Physical sign data monitoring model construction method and device, medium and equipment

Abstract

The invention discloses a physical sign data monitoring model construction method and device, a medium and equipment, and relates to the technical field of health monitoring. The method comprises the steps of performing feature extraction of physiological sign data by adopting a fractional order reverse scatter entropy method, replacing a Sigmoid activation function in a traditional extreme learning machine with a DiTAC activation function, modifying an L2 norm of an output layer weight coefficient in a loss function in the extreme learning machine into an L1 norm, establishing an optimization model based on the improved extreme learning machine, and performing feature extraction of the physiological sign data based on the optimized model. And solving the optimization model by using an Aribus algorithm and a forty-thief algorithm which introduce discrete fractional order Caputo, so as to obtain a physical sign data monitoring model. The extreme learning machine is constrained by the trainable parameters introduced by the DiTAC activation function and the L1 norm of the output layer weight coefficient in the training process, so that the trained extreme learning machine can be more suitable for physical sign data monitoring, and a more accurate monitoring result is obtained.

Description

Technical Field

[0001] The present invention relates to the field of health monitoring technology, and in particular to a method, device, medium and equipment for constructing a vital sign data monitoring model. Background Art

[0002] As the scale of power grids continues to expand, a large number of new and large-scale equipment have been put into use, which has greatly increased the complexity of power grid equipment installation and operation and maintenance work, and has also put forward higher requirements for the safety of power workers. The physiological signs of power workers on site are monitored in order to detect physiological indicators such as blood pressure or heart rate of workers during the operation to achieve early warning, thereby reducing sudden safety accidents and improving the ability to control the risks of high-risk operations.

[0003] At present, in the field of vital sign index detection of power workers based on smart wearable devices, there is a need for a lightweight model that is efficient, lightweight, simple in principle and easy to implement on equipment with limited resources. This model should be able to quickly and accurately issue warnings and be interpretable. Among them, the extreme learning machine can be deployed on smart wearable devices for health detection due to its lightweight, interpretable and good predictive performance. When monitoring and evaluating health based on the extreme learning machine, it is usually necessary to construct an extreme learning machine model, then conduct supervised training on the weight coefficients of the extreme learning machine, and finally use the trained model to monitor vital signs.

[0004] However, the physiological signs produced by different power workers or the same power worker facing different working conditions may vary greatly, that is, the data used to train the extreme learning machine has the characteristics of large data fluctuations in the same sample and multiple data representations of different samples. Traditional extreme learning machines are sensitive to data fluctuations or noise, and due to the limited nonlinear expression ability of traditional extreme learning machines, their learning ability for data in multiple representations is limited. The above problems will cause the accuracy of the vital sign data monitoring model obtained by training based on traditional extreme learning machines to be improved. Summary of the invention

[0005] Based on this, in order to solve the technical problems in the prior art, the present invention provides a method, device, medium and equipment for constructing a vital sign data monitoring model.

[0006] The present invention provides a method for constructing a vital sign data monitoring model, comprising:

[0007] An improved hidden layer is formed by replacing the Sigmoid activation function of the hidden layer in the original extreme learning machine with a DiTAC activation function, wherein the DiTAC activation function performs a non-fixed nonlinear operation on the input data by introducing a trainable parameter; and an improved extreme learning machine including an input layer, an improved hidden layer and an output layer is constructed;

[0008] Collect physiological sign data and corresponding real health status labels to construct a data set; input the physiological sign data in the data set into an improved extreme learning machine, take minimizing the error between the predicted health status label and the real health status label corresponding to the physiological sign data output by the improved extreme learning machine as the objective function, and take the L1 norm of the output layer weight coefficient and the L2 norm of the trainable parameter introduced by the DiTAC activation function as constraints to construct an optimization model;

[0009] The optimization model is solved to obtain the optimal weight coefficient and the optimal parameters of the DiTAC activation function that minimize the error between the predicted health status label and the true health status label; based on the optimal weight coefficient and the optimal parameters of the DiTAC activation function, a vital sign data monitoring model for mapping physiological sign data to health status labels is obtained.

[0010] Furthermore, the construction of the optimization model specifically includes:

[0011]

[0012] in, is the objective function, H is the output matrix of the improved hidden layer in the improved extreme learning machine, β is the weight matrix of the output layer in the improved extreme learning machine, Hβ is the predicted health status label output by the improved extreme learning machine; T is the predicted health status label, and the superscript T represents the transposition operation; θ is the trainable parameter introduced by the DiTAC activation function in the improved hidden layer; λ1||β||1+λ2||θ||2 is the constraint condition, ||β||1 is the L1 norm of β, and λ1 is the L1 regularization term coefficient; ||θ||2 is the L2 norm of θ, and λ2 is the L2 regularization term coefficient.

[0013] Furthermore, H is an output matrix of an improved hidden layer in an improved extreme learning machine, specifically comprising:

[0014]

[0015] Among them, w1~w l is the weight coefficient vector of the input layer; b1~b l is the bias vector of the input layer; l is the number of neurons in the improved hidden layer; x1~x Q Training data for group Q; is the DiTAC activation function that introduces the trainable parameter θ.

[0016] Furthermore, the solving of the optimization model is to solve the optimization model by a meta-heuristic swarm intelligence method in derivative-free optimization technology to solve the problem that the optimization model is not differentiable due to the introduction of the L1 norm;

[0017] The meta-heuristic swarm intelligence method in the derivative-free optimization technology is to improve the Alibaba and Forty Thieves algorithm. The improved Alibaba and Forty Thieves algorithm uses discrete fractional order Caputo to update the thieves' position information carrying β and θ in the optimization process. The discrete fractional order Caputo assists the Alibaba and Forty Thieves algorithm to record historical search experience through its long-term memory effect characteristics, so as to reduce the probability of falling into local optimality.

[0018] Furthermore, before inputting the physiological sign data in the data set into the improved extreme learning machine, feature extraction is performed on the physiological sign data to obtain a fractional order reverse scatter entropy for characterizing the temporal complexity of the physiological sign data and the data mutation behavior; specifically including:

[0019] The physiological sign data with a time length of T is mapped to the normal distribution space through the normal distribution function:

[0020] X={x(i),i=1,2,...,T}→Y={y(i),i=1,2,...,T}

[0021] Among them, x(i) is the physiological sign data at the i-th moment; y(i) is the mapping value corresponding to x(i);

[0022] Through linear changes Map Y to a categorical form to obtain the category labels corresponding to each component in Y:

[0023] Y={y(i),i=1,2,…,T}→Z={z(i),i=1,2,…,T}

[0024] Among them, int() is the rounding function, c is the number of categories, and z(i) is a positive integer ranging from 1 to c;

[0025] Reconstruct the category form Z into L embedding vectors with time delay d and embedding dimension m, each of which includes m components ranging from 1 to c;

[0026] Map each embedding vector to a scatter pattern π i Calculate the probability P(π i ):

[0027]

[0028] Among them, P(π i ) is represented by the ratio of the number of the ith scatter pattern to the number of embedded vectors; Number(π i ) is the number of embedding vectors mapped to the i-th scatter pattern; c m is the number of all scatter modes;

[0029] Based on the fractional order probability P(π i ) is weighted to obtain the fractional reverse diffusion entropy FRDE:

[0030]

[0031] Among them, α is the fractional order; Γ(·) is the gamma function; ψ(·) is the digamma function.

