Method and system for estimating air-ground interconnection bandwidth under high dynamic link

Abstract

This invention discloses a method and system for estimating bandwidth of air-to-ground interconnection under highly dynamic links. By introducing a bidirectional long short-term memory network based on an attention mechanism, it can not only extract temporal features from both forward and backward directions, but also automatically identify key historical moments using attention weight factors. This solves the problem of insufficient capture of long-period dependence factors of highly dynamic links in the prior art, enabling the system to more accurately grasp the macroscopic change law of network bandwidth over a long time scale and avoid estimation deviations caused by ignoring historical trends.

Description

Technical Field

[0001] This invention relates to the field of mobile communications, and in particular to a method and system for estimating bandwidth for air-to-ground interconnection under high dynamic links. Background Technology

[0002] In highly dynamic link network scenarios, many bandwidth estimation methods focus on traditional methods, typically incorporating historical bandwidth fluctuations for estimation. These methods do not rely on large amounts of training data, but rather extrapolate from the statistical regularities of historical throughput. In recent years, deep learning models have made significant progress in bandwidth estimation, usually combining network state with machine learning modules to estimate network bandwidth, thereby meeting high bandwidth demands and making effective decisions. However, in complex network environments with drastic bandwidth fluctuations, existing traditional methods and deep learning-based methods cannot achieve high-precision bandwidth estimation. Existing methods lack the ability to capture long-term dependencies in highly dynamic network links, making it difficult to grasp long-term macro-level trends in network bandwidth. Furthermore, they are slow to respond to short-term mutations in highly dynamic link network scenarios and lack bandwidth adaptation mechanisms. Most existing research focuses on optimization at a single time scale, lacking a unified framework that can collaboratively handle long-term trend prediction and short-term mutation tracking.

[0003] In summary, a method and system for estimating air-to-ground interconnection bandwidth under highly dynamic links is needed to address the shortcomings of existing technologies. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method and system for estimating air-to-ground interconnection bandwidth under highly dynamic links, aiming to solve the aforementioned problems.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for estimating air-to-ground interconnection bandwidth under high dynamic links, comprising the following steps:

[0006] Step S1: Raw broadband data acquisition, acquiring raw measurement data of the network link at a fixed sampling period;

[0007] Step S2: Multivariate feature vector construction. Based on the original measurement data, calculate the second-level bandwidth difference, combine the original measurement data to form a multivariate feature vector, and detect fine-grained bandwidth changes.

[0008] Step S3: Long-term trend prediction. Input the multivariate feature vector into the bidirectional long short-term memory network model based on the attention mechanism, use the forward and backward LSTM networks to extract the bandwidth temporal features, and combine the attention weight factor to calculate the predicted value of bandwidth change trend over a long time scale.

[0009] Step S4: Network status determination. Calculate the network status quantification value based on the original measurement data, and classify the current network status into light load, full load, or congestion state according to the preset threshold.

[0010] Step S5: Short-term incremental adaptive estimation: Using the long-term trend prediction value as the initial value, and combining it with the network state, the bandwidth is adjusted using the incremental adaptive AIMD improved algorithm, and the final bandwidth estimation value is output.

[0011] Optionally, the raw measurement data in step S1 includes the current bandwidth value representing the currently available network bandwidth, the packet loss rate representing the proportion of data packets lost per unit time, and the delay jitter representing the variance of the transmission delay of adjacent data packets.

[0012] Optionally, step S2 is implemented in the following manner:

[0013] Step S21: Calculate the second-level bandwidth difference. Based on the collected raw measurement data, calculate the bandwidth difference between the current time and the previous time.

[0014] Step S22: Combine the multivariate feature vectors by concatenating the original measurement data with the second-level bandwidth difference to form a multivariate feature vector;

[0015] Step S23: Feature normalization processing, normalize each feature dimension in the combined feature vector.

[0016] Optionally, the second-level bandwidth difference bdps(t) in step S21 is calculated in the following way:

[0017] bdps(t) = |b(t) - b(t-1)|,

[0018] In the formula, b(t) is the current measured bandwidth value, and b(t-1) is the bandwidth measured value at the previous moment;

[0019] The fine-grained change in bandwidth is detected by comparing the second-level bandwidth difference with the bandwidth measurement value at the previous moment.

