result as a health early warning category, and carrying out health early warning on the operator. According to the method, the extreme condition of the features can be processed through t distribution, so that statistical inference is more stable, the probability that the result falls into the local extreme value is reduced through global optimization, optimization of the t distribution probability neural network model is realized, and the accuracy of the subsequent output result is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of health monitoring, and in particular to a health early warning method, system, equipment and medium for operators. Background Art

[0002] Safety in production is the most important prerequisite for power operations. The working environment of the State Grid production department is complex, with high-altitude operations and proximity to high-voltage equipment, which poses great safety risks. Safety accidents may cause serious injuries or even life-threatening situations. Therefore, in power operations, it is particularly important to monitor the physical health of operators. Due to restrictions such as the environment and terrain, communication signals are limited and cannot be accurately positioned at the work site, and medical emergency response is slow. Improving the safety of employees working in dangerous environments has become an urgent need in the power industry. During power operations, operators can obtain detailed information about their current status through smart wearable devices, and quickly analyze and warn of their own physical signs in advance, so as to facilitate the work tasks of the staff and ensure the reliability and efficiency of safety and dispatching and command.

[0003] The sensors of smart wearable devices for power workers generally collect time series data such as body temperature, heart rate, blood oxygen, and brain waves. When extracting time series data features, a variety of algorithms and techniques are usually used. These algorithms can be roughly divided into time domain feature extraction and frequency domain feature extraction methods, but time domain and frequency domain feature extraction methods each have their own unique advantages and limitations. In particular, feature extraction of brain wave data is a key link in EEG signal processing, which aims to extract useful information from complex EEG signals for subsequent analysis and classification. Traditional time and frequency domain feature extraction methods often cannot effectively extract hidden data features, resulting in a lower health warning recognition rate. Due to the different system responses caused by the impact of different health warning categories, the degree of confusion of the monitoring data varies.

[0004] In the health monitoring and early warning of power workers, the traditional time and frequency domain feature extraction method uses a deep learning method that relies on gradient information for model training when obtaining the health warning category. However, extreme situations are prone to occur when extracting features during the training process, which makes it difficult for existing optimization methods to obtain the optimal health warning category output results, resulting in large output result deviations. Summary of the invention

[0005] The purpose of the present invention is to address the deficiencies of the above-mentioned prior art and provide a health warning method, system, device and medium for operators, so as to solve the problem that the deep learning method used in the prior art to obtain health warning categories has many parameters, high energy consumption, requires a large amount of labeled data for training, and the features extracted during the training process are prone to extreme situations, resulting in the existing optimization method that relies on gradient information being difficult to obtain the optimal health warning category output result, causing large deviations in the output results.

[0006] The present invention specifically provides the following technical solutions:

[0007] A health warning method for operators comprises the following steps:

[0008] Collect various types of data representing health warning categories from workers' wearable devices, and extract time domain and frequency domain features from various types of data;

[0009] All features are input into the t-distribution probability neural network, and the activation function operation is performed on each type of feature through the t-distribution probability density function represented by the Euclidean distance. The results of each operation are added by category, and the categories after the addition are judged. The category with the largest probability value in the judgment result is taken as the output result, and the fractional cockroach swarm optimization algorithm is used to globally optimize the parameters in the output result, and the optimal t-distribution probability neural network is obtained through the optimized parameters;

[0010] Various types of data collected in real time by the workers' wearable devices are input into the optimal t-distribution probability neural network to obtain the optimal output result, and the optimal output result is output as the health warning category to provide health warnings to the workers.

[0011] Preferably, the extracting of time domain and frequency domain features from various types of data includes:

[0012] The fractional calculus theory is introduced into the Boltzmann-Shannon interaction entropy to obtain the fractional Boltzmann-Shannon interaction entropy.

[0013] The fractional Boltzmann-Shannon interaction entropy is used to divide various types of data representing health warning categories of workers' wearable devices into several partitions according to categories, and different probability distributions of the same partition are extracted based on frequency and measurement technology, and the dynamic changes of fractional time series in various types of data are measured;

[0014] The characteristics of time domain and frequency domain in various types of data are obtained through the dynamic changes of different probability distributions and fractional-order time series.

[0015] Preferably, the step of introducing the fractional calculus theory into the Boltzmann-Shannon interaction entropy to obtain the fractional Boltzmann-Shannon interaction entropy comprises:

[0016] The specific expression of fractional Boltzmann-Shannon interaction entropy is:

[0017] FrBSIE(X)=1-JSD α (p‖q);

[0018] Where Γ(·) is the gamma function, X is the bounded region sample sequence, α is the fractional order, and JSD α (p||q) fractional JS divergence, the specific expression is:

[0019]

[0020] Among them, p and q are the probabilities of the two probability distributions P and Q under a specific random variable value, P is the equal-length histogram segmentation distribution, Q is the quantile segmentation distribution, p i is the equal length histogram segmentation distribution under the random variable i, q i is the quantile split distribution of random variable i, m i is the mean of the sum of the probability distributions P and Q.