[0032] Furthermore, the physiological sign data includes body temperature data, heart rate data, blood oxygen data and brain wave data.

[0033] Furthermore, the real health status label includes a normal label, a low-risk label and a high-risk label.

[0034] The present invention provides a device for constructing a vital sign data monitoring model, comprising:

[0035] An extreme learning machine improvement module is used to form an improved hidden layer by replacing the Sigmoid activation function of the hidden layer in the original extreme learning machine with a DiTAC activation function, wherein the DiTAC activation function performs a non-fixed nonlinear operation on the input data by introducing a trainable parameter; and construct an improved extreme learning machine including an input layer, an improved hidden layer and an output layer;

[0036] The optimization model construction module is used to collect physiological sign data and corresponding real health status labels to construct a data set; the physiological sign data in the data set is input into the improved extreme learning machine, and the objective function is to minimize the error between the predicted health status label and the real health status label corresponding to the physiological sign data output by the improved extreme learning machine, and the L1 norm of the output layer weight coefficient and the L2 norm of the trainable parameter introduced by the DiTAC activation function are used as constraints to construct the optimization model;

[0037] The vital sign data monitoring model acquisition module is used to solve the optimization model to obtain the optimal weight coefficient and the optimal parameters of the DiTAC activation function that minimize the error between the predicted health status label and the true health status label; based on the optimal weight coefficient and the optimal parameters of the DiTAC activation function, a vital sign data monitoring model for mapping physiological vital sign data into health status labels is obtained.

[0038] The present invention provides a computer-readable storage medium, wherein the storage medium stores a computer program, and when the computer program is executed by a processor, the method for constructing the above-mentioned vital sign data monitoring model is implemented.

[0039] The present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the method for constructing the above-mentioned vital sign data monitoring model when executing the program.

[0040] At least one of the above technical solutions adopted by the present invention can achieve the following beneficial effects:

[0041] In the method for constructing the vital sign data monitoring model provided by the present invention, the Sigmoid activation function of the hidden layer in the original extreme learning machine is replaced with the DiTAC activation function for improvement, and the output layer weight parameters are subjected to L1 regularization to improve the accuracy of the vital sign data monitoring model obtained based on the extreme learning machine. Specifically: the trainable parameters introduced by the DiTAC activation function can automatically adjust the shape of the activation function according to the specific training data, which enhances the nonlinear expression ability of the extreme learning machine and enables it to adapt to physiological sign data with many forms of expression; compared with the L2 norm, the L1 norm is less sensitive to outliers, which makes it more stable in physiological sign data with large data fluctuations, which helps to improve the performance of the classifier. During the training process, the trainable parameters introduced by the DiTAC activation function and the L1 norm of the output layer weight coefficient are constrained, so that the trained extreme learning machine can be more suitable for vital sign data monitoring and obtain more accurate monitoring results. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0043] Figure 1 A schematic diagram of a process for constructing a vital sign data monitoring model provided by the present invention;

[0044] Figure 2 A schematic diagram of the algorithm structure of a vital sign data monitoring model provided by the present invention. DETAILED DESCRIPTION

[0045] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in combination with the specific embodiments of the present invention and the corresponding drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments in the specification, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0046] Extreme Learning Machine (ELM) is a learning algorithm for a single hidden layer feedforward neural network. It has the characteristics of fast learning speed and strong generalization ability, and is widely used in classification, regression, feature learning and other fields. ELM first passes the data to the hidden layer through the input layer. Each hidden layer node performs a nonlinear transformation on the input data. Then the output of the hidden layer node is directly connected to the output layer through randomly generated weights. These weights are no longer adjusted during the training process. The output weights of the network are learned by minimizing the error of the output layer. This step is usually completed by solving a linear system.

[0047] Based on the defects of traditional extreme learning machines when detecting the health status of power workers, the present invention provides a method for constructing a vital sign data monitoring model. First, the fractional inverse spread entropy of key physiological data such as body temperature, heart rate, blood oxygen saturation and brain waves obtained by smart wearable devices is extracted; then the Sigmoid activation function in the traditional extreme learning machine is replaced by the DiTAC activation function, which is essentially replacing the fixed activation function with an activation function containing trainable parameters. Although the trainable parameters of the hidden layer are increased, the fitting ability and learning ability of the network are also greatly improved; at the same time, L1 regularization is performed on the weight parameters β of the hidden layer and the output layer; finally, the Alibaba and Forty Thieves algorithm based on the discrete fractional-order Caputo method is used to solve the weight coefficients in the improved extreme learning machine to obtain the vital sign data monitoring model.

[0048] Figure 1 and Figure 2 The construction method flow of the vital sign data monitoring model of this embodiment is shown. Figure 1 and Figure 2 The construction method of the model is described in detail, which includes the following steps:

[0049] S1: Collect physiological sign data including body temperature data, heart rate data, blood oxygen data and brain wave data.

[0050] Collect the historical body temperature time series data, historical heart rate time series data, historical blood oxygen saturation time series data and historical brain wave time series data of several monitored objects, as well as the corresponding health status: normal, mild risk, high risk, and build a data set based on several one-to-one corresponding physiological sign data and health status labels.

[0051] S2: The fractional inverse scatter entropy method is used to extract features from physiological sign data.

[0052] Extract the fractional inverse scatter entropy of historical body temperature time series data, historical heart rate time series data, historical blood oxygen saturation time series data, and historical brain wave time series data. When processing time series data of key physiological parameters such as body temperature, heart rate, blood oxygen saturation, and brain waves, a series of advanced algorithms and techniques are usually used for feature extraction. These methods can be mainly divided into two categories: time domain feature extraction and frequency domain feature extraction, each of which has unique advantages and limitations. Especially in the processing of brain wave data, feature extraction plays a vital role. Its core goal is to mine valuable information from complex EEG signals and lay the foundation for subsequent analysis and classification. However, traditional feature extraction methods have certain limitations in both time domain (such as mean, root mean square, kurtosis, and skewness) and frequency domain (such as power spectrum, envelope spectrum analysis, and fast Fourier transform). Its specific shortcomings are as follows:

[0053] (1) Although time-domain feature extraction methods can track the evolution of signals over time, they often fail to fully capture the frequency components and dynamic changes of signals. For example, statistics such as mean and root mean square can describe the overall strength or energy level of a signal, but cannot reveal the spectral distribution of the signal. (2) Certain features, such as kurtosis and skewness, are sensitive to specific distribution characteristics of the signal (e.g., asymmetry or spikes), but may not be stable enough in some cases and are easily disturbed by noise, thus limiting their effectiveness in abnormal signal detection. (3) When processing large-scale or long-term data sets, time-domain feature extraction may impose a significant computational burden, especially when real-time processing is required, which may increase the computational load and response time of the system. As for frequency-domain feature extraction methods, such as fast Fourier transform (FFT) and power spectrum analysis, although they can reveal the spectral information of the signal, these algorithms are relatively complex and require more computing resources and storage space. Although FFT is computationally efficient, it requires that the signal length must be a power of 2, which may cause certain limitations in practical applications. (4) The frequency resolution of frequency-domain analysis methods is limited by the length of the signal. For short-duration or high-frequency signals, it may not be possible to obtain sufficient resolution to accurately describe the spectral characteristics of the signal. In addition, the choice and parameter setting of the window function will also have a significant impact on the frequency resolution. (5) Frequency domain analysis methods face challenges when processing non-periodic signals. The spectral analysis of non-periodic signals needs to consider issues such as boundary effects and truncation errors, which may lead to inaccurate analysis results. (6) In the process of converting the signal from the time domain to the frequency domain, some key time domain information may be lost, such as the instantaneous characteristics and phase information of the signal, which may be crucial in some application scenarios.

