Satellite internet traffic QoS classification method based on deep learning

Abstract

The invention discloses a satellite internet traffic QoS classification method based on deep learning, and belongs to the technical field of satellite communication networks. The method comprises the steps of designing a'data preprocessing-pre-training-fine tuning 'full-process optimization mechanism by adapting to long time delay, high bit error rate and MTU fragmentation characteristics of a satellite link, realizing high-precision prediction of five types of QoS requirements (high bandwidth and low delay, middle bandwidth and middle delay, low bandwidth and low delay, high reliability and unknown types) in an environment with limited satellite terminal resources, and improving the quality of the satellite terminal. The problems that an existing model is poor in migration performance, large in reasoning overhead and high in labeling dependence are solved, and on-satellite resource dynamic allocation and QoS routing optimization are directly supported.

Description

Technical Field

[0001] This invention relates to the field of satellite communication network technology, specifically to a deep learning-based method for classifying QoS requirements of satellite internet traffic. This method is specifically designed for scenarios characterized by long latency, high bit error rate, MTU limitations, and limited terminal resources in satellite links. Through lightweight pre-trained models and scenario adaptation techniques, it achieves high-precision, low-overhead prediction of QoS requirements for encrypted traffic, providing key technical support for dynamic allocation of satellite network resources and QoS routing optimization. Background Technology

[0002] With the accelerated construction of integrated space-air-ground networks, low Earth orbit (LEO) and medium Earth orbit (MEO) satellite internet has become a core support for global communication coverage, with encrypted traffic accounting for over 85% of satellite networks (such as TLS 1.3 encrypted video calls and cloud service transmissions). Accurate QoS requirement classification of traffic is a core prerequisite for achieving dynamic bandwidth allocation, link priority adjustment, and quality of service assurance in satellite networks; however, its implementation faces unique challenges in satellite scenarios.

[0003] Existing technologies face three major bottlenecks in classifying QoS requirements for satellite traffic: Insufficient scenario adaptability: Current traffic QoS classification methods mainly rely on the design of terrestrial network characteristics and have not yet been designed for satellite network traffic. They cannot cope with the sparsity of traffic characteristics and timing distortion caused by long latency (200-600ms) and MTU fragmentation (1440 bytes) of satellite links.

[0004] Excessive model resource consumption: The mainstream classification methods that use pre-trained models (such as ET-BERT and NetMamba) generally have more than 100M parameters. Satellite terminals have limited computing and storage resources, which cannot bear the inference overhead and cannot meet the real-time scheduling requirements.

[0005] Data dependency and semantic gap: Satellite traffic QoS annotation requires manual analysis of link characteristics, which is costly and has a limited sample size. Traditional supervised learning models are prone to underfitting. At the same time, existing pre-training tasks are not adapted to the structural patterns of satellite traffic, making it difficult for models to capture QoS-related features.

[0006] Furthermore, publicly available datasets (such as ISCX-VPN and Mobile-App) are all built on terrestrial Ethernet and do not include satellite link features. Direct transfer of these datasets can lead to "feature learning bias." Therefore, a satellite internet traffic QoS classification scheme that balances lightweight design, scenario adaptability, and high accuracy is currently needed. Summary of the Invention

[0007] The purpose of this invention is to provide an efficient deep learning-based method for QoS classification of satellite internet traffic.

[0008] To address the aforementioned technical problems, this invention provides a deep learning-based satellite internet traffic QoS classification method, comprising the following steps: Obtain encrypted satellite traffic; Preprocess the encrypted satellite traffic to obtain preprocessed encrypted satellite traffic; Obtain the pre-trained DD-DistilBERT model; The preprocessed encrypted satellite traffic is input into the pre-trained DD-DistilBERT model for fine-tuning to obtain the optimal fine-tuned model. Obtain the encrypted traffic of the satellite to be predicted; The encrypted traffic of the satellite to be predicted is input into the optimal fine-tuning model to obtain the prediction results.

[0009] Preferably, the satellite encrypted traffic is preprocessed to obtain preprocessed satellite encrypted traffic, specifically including the following steps: The encrypted satellite traffic is segmented into several independent traffic streams; The independent flow streams are feature-processed to obtain the target independent flow stream; Feature extraction and normalization are performed on the target independent traffic flow to obtain the preprocessed satellite encrypted traffic.

