MAC protocol intelligent identification method based on deep ensemble learning

Abstract

The invention particularly relates to an MAC protocol intelligent identification method based on deep ensemble learning, which comprises the following steps: preprocessing collected signals under multiple types of MAC protocols, extracting time domain features, time-frequency domain features and channel features, and dividing feature data into a training set and a test set; inputting the training set data into a deep integrated learning network for training, processing time domain features by using a Transform-BiLSTM network, processing time-frequency domain features by using a ResNet network, and processing channel features by using an LSTM-RBF network; splicing output results of the three networks to form a new feature, inputting the new feature into a meta-learner for decision fusion, outputting a final recognition result, and storing a trained deep integrated learning network model; and inputting the test set data into the trained deep integrated learning network to obtain an identification result. The method can effectively realize intelligent identification of various different MAC protocols, shows excellent identification performance under a low signal-to-noise ratio, and can be widely applied to the technical field of cognitive radio.

Description

Technical Field

[0001] The present invention relates to the field of MAC protocol identification in cognitive radio technology, and in particular to a MAC protocol intelligent identification method based on deep ensemble learning. Background Art

[0002] In cluster communications, common MAC protocols include TDMA, CSMA / CA, Pure ALOHA, and SlottedALOHA. Identifying the MAC protocol in cluster targets can lay the foundation for spectrum management, optimized scheduling, security protection, and covert communications. Furthermore, in a communications adversarial environment, identifying the MAC protocol used by enemy cluster targets allows for the selection of specific jamming methods to improve the jamming effect. However, most existing methods rely on partial data from the signal's physical layer as features for analysis and identification, and only use a single classifier to process the input features. This results in insufficient feature extraction and is prone to overfitting, making it difficult to adapt to complex communication environments. Therefore, it is particularly important to design intelligent and reliable MAC protocol identification methods.

[0003] Currently, many MAC protocol identification methods have been proposed. Classic MAC protocol identification methods include the power feature-based algorithm (Yang Z, Yao YD, Chen S, et al. MAC protocol classification in acognitive radio network [C]. The 19th Annual Wireless and Optical Communications Conference (WOCC 2010). IEEE, 2010: 1-5) and the power mean and variance-based method (Hu S, Yao YD, Yang Z. MAC protocol identification using support vector machines for cognitive radio networks [J]. IEEE Wireless Communications, 2014, 21(1): 52-60).With the development of deep learning, new MAC protocol identification methods include MAC protocol identification method based on time-frequency graph and CNN network (hang X, Shen W, Xu J, et al. A mac protocol identification approachbased on convolutional neural network[C]. 2020 International Conference on Wireless Communications and Signal Processing(WCSP). IEEE, 2020: 534-539), MAC protocol identification method based on short-time energy feature and random forest (Ma G, Shen X, Wang H, et al. Clustering Quantization Short-Time Energy Feature Extraction Method for MAC Protocol Identification in Non-cooperative UWANs[C]. 2022 IEEE International Conference on Signal Processing, Communications and Computing(ICSPCC). IEEE, 2022: 1-6), and MAC protocol identification method based on squeeze and excitation mechanism improvement (Zhang X, Ding G, Li J, et al. Deep learning empowered MAC protocol identification with squeeze-and-excitation networks[J]. IEEE Transactions on Cognitive Communications and Networking,2021,8(2):683-693).

[0004] Most of the aforementioned MAC protocol identification methods rely on partial data from the signal's physical layer as features for analysis and identification, failing to fully extract the signal's characteristic information, resulting in low MAC protocol recognition accuracy. Furthermore, while deep learning-based methods offer improved performance compared to traditional SVM methods, they typically utilize only a single classifier, which inadequately extracts input features and is prone to overfitting, making it difficult to adapt to complex communication environments. Therefore, it is possible to design an intelligent MAC protocol recognition method based on deep ensemble learning by comprehensively analyzing multiple signal features to improve MAC protocol recognition performance.

[0005] Through the above analysis, the problems and defects of the existing technology are as follows:

[0006] (1) Most existing MAC protocol identification methods rely on partial data of the signal physical layer as features for analysis and identification, and do not fully extract the characteristic information of the signal.

