Neural network model signal classification method and system based on fractional order interval extreme learning machine

Abstract

The invention discloses a signal classification method and system for a neural network model based on a fractional order interval extreme learning machine, and the method comprises the steps: converting collected uncertain signal data into interval type data, and carrying out the normalization preprocessing of the data; establishing a three-layer feedforward neural network model based on a fractional order interval extreme learning machine, and setting corresponding parameters such as a fractional order number, a learning rate and a hidden node number according to preprocessed data features; taking the preprocessed data as training data, calculating actual output of a model through features of the training data, and constructing an error function E in combination with the actual output and ideal output; training a weight minimization error function E in the network model in a multi-iteration process; the fractional order interval extreme learning machine can effectively process interval uncertainty signal data, the classification performance and convergence speed of the neural network model are further improved through introduction of the fractional order derivative, and therefore the excellent purpose is achieved in a complex signal classification task.

Description

Technical Field

[0001] The present invention relates to the field of classification methods, and in particular to a signal classification method and system based on a neural network model of a fractional-order interval extreme learning machine. Background Art

[0002] At present, intelligent communication systems are usually composed of multiple modules: (1) Signal generation and processing module, whose main task is to collect and preprocess data. (2) Feature extraction module, whose main task is to extract representative features from the collected data as input to the neural network classifier. (3) Neural network classifier module, whose main task is to use the neural network to classify the extracted features and identify the modulation format or channel conditions of the signal. (4) Decision module, which makes the final decision based on the output of the neural network.

[0003] Neural network models play an important role as decision makers in signal processing. With the development of deep learning and neural networks, many problems in signal processing (such as channel estimation, equalization, noise reduction, demodulation, etc.) can be solved by well-trained neural networks. However, in the real world, communication data is not completely certain, but has uncertainty and high noise. This uncertainty may come from factors such as multipath effects, channel fading, interference, and equipment errors, which makes intelligent communication systems face many challenges when processing data, especially the classification accuracy of neural network models for uncertain data.

[0004] Taking the background of determining the user's indoor location by observing uncertain signal data on a smartphone, the following is explained: First, the signal generation and processing module collects and preprocesses the uncertain signal data. Secondly, the feature extraction module extracts features for neural network model training. After that, the neural network classifier trains, verifies and tests the data set, and optimizes the model through multiple iterations. Finally, the decision module determines the indoor location of the smartphone user based on the output of the neural network. In addition, the signal generation and processing module allows the collection of multiple types of uncertain signals, such as electrocardiograms (collected by biomedical sensors), interference signals (interference signals in wireless communications), radar reflection signals (echo signals collected by radar systems), mechanical vibration signals, sound wave signals (vibration or sound wave signals generated by industrial equipment), etc.

[0005] However, in the existing technology, traditional neural networks have limitations when processing complex and uncertain signals, while interval neural networks are better at processing fuzzy and noisy data. Among the many types of neural networks, interval neural networks have advantages in processing uncertainty and noise, and can better simulate and cope with the complexity of the real world. Interval neural networks can better cope with fuzzy and noisy data in the real world by introducing interval values ​​to represent the uncertainty of data during the training process. The model performs well in processing classification problems with interval uncertainty and has been verified in a variety of application scenarios. However, traditional interval neural networks have high computational costs when facing high-dimensional and complex classification tasks, and there is still room for optimization in terms of convergence speed and model generalization ability.

[0006] Extreme learning machine is a new type of fast learning algorithm in recent years, which has the characteristics of fast training speed and global optimality. The traditional extreme learning machine can complete the training of the neural network with only one calculation through Moore-Penrose generalized inverse, and has high-speed computing power and good generalization performance when processing large-scale data. However, the traditional extreme learning machine cannot be directly introduced into the interval neural network due to nonlinear constraints. When the extreme learning machine uses Moore-Penrose generalized inverse to calculate the output layer weights of the interval neural network: θ = H + U, where θ is the output weight, H + is the Moore-Penrose generalized inverse of the hidden layer output matrix, and U is the ideal output. There are nonlinear constraints: Among them, I L is the L×L identity matrix, 0 L is an L×L zero matrix, is the least squares solution matrix. Summary of the invention

[0007] The purpose of the present invention is to solve the problems in the prior art that when faced with high-dimensional, complex, and highly uncertain signal classification tasks, the computational cost is high, and the convergence speed and generalization ability of the model still need to be optimized. At the same time, in order to solve the problem that the extreme learning machine cannot be applied to the interval neural network due to nonlinear constraints and to ensure that the optimized model has a faster learning speed and significant performance.