[0010] Preferably, the encrypted satellite traffic is segmented into several independent traffic streams, specifically including the following steps: The satellite encrypted traffic is split into several independent traffic streams by dividing it into quintuples.

[0011] Preferably, the independent traffic streams are subjected to feature processing to obtain the target independent traffic stream, specifically including the following steps: Merge consecutive independent traffic streams that are less than a first preset byte threshold; Remove independent traffic streams whose length is less than the second preset byte threshold; The first few independent traffic flows are retained as the target independent traffic flow.

[0012] Preferably, feature extraction and normalization are performed on the target independent traffic stream to obtain preprocessed satellite encrypted traffic, specifically including the following steps: Extract the loads of all target independent flow streams, convert them into a list of decimal integers, and form a load feature matrix; Calculate the length of all independent target traffic flows to form packet length sequence characteristics; Flatten the load feature matrix and concatenate it with the packet length sequence features to obtain the original feature vector; The eigenvalues ​​of the original feature vector are scaled to the [0,1] interval by normalization and used as the preprocessed satellite encrypted traffic.

[0013] Preferably, the method further includes the following steps: QoS labels are applied to preprocessed encrypted satellite traffic.

[0014] Preferably, the pre-training of the DD-DistilBERT model specifically includes the following steps: TU generation: The SWOBG algorithm is used to process unlabeled satellite traffic samples, with a window duration of 100ms and an overlap ratio of 0.25. Each traffic stream is divided into 5-8 transmission units; adjacent TUs overlap by 25% of their data packets. TU Encoding: TU is encoded using the Tu2Token module: first, the hexadecimal byte sequence of TU is split into two parts according to a binary model, then byte-pair encoding is used to generate a token; a token dictionary is constructed. Position division: Divide the TU token sequence into head, middle and tail according to the ratio of 15%:70%:15%; Differentiated masking strategy: Set the probability of masking the header token to 20% and the tail token to 10%; replace 80% of the selected tokens with [MASK], 10% with a random token, and leave 10% unchanged; Pre-training parameter settings: batch size = 16, learning rate = 2e-5, dropout rate = 0.5, number of iterations = 1000; start training in the P100 16GGPU environment, monitor the model's prediction accuracy for the mask token in real time, stop training when the accuracy does not improve for 10 consecutive rounds, and save the pre-trained model.

[0015] Preferably, the preprocessed encrypted satellite traffic is input into the pre-trained DD-DistilBERT model for fine-tuning to obtain the optimal fine-tuned model. This process includes the following steps: Fine-tuning data preparation: Select preprocessed satellite encrypted traffic labeled with QoS tags that meet the requirements of satellite scene adaptation, and divide it into training set, validation set and test set. Introduce heterogeneous samples into the training set. Model structure adjustment: Load the pre-trained DD-DistilBERT model, delete the original output layer and replace it with a QoS classification layer, and use softmax as the activation function; retain the parameters of the bottom Transformer encoder, and update the parameters of the classification layer and the upper encoder. Fine-tuning parameter settings: batch size = 16, learning rate = 1e-5, number of iterations = 500, loss function is cross-entropy; start training on the training set, and evaluate the model's F1 score using the validation set after each iteration; Early stopping mechanism trigger: When the F1 score on the validation set does not improve for 10 consecutive rounds, the early stopping mechanism is triggered, training is stopped and the optimal fine-tuned model is saved.

[0016] Preferably, the encrypted traffic of the satellite to be predicted is input into the optimal fine-tuning model to obtain the prediction result, specifically including the following steps: The encrypted traffic of the satellite to be predicted is generated into TUs according to the SWOBG algorithm, and each TU is encoded into a token sequence using Tu2Token. The encoded token sequence is input into the optimal fine-tuning model, which extracts features through the bottom encoder, calculates probabilities through the upper classification layer, and outputs a probability distribution. The category with the highest probability in the probability distribution is selected as the final QoS requirement prediction result.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. Deep adaptation to satellite scenarios, resulting in more stable classification accuracy: Existing technologies are mostly designed based on terrestrial network characteristics, without considering the long latency (200-600ms), MTU fragmentation (1440 bytes), and feature sparsity of satellite links. When migrated to satellite scenarios, the classification accuracy drops significantly. This invention uses a specially designed SWOBG sliding window overlap mechanism (0.25 overlap ratio) and PMBM location-differentiated masking strategy (15:70:15 location division + layered masking ratio) to accurately capture the "protocol interaction mode - timing jitter - fragmentation pattern" of satellite traffic. The QoS classification F1 value reaches 86.61%, solving the core pain point of "poor adaptability of terrestrial models to satellite migration".