[0007] (2) Existing MAC protocol identification methods usually only use a single classifier. A single classifier does not extract input features sufficiently and is prone to overfitting, making it difficult to adapt to complex communication environments.

[0008] It should be noted that the information disclosed in the above background technology section is only used to enhance the understanding of the background of the present invention, and therefore may include information that does not constitute prior art known to ordinary technicians in this field. Summary of the Invention

[0009] The present invention provides a MAC protocol intelligent recognition method based on deep ensemble learning, which can effectively realize the intelligent recognition of four different MAC protocols and show excellent recognition performance under low signal-to-noise ratio. It can be widely used in the field of cognitive radio technology.

[0010] Other features and advantages of the present invention will become apparent from the following detailed description, or may be learned in part by practice of the present invention.

[0011] According to a first aspect of the present invention, a method for intelligently identifying a MAC protocol based on deep ensemble learning is provided, the method comprising:

[0012] The signals collected under various MAC protocols are preprocessed using an improved wavelet threshold denoising method.

[0013] Extract time domain features, time-frequency domain features, and channel features from the preprocessed MAC protocol signal, and divide the feature data into a training set and a test set;

[0014] Input the training set data into a deep ensemble learning network for training. The deep ensemble learning network includes a Transformer-BiLSTM network, a ResNet network, an LSTM-RBF network, and a meta-learner. The Transformer-BiLSTM network is used to process time domain features, the ResNet network is used to process time-frequency domain features, and the LSTM-RBF network is used to process channel features. The output results of the Transformer-BiLSTM network, the ResNet network, and the LSTM-RBF network are concatenated to form new features, which are input into the meta-learner for decision fusion, and the trained deep ensemble learning network model is saved.

[0015] The test set data is input into the trained deep ensemble learning network to obtain the recognition result.

[0016] In some exemplary embodiments, the signals under the multiple types of MAC protocols include signals under four different MAC protocols, specifically: time division multiple access signals, carrier sense multiple access / collision avoidance protocol signals, pure random access wireless data transmission protocol signals, and slotted ALOHA signals.

[0017] In some exemplary embodiments, the preprocessing using an improved wavelet threshold denoising method specifically includes:

[0018] The wavelet coefficients of each layer are processed according to the preset threshold and threshold function to obtain the denoised coefficients. The preset threshold is a correction factor based on the kurtosis to improve the general threshold. The kurtosis of the coefficient of the j-th layer is:

[0019]

[0020] Among them, D j,k is the wavelet coefficient, μ is D j,k The mean of , E represents the expectation; the corrected threshold is:

[0021]

[0022] Among them, σ j is the noise standard deviation of the jth layer, N is the signal length, e is the natural logarithm, and Median represents the median;

[0023] Improved wavelet threshold function based on preset threshold:

[0024]

[0025] Where λ is the threshold, is the processed wavelet coefficient, and sgn represents the sign function.

[0026] In some exemplary embodiments, extracting time domain features, time-frequency domain features, and channel features from the preprocessed MAC protocol signal specifically includes:

[0027] The sliding window method is used to divide the MAC protocol signal into multiple short sequences. The instantaneous power of each short sequence is calculated and normalized to obtain the instantaneous power feature, i.e., the time domain feature.

[0028] After performing wavelet packet decomposition on the MAC protocol signal to be identified, the energy of each frequency band is calculated and normalized to obtain the time-frequency domain features;

[0029] Calculate the channel busy duration and the channel idle duration, establish a channel busy duration array and a channel idle duration array according to the channel busy duration and the channel idle duration respectively, and calculate the minimum, maximum and median of the channel busy duration according to the channel busy duration array; calculate the minimum, maximum and median of the channel idle duration according to the channel idle duration array, and combine the six features of the minimum, maximum and median of the channel busy duration and the minimum, maximum and median of the channel idle duration to form a channel feature.

[0030] In some exemplary embodiments, processing the time domain features using a Transformer-BiLSTM network specifically includes:

[0031] After the time domain features are processed by the Transformer-BiLSTM network, the output result is the first probability that the discrete received signal belongs to the signal under the i-th type of MAC protocol.

