Internet of vehicles data transmission quality evaluation method based on 5G

Abstract

The invention relates to the field of big data processing, and particularly discloses a 5G-based Internet of Vehicles data transmission quality evaluation method, which takes dynamic network state time sequence data and semantic metadata of a V2X message as double-path input, and extracts deep features of the dynamic network state time sequence data and the semantic metadata respectively. Furthermore, through a cross-modal fusion mechanism, semantic information of the message is used as a probe to dynamically view a network state sequence and focus on network fluctuation forming the most critical influence on the current message, so that a semantically related network state context is generated. And finally, the model accurately predicts the semantic transmission quality of the specific message based on the deeply fused context information and outputs an operable instruction. Through the mode, evaluation is improved from a physical level that the network is good or bad to a semantic level that the network is not good enough to a specific message, application perception evaluation of transmission quality is realized, and a communication quality guarantee with higher decision-making value is provided for upper-layer applications.

Description

Technical Field

[0001] This application relates to the field of big data processing, and more specifically, to a method for evaluating the quality of data transmission in 5G-based vehicle-to-everything (V2X) networks. Background Technology

[0002] With the rapid deployment of 5G and its deep integration with vehicle-to-everything (V2X) applications, intelligent transportation and autonomous driving are experiencing revolutionary development. 5G networks, with their ultra-high bandwidth, ultra-low latency, and massive connectivity, provide unprecedented technical support for efficient information exchange between vehicles, between vehicles and infrastructure, and between vehicles and networks. In safety-critical application scenarios such as autonomous driving, V2X communication not only carries conventional vehicle status information but also transmits critical data with extremely stringent transmission quality requirements, such as collaborative perception, intent sharing, and emergency braking warnings. Therefore, accurately and in real-time evaluating the data transmission quality of 5G networks for specific V2X services in complex and dynamic environments has become a core technical prerequisite for ensuring driving safety and improving traffic efficiency.

[0003] However, existing data transmission quality assessment methods primarily focus on network layer quality of service metrics, such as measuring network latency, jitter, throughput, and packet loss rate to determine network performance. These methods are essentially application-agnostic; they treat all transmitted data bitstreams the same, failing to perceive and differentiate the differences and importance of different V2X messages at the business logic level. For example, a low packet loss rate might be perfectly acceptable for in-vehicle infotainment services, but if the lost data packets happen to be core data describing critical obstacles in a collaborative awareness message, it could directly lead to a catastrophic security incident. Even some advanced methods that introduce machine learning models to predict future network QoS metrics still essentially remain at the level of predicting network physical performance, failing to address the semantic gap between assessment results and the actual needs of upper-layer applications. The fundamental flaw of this assessment mechanism is that it can only answer the question of whether the current network is good or bad, but it cannot answer the more crucial question of whether the current network's transmission quality is good enough for this specific semantic V2X message.

[0004] Therefore, an optimized 5G-based vehicle-to-everything (V2X) data transmission quality assessment scheme is desired. Summary of the Invention

[0005] To address the aforementioned technical problems, this application is proposed. Embodiments of this application provide a method for evaluating the data transmission quality of a 5G-based vehicle-to-everything (V2X) network.

[0006] According to one aspect of this application, a method for evaluating the data transmission quality of a 5G-based vehicle-to-everything (V2X) network is provided, comprising: Network feature engineering is performed on the acquired raw network data and raw vehicle data to obtain network feature vectors; Semantic feature encoding is performed on the acquired V2X message metadata to obtain the message semantic feature vector; Network and semantic spatiotemporal features are extracted from the time series of network feature vectors and message semantic feature vectors to obtain network state encoding sequences and message semantic content encoding feature vectors. Cross-modal feature fusion is performed on the network state encoding sequence and the message semantic content encoding feature vector to obtain a semantically related network state context vector; Semantic quality prediction is performed based on semantically relevant network state context vectors and message semantic content encoding feature vectors to obtain predicted semantic quality scores. The predicted semantic quality score is input into the threshold judgment engine to obtain actionable instructions.

[0007] Compared with existing technologies, this application provides a 5G-based method for assessing the data transmission quality of vehicle-to-everything (V2X) networks. It uses dynamic network state time-series data and semantic metadata of V2X messages as dual inputs, extracting their deep features separately. Furthermore, through a cross-modal fusion mechanism, it utilizes the semantic information of the messages as probes to dynamically examine the network state sequence, focusing on network fluctuations that have the most critical impact on the current message, thereby generating a semantically relevant network state context. Finally, based on this deeply fused context information, the model accurately predicts the semantic transmission quality of specific messages and outputs actionable instructions. In this way, the assessment is elevated from the physical level of network quality to the semantic level of network performance for specific messages, achieving application-aware assessment of transmission quality and providing more decision-making-value communication quality assurance for upper-layer applications. Attached Figure Description