[0054] In view of the defects of existing methods based on time domain and frequency domain, the present invention will use Fractional Reverse Dispersion Entropy (FRDE) to extract features from data. On the one hand, Reverse Dispersion Entropy (RDE) is an emerging complexity measurement method, which has many advantages in feature extraction. It improves the ability to detect the complexity of time series by introducing amplitude information and distance information. Compared with the traditional permutation entropy (PE), RDE performs better in detecting mutation signals and noise robustness testing, and has better stability for analog and sensor signals. RDE is superior to the PE method in distinguishing different data sets, which makes it more effective in processing data with spike features or sudden amplitude changes. In addition, it has high stability for time series of different lengths, which is particularly important for processing variable-length signals that may be encountered in practical applications. At high noise levels, RDE also shows good robustness and can extract complexity information from data, even if there are spike features or sudden amplitude changes in the signal. On the other hand, fractional entropy is an entropy measurement method based on fractional calculus theory, which has significant advantages in feature extraction and system complexity analysis. Enhanced complexity measurement capability: Fractional entropy enhances the complexity measurement capability by introducing fractional derivatives, and can capture the complexity and chaotic characteristics of time series more finely. This method is not only applicable to periodic data, but also can measure chaotic series well, providing a deeper understanding of the degree of randomness or disorder in the macroscopic structure of the system. Fractional entropy can distinguish different types of dynamic behaviors and has better distinguishability and recognition capabilities. In the calculation process of fractional entropy, the selection of parameters provides more flexibility. By adjusting the fractional parameters, it can be optimized for different data characteristics and analysis requirements, thereby improving the adaptability and accuracy of entropy measurement. The introduction of fractional entropy provides a new perspective for entropy theory and has the potential to be combined with existing methods. This combination may lead to more efficient and more accurate feature extraction methods.

[0055] Therefore, the present invention proposes fractional reverse scatter entropy, which is a complexity measurement method based on fractional calculus and entropy theory, and combines the flexibility and history dependence of fractional calculus with the advantages of reverse scatter entropy.

[0056] Reverse Dispersion Entropy (RDE) is a concept used in signal processing and nonlinear dynamics analysis, especially in fault diagnosis and signal complexity evaluation. As a new nonlinear dynamic analysis method, RDE is based on the theory of permutation entropy (PE), combines the advantages of dispersion entropy (DE) and reverse permutation entropy (RPE), and performs well in stability and discrimination.

[0057] For a physiological sign time series X = {x(1), x(2), …, x(T)} of length T, the fractional-order RDE calculation steps are as follows:

[0058] (1) Through the normal distribution function:

[0059]

[0060] Map the time series X = {x(i), i = 1, 2, ..., T} to the normal distribution space Y = {y(i), i = 1, 2, ..., T}, where y(i) ranges from 0 to 1, where μ is the expected value and σ is the average value. 2 is the variance.

[0061] (2) Through linear transformation:

[0062]

[0063] Map the new sequence Y = {y(i), i = 1, 2, ..., T} to the range of Z = {z(i), i = 1, 2, ..., T}, where int is the rounding function, c is the number of categories, and z(i) is a positive integer from 1 to c. The purpose of mapping Y to Z is to clarify the category labels corresponding to each component in Y, so as to facilitate the subsequent calculation of probability by counting.

[0064] (3) Phase space reconstruction:

[0065] The category form Z is reconstructed into L embedding vectors with a delay of d and an embedding dimension of m. The matrix composed of all embedding vectors is:

[0066]

[0067] In the formula, the number of embedded vectors L is T-(m-1)d.

[0068] (4) Calculate the scatter pattern: Each embedding vector can be mapped to a scatter pattern π. The total number of possible scatter patterns is c mBecause each embedding vector has m components, each component can be an integer between 1 and c.

[0069] (5) Calculate the probability of each scattering mode: The relative probability of the i-th scattering mode can be expressed as

[0070]

[0071] Where: P(π i ) is the ratio of the number of the ith scatter pattern to the number of embedded vectors; Number(π i ) refers to the number of embedding vectors mapped to the i-th scatter pattern.

[0072] (6) Based on the fractional order probability P(π i ) is weighted to obtain the fractional reverse diffusion entropy FRDE:

[0073]

[0074] Among them, α is the fractional order; Γ(·) is the gamma function; ψ(·) is the digamma function.

[0075] S3: An improved hidden layer is formed by replacing the Sigmoid activation function of the hidden layer in the original extreme learning machine with a DiTAC activation function. The DiTAC activation function performs a non-fixed nonlinear operation on the input data by introducing training parameters; and an improved extreme learning machine including an input layer, an improved hidden layer and an output layer is constructed.

[0076] The health warning algorithm for smart wearable devices is essentially a multi-classification problem. At present, two technical means are mainly used in this field to achieve this goal: the first method is the weighted average method, which fuses the data by adjusting the weights. This method combines the data collected by each sensor to form a comprehensive assessment of the health status of the operator. However, this method has two main disadvantages: first, in real applications, the relationship between different features may not always be linear, which may cause the comprehensive assessment model to be inconsistent with the actual situation, thereby affecting the accuracy of the assessment. Second, the weights as hyperparameters need to be set manually, and manual selection often has large errors. Although these weights can be determined by hyperparameter optimization methods, this will take a lot of time and reduce the timeliness of the warning. In general, the weighted average method ignores the nonlinear associations implied by each feature and roughly performs a low-order method of linear fitting of the inherent rules. The second method is to use deep learning technology to achieve multi-classification. However, the performance of deep learning is highly dependent on the quantity and quality of training data. If the training data is insufficient, unevenly distributed, or contains noise, the performance of the deep learning model may be seriously affected. In addition, deep learning models may be difficult to generalize to new data with a large distribution difference from the training data. Training deep learning models usually requires a lot of computing resources, which not only increases the cost of training, but also limits the application of models on devices with limited resources. Deep learning models, especially deep neural networks, are often regarded as "black box" models because their internal working mechanisms are difficult to explain intuitively, which limits their application in fields that require a high degree of transparency and interpretability. Deep learning models mainly rely on automatically learning features from data, and rarely make use of human prior knowledge. In some cases, combining human prior knowledge can solve problems more effectively, but deep learning models have limitations in this regard. And hyperparameters play a vital role in deep learning models, but improper selection and adjustment may have a negative impact on the model. The learning rate is a key hyperparameter that controls the learning speed of the model. If the learning rate is set too high, the model may not converge, resulting in high training error and generalization error, and underfitting. On the contrary, if the learning rate is too low, the model will be trained slowly and may even fall into a local optimal solution, affecting the effective capacity of the model. The hyperparameters of the loss function affect the optimization process. If set improperly, even a suitable loss function may have difficulty optimizing the model, thereby reducing the effective capacity of the model. Similarly, the size of the batch sample number affects the convergence speed and stability of the model. Too large or too small a batch size may reduce the effective capacity of the model. In addition, the choice of hyperparameters will directly affect the generalization ability of the model. Too large a regularization parameter may make the model too simple, while too small a regularization parameter may not effectively prevent overfitting.Therefore, the learner (classifier) ​​used in the health warning algorithm of smart wearable devices should minimize the number of hyperparameters, otherwise it will require a lot of labor and time costs.