[0032] In some exemplary embodiments, processing time-frequency domain features using a ResNet network specifically includes:

[0033] In the RseNet network, after the time-frequency domain features undergo convolution and pooling operations, the output result is the second probability that the discrete received signal belongs to the signal under the i-th type MAC protocol.

[0034] In some exemplary embodiments, processing channel features using an LSTM-RBF network specifically includes:

[0035] After the channel characteristics are processed by the LSTM-RBF network, the output result is the third probability that the discrete received signal belongs to the signal under the i-th type MAC protocol.

[0036] In some exemplary embodiments, the meta-learner adopts a fully connected neural network.

[0037] According to a second aspect of the present invention, a storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the method for intelligently identifying a MAC protocol based on deep integrated learning according to the first aspect is implemented.

[0038] According to a third aspect of the present invention, a computer program product is provided, on which a computer program is stored. When the computer program is executed by a processor, the MAC protocol intelligent identification method based on deep integrated learning described in the first aspect is implemented.

[0039] According to a fourth aspect of the present invention, there is provided an electronic device, comprising:

[0040] processor; and

[0041] a memory for storing executable instructions of the processor;

[0042] Among them, the processor is configured to implement the MAC protocol intelligent identification method based on deep integrated learning described in the first aspect above by executing the executable instructions.

[0043] The MAC protocol intelligent identification method based on deep ensemble learning provided by the embodiments of the present invention uses an improved wavelet threshold denoising method to pre-process different MAC protocol signals for noise reduction, and extracts time domain features, time-frequency domain features, and channel features from the pre-processed MAC protocol signals. Then, based on the concept of stacking ensemble learning, a Transformer-BiLSTM network is used to process time domain features, a ResNet network is used to process time-frequency domain features, and an LSTM-RBF network is used to process channel features. Finally, the output results of the three deep networks are concatenated to form new features, which are then input into a meta-learner, which performs decision fusion and outputs the final classification and recognition results. This is specifically reflected as follows:

[0044] 1. The present invention pre-processes the received signal by using an improved wavelet threshold noise reduction method, thereby reducing the impact of noise on the signal, helping the algorithm to better extract signal features, and enhancing the robustness and generalization ability of the algorithm.

[0045] 2. The present invention overcomes the limitations of using a single feature by extracting multi-dimensional features of the signal, helping the algorithm to better understand the essential characteristics of the data, thereby reducing the risk of overfitting and improving the recognition performance of the algorithm.

[0046] 3. This invention leverages the concept of stacking ensemble learning, using three different deep neural networks as base learners to process signal features separately. A meta-learner then performs decision fusion to output the final recognition result. This overcomes the problem of a single classifier's insufficient input feature extraction and prone to overfitting. This technology can provide technical support for cognitive radio networks in complex communication environments and expand their application scope.

[0047] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] The accompanying drawings are incorporated into and constitute a part of this specification, illustrate embodiments consistent with the present invention, and together with the description, serve to explain the principles of the present invention. Obviously, the drawings described below are only some embodiments of the present invention, and it is clear that those skilled in the art can derive other drawings based on these drawings without inventive effort.

[0049] Figure 1 This is a flow chart of a MAC protocol intelligent identification method based on deep ensemble learning according to an embodiment of the present invention;

[0050] Figure 2 This is a graph showing the MAC protocol recognition accuracy under different signal-to-noise ratio conditions of the present invention;

[0051] Figure 3 This is a performance comparison chart of the average recognition accuracy of the MAC protocol of the present invention and a single learner;

[0052] Figure 4 This is a diagram showing the effect of the improved wavelet threshold noise reduction signal preprocessing on recognition performance proposed in the method of the present invention;

[0053] Figure 5 This is a comparison chart of the average recognition accuracy of the MAC protocol based on the present invention and the SVM method, LSTM method, CNN method and Transformer method. DETAILED DESCRIPTION

[0054] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be embodied in many forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided so that this disclosure will be thorough and complete and will fully convey the concepts of the example embodiments to those skilled in the art. The described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0055] In addition, the accompanying drawings are merely schematic illustrations of the present invention and are not necessarily drawn to scale. Identical reference numerals in the figures denote identical or similar parts, and thus repetitive descriptions thereof will be omitted. Some of the blocks shown in the accompanying drawings are functional entities that do not necessarily correspond to physically or logically separate entities. These functional entities may be implemented in software, in one or more hardware modules or integrated circuits, or in different networks and / or processor devices and / or microcontroller devices.