[0008] In order to solve the above problems, the present invention proposes a signal classification method based on a neural network model of a fractional-order interval extreme learning machine, which includes the following steps, all of which can be implemented using Matlab:

[0009] Step S1: The uncertain signal is converted into uncertain data by orthogonal frequency division multiplexing combined with multi-level orthogonal amplitude modulation;

[0010] Uncertain data consists of multiple samples s j Composition, each sample s j It includes N feature values ​​and 1 label class. The specific information of the uncertain signal is:

[0011]

[0012] Step S2: converting the uncertain data into interval type data and performing normalization preprocessing on the interval type data. The specific steps include:

[0013] Step S2-1: Convert uncertain data into interval type data (xj,uj);

[0014] x j =[x j1 ,...,x jn ] T , x ji = <x C ,x R >,u j =[u j1 ,...,u jm ] T ,u jt = C ,u R >, j=1,...,J, i=1,...,n, t=1,...,m

[0015] In the formula, x represents the input of interval data, u represents the ideal output of interval data, J represents the total number of interval data samples, n represents the number of input features of interval data, m represents the dimension of interval output, C superscript represents the center of interval data, and R superscript represents the radius of interval data;

[0016] Step S2-2: Normalization preprocessing, maximum and minimum normalization formula:

[0017]

[0018] In the formula, max is the interval type data x j The maximum value of the eigenvalue, min is the interval type data x j The minimum eigenvalue of x is x j An eigenvalue of * To normalize the preprocessed eigenvalues, the eigenvalues ​​of each sample are preprocessed to the range [0,1] using the maximum and minimum normalization formula, and the label data remains unchanged;

[0019] Step S3: Establishing a three-layer feedforward neural network model based on fractional-order interval extreme learning machine;

[0020] ​Step S3-1: using the interval data preprocessed according to the method of step S3 as input data to set the corresponding parameters of fractional order α, learning rate η and number of hidden nodes L;

[0021] Step S3-1-1: Establish a three-layer neural network including an input layer, a hidden layer, and an output layer, where the input of the input layer is:

[0022] x j =[x j1 ,...,x jn ,...,x jN ]

[0023] In the formula, x jN Represents the N feature data of the jth sample, and the weight and threshold between the input layer and the hidden layer are:

[0024] (w,b)=[(w1,b1),...,(w L ,b L )]

[0025] In the formula, L is the number of hidden nodes, w represents the weight between the input layer and the hidden layer, b represents the threshold between the input layer and the hidden layer, and the output of the hidden layer is:

[0026] h l =[h1,...,h l ,...,h L ]

[0027] In the formula, h l Represents the output of the lth hidden node, and the weight between the hidden layer and the output layer is θ l =[θ1,...,θ L ], where L is the number of hidden nodes, and the output of the output layer is:

[0028] o j =[o j1 ,...,o jm ,...,o jM ]

[0029] In the formula, o jM represents the M output dimensions of the j-th sample;

[0030] Step S3-1-2: Set parameters, fractional order α value: 9 / 11, learning rate η value: grid search method to find the optimal learning rate, hidden layer node number L value: grid search method to find the optimal number of nodes, minimum error E_min value: 0.00001 and maximum iteration step number K value: 300, randomly generate and fix the weight w between the input layer and the hidden layer value: range [0,1] and threshold b value: range [0,1], initialize the weight θ between the hidden layer and the output layer value: range [0,1];

[0031] Step S4: Use the interval data preprocessed in step S2 as training data, calculate the actual output of the model through the characteristics of the training data, and construct the error function E by combining the actual output with the ideal output:

[0032] Step S4-1: Calculate the actual output:

[0033]

[0034] In the formula, θ l =[θ1,...,θ L ] is the weight between the hidden layer and the output layer, and the hidden node output is:

[0035]

[0036] Where: l is the number of hidden nodes, g(x) is the sigmoid activation function, and the error function E is calculated:

[0037]

[0038] Step S5: Minimize the error function E by training the network model through K iterations;

[0039] Step S5-1: Weight update formula:

[0040]

[0041]

[0042] Where D is the abbreviated symbol of the fractional derivative of the error function, caputo is the Caputo fractional derivative, k represents the current iteration number, and sgn(θ) is the symbolic function:

[0043]

[0044] Γ(x) is denoted as the gamma function;

[0045] Step S5-2: Repeat step S5 for K times to minimize the error function E and obtain the optimal configuration of the neural network model.

