Wireless network energy prediction and clustering routing method based on attention mechanism and improved SCINet

Abstract

The invention relates to the technical field of wireless communication, and provides a wireless network energy prediction and clustering routing method based on an attention mechanism and improved SCINet, which comprises the following steps of: performing data normalization and feature extraction on energy data of a wireless sensor network; an ATT-Fuse-SCINet model is adopted to carry out energy prediction on the processed energy data, an energy prediction result is obtained, and the ATT-Fuse-SCINet model comprises an SCI-Block structure, a probability sparse self-attention mechanism and a residual fusion structure; and on the basis of an energy prediction result, an EP-LEACH clustering routing algorithm is adopted to perform clustering head election, and residual energy and future prediction energy of the nodes are comprehensively considered in the election process. The energy prediction model provided by the invention has higher prediction accuracy and generalization ability, and the improved clustering routing algorithm based on the energy prediction model can effectively prolong the service life of the wireless sensor network.

Description

Technical Field

[0001] This invention relates to the field of wireless communication technology, and more specifically, to a wireless network energy prediction and clustering routing method based on an attention mechanism and an improved SCINet. Background Technology

[0002] In recent years, with the rapid development of Internet of Things (IoT) technology, the application of wireless sensor networks (WSNs) has become increasingly widespread, including environmental monitoring, habitat monitoring, smart agriculture, and intelligent transportation. In traditional WSNs, when the sensor battery runs out of power, the sensor enters a dead state, leading to a decline in network performance or even shutdown. Limited battery capacity restricts the lifespan of sensor nodes, significantly increasing the maintenance costs of WSNs.

[0003] To meet the growing application demands of future wireless sensor networks, researchers are constantly seeking suitable energy sources and convenient power supply solutions to power sensor nodes. Energy-harvesting wireless sensor networks (EH-WSNs), which can harvest energy from the surrounding environment to maintain normal network operation, have therefore attracted widespread attention. Currently, energy sources that EH-WSNs can harvest include solar energy, radio frequency (RF), thermal energy, wind energy, and mechanical energy.

[0004] Solar energy is widely used in EH-WSNs due to its high energy density, cleanliness, and ease of access. Extending WSN lifetime using solar energy involves technologies such as energy management, adaptive power adjustment, and intelligent routing. However, solar energy is significantly affected by weather conditions and exhibits strong instability. To adapt to variable solar energy and improve energy efficiency, researchers have conducted extensive research on energy prediction algorithms. Current energy prediction methods are based on statistical principles, such as EWMA, WCMA, and Pro-Energy, as well as machine learning principles, such as Knowledge-Based Neural Networks (KBNN), Recurrent Neural Networks (RNN), and Long Short-Term Memory Networks (LSTM). These energy prediction methods face challenges in terms of accuracy under complex weather conditions and long-term forecasting. Low prediction accuracy leads to poor energy utilization and premature node death in WSNs; therefore, a prediction scheme with higher accuracy and generalization ability is urgently needed to extend the lifetime of WSNs. Summary of the Invention

[0005] This invention proposes a wireless network energy prediction and clustering routing method based on an attention mechanism and an improved SCINet. By introducing a probabilistic sparse self-attention mechanism, attention weights are calculated at each time step of the new feature sequence to capture important features and improve the model's prediction accuracy. Finally, the clustering routing algorithm is improved based on the node's remaining energy and the predicted future solar energy.

[0006] To achieve the above objectives, this invention proposes a wireless network energy prediction and clustering routing method based on an attention mechanism and an improved SCINet, comprising: Data normalization and feature extraction are performed on energy data from wireless sensor networks; The ATT-Fuse-SCINet model is used to predict energy from the processed energy data and obtain the energy prediction results. The ATT-Fuse-SCINet model includes an SCI-Block structure, a probabilistic sparse self-attention mechanism, and a residual fusion structure. Based on the energy prediction results, the EP-LEACH clustering routing algorithm is used to elect the cluster head, taking into account both the remaining energy and the future predicted energy of the nodes during the election process.

[0007] Furthermore, the data normalization method is as follows:

[0008] in, It is a feature The normalized value; It is a feature The original value; and These are the maximum and minimum values ​​of the feature, respectively.