[0021] Preferably, all features are input into the t-distribution probability neural network, and the activation function operation is performed on each type of feature through the t-distribution probability density function represented by the Euclidean distance, and the various operation results are added by category, and the added categories are judged, and the category with the largest probability value in the judgment result is used as the output result, including:

[0022] The feature is input into the t-distribution probability neural network, and the distance between the input vector and all training sample vectors is obtained in the input layer; wherein the input feature data is X=(x1, x2,…, x n ), n is the number of categories;

[0023] At the sample layer, the feature data is processed by the activation function and then output; the specific expression is:

[0024]

[0025] Among them, the activation function is the probability density function of the t distribution represented by the Euclidean distance, X ik is the kth training sample of the i-th class, ζ is the smoothing parameter, Y ik is the output of the k-th feature data of the i-th category after the activation function, that is, the output of the sample layer, represents the square of the 2-norm;

[0026] In the summation layer, each sample is added by class; the specific expression is:

[0027]

[0028] where g i (·) is the adder of the i-th type, n i is the total number of training samples of the i-th category;

[0029] The added categories are judged to obtain the output result with the maximum output probability value; the specific expression is:

[0030]

[0031] in, is the output result with the largest output probability value.

[0032] Preferably, the fractional cockroach swarm optimization algorithm is used to globally optimize the parameters in the output result, and the optimal t-distribution probability neural network is obtained through the optimized parameters, including:

[0033] Construct a target optimization function, optimize the smoothing parameters, select the root mean square error formula for error calculation, and optimize the model according to the error; the target optimization function is the fractional cockroach swarm optimization algorithm, and the specific expression for error E calculation is:

[0034]

[0035] Among them, a is the total number of test data, j is any test data;

[0036] Taking the error E as the minimum problem, the error E is transformed into a fitness function, and the adaptive function is obtained as f(z); the specific expression is:

[0037]

[0038] Preferably, the optimizing the smoothing parameters comprises:

[0039] The fractional cockroach swarm optimization algorithm is used to optimize the smoothing parameters, specifically:

[0040] Get each individual of the kth iteration and convert each individual X k The position of (i) is used as an initial smoothing parameter, and the local optimal individual P is obtained within each individual visual range. k (i);

[0041] Within a certain time interval, each individual is randomly dispersed, and the current local optimal individual is used to replace the randomly selected individual, and finally the global optimal individual is obtained, and the position of the global optimal individual is used as the optimal smoothing parameter.

[0042] Preferably, the step of randomly dispersing each individual and replacing the randomly selected individual with the current local optimal individual comprises:

[0043] Each individual is randomly dispersed; the specific expression is:

[0044]

[0045] Among them, rand(1,D) is a D-dimensional random vector set within a certain range, real number α∈[0,1], X k+1 (i) is each individual in the k+1th iteration, X k (i) for each individual in the kth iteration;

[0046] Use the current local optimal individual to replace the randomly selected individual, and finally obtain the global optimal individual; the specific expression is:

[0047] X k (i) = Pg k ;

[0048] Among them, k belongs to [1, N], N is the maximum number of categories, Pg k is the global optimal individual position.

[0049] The present invention provides a health warning system for operators, comprising:

[0050] The acquisition module is used to collect various data representing health warning categories from the wearable devices of the operators, and to extract the time domain and frequency domain features of various data;

[0051] The model processing module is used to input all features into the t-distribution probability neural network, perform activation function operation on each type of feature through the t-distribution probability density function represented by the Euclidean distance, and add the results of each operation by category, judge the added categories, and take the category with the largest probability value in the judgment result as the output result, and use the fractional cockroach swarm optimization algorithm to globally optimize the parameters in the output result, and obtain the optimal t-distribution probability neural network through the optimized parameters;

[0052] The early warning module is used to input various data collected in real time by the wearable devices of the operators into the optimal t-distribution probability neural network to obtain the optimal output result, and output the optimal output result as the health warning category to provide health warning to the operators.

[0053] The present invention provides a computer device, including a memory and a processor, wherein a program is stored in the memory, and when the program is executed by the processor, the processor executes the steps of the above-mentioned health warning method for operators.

[0054] The present invention provides a storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of the above-mentioned health warning method for operators are implemented.