[0046] In a preferred embodiment, the feature value is a feature of the data type of the sample, and the label class is a category of the sample label.

[0047] A signal classification system based on a neural network model of a fractional-order interval extreme learning machine includes the following modules: a data conversion module, a normalization module, a model creation module, and a training module;

[0048] Step M1: The data conversion module converts the uncertain signal into uncertain data by combining orthogonal frequency division multiplexing with multi-level orthogonal amplitude modulation;

[0049] Uncertain data consists of multiple samples s j Composition, each sample s j Includes N feature values ​​and 1 label class;

[0050] Step M2: The data conversion module converts the uncertain data into interval type data, and performs normalization preprocessing on the interval type data through the normalization module. The specific steps include:

[0051] Step M2-1: Convert uncertain data into interval type data (xj,uj);

[0052] x j =[x j1 ,...,x jn ] T , x ji = <x C ,x R >,u j =[u j1 ,...,u jm ] T ,u jt = C ,u R >, j=1,...,J, i=1,...,n, t=1,...,m

[0053] In the formula, x represents the input of interval data, u represents the ideal output of interval data, J represents the total number of interval data samples, n represents the number of input features of interval data, m represents the dimension of interval output, C superscript represents the center of interval data, and R superscript represents the radius of interval data;

[0054] Step M2-2: Normalization preprocessing, maximum and minimum normalization formula:

[0055]

[0056] In the formula, max is the interval type data x j The maximum value of the eigenvalue, min is the interval type data x j ​The minimum eigenvalue of x is x j An eigenvalue of * To normalize the preprocessed eigenvalues, the eigenvalues ​​of each sample are preprocessed to the range [0,1] using the maximum and minimum normalization formula, and the label data remains unchanged;

[0057] Step M3: The model creation module establishes a three-layer feedforward neural network model based on the fractional-order interval extreme learning machine;

[0058] Step M3-1: Use the interval data preprocessed according to the method of step M2 as input data to set the corresponding parameters of fractional order α, learning rate η and number of hidden nodes L:

[0059] Step M3-1-1: Establish a three-layer neural network including input layer, hidden layer, and output layer, where the input of the input layer is:

[0060] x j =[x j1 ,...,x jn ,...,x jN ]

[0061] In the formula, x jN Represents the N feature data of the jth sample, and the weight and threshold between the input layer and the hidden layer are:

[0062] (w,b)=[(w1,b1),...,(w L ,b L )]

[0063] In the formula, L is the number of hidden nodes, w represents the weight between the input layer and the hidden layer, b represents the threshold between the input layer and the hidden layer, and the output of the hidden layer is:

[0064] h l =[h1,...,h l ,...,h L ]

[0065] In the formula, h l Represents the output of the lth hidden node, and the weight between the hidden layer and the output layer is θ l =[θ1,...,θ L ], where L is the number of hidden nodes, and the output of the output layer is:

[0066] o j =[o j1 ,...,o jm ,...,o jM ]

[0067] In the formula, o jM represents the M output dimensions of the j-th sample;

[0068] Step M3-1-2: Set parameters, fractional order α value: 9 / 11, learning rate η value: grid search method to find the optimal learning rate, hidden layer node number L value: grid search method to find the optimal number of nodes, minimum error E_min value: 0.00001 and maximum iteration step number K value: 300, randomly generate and fix the weight w between the input layer and the hidden layer value: range [0,1] and threshold b value: range [0,1], initialize the weight θ between the hidden layer and the output layer value: range [0,1];

[0069] Step M4: Using the interval data preprocessed in step M2 as training data, the training module calculates the actual output of the model through the characteristics of the training data, and constructs the error function E by combining the actual output with the ideal output:

[0070] Step M4-1: Calculate the actual output:

[0071]