[0018] 2. Lightweight architecture design for more user-friendly terminal deployment: Existing mainstream pre-trained models have large parameter counts and high inference overhead, which cannot adapt to the resource constraints of satellite terminals. This invention optimizes the model through DistilBERT knowledge distillation, compressing the number of model parameters to 66M (40% less than BERT-base) and occupying only 256MB of storage. Combined with INT8 quantization and parallel computing optimization, it reduces the inference time per sample and the consumption of computing power, achieving an optimal balance between "lightweight" and "high precision". It can be deployed on satellite terminals and meets scheduling requirements.

[0019] 3. Stronger generalization ability and better cross-scenario adaptability: Existing technologies are mostly limited to a single dataset or specific scenario, and their generalization ability is weak when facing heterogeneous traffic (such as ground apps transmitting through satellite links). This invention introduces ground-fused heterogeneous samples during the fine-tuning stage, and verifies them on three types of public datasets: ISCX-VPN-Service, Mobile-App, and CrossPlatform (iOS). The average F1 score reaches 93.04%, which can adapt to five types of core QoS requirements and typical satellite applications. It can cover both pure satellite scenario services and support the QoS classification requirements of satellite-ground fusion networks, making it more widely applicable. Attached Figure Description

[0020] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0021] Figure 1 This is a schematic diagram of the DD-DistilBERT module pre-training process, showing the key steps of SWOBG generation of TU, ​​Tu2Token encoding, and PMBM mask training.

[0022] Figure 2 The graph shows the experimental results of SWOBG overlap ratio. The horizontal axis represents the overlap ratio (0.05~0.45), and the vertical axis represents the QoS prediction score. It shows that the performance is optimal when the overlap ratio is 0.25.

[0023] Figure 3 The graph shows the results of the token location partitioning experiment. The horizontal axis represents the token location partitioning ratio, and the vertical axis represents the QoS prediction score. It shows that the performance is optimal when the location partitioning ratio is 15:70:15.

[0024] Figure 4 A comparison of various performance metrics of the DD-DistilBERT model on constructed satellite traffic data.

[0025] Figure 5 The experimental results of the DD-DistilBERT model on different datasets show that it has strong generalization ability and excellent overall performance metrics. Detailed Implementation

[0026] Numerous specific details are set forth in the following description to provide a full understanding of the invention. However, the invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0027] The terminology used in one or more embodiments of this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the one or more embodiments of this specification. The singular forms “a,” “described,” and “the” as used in one or more embodiments of this specification and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in one or more embodiments of this specification refers to and includes any or all possible combinations of one or more associated listed items.

[0028] It should be understood that although the terms first, second, etc., may be used to describe various information in one or more embodiments of this specification, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, first may also be referred to as second without departing from the scope of one or more embodiments of this specification, and similarly, second may also be referred to as first. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to a determination."

[0029] The present invention will now be described in further detail with reference to the accompanying drawings: This invention provides a deep learning-based satellite internet traffic QoS classification method, comprising the following steps: Step 1: Preprocessing of encrypted satellite traffic.

[0030] 1.1 Session Segmentation: Process the original PCAP file and split it into N independent traffic streams according to the five-tuple, ensuring that each stream corresponds to a single business scenario; 1.2 Feature Cleaning and MTU Adaptation: Merge consecutive small data packets to the satellite MTU specification, remove invalid fragment packets and records exceeding the interval, and retain the core valid packets of each flow; 1.3 Two-dimensional feature extraction: Extract the payload data and packet length sequence of the core effective packets, construct the original feature vector, and normalize it to the specified interval; 1.4 QoS Labeling: Combining satellite link latency and jitter characteristics, the QoS category of each preprocessed sample is manually labeled to construct a satellite traffic dataset.