[0077] In the field of health warning for smart wearable devices for power workers, there is an urgent need for a lightweight model that is efficient, lightweight, simple in principle, and easy to implement on devices with limited resources. This model should be able to provide health warnings quickly and accurately, while also having good generalization and interpretability.

[0078] The present invention achieves the purpose of establishing a lightweight model that is efficient, lightweight, simple in principle, and easy to implement on a resource-limited device. The extreme learning machine is combined with a cutting-edge trainable activation function based on differential homeomorphism, and an extreme learning machine (DiTAC Extreme Learning Machine, DELM) based on a trainable activation function (Diffeomorphism-based TrainableActivation Function, DiTAC) is proposed. On the one hand, a major advantage of ELM is its extremely fast learning speed. Compared with the traditional gradient descent-based algorithm, it does not require iterative weight adjustment in the hidden layer, which greatly reduces the training time. The implementation of ELM is relatively simple because it randomly generates the weights and biases of the hidden layer without the need for complex parameter tuning. During its training, it only needs to set the number of neurons in the hidden layer, which reduces human intervention and makes model training more automated. Even when the parameters of the hidden layer neurons are randomly generated, ELM can still maintain good interpolation ability, universal approximation ability and classification ability. On the other hand, DiTAC is based on efficient differential homeomorphism transformation and has high expressiveness. Even if only a very small number of trainable parameters are introduced, it can significantly enhance the expressiveness and performance of the model. DiTAC is able to learn activation functions of various shapes, which is in contrast to existing trainable activation functions, which are usually limited to learning a certain shape or only convex functions. DiTAC only adds negligible trainable parameters while maintaining high expressive power. The present invention replaces the Sigmoid activation function in the traditional extreme learning machine with the DiTAC activation function, which essentially replaces the fixed activation function with an activation function containing trainable parameters. Although the trainable parameters of the hidden layer are increased, the fitting and learning capabilities of the network are also greatly improved.

[0079] S4: Construct a data set with the extracted fractional inverse scatter entropy features and the corresponding true health status labels; input the fractional inverse scatter entropy features in the data set into the improved extreme learning machine, and take minimizing the gap between the predicted health status label and the true health status label corresponding to the fractional inverse scatter entropy features output by the improved extreme learning machine as the objective function, and take the L1 norm of the output layer weight coefficient and the L2 norm of the training parameters introduced by the DiTAC activation function as constraints to construct an optimization model including the objective function and constraints.

[0080] In order to suppress overfitting, the loss function of the traditional extreme learning machine often performs L2 regularization on the weight parameters β of the hidden layer and the output layer, but this will amplify and reduce the error, that is, the L2 norm is highly sensitive to outliers. In order to promote the sparsity of model parameters and increase the inherent robustness to outliers, the present invention performs L1 regularization on the weight parameters β of the hidden layer and the output layer in the DELM loss function. When the L1 norm is used as a regularization term, the sparsity of the model parameters can be promoted. This means that many coefficients will be compressed to zero, thereby achieving feature selection. This is particularly useful in high-dimensional data because it can reduce the complexity of the model and improve interpretability. And compared with the L2 norm, the L1 norm is less sensitive to outliers, which makes it more stable in data sets with noise or outliers. In multi-class classification problems, the L1 norm can promote the sparsity of the boundaries between categories, which helps to improve the performance of the classifier. The present invention replaces the Sigmoid activation function in the traditional extreme learning machine with the DiTAC activation function, so a trainable parameter θ is introduced in H, which is the hidden layer output matrix. The trainable parameter θ is learned through network training and is not easy to be sparse, which will affect the parameter learning effect. Furthermore, the present invention performs L2 regularization on the trainable parameter in the DELM loss function.

[0081] The traditional extreme learning machine (ELM) is a single-layer feedforward neural network with the characteristics of fast learning speed and good generalization performance. The input weights and hidden layer biases of ELM can be randomly initialized and do not need to be adjusted during the training process. This makes ELM learn very quickly, especially when dealing with large-scale data sets. ELM generates input weights and hidden layer biases randomly, and then calculates the output weights using algorithms such as least squares. This mechanism enables ELM to approximate complex functions while maintaining a simple structure. Activation functions (AFs) play an important role in the success of deep neural networks because they give the latter nonlinear characteristics. In fact, activation functions are crucial for the network to approximate almost arbitrarily complex functions, learn meaningful feature representations, and achieve high prediction performance. In the framework of artificial neural network neurons, the role of the activation function is to put the weighted sum into a nonlinear function for nonlinearization. Its essence is: the weighted sum can be regarded as a linear hyperplane, which divides the spatial data into two categories and can be regarded as a binary linear classifier. However, most real-world problems are nonlinear, and linear classifiers are not universally meaningful at the application level. Therefore, if we need to make the linear classifier nonlinear, we need to use the activation function.

[0082] The choice of nonlinear activation function in a neural network architecture is crucial and seriously affects the performance of the neural network. An ideal activation function needs to meet many conditions, such as nonlinearity, non-saturation, and almost everywhere differentiability. Existing traditional activation functions all have their own shortcomings. In addition to their nonlinear properties, activation functions have various characteristics that directly affect the performance of the network. Traditional activation functions, such as Logistic Sigmoid and Tanh Unit, map the input values ​​to a small range, which may cause the network gradient to approach zero, thus affecting the training performance. The Rectified Linear Unit (ReLU) and its variants (e.g., LReLU and PReLU) partially solve this problem by mapping the input to a range that is unrestricted in one or two directions. Exponential activation functions such as ELU inherit the advantages of ReLU, but also push the response of the activation function towards zero mean to improve performance. In general, fixed activation functions (AFs) have limited nonlinearity (and therefore limited expressive power) and impose learning biases on the network. Therefore, it is challenging to adapt them to different problem types and data complexities.

[0083] Trainable activation functions (TAFs) such as PReLU, Swish, and PELU adjust the shape of standard fixed activation functions by adding a few learnable parameters. Such functions have only achieved a marginal improvement in expressiveness, as the performance of these TAFs is often similar to their base non-trainable activation functions. Maxout units provide an alternative approach to activation functions. Despite showing significant improvements over ReLU in solving classification tasks, the number of parameters in a Maxout layer grows with the number of neurons in the network. A diffeomorphism is a differentiable reversible function that has a differentiable inverse. Trainable activation functions based on diffeomorphisms are a differentiable parameterization based on efficient and highly-expressive parametric diffeomorphisms (CPAB). Although DiTAC adds only negligible trainable parameters, it is still very expressive. This is in contrast to existing trainable activation functions such as PReLU, Swish, and PELU, which are typically only able to learn a certain shape or only convex functions, while DiTAC is able to learn a variety of shapes to accommodate different types of data and tasks. The working principle of the DiTAC activation function is as follows:

[0084] Assume T θ are diffeomorphisms parameterized by θ. However, using diffeomorphisms usually involves expensive computations. The main reason for adopting CPAB transforms as the family of diffeomorphisms to be used in the DiTAC method is that they are both expressive and efficient. All CPAB transforms are assumed to be one-dimensional.