[0072] Among them, θ l =[θ1,...,θ L ] is the weight between the hidden layer and the output layer, and the hidden node output is:

[0073]

[0074] Where: l is the number of hidden nodes, g(x) is the sigmoid activation function, and the error function E is calculated:

[0075]

[0076] Step M5: Minimize the error function E by training the network model through K iterations;

[0077] Step M5-1: Weight update formula:

[0078]

[0079]

[0080] Where D is the abbreviated symbol of the fractional derivative of the error function, caputo is the Caputo fractional derivative, k represents the current iteration number, and sgn(θ) is the symbolic function:

[0081]

[0082] Γ(x) is denoted as the gamma function;

[0083] Step M5-2: Repeat step M5 for K times to minimize the error function E, thereby obtaining the optimal configuration of the neural network model.

[0084] Orthogonal frequency division multiplexing combined with multi-level quadrature amplitude modulation, the specific reference is: Oruthota, U. et al "SER / BER expression for M-QAM OFDM systems with imperfect channel estimation and I / Q imbalance." EURASIP Journal on Wireless Communications and Networking 2012 (2012): 1-14.

[0085] Beneficial effects of the invention: The invention solves the problem that traditional interval neural networks have high computational costs when facing high-dimensional and complex signal classification tasks, and the convergence speed and generalization ability of the model still need to be optimized, which in turn leads to a decrease in the overall performance of the intelligent communication system and affects its practicality in the real world. Based on the convergence speed and classification accuracy as optimization goals, the corresponding neural network model configuration is given to complete the establishment of the model for the classification task. In addition, two experimental results are given to verify that this innovation improves the convergence speed and classification accuracy of the interval neural network model. The model can effectively process interval uncertainty and high-noise signal data, and the classification performance and convergence speed of the model are further improved by the introduction of fractional-order derivatives, thereby showing excellent performance in complex classification tasks.

[0086] A fractional-order extreme learning machine is designed to optimize the existing interval neural network model. This optimization method solves the problem that the extreme learning machine cannot be applied to the interval neural network due to nonlinear constraints, and ensures that the optimized model has a faster learning speed and significant classification accuracy. The fractional-order extreme learning machine uses the gradient descent method to update the output layer weights, which effectively solves the nonlinear constraints of the interval neural network. At the same time, with the characteristics of fractional-order gradient descent, such as memory and non-locality, the fractional-order extreme learning machine can provide richer dynamic behaviors and better convergence characteristics when optimizing the neural network model.

[0087] In addition, the model of the prior art uses the fractional gradient descent method to train weights, which is mainly applied to the classification of non-interval data. The model of the present invention is improved to realize the application of the fractional gradient descent method to train weights, thereby meeting the use of interval data classification and achieving the above-mentioned beneficial effects. BRIEF DESCRIPTION OF THE DRAWINGS

[0088] Figure 1It is a communication system module diagram of a signal classification method of a neural network model based on a fractional-order interval extreme learning machine of the present invention;

[0089] Figure 2 It is a flow chart of a signal classification method of a neural network model based on a fractional-order interval extreme learning machine of the present invention;

[0090] Figure 3 It is an algorithm flow chart of a signal classification method based on a neural network model of a fractional-order interval extreme learning machine of the present invention;

[0091] Figure 4 It is a neural network structure diagram of a signal classification method of a neural network model based on a fractional-order interval extreme learning machine of the present invention.

[0092] Figure 5 It is a graph showing the classification accuracy changes of the fractional-order interval extreme learning machine and the interval extreme learning machine of the present invention on the task 1 data set;

[0093] Figure 6 It is a graph showing the changes in classification accuracy of the fractional-order interval extreme learning machine and the interval extreme learning machine of the present invention on the task 2 data set. DETAILED DESCRIPTION

[0094] Embodiment 1:

[0095] Step 1: First, the signal generation and processing module is responsible for collecting uncertain signals through sensors. The specific information of uncertain signals is shown in Table 1.

[0096] Table 1 Specific information of uncertain signals

[0097]

[0098] Secondly, the collected uncertain signals are converted into uncertain data through orthogonal frequency division multiplexing combined with multi-level orthogonal amplitude modulation. Among them, the uncertain data pattern consists of multiple samples xj, each sample xj includes N feature values ​​and 1 label class, as shown in Table 2. Among them, the features need to be converted into feature values ​​that can be represented by data, and the labels need to be converted into corresponding label classes.