[0031] Step 2: Pre-training the DD-DistilBERT model.

[0032] 2.1 TU Generation and Encoding: Select unlabeled satellite traffic samples, generate transmission units (TUs) through a sliding window overlap generator (SWOBG), and then encode the TUs into token sequences through the Tu2Token module to build a token dictionary adapted to satellite traffic; 2.2PMBM pre-training task execution: Divide the token sequence of TU into head, middle and tail according to the proportion, perform differential masking operation, train the model to predict mask tokens based on context, set pre-training parameters to start training, monitor the training effect and stop when appropriate; 2.3 Validation of pre-trained model: Validate the ability of the pre-trained model to extract the structural features of satellite traffic on unlabeled samples to ensure that the model can capture the key patterns of satellite traffic.

[0033] Step 3: Fine-tuning of the DD-DistilBERT model and QoS prediction inference.

[0034] 3.1 Fine-tuning data preparation: The labeled satellite traffic dataset is divided into training set, validation set and test set according to the proportion. All samples are processed by the preprocessing process in step 1. 3.2 Model Structure Adjustment and Training: Load the pre-trained model, delete the original output layer and replace it with a 5-class QoS classification layer, set the fine-tuning parameters and start training, trigger the early stop mechanism by monitoring the performance of the validation set, and save the optimal fine-tuned model; 3.3 QoS Prediction Inference: The traffic flow to be predicted is generated into a TU according to SWOBG and encoded as the model input. The probability distribution of 5 types of QoS is output through the fine-tuned model to determine the final QoS demand prediction result.

[0035] The core of this invention is to construct a lightweight satellite QoS classification model based on DD-DistilBERT. Through a four-layer technical architecture of "satellite scene adaptation preprocessing + sliding window overlap generator (SWOBG) + location mask pre-training (PMBM) + small sample fine-tuning", the QoS requirement classification achieves triple optimization of "accuracy-efficiency-scene adaptation".

[0036] Specifically, the main inventive contents of this invention include: 1. DD-DistilBERT Lightweight QoS Classification Model Architecture: Addressing the limitations of satellite terminal computing resources and the sparsity of satellite traffic features, this invention constructs a core architecture integrating "Tu2Token encoding + dual-dimensional feature input + knowledge distillation optimization." Based on DistilBERT, the number of Transformer layers is compressed from 12 to 6 through knowledge distillation technology, with the number of parameters controlled to 66M (a 40% reduction compared to BERT-base), and model storage usage of 256MB. Simultaneously, it incorporates dual-dimensional features of "payload matrix + packet length sequence," and uses the Tu2Token module to convert satellite traffic into language-like tokens, solving the problems of invisible and sparsity of encrypted traffic features, reducing single-sample inference time, and adapting to the deployment requirements of satellite terminals.

[0037] 2. Pre-training strategy for satellite scene adaptation: To address the problem of insufficient model generalization caused by scarce labeled data, this invention constructs a collaborative mechanism of "Tu2Token encoding + sliding window overlap generation (SWOBG) + location mask pre-training (PMBM)". SWOBG generates Transmission Units (TUs) with a 100ms window and a 0.25 overlap ratio, preserving traffic context information; PMBM divides the TU's token sequence into header, middle, and tail sections in a 15:70:15 ratio, performing differential masking (20% for header, 15% for middle, and 10% for tail) to force the model to capture the "protocol interaction mode - timing jitter - fragmentation pattern" of satellite traffic; Tu2Token generates tokens through binary model splitting and byte-pair encoding (BPE), constructing a 65536-byte dictionary to adapt to the byte-level feature representation of encrypted traffic.

[0038] 3. Satellite-adapted three-level QoS classification process of "pre-training-fine-tuning-inference optimization": As the implementation carrier of the aforementioned DD-DistilBERT lightweight architecture and satellite pre-training strategy, this process focuses on the three major pain points of satellite scenarios: "scarcity of labeled samples, difficulty in cross-scenario deployment, and high terminal inference overhead". It achieves the engineering adaptation of technical solutions through three-level collaborative design.