[0008] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0009] Figure 1 This is a flowchart of a 5G-based vehicle-to-everything (V2X) data transmission quality assessment method according to an embodiment of this application; Figure 2 This is a schematic diagram of data flow in a 5G-based vehicle-to-everything (V2X) data transmission quality assessment method according to an embodiment of this application. Figure 3This is a flowchart illustrating the semantic feature encoding of acquired V2X message metadata to obtain a message semantic feature vector, according to the 5G-based vehicle-to-everything (V2X) data transmission quality assessment method of this application. Figure 4 This is a flowchart illustrating the cross-modal feature fusion of network state coding sequences and message semantic content coding feature vectors to obtain semantically related network state context vectors, according to the 5G-based vehicle network data transmission quality assessment method of this application. Figure 5 This is a flowchart illustrating the semantic quality prediction based on semantically related network state context vectors and message semantic content encoding feature vectors in the 5G-based vehicle network data transmission quality assessment method according to embodiments of this application, to obtain a predicted semantic quality score. Detailed Implementation

[0010] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0011] As indicated in this application and claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" are not specifically singular and may include plural forms. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of explicitly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.

[0012] While this application makes various references to certain modules of the systems according to embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The modules described are merely illustrative, and different aspects of the systems and methods may use different modules.

[0013] Flowcharts are used in this application to illustrate the operations performed by the system according to embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from them.

[0014] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0015] Existing technologies for evaluating the quality of vehicle-to-everything (V2X) data transmission generally suffer from core technical problems: their evaluation methods are limited to network physical layer indicators, exhibiting application-independent characteristics and failing to measure the true impact of network fluctuations on the semantic integrity of different V2X messages, thus creating a semantic gap between the evaluation results and actual application needs. To address this issue, this application proposes a 5G-based method for evaluating the quality of V2X data transmission. This method achieves a leap from evaluating bitstream quality to evaluating information quality by constructing an intelligent evaluation model that deeply integrates network state and message semantics. In specific implementation, this method first processes two heterogeneous data streams in parallel: one stream performs time-series modeling on continuous 5G network measurement data to capture the dynamic evolution of network state and form a network state encoding sequence; the other stream encodes metadata such as the type and payload of V2X messages to generate message semantic content encoding feature vectors that characterize their service attributes and inherent requirements. The key innovation of this concept lies in its approach: instead of simply concatenating two features, it employs a cross-modal fusion mechanism. This mechanism uses the message semantic vector as a query to dynamically examine and focus on the network state encoding sequence. It intelligently identifies which fluctuations in the network's history over a period of time pose the greatest potential threat to the current message with specific semantics, thereby generating a highly relevant semantic network state context vector. Finally, this context vector is combined with the original message semantic vector for final prediction, outputting a predicted semantic quality score that directly reflects whether the message content can be effectively received. Based on this score, actionable instructions are generated to guide adjustments to communication strategies, thus precisely bridging the gap between network performance and application requirements.

[0016] Figure 1 This is a flowchart of a 5G-based vehicle-to-everything (V2X) data transmission quality assessment method according to an embodiment of this application. Figure 2 The data flow diagram of the 5G-based vehicle-to-everything (V2X) data transmission quality assessment method according to embodiments of this application is shown below. Figure 1 and Figure 2As shown, the 5G-based vehicle-to-everything (V2X) data transmission quality assessment method according to an embodiment of this application includes: S100, performing network feature engineering on the acquired raw network data and raw vehicle data to obtain a network feature vector; S200, performing semantic feature encoding on the acquired V2X message metadata to obtain a message semantic feature vector; S300, performing network and semantic spatiotemporal feature extraction on the time series of the network feature vector and the message semantic feature vector to obtain a network state encoding sequence and a message semantic content encoding feature vector; S400, performing cross-modal feature fusion on the network state encoding sequence and the message semantic content encoding feature vector to obtain a semantically relevant network state context vector; S500, performing semantic quality prediction based on the semantically relevant network state context vector and the message semantic content encoding feature vector to obtain a predicted semantic quality score; and S600, inputting the predicted semantic quality score into a threshold judgment engine to obtain an operable instruction.

[0017] Specifically, in step S100, network feature engineering is performed on the acquired raw network data and raw vehicle data to obtain network feature vectors. It should be understood that since the acquired raw network data is instantaneous, high-frequency, and noisy, while the raw vehicle data reflects the physical dynamic context affecting the wireless channel environment, these two types of raw data streams differ significantly in format, frequency, and physical meaning, and cannot be directly used for effective analysis in downstream machine learning models. Therefore, in the technical solution of this application, network feature engineering is performed on the acquired raw network data and raw vehicle data to transform these heterogeneous, unstructured raw data into a structured mathematical representation that can comprehensively and stably characterize the network communication environment and its dynamic changing trends. This provides a higher information density, stronger robustness, and dimensionally unified input feature for subsequent transmission quality assessment models, thereby significantly improving the accuracy and reliability of the assessment.

