A dual-mode communication dynamic routing optimization method and system based on AI link prediction

Abstract

The application provides a dual-mode communication dynamic routing optimization method and system based on AI link prediction. Through deep mining of historical data, a prediction model is constructed to make multi-step prediction of future link quality. Based on the prediction results, forward-looking dynamic routing optimization is achieved. Under the premise of protecting data privacy, AI technology is used to dynamically adapt to the complex time-varying environment of low-voltage power grids, significantly improving the communication reliability and long-term operation adaptability of the dual-mode communication module in complex power grid environments.

Description

Technical Field

[0001] This invention relates to the field of power Internet of Things / smart grid communication technology, and in particular to a dual-mode communication dynamic routing optimization method and system based on AI link prediction. Background Technology

[0002] Dual-mode communication technology, combining high-speed power line carrier (HPLC) and high-speed radio frequency (HRF), is currently the mainstream solution for low-voltage distribution area power consumption information acquisition systems. This solution features high real-time performance and reliability, and can be used in a wide range of business scenarios.

[0003] However, the power grid operating environment is highly dynamic and uncertain. With the influence of factors such as equipment aging, seasonal load changes, and the integration of new energy sources, the characteristics of power line carrier channels continuously change, making them susceptible to variations in grid impedance and harmonic interference, leading to decreased communication reliability. Meanwhile, wireless communication is affected by transmission distance limitations and changes in the electromagnetic environment, resulting in poor reliability over long distances. Traditional dual-mode communication systems typically employ static routing or channel switching strategies based on simple signal strength thresholds, which are ill-suited to the dynamic evolution of the power grid environment during long-term operation, exhibiting the following technical shortcomings:

[0004] 1. Static routing cannot respond to changes in channel quality in real time, and the communication success rate decreases significantly over time.

[0005] 2. Channel selection strategies based on real-time measurements lack foresight, triggering handover only when channel quality deteriorates sharply. This handover lag leads to data transmission failures and increased retransmission overhead.

[0006] 3. Existing routing methods mainly rely on the link state parameters at the current moment for decision-making, failing to make full use of historical data to mine the patterns of channel quality evolution, and thus have limited predictive capabilities.

[0007] Therefore, there is an urgent need for a method that can use historical data to make forward-looking predictions of the quality of dual-mode communication links and achieve dynamic route optimization based on the prediction results. Summary of the Invention

[0008] The purpose of this application is to provide a dynamic routing optimization method and system for dual-mode communication based on AI link prediction, so as to solve the problem of existing dual-mode communication modules relying on static routing tables or real-time measurement values ​​for channel selection.

[0009] To address the aforementioned technical problems, this application provides a dynamic routing optimization method for dual-mode communication based on AI link prediction, comprising the following steps:

[0010] Step 1, Data Acquisition Step: Each node periodically collects multi-dimensional link indicators of the carrier and wireless channel. The multi-dimensional link indicators include communication success rate (PSR), signal-to-noise ratio (SNR), received signal strength (RSSI), and transmitted power (TX Power), forming a historical time-series data sequence of fixed length L.

[0011] Step 2, AI prediction engine steps: Based on the encoder-decoder architecture model built on LSTM, the multi-dimensional link indication features within the historical time window are used as output to predict the link quality of future time steps in multiple steps, and the comprehensive link quality index of carrier and wireless channel at each time step is output as the prediction result.

[0012] Step 3, Routing Decision Step: Based on the prediction results, calculate the total cost of each path. The total cost consists of a weighted average of link quality cost, channel handover cost, and hop count penalty. Use Dijkstra's algorithm to calculate the minimum cost path from the Central Coordinator (CCO) to each STA. Assign the predicted optimal channel to each link, generate a dynamic routing table, and send it to each STA via beacon frames.

[0013] Step 4, Model Update Step: The Central Coordinator (CCO) acts as the federated learning server, and the site (STA) acts as the client. Under the premise of protecting the privacy of the original data, the local training gradients are aggregated to update the global prediction model.