[0039] This invention adopts a three-layer collaborative design of "infrastructure-optimization strategy-implementation process" to form a lightweight traffic QoS requirement classification method adapted to satellite scenarios. The three layers are progressive, from "technical architecture construction" to "strategy optimization" and then to "process implementation", forming a closed loop. This ensures high accuracy in satellite Internet traffic QoS requirement classification and meets the requirements of lightweight terminals and scenario adaptation.

[0040] To better illustrate the technical effects of the present invention, the present invention provides the following specific embodiments to illustrate the above technical process: Example 1: A deep learning-based satellite internet traffic QoS classification method, comprising the following steps: Step 1: Preprocessing of encrypted satellite traffic. This step cleans and extracts features from the raw traffic, taking into account the link characteristics of encrypted satellite traffic, to provide input data adapted to the satellite scenario for subsequent model training and inference.

[0041] Step 1-1: Session segmentation to obtain independent traffic streams.

[0042] The SplitCap tool is used to split the PCAP file into N independent traffic streams by the five-tuple of "source IP address, source port number, destination IP address, destination port number, and transport layer protocol", ensuring the independence of each traffic stream.

[0043] Step 1-2: Feature processing to obtain the target independent flow.

[0044] Merging process: Consecutive packets <500 bytes are merged into 1440-byte packets to adapt to the satellite MTU fragmentation pattern and reduce feature sparsity caused by too many small packets; Invalid data removal: Remove invalid fragment packets with a length of <64 bytes (mostly link noise), and delete records with a data packet arrival time interval >1s (to avoid abnormal interference caused by link interruption); Core feature retention: Only the first 20 valid packets of each session flow are retained. Experiments have verified that this operation can cover more than 85% of QoS key features, while avoiding subsequent computational redundancy.

[0045] Steps 1-3: Feature extraction and normalization.

[0046] A two-dimensional feature extraction strategy is adopted to construct a feature representation adapted to encrypted traffic: Payload feature extraction: Extract the 32-byte payload of the first 20 packets of each stream, convert it into a list of decimal integers, and form a 20×32 payload feature matrix (covering the byte-level potential features of the encrypted traffic). Packet length sequence extraction: Calculate the length (in bytes) of the first 20 packets of each flow to form a 1×20 packet length sequence feature (reflecting the transmission rhythm of traffic and its association with service type). Feature integration and normalization: The payload feature matrix is ​​flattened (dimension converted to 640) and concatenated with the packet length sequence features (dimension 20) to obtain a 660-dimensional original feature vector; the feature values ​​are scaled to the [0,1] interval by Min-Max normalization to eliminate the influence of dimensional differences on model training.

[0047] Steps 1-4: QoS labeling.

[0048] By combining satellite link latency and jitter characteristics, the QoS category of each preprocessed sample is manually labeled to construct a satellite traffic dataset.

[0049] Step 2: DD-DistilBERT module pre-training, fine-tuning, and QoS requirement prediction. The DD-DistilBERT module is based on DistilBERT optimization (see attached). Figure 1 As shown, QoS requirement prediction with a small amount of labeled data is achieved by using "Sliding Window Overlap Generator (SWOBG) + Location Mask Pre-training (PMBM)".

[0050] Step 2-1: Model pre-training.

[0051] During the pre-training phase, unlabeled satellite traffic is used as input. The model learns the "protocol interaction mode - timing jitter - fragmentation pattern" of satellite traffic through scene-specific pre-training tasks, laying the feature foundation for subsequent fine-tuning.

[0052] 2.1.1 TU generation and encoding.

[0053] TU generation: Unlabeled satellite traffic samples were processed using the SWOBG algorithm, with a window duration of 100ms and an overlap ratio of 0.25 (see attached). Figure 2 As shown in the figure, each traffic stream is divided into 5-8 transmission units (TUs); adjacent TUs overlap by 25% of the data packets, which can effectively preserve the timing context information of the traffic and avoid feature loss due to window fragmentation.

[0054] TU Encoding: TU is encoded using the Tu2Token module: First, the hexadecimal byte sequence of TU is split into binary bytes (e.g., "0x1A0x2B" is split into "1A2B"), and then a token is generated using byte pair encoding (BPE); a token dictionary of size 65536 is constructed, and special tokens such as [BOS] (sequence start), [EOS] (sequence end), and [MASK] (mask) are added to achieve a language-like expression of encrypted traffic.