[0055] Compared with the prior art, the present invention has the following significant advantages:

[0056] The present invention performs activation function operation on each type of feature through the t-distribution probability density function represented by the Euclidean distance, wherein the t-distribution can process the extreme cases of the feature, so that the statistical inference of the warning result is more robust, and at the same time, the output result is obtained by adding and judging the various operation results by category, and the fractional-order cockroach swarm optimization algorithm is used to perform global optimization on the parameters in the output result, and the probability of the result falling into the local extreme value is reduced through global optimization, so as to achieve the optimization of the t-distribution probability neural network model, and the deviation of the subsequent output result is reduced through the processing of the extreme cases of the feature and the global optimization of the parameters, thereby improving the warning accuracy of the health warning category output. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 It is an overall flow chart of a health warning method for workers of the present invention;

[0058] Figure 2 This is a diagram of the t-PNN network structure of the present invention;

[0059] Figure 3 This is a structural diagram of the health warning algorithm of the smart wearable device for power workers in the present invention. DETAILED DESCRIPTION

[0060] The following is a clear and complete description of the technical solutions of the embodiments of the present invention in conjunction with the drawings in the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0061] like Figure 1 As shown, the main idea of ​​the present invention is to use the fractional Boltzmann-Shannon interaction entropy to perform comprehensive feature extraction on the fractional Boltzmann-Shannon interaction entropy, and then input the obtained features into the probabilistic neural network with t distribution for health warning multi-classification processing, and finally use the fractional cockroach swarm optimization algorithm to optimize the network parameters. The present invention provides a health warning method for operators, which specifically includes the following steps:

[0062] Step S1: Collect various data representing health warning categories from the wearable devices of the workers, and extract the time domain and frequency domain features of each type of data.

[0063] The present invention uses fractional order Boltzmann-Shannon interaction entropy (BSIE) to extract features from data collected by smart wearable devices, without decomposing or transforming the data, and directly measures the complexity of the data. Boltzmann-Shannon interaction entropy is defined as the standard divergence between measurement-based and frequency-based distributions, which can be associated with two historical data. This parameter-free metric can effectively calculate the estimated value and can be regarded as a complementary fusion of differential Shannon entropy and geometric partition entropy. This new metric provides an effective method for unbiased entropy estimation obtained from samples in the state space. Thereby overcoming the shortcomings that traditional feature extraction methods in the time and frequency domains often cannot effectively extract hidden features. Furthermore, the present invention introduces fractional calculus theory into Boltzmann-Shannon interaction entropy. This parameter-free measure can be accurately estimated in a computationally efficient manner, further considering fractional order information, and can measure the dynamic changes of fractional order time series, so that more comprehensive features can be extracted from data collected by smart wearable devices.

[0064] Fractional Boltzmann-Shannon interaction entropy: The entropy estimation method based on frequency and measurement is used to analyze the distribution of samples in the state space. JDS is the JS divergence (Jensen-Shannon Divergence), which represents the standard asymmetric KL divergence (Kullbeck-Leibler Divergence) between two distributions P and Q. Before training the health warning classification algorithm, the data collected by smart wearable devices are generally divided into several partitions according to the health warning category. P and Q are different probability distributions extracted from the same partition based on frequency and measurement technology, respectively. P is an equal-length histogram segmentation distribution, and Q is a quantile segmentation distribution.

[0065] In practical applications, equal-length histogram segmentation distribution usually requires the following steps:

[0066] (1) Collect data and determine the range of the data (maximum value minus minimum value).

[0067] (2) Determine the bin width, that is, the width of each bin of the histogram, which is usually determined by dividing the range by the number of bins in the histogram.

[0068] (3) Determine the boundary values ​​for each group, ensuring that both the maximum and minimum values ​​in the data table are included.

[0069] (4) Prepare a frequency distribution table and count the frequency of each group.

[0070] (5) Draw a histogram. The height of each rectangle represents the number of data in the group, and the widths are all equal.

[0071] Quantile split distribution is a method in statistics that involves dividing a dataset into several parts, each containing an equal number of data points.

[0072] The application steps of quantile split distribution can be summarized as follows:

[0073] (1) Collect data: Collect the grayscale value data of the image. The amount of this data should be sufficient for effective analysis.

[0074] (2) Determine the range of the data: calculate the maximum and minimum values ​​of the data to obtain the range.

[0075] (3) Determine the class interval: Based on the range and the preset number of groups, calculate the width of each group, i.e., the class interval.

[0076] (4) Determine the boundary values ​​of each group: In order to avoid overlap between data values ​​and group boundary values, the unit of the group boundary value should be half of the smallest measurement unit.

[0077] (5) Count the frequency of occurrence of each group of data: calculate the number of pixel values ​​in each quantile interval.