[0056] In view of the shortcomings and deficiencies of the existing technology, this example embodiment provides a MAC protocol intelligent identification method based on deep integrated learning, which preprocesses the signals under four types of MAC protocols, namely Time Division Multiple Access (TDMA), Carrier Sense Multiple Access with Collision Avoidance (CSMA / CA), Pure Additive Links On-line Hawaii Area (Pure ALOHA), and Slotted ALOHA, using an improved wavelet threshold denoising method; extracts time domain features, time-frequency domain features, and channel features from the preprocessed MAC protocol signals; uses a residual network (Residual Network, ResNet) to process time-frequency domain features, and uses a Transformer network (Transformer with BiLSTM Enhancement, Transformer-BiLSTM) based on a bidirectional long short-term memory network (BiLSTM) to process time domain features, and uses a long short-term memory network (Long Short-Term Memory Network Improved Based on Radial Basis Function (RBF)) to improve the long short-term memory network. By RBF, LSTM-RBF) to process channel features; based on the idea of staking ensemble learning, the output results of the above three deep neural networks are input into the meta-learner fully connected neural network (Fully Connected Neural Network, FCNN) for decision fusion, and the meta-learner outputs the final recognition result. Figure 1 As shown, the following steps may be specifically included:

[0057] S101, preprocessing the collected signals under multiple types of MAC protocols using an improved wavelet threshold noise reduction method;

[0058] S102, extracting time domain features, time-frequency domain features, and channel features from the pre-processed MAC protocol signal, and dividing the feature data into a training set and a test set;

[0059] S103, inputting the training set data into the deep ensemble learning network for training, using the Transformer-BiLSTM network to process time domain features, using the ResNet network to process time-frequency domain features, and using the LSTM-RBF network to process channel features;

[0060] S104, concatenate the output results of the three deep neural networks in S103 to form a new feature, input it into the meta-learner for decision fusion, and save the trained network model;

[0061] S105: Input the test set data into the trained deep integration learning network to obtain the recognition result.

[0062] Below, each step in this exemplary implementation will be described in more detail with reference to the accompanying drawings and embodiments.

[0063] S101 , preprocessing the collected signals under multiple types of MAC protocols using an improved wavelet threshold noise reduction method.

[0064] Exemplarily, a total of C discrete received signals including M signal categories in the electromagnetic environment are obtained to obtain a discrete signal set X={x1, x2, ...x c ,...x C}, where x c ={x c (1),x c (2),...x c (n),...x c (N)} represents the cth discrete received signal subset, x c (n) represents x c The nth discrete point of X, n = 1, 2, ..., N, M = 4, C = 34000, that is, X includes signals under four different MAC protocols: TDMA signal, CSMA / CA signal, PureALOHA signal and SlottedALOHA signal.

[0065] Furthermore, an improved wavelet threshold denoising method is used to pre-process the received signal, thereby improving the recognition effect of the MAC protocol. Specifically, the appropriate wavelet basis function and decomposition layer number are selected, and for each discrete received signal x c Perform wavelet decomposition:

[0066] ∫x c (t)ω j,k (t)dt=∫s c (t)ω j,k (t)dt+∫n(t)ω j,k (t)dt

[0067] Among them, x c (t) represents the noisy signal, s c (t) represents the original signal, n(t) is Gaussian white noise, ω j,k (t) is the wavelet basis function. Convert the above formula into coefficient form:

[0068] D j,k =V j,k +U j,k

[0069] Among them, D j,k Represents x c (t) The coefficients of each layer after wavelet transformation, V j,k Indicates s c (t) wavelet transform coefficient, U j,k Represents the wavelet transform coefficients of the noise n(t).