[0055] 2.1.2 PMBM pre-training task.

[0056] Design a location-differentiated masking task to force the model to focus on key location features of satellite traffic: Position Division: The TU token sequence is divided into a header (protocol start information, reflecting the initialization characteristics of business interaction), a middle section (payload body, carrying the core potential characteristics of encrypted traffic), and a tail section (end marker, associated with the ending rules of business transmission) in a ratio of 15%:70%:15%. (Appendix) Figure 3 As shown in the figure, the model has the best ability to capture satellite traffic characteristics under this division ratio; Differentiated masking strategy: The probability of masking the head token is set to 20% (focusing on learning protocol interaction patterns), the middle token to 15% (focusing on potential correlations in the payload), and the tail token to 10% (while also taking into account the recognition of the end marker); the selected tokens are subjected to the operation of "80% replaced with [MASK], 10% replaced with random tokens, and 10% left unchanged" to improve the feature generalization ability of the model.

[0057] Pre-training parameter settings: batch size = 16, learning rate = 2e-5, dropout rate = 0.5, number of iterations = 1000; start training in the P100 16GGPU environment, monitor the model's prediction accuracy for the mask token in real time, stop training when the accuracy does not improve for 10 consecutive rounds, and save the pre-trained model.

[0058] Step 2-2: Model fine-tuning.

[0059] The fine-tuning phase uses satellite traffic labeled with QoS tags as input. By fine-tuning the parameters, the model is adapted to classification tasks with five types of QoS requirements (high bandwidth and low latency, medium bandwidth and medium latency, low bandwidth and low latency, high reliability, and unknown category). At the same time, heterogeneous samples are introduced to enhance cross-scenario generalization ability.

[0060] 2.2.1 Fine-tuning data preparation.

[0061] Dataset selection: Select satellite traffic datasets that meet the requirements of satellite scene adaptation and are labeled with 5 types of QoS tags (all samples have been processed by the preprocessing process in step 1 to ensure feature consistency). Dataset partitioning: The dataset is divided into a training set (for model parameter updates), a validation set (for training effect monitoring), and a test set (for final performance evaluation) in an 8:1:1 ratio. Generalization enhancement processing: 300 additional heterogeneous samples from ground fusion scenes are introduced and added to the training set to improve the model's cross-scene adaptability.

[0062] 2.2.2 Fine-tuning process.

[0063] Model structure adjustment: Load the pre-trained DD-DistilBERT model, delete the original output layer and replace it with a 5-class QoS classification layer (number of neurons = 5), and use softmax as the activation function (to achieve multi-class probability output); retain the parameters of the bottom Transformer encoder (inheriting the satellite traffic feature extraction capability learned in pre-training), and only update some parameters of the classification layer and the upper encoder to reduce training overhead.

[0064] Fine-tuning parameter settings: batch size = 16, learning rate = 1e-5 (lower than the pre-training learning rate to avoid parameter oscillation), number of iterations = 500, loss function uses cross-entropy (adapted to loss calculation for multi-class tasks); start training on the training set, and evaluate the model's F1 score using the validation set after each iteration.

[0065] Early stopping mechanism triggered: When the F1 score on the validation set shows no improvement for 10 consecutive rounds, the early stopping mechanism is triggered, training is stopped, and the optimal fine-tuned model is saved; the performance metrics of this model on the constructed satellite traffic test set (see appendix) Figure 4 As shown in the figure, the F1 value of the validation set reaches 86.61%, which can effectively meet the QoS classification accuracy requirements of satellite scenarios.

[0066] Steps 2-3: QoS prediction inference.

[0067] The inference phase takes actual encrypted satellite traffic as input and quickly outputs QoS requirement prediction results through a lightweight inference process, adapting to the real-time scheduling requirements of satellite terminals.

[0068] 2.3.1 Input Construction.

[0069] The satellite encrypted traffic stream to be predicted is generated into 5 TUs according to the SWOBG parameters (window 100ms, overlap ratio 0.25) in step 2.1.1. Each TU is encoded into a 512-dimensional token sequence by Tu2Token.

[0070] 2.3.2 Model reasoning and result output.