[0078] (6) Draw a histogram: Use the group interval as the base length and the frequency as the height to draw a rectangular graph of each group, thereby obtaining a histogram of the quantile segmentation distribution.

[0079] Extract time and frequency domain features from various types of data, including:

[0080] The fractional calculus theory is introduced into the Boltzmann-Shannon interaction entropy to obtain the fractional Boltzmann-Shannon interaction entropy.

[0081] The fractional Boltzmann-Shannon interaction entropy is used to divide various types of data representing health warning categories of workers' wearable devices into several partitions according to categories, and different probability distributions of the same partition are extracted based on frequency and measurement technology, and the dynamic changes of fractional time series in various types of data are measured.

[0082] The time domain and frequency domain characteristics of various types of data are obtained through the dynamic changes of different probability distributions and fractional-order time series.

[0083] The fractional calculus theory is introduced into the Boltzmann-Shannon interaction entropy to obtain the fractional Boltzmann-Shannon interaction entropy, including:

[0084] The specific expression of fractional Boltzmann-Shannon interaction entropy is:

[0085] FrBSIE(X)=1-JSD α (p‖q) (1);

[0086] Where Γ(·) is the gamma function, BSIE(X) is the Boltzmann-Shannon interaction entropy, X is the bounded region sample sequence, α is the fractional order, and JSD α (p||q) fractional JS divergence, JSD α The specific expression of (p||q) is:

[0087]

[0088] Among them, p and q are the probabilities of the two probability distributions P and Q under a specific random variable value, P is the equal-length histogram segmentation distribution, Q is the quantile segmentation distribution, p i is the equal length histogram segmentation distribution under the random variable i, q i is the quantile split distribution of random variable i, m i It is the mean of the sum of the probability distributions of P and Q, -1≤α≤1.

[0089] The Boltzmann-Shannon interaction entropy is defined as follows:

[0090]

[0091] The formula is further refined as follows:

[0092]

[0093] The KL divergence is defined as follows:

[0094]

[0095] JS divergence JSD(P||Q) is defined as follows:

[0096]

[0097] in Therefore, the calculation formula of JSD(P||Q) is as follows:

[0098]

[0099] Step S2: All features are input into the t-distribution probability neural network, and the activation function operation is performed on each type of feature through the t-distribution probability density function represented by the Euclidean distance, and the results of each operation are added by category, and the categories after the addition are judged, and the category with the largest probability value in the judgment result is taken as the output result, and the fractional cockroach swarm optimization algorithm is used to globally optimize the parameters in the output result, and the optimal t-distribution probability neural network is obtained through the optimized parameters.

[0100] As a special radial basis neural network, probabilistic neural network (PNN) has a series of significant advantages:

[0101] Simple learning process and fast training speed: PNN's learning process is relatively simple and its training speed is fast, which is very suitable for real-time processing. This is mainly due to its supervised classification mechanism based on Bayesian minimum risk criterion and its ability to convert nonlinear learning algorithms into linear learning algorithms, thereby maintaining high accuracy.

[0102] High classification accuracy: PNN performs well in classification problems, especially when there are enough training samples, it can converge to the Bayesian classifier and achieve high classification accuracy. This feature makes it widely used in fields such as pattern recognition and fault diagnosis.

[0103] Easy to expand and adjust: The network structure of PNN is relatively fixed, and the number of neurons in each layer is relatively fixed, so it is easy to implement in hardware. At the same time, when it is necessary to increase or decrease the category mode, there is no need to re-train for a long time, which makes PNN have good flexibility and scalability in practical applications.

[0104] Strong nonlinear approximation capability: PNN can achieve arbitrary nonlinear approximation, and the decision surface formed by the PNN network is very close to the surface under the Bayesian optimal criterion. This makes PNN have a great advantage in dealing with complex nonlinear problems.

[0105] However, the activation function of PNN is the Gaussian distribution probability density function. Although the Gaussian distribution is very useful in many cases, it is not the best choice in all cases. For non-continuous data, distributions with multiple peaks, or data with obvious asymmetry, other types of distributions may need to be considered. In addition, the Gaussian distribution assumes that the data is symmetrical and the data points are mainly concentrated near the mean. However, in practical applications, the data may contain extreme values ​​or outliers, which may be inappropriately emphasized or ignored in the Gaussian distribution. Outliers may significantly affect the mean and variance of the Gaussian distribution, thereby affecting the predictive ability and stability of the model. However, the t distribution has a certain tolerance for outliers (extreme values) in the data. Since the tail of the t distribution is "thicker" than the normal distribution (i.e., the tail is higher), it can better handle extreme cases in the data, making statistical inference more robust. Therefore, the present invention combines the advantages of the t distribution and PNN to propose a t-distribution probabilistic neural network (t-Distribution Probabilistic Neural Network, t-PNN), that is, the Gaussian distribution probability density function is replaced by the t distribution probability density function represented by the Euclidean distance.