[0085] set up is a finite interval, let V be from Ω to is a space of continuous functions that are also piecewise affine with respect to some fixed partition of Ω into subintervals. V is a finite-dimensional linear space. Let d = dim(V), let Let v θ ∈V represents a universal element of V parameterized by θ. If x∈Ω, then v θ ∈V defines a trajectory So that φ θ (x,0)=x and φ θ (x, t). The CPAB transformation space definition and its integral equation obtained by integrating the elements in V are:

[0086]

[0087] The equivalent ordinary differential equation is:

[0088] dφ θ(x,t) / dt=v θ (φ θ (x,t)). (2)

[0089] Each T θ ∈T is monotonically increasing and diffeomorphic.

[0090] Unlike existing TAFs, which assign a parameter to each input channel, DiTAC derives its expressive power from the expressive power of the CPAB transform. The expressive power (and differentiability) of DiTAC is apparent from the first data transformation experienced by each neuron. Importantly, the availability of closed-form expressions for the CPAB transform and its gradients allows DiTAC to be easily used as a drop-in replacement for any activation function in any neural network architecture. θ is defined on a finite interval [a, b]. Its range is also a finite interval, which may or may not overlap with Ω. Since some AFs inputs may fall outside [a, b], the mainstream version of DiTAC used in this invention is to convert T θ Combined with GELU, GELU is a widely used AFs in state-of-the-art models. GELU is given by GELU(x) = x·Φ(x), where Φ(x) is the cumulative distribution function of the standard normal distribution. The mainstream form of DiTAC is:

[0091]

[0092] in

[0093]

[0094] In view of the superiority of the DiTAC activation function, the present invention will use it to replace the traditional activation function of the extreme learning machine and propose an extreme learning machine (DiTAC Extreme Learning Machine, DELM) based on a differential homeomorphic trainable activation function.

[0095] DELM algorithm principle:

[0096]

[0097] w is the weight parameter matrix of the input layer and the hidden layer, n is the number of nodes in the input layer, and l is the number of neurons in the hidden layer.

[0098]

[0099] β is the weight parameter matrix of the hidden layer and the output layer, m is the number of nodes in the output layer, and l is the number of neurons in the hidden layer.

[0100]

[0101] b is the hidden layer bias vector.

[0102]

[0103] X is the training input data matrix, that is, the feature matrix. Each column is a set of training input data, with a total of Q sets of training data.

[0104]

[0105] T is the training output target matrix, which is also the label matrix, that is, each column corresponds to the label vector of each set of training input data (a total of Q groups).

[0106]

[0107] in is the DiTAC activation function, and H is the hidden layer output matrix. In the extreme learning machine algorithm, the parameter matrix or vector w, b is randomly generated during training, and the training input data (x1, x2, …, x Q ) is a known quantity. The present invention replaces the Sigmoid activation function in the traditional extreme learning machine with the DiTAC activation function, that is, a trainable parameter θ is introduced in H, which is the hidden layer output matrix. It is learned through network training, so H depends on the trainable parameter θ during model training.

[0108] Ideally, we hope that the hidden layer output matrix H multiplied by the hidden layer and output layer weight parameter matrix β is exactly equal to the transpose of the output target matrix T, which means that the algorithm has completely learned the intrinsic relationship between features and labels and has the ability to classify data:

[0109] Hβ=T T (11)

[0110] But in practical applications, this is extremely rare, and generally only Hβ and T are required. T Close enough. Among them, H and T T is a determined quantity, β and θ are unknown quantities. This model is equivalent to solving the following optimization problem:

[0111]

[0112] That is, by solving the optimal β,θ, the error between the model predicted label and the true label is minimized. Among them, ||·|| is the Frobenius norm of the matrix elements. Further prevent problems such as overfitting and data collinearity, and improve the generalization ability of the model. In order to suppress overfitting, the loss function of the traditional extreme learning machine often performs L2 regularization on the hidden layer and output layer weight parameters β, but this will amplify and reduce the error, that is, the L2 norm is highly sensitive to outliers. In order to promote the sparsity of model parameters and increase the inherent robustness to outliers, the present invention performs L1 regularization on the hidden layer and output layer weight parameters β in the DELM loss function. When the L1 norm is used as a regularization term, the sparsity of model parameters can be promoted. This means that many coefficients will be compressed to zero, thereby achieving feature selection. This is particularly useful in high-dimensional data because it can reduce the complexity of the model and improve interpretability. And compared with the L2 norm, the L1 norm is less sensitive to outliers, which makes it more stable in datasets with noise or outliers. In multi-class classification problems, the L1 norm can promote sparsity of boundaries between categories, which helps improve the performance of the classifier. The invention performs L2 regularization on the trainable parameters in the DELM loss function.

[0113] Therefore, the present invention rewrites the original optimization problem as follows:

[0114]

[0115] Among them, ||·||1 is the L1 norm, λ1 and λ2 are the coefficients of L1 and L2 regularization terms, respectively, and are all hyperparameters. Since the L1 regularization term is a non-smooth term, it is not differentiable. Therefore, this optimization problem is a non-smooth optimization problem, and all the optimization methods commonly used in machine learning that rely on gradient information are invalid. Furthermore, the present invention will use the meta-heuristic swarm intelligence method in derivative-free optimization technology to solve it.

[0116] S5: The Alibaba and the Forty Thieves algorithm with discrete fractional order Caputo is used to solve the optimization model to solve the non-differentiable problem of the optimization model caused by the introduction of the L1 norm, and the optimal weight coefficient and the optimal parameters of the DiTAC activation function that minimize the error between the predicted health status label and the true health status label are obtained; based on the optimal weight coefficient and the optimal parameters of the DiTAC activation function, the optimal parameters of the DiTAC activation function of the sign data monitoring model for mapping physiological sign data to health status labels are obtained.

[0117] Since the DELM loss function contains L1 regularization terms, which are non-smooth terms and non-differentiable. Therefore, this optimization problem is a non-smooth optimization problem, and all optimization methods commonly used in machine learning that rely on gradient information are invalid. Furthermore, the present invention will use the meta-heuristic swarm intelligence method in derivative-free optimization technology to solve it. Specifically, the improved Alibaba and Forty Thieves algorithm based on the discrete fractional-order Caputo method is used to solve the optimization model, and the weight coefficients of the hidden layer and the output layer in the improved extreme learning machine, as well as the trainable parameters in the DiTAC activation function, are obtained.

[0118] In order to optimize the parameters of non-smooth loss functions, the present invention proposes a novel derivative-free optimization method - Alibaba and the Forty Thieves algorithm based on the discrete fractional-order Caputo method. The present invention adopts fractional calculus using Caputo fractional difference and operator to enhance the exploration and development performance of the traditional Alibaba and the Forty Thieves algorithm by utilizing the historical dependence of fractional-order derivatives. The improved Alibaba and the Forty Thieves algorithm uses discrete fractional-order Caputo to update the thief position information carrying β and θ during the optimization process. The discrete fractional-order Caputo assists the Alibaba and the Forty Thieves algorithm in recording historical search experience through its long-term memory effect characteristics to reduce the probability of falling into local optimality.