[0020] Optionally, step S3 is implemented in the following manner:

[0021] Step S31: Feature vector fusion and input. The multivariate feature vectors are weighted and fused to generate a single fused value, which is used as the input vector of the bidirectional long short-term memory network model.

[0022] Step S32: Bidirectional temporal feature extraction. Feature extraction is performed using a bidirectional long short-term memory network. The cell state and hidden state are updated through the forget gate, input gate, and output gate. The forward hidden vector and the backward hidden vector are calculated respectively, and the final hidden vector is calculated by weighting.

[0023] Step S33: Attention weighting and trend prediction. The hidden vector output by the bidirectional long short-term memory network model is used as input. Combined with the attention weight factor, key historical moments are automatically identified and given higher weights. Based on the calculated attention weights and hidden vectors, the bandwidth change trend prediction value over a long time scale is calculated and output through a fully connected layer.

[0024] Optionally, the bandwidth change trend prediction value in step S33 Calculated in the following way:

[0025] ,

[0026] In the formula, It is a multivariate prediction function. To predict the step size, For multivariate time series based on historical bandwidth, Let be a feature vector containing m kinds of historical information.

[0027] Optionally, step S4 is implemented in the following manner:

[0028] Step S41: Calculate the network state quantization value. Based on the packet loss rate and latency jitter, a comprehensive quantization value is obtained through weighted calculation.

[0029] Step S42: Determine the network state category. Based on the calculated quantization value, compare it with a predefined threshold to classify the current network state into light load, full load, and congestion state.

[0030] Optionally, step S5 is implemented in the following manner:

[0031] Step S51: Initialize the input by using the macroscopic trend prediction value output from the long-term bandwidth trend prediction as the initial baseline value for the short-term estimation.

[0032] Step S52: Strategy selection, switching and adjusting strategies in real time according to different network conditions;

[0033] Step S53: Incremental adaptive calculation and output. Based on the measured bandwidth at time t, the bandwidth threshold at the time of the last congestion, and the average bandwidth of the previous n times, the incremental value is dynamically calculated, and the final bandwidth estimate is output.

[0034] A high-dynamic link air-to-ground interconnection bandwidth estimation system, employing the high-dynamic link air-to-ground interconnection bandwidth estimation method, includes a raw data acquisition module, a multi-dimensional feature construction module, a long-term trend prediction module, a network status determination module, and a short-term incremental adaptive estimation module.

[0035] The raw data acquisition module is used to collect raw measurement data of the network link in real time at a fixed sampling period. It collects the bandwidth value representing the current available network bandwidth, the packet loss rate representing the proportion of data packets lost per unit time, and the delay jitter representing the variance of the transmission delay of adjacent data packets.

[0036] The multi-feature construction module is used to calculate the second-level bandwidth difference based on the original measurement data, thereby detecting fine-grained changes in bandwidth. The original measurement data and the calculated second-level bandwidth difference are concatenated to form a multi-feature vector. The combined feature vector is then normalized to eliminate the dimensional differences between features of different dimensions.

[0037] The long-term trend prediction module utilizes a bidirectional long short-term memory network model based on an attention mechanism to process multivariate feature vectors. Through forward and backward LSTM networks, it extracts the temporal features of bandwidth over long time scales, and combines these with attention weight factors to automatically identify key historical moments and calculate the predicted bandwidth change trend over long time scales.

[0038] The network status determination module calculates a comprehensive network status quantification value based on the original measurement data, compares the calculated quantification value with a preset threshold, and classifies the current network operating status into three states in real time: light load, full load, or congestion.

[0039] The short-term incremental adaptive estimation module uses the macroscopic prediction value output by the long-term trend prediction module as the initial benchmark, and combines it with the current state output by the network state determination module. It then uses the incremental adaptive AIMD improved algorithm to adjust the bandwidth. Based on the measured bandwidth at time t, the bandwidth threshold at the time of the last congestion, and the average bandwidth of the previous n times, it dynamically calculates the incremental value and outputs the final bandwidth estimate.