[0009] Furthermore, the feature extraction method is as follows:

[0010] in, It is the correlation coefficient; It is the sample size; , It consists of the target sequence values ​​and their average values. and These are other feature sequence values ​​and their average values.

[0011] Furthermore, the SCI-Block structure includes: The input sequence is downsampled into odd and even sequences; One-dimensional convolutional modules are used to transform odd and even sequences, and interactive learning is performed through Hadamard products. New subsequences are generated through convolution mapping and addition / subtraction operations to compensate for information loss during downsampling.

[0012] Furthermore, the probabilistic sparse self-attention mechanism calculates the sparsity of the Query using the following formula:

[0013] in, It is a measure of sparsity; yes The first in One element; yes The first in One element; yes The length.

[0014] Furthermore, the residual fusion structure adopts the following formula:

[0015] Among them, weight and The importance of the output and input in the fusion process is indicated by the Swish function and normalization.

[0016] Furthermore, the weight calculation and Swish function of the residual fusion structure are shown in the following equation: in, This represents the calculated weight parameters. This represents the original weight parameters, with an initial value of 1. It is to prevent extremely small numbers with a denominator of 0.

[0017] Furthermore, the method for cluster head election using the EP-LEACH clustering routing algorithm adopts the following formula:

[0018]

[0019] in, It is the current node. It is the percentage of cluster heads. It is the candidate cluster head set. This is the current round. It is the remaining energy of node i. It is the predicted energy of the node at the end of the current round. It is the initial energy of node i.

[0020] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention proposes a novel energy prediction model based on an attention mechanism and an improved SCINet model. The improved SCINet model replaces the residual connection structure with a residual fusion structure and introduces an attention mechanism to adaptively adjust the contributions of input and output in the final result, thereby improving the model's prediction accuracy and generalization ability.

[0021] This invention proposes a clustering routing algorithm based on energy prediction. In this algorithm, the remaining energy of a node and the predicted energy harvesting in the future are used as key factors for cluster head election, which significantly improves energy utilization and extends the lifetime of nodes and the network.

[0022] The effectiveness of the proposed energy prediction model and clustering routing algorithm was evaluated through simulation experiments, which proved that the method can effectively extend the life cycle of WSN. Attached Figure Description

[0023] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. In the drawings: Figure 1 This is a schematic diagram of the overall process of the method proposed in this invention; Figure 2 This is a structural diagram of the ATT-Fuse-SCINet model proposed in this invention; Figure 3 is a schematic diagram comparing the generalization ability of the ATT-Fuse-SCINet model of the present invention with other models under three different weather conditions, where (a) is an 8-day continuous forecast, (b) is a sunny forecast, (c) is a cloudy forecast, and (d) is a rainy forecast. Figure 4 This diagram illustrates the comparison of the number of surviving nodes between the LEACH algorithm of this invention and the EP-LEACH algorithm under three different prediction models. Figure 5 This is a schematic diagram comparing the total network energy of the LEACH algorithm of this invention and the EP-LEACH algorithm under three different prediction models; Figure 6 This is a schematic diagram of the death node statistics in an embodiment of the present invention. Detailed Implementation

[0024] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0025] This embodiment proposes a wireless network energy prediction and clustering routing method based on an attention mechanism and an improved SCINet, such as... Figure 1 As shown, it includes: Data normalization and feature extraction are performed on energy data from wireless sensor networks; The ATT-Fuse-SCINet model is used to predict energy from the processed energy data and obtain the energy prediction results. The ATT-Fuse-SCINet model includes an SCI-Block structure, a probabilistic sparse self-attention mechanism, and a residual fusion structure. Based on the energy prediction results, the EP-LEACH clustering routing algorithm is used to elect the cluster head, taking into account both the remaining energy and the future predicted energy of the nodes during the election process.

[0026] The solution of this embodiment will be described in detail below, based on the above steps: The implementation process of this embodiment includes three parts. The first part processes the raw data; the second part proposes the ATT-Fuse-SCINet (SCINet with attention and residual fusion structure) energy prediction model; and the third part proposes the EP-LEACH clustering routing algorithm (LEACH algorithm based on energy prediction) based on the high-precision energy prediction model.