[0119] In the proposed new derivative-free optimization algorithm, which is also a new Alibaba and Forty Thieves algorithm, fractional-order calculus using Caputo fractional difference and operator is adopted to enhance the exploration and exploitation performance of the traditional Alibaba and Forty Thieves algorithm by utilizing the historical dependence of fractional-order calculus (FC). Enhance the search ability of the algorithm: Fractional-order derivatives can simulate long-term memory effects, which is particularly important for the search process in swarm intelligence algorithms. It can help the algorithm remember the past search history, thereby achieving a better balance between global search and local search, and improving the global search ability and convergence performance of the algorithm. Through fractional-order derivatives, the dynamic characteristics of the algorithm can be adjusted to make it more flexible to adapt to different optimization problems. The introduction of fractional-order derivatives allows the algorithm to dynamically adjust its behavior during the search process to adapt to changes in the problem. Fractional-order derivatives can be used to control the convergence speed and accuracy of the algorithm. By adjusting the order of the fractional order, the search depth and breadth of the algorithm can be controlled, thereby avoiding the algorithm from falling into the local optimal solution too early and improving the reliability and convergence of the algorithm. The introduction of fractional derivatives can improve the robustness of the algorithm to different types of problems and different initial conditions. It can help the algorithm maintain stable performance when facing complex and changing environments.

[0120] First, the discrete fractional order Caputo used in the following steps 6 to 8 to update the thief's position is explained:

[0121] The Caputo form of fractional backward differencing (C-FOBD) is:

[0122]

[0123] The finite memory window version of C-FOBD with window size can be written as:

[0124]

[0125] For 0<v<1, equation (15) can be written as:

[0126]

[0127] Initialize by randomly initializing the positions of a number of n individuals in the d-dimensional search space.

[0128] The optimization process using Alibaba and the Forty Thieves algorithm specifically includes the following steps:

[0129] Step 1. Set the parameters of the Alibaba and the Forty Thieves algorithm: determine the number of thieves (i.e., the number of solutions in the search space); set algorithm parameters such as the number of iterations and the search interval; define the objective function, i.e., improve the objective function of the extreme learning machine.

[0130] Step 2, initialize the thief position: randomly generate a set of initial solutions in the search space, each solution represents a set of possible weight coefficients β and DiTAC activation function trainable parameters θ; set n individuals, each of which includes a set of w, b, θ and β, where w and b (randomly generated by the improved extreme learning machine, once generated, they will not change and are known parameters) are the weight parameter matrix and bias of the input layer respectively:

[0131]

[0132] Where x is the location of all gang members, d is the number of variables in a given problem, and xij represents the jth dimension component of the i-th member.

[0133] The initial position of the population can be generated by the following equation:

[0134] x i = l j +r0×(u j -l j ) (18)

[0135] The position of the i-th member of xi represents a candidate solution to a problem, lj and uj are the lower and upper bounds of the j-th dimension component, and r0 is a random number generated in (0,1) that satisfies a uniform distribution.

[0136] Step 3: For each thief (i.e., each set of β and θ), initialize the best position for the thief. i t ) and the global best position (gbest t ).

[0137] Step 4: Initialize the maid's intelligence level on all thieves.

[0138] Step 5. Use the objective function to evaluate the positions of all thieves.

[0139] There are three basic scenarios that can occur as the algorithm searches for a solution. In each case, we assume that each member is effectively exploring the surrounding search space, and that there is also a portion of the maid’s intelligence and planning that forces members to explore in random locations. These three scenarios are as follows.

[0140] Step 6, update the thief's position: according to the "thief gets information and pursues Alibaba" strategy (exploration phase) of Alibaba and the Forty Thieves algorithm, update the thief's position (i.e. β and θ); in each iteration, record the optimal solution currently found (i.e. β and θ with the minimum objective function value).

[0141] The thieves got the information and went after Alibaba.

[0142] The thieves may track down Alibaba with the help of information they got from someone, in which case the thieves' new location is as follows:

[0143]

[0144] Rewritten as discrete form:

[0145]

[0146] is equivalent to:

[0147]

[0148] Where xit+1 represents the position of thief member i at iteration (t+1), yit represents the position of member i relative to Alibaba at iteration t, bestit represents the best position reached by member i at iteration t, and gbestt represents the global best position obtained by any member at iteration t. λ is the global information parameter, which is a hyperparameter.

[0149] ma(i)t represents the wisdom level of the maid used to disguise member i in the tth iteration. The a of ma(i)t is:

[0150]

[0151] where rand(n, 1) represents a random number vector generated with uniform distribution in the range [0, 1]. The symbol is rounded up. If the quality of the new solution that the thieves come up with is better than their previous position, the maid will update her shrewd plan. In this case, Equation 7 can be used to update her intelligence:

[0152]

[0153] Where f() represents the score of the fitness function.

[0154] Tdt represents the search distance of the member at the tth iteration:

[0155]

[0156] Where t and T represent the current iteration number and the maximum iteration number respectively, α0 (α0=1) represents the initial estimate of the tracking distance at the first iteration, and α1 is a constant value used to manage exploration and development capabilities.

[0157] Pp t represents the member's perceived potential of Alibaba at the tth iteration:

[0158]

[0159] where β0 (β0=0.1) represents the final rough estimate of the probability that the thief will achieve his goal at the end of the iterative process, and β1 is a constant value used to manage exploration and exploitation capabilities.

[0160] r1, r2, r4, and rand are random numbers that are uniformly distributed in the interval (0, 1). r3 is greater than or less than 0.5, indicating whether the information obtained by the thieves is true or false. sgn (rand-0.5) randomly gives 1 or -1 to change the direction of the search process.

[0161] Step 7, update the thief's position: according to the "thief is deceived, randomly explore Alibaba" strategy (development phase) of Alibaba and the Forty Thieves algorithm, update the thief's position (i.e. β and θ); in each iteration, record the currently found optimal solution (i.e. β and θ with the minimum objective function value).

[0162] The thieves may realize that they have obtained a wrong information, so they will randomly explore the search space to find Alibaba's location. In this case, the new positions of the members that can be obtained are as follows:

[0163]

[0164] Because the thieves have a high level of intelligence to identify the most suitable area in the search space, they can further infer the location of Alibaba's home.

[0165] Step 8. Update the thief's position: According to the "balanced global and local optimization" strategy (balance phase) of Alibaba and the Forty Thieves algorithm, update the thief's position (i.e., β and θ); in each iteration, record the currently found optimal solution (i.e., β and θ with the minimum objective function value).

[0166] S405: Balancing global and local optimization

[0167] In order to further improve the feasibility of optimization and the accuracy of exploration, this paper also considers exploration in other positions. In this case, the following new positions of members can be obtained:

[0168]

[0169] Step 9: Check whether the preset number of iterations is reached or a solution is found that satisfies the objective function value less than a certain threshold. If so, return the individual carrying the weight coefficient; otherwise, return to step 3.

[0170] After many iterations, the algorithm will eventually output the most suitable member position, which is also the position of Alibaba as the optimal solution to the problem.