[0014] As a preferred embodiment, a dual-mode communication dynamic routing optimization method based on AI link prediction calculates the comprehensive link quality index using the following formula:

[0015]

[0016]

[0017] in, , , , For configurable weighting coefficients, satisfying The default value is , , , ; For carrier-integrated link quality, , , , , , , These are the communication success rate, signal-to-noise ratio, maximum signal-to-noise ratio, received signal strength, maximum received signal strength, transmit power, and minimum transmit power of the carrier channel. For wireless integrated link quality, , , , , , , These are the wireless channel's communication success rate, signal-to-noise ratio, maximum signal-to-noise ratio, received signal strength, maximum received signal strength, transmit power, and minimum transmit power.

[0018] As a preferred embodiment, a dual-mode communication dynamic routing optimization method based on AI link prediction is provided, wherein the link quality cost is defined as the reciprocal of the effective comprehensive quality index:

[0019]

[0020] Among them, the effective comprehensive quality index This is the weighted average of the next M time steps, with the weights of the most recent time steps increasing exponentially by a decay factor γ = 0.9. To prevent the minimum value from being divided by zero.

[0021] As a preferred embodiment, a dual-mode communication dynamic routing optimization method based on AI link prediction, wherein the channel switching cost is:

[0022]

[0023] in, This is a channel switching indication function, used when the i-th hop and the (i+1)-th hop in the path use different channels. Otherwise, it is 0; β=0.2 is the switching penalty coefficient, and H is the number of hops.

[0024] As a preferred embodiment, a dual-mode communication dynamic routing optimization method based on AI link prediction, wherein the hop count penalty is:

[0025]

[0026] in, = 0.05 is the fixed cost per hop, and H is the number of hops.

[0027] As a preferred embodiment, a dual-mode communication dynamic routing optimization method based on AI link prediction is provided, wherein the communication success rate (PSR) is calculated as follows:

[0028]

[0029] In the above formula, The number of data packets for which ACK confirmations were successfully received; This represents the total number of data packets sent.

[0030] The signal-to-noise ratio (SNR) is calculated as follows:

[0031]

[0032] In the above formula, The signal power measured at the receiving end; The signal-to-noise ratio (SNR) is measured in dB. The carrier and wireless channel are periodically calculated based on the configuration items to obtain the noise floor.

[0033] To address the aforementioned technical issues, this application also provides a dual-mode communication dynamic routing optimization system based on AI link prediction, comprising a data acquisition module deployed at the Central Coordinator (CCO) and each STA site, an AI prediction engine module and a routing decision module deployed at the CCO, and a model update module deployed at the CCO and each STA site.

[0034] The data acquisition module is used to periodically collect multi-dimensional link indicators of the carrier and wireless channel at each node. The multi-dimensional link indicators include communication success rate (PSR), signal-to-noise ratio (SNR), received signal strength (RSSI), and transmitted power (TX_Power), forming a historical time-series data sequence of fixed length L.

[0035] The AI ​​prediction engine module is used for an encoder-decoder architecture model based on LSTM. It takes multi-dimensional link features within the historical time window as input, performs multi-step prediction of link quality in future time steps, and outputs the comprehensive link quality index of carrier and wireless channel at each time step as the prediction result.

[0036] The routing decision module is used to calculate the total cost of each path based on the prediction results. The total cost consists of a weighted average of link quality cost, channel switching cost, and hop count penalty. The Dijkstra algorithm is used to calculate the minimum cost path from the Central Coordinator (CCO) to each STA. The module also allocates the predicted optimal channel to each link, generates a dynamic routing table, and sends it to each STA via beacon frames.

[0037] The model update module is used to aggregate local training gradients and update the global prediction model, with the central coordinator (CCO) acting as the federated learning server and the site (STA) acting as the client, while protecting the privacy of the original data.

[0038] As a preferred embodiment, a dual-mode communication dynamic routing optimization system based on AI link prediction is provided, wherein the data acquisition module is further used to perform sliding window normalization processing, and the normalization parameters are dynamically updated in a sliding window manner to adapt to the index drift during long-term operation.

[0039] As a preferred embodiment, a dual-mode communication dynamic routing optimization system based on AI link prediction is characterized in that the routing decision module is also used to maintain a neighbor discovery table, which includes the network access status, neighbor relationships and hop count limits of all STAs, and supports the maximum hop count constraint of the path.