[0099] Table 2 Classification results of fractional-order interval extreme learning machine and interval extreme learning machine on task 1 dataset

[0100]

[0101] Then, the collected uncertain data is converted into interval data in two steps, and the data is normalized and preprocessed:

[0102] Uncertain data is converted into interval data (xj ,u j ).

[0103] where x j =[x j1 ,...,x jn ] T , x ji = <x C ,x R >,u j =[u j1 ,...,u jm ] T ,u jt = C ,u R >, j=1,...,J, i=1,...,n, t=1,...,m.

[0104] Where: x represents the input of the interval data, u represents the ideal output of the interval data. J represents the total number of interval data samples, n represents the number of input features of the interval data, and m represents the dimension of the interval output. The C superscript represents the center of the interval data, and the R superscript represents the radius of the interval data.

[0105] Normalization preprocessing: Maximum and minimum normalization formula: max is the interval type data xj The maximum value of the eigenvalue, min is the minimum value of the eigenvalue of interval type data xj, x is an eigenvalue of xj, and x* is normalized The eigenvalues ​​after preprocessing. The feature data of each sample is preprocessed to the range [0,1] using the maximum and minimum normalization formula, and the label data remains unchanged.

[0106] Step 2: Establish a three-layer feedforward neural network model based on fractional-order interval extreme learning machine, which is divided into two steps. According to the input data preprocessed by the method in step 1, set the corresponding parameters of fractional order α, learning rate η and number of hidden nodes L:

[0107] Use Matlab to build a three-layer neural network including input layer, hidden layer and output layer (see attached) Figure 4 As shown in the figure, this figure shows the process from input to output of a sample). The input of the input layer is x j =[x j1 ,...,x jn ,...,x jN ], representing the N feature data of the jth sample. The weight and threshold between the input layer and the hidden layer are (w,b)=[(w1,b1),...,(w L ,b L )], where L is the number of hidden nodes. The output of the hidden layer is h l =[h1,...,h l ,...,h L ],h​l represents the output of the lth hidden node. The weight and threshold between the input layer and the hidden layer are θ=[θ1,...,θ L ], where L is the number of hidden nodes. The output of the output layer is o j =[o j1 ,...,o jm ,...,o jM ], indicating the M output dimensions of the j-th sample.

[0108] Set parameters: fractional order α (value: 9 / 11), learning rate η (value: grid search method to find the optimal learning rate), number of hidden layer nodes L (value: grid search method to find the optimal number of nodes), minimum error E_min (value: 0.00001) and maximum number of iterations K (value: 300). Randomly generate and fix the weight w (value: range [0,1]) and threshold b (value: range [0,1]) between the input layer and the hidden layer. Initialize the weight θ (value: range [0,1]) between the hidden layer and the output layer.

[0109] Step 3: Use the data preprocessed in step 1 as training data, calculate the actual output of the model through the characteristics of the training data, and construct the error function E by combining the actual output with the ideal output:

[0110] Calculate the actual output:

[0111]

[0112] Among them, the hidden node output is:

[0113]

[0114] Where: is the number of hidden nodes, and g(x) is the sigmoid activation function.

[0115] Calculate the error function E:

[0116]

[0117] Step 4: Train the network model through K iterations. The ultimate goal of training the neural network model is to minimize the error function E:

[0118] Weight update formula:

[0119]

[0120]

[0121] Where D is the abbreviation symbol of the fractional derivative of the error function, c is the abbreviation symbol of the Caputo fractional derivative, k represents the current iteration step, and sgn(θ) is the symbolic function Γ(x) is the gamma function.

[0122] Step 5: The configuration after minimizing the error function E is the configuration of the trained neural network model, and the trained neural network model is used for the signal data classification task with the same characteristics as the training signal data:

[0123] Repeat steps 3 and 4 for K iterations to minimize the error function E, and obtain the optimal configuration of the neural network model for the subsequent classification of signal data. Regarding the classification process of subsequent signal data, the input samples are obtained after the data preprocessing shown in step 1, and then the input samples are input into the trained neural network model to obtain the output results, and the final classification decision is made according to the category obtained by the result.