[0106] Probabilistic neural network is a kind of neural network commonly used for pattern classification. It uses Gaussian function to estimate probability density and realize pattern classification. PNN has the characteristics of simple training, fast convergence, and no local minimum problem. It is suitable for real-time processing and nonlinear approximation. Like PNN, t-PNN contains four network layers: input layer, sample layer, summation layer, and competition layer. Including:

[0107] Input layer: Input the features into the t-distribution probability neural network, and obtain the distance between the input vector and all training sample vectors in the input layer; where the input feature data is X = (x1, x2, ..., x n ), n is the number of categories.

[0108] Sample layer: The number of neurons is the number of training samples. In the sample layer, the feature data is processed by the activation function and then output; the specific expression is:

[0109]

[0110] Among them, the activation function is the probability density function of the t distribution represented by the Euclidean distance, X ik is the kth training sample of the i-th class, ζ is the smoothing parameter, Y ik is the output of the k-th feature data of the i-th category after the activation function, that is, the output of the sample layer, Represents the square of the 2-norm.

[0111] Summation layer: The number of neurons is the number of categories. Each sample is added by category, which is equivalent to C adders; the specific expression is:

[0112]

[0113] where g i (·) is the adder of the i-th type, n i is the total number of training samples in the i-th category.

[0114] Competition layer: The number of neurons is 1. The added categories are judged to obtain the output result with the largest output probability value, that is, there is only one 1 in the output result, and the rest are 0. The output result of the category with the largest probability value is 1; the specific expression is:

[0115]

[0116] Among them, y~ is the output result with the largest output probability value.

[0117] The fractional cockroach swarm optimization algorithm is used to globally optimize the parameters in the output results, and the optimal t-distribution probability neural network is obtained through the optimized parameters, including:

[0118] In this method, the smoothing parameter ζ is an important parameter that affects the algorithm results. It is necessary to construct a target optimization function, optimize the smoothing parameter, and optimize the model according to the error. The root mean square error formula is selected for error calculation. The model is optimized according to the error. Suppose the true value corresponding to the jth group of test data is y j , the predicted value is Then the objective function is the error E; the objective optimization function is the fractional cockroach swarm optimization algorithm, and the specific expression for calculating the error E is:

[0119]

[0120] Among them, a is the total number of test data, and j is any test data.

[0121] In order to facilitate the subsequent swarm intelligence method to optimize parameters, the error E is taken as the minimum problem, and the error E is transformed into a fitness function to obtain the adaptive function f(z); the specific expression is:

[0122]

[0123] Its essence is equivalent to a strongly nonlinear non-convex optimization problem. Taking the three-class classification problem with 2 training samples in each class as an example, its network structure Figure 2 shown.

[0124] Optimize the smoothing parameters, including:

[0125] The Fractional Order Cockroach Swarm Optimization (FCSO) algorithm is used to optimize the parameters. The FCSO algorithm is inspired by the predation behavior of cockroach colonies. The algorithm is established by imitating the chasing behavior of individual cockroaches to find the overall optimal value. Based on the Cockroach Swarm Optimization (CSO) algorithm, the present invention extends the theory through fractional derivatives and proposes a Fractional Order Cockroach Swarm Optimization algorithm for the first time.

[0126] The standard CSO assumes R D is a D-dimensional search space. A cockroach colony consists of N cockroach individuals. The i-th cockroach individual is represented by a D-dimensional vector X(i) = (x i1 ,x i2 ,...,x iD )(i=1,2,...,N), each individual position is a potential solution. Specifically:

[0127] Aggregation behavior: Get each individual X at the kth iteration k (i), for each individual X kThe position of (i) is used as an initial smoothing parameter and will chase the local optimal individual P within each individual’s visual range. k (i). Of course, this chasing behavior also produces clusters. When an individual is the best individual within its visual range, it will chase the global optimal individual Pg k , the standard CSO formula describing this behavior is:

[0128]

[0129] Where τ is a fixed value and rand is a random number in the range [0,1].

[0130] P k (i) = Opt j {X k (j)||X k (i)-X k (j)|≤Θ,j=1,2,...,N}(i=1,2,...,N) (14);

[0131] Pg k =Opt i {X k (i), i=1,2,...,N} (15);

[0132] Among them, P k (i) and Pg k are the individual optimal and global optimal cockroach positions respectively, and Θ is the field of view radius which is a constant.