[0070] The wavelet coefficients of each layer are processed according to the preset threshold and threshold function to obtain the denoised coefficients. Since the traditional fixed threshold uniformly processes the wavelet coefficients of each layer, it cannot be adjusted according to the different characteristics of the signal and lacks adaptability. A correction factor based on kurtosis (Ks) is proposed to improve the universal threshold. This allows the threshold selection to better adapt to the characteristics of the noise distribution and the distribution characteristics of different MAC protocol signals in each layer. The kurtosis of the coefficients of the jth layer is:

[0071]

[0072] Among them, D j,k is the wavelet coefficient, μ is D j,k The mean of , E represents the expectation. The corrected threshold is:

[0073]

[0074] Among them, σ j is the noise standard deviation of the jth layer, N is the signal length, e is the natural logarithm, and Median represents the median.

[0075] In order to ensure the continuity of the threshold function at the threshold point and reduce the constant deviation of the reconstructed signal, an improved wavelet threshold function is proposed:

[0076]

[0077] Among them, λ is the threshold, D j,k is the kth wavelet coefficient of the jth layer after decomposition, is the processed wavelet coefficient, and sgn represents the sign function.

[0078] S102 , extracting time domain features, time-frequency domain features, and channel features from the pre-processed MAC protocol signal, and dividing the feature data into a training set and a test set.

[0079] First, for the time domain features, a sliding window method is used to convert the longer MAC protocol signal Split into N shorter sequences of length M There is no overlap between adjacent shorter sequences, where N = L / M represents the number of short sequences after segmentation. Each segment of data is Calculate the instantaneous power for each data segment:

[0080] P′ Mi (n) = |x′ Mi (n)| 2

[0081] Among them, P′ Mi (n) is the instantaneous power of the i-th data segment at time n, x′ Mi (n) represents the i-th data segment.

[0082] Get x′ c The instantaneous power of all data segments According to the Min-Max normalization criterion, P′ is normalized to obtain the normalized instantaneous power characteristics. The normalization formula is:

[0083]

[0084] Among them, P′(n) represents the original eigenvalue, P′ max (n) and P′ min (n) represents the minimum and maximum eigenvalues respectively, and P(n) represents the normalized eigenvalue, whose value range is [0,1].

[0085] Calculate each discrete subset of signals to be identified x c The instantaneous power P of the received signal is used as a feature of the received signal.

[0086] Then, each discrete subset of signals to be identified x is calculated based on wavelet packet decomposition c The time-frequency domain characteristics of:

[0087]

[0088] Among them, d j,2m (n) and d j,2m+1 (n) is the low-frequency and high-frequency components of the jth layer, m represents the node index, h(k) and g(k) represent the low-pass filter and high-pass filter respectively. After the MAC protocol signal to be identified is decomposed by wavelet packet, the energy of each frequency band is calculated, d 3,j Energy E 3,j (j=0,1,...,7) is:

[0089]

[0090] Among them, a jk(j=0,1,...,7;k=1,2,...,n) is d 3,j The amplitude of the kth point, n is the length of the sequence.

[0091] Normalize the energy of each frequency band to obtain the eigenvector T:

[0092] T=[E 3,0 / E,E 3,1 / E,E 3,2 / E,E 3,3 / E,E 3,4 / E,E 3,5 / E,E 3,6 / E,E 3,7 / E]

[0093]

[0094] Among them, E 3,j Represents the energy of the j-th wavelet packet coefficient in the third layer.

[0095] Calculate each discrete subset of signals to be identified x c The time-frequency domain T is used as another feature of the received signal.

[0096] Then, calculate the channel busy duration A busy,i,j and idle duration A idle,i,j . A busy,i,j is the longest sequence length of consecutive sampling points whose signal power intensity is greater than the threshold T in the jth frame:

[0097]

[0098] in, represents the indicator function, when |x i [n]| 2 The value is 1 if ≥T is true, and 0 otherwise.

[0099] A idle,i,j is the longest sequence length of consecutive sampling points whose signal power intensity is less than the threshold T in the jth frame:

[0100]

[0101] in, represents the indicator function, if |x i [n]| 2 If <T is true, the value is 1, otherwise it is 0.