[0040] The present application provides a dynamic routing optimization method and system for dual-mode communication based on AI link prediction. By deeply mining historical data, a prediction model is constructed to predict the future link quality in multiple steps, and dynamic routing optimization is achieved based on the prediction results. This enables the use of AI technology to dynamically adapt to the complex time-varying environment of low-voltage power grids while protecting data privacy, and significantly improves the communication reliability and long-term operational adaptability of dual-mode communication modules in complex power grid environments. Attached Figure Description

[0041] To more clearly illustrate the technical solution of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0042] Figure 1 This invention provides an overall architecture and module deployment diagram of a dual-mode communication dynamic routing optimization system based on AI link prediction.

[0043] Figure 2 This invention provides a data acquisition and preprocessing flowchart;

[0044] Figure 3 This invention provides a dual-mode communication network topology diagram;

[0045] In the diagram: 1. Data acquisition module; 2. AI prediction engine module; 3. Routing decision module; 4. Model update module. Detailed Implementation

[0046] To enable those skilled in the art to better understand the technical solutions in this application, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings.

[0047] The core of this application is to provide a dynamic routing optimization method and system for dual-mode communication based on AI link prediction, which solves the problem of existing dual-mode communication modules relying on static routing tables or real-time measurement values ​​for channel selection.

[0048] Figure 1 This invention provides an overall architecture and module deployment diagram of a dual-mode communication dynamic routing optimization system based on AI link prediction. Figure 2 This invention provides a data acquisition and preprocessing flowchart; Figure 3 See the dual-mode communication network topology diagram provided by this invention. Figures 1 to 3 As shown.

[0049] Example 1

[0050] A dual-mode communication dynamic routing optimization system based on AI link prediction is applicable to intelligent routing decision-making and adaptive optimization in high-speed power line carrier (HPLC) and high-speed radio frequency (HRF) dual-mode communication networks. The system includes a data acquisition module 1, an AI prediction engine module 2, a routing decision module 3, and a model update module 4. Figure 1 As shown, data acquisition module 1 is deployed on the CCO and all STAs, with each node collecting its own observed link status indicators; AI prediction engine module 2 is deployed on the CCO, using an LSTM temporal neural network to perform multi-step predictions of the future quality of each link; routing decision module 3 is deployed on the CCO, where the routing table is centrally calculated, dynamically calculating the optimal routing table based on the prediction results, and distributing it to each STA via beacons or configuration messages; model update module 4 is deployed on the CCO and STAs, employing a federated learning architecture, with the CCO acting as the server and the STAs as clients. The CCO is responsible for aggregating the gradients uploaded by each STA and updating the global model; the STAs are responsible for adjusting the model locally using the collected data and uploading the gradients.

[0051] Example 2

[0052] A dynamic routing optimization method for dual-mode communication based on AI link prediction includes the following steps:

[0053] Step 1, the data acquisition steps include steps 1.1 and 1.2:

[0054] Step 1.1: The CCO or STA node collects link quality indicators on the carrier channel and the wireless channel at a sampling period of 15 minutes. The multi-dimensional link quality parameters include communication success rate (PSR); signal-to-noise ratio (SNR), the ratio of signal power to noise power of the carrier or wireless channel; received signal strength (RSSI), the signal power strength measured at the receiver; and transmit power (TX Power), the transmit power used by the transmitter of the carrier or wireless channel at the current link.

[0055] Within a unit statistical time window ΔT, the communication success rate (PSR) is calculated as follows:

[0056]

[0057] In the above formula, The number of data packets for which ACK confirmations were successfully received; This represents the total number of data packets sent.

[0058] The signal-to-noise ratio (SNR) is calculated as follows:

[0059]

[0060] In the above formula, The signal power measured at the receiving end; The signal-to-noise ratio (SNR) is measured in dB. The carrier and wireless channel are periodically calculated based on the configuration items to obtain the noise floor.

[0061] The collected data is recorded according to a unified timestamp, forming a multidimensional time-series data sequence, such as... Multidimensional time series data sequence x(t) can be expressed by the following formula:

[0062]

[0063] In the above formula, the acquired data from carrier and wireless channels are distinguished by Channel_Type, where HPLC represents the carrier channel and HRF represents the wireless channel. , , , , , , , .