[0133] Rewrite the first line of formula (13) as follows:

[0134] X k+1 (i)-X k (i) = τ·rand·(P k (i)-X k (i)) (16);

[0135] The left side is the discrete form of the differential with order 1. In order to enhance the historical dependence of the algorithm, (16) is generalized to the fractional order case. Because the fractional order derivative has global correlation, it can better reflect the historical dependence process of the system function development. This means that the fractional order derivative not only focuses on the change of the current state, but also considers the impact of the past historical state on the current state of the system. In contrast, the integer order derivative has locality and is more suitable for describing local change processes that do not depend on the historical state.

[0136] The Grünwald-Letnikov fractional derivative of X(t) is defined as:

[0137]

[0138] where α∈(0,1], Γ(·) is a gamma function.

[0139] While integer derivatives imply only a finite sequence, fractional derivatives require an infinite number of terms; integer derivatives are "local" operators, whereas fractional derivatives implicitly contain a "memory" of all past events.

[0140] The approximate expression of formula (17) is:

[0141]

[0142] Where T is the sampling period and r is the truncation order.

[0143] X on the left side of formula (16) k+1 (i)-X k (i) is the discrete form of the differential when the order α = 1 (assuming T = 1), so:

[0144] D α [X k+1 (i)] = τ·rand·(P k (i)-X k (i)) (19);

[0145] If the characteristics of fractional calculus are considered, the order of velocity differential can be extended to real number α∈[0,1], which can obtain smoother changes and longer memory effects. Considering the first r=4 term of the differential given by expression (18), expression (19) can be rewritten as:

[0146]

[0147] When r≥4, the degree of approximation does not change much, so it is discarded. Therefore, we have:

[0148]

[0149] The second line of formula (13) is also processed in the same way, and the formula is rewritten as:

[0150]

[0151] This formula is used to describe the clustering behavior of FCSO.

[0152] Dispersed behavior: Within a certain time interval, each individual is randomly dispersed so that it can maintain the current individual diversity. The standard CSO model is:

[0153] X k+1 (i) = X k(i)+rand(1,D),i=1,2,...,N (23);

[0154] where rand(1,D) is a D-dimensional random vector that can be set within a certain range.

[0155] Formula (23) can also be rewritten in fractional order form:

[0156]

[0157] Among them, rand(1,D) is a D-dimensional random vector set within a certain range, real number α∈[0,1], X k+1 (i) is each individual in the k+1th iteration, X k (i) is each individual in the kth iteration. This formula is used to describe the decentralized behavior of FCSO.

[0158] Cruel behavior: In a certain time interval, the current local optimal individual is used to replace the randomly selected individual, and finally the global optimal individual is obtained. The position of the global optimal individual is used as the optimal smoothing parameter, that is, the strong eat the weak. The standard CSO model is:

[0159] X k (i) = Pg k (25);

[0160] where k is a random integer in [1, N], Pg k For the global optimal individual position, the mathematical expressions of the cruel behaviors of FCSO and CSO are consistent.

[0161] Step S3: Input various data collected in real time by the wearable devices of the operators into the optimal t-distribution probability neural network to obtain the optimal output result, and output the optimal output result as the health warning category to provide health warning for the operators.

[0162] In summary, the present invention uses fractional-order Boltzmann-Shannon interaction entropy to perform comprehensive feature extraction on various time series data collected by sensors of smart wearable devices of power workers, and then inputs the obtained feature data into the t-distribution probability neural network for health warning multi-classification processing. The optimizer of this network is the fractional-order cockroach swarm optimization algorithm.

[0163] The health warning algorithm for smart wearable devices for power workers should include an input layer of 4 neurons, which input the 4 characteristic data of body temperature, heart rate, blood oxygen and brain waves respectively; a sample layer neuron with the same number of training samples; a summation layer neuron with the same number of categories. The present invention assumes that there are three warning categories, namely normal, mild risk and high risk; a competition layer neuron with 1 neuron, which outputs the category with the highest probability. Taking the 3-class classification problem with 2 training samples in each category as an example, its network structure is Figure 3 shown.

[0164] The training process of the t-distribution probability neural network is equivalent to solving a strongly nonlinear, non-convex optimization problem. The optimization methods that rely on gradient information in the traditional machine learning field often fall into the trap of local extreme values ​​and are difficult to find the global optimum. Based on the cockroach swarm optimization algorithm, the present invention uses the fractional-order derivative theory to transform it into a fractional-order cockroach swarm optimization algorithm with long-term memory dependence and global optimization capabilities.