[0018] More specifically, the implementation process of this network feature engineering first involves time synchronization and windowing of data from different sources. Then, statistical features reflecting the dynamic characteristics of the network state are extracted and integrated with vehicle physical state information. Finally, these features are concatenated and normalized to generate a network feature vector. In a specific example of this application, when it is necessary to evaluate the transmission quality of a V2X message, the generation timestamp of the message is used as a benchmark to backtrack and extract all raw network data (including RSRP, SINR, etc.) and corresponding raw vehicle data (including speed, acceleration, etc.) within the past 200 milliseconds. Next, for the network data sequence within this time window, the mean, variance, and linear regression slope of each indicator are calculated to quantify its average performance, stability, and trend. Subsequently, these calculated statistical features are concatenated with the instantaneous vehicle speed and acceleration values ​​at the end of the time window to form a temporary multidimensional raw feature vector. Finally, a preset minimum-maximum normalization method is applied to each dimension of this temporary vector to linearly scale all its components to the range of 0 to 1, thereby obtaining the final network feature vector.

[0019] Specifically, in step S200, the acquired V2X message metadata is semantically encoded to obtain a message semantic feature vector. It should be understood that since the original V2X message metadata consists of non-numerical category labels (such as message type) and numerical values ​​with varying dimensions (such as load size), its format is inconsistent and it lacks direct computability, making it unsuitable as effective input to downstream machine learning models. Therefore, in the technical solution of this application, the acquired V2X message metadata is further semantically encoded to transform these heterogeneous, descriptive metadata into a standardized, machine-understandable, and machine-processable quantitative numerical vector. This enables the model not only to identify the message's identity but also to quantify its inherent business attributes and communication resource requirements, providing a crucial semantic foundation for subsequent application-aware cross-modal feature fusion.

[0020] Figure 3 This is a flowchart illustrating the process of semantically encoding acquired V2X message metadata to obtain a message semantic feature vector, according to an embodiment of this application. For example... Figure 3 As shown, step S200 includes: S210, extracting message category features and message numerical features from V2X message metadata; S220, performing category feature encoding on the message category features to obtain a message category feature encoding vector; S230, performing numerical feature quantization on the message numerical features to obtain a message numerical feature encoding vector; S240, concatenating the message category feature encoding vector and the message numerical feature encoding vector to obtain a message semantic feature vector.

[0021] More specifically, the implementation process of this semantic feature encoding first separates categorical and numerical features from the message metadata, then performs appropriate encoding and quantization processing on them respectively, and finally concatenates the processed results into a unified semantic feature vector. In a specific example of this application, when a collaborative awareness message to be evaluated is received, with a payload size of 1200 bytes, the message categorical feature (i.e., message type CPM) and the message numerical feature (i.e., payload size 1200) are first extracted from the V2X message metadata. Next, the message categorical feature is categorically encoded. Based on the preset message type dictionary {BSM, CAM, CPM}, one-hot encoding is used to convert CPM into a message categorical feature encoding vector [0, 0, 1]. At the same time, the message numerical feature is quantized. Based on the preset payload size range (e.g., 0 to 2048 bytes), the 1200 bytes are converted into a message numerical feature encoding vector [0.586] using the minimum-maximum normalization method. Finally, the message category feature encoding vector and the message numerical feature encoding vector are concatenated, and [0, 0, 1] and [0.586] are merged to obtain the final message semantic feature vector [0, 0, 1, 0.586].

[0022] Specifically, in step S300, network and semantic spatiotemporal feature extraction is performed on the time series of the network feature vector and the message semantic feature vector to obtain the network state encoding sequence and the message semantic content encoding feature vector. It should be understood that since the network feature vector generated in the previous step is a time series, it contains the dynamic patterns and dependencies of the network state evolving over time, while the message semantic feature vector is a static vector describing the inherent attributes of the message. The two differ fundamentally in data structure, information dimension, and the connotation they represent. Therefore, in the technical solution of this application, network and semantic spatiotemporal feature extraction is further performed on the time series of the network feature vector and the message semantic feature vector. This allows the two heterogeneous features to be mapped to a unified high-dimensional feature space through their respective dedicated deep learning models, and the temporal context information of the network state and the deep abstract representation of the message semantics are extracted respectively. This provides high-quality input with structural equivalence and semantic alignment for subsequent cross-modal feature fusion, enabling the model to learn and understand the complex interaction between the dynamic network environment and the static message requirements.

[0023] More specifically, in this embodiment of the application, network and semantic spatiotemporal feature extraction is performed on the time series of network feature vectors and message semantic feature vectors to obtain network state encoding sequences and message semantic content encoding feature vectors, including: inputting the time series of network feature vectors into a recurrent neural network to obtain a network state encoding sequence; and inputting the message semantic feature vectors into a multilayer perceptron to obtain message semantic content encoding feature vectors.