[0040] The beneficial effects of this invention are:

[0041] 1. In this invention, by introducing a bidirectional long short-term memory network based on an attention mechanism, it is possible not only to extract temporal features from both forward and backward directions, but also to automatically identify key historical moments using attention weight factors. This solves the problem of insufficient capture of long-term dependence factors of high-dynamic links in the prior art, enabling the system to more accurately grasp the macroscopic change law of network bandwidth over a long time scale and avoid estimation deviations caused by ignoring historical trends.

[0042] 2. This invention employs multi-feature construction, introduces second-level bandwidth difference, and combines it with an incremental adaptive AIMD improved algorithm. It detects fine-grained changes in bandwidth through second-level difference and dynamically adjusts the bandwidth increment based on the current network status. This enables the system to react quickly to drastic fluctuations in the network environment, achieving high-precision real-time estimation.

[0043] 3. In this invention, during the feature construction stage, the original measurement data and the calculated second-level bandwidth difference are spliced ​​and normalized, overcoming the defect of inaccurate estimation of single-dimensional data. This multi-dimensional feature vector contains both the current network state and the rate of change of bandwidth, providing a more comprehensive data foundation for subsequent accurate prediction. Attached Figure Description

[0044] Figure 1 This is a schematic diagram of a method flow of the present invention.

[0045] Figure 2 This is a schematic diagram of step S2 of the present invention.

[0046] Figure 3 This is a schematic diagram of step S3 of the present invention.

[0047] Figure 4 This is a schematic diagram of step S4 of the present invention.

[0048] Figure 5 This is a schematic diagram of step S5 of the present invention.

[0049] Figure 6 This is a schematic diagram of a system structure according to the present invention. Detailed Implementation

[0050] To more clearly illustrate the technical solutions in the embodiments of the invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0051] like Figures 1 to 5 As shown, a method for estimating air-to-ground interconnect bandwidth under highly dynamic links includes the following:

[0052] Using current bandwidth, packet loss rate, latency jitter, and second-level bandwidth difference (bdps) as data sources, and addressing the heterogeneous characteristics of these metrics—including significant fluctuations, strong nonlinear dependencies, and frequent short-term mutations—in highly dynamic link environments, a complete closed-loop processing flow from raw bandwidth data acquisition to unified bandwidth estimation output is implemented. This solution is suitable for video transmission and real-time communication applications in highly dynamic mobile environments such as high-speed rail, highways, and subways, and is particularly suitable for bandwidth prediction in highly dynamic link scenarios with frequent base station switching and unstable signal quality. The estimation results can be directly output to the bitrate decision module or network congestion control module of an adaptive video transmission system, supporting adaptive video bitrate adjustment and transmission strategy optimization. The algorithm implementation steps are as follows:

[0053] Step S1: Raw bandwidth data acquisition.

[0054] Raw measurement data of the network link is collected at a fixed sampling period of 1 second. The collected data includes three categories: the first is the current bandwidth value b(t), representing the available network bandwidth at the current moment, in Mbps; the second is the packet loss rate λ, representing the proportion of data packets lost per unit time, in percentage; and the third is the delay jitter J, representing the variance of the transmission delay between adjacent data packets, in milliseconds. The collected data is stored in chronological order as a sliding window sequence of length L, where L is 30. When new data arrives, the oldest data in the window is removed, while maintaining the same window length.

[0055] Step S2: Construction of multivariate feature vectors.

[0056] A second-level bandwidth difference (bpds) feature is constructed and a multivariate feature vector is formed. The bpds feature is used to characterize the rapid changes in bandwidth; when bdps(t) approaches 0, it indicates stable bandwidth, and when bdps(t) is large, it indicates drastic bandwidth changes. Then, the data obtained from the original measurement in step 1 is combined with the bpds feature to form a multivariate feature vector X(t), which is a vector composed of four elements: b(t), λ, J, and bdps(t). Finally, each feature dimension is normalized so that all feature values ​​fall within the range of 0 to 1. The multivariate time series based on historical bandwidth is defined as... ,in This represents a feature vector containing m pieces of historical information, including measurement bandwidth b(t) and packet loss rate. The latency jitter J and the second-level bandwidth difference bdps(t) are defined as follows:

[0057] bdps(t) = |b(t) - b(t-1)|,

[0058] In the formula, b(t) is the current measured bandwidth value, and b(t-1) is the bandwidth measured value at the previous moment.