[0102] According to A busy,i,j and A idle,i,j Establish channel busy duration array D respectively busyand channel idle duration array D idle , and according to D busy Calculate the minimum channel busy duration d busy,min , maximum value d busy,max , median d busy,median and mean d busy,mean According to D idle Calculate the minimum channel idle duration d idle,min , maximum value d idle,max , median d idle,median and mean d idle,mean The eight-element feature group consisting of these eight features is obtained to characterize the channel feature C:

[0103] C=[d busy,min ,d busy,max ,d busy,median ,d busy,mean ,d idle,min ,d idle,max ,d idle,median ,d idle,mean ]

[0104] Calculate each discrete subset of signals to be identified x c The channel characteristic C is used as another feature of the received signal.

[0105] Finally, the three feature data sets obtained above are divided into training sets and test sets.

[0106] S103: Input the training data into the deep integrated learning network for training, use the Transformer-BiLSTM network to process time domain features, use the ResNet network to process time-frequency domain features, and use the LSTM-RBF network to process channel features.

[0107] First, all learnable parameters in the deep ensemble learning network model were randomly initialized. The Adaptive Moment Estimation (Adam) optimizer was used for the ResNet and LSTM-RBF network models, and the AdamW optimizer was used for the Transformer-BiLSTM. The maximum number of initial iterations was set to 100, the learning rate was 0.001, and the cross-entropy (CE) loss function was used. The training data was randomly shuffled and fed into the network model in batches for training. The training error between the predicted and true values of each batch was calculated and back-propagated to optimize all learnable parameters. The CE loss function is calculated as follows:

[0108]

[0109] Among them, M represents the number of categories, y i Indicates the mark of the true category in category i, represents the probability of the i-th category predicted by the model.

[0110] Then, the time domain features, time-frequency domain features and channel features of the discrete received signal obtained in S102 are input into the Transformer-BiLSTM network, the RseNet network and the LSTM-RBF network for processing respectively. In the Transformer-BiLSTM network, BiLSTM is integrated into the Transformer model to improve the traditional Transformer architecture. In the present invention, since it is a classification task, the Transformer encoder is used to extract features. Considering the advantages of BiLSTM in capturing time series features, the present invention adds a BiLSTM layer to the Transformer encoder, thereby constructing a new Transformer-BiLSTM fusion model. After the time domain features are processed by the Transformer-BiLSTM network, the output result is P1=[P 11 ,P 12 ,P 13 ,...,P 1M ], where M=4 represents the number of signal categories, P 1i It represents the probability that the discrete received signal output by the Transformer-BiLSTM network belongs to the signal under the i-th type MAC protocol. In the RseNet network, after the time-frequency domain features undergo convolution, pooling and other operations, the output result is P2 = [P 21 ,P 22 ,P 23 ,...,P 2M ], where M=4 represents the number of signal categories, P 2i Indicates the probability that the discrete received signal output by the ResNet network belongs to the signal under the i-th type of MAC protocol. In the LSTM-RBF network, based on the advantages of LSTM in processing sequence data and the ability of RBF to perform nonlinear mapping of data through Gaussian kernel function, the LSTM network is combined with the RBF network to achieve an improvement on the LSTM network. After the channel features are processed by the LSTM-RBF network, the output result is P3 = [P 31 ,P 32 ,P 33 ,...,P 3M ], where M=4 represents the number of signal categories, P 3i It represents the probability that the discrete received signal output by the LSTM-RBF network belongs to the signal under the i-th type of MAC protocol.

[0111] S104, splicing the output results of the three deep neural networks in S103 to form a new feature, inputting it into the meta-learner for decision fusion, and saving the trained deep ensemble learning network model.