[0064] Step 1.2, as follows Figure 2 As shown, all samples are stored in the local circular buffer of each node in chronological order, forming a fixed-length L historical multidimensional time-series data sequence set. This is used for subsequent LSTM model training and prediction, selecting a historical time window T. history All eight-dimensional link features within the data are used as input, with L preferably having a length of 96, i.e., T hhistory The preferred sampling period is 24 hours, 15 minutes, corresponding to 96 sampling points. It can be expressed by the following formula:

[0065]

[0066] Normalizing the data for each feature dimension and mapping it to the [0,1] interval can be expressed by the following formula:

[0067]

[0068] In the above formula, the normalization parameter ( , The indicator is updated in a sliding window manner at the CCO end to adapt to indicator drift during long-term operation.

[0069] Training samples are generated using a sliding window method, with each sample having a shape (T) history The time series matrix of (8) corresponds to the label of future T. future The target for predicting link quality within the T region. future The preferred sampling period is 4 hours, 15 minutes, and 16 sampling points.

[0070] Step 2, the AI ​​prediction engine steps include steps 2.1 to 2.5:

[0071] Step 2.1: The LSTM prediction model adopts an encoder-decoder architecture, consisting of an input layer, encoder, decoder, and output layer. For example... , Normalized This yields historical time-series data for the input layer of the LSTM prediction model.

[0072] The input layer receives preprocessed historical time-series data, and the input matrix is ​​expressed as follows:

[0073]

[0074] In the above formula, For the link feature dimension, i.e. ; For historical time window, i.e. T history N represents the 96 sampling points.

[0075] Step 2.2: The encoder consists of two stacked LSTM layers, each containing 128 hidden units. The encoder processes the input sequence sequentially according to time steps. Each feature vector in the matrix. Let Xi be the input of the encoder at time i (i=1,2,...,N), and its calculation process can be expressed as:

[0076] (7)

[0077] In the above formula, Let be the hidden state at time i. In cellular state. Initial hidden state. and initial cell state All vectors are initialized to zero.

[0078] At time i=0, the initial hidden state Initial cell state At time i=1, the encoder receives the feature vector of the first time step according to expression (7). Combined with the initial hidden state and initial cell state Output the hidden state of the current step. and cell state .

[0079] Step 2.3, based on step 2.2, at time i=2, the encoder receives... and the previous step As the initial state, output This process continues until the last time step i=N is completed. After the encoder has processed all N time steps, it outputs the final hidden state. and final cell state These two vectors together constitute the context encoding of the input historical sequence, condensing all the temporal feature information of N time steps.

[0080] Step 2.4: The decoder's initialization state is not initialized with zero, but is directly initialized with the encoder's final state.

[0081] (8)

[0082] (9)

[0083] The decoder also consists of two stacked LSTM layers, each with 128 hidden units, and its structure is symmetrical to the encoder. The decoder uses an autoregressive approach, generating predictions for the next M=16 time steps, i.e., T, step by step. future In the first decoding step j=1, the decoder receives the start symbol vector. The start symbol is taken from the actual observation xi at the last time step of the history window. The decoder combines the initial state. With input Calculate the hidden state of the current step. and cell state The following expression:

[0084] (10)

[0085] In subsequent decoding steps j=2, 3, ..., M, the decoder receives the prediction output from the previous step. As the input for the current step, combined with the hidden state from the previous step. Output the hidden state of the current step. and cell state The following expression:

[0086] (11)

[0087] Step 2.5: The output layer is a fully connected layer followed by a sigmoid activation function. This corresponds to the hidden state output by the decoder at each time step. The output layer linearly maps it to a d-dimensional output vector and constrains the output value to the [0,1] interval using the Sigmoid activation function. The expression is as follows:

[0088] (12)

[0089] In the above formula, This is the weight matrix. For bias vectors, The sigmoid activation function is used. In the LSTM prediction model described in this invention, both the weight matrix and the bias vector are trainable parameters, and their optimal values ​​are automatically learned during the model training process, requiring no manual setting.

[0090] For any channel, define its composite link quality index. This is the weighted fusion result of the various indicators, with a value range of [0, 1]. A larger value indicates better link quality. The expression is as follows:

[0091] (13)

[0092] (14)

[0093] In the above formula, , , , For configurable weighting coefficients, satisfying Based on expressions (13) and (14), the combined link quality index of the carrier and the wireless channel is calculated. and The default value in this application is , , , The weighting coefficients are allocated based on the importance of each indicator in the link quality assessment.