[0027] Data processing Data normalization The original data is normalized according to formula (1) so that all features are on the same scale, thereby accelerating the convergence of the learning algorithm.

[0028] (1) in, It is a feature The normalized value; It is a feature The original value; and These are the maximum and minimum values ​​of the feature, respectively.

[0029] Feature extraction This embodiment uses the publicly available solar irradiance dataset from the Tennant Creek station of the Desert Knowledge Solar Centre (DKASC) in Australia. In addition to direct solar irradiance, the dataset includes environmental features such as horizontal irradiance, tilted irradiance, temperature, wind direction, wind speed, humidity, and rainfall. However, some of these features may be redundant or irrelevant to the prediction task. To reduce computational costs and improve training efficiency, this paper employs the Pearson correlation coefficient method in Equation (2) to analyze the correlation between the remaining features and direct solar irradiance.

[0030] (2) in, It is the correlation coefficient; It is the sample size; , It consists of the target sequence values ​​and their average values. and These are other feature sequence values ​​and their average values.

[0031] ATT-Fuse-SCINet Predictive Modeling This embodiment proposes the ATT-Fuse-SCINet prediction model. First, a probabilistic sparse attention mechanism is introduced, adaptively allocating attention weights to the two new subsequences generated by SCI-Block interactive learning, making the model focus more on important features. After multiple SCI-Block layers, the subsequences are reassembled into a new sequence, which is then added to the original sequence for prediction via residual fusion rather than residual connections. The ATT-Fuse-SCINet prediction model adaptively adjusts the contributions of input and output in the final result, improving the model's generalization ability. The structure diagram of the ATT-Fuse-SCINet model is shown below. Figure 2 As shown.

[0032] SCI-Block SCI-Block is the basic unit of SCINet. It downsamples the input data or features into odd sequences. Even sequence Then use a set of one-dimensional convolutional modules and Will and Convert to An exponential function with base 0, and with respect to... and Performing element-wise interactive Hadamard product operation yields... and , where ⊙ represents the Hadamard product.

[0033] ⊙ (3) ⊙ (4) Then, use a one-dimensional convolution module. and Two subsequences and Map to the hidden state, then with and Perform interactive addition and subtraction operations. This process ultimately produces two new sub-features. and This interactive learning strategy can compensate for information loss during the downsampling process.

[0034] (5) (6) SCINet is composed of multiple SCI-Blocks arranged in a binary tree structure, with the i-th layer (i=1,2,3,....,L) having 2 i -1 SCI-Block. After L layers of downsampling, convolution, and interactive learning operations, all elements in the resulting subsequence are reverse-parity split, rearranged, and concatenated into a new sequence. This new sequence is then added to the original sequence using residual connections, and finally decoded through a fully connected layer to obtain the final prediction result. The significant advantage of this design is that each SCI block has both a local view and a global view of the entire time series, which is beneficial for extracting useful temporal features and maintaining stability in long-term forecasts.

[0035] ProbSparse self-attention Attention mechanism dynamically adjusts output and The weights are assigned to highlight key information and enhance the network's attention to that information, thereby improving prediction accuracy. Traditional self-attention mechanisms essentially assign weights to samples, using an attention function to map the query vector and a set of key-value vectors to the output vector.

[0036] Probabilistic sparse self-attention mechanisms are an evolution of traditional self-attention mechanisms. Sparse self-attention mechanisms exhibit a long-tail effect, where the main attention is concentrated on a few dot products, while the contributions of other dot products are relatively small, or even negligible. Therefore, reducing the number of connections between elements can effectively reduce computational complexity. This embodiment proposes a method to measure the sparsity of a query and then selects the important queries. The sparsity metric for the i-th query is: (7) in, It is a measure of sparsity; yes The first in One element; yes The first in One element; yes The length of the input sequence X is calculated as follows: (8) in It is a self-attention score matrix; It is a query vector; It is a probabilistic sparse key vector; yes The corresponding value vector. , , It is a trainable projection matrix. It is the dimension of the input key vector.

[0037] Corresponding input element Output It is all The weighted sum, where the weights are given by the corresponding attention scores: (9) in It is an attention score matrix The Middle Line number Column weights. It is the first sparse value vector of probability. Each element. It is the number of computational elements selected under the probabilistic sparse self-attention mechanism.