[0112] First, set the optimizer of the meta-learner to Adam, initialize the maximum number of iterations to 100, and the learning rate to 0.001;

[0113] Then, the three deep neural network output results P1, P2 and P3 obtained in S103 are concatenated to obtain a new feature P = [P 11 ,P 12 ,...,P 1M ,P 21 ,P 22 ,...,P 2M ,P 31 ,P 32 ,...,P 3M ];

[0114] Finally, according to the idea of stacking ensemble learning, the new feature P is input into the meta-learner for training, and the meta-learner performs decision fusion and outputs the final recognition result. In this invention, considering the nonlinear modeling ability and structural simplicity of the meta-learner, a fully connected neural network (FCNN) is used as the meta-learner. When the data of the training set undergoes forward output and back propagation, an iteration is completed. When the number of iterations reaches the maximum number of iterations, the training is completed and the deep ensemble learning network model is saved.

[0115] S105: Input the test set data into the trained deep integration learning network to obtain the recognition result.

[0116] First, the data of the test set obtained in S102 is input into the trained deep integrated learning network model. The time domain features in the test set pass through the Transformer-BiLSTM network module in the deep integrated learning network, the time-frequency domain features pass through the ResNet network module in the deep integrated learning network, and the channel features pass through the LSTM-RBF network module in the deep integrated learning network.

[0117] Finally, the meta-learner module FCNN in the deep integrated learning network outputs the final classification and recognition results.

[0118] The technical effects of the present invention are described in detail below in conjunction with simulation experiments.

[0119] To evaluate the performance of the present invention, simulations were conducted. Matlab software was used to simulate signals under four different MAC protocols, including TDMA, CSMA / CA, Pure Aloha, and Slotted Aloha. All four MAC protocols employed fixed-frequency communication and BPSK modulation. The noise was Gaussian white noise with a mean of 0 and a variance of 1, and the signal-to-noise ratio (SNR) range was set to -6dB to 10dB. The signal bit rate was 1000 bit / s, the sampling frequency was 100 kHz, and the transmission probability was 0.1-1.0. For each MAC protocol signal, 500 sets of data were generated at each SNR, for a total of 34,000 sets of data. Time domain features, time-frequency domain features, and channel features were extracted for each of these signals. Labels were assigned to each of the three features, with the labels [0, 1, 2, 3] representing TDMA, CSMA / CA, Pure Aloha, and Slotted Aloha, respectively. Training and test datasets were then constructed. The network model uses cross-entropy as the loss function and dynamically adjusts parameters through the backpropagation algorithm. After multiple cycles, the optimal parameters are obtained. During training, the Adam optimizer is used for ResNet and LSTM-RBF, and the AdamW optimizer is used for Transformer-BiLSTM. The trained network model is finally obtained.

[0120] The recognition accuracy of the method proposed in this invention for different MAC protocols under different signal-to-noise ratio conditions is as follows: Figure 2 As shown in the figure, it can be seen that the present invention can effectively identify signals under different types of MAC protocols, and as the signal-to-noise ratio increases, the recognition accuracy of each type of MAC protocol signal first increases and then tends to stabilize. Figure 3 As shown in the figure, it can be seen that the overall performance of the deep ensemble learning model is significantly better than that of a single learner, proving the effectiveness of the deep ensemble learning method. The effect of the improved wavelet threshold denoising signal preprocessing proposed in the method of the present invention on the recognition performance is shown in the figure. Figure 4 As shown in the figure, it can be seen that the recognition performance of the model is significantly improved after the noise reduction preprocessing. The average recognition accuracy of the MAC protocol based on the method proposed in this invention is compared with that based on the SVM method, LSTM method, CNN method and Transformer method. Figure 5 As shown in the figure, it can be seen that under different transmission probability conditions, the average recognition accuracy of the MAC protocol of the present invention is better than that of the SVM method, LSTM method, CNN method and Transformer method, and as the transmission probability continues to increase, the average recognition accuracy of the five methods gradually increases and then tends to be stable.

[0121] Other embodiments of the present invention will readily occur to those skilled in the art after considering the specification and practicing the invention herein. This application is intended to cover any variations, uses, or adaptations of the present invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein. The description and examples are to be considered as exemplary only, with the true scope and spirit of the invention being indicated by the claims.

[0122] It should be understood that the present invention is not limited to the exact construction described above and shown in the drawings and that various modifications and variations can be made without departing from the scope thereof, which is limited only by the appended claims.