[0124] The following takes two classification tasks of signal data sets in the real world as examples to further explain the practical application of the present invention.

[0125] Classification Task 1: A classification task to determine the user’s indoor location by observing signal data on a smartphone.

[0126] This dataset comes from the UCI Machine Learning Repository. The total number of samples in the dataset is 2000, the number of features of each sample is 7, and the decision variable is the user's indoor location and the number is 4.

[0127] 60% of the total samples after random shuffling are selected as the training set, and the remaining 40% are used as the test set. The center of the neural network model input is the sample feature value, and the input radius is randomly selected from the interval [0,0.1]. The center of the output is the decision variable, and the output radius is randomly selected from the interval [0.05,0.1]. The initial weight and threshold are randomly selected in the intervals [-1,1] and [0,1] respectively and fixed throughout the training process. The initial weight of the output layer is randomly selected in [-1,1]. The remaining neural network model parameter settings are shown in the result table.

[0128] Table 3 Classification results of fractional-order interval extreme learning machine and interval extreme learning machine on task 1 dataset

[0129]

[0130] Analysis of the classification results in Table 3 shows that the fractional-order interval extreme learning machine has a higher classification accuracy than the existing interval extreme learning machine. It can be seen that the fractional-order interval extreme learning machine of the present invention fits complex data structures more flexibly than the existing interval extreme learning machine and is more suitable for processing nonlinear and non-stationary data features.

[0131] like Figure 5 It can be seen that the fractional-order interval extreme learning machine of the present invention not only has a higher classification accuracy than the existing interval extreme learning machine, but also tends to be stable with fewer iteration steps, that is, a faster convergence speed.

[0132] Classification Task 2: Determine the spam classification task by observing the signal data.

[0133] This dataset comes from the UCI Machine Learning Repository. The total number of samples in the dataset is 4601, the number of features of each sample is 57, and the decision variable is the user's indoor location and the number is 2.

[0134] 60% of the total samples after random shuffling are selected as the training set, and the remaining 40% are used as the test set. The center of the neural network model input is the sample feature value, and the input radius is randomly selected from the interval [0,0.1]. The center of the output is the decision variable, and the output radius is randomly selected from the interval [0.05,0.1]. The initial weight and threshold are randomly selected in the intervals [-1,1] and [0,1] respectively and fixed throughout the training process. The initial weight of the output layer is randomly selected in [-1,1]. The remaining neural network model parameter settings are shown in the result table.

[0135] Table 4 Classification results of fractional-order interval extreme learning machine and interval extreme learning machine on task 2 dataset

[0136]

[0137] Analyzing the classification results in Table 4, it can be seen that the classification accuracy of the fractional-order interval extreme learning machine in the classification data set is always higher than that of the existing interval extreme learning machine. It can be seen that the fractional-order interval extreme learning machine method of the present invention has more accurate classification decisions than the prior art, and improves the classification performance of the intelligent communication system.

[0138] like Figure 6 It can be seen that the fractional-order interval extreme learning machine of the present invention is not only more reliable than the classification decision of the existing interval extreme learning machine. The ultimate purpose of the intelligent communication system is to identify the most accurate signal results in the shortest time. The fractional-order interval extreme learning machine of the present invention not only has a faster convergence speed than the prior art, but also has a higher accuracy, which undoubtedly meets the requirements of the intelligent communication system.

[0139] Through in-depth analysis of the experimental results of the two data sets and comparison with the classification performance of the prior art, the results show that the neural network model based on the fractional-order interval extreme learning machine proposed in the present invention shows higher effectiveness and stability in the signal classification task in the intelligent communication system. Specifically, this method can better cope with the complex channel environment and uncertain information in the intelligent communication system. By flexibly adjusting the fractional-order parameters of the model and using the interval neural network to process the interval value, the model's ability to extract signal features and classification accuracy are improved. Compared with the existing methods, this model not only has a significant improvement in classification accuracy, but also performs well in terms of the convergence speed of the model, fully demonstrating its potential for wide application in intelligent communication systems.

[0140] In order to further improve the performance of interval neural networks and improve their applicability in intelligent communication systems, the present invention proposes a neural network model signal classification method and system based on a fractional-order interval extreme learning machine, which not only solves the problem that extreme learning machines cannot be applied to interval neural networks due to nonlinear constraints, but also ensures that the optimized model has a faster learning speed and significant performance. In addition, the model can effectively process interval uncertainty data, and the classification performance and convergence speed of the model are further improved by the introduction of fractional-order derivatives, thereby showing excellent performance in complex classification tasks.