[0171] The main idea of ​​the present invention is to use fractional-order reverse spread entropy for comprehensive feature extraction, and then input the obtained features into an extreme learning machine with a trainable activation function based on differential homeomorphism for health warning multi-classification processing, and finally use Alibaba and Forty Thieves algorithm based on discrete fractional-order Caputo method to optimize the network parameters. The health warning algorithm of the smart wearable device for power workers should include an input layer of 4 neurons, which input 4 feature data of body temperature, heart rate, blood oxygen, and brain waves respectively. There are three warning categories, namely normal, mild risk, and high risk. Of course, the number of features and categories can be increased or decreased according to the actual application scenario. The model construction method provided by the present invention has the following advantages:

[0172] (1) In view of the defects of existing methods based on time domain and frequency domain, the present invention will use fractional reverse spread entropy to extract data features. On the one hand, reverse spread entropy is an emerging complexity measurement method, which has many advantages in feature extraction. It improves the detection ability of time series complexity by introducing amplitude information and distance information. Compared with traditional permutation entropy, RDE performs better in detecting mutation signals and noise robustness testing, and has better stability for analog and sensor signals. RDE is better than PE method in distinguishing different data sets, which makes it more effective in processing data with spike features or sudden amplitude changes. In addition, it has high stability for time series of different lengths, which is particularly important for processing variable length signals that may be encountered in practical applications. RDE also shows good robustness under high noise levels and can extract complexity information from data even if there are spike features or sudden amplitude changes in the signal. On the other hand, fractional entropy is an entropy measurement method based on fractional calculus theory, which has significant advantages in feature extraction and system complexity analysis. Enhanced complexity measurement capability: Fractional entropy enhances the complexity measurement capability by introducing fractional derivatives, and can capture the complexity and chaotic characteristics of time series more finely. This method is not only applicable to periodic data, but also can measure chaotic sequences well, providing a deeper understanding of the degree of randomness or disorder in the macroscopic structure of the system. Fractional entropy can distinguish different types of dynamic behaviors and has better distinguishing and identifying capabilities. In the calculation process of fractional entropy, the selection of parameters provides more flexibility. By adjusting the fractional parameters, it can be optimized for different data characteristics and analysis requirements, thereby improving the adaptability and accuracy of entropy measurement. The introduction of fractional entropy provides a new perspective for entropy theory and has the potential to be combined with existing methods. This combination may lead to a more efficient and more accurate feature extraction method. Therefore, the present invention proposes fractional reverse spread entropy, which is a complexity measurement method based on fractional calculus and entropy theory, which combines the flexibility, historical dependence and advantages of reverse spread entropy of fractional calculus.

[0173] (2) The present invention achieves the purpose of establishing a lightweight model that is efficient, lightweight, simple in principle, and easy to implement on resource-limited devices. It combines the extreme learning machine with the cutting-edge trainable activation function based on differential homeomorphism, and proposes an extreme learning machine based on differential homeomorphism trainable activation function. On the one hand, a major advantage of ELM is its extremely fast learning speed. Compared with traditional gradient descent-based algorithms, it does not require iterative weight adjustment in the hidden layer, which greatly reduces the training time. The implementation of ELM is relatively simple because it randomly generates the weights and biases of the hidden layer without the need for complex parameter tuning. Its training process only requires setting the number of hidden layer neurons, which reduces human intervention and makes model training more automated. Even when the parameters of the hidden layer neurons are randomly generated, ELM can still maintain good interpolation ability, universal approximation ability and classification ability. On the other hand, DiTAC is based on an efficient differential homeomorphism transformation (called CPAB) and has high expressiveness. Even if only a very small number of trainable parameters are introduced, it can significantly enhance the expressiveness and performance of the model. DiTAC is able to learn activation functions of various shapes, which is in contrast to existing trainable activation functions, which are usually limited to learning a certain shape or only convex functions. DiTAC only adds negligible trainable parameters while maintaining high expressive power. The present invention replaces the Sigmoid activation function in the traditional extreme learning machine with the DiTAC activation function, which essentially replaces the fixed activation function with an activation function containing trainable parameters. Although the trainable parameters of the hidden layer are increased, the fitting and learning capabilities of the network are also greatly improved.

[0174] (3) In order to suppress overfitting, the loss function of the traditional extreme learning machine often performs L2 regularization on the weight parameters of the hidden layer and the output layer, but this will amplify and reduce the error, that is, the L2 norm is highly sensitive to outliers. In order to promote the sparsity of model parameters and increase the inherent robustness to outliers, the present invention performs L1 regularization on the weight parameters of the hidden layer and the output layer in the DELM loss function. When the L1 norm is used as a regularization term, the sparsity of the model parameters can be promoted. This means that many coefficients will be compressed to zero, thereby achieving feature selection. This is particularly useful in high-dimensional data because it can reduce the complexity of the model and improve interpretability. And compared with the L2 norm, the L1 norm is less sensitive to outliers, which makes it more stable in data sets with noise or outliers. In multi-class classification problems, the L1 norm can promote the sparsity of the boundaries between categories, which helps to improve the performance of the classifier. The present invention replaces the Sigmoid activation function in the traditional extreme learning machine with the DiTAC activation function, so a trainable parameter is introduced into the output matrix of the hidden layer. The trainable parameters are learned through network training and are not easy to be sparse, which will affect the parameter learning effect. Furthermore, the present invention performs L2 regularization on the trainable parameters in the DELM loss function. Since the regularization term contained in the DELM loss function is a non-smooth term, it is not differentiable. Therefore, this optimization problem is a non-smooth optimization problem, and the optimization methods commonly used in machine learning that rely on gradient information are all invalid. Furthermore, the present invention will adopt the meta-heuristic swarm intelligence method in the derivative-free optimization technology to solve it.

[0175] (4) In order to optimize the parameters of non-smooth loss functions, the present invention proposes a novel derivative-free optimization method, namely, Alibaba and the Forty Thieves algorithm based on discrete fractional-order Caputo method. The present invention adopts fractional calculus using Caputo fractional difference sum operator to enhance the exploration and exploitation performance of the traditional Alibaba and the Forty Thieves algorithm by utilizing the historical dependence of fractional derivatives.

[0176] The above is a method for constructing a vital sign data monitoring model provided by one or more embodiments of the present invention. Based on the same idea, the present invention also provides a corresponding device for constructing a vital sign data monitoring model, including:

[0177] The extreme learning machine improvement module is used to form an improved hidden layer by replacing the Sigmoid activation function of the hidden layer in the original extreme learning machine with a DiTAC activation function, wherein the DiTAC activation function performs a non-fixed nonlinear operation on the input data by introducing a trainable parameter; and construct an improved extreme learning machine including an input layer, an improved hidden layer and an output layer.

[0178] The optimization model construction module is used to collect physiological sign data and the corresponding real health status labels to construct a data set; the physiological sign data in the data set is input into the improved extreme learning machine, and the objective function is to minimize the error between the predicted health status label and the real health status label corresponding to the physiological sign data output by the improved extreme learning machine. The L1 norm of the output layer weight coefficient and the L2 norm of the trainable parameters introduced by the DiTAC activation function are used as constraints to construct an optimization model.