[0038] ATT-Fuse-SCINet In the SCINet model, residual connections are used after the output sequence of the L-layer SCI-Block. Residual connections are a common neural network architecture that directly adds the input to the network output. This method brings efficient gradient backflow in the backpropagation of deep networks, reduces error accumulation, and alleviates gradient vanishing. The formula for residual connections is expressed as Equation (10).

[0039] (10) in It is the input sequence. This represents the output sequence of the L-layer SCI-Block processing. This represents the sequence after residual connection.

[0040] The residual fusion proposed in this embodiment is a more flexible structure that trains the input. and output Find the optimal weight fusion based on data features and errors. and The input and output contributions in the final result are adaptively adjusted to enhance the representational power of the model, as shown in Equation (11).

[0041] (11) Among them, weight and The importance of the output and input in the fusion process is determined by the Swish function and normalization. The weight calculation and Swish function are shown in Equation (12) and Equation (13), respectively.

[0042] (12) (13) in This represents the original weight parameters, with an initial value of 1. This is to prevent extremely small numbers with a denominator of 0, ensuring the stability of values ​​during training. The fused sequence result is represented as... .

[0043] Clustering Routing Based on Energy Prediction Clustering routing algorithm The LEACH protocol proposed in existing literature effectively reduces the energy consumption of WSN by introducing the concept of clustering. During the clustering process, the protocol adopts a cyclic cluster head election mechanism to ensure that each node has an equal probability of becoming a cluster head. Specifically, each sensor node in the network selects a random number from (0, 1). If the number is less than the threshold T(n), the node will be selected as the cluster head. The threshold is calculated as shown in formula (14).

[0044] (14) in It is the current node. It is the percentage of cluster heads. This is the current round. It is the set of nodes that did not become cluster heads.

[0045] Existing literature improves the LEACH protocol by adding the remaining energy to the influencing factors of cluster head election, thereby increasing the probability that nodes with more remaining energy will be elected as cluster heads. Its cluster head election formula is shown in formula (15).

[0046] (15) (16) in It is the remaining energy of the node. It is the average energy of the network.

[0047] EP-LEACH Protocol Algorithm For EH-WSNs, routing protocols that focus on the remaining energy of nodes while ignoring future energy harvesting lead to energy waste and inefficiency. Therefore, based on a high-precision energy prediction model, this embodiment proposes a clustering routing protocol. This protocol uses remaining energy and expected future harvesting energy as key factors for cluster head election. Nodes with higher energy have a greater probability of being elected as cluster heads. The improved cluster head election formula is expressed as Equation (17).

[0048] (17) (18) in It is the candidate cluster head set. This is the current round. It is the remaining energy of node i. It is the predicted energy of the node at the end of the current round. It is the initial energy of node i.

[0049] Dataset Introduction This experiment uses the publicly available irradiance dataset from the Tennant Creek station of the Australian Solar Centre (DKASC). The experimental sample includes data from September 29, 2023 to December 17, 2023. The dataset is divided into training, testing, and validation sets. The first 80% of the data is used as the training set, the last 10% as the testing set, and the remaining 10% as the validation set. The Pearson correlation coefficient method is used to analyze the correlation between meteorological characteristics such as horizontal irradiance, tilted irradiance, temperature, wind direction, wind speed, humidity, and direct irradiance. The correlation coefficients are as follows: horizontal irradiance 0.9583, tilted irradiance 0.9444, temperature 0.7124, humidity -0.5002, wind speed 0.1259, and wind direction 0.0833. The four features with the highest absolute correlation values ​​are selected to form a new feature set.

[0050] Evaluation indicators Solar energy prediction experiments use root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²). 2 The predictive accuracy of the model is evaluated by the mean absolute percentage error (MAPE) and calculated according to formulas (19), (20), (21), and (22), respectively.

[0051] (19) (20) (twenty one) (twenty two) in It is the actual value; It is a predicted value. This is the number of test samples.

[0052] Performance evaluation of the ATT-Fuse-SCINet model To effectively evaluate the performance of the proposed ATT-Fuse-SCINet model, it was compared with other models: LSTM, ARIMA-LSTM, SCINet, ATT-SCINet, and Fuse-SCINet. These models were used for multi-day continuous rolling forecasts to verify their long-term predictive performance. Furthermore, this experiment selected three different weather conditions (sunny, cloudy, and rainy) to evaluate the model's generalization ability, as shown in Figure 3. The experimental results for the evaluation metrics are presented in Tables 1-4.

