Channel data evaluation method and device, equipment, storage medium and product

Abstract

The invention discloses a channel data evaluation method and device, equipment, a storage medium and a product. The method comprises the following steps: acquiring initial wireless channel data; the initial wireless channel data is used for measuring the quality of wireless communication; performing key channel feature extraction on the initial wireless channel data to obtain a plurality of channel features and a plurality of corresponding initial channel feature values; for each initial channel characteristic value, discretizing the initial channel characteristic value by adopting a residual vector quantization algorithm to obtain a characteristic marking sequence; based on the feature mark sequence and a preset instruction prompt word, constructing an input sequence for large language model recognition; the instruction prompt word is used for indicating a to-be-executed evaluation task of the large language model; and inputting the input sequence into a pre-trained large language model to obtain an evaluation result of the initial wireless channel data. The method can improve the accuracy of the channel data evaluation result.

Description

Technical Field

[0001] This application relates to the field of communication technology, and in particular to a channel data evaluation method, apparatus, device, computer storage medium, and computer program product. Background Technology

[0002] In wireless communication systems, channel data evaluation is a key step in optimizing network performance and improving communication quality.

[0003] Existing technologies typically employ Vector Quantization (VQ) to map continuous features of channel data to discrete labels when evaluating channel data. However, traditional VQ uses only one codebook, mapping the input vector to the nearest code vector in the codebook. This leads to codebook collapse (i.e., many vectors in the codebook are rarely used, while only a few are frequently used), thus failing to fully utilize the codebook. This results in low label utilization, reduces the accuracy of channel data quantization, and further reduces the accuracy of channel data evaluation results.

[0004] Therefore, how to provide a channel data evaluation method to improve the accuracy of channel data quantization and thus improve the accuracy of channel data evaluation has become an urgent problem to be solved. Summary of the Invention

[0005] This application provides a channel data evaluation method, apparatus, device, computer storage medium, and computer program product, which can improve the accuracy of channel data evaluation results.

[0006] Firstly, this application provides a channel data evaluation method, which includes: Acquire initial wireless channel data; the initial wireless channel data is used to measure the quality of wireless communication. Key channel features are extracted from the initial wireless channel data to obtain multiple channel features and corresponding initial channel feature values; the channel features include at least one of delay spread, Doppler spread, path loss, and coherence bandwidth. For each initial channel feature value, the residual vector quantization algorithm is used to discretize the initial channel feature value to obtain a feature label sequence; Based on the feature tag sequence and preset instruction prompts, an input sequence for large language model recognition is constructed; the instruction prompts are used to indicate the evaluation task to be performed by the large language model. The input sequence is fed into a pre-trained large language model to obtain the evaluation results of the initial wireless channel data.

[0007] In some possible implementations, the residual vector quantization algorithm is used to discretize the channel feature values ​​to obtain a feature label sequence, including: Based on the initial channel feature values, determine the initial residual vector; Based on the initial residual vector, a first codebook is selected from a plurality of pre-trained codebooks; the codebook includes a plurality of code vectors and a feature label corresponding to each code vector. Select a first code vector corresponding to the initial residual vector and a first feature tag corresponding to the first code vector from the first codebook; Based on the initial residual vector and the first code vector, a second residual vector is obtained; Select the second code vector corresponding to the second residual vector and the second feature tag corresponding to the second code vector from the second codebook; Based on the second residual vector and the second code vector, the third residual vector is obtained; Select the third code vector corresponding to the third residual vector and the third feature tag corresponding to the third code vector from the third codebook; The third residual vector is updated to the initial residual vector, the third codebook is updated to the first codebook, and the process of selecting the first code vector corresponding to the initial residual vector and the first feature tag corresponding to the first code vector from the first codebook is repeated until the residual vector converges, resulting in a feature tag sequence. The feature tag sequence includes the first feature tag, the second feature tag, the third feature tag, and the feature tags obtained in each subsequent iteration.

[0008] In some possible implementations, selecting the first code vector corresponding to the initial residual vector from the first codebook includes: Calculate the similarity between the initial residual vector and each code vector in the first codebook; If the similarity satisfies the preset similarity conditions, the code vector that satisfies the preset similarity conditions will be used as the first code vector corresponding to the initial residual vector.

[0009] In some possible implementations, the method further includes: Obtain the preset key position markers in the sequence; The step of constructing an input sequence for large language model recognition based on the feature marker sequence and preset instruction prompts includes: Based on the feature tag sequence, preset instruction prompts, and key position tags of the sequence, an input sequence for large language model recognition is constructed.

[0010] In some possible implementations, the sequence key position markers include: a sequence start marker, a sequence segmentation marker, and a sequence end marker; the sequence start marker is used to mark the beginning of the input sequence, the sequence end marker is used to mark the end of the input sequence, and the sequence segmentation marker is used to segment the feature marker sequence and the instruction prompt word; The step of constructing an input sequence for large language model recognition based on the feature-labeled sequence, preset instruction prompts, and key position markers of the sequence includes: An input sequence template is constructed based on the sequence start marker, the feature marker sequence, the sequence segmentation marker, the instruction prompt word, and the sequence end marker; Based on the input sequence template, an input sequence for large language model recognition is constructed.