[0059] The fine-grained change in bandwidth is then detected by comparing the second-level bandwidth difference with the bandwidth measurement value from the previous moment, expressed as: .

[0060] Step S3: Long-term bandwidth trend prediction.

[0061] The feature vector X(t) obtained from historical data is weighted and fused to form a single fused value xt, which serves as the input vector to the LSTM network. Then, in the hidden layer, the weight vector and the input vector are combined to generate a new vector. The output vector is then fully connected to obtain the prediction result. A BiLSTM model, incorporating both forward and backward LSTM networks, is employed to improve bandwidth prediction accuracy and extract bandwidth features more comprehensively. The BiLSTM model utilizes two independent LSTM networks to process the fused values ​​of multiple types of time-series data from the forward and backward directions, respectively. A one-way LSTM network includes a forget gate, an input gate, and an output gate, and a hidden vector. The update is as follows:

[0062] ,

[0063] In the formula, For the Gate of Oblivion For output gate, and Input gate, For cell vectors, The previous hidden vector The weight matrix, Indicates fusion value The weight matrix, It is the Sigmoid activation function. Forgotten Gate The bias term, For input gate The bias term, For input gate The bias term, For output gate The bias term.

[0064] Based on the above formulas, the forward hidden vectors are obtained respectively. and reverse hidden vector The final hidden vector is calculated using weighted averages. :

[0065] ,

[0066] The attention mechanism transforms the original random weight allocation into a probability allocation, and uses the hidden vectors obtained from the BiLSTM model as input to the attention mechanism's processing stage. This is combined with attention weight factors. Capture key temporal features affecting bandwidth, and calculate bandwidth change feature values ​​based on n prediction steps (default unit is 1 second). :

[0067] , ,

[0068] ,

[0069] In the formula, Represents the hidden vector The weight matrix, This represents the attention weight.

[0070] Finally, the bandwidth change trend prediction value is calculated and output by the fully connected layer. This value is the long-term bandwidth prediction value, and also serves as the initial value for bandwidth estimation on a short-term time scale. It is calculated as follows:

[0071] .

[0072] In the formula, It is a multivariate prediction function. To predict the step size, For multivariate time series based on historical bandwidth, Let be a feature vector containing m kinds of historical information.

[0073] Step S4: Calculate network congestion status.

[0074] Overall packet loss rate and delay jitter The calculated network state quantization value *r* is used to classify the network into three states: light load, full load, and congestion, based on two predefined thresholds. The estimated bandwidth is adjusted according to each state. When the network is lightly loaded, the estimated bandwidth is increased by an additive factor; when the network is congested, the estimated bandwidth is decreased by a multiplicative factor; and when the network is full load, the estimated bandwidth remains unchanged. The network state quantization value *r* is set with parameters based on the impact of packet loss rate and latency jitter. The specific calculations are as follows:

[0075] ,

[0076] in, and These represent the maximum packet loss rate and the maximum latency jitter, respectively.

[0077] Let r1 and r2 be the thresholds for the network state transitioning from idle to normal and from normal to congested, respectively. Then, the current network state... The description is as follows:

[0078]

[0079] Here, 1, 0, and -1 represent network states such as light load, full load, and congestion, respectively.

[0080] Step S5: Short-timescale bandwidth estimation. The short-timescale network state partitioning module calculates the final bandwidth value within a short timescale. The long-timescale bandwidth prediction value obtained in Step 3 is then used. Using the network state Q(r) obtained in step four as the initial input, the improved AIMD algorithm based on incremental adaptation is used to adjust the bandwidth, and finally outputs an accurate bandwidth estimate. .