[0094] For the j-th future time step, the maximum of the two values ​​is taken as the optimal channel composite quality index for that time step. The following expression:

[0095] (15)

[0096] And record the corresponding optimal channel The following expression:

[0097] (16)

[0098] After calculation for all j=1,...,16, the comprehensive quality index prediction sequence for the next 4 hours is obtained. and The expression is as follows:

[0099] (17)

[0100] (18)

[0101] Step 3, the routing decision steps include steps 3.1 to 3.5:

[0102] Step 3.1, as follows Figure 3 The neighbor discovery table maintained by the CCO contains the network entry status, neighbor relationships, and hop count limits for all STAs. For any candidate path p = [n0, n1, ..., nH], where n0 = CCO, nH is the target STA, H is the number of paths, and the directed link corresponding to the i-th hop in the path is li = (ni-1, ni). The path cost consists of three weighted components: link quality cost, channel handover cost, and hop count penalty.

[0103] Step 3.2 defines the effective overall quality index of the i-th hop link as the weighted average of the next M time steps, with higher weights for recent times to reflect the recent trend of link quality, as shown in the following expression:

[0104] (19)

[0105] In the above formula, As the attenuation factor, the present invention preferably... This results in a higher weight for more recent time steps. A Qeff value closer to 1 indicates better link quality.

[0106] The link quality cost is defined as the reciprocal of the effective quality index, as shown in the following formula:

[0107] (20)

[0108] Step 3.3: If two consecutive paths use different channels, a channel handover cost is introduced. A channel handover indication function is defined between the i-th hop and the (i+1)-th hop:

[0109] (twenty one)

[0110] The channel switching cost is:

[0111] (twenty two)

[0112] In the above formula, To switch the penalty coefficient, the present invention preferably... .

[0113] Meanwhile, to avoid excessive end-to-end latency due to excessively long paths, a hop count penalty is introduced, as shown in the following expression:

[0114] (twenty three)

[0115] In the above formula, For each fixed cost, the present invention preferably .

[0116] The total cost of candidate path p is:

[0117] (twenty four)

[0118] Step 3.4: Using CCO as the source, Dijkstra's shortest path algorithm is used to calculate the minimum cost path to each STA, and the predicted optimal channel is assigned to each link.

[0119] Step 3.5: The calculation results are encapsulated into a routing table, containing the destination address, next path, preferred channel, backup channel, hop count, and validity period. This table is then broadcast to each STA via beacon fragmentation. This enables proactive and dynamic switching of communication paths, effectively mitigating link quality degradation.

[0120] Step 4, Model Update Step: When the CCO detects that the cumulative prediction error of the global model exceeds a threshold or reaches a preset time window, the CCO randomly selects 30% of the STAs to participate in this federated learning based on their computing power, current channel status, and historical participation, in order to balance the computational load.

[0121] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the application disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and incorporate common knowledge or customary techniques in the art disclosed herein. The specification and examples are to be considered exemplary only, and the true scope of this application is indicated by the claims.

[0122] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The embodiments of this application described above do not constitute a limitation on the scope of protection of this application.

Claims

1. A dynamic routing optimization method for dual-mode communication based on AI link prediction, characterized in that, Includes the following steps: Step 1, Data Acquisition Step: Each node periodically collects multi-dimensional link indicators of the carrier and wireless channel. The multi-dimensional link indicators include communication success rate (PSR), signal-to-noise ratio (SNR), received signal strength (RSSI), and transmitted power (TX Power), forming a historical time-series data sequence of fixed length L. Step 2, AI prediction engine steps: Based on the encoder-decoder architecture model built on LSTM, the multi-dimensional link indication features within the historical time window are used as output to predict the link quality of future time steps in multiple steps, and the comprehensive link quality index of carrier and wireless channel at each time step is output as the prediction result. Step 3, Routing Decision Step: Based on the prediction results, calculate the total cost of each path. The total cost consists of a weighted average of link quality cost, channel handover cost, and hop count penalty. Use Dijkstra's algorithm to calculate the minimum cost path from the Central Coordinator (CCO) to each STA. Assign the predicted optimal channel to each link, generate a dynamic routing table, and send it to each STA via beacon frames. Step 4, Model Update Step: The Central Coordinator (CCO) acts as the federated learning server, and the site (STA) acts as the client. Under the premise of protecting the privacy of the original data, the local training gradients are aggregated to update the global prediction model.