[0024] In other words, the implementation process of this network and semantic spatiotemporal feature extraction adopts a dual-parallel processing architecture. One path uses a recurrent neural network to process the time series of network features, while the other path uses a multilayer perceptron to process the static message semantic feature vector. In a specific example of this application, the network feature vector sequence generated in the previous step, consisting of 20 time steps, is input into a pre-defined gated recurrent unit network containing two layers with 64 hidden units in each layer. The recurrent neural network processes each network feature vector in the sequence in chronological order, and its gating mechanism captures and remembers the temporal dependencies of network states (such as continuous deterioration or sudden improvement of channel quality), ultimately outputting a network state encoding sequence containing 20 64-dimensional vectors. At the same time, the message semantic feature vector [0,0, 1, 0.586] corresponding to the collaboratively perceived message is input into a pre-defined multilayer perceptron with two hidden layers (with 32 and 64 neurons respectively) and using the ReLU activation function. Through layer-by-layer nonlinear transformation of the multilayer perceptron, the original sparse features are mapped to a 64-dimensional high-dimensional dense semantic space, resulting in a message semantic content encoding feature vector that can more richly represent the inherent needs of the message.

[0025] Specifically, in step S400, cross-modal feature fusion is performed on the network state encoding sequence and the message semantic content encoding feature vector to obtain a semantically relevant network state context vector. It should be understood that since the network state encoding sequence and the message semantic content encoding feature vector obtained in the previous step belong to different modalities—the former describing the time-varying network physical environment and the latter describing static application layer requirements—simply concatenating or adding them together cannot effectively reveal the nonlinear impact of network state fluctuations at a specific point in time on the semantics of a specific message. Furthermore, existing attention mechanisms suffer from a core flaw of homogeneous query generation when processing the semantic content encoding of vehicular network messages. Specifically, this mechanism indiscriminately generates a query vector from a semantic vector encoded by a mixture of categorical features such as message type and numerical features such as load size through a single linear transformation. This approach ignores the differentiated roles played by different types of semantic features in guiding the attention mechanism to focus on the network state. In vehicle-to-everything (V2X) applications, the category attributes of a message (e.g., whether it is an emergency braking message) typically determine its baseline paradigm for network quality requirements (e.g., latency, jitter), while numerical attributes (e.g., load size) are fine-tuned and modulated on top of this baseline. The original mechanism mixes these two different granularities of influence into a homogeneous query vector, making it difficult for the model to learn refined attention patterns. It cannot clearly decouple and characterize whether the current attention weight is dominated by the message type or its size, thus limiting the model's representational ability and the accuracy of the final transmission quality assessment. To address the aforementioned query generation homogeneity problem, in a preferred embodiment of this solution, a dual-path decoupled attention mechanism based on information divergence is proposed. Specifically, in the technical solution of this application, cross-modal feature fusion is further performed on the network state encoding sequence and the message semantic content encoding feature vector to quantify the degree of difference between the actual network state at each time point and the ideal network state expected by the current message semantics, and dynamic feature aggregation is performed based on this difference. This allows the model to accurately identify and focus on the critical moments that are least compatible with the current message requirements and pose the greatest potential risk from the historical sequence of network states, thereby generating a more representative and risk-focused context vector.

[0026] Figure 4 This is a flowchart illustrating the cross-modal feature fusion of network state encoding sequences and message semantic content encoding feature vectors to obtain semantically relevant network state context vectors, according to an embodiment of this application's 5G-based vehicle-to-everything (V2X) data transmission quality assessment method. Figure 4As shown, step S400 includes: S410, performing semantic feature decoupling and distribution parameterization on the message semantic content encoding feature vector to obtain the target Gaussian distribution; S420, performing network state distribution on each network state encoding in the network state encoding sequence to obtain the observation distribution sequence; S430, calculating the KL divergence between the target Gaussian distribution and each observation distribution in the observation distribution sequence to obtain the divergence score sequence; S440, performing context aggregation on the network state encoding sequence based on the divergence score sequence to obtain the semantically related network state context vector.

[0027] Accordingly, in step S410, the semantic feature decoupling and distribution parameterization of the message semantic content encoding feature vector are performed to obtain a target Gaussian distribution. It should be understood that a single message semantic content encoding feature vector, formed by mixing message categorical and numerical features, suffers from query generation homogeneity, failing to distinguish the differentiated roles played by different semantic dimensions in defining network quality requirements. In the vehicle-to-everything (V2X) scenario, the message's categorical attributes (such as emergency braking messages) determine its baseline paradigm for network quality requirements (such as extremely low latency), while numerical attributes (such as load size) are fine-tuned and modulated on top of this baseline (such as requiring higher bandwidth). Therefore, in the technical solution of this application, the message semantic content encoding feature vector is semantically decoupled and distributed parameterized to obtain a target Gaussian distribution, thereby explicitly mapping an abstract V2X message semantic requirement to a probability distribution representing the ideal network state expected or preferred by the message. In this way, the model can learn that message categories define the baseline position of the ideal network state distribution, while the numerical attributes of messages fine-tune this baseline, thereby achieving a more refined and physically meaningful mathematical representation of semantic requirements.