[0011] In some possible implementations, the method further includes: Based on the relationship between feature tags and preset channel feature index attributes, the feature tag sequence is converted into multiple channel feature relationship pairs containing channel features and index attributes; The step of constructing an input sequence for large language model recognition based on the input sequence template includes: Based on the sequence start marker, the channel feature relationship pair, the sequence segmentation marker, the instruction prompt word, and the sequence end marker, an input sequence for large language model recognition is constructed.

[0012] In some possible implementations, the input sequence template is arranged in the order of the sequence start marker, the feature marker sequence, the sequence segmentation marker, the instruction prompt word, and the sequence end marker.

[0013] In some possible implementations, the step of inputting the input sequence into a pre-trained large language model to obtain the evaluation result of the initial wireless channel data includes: Obtain the input sequence corresponding to each channel feature; Based on each input sequence, the evaluation result of each channel feature is obtained; The evaluation results of the initial wireless channel data are obtained by combining the evaluation results of multiple channel characteristics.

[0014] In some possible implementations, the method further includes: Based on the evaluation results of the initial wireless channel data, feature tags corresponding to channel features that are related to the evaluation results are obtained; Based on the mapping relationship between the feature marker and the channel feature data interval, the channel feature data interval corresponding to the feature marker is obtained; Given that the initial channel characteristic values ​​are within the channel characteristic data range, the evaluation result is deemed to have passed verification.

[0015] In some possible implementations, the method further includes: Based on the evaluation results of the initial wireless channel data, obtain the causal relationship pairs in the evaluation results; The causal relationship pairs are matched with a preset knowledge base to obtain the verification result of the causal relationship; the knowledge base stores causal relationships between channel characteristics and communication results of different indicator attributes in advance. If the causal relationship represented by the causal relationship pair is consistent with the causal relationship in the knowledge base, the evaluation result is determined to be verified.

[0016] In some possible implementations, the large language model is trained in the following manner: Acquire training samples, which include channel feature label data samples and evaluation result samples corresponding to the channel feature label data samples; The training samples are input into the initial large language model to generate initial evaluation results; Based on the initial evaluation results and the evaluation result samples, determine the loss function value of the initial large language model; If the loss function value satisfies the preset training stopping condition, the initial large language model is determined as the trained large language model. If the loss function value does not meet the preset training stopping condition, adjust the model parameters of the initial large language model, update the initial large language model, and return to the step of inputting the training samples into the initial large language model to generate the initial evaluation result.

[0017] Secondly, this application also provides a channel data evaluation apparatus, the apparatus comprising: An acquisition module is used to acquire initial wireless channel data; the initial wireless channel data is used to measure the quality of wireless communication. The extraction module is used to extract key channel features from the initial wireless channel data to obtain multiple channel features and corresponding initial channel feature values; the channel features include at least one of delay spread, Doppler spread, path loss, and coherence bandwidth. The discretization module is used to discretize the initial channel feature value for each initial channel feature value using a residual vector quantization algorithm to obtain a feature label sequence. The construction module is used to construct an input sequence for large language model recognition based on the feature tag sequence and preset instruction prompts; the instruction prompts are used to indicate the evaluation task to be performed by the large language model. The input module is used to input the input sequence into a pre-trained large language model to obtain the evaluation result of the initial wireless channel data.

[0018] Thirdly, this application also provides a channel data evaluation device, the device comprising: a processor and a memory storing computer program instructions; the processor, when executing the computer program instructions, implements the channel data evaluation method as described in the first aspect above.

[0019] Fourthly, this application also provides a computer-readable storage medium storing computer program instructions that, when executed by a processor, implement the channel data evaluation method as described in the first aspect above.

[0020] Fifthly, this application also provides a computer program product in which instructions, when executed by a processor of an electronic device, enable the electronic device to perform the channel data evaluation method as described in the first aspect above.

[0021] The channel data evaluation method provided in this application, after acquiring initial wireless channel data, firstly extracts key channel features from the initial wireless channel data to obtain multiple channel features and corresponding initial channel feature values; then, for each initial channel feature value, a residual vector quantization algorithm is used to discretize the initial channel feature value to obtain a feature label sequence; then, based on the feature label sequence and preset instruction prompts, an input sequence for large language model recognition is constructed; finally, the input sequence is input into a pre-trained large language model to obtain the evaluation result of the initial wireless channel data. On the one hand, this method uses a residual vector quantization algorithm to discretize continuous features into a labeled sequence with semantic information. By progressively refining the data, the residual vector quantization algorithm improves the accuracy of channel data quantization, solves the codebook collapse problem caused by traditional vector quantization methods, and improves the label utilization rate, thereby ensuring the effective transmission of feature information and improving the accuracy of channel data evaluation results. On the other hand, based on the feature labeled sequence and preset instruction prompts, this method constructs an input sequence for large language model recognition, inputs the input sequence into the large language model, and finally generates natural language evaluation results, solving the problem that traditional numerical results are difficult to understand and improving practicality. Attached Figure Description

[0022] The features, advantages, and technical effects of exemplary embodiments of this application will now be described with reference to the accompanying drawings.