[0165] The present invention uses fractional Boltzmann-Shannon interaction entropy to extract features from data collected by smart wearable devices, without decomposing or transforming the data, and directly measures the complexity of the data. This overcomes the shortcomings of traditional time and frequency domain feature extraction methods that often cannot effectively extract hidden data features. It can also measure the dynamic changes of fractional time series and extract more comprehensive signal features from voiceprint data.

[0166] As a special radial basis neural network, the probabilistic neural network has a series of significant advantages: simple learning process, fast training speed, high classification accuracy, easy expansion and adjustment, and strong nonlinear approximation ability. Therefore, the present invention combines the advantages of t distribution and PNN to propose a t distribution probabilistic neural network.

[0167] The training process of the t-distribution probability neural network is equivalent to solving a strongly nonlinear, non-convex optimization problem. The optimization methods that rely on gradient information in the traditional machine learning field often fall into the trap of local extreme values ​​and are difficult to find the global optimum. Based on the cockroach swarm optimization algorithm, the present invention uses the fractional-order derivative theory to transform it into a fractional-order cockroach swarm optimization algorithm with long-term memory dependence and global optimization capabilities.

[0168] Based on the above method, the present invention provides a health warning system for operators, including: a collection module, a model processing module and a warning module.

[0169] Among them, the acquisition module is used to collect various types of data representing health warning categories from the wearable devices of the operators, and to extract the time domain and frequency domain features from various types of data; the model processing module is used to input all features into the t-distribution probability neural network, and perform activation function operations on each type of feature through the t-distribution probability density function represented by the Euclidean distance, and add the results of each operation by category, judge the added category, and take the category with the largest probability value in the judgment result as the output result, and use the fractional-order cockroach swarm optimization algorithm to globally optimize the parameters in the output result, and obtain the optimal t-distribution probability neural network through the optimized parameters; the warning module is used to input various types of data collected in real time by the operators' wearable devices into the optimal t-distribution probability neural network, obtain the optimal output result, and output the optimal output result as the health warning category to provide health warning for the operators.

[0170] The present invention also provides a computer device, including a memory and a processor, wherein a program is stored in the memory, and when the program is executed by the processor, the processor executes the steps of a health warning method for operators.

[0171] In accordance with the disclosed embodiments, a computing device may communicate with one or more external devices (e.g., keyboards, pointing devices, Bluetooth communications, etc.), or with any device (e.g., routers, modems, etc.) that enables a computing device to communicate with one or more other computing devices.

[0172] The present invention also provides a storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of a health warning method for operators are implemented.

[0173] According to the disclosed embodiments, the storage medium may be a non-volatile computer-readable storage medium, such as but not limited to: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof. In the present invention, the storage medium may be any tangible medium containing or storing a program that may be used by or in conjunction with an instruction execution system, apparatus, or device.

[0174] The above content is a further detailed description of the present invention in combination with a specific preferred embodiment. For technicians in the technical field to which the present invention belongs, several simple deductions or substitutions can be made without departing from the concept of the present invention, which should be regarded as belonging to the protection scope of the present invention.

Claims

1. A health warning method for workers, characterized in that: include: Collect various types of data representing health warning categories from workers' wearable devices, and extract time domain and frequency domain features from various types of data; All features are input into the t-distribution probability neural network, and the activation function operation is performed on each type of feature through the t-distribution probability density function represented by the Euclidean distance. The results of each operation are added by category, and the categories after the addition are judged. The category with the largest probability value in the judgment result is taken as the output result, and the fractional cockroach swarm optimization algorithm is used to globally optimize the parameters in the output result, and the optimal t-distribution probability neural network is obtained through the optimized parameters; Various types of data collected in real time by the workers' wearable devices are input into the optimal t-distribution probability neural network to obtain the optimal output result, and the optimal output result is output as the health warning category to provide health warnings to the workers.

2. A health warning method for workers as claimed in claim 1, characterized in that: The extraction of time domain and frequency domain features in various types of data includes: The fractional calculus theory is introduced into the Boltzmann-Shannon interaction entropy to obtain the fractional Boltzmann-Shannon interaction entropy. The fractional Boltzmann-Shannon interaction entropy is used to divide various types of data representing health warning categories of workers' wearable devices into several partitions according to categories, and different probability distributions of the same partition are extracted based on frequency and measurement technology, and the dynamic changes of fractional time series in various types of data are measured; The characteristics of time domain and frequency domain in various types of data are obtained through the dynamic changes of different probability distributions and fractional-order time series.