[0053] Table 1

[0054] Table 2

[0055] Table 3

[0056] Table 4

[0057] The comparison in Table 1-4 shows that the ATT-Fuse-SCINet model proposed in this invention achieved the highest accuracy in the 8-day continuous prediction experiment. The RMSE, MAE, and MAPE were 0.0521, 0.0202, and 10.5804%, respectively. Compared with SCINet, RMSE decreased by 5.96%, MAE decreased by 11.79%, MAPE decreased by 13.70%, and R... 2 The maximum value is 0.9733. Experiments show that the ATT-Fuse-SCINet model has significant advantages in long-term prediction.

[0058] Due to the lack of complex nonlinear relationships under stable weather conditions, most models perform well under clear weather conditions. ATT-Fuse-SCINet achieves the highest prediction accuracy, with RMSE, MAE, and MAPE values ​​of 0.0106, 0.0076, and 2.081%, respectively. These are the minimum error values. R0 2 The maximum value is 0.9991. Due to the time lag issue of ARIMA, the accuracy of the ARIMA-LSTM model is compromised, with RMSE, MAE, and MAPE reaching 0.0376, 0.0327, and 8.9983%, respectively. 2The R² value is 0.9889. Under unstable weather conditions, such as cloudy and rainy days, LSTM performs the worst. ATT-Fuse-SCINet achieves the best performance across all metrics, followed by the Fuse-SCINet model. Taking the most complex rainy day example, ATT-Fuse-SCINet's RMSE, MAE, and MAPE values ​​are 0.0758, 0.0477, and 15.6985%, respectively, which are the smallest error values ​​among all models. Compared to SCINet, these values ​​are reduced by 18.05%, 25.71%, and 26.50%, respectively. 2 The maximum value is 0.9017.

[0059] Network lifespan comparison Simulation experiments were conducted using MATLAB 2022(a) software, and the simulation parameters are shown in Table 5. The proposed EP-LEACH clustering routing protocol was integrated with LSTM, SCINet, and ATT-Fuse-SCINet prediction models, and compared with the LEACH algorithm in terms of the number of surviving nodes, total network energy, and number of dead nodes. The experiments aimed to evaluate the effect of the proposed EP-LEACH algorithm on extending network lifetime and to analyze the impact of different energy prediction models on EP-LEACH performance.

[0060] Table 5

[0061] Experimental results are as follows Figure 4 , Figure 5 and Figure 6 As shown. Figure 4 This chart compares the number of surviving nodes for the LEACH algorithm and the EP-LEACH algorithm under three different prediction models. In the first 158 ​​rounds, none of the four algorithms experienced any dead nodes. The number of surviving nodes decreased with each round. By round 400, the number of surviving nodes for the LEACH algorithm, EP-LEACH-LSTM, and EP-LEACH-SCINet were 79.3%, 89.7%, and 93.1% of those for EP-LEACH-ATT-Fuse-SCINet, respectively; by round 800, these values ​​were 27.8%, 50%, and 90.7%.

[0062] Figure 5This chart compares the total network energy of the LEACH algorithm and the EP-LEACH algorithm under three different prediction models. The total network energy decreases with increasing rounds. The total network energy decreases fastest under the LEACH algorithm, while the decrease is slowest under the EP-LEACH-ATT-Fuse-SCINet algorithm. By round 400, the total network energy of the LEACH algorithm, EP-LEACH-LSTM, and EP-LEACH-SCINet are 63.6%, 80.3%, and 89.1% of that of EP-LEACH-ATT-Fuse-SCINet, respectively; by round 800, these values ​​are 15.4%, 34.6%, and 76.9%.