[0081] The incremental adaptive AIMD improved algorithm considers two different network bandwidth increment values ​​for two different networks. One case is where the network has not experienced congestion, in which the addition factor is taken. The estimated bandwidth is increased gradually to smoothly approximate the actual network bandwidth, avoiding frequent network congestion; when network congestion occurs, a multiplication factor is applied. This causes the estimated bandwidth to decrease rapidly, responding promptly to changes in actual bandwidth until the network link returns to normal. Specifically, when the network is not congested, i.e., Q(r) is 0 or 1, the estimated bandwidth... The description is as follows:

[0082] ,

[0083] in, This represents the initial increment value.

[0084] Another scenario is when network congestion has already occurred. Setting the estimated bandwidth too high can cause fluctuations in the packet transmission rate and potentially trigger further congestion. Setting it too low results in slow growth of the transmission rate under light network load conditions, which is detrimental to real-time network bandwidth detection and improved bandwidth resource utilization. Therefore, the estimated bandwidth should be determined based on real-time changes in network status, i.e., light load, full load, and critical congestion. The adjustments are as follows:

[0085] ,

[0086] in, Indicates an incremental adaptive value. This represents the bandwidth threshold for the most recent near-congestion state.

[0087] To avoid frequent changes in network status caused by excessively rapid bandwidth adjustments, an incremental adaptive value is calculated. We need to consider the measured bandwidth b(t) at time t and the bandwidth during the last transmission congestion. and the average bandwidth of the first n time steps :

[0088] ,

[0089] Based on the above formula, bandwidth is estimated under two different scenarios: before and after network congestion, and the final bandwidth estimate is output. .

[0090] like Figure 5 As shown, a high-dynamic link air-to-ground interconnection bandwidth estimation system adopts the high-dynamic link air-to-ground interconnection bandwidth estimation method, including a raw data acquisition module, a multi-feature construction module, a long-term trend prediction module, a network status determination module, and a short-term incremental adaptive estimation module.

[0091] The raw data acquisition module is used to collect raw measurement data of the network link in real time at a fixed sampling period. It collects the bandwidth value representing the current available network bandwidth, the packet loss rate representing the proportion of data packets lost per unit time, and the delay jitter representing the variance of the transmission delay of adjacent data packets.

[0092] The multi-feature construction module is used to calculate the second-level bandwidth difference based on the original measurement data, thereby detecting fine-grained changes in bandwidth. The original measurement data and the calculated second-level bandwidth difference are concatenated to form a multi-feature vector. The combined feature vector is then normalized to eliminate the dimensional differences between features of different dimensions.

[0093] The long-term trend prediction module utilizes a bidirectional long short-term memory network model based on an attention mechanism to process multivariate feature vectors. Through forward and backward LSTM networks, it extracts the temporal features of bandwidth over long time scales, and combines these with attention weight factors to automatically identify key historical moments and calculate the predicted bandwidth change trend over long time scales.

[0094] The network status determination module calculates a comprehensive network status quantification value based on the original measurement data, compares the calculated quantification value with a preset threshold, and classifies the current network operating status into three states in real time: light load, full load, or congestion.

[0095] The short-term incremental adaptive estimation module uses the macroscopic prediction value output by the long-term trend prediction module as the initial benchmark, and combines it with the current state output by the network state determination module. It then uses the incremental adaptive AIMD improved algorithm to adjust the bandwidth. Based on the measured bandwidth at time t, the bandwidth threshold at the time of the last congestion, and the average bandwidth of the previous n times, it dynamically calculates the incremental value and outputs the final bandwidth estimate.

[0096] This application utilizes an attention mechanism to fuse a bidirectional long short-term memory network, significantly improving the ability to accurately capture macroscopic bandwidth trends over long time scales. Traditional methods often struggle to handle the nonlinear dependencies of multidimensional features such as packet loss rate and latency jitter in highly dynamic scenarios. This algorithm, by constructing a bidirectional propagation structure, not only extracts historical bandwidth data forward but also mines future contextual information backward, achieving comprehensive learning of temporal features. Simultaneously, the introduction of an attention mechanism automatically identifies and assigns higher weights to key historical moments, filtering out short-term interference, thereby accurately predicting bandwidth evolution trends over long periods and providing reliable initial benchmark values ​​for subsequent transmission control.