[0028] Specifically, in this embodiment of the application, semantic feature decoupling and distribution parameterization are performed on the message semantic content encoding feature vector to obtain a target Gaussian distribution, including: decoupling the message semantic content encoding feature vector to obtain a category feature encoding part and a numerical feature encoding part; and performing distribution parameterization on the category feature encoding part and the numerical feature encoding part to obtain a target Gaussian distribution.

[0029] More specifically, the implementation process of semantic feature decoupling and distributed parameterization first decomposes the unified message semantic encoding vector back into its category and numerical components, and then generates the statistical parameters of the target Gaussian distribution through parallel nonlinear transformation and subsequent fusion parameterization. In a specific example of this application, the 64-dimensional message semantic content encoding feature vector generated in the previous step is first processed... Feature decoupling is performed by explicitly decomposing it into its original components, namely the category feature encoding part representing the message category. and the numerical feature encoding part representing the numerical attributes of the message Subsequently, the decoupled features are nonlinearly transformed using two independent multilayer perceptrons, and their outputs are fused to parameterize and generate a target Gaussian distribution. The mean and logarithmic variance. Specifically, the categorical feature encoding is input into a separate, pre-defined multilayer perceptron network. The numerical feature encoding part is input into another independent multilayer perceptron network. Both undergo nonlinear transformations to extract their deep representations. Then, and The output vectors are concatenated to form an intermediate feature vector that integrates category and numerical information. Finally, this intermediate feature vector is input into a third independent multilayer perceptron network. The network's output layer has a dimension of 128. This 128-dimensional output vector is parsed into two 64-dimensional vectors, serving as the mean vector and log-variance vector of the target Gaussian distribution, respectively. The categorical feature encoding and numerical feature encoding parts are parameterized for distribution using the following formula:

[0030] in, The vector representing the mean of the target Gaussian distribution. The variance vector representing the target Gaussian distribution. They are three independent multilayer perceptron networks. This represents a vector concatenation operation. For the categorical feature encoding part, This is the part encoded for numerical features. In this way, an abstract V2X message semantic (e.g., a high-load collaborative sensing message) is successfully mapped to a Gaussian distribution representing the ideal network state expected or preferred by that message.

[0031] Accordingly, in step S420, the network state codes in the network state coding sequence are distributed to obtain the observation distribution sequence. It should be understood that since the semantic requirements of the message have been parameterized as a target probability distribution in the previous step, while the network state coding sequence is still composed of a series of deterministic high-dimensional vectors, there is a lack of a unified, directly information-theoretic benchmark for comparison between these two different forms of mathematical representation. Therefore, in order to quantify the degree of mismatch between the actual network state and the aforementioned semantic expectation, it is necessary to distribute the coding at each time step in the network state sequence. In the technical solution of this application, the network state codes in the network state coding sequence are further distributed to map the deterministic network state vector at each time point to a probability distribution, thereby constructing a framework capable of quantifying the degree of mismatch between the actual network state and the semantic expectation within a unified probability space. In this way, the subsequent attention score calculation can be upgraded from the traditional dot product operation that measures the geometric similarity between vectors to the KL divergence calculation that measures the information difference between probability distributions. This directly quantifies the risk of the real network state deviating from the ideal state and enables accurate identification of key risk points in the network state time series.

[0032] Specifically, the implementation process of this network state distribution involves independently transforming each vector in the network state encoding sequence through a neural network to generate the observation distribution parameters at the corresponding time step. In a specific example of this application, the 64-dimensional network state encoding sequence generated in the previous step, consisting of 20 time steps, is transformed... Each network state code Each input is sequentially fed into an independent multilayer perceptron network, and mapped to the Gaussian distribution of the network state observed at that moment. The mean and log-variance of the network are calculated. The output layer of the network has a dimension of 128, and its output is parsed into two independent 64-dimensional vectors, which serve as the mean vector and log-variance vector of the observation distribution at that time, respectively. After this operation is completed for encoding all 20 network states in the sequence, a sequence of observation distributions consisting of 20 Gaussian distributions is obtained, where each distribution represents a probabilistic representation of the network state observed at that specific time.