[0063] according to Figure 6 The statistics of dead nodes show that the EP-LEACH-LSTM algorithm encountered its first dead node in round 158, while the LEACH algorithm encountered its first dead node in round 168. The low accuracy of energy prediction negatively impacts the selection of cluster heads. The EP-LEACH-SCINet and EP-LEACH-ATT-Fuse-SCINet algorithms encountered their first dead nodes in rounds 186 and 191, respectively. The LEACH algorithm had half of its nodes die in round 490, while the EP-LEACH-LSTM, EP-LEACH-SCINet, and EP-LEACH-ATT-Fuse-SCINet algorithms had half of their nodes die in rounds 516, 546, and 570, respectively. The LEACH algorithm had all its nodes die in round 949, while the EP-LEACH-LSTM, EP-LEACH-SCINet, and EP-LEACH-ATT-Fuse-SCINet algorithms had all their nodes die in rounds 1107, 1153, and 1193, respectively.

[0064] Experimental results demonstrate that, compared to the LEACH algorithm, the EP-LEACH algorithm incorporating an energy prediction model exhibits lower energy consumption and a slower node mortality rate. Specifically, the EP-LEACH algorithm incorporating the ATT-Fuse-SCINet energy prediction model significantly outperforms algorithms integrating other prediction models in extending network lifetime.

[0065] in conclusion To better utilize solar energy and extend the lifespan of WSNs, this embodiment proposes an energy prediction and clustering routing algorithm for EH-WSNs based on an attention mechanism and an improved SCINet model. Based on experimental analysis, the following conclusions can be drawn: 1) To address the issue of poor accuracy in energy prediction algorithms within EH-WSNs, the proposed ATT-Fuse-SCINet model exhibits better stability and generalization ability compared to other prediction models. Solar energy prediction is significantly improved under both long-term forecasting and unstable weather conditions.

[0066] 2) Furthermore, addressing the issues of limited node energy and poor sustainability in WSNs, this embodiment proposes an EP-LEACH clustering routing algorithm based on high-precision energy prediction. By introducing a high-precision energy prediction algorithm into the network, EP-LEACH effectively optimizes energy resource utilization, alleviates node mortality and network energy consumption, and extends the overall network lifetime.

[0067] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A wireless network energy prediction and clustering routing method based on attention mechanism and improved SCINet, characterized in that, include: Data normalization and feature extraction are performed on energy data from wireless sensor networks; The ATT-Fuse-SCINet model is used to predict energy from the processed energy data and obtain the energy prediction results. The ATT-Fuse-SCINet model includes an SCI-Block structure, a probabilistic sparse self-attention mechanism, and a residual fusion structure. Based on the energy prediction results, the EP-LEACH clustering routing algorithm is used to elect the cluster head, taking into account both the remaining energy and the future predicted energy of the nodes during the election process.

2. The method according to claim 1, characterized in that, The data normalization method is as follows: ， in, It is a feature The normalized value; It is a feature The original value; and These are the maximum and minimum values ​​of the feature, respectively.

3. The method according to claim 1, characterized in that, The feature extraction method is as follows: ， in, It is the correlation coefficient; It is the sample size; , It consists of the target sequence values ​​and their average values. and These are other feature sequence values ​​and their average values.

4. The method according to claim 1, characterized in that, The SCI-Block structure includes: The input sequence is downsampled into odd and even sequences; One-dimensional convolutional modules are used to transform odd and even sequences, and interactive learning is performed through Hadamard products. New subsequences are generated through convolution mapping and addition / subtraction operations to compensate for information loss during downsampling.

5. The method according to claim 1, characterized in that, The probabilistic sparse self-attention mechanism calculates the sparsity of the Query using the following formula: ， in, It is a measure of sparsity; yes The first in One element; yes The first in One element; yes The length.

6. The method according to claim 1, characterized in that, The residual fusion structure adopts the following formula: ， Among them, weight and The importance of the output and input in the fusion process is indicated by the Swish function and normalization.

7. The method according to claim 1, characterized in that, The weight calculation and Swish function of the residual fusion structure are shown in the following formula: ， ， in, This represents the calculated weight parameters. This represents the original weight parameters, with an initial value of 1. It is to prevent extremely small numbers with a denominator of 0.

8. The method according to claim 1, characterized in that, The method for cluster head election using the EP-LEACH clustering routing algorithm adopts the following formula: ， ， in, It is the current node. It is the percentage of cluster heads. It is the candidate cluster head set. This is the current round. It is the remaining energy of node i. It is the predicted energy of the node at the end of the current round. It is the initial energy of node i.

Images