3. A health warning method for workers as claimed in claim 2, characterized in that: The fractional-order calculus theory is introduced into the Boltzmann-Shannon interaction entropy to obtain the fractional-order Boltzmann-Shannon interaction entropy, including: The specific expression of fractional Boltzmann-Shannon interaction entropy is: FrBSIE(X)=1-JSD α (p‖q); Where Γ(·) is the gamma function, X is the bounded region sample sequence, α is the fractional order, and JSD α (p||q) fractional JS divergence, the specific expression is: Among them, p and q are the probabilities of the two probability distributions P and Q under a specific random variable value, P is the equal-length histogram segmentation distribution, Q is the quantile segmentation distribution, p i is the equal length histogram segmentation distribution under the random variable i, q i is the quantile split distribution of random variable i, m i is the mean of the sum of the probability distributions P and Q.

4. A health warning method for workers as claimed in claim 1, characterized in that: All features are input into the t-distribution probability neural network, and the activation function operation is performed on each type of feature through the t-distribution probability density function represented by the Euclidean distance. The results of each operation are added by category, and the categories after addition are judged. The category with the largest probability value in the judgment result is taken as the output result, including: The feature is input into the t-distribution probability neural network, and the distance between the input vector and all training sample vectors is obtained in the input layer; wherein the input feature data is X=(x1, x2,…, x n ), n is the number of categories; At the sample layer, the feature data is processed by the activation function and then output; the specific expression is: Among them, the activation function is the probability density function of the t distribution represented by the Euclidean distance, X ik is the kth training sample of the i-th class, ζ is the smoothing parameter, Y ik is the output of the k-th feature data of the i-th category after the activation function, that is, the output of the sample layer, represents the square of the 2-norm; In the summation layer, each sample is added by class; the specific expression is: where g i (·) is the adder of the i-th type, n i is the total number of training samples of the i-th category; The added categories are judged to obtain the output result with the maximum output probability value; the specific expression is: in, is the output result with the largest output probability value.

5. A health warning method for workers as claimed in claim 4, characterized in that: The fractional cockroach swarm optimization algorithm is used to globally optimize the parameters in the output result, and the optimal t-distribution probability neural network is obtained through the optimized parameters, including: Construct a target optimization function, optimize the smoothing parameters, select the root mean square error formula for error calculation, and optimize the model according to the error; the target optimization function is the fractional cockroach swarm optimization algorithm, and the specific expression for error E calculation is: Among them, a is the total number of test data, j is any test data; Taking the error E as the minimum problem, the error E is transformed into a fitness function, and the adaptive function is obtained as f(z); the specific expression is:

6. A health warning method for workers as claimed in claim 5, characterized in that: The step of optimizing the smoothing parameters comprises: The fractional cockroach swarm optimization algorithm is used to optimize the smoothing parameters, specifically: Get each individual of the kth iteration and convert each individual X k The position of (i) is used as an initial smoothing parameter, and the local optimal individual P is obtained within each individual visual range. k (i); Within a certain time interval, each individual is randomly dispersed, and the current local optimal individual is used to replace the randomly selected individual, and finally the global optimal individual is obtained, and the position of the global optimal individual is used as the optimal smoothing parameter.

7. A health warning method for workers as claimed in claim 6, characterized in that: The method of randomly dispersing each individual and replacing the randomly selected individual with the current local optimal individual includes: Each individual is randomly dispersed; the specific expression is: Among them, rand(1,D) is a D-dimensional random vector set within a certain range, real number α∈[0,1], X k+1 (i) is each individual in the k+1th iteration, X k (i) for each individual in the kth iteration; Use the current local optimal individual to replace the randomly selected individual, and finally obtain the global optimal individual; the specific expression is: X k (i)=Pg k ; Among them, k belongs to [1, N], N is the maximum number of categories, Pg k is the global optimal individual position.

8. A health warning system for workers, characterized in that: include: The acquisition module is used to collect various data representing health warning categories from the wearable devices of the operators, and to extract the time domain and frequency domain features of various data; The model processing module is used to input all features into the t-distribution probability neural network, perform activation function operation on each type of feature through the t-distribution probability density function represented by the Euclidean distance, and add the results of each operation by category, judge the added categories, and take the category with the largest probability value in the judgment result as the output result, and use the fractional cockroach swarm optimization algorithm to globally optimize the parameters in the output result, and obtain the optimal t-distribution probability neural network through the optimized parameters; The early warning module is used to input various data collected in real time by the wearable devices of the operators into the optimal t-distribution probability neural network to obtain the optimal output result, and output the optimal output result as the health warning category to provide health warning to the operators.

9. A computer device, characterized in that: It includes a memory and a processor, wherein a program is stored in the memory, and when the program is executed by the processor, the processor executes the steps of a health warning method for operators as described in any one of claims 1 to 7.

10. A storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of a health warning method for operators described in any one of claims 1 to 7 are implemented.

Images