[0033] Accordingly, in step S430, the KL divergence between the target Gaussian distribution and each observed distribution in the observed distribution sequence is calculated to obtain a divergence score sequence. It should be understood that since the previous step has mapped both message semantics and network state to the probability distribution space, the model still lacks a quantitative metric with clear physical meaning to assess the degree of mismatch between the two. Therefore, in the technical solution of this application, the KL divergence between the target Gaussian distribution and each observed distribution in the observed distribution sequence is further calculated to obtain a divergence score sequence, thereby generating a numerical score for each time point in the network state sequence that directly quantifies the degree of risk of deviating from the ideal state. In this way, the calculation of attention scores can be upgraded from the traditional dot product operation that measures the geometric similarity between vectors to the KL divergence calculation that measures the information difference between probability distributions, enabling the model to accurately identify the network state moment that poses the greatest threat to the current message transmission based on clear risk signals.

[0034] Specifically, the implementation process of this KL divergence calculation involves traversing the observed distribution sequence and calculating the information divergence between each distribution and a single target distribution, and then calculating the target distribution. Observational distribution at each time step The differences between the target and target distributions are used to form a score sequence of the same length as the time series. In a specific example of this application, a single target Gaussian distribution (defined by its mean and variance vectors) representing the demand for Collaborative Perception Messages (CPM) is obtained, along with a sequence of 20 observed Gaussian distributions representing the network state over the past 200 milliseconds. First, the observed distribution at the first time step is extracted from the observed distribution sequence, and its mean and variance are compared with those of the target distribution. These are then substituted into a pre-defined KL divergence calculation formula to obtain a scalar value. This value directly quantifies the difference between the network state at the first time point and the ideal demand of the CPM message. Subsequently, the KL divergence between the target distribution and the remaining 19 observed distributions is calculated sequentially in the same manner. Finally, these 20 calculated KL divergence scalar values ​​are arranged in chronological order to form a divergence score sequence containing 20 elements. This is expressed by the following formula:

[0035] in, This represents the KL divergence value between two distributions. and It is the target distribution The mean and standard deviation, h and The observed distribution at time step t The mean and standard deviation are then calculated. This elevates the calculation of attention scores from the traditional dot product operation, which measures geometric similarity between vectors, to the KL divergence calculation, which measures the information difference between probability distributions. It can be understood that the KL divergence value directly quantifies the degree of risk of the actual network state deviating from the ideal state. If the distribution generated by the network state at a certain moment (such as high latency, high jitter) differs significantly from the expected distribution (low latency, low jitter) of the current emergency message, i.e., the KL divergence value is high, the model will identify that moment as a dangerous signal requiring high attention, achieving accurate identification of key risk points in the network state timeline.

[0036] Accordingly, in step S440, context aggregation is performed on the network state encoding sequence based on the divergence score sequence to obtain a semantically relevant network state context vector. It should be understood that the divergence score sequence generated in the previous step only provides a quantitative assessment of the degree of mismatch between the network state and the message semantic requirements at each time point, but does not effectively combine these discrete risk signals with the specific content of the network state itself to form a unified feature representation that can be directly used by downstream tasks. Therefore, in the technical solution of this application, context aggregation is further performed on the network state encoding sequence based on the divergence score sequence to convert these quantified risk level scores into effective attention weights, and these weights are used to perform a weighted summation on the original network state encoding sequence, thereby compressing the entire temporal information into a single context vector.

[0037] Specifically, in this embodiment of the application, context aggregation is performed on the network state encoding sequence based on the divergence score sequence to obtain a semantically relevant network state context vector, including: weighting the divergence score sequence based on the Softmax function to obtain a weight sequence; and performing a weighted summation on the network state encoding sequence based on the weight sequence to obtain the semantically relevant network state context vector.

[0038] More specifically, the implementation of this context aggregation first normalizes the divergence score sequence to generate attention weights, and then uses these weights to perform a weighted summation of the network state encoding sequence. In a specific example of this application, the divergence score sequence generated in the previous step, containing 20 elements (where each element...) is first... That is A weighting process based on the Softmax function is performed. This Softmax function amplifies terms with higher scores, i.e., those with larger KL divergence values ​​and higher risk levels, thus generating a weight sequence of 20 elements with a sum of 1. This sequence represents the attention weights rich in risk information. In other words, the divergence score sequence is first input into the Softmax function, which amplifies terms with higher scores (i.e., those with larger KL divergence values ​​and higher risk levels), thereby generating an effective probability distribution as the final attention weights. This can be expressed by the following formula: in, This is a semantically relevant network state context vector. This ensures that the final generated context vector... Instead of a vague, uniform average network state across all moments, it becomes a highly focused risk profile vector dominated by the network state moments that are least semantically relevant to the current message and pose the highest risk. This provides downstream transmission quality prediction tasks with inputs that are more information-dense and valuable for decision-making.

[0039] Through the above preferred embodiments, the attention weight generation process can be transformed from a black-box computation relying on fuzzy vector similarity into a white-box metric process based on probability distribution and information divergence, possessing clear physical meaning and interpretability. This mechanism not only solves the problem of insufficient representational ability caused by the homogenization of semantic feature processing in the original mechanism, but more importantly, it can dynamically and adaptively identify the key time points in the network state time series that pose the greatest potential threat to the current message based on the specific category and numerical attributes of the V2X message. In this way, the model's ability to capture the interaction between network dynamics and application requirements in complex vehicle-to-everything (V2X) scenarios is improved, making the final output transmission quality assessment results more accurate and reliable, thereby providing a more robust communication quality guarantee for upper-layer autonomous driving applications.