[0179] The vital sign data monitoring model acquisition module is used to solve the optimization model to obtain the optimal weight coefficient and the optimal parameters of the DiTAC activation function that minimize the error between the predicted health status label and the true health status label; based on the optimal weight coefficient and the optimal parameters of the DiTAC activation function, a vital sign data monitoring model for mapping physiological vital sign data into health status labels is obtained.

[0180] For the specific definition of the construction device of the vital sign data monitoring model, please refer to the definition of the construction method of the vital sign data monitoring model in the above text, which will not be repeated here. Each module in the construction device of the above-mentioned vital sign data monitoring model can be implemented in whole or in part by software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0181] The present invention also provides a computer-readable storage medium, which stores a computer program, which can be used to execute the above Figure 1 A method for constructing a vital sign data monitoring model is provided.

[0182] The present invention also provides a computer device structure. At the hardware level, the computer device includes a processor, an internal bus, a network interface, a memory, and a non-volatile memory. Of course, it may also include other hardware required for the business. The processor reads the corresponding computer program from the non-volatile memory into the memory and then runs it to achieve the above Figure 1 A method for constructing a vital sign data monitoring model is provided.

[0183] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided by the present invention can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory or optical memory, etc. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).

[0184] The technical features of the above embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of the present invention.

Claims

1. A method for constructing a vital sign data monitoring model, characterized in that: include: An improved hidden layer is formed by replacing the Sigmoid activation function of the hidden layer in the original extreme learning machine with a DiTAC activation function, wherein the DiTAC activation function performs a non-fixed nonlinear operation on the input data by introducing a trainable parameter; and constructing an improved extreme learning machine including an input layer, an improved hidden layer, and an output layer; Collect physiological sign data and corresponding real health status labels to construct a data set; input the physiological sign data in the data set into an improved extreme learning machine, take minimizing the error between the predicted health status label and the real health status label corresponding to the physiological sign data output by the improved extreme learning machine as the objective function, and take the L1 norm of the output layer weight coefficient and the L2 norm of the trainable parameter introduced by the DiTAC activation function as constraints to construct an optimization model; The optimization model is solved to obtain the optimal weight coefficient and the optimal parameters of the DiTAC activation function that minimize the error between the predicted health status label and the true health status label; based on the optimal weight coefficient and the optimal parameters of the DiTAC activation function, a vital sign data monitoring model for mapping physiological sign data to health status labels is obtained.

2. The method for constructing a vital sign data monitoring model according to claim 1, characterized in that: The construction of the optimization model specifically includes: in, is the objective function, H is the output matrix of the improved hidden layer in the improved extreme learning machine, β is the weight matrix of the output layer in the improved extreme learning machine, Hβ is the predicted health status label output by the improved extreme learning machine; T is the predicted health status label, and the superscript T represents the transposition operation; θ is the trainable parameter introduced by the DiTAC activation function in the improved hidden layer; λ1||β||1+λ2||θ||2 is the constraint condition, ||β||1 is the L1 norm of β, and λ1 is the L1 regularization term coefficient; ||θ||2 is the L2 norm of θ, and λ2 is the L2 regularization term coefficient.

3. The method for constructing a vital sign data monitoring model according to claim 2, characterized in that: The H is the output matrix of the improved hidden layer in the improved extreme learning machine, specifically including: Among them, w1~w l is the weight coefficient vector of the input layer; b1~b l is the bias vector of the input layer; l is the number of neurons in the improved hidden layer; x1~x Q Training data for group Q; is the DiTAC activation function that introduces the trainable parameter θ.

4. The method for constructing a vital sign data monitoring model according to claim 2, characterized in that: The solving of the optimization model is to solve the optimization model by using a meta-heuristic swarm intelligence method in derivative-free optimization technology to solve the problem that the optimization model is not differentiable due to the introduction of the L1 norm; The meta-heuristic swarm intelligence method in the derivative-free optimization technology is to improve the Alibaba and Forty Thieves algorithm. The improved Alibaba and Forty Thieves algorithm uses discrete fractional order Caputo to update the thieves' position information carrying β and θ in the optimization process. The discrete fractional order Caputo assists the Alibaba and Forty Thieves algorithm to record historical search experience through its long-term memory effect characteristics, so as to reduce the probability of falling into local optimality.

5. The method for constructing a vital sign data monitoring model according to claim 1, characterized in that: Before inputting the physiological sign data in the data set into the improved extreme learning machine, feature extraction is performed on the physiological sign data to obtain a fractional order reverse scatter entropy for characterizing the temporal complexity of the physiological sign data and the data mutation behavior; specifically including: The physiological sign data with a time length of T is mapped to the normal distribution space through the normal distribution function: X={x(i),i=1,2,…,T}→Y={y(i),i=1,2,…,T} Among them, x(i) is the physiological sign data at the i-th moment; y(i) is the mapping value corresponding to x(i); Through linear changes Map Y to a categorical form to obtain the category labels corresponding to each component in Y: Y={y(i),i=1,2,...,T}→Z={z(i),i=1,2,...,T} Among them, int() is the rounding function, c is the number of categories, and z(i) is a positive integer ranging from 1 to c; Reconstruct the category form Z into L embedding vectors with time delay d and embedding dimension m, each of which includes m components ranging from 1 to c; Map each embedding vector to a scatter pattern π i Calculate the probability P(π i ): Among them, P(π i ) is represented by the ratio of the number of the ith scatter pattern to the number of embedded vectors; Number(π i ) is the number of embedding vectors mapped to the i-th scatter pattern; c m is the number of all scatter modes; Based on the fractional order probability P(π i ) is weighted to obtain the fractional reverse diffusion entropy FRDE: Among them, α is the fractional order; Γ(·) is the gamma function; ψ(·) is the digamma function.

6. The method for constructing a vital sign data monitoring model according to claim 1, characterized in that: The physiological sign data include body temperature data, heart rate data, blood oxygen data and brain wave data.

7. The method for constructing a vital sign data monitoring model according to claim 1, characterized in that: The real health status labels include normal labels, low-risk labels and high-risk labels.

8. A device for constructing a vital sign data monitoring model, characterized in that: include: An extreme learning machine improvement module, used for forming an improved hidden layer by replacing the Sigmoid activation function of the hidden layer in the original extreme learning machine with a DiTAC activation function, wherein the DiTAC activation function performs a non-fixed nonlinear operation on the input data by introducing a trainable parameter; and constructing an improved extreme learning machine including an input layer, an improved hidden layer, and an output layer; The optimization model construction module is used to collect physiological sign data and corresponding real health status labels to construct a data set; the physiological sign data in the data set is input into the improved extreme learning machine, and the objective function is to minimize the error between the predicted health status label and the real health status label corresponding to the physiological sign data output by the improved extreme learning machine, and the L1 norm of the output layer weight coefficient and the L2 norm of the trainable parameter introduced by the DiTAC activation function are used as constraints to construct the optimization model; The vital sign data monitoring model acquisition module is used to solve the optimization model to obtain the optimal weight coefficient and the optimal parameters of the DiTAC activation function that minimize the error between the predicted health status label and the true health status label; based on the optimal weight coefficient and the optimal parameters of the DiTAC activation function, a vital sign data monitoring model for mapping physiological vital sign data into health status labels is obtained.

9. A computer-readable storage medium, characterized in that: The storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

10. A computer device, characterized in that: The method comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the method described in any one of claims 1 to 7 is implemented.

Images