[0097] A short-time estimation model based on incremental adaptive bandwidth was established, achieving millisecond-level rapid response and smooth approximation to sudden bandwidth changes. Addressing the "cliff-like" drop or "step-like" rise in bandwidth caused by frequent base station handovers in highly dynamic links, this algorithm abandons the fixed-parameter prediction mode and instead adopts a dynamic adjustment strategy similar to "steady rise and rapid fall" (AIMD): under light or full network load conditions, an additive factor is used to gradually increase the estimated bandwidth, avoiding frequent attempts that could lead to congestion; once congestion is detected (based on quantized thresholds of packet loss rate and latency jitter), a multiplicative factor is immediately switched to rapidly decrease the estimated bandwidth, promptly matching the actual available bandwidth. This mechanism can also dynamically calculate the incremental adaptive value based on the most recent congestion threshold and average bandwidth, effectively preventing repeated oscillations of the estimated value under critical conditions and ensuring high sensitivity and stability on a short-time scale.

[0098] By organically integrating long-term trend prediction with short-term adaptive correction, robust estimation of all states (light load, full load, congestion) in complex scenarios is achieved. This algorithm does not view long-term or short-term features in isolation, but rather uses the trend prediction value output from the long-term scale as the initial input for short-term scale adjustments, forming a closed-loop control of "macro guiding micro, micro correcting macro." This dual-scale strategy overcomes the shortcomings of traditional statistical methods in handling non-stationary sequences and solves the problem of slow response of single deep learning models under extreme fluctuations. As a result, the algorithm demonstrates superior overall performance compared to existing typical algorithms on real high-dynamic datasets such as high-speed rail and air-to-ground interconnection, both in terms of single-step / multi-step prediction accuracy and video experience quality (QoE) assurance in scenarios with sudden bandwidth changes.

[0099] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions or improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A high dynamic link under-bandwidth estimation method, characterized in that, Includes the following steps: Step S1: Raw broadband data acquisition, acquiring raw measurement data of the network link at a fixed sampling period; Step S2: Multivariate feature vector construction. Based on the original measurement data, calculate the second-level bandwidth difference, combine the original measurement data to form a multivariate feature vector, and detect fine-grained bandwidth changes. Step S3: Long-term trend prediction. Input the multivariate feature vector into the bidirectional long short-term memory network model based on the attention mechanism, use the forward and backward LSTM networks to extract the bandwidth temporal features, and combine the attention weight factor to calculate the predicted value of bandwidth change trend over a long time scale. Step S4: Network status determination. Calculate the network status quantification value based on the original measurement data, and classify the current network status into light load, full load, or congestion state according to the preset threshold. Step S5: Short-term incremental adaptive estimation: Using the long-term trend prediction value as the initial value, and combining it with the network state, the bandwidth is adjusted using the incremental adaptive AIMD improved algorithm, and the final bandwidth estimation value is output.

2. The method of claim 1, wherein, The raw measurement data in step S1 includes the current bandwidth value representing the currently available network bandwidth, the packet loss rate representing the proportion of data packets lost per unit time, and the delay jitter representing the variance of the transmission delay of adjacent data packets.

3. The method of claim 2, wherein, Step S2 is implemented in the following manner: Step S21: Calculate the second-level bandwidth difference. Based on the collected raw measurement data, calculate the bandwidth difference between the current time and the previous time. Step S22: Combine the multivariate feature vectors by concatenating the original measurement data with the second-level bandwidth difference to form a multivariate feature vector; Step S23: Feature normalization processing, normalize each feature dimension in the combined feature vector.

4. The method of claim 3, wherein, In step S21, the second-level bandwidth difference bdps(t) is calculated in the following way: bdps(t) = |b(t) - b(t-1)|, In the formula, b(t) is the current measured bandwidth value, and b(t-1) is the bandwidth measured value at the previous moment; The fine-grained change in bandwidth is detected by comparing the second-level bandwidth difference with the bandwidth measurement value at the previous moment.