Claims

1. A MAC protocol intelligent identification method based on deep ensemble learning, characterized in that: The method comprises: The signals collected under various MAC protocols are preprocessed using an improved wavelet threshold denoising method. Extract time domain features, time-frequency domain features, and channel features from the preprocessed MAC protocol signal, and divide the feature data into a training set and a test set; Input the training set data into a deep ensemble learning network for training. The deep ensemble learning network includes a Transformer-BiLSTM network, a ResNet network, an LSTM-RBF network, and a meta-learner. The Transformer-BiLSTM network is used to process time domain features, the ResNet network is used to process time-frequency domain features, and the LSTM-RBF network is used to process channel features. The output results of the Transformer-BiLSTM network, the ResNet network, and the LSTM-RBF network are concatenated to form new features, which are input into the meta-learner for decision fusion, and the trained deep ensemble learning network model is saved. The test set data is input into the trained deep ensemble learning network to obtain the recognition result.

2. The method according to claim 1, characterized in that The multiple types of MAC protocol signals include four types of signals under different MAC protocols, specifically: time division multiple access signals, carrier sense multiple access / collision avoidance protocol signals, pure random access wireless data transmission protocol signals, and slotted ALOHA signals.

3. The method according to claim 1, characterized in that The preprocessing using the improved wavelet threshold noise reduction method specifically includes: The wavelet coefficients of each layer are processed according to the preset threshold and threshold function to obtain the denoised coefficients. The preset threshold is a correction factor based on the kurtosis to improve the general threshold. The kurtosis of the coefficient of the j-th layer is: Among them, D j,k is the wavelet coefficient, μ is D j,k The mean of , E represents the expectation; the corrected threshold is: Among them, σ j is the noise standard deviation of the jth layer, N is the signal length, e is the natural logarithm, and Median represents the median; Improved wavelet threshold function based on preset threshold: Where λ is the threshold, is the processed wavelet coefficient, and sgn represents the sign function.

4. The method according to claim 1, wherein The extracting of time domain features, time-frequency domain features, and channel features from the preprocessed MAC protocol signal specifically includes: The sliding window method is used to divide the MAC protocol signal into multiple short sequences. The instantaneous power of each short sequence is calculated and normalized to obtain the instantaneous power feature, i.e., the time domain feature. After performing wavelet packet decomposition on the MAC protocol signal to be identified, the energy of each frequency band is calculated and normalized to obtain the time-frequency domain features; Calculate the channel busy duration and the channel idle duration, establish a channel busy duration array and a channel idle duration array according to the channel busy duration and the channel idle duration respectively, and calculate the minimum, maximum, median and mean of the channel busy duration according to the channel busy duration array; calculate the minimum, maximum, median and mean of the channel idle duration according to the channel idle duration array, and combine the eight features of the minimum, maximum, median and mean of the channel busy duration and the minimum, maximum, median and mean of the channel idle duration to form a channel feature.

5. The method according to claim 1, wherein The use of the Transformer-BiLSTM network to process time domain features specifically includes: After the time domain features are processed by the Transformer-BiLSTM network, the output result is the first probability that the discrete received signal belongs to the signal under the i-th type of MAC protocol.

6. The method according to claim 1, characterized in that The use of the ResNet network to process time-frequency domain features specifically includes: In the RseNet network, after the time-frequency domain features undergo convolution and pooling operations, the output result is the second probability that the discrete received signal belongs to the signal under the i-th type MAC protocol.

7. The method according to claim 1, characterized in that The channel feature processing using the LSTM-RBF network specifically includes: After the channel characteristics are processed by the LSTM-RBF network, the output result is the third probability that the discrete received signal belongs to the signal under the i-th type MAC protocol.

8. The method according to claim 1, characterized in that The meta-learner adopts a fully connected neural network.

9. A storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the MAC protocol intelligent identification method based on deep integrated learning according to any one of claims 1 to 8 is implemented.

10. An electronic device, characterized in that: include: processor; as well as a memory for storing executable instructions of the processor; The processor is configured to execute the MAC protocol intelligent identification method based on deep integrated learning according to any one of claims 1 to 8 by executing the executable instructions.

Images