[0040] Specifically, in step S500, semantic quality prediction is performed based on the semantically relevant network state context vector and the message semantic content encoding feature vector to obtain a predicted semantic quality score. It should be understood that although the semantically relevant network state context vector generated in the previous step is highly focused on network risks related to message semantics, it only represents the overall state of the network side, while the message semantic content encoding feature vector represents the inherent attributes of the message itself. These two need to be jointly decided to arrive at a conclusive prediction about transmission quality. Therefore, in the technical solution of this application, semantic quality prediction is further performed based on the semantically relevant network state context vector and the message semantic content encoding feature vector. This allows a final regression model to learn and establish a nonlinear mapping relationship from the joint feature representation to a specific semantic quality score. In this way, complex, multi-dimensional risk and demand information can ultimately converge into an intuitive, quantitative prediction score, thereby providing direct and clear communication quality assessment results for upper-layer applications and effectively bridging the information gap between network physical performance and application semantic requirements.

[0041] Figure 5 This document presents a flowchart illustrating the semantic quality prediction process based on semantically relevant network state context vectors and message semantic content encoding feature vectors in a 5G-based vehicle-to-everything (V2X) data transmission quality assessment method according to embodiments of this application. The flowchart illustrates this process to obtain a predicted semantic quality score. Figure 5 As shown, step S500 includes: S510, performing feature concatenation on the semantically related network state context vector and the message semantic content encoding feature vector to obtain the final fused feature vector; S520, inputting the final fused feature vector into the MLP regression network to obtain the predicted semantic quality score.

[0042] More specifically, the semantic quality prediction process first concatenates the context vector representing dynamic network risk with the semantic vector representing static message demand to form a comprehensive joint feature representation. This joint feature is then input into a regression network to output the final prediction score. In a specific example of this application, the 64-dimensional semantically related network state context vector generated in the previous step and the 64-dimensional message semantic content encoding feature vector corresponding to the collaboratively perceived message are first concatenated, and then merged along the feature dimensions to obtain a 128-dimensional final fused feature vector. Subsequently, this 128-dimensional final fused feature vector is input into a pre-defined MLP regression network, which contains a hidden layer with 32 neurons and a linear output layer with an output dimension of 1. The activation value of this output layer is the predicted semantic quality score, ranging from 0 to 1, where a higher value indicates a higher probability that the semantic content of the collaboratively perceived message is successfully received and parsed.

[0043] Specifically, in step S600, the predicted semantic quality score is input into the threshold judgment engine to obtain operable instructions. It should be understood that since the predicted semantic quality score obtained in the previous step is a continuous value, it does not directly constitute a discrete decision instruction that can be executed by upper-layer V2X applications or communication protocol stacks, creating a gap between evaluation and execution. Therefore, in the technical solution of this application, the predicted semantic quality score is further input into the threshold judgment engine to map the quantified evaluation result into predefined control instructions with clear operational meaning. This transforms abstract quality evaluation into specific communication strategy adjustments or application-layer behavioral interventions, thereby forming a complete perception-decision-execution closed loop and achieving intelligent and adaptive management of V2X communication.

[0044] More specifically, the generation process of this operable instruction first presets multiple semantic quality score thresholds based on the service priority and security level of different V2X messages. Then, it compares the real-time predicted score with these thresholds and finally outputs the corresponding instruction based on its range. In a specific example of this application, two thresholds are preset for collaboratively aware messages: a high-quality threshold of 0.9 and a low-quality threshold of 0.6. When the semantic quality score predicted for a CPM to be sent is 0.95, since this score is higher than the high-quality threshold of 0.9, the threshold judgment engine will output the instruction "send normally". If the predicted score is 0.75, since this score is between 0.6 and 0.9, the engine will output the instruction "compress data load before sending" to improve the transmission success rate while ensuring critical information. If the predicted score is only 0.5, since this score is lower than the low-quality threshold of 0.6, the engine will output the instruction "delay sending" and wait for the next evaluation cycle to seek better network conditions.

[0045] In summary, the 5G-based vehicle-to-everything (V2X) data transmission quality assessment method according to the embodiments of this application is clarified. It uses dynamic network state time-series data and semantic metadata of V2X messages as dual inputs, and extracts their deep features separately. Furthermore, through a cross-modal fusion mechanism, the semantic information of the message is used as a probe to dynamically examine the network state sequence, focusing on network fluctuations that have the most critical impact on the current message, thereby generating a semantically relevant network state context. Finally, based on this deeply fused context information, the model accurately predicts the semantic transmission quality of a specific message and outputs actionable instructions. In this way, the assessment is elevated from the physical level of network quality to the semantic level of whether the network is good enough for a specific message, realizing application-aware assessment of transmission quality and providing more decision-making-value communication quality assurance for upper-layer applications.