5. The method of claim 1, wherein, Step S3 is implemented in the following manner: Step S31: Feature vector fusion and input. The multivariate feature vectors are weighted and fused to generate a single fused value, which is used as the input vector of the bidirectional long short-term memory network model. Step S32: Bidirectional temporal feature extraction. Feature extraction is performed using a bidirectional long short-term memory network. The cell state and hidden state are updated through the forget gate, input gate, and output gate. The forward hidden vector and the backward hidden vector are calculated respectively, and the final hidden vector is calculated by weighting. Step S33: Attention weighting and trend prediction. The hidden vector output by the bidirectional long short-term memory network model is used as input. Combined with the attention weight factor, key historical moments are automatically identified and given higher weights. Based on the calculated attention weights and hidden vectors, the bandwidth change trend prediction value over a long time scale is calculated and output through a fully connected layer.

6. The method for estimating air-to-ground interconnection bandwidth under high dynamic link conditions according to claim 5, characterized in that, The bandwidth change trend prediction value in the step S33 is calculated by ， wherein, is a multi-variate prediction function, is a prediction step value, is a multi-variate time series based on historical bandwidths, is a feature vector with m kinds of historical information.

7. The method of claim 1, wherein, Step S4 is implemented in the following manner: Step S41: Calculate the network state quantization value. Based on the packet loss rate and latency jitter, a comprehensive quantization value is obtained through weighted calculation. Step S42: Determine the network state category. Based on the calculated quantization value, compare it with a predefined threshold to classify the current network state into light load, full load, and congestion state.

8. The method of claim 1, wherein, Step S5 is implemented in the following manner: Step S51: Initialize the input by using the macroscopic trend prediction value output from the long-term bandwidth trend prediction as the initial baseline value for the short-term estimation. Step S52: Strategy selection, switching and adjusting strategies in real time according to different network conditions; Step S53: Incremental adaptive calculation and output. Based on the measured bandwidth at time t, the bandwidth threshold at the time of the last congestion, and the average bandwidth of the previous n times, the incremental value is dynamically calculated, and the final bandwidth estimate is output.

9. A system for estimating the bandwidth of a high dynamic link under space-ground interconnection, which adopts the method for estimating the bandwidth of a high dynamic link under space-ground interconnection according to any one of claims 1-8. It includes a raw data acquisition module, a multi-feature construction module, a long-term trend prediction module, a network state determination module, and a short-term incremental adaptive estimation module; The raw data acquisition module is used to collect raw measurement data of the network link in real time at a fixed sampling period. It collects the bandwidth value representing the current available network bandwidth, the packet loss rate representing the proportion of data packets lost per unit time, and the delay jitter representing the variance of the transmission delay of adjacent data packets. The multi-feature construction module is used to calculate the second-level bandwidth difference based on the original measurement data, thereby detecting fine-grained changes in bandwidth. The original measurement data and the calculated second-level bandwidth difference are concatenated to form a multi-feature vector. The combined feature vector is then normalized to eliminate the dimensional differences between features of different dimensions. The long-term trend prediction module is used to process multivariate feature vectors using a bidirectional long short-term memory network model based on an attention mechanism. Through forward and backward LSTM networks, it extracts the temporal features of bandwidth over a long time scale. Combined with attention weight factors, it automatically identifies key historical moments and calculates the predicted value of bandwidth change trend over a long time scale. The network status determination module calculates a comprehensive network status quantification value based on the original measurement data, compares the calculated quantification value with a preset threshold, and classifies the current network operating status into three states in real time: light load, full load, or congestion. The short-term incremental adaptive estimation module uses the macroscopic prediction value output by the long-term trend prediction module as the initial benchmark, and combines it with the current state output by the network state determination module. It then uses the incremental adaptive AIMD improved algorithm to adjust the bandwidth. Based on the measured bandwidth at time t, the bandwidth threshold at the time of the last congestion, and the average bandwidth of the previous n times, it dynamically calculates the incremental value and outputs the final bandwidth estimate.

Images