[0071] Inference calculation: The encoded token sequence is input into the optimal fine-tuning model. The model extracts features through the bottom encoder and calculates probabilities through the upper classification layer, outputting the probability distribution of 5 types of QoS (e.g., "high bandwidth and low latency" probability = 0.85, "medium bandwidth and medium latency" probability = 0.12, and other categories probability < 0.03). Result determination: The category with the highest probability was selected as the final QoS requirement prediction result; Meanwhile, the experimental results of this model on three public datasets—ISCX-VPN-Service, Mobile-App, and CrossPlatform (iOS)—are attached. Figure 5 As shown in the figure, the average F1 score reached 93.04%, proving that it has strong cross-scenario generalization ability and can adapt to the QoS classification requirements of satellite pure scenario and satellite-ground integrated scenario.

[0072] This invention presents a deep learning-based QoS classification method for satellite internet traffic, specifically designed for satellite scenarios. Utilizing the lightweight DD-DistilBERT architecture, it achieves high-precision QoS demand prediction in resource-constrained satellite terminal environments. Its "QoS demand prediction - resource scheduling" functional logic directly supports dynamic allocation of onboard resources, QoS routing optimization, and security anomaly detection. Leveraging DD-DistilBERT's feature capture capabilities for encrypted satellite traffic (e.g., preserving temporal context through a SWOBG sliding window and focusing key features using PMBM location masks), it outputs accurate classification results for five types of QoS demands (high bandwidth, low latency, etc.) without relying on additional application recognition modules (validation set F1 score of 86.61%). Furthermore, this architecture incorporates heterogeneous ground-based samples during the fine-tuning phase, exhibiting strong cross-scenario generalization capabilities and adapting to both pure satellite scenarios and satellite-ground fusion scenarios (average F1 score of 93.04% across three publicly available datasets). Furthermore, the design concept and inference optimization mechanism of "satellite traffic preprocessing - lightweight pre-training - small sample fine-tuning" built around DD-DistilBERT in this invention can be extended to other time series data classification tasks (such as industrial equipment status monitoring and low-altitude communication traffic demand prediction), and has significant engineering application value.

[0073] In the several embodiments provided by this invention, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For instance, the division of modules, units, or units is merely a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units, modules, or components may be combined or integrated into another device, or some features may be ignored or not executed.

[0074] The units may or may not be physically separate. The components shown as units can be one or more physical units, meaning they can be located in one place or distributed in multiple different locations. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0075] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0076] In particular, according to embodiments disclosed in this invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication component, and / or installed from a removable medium. When the computer program is executed by a central processing unit (CPU), it performs the functions defined in the methods of this invention. It should be noted that the computer-readable medium described above in this invention can be a computer-readable signal medium or a computer-readable storage medium, or any combination of the two. The computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof.

[0077] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0078] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions within the technical scope disclosed in the present invention should be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A deep learning based satellite internet traffic QoS classification method, characterized in that, The method comprises the following steps: Obtain satellite encrypted traffic; Preprocess the satellite encrypted traffic to obtain preprocessed satellite encrypted traffic; Obtain a pre-trained DD-DistilBERT model; Input the preprocessed satellite encrypted traffic into the pre-trained DD-DistilBERT model for fine-tuning to obtain an optimal fine-tuning model; Obtain satellite encrypted traffic to be predicted; Input the satellite encrypted traffic to be predicted into the optimal fine-tuning model to obtain a prediction result.

2. The deep learning based satellite internet traffic QoS classification method according to claim 1, characterized in that, The satellite encrypted traffic is preprocessed to obtain preprocessed satellite encrypted traffic, which comprises the following steps: Session segmentation is performed on the satellite encrypted traffic to obtain a plurality of independent traffic flows; Feature processing is performed on the independent traffic flows to obtain target independent traffic flows; Feature extraction and normalization are performed on the target independent traffic flows to obtain the preprocessed satellite encrypted traffic.

3. The deep learning based satellite internet traffic QoS classification method according to claim 2, characterized in that, Session segmentation is performed on the satellite encrypted traffic to obtain a plurality of independent traffic flows, which comprises the following steps: The satellite encrypted traffic is split according to the five-tuple to obtain a plurality of independent traffic flows.