[0141] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A signal classification method based on a neural network model of a fractional-order interval extreme learning machine, characterized in that: The following steps are involved: Step S1: The uncertain signal is converted into uncertain data by combining orthogonal frequency division multiplexing with multi-level orthogonal amplitude modulation; The uncertain data consists of multiple samples s j Composition, each sample s j Includes N feature values ​​and 1 label class; Step S2: converting the uncertain data into interval type data, and performing normalization preprocessing on the interval type data. The specific steps include: Step S2-1: The uncertain data is converted into the interval type data (x j ,u j ); x j =[x j1 ,...,x jn ] T ,x ji =<x C ,x R >,u j =[u j1 ,...,u jm ] T ,u jt =<u C ,u R >,j=1,...,J,i=1,...,n, t=1,...,m In the formula, x represents the input of interval data, u represents the ideal output of interval data, J represents the total number of interval data samples, n represents the number of input features of interval data, m represents the dimension of interval output, C superscript represents the center of interval data, and R superscript represents the radius of interval data; Step S2-2: The interval type data is normalized and preprocessed, and the maximum and minimum normalization formula is: Wherein, max is the interval type data x j The maximum value of the characteristic value, min is the interval type data x j The minimum value of the eigenvalue of x is x j One of the eigenvalues, x * To normalize the preprocessed eigenvalues, the eigenvalues ​​of each sample are preprocessed to the range [0, 1] by using the maximum and minimum normalization formula, and the label data remains unchanged; Step S3: Establishing a three-layer feedforward neural network model based on fractional-order interval extreme learning machine; Step S3-1: using the interval type data preprocessed according to the method of step S2 as input data to set corresponding parameters of fractional order α, learning rate η and number of hidden nodes L; Step S3-1-1: Establish a three-layer neural network including an input layer, a hidden layer, and an output layer, where the input of the input layer is: x j =[x j1 ,...,x jn ,...,x jN ] In the formula, x jN Represents the N feature data of the jth sample, and the weight and threshold between the input layer and the hidden layer are: (w,b)=[(w1,b1),...,(w L ,b L )] Where L is the number of hidden nodes, w represents the weight between the input layer and the hidden layer, b represents the threshold between the input layer and the hidden layer, and the output of the hidden layer is: h l =[h1,...,h l ,...,h L ] In the formula, h l represents the output of the lth hidden node, and the weight between the hidden layer and the output layer is θ l =[θ1,...,θ L ], the output of the output layer is: the j =[o j1 ,...,o jm ,...,o jM ] In the formula, o jM represents the M output dimensions of the j-th sample; Step S3-1-2: Set parameters, the fractional order α is 9 / 11, the learning rate η is the optimal learning rate found by the grid search method, the number of hidden layer nodes L is the optimal number of nodes found by the grid search method, the minimum error E_min is 0.00001 and the maximum number of iterations K is 300, randomly generate and fix the weight w between the input layer and the hidden layer: range [0,1] and the threshold b: range [0,1], initialize the weight θ between the hidden layer and the output layer: range [0,1]; Step S4: Use the interval data preprocessed in step S2 as training data, calculate the actual output of the model through the characteristics of the training data, and construct an error function E by combining the actual output with the ideal output: Step S4-1: Calculate the actual output: In the formula, θ l =[θ1,...,θ L ] is the weight between the hidden layer and the output layer, and the hidden node output is: Where: l is the number of hidden nodes, g(x) is the sigmoid activation function, and the error function E is calculated: Step S5: Minimize the error function E by training the network model through K iterations; Step S5-1: Weight update formula: i k+1 =θ k +Δθ k , Where D is the abbreviated symbol of the fractional derivative of the error function, caputo is the Caputo fractional derivative, k represents the current iteration number, and sgn(θ) is the symbolic function: Γ(x) is denoted as the gamma function; Step S5-2: Repeat step S5 for K times to minimize the error function E and obtain the optimal configuration of the neural network model.