[0046] The various embodiments of this disclosure have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A method for evaluating the data transmission quality of 5G-based vehicle-to-everything (V2X) networks, characterized in that, include: Network feature engineering is performed on the acquired raw network data and raw vehicle data to obtain network feature vectors; Semantic feature encoding is performed on the acquired V2X message metadata to obtain the message semantic feature vector; Network and semantic spatiotemporal features are extracted from the time series of network feature vectors and message semantic feature vectors to obtain network state encoding sequences and message semantic content encoding feature vectors. Cross-modal feature fusion is performed on the network state encoding sequence and the message semantic content encoding feature vector to obtain a semantically related network state context vector; Semantic quality prediction is performed based on semantically relevant network state context vectors and message semantic content encoding feature vectors to obtain predicted semantic quality scores. The predicted semantic quality score is input into the threshold judgment engine to obtain actionable instructions.

2. The method for evaluating the data transmission quality of 5G-based vehicle-to-everything (V2X) networks according to claim 1, characterized in that, The acquired V2X message metadata is semantically encoded to obtain a message semantic feature vector, including: Extract message category features and message numerical features from V2X message metadata; The message category features are categorically encoded to obtain the message category feature encoding vector; Numerical feature quantization is performed on the numerical features of the message to obtain the message numerical feature encoding vector; The message category feature encoding vector and the message numerical feature encoding vector are concatenated to obtain the message semantic feature vector.

3. The method for evaluating the data transmission quality of 5G-based vehicle-to-everything (V2X) networks according to claim 1, characterized in that, Network and semantic spatiotemporal features are extracted from the time series of network feature vectors and message semantic feature vectors to obtain network state encoding sequences and message semantic content encoding feature vectors, including: The time series of network feature vectors are input into a recurrent neural network to obtain the network state encoding sequence; The message semantic feature vector is input into a multilayer perceptron to obtain the message semantic content encoding feature vector.

4. The method for evaluating the data transmission quality of 5G-based vehicle-to-everything (V2X) networks according to claim 1, characterized in that, Cross-modal feature fusion is performed on the network state encoding sequence and the message semantic content encoding feature vector to obtain a semantically relevant network state context vector, including: Semantic feature decoupling and distribution parameterization are performed on the message semantic content encoding feature vector to obtain the target Gaussian distribution; The network state codes in the network state coding sequence are distributed to obtain the observation distribution sequence; Calculate the KL divergence between the target Gaussian distribution and each observed distribution in the observed distribution sequence to obtain the divergence score sequence; Based on the divergence score sequence, context aggregation is performed on the network state encoding sequence to obtain semantically relevant network state context vectors.

5. The method for evaluating the data transmission quality of 5G-based vehicle-to-everything (V2X) networks according to claim 4, characterized in that, The semantic feature decoupling and distribution parameterization of the message semantic content encoding feature vector are performed to obtain the target Gaussian distribution, including: The semantic content encoding feature vector of the message is decoupled to obtain the category feature encoding part and the numerical feature encoding part; The categorical feature encoding part and the numerical feature encoding part are distributed parameterized to obtain the target Gaussian distribution.

6. The method for evaluating the data transmission quality of 5G-based vehicle-to-everything (V2X) networks according to claim 5, characterized in that, The distribution parameterization of the categorical feature encoding part and the numerical feature encoding part to obtain the target Gaussian distribution includes: performing distribution parameterization on the categorical feature encoding part and the numerical feature encoding part using the following formula: ； in, The vector representing the mean of the target Gaussian distribution. The variance vector representing the target Gaussian distribution. They are three independent multilayer perceptron networks. This represents a vector concatenation operation. For the categorical feature encoding part, This is the numerical feature encoding part. It is the value of a logarithmic function with base 2.

7. The method for evaluating the data transmission quality of 5G-based vehicle-to-everything (V2X) networks according to claim 5, characterized in that, Based on the divergence score sequence, context aggregation is performed on the network state encoding sequence to obtain a semantically relevant network state context vector, including: The divergence score sequence is weighted using the Softmax function to obtain a weighted sequence; Based on the weight sequence, the network state encoding sequence is weighted and summed to obtain the semantically related network state context vector.

8. The method for evaluating the data transmission quality of 5G-based vehicle-to-everything (V2X) networks according to claim 1, characterized in that, Semantic quality prediction is performed based on the semantically relevant network state context vector and message semantic content encoding feature vector to obtain the predicted semantic quality score, including: The semantically related network state context vector and message semantic content encoding feature vector are concatenated to obtain the final fused feature vector; The final fused feature vector is input into the MLP regression network to obtain the predicted semantic quality score.

Images