2. The dual-mode communication dynamic routing optimization method based on AI link prediction according to claim 1, characterized in that, The comprehensive link quality index is calculated using the following formula: in, , , , For configurable weighting coefficients, satisfying The default value is , , , ; For carrier-integrated link quality, , , , , , , These are the communication success rate, signal-to-noise ratio, maximum signal-to-noise ratio, received signal strength, maximum received signal strength, transmit power, and minimum transmit power of the carrier channel. For wireless integrated link quality, , , , , , , These are the wireless channel's communication success rate, signal-to-noise ratio, maximum signal-to-noise ratio, received signal strength, maximum received signal strength, transmit power, and minimum transmit power.

3. The dual-mode communication dynamic routing optimization method based on AI link prediction according to claim 1, characterized in that, The link quality cost is defined as the reciprocal of the effective overall quality index: Among them, the effective comprehensive quality index This is a weighted average of the next M time steps, with the weights of the most recent time steps adjusted by a decay factor. The exponential increases. To prevent the minimum value from being divided by zero.

4. The dual-mode communication dynamic routing optimization method based on AI link prediction according to claim 1, characterized in that, The channel switching cost is: in, This is a channel switching indication function, used when the i-th hop and the (i+1)-th hop in the path use different channels. Otherwise, it is 0; =0.2 is the switching penalty coefficient, and H is the number of jumps.

5. The dual-mode communication dynamic routing optimization method based on AI link prediction according to claim 1, characterized in that, The jump count penalty is: in, =0.05 is the fixed cost per line. This represents the number of jumps.

6. The dual-mode communication dynamic routing optimization method based on AI link prediction according to claim 1, characterized in that, The communication success rate (PSR) is calculated using the following formula: In the above formula, The number of data packets for which ACK confirmations were successfully received; This represents the total number of data packets sent. The signal-to-noise ratio (SNR) is calculated as follows: In the above formula, The signal power measured at the receiving end; The signal-to-noise ratio (SNR) is measured in dB. The carrier and wireless channel are periodically calculated based on the configuration items to obtain the noise floor.

7. A dual-mode communication dynamic routing optimization system based on AI link prediction, characterized in that, It includes a data acquisition module (1) deployed in the Central Coordinator (CCO) and each STA, an AI prediction engine module (2) and a routing decision module (3) deployed in the Central Coordinator (CCO), and a model update module (4) deployed in the Central Coordinator (CCO) and each STA: The data acquisition module (1) is used to periodically collect multi-dimensional link indicators of carrier and wireless channel at each node. The multi-dimensional link indicators include communication success rate (PSR), signal-to-noise ratio (SNR), received signal strength (RSSI), and transmitted power (TX_Power), forming a historical time-series data sequence of fixed length L. The AI ​​prediction engine module (2) is used for the encoder-decoder architecture model based on LSTM. It takes the multi-dimensional link features within the historical time window as input, performs multi-step prediction of the link quality of future time steps, and outputs the comprehensive link quality index of the carrier and wireless channel at each time step as the prediction result. The routing decision module (3) is used to calculate the total cost of each path based on the prediction results. The total cost consists of link quality cost, channel switching cost and hop count penalty weighting. The Dijkstra algorithm is used to calculate the minimum cost path from the central coordinator (CCO) to each station (STA), and the predicted optimal channel is allocated to each link. A dynamic routing table is generated and sent to each station (STA) through beacon frames. The model update module (4) is used to aggregate local training gradients and update the global prediction model under the premise of protecting the privacy of the original data, with the central coordinator (CCO) acting as the federated learning server and the site (STA) acting as the client.

8. The dual-mode communication dynamic routing optimization system based on AI link prediction according to claim 7, characterized in that, The data acquisition module (1) is also used to perform sliding window normalization processing, and the normalization parameters are dynamically updated in a sliding window manner to adapt to the drift of indicators during long-term operation.

9. The dual-mode communication dynamic routing optimization system based on AI link prediction according to claim 7, characterized in that, The routing decision module (3) is also used to maintain a neighbor discovery table, which includes the network entry status, neighbor relationships and hop count limits of all STAs, and supports the maximum hop count constraint of the path.

Images