2. The signal classification method based on the neural network model of fractional-order interval extreme learning machine according to claim 1 is characterized in that: The feature value is the characteristic of the sample data type, and the label class is the category of the sample label.

3. A signal classification system based on a neural network model of a fractional-order interval extreme learning machine, characterized in that: Includes the following modules: data conversion module, normalization module, model creation module, training module; Step M1: the data conversion module converts the uncertain signal into the uncertain data through the orthogonal frequency division multiplexing combined with multi-level orthogonal amplitude modulation; The uncertain data is composed of a plurality of samples s j Composition, each of the samples j including N said feature values ​​and 1 said label class; Step M2: the data conversion module converts the uncertain data into the interval type data, and performs normalization preprocessing on the interval type data through the normalization module. The specific steps include: Step M2-1: converting the uncertain data into the interval type data (xj, uj); x j =[x j1 ,...,x jn ] T， x ji =<x C ,x R >,u j =[u j1 ,...,u jm ] T， u jt =<u C ,u R >,j=1,...,J,i=1,...,n, t=1,...,m In the formula, x represents the input of interval data, u represents the ideal output of interval data, J represents the total number of interval data samples, n represents the number of input features of interval data, m represents the dimension of interval output, C superscript represents the center of interval data, and R superscript represents the radius of interval data; Step M2-2: Normalization preprocessing, maximum and minimum normalization formula: Wherein, max is the interval type data x j The maximum value of the characteristic value, min is the interval type data x j The minimum value of the eigenvalue of x is x j One of the eigenvalues, x * To normalize the preprocessed eigenvalues, the eigenvalues ​​of each sample are preprocessed to a range of [0, 1] using a maximum and minimum normalization formula, and the label data remains unchanged; Step M3: the model creation module establishes the three-layer feedforward neural network model based on the fractional-order interval extreme learning machine; Step M3-1: Using the interval data preprocessed according to the method of step M2 as the input data, the corresponding parameters of the fractional order α, the learning rate η and the number of hidden nodes L are set: Step M3-1-1: Establish a three-layer neural network including the input layer, the hidden layer, and the output layer, wherein the input of the input layer is: x j =[x j1 ,...,x jn ,...,x jN ] In the formula, x jN The N feature data representing the jth sample, the weight and threshold between the input layer and the hidden layer are: (w,b)=[(w1,b1),...,(w L ,b L )] Wherein, L is the number of hidden nodes, w represents the weight between the input layer and the hidden layer, b represents the threshold between the input layer and the hidden layer, and the output of the hidden layer is: h l =[h1,...,h l ,...,h L ] In the formula, h l represents the output of the lth hidden node, and the weight between the hidden layer and the output layer is θ l =[θ1,...,θ L ], where L is the number of hidden nodes, and the output of the output layer is: the j =[o j1 ,...,o jm ,...,o jM ] In the formula, o jM represents the M output dimensions of the j-th sample; Step M3-1-2: Set the parameters, the fractional order α is 9 / 11, the learning rate η is the optimal learning rate found by the grid search method, the number of hidden layer nodes L is the optimal number of nodes found by the grid search method, the minimum error E_min is 0.00001 and the maximum number of iterations K is 300, randomly generate and fix the weight w between the input layer and the hidden layer: range [0,1] and the threshold b: range [0,1], initialize the weight θ between the hidden layer and the output layer: range [0,1]; Step M4: Using the interval data preprocessed in step M2 as the training data, the training module calculates the actual output of the model through the features of the training data, and constructs an error function E by combining the actual output with the ideal output: Step M4-1: Calculate the actual output: Among them, θ l =[θ1,...,θ L ] is the weight between the hidden layer and the output layer, and the hidden node output is: Where: l is the number of hidden nodes, g(x) is the sigmoid activation function, and the error function E is calculated: Step M5: Minimize the error function E by training the network model through K iterations; Step M5-1: Weight update formula: i k+1 =θ k +Δθ k , Where D is the abbreviated symbol of the fractional derivative of the error function, caputo is the Caputo fractional derivative, k represents the current iteration step, and sgn(θ) is the sign function: Γ(x) is denoted as the gamma function; Step M5-2: Repeat step S5 for K times to minimize the error function E, that is, obtain the optimal configuration of the neural network model, and complete the classification task of uncertain signals based on the neural network model with the optimal configuration.

Images