4. The deep learning based satellite internet traffic QoS classification method according to claim 3, characterized in that, Feature processing is performed on the independent traffic flows to obtain target independent traffic flows, which comprises the following steps: Merge the independent traffic flows that are continuous and less than a first preset byte threshold; Remove the independent traffic flows with a length less than a second preset byte threshold; Keep the first plurality of independent traffic flows as target independent traffic flows.

5. The deep learning based satellite internet traffic QoS classification method according to claim 1, characterized in that, Feature extraction and normalization are performed on the target independent traffic flows to obtain the preprocessed satellite encrypted traffic, which comprises the following steps: Extract the payloads of all target independent traffic flows, convert them into a list of decimal integers, and form a payload feature matrix; Count the lengths of all target independent traffic flows to form a packet length sequence feature; Flatten the payload feature matrix and concatenate the packet length sequence feature to obtain an original feature vector; Scale the feature values of the original feature vector to the interval [0, 1] through normalization to obtain the preprocessed satellite encrypted traffic.

6. The deep learning based satellite internet traffic QoS classification method according to claim 5, characterized in that, The method further comprises the following steps: Label the preprocessed satellite encrypted traffic with QoS labels.

7. The deep learning based satellite internet traffic QoS classification method according to claim 6, characterized in that, The pre-training of the DD-DistilBERT model comprises the following steps: TU generation: process the unlabeled satellite traffic samples using the SWOBG algorithm, set the window length to 100 ms and the overlap ratio to 0.25, divide each traffic flow into 5-8 transmission units (TUs), and overlap 25% of the data packets of adjacent TUs; TU encoding: encode the TUs through the Tu2Token module: first, split the hexadecimal byte sequence of the TU according to the binary model, then use byte pair encoding to generate tokens, and construct a token dictionary; Position division: divide the token sequence of the TU into a head, a middle, and a tail according to a ratio of 15:70:15; Differential masking strategy: set the masking probability of the head tokens to 20% and the tail tokens to 10%; replace 80% of the selected tokens with [MASK], 10% with random tokens, and 10% remain unchanged; Pre-training parameter settings: batch size = 16, learning rate = 2e-5, dropout rate = 0.5, number of iterations = 1000; start training in a P100 16G GPU environment, monitor the prediction accuracy of the model for mask tokens in real time, stop training when the accuracy does not improve for 10 consecutive rounds, and save the pre-trained model.

8. The deep learning based satellite internet traffic QoS classification method according to claim 7, characterized in that, Input the preprocessed satellite encrypted traffic into the DD-DistilBERT model after pre-training to fine-tune and obtain the optimal fine-tuned model, which includes the following steps: Fine-tuning data preparation: select preprocessed satellite encrypted traffic with labeled QoS tags that meet the requirements of satellite scenario adaptation, and divide them into training set, validation set and test set, and introduce heterogeneous samples into the training set; Model structure adjustment: load the pre-trained DD-DistilBERT model, delete the original output layer and replace it with a QoS classification layer with a softmax activation function; keep the parameters of the bottom Transformer encoder, and update the parameters of the classification layer and the upper encoder; Fine-tuning parameter settings: batch size = 16, learning rate = 1e-5, number of iterations = 500, and cross-entropy loss function; start training on the training set, and evaluate the model's F1 score on the validation set after each iteration; Early stopping mechanism trigger: when the validation set F1 score does not improve for 10 consecutive rounds, trigger the early stopping mechanism, stop training and save the optimal fine-tuned model.

9. The deep learning based satellite internet traffic QoS classification method according to claim 8, characterized in that, Input the satellite encrypted traffic to be predicted into the optimal fine-tuned model to obtain the prediction results, which includes the following steps: Input the satellite encrypted traffic to be predicted into the optimal fine-tuned model to obtain the prediction results, which includes the following steps: Generate TUs according to the SWOBG algorithm for the satellite encrypted traffic to be predicted, and encode each TU into a token sequence using Tu2Token; Input the encoded token sequence into the optimal fine-tuned model, which extracts features through the bottom encoder and calculates probabilities through the upper classification layer, and outputs the probability distribution; Select the class with the highest probability in the probability distribution as the final QoS demand prediction result.

Images