[0023] Figure 1 A flowchart of a channel data evaluation method provided in one embodiment of this application; Figure 2 A flowchart of S130 provided in one embodiment of this application; Figure 3 A schematic diagram of the structure of a channel data evaluation device provided in another embodiment of this application; Figure 4 This is a schematic diagram of the structure of a channel data evaluation device provided in another embodiment of this application. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0025] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the application; the terms “comprising” and “having”, and any variations thereof, in the specification, claims, and foregoing description of the drawings are intended to cover non-exclusive inclusion.

[0026] In the description of the embodiments of this application, technical terms such as "first" and "second" are used only to distinguish different objects and should not be construed as indicating or implying relative importance or implicitly specifying the number, specific order, or primary and secondary relationship of the indicated technical features. In the description of the embodiments of this application, "multiple" means two or more, unless otherwise explicitly defined.

[0027] In this application, the term "embodiment" is used to mean that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described in this application can be combined with other embodiments.

[0028] In the description of the embodiments in this application, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this document generally indicates that the preceding and following related objects have an "or" relationship.

[0029] Existing technologies typically employ Vector Quantization (VQ) to map continuous features of channel data to discrete labels when evaluating channel data. However, traditional VQ uses only one codebook, mapping the input vector to the nearest code vector in the codebook. This leads to codebook collapse (i.e., many vectors in the codebook are rarely used, while only a few are frequently used), thus failing to fully utilize the codebook. This results in low label utilization, reduces the accuracy of channel data quantization, and further reduces the accuracy of channel data evaluation results.

[0030] To address the problems of the prior art, embodiments of the present invention provide a channel data evaluation method, apparatus, device, computer storage medium, and computer program product.

[0031] The following is combined Figure 1 The channel data evaluation method provided in the embodiments of this application will be described in detail.

[0032] Figure 1 A schematic flowchart of a channel data evaluation method provided in an embodiment of the present invention is shown, as follows: Figure 1 As shown, the method may include the following steps: S110, Acquire initial wireless channel data.

[0033] In this embodiment, the initial wireless channel data is used to measure the quality of wireless communication.

[0034] For example, initial wireless channel data can be generated through MATLAB simulation or other simulation systems, enabling the simulation of different scenarios and signal conditions. For instance, different scenarios could be dense cities, open suburbs, or inside a high-speed train carriage; different signal conditions could include signal strength, multipath reflection, noise interference, etc.

[0035] S120. Extract key channel features from the initial wireless channel data to obtain multiple channel features and corresponding initial channel feature values.

[0036] In this embodiment of the application, after acquiring the initial wireless channel data, the initial wireless channel data is preprocessed. For example, the preprocessing may involve noise reduction, segmentation, and normalization of the original data. After processing, key channel features are extracted from the initial wireless channel data. These key channel features are quantifiable indicators representing the core health status of the channel.

[0037] For example, channel characteristics may include at least one of delay spread, Doppler spread, path loss, and coherence bandwidth.

[0038] Delay spread is used to measure the severity of signal echo. Due to multipath effects, signals arrive at the receiver via different paths at varying times. The larger the delay spread value, the more likely it is to cause inter-symbol interference (ISI), where the tail of one symbol interferes with the head of the next, resulting in data errors.

[0039] Doppler spread measures the degree of frequency "drift" or "spread" of a signal caused by the relative motion between the transmitter and receiver. A larger Doppler spread value indicates a faster channel change and makes it more difficult for the receiver to lock onto the signal.

[0040] Path loss measures the degree of attenuation of average signal power during propagation. The higher the path loss value, the worse the signal coverage.

[0041] Coherence bandwidth measures whether signals within a frequency range can pass through a channel without distortion. Coherence bandwidth is closely related to delay spread. The smaller the coherence bandwidth, the greater the difference in the channel's response to different frequency components, which can easily lead to signal distortion.

[0042] Initial channel characteristic values ​​refer to the numerical values ​​corresponding to channel characteristics, which are continuous values ​​used to quantitatively describe a certain dimension of channel characteristics. For example, delay spread = 650 ns, Doppler spread = 120 Hz, path loss = 105 dB, coherence bandwidth = 200 kHz.

[0043] S130. For each initial channel feature value, the residual vector quantization algorithm is used to discretize the initial channel feature value to obtain the feature label sequence.

[0044] For example, discretization (discretization) refers to the process of mapping data that originally belonged to a continuous value range to a predetermined set consisting of a finite number of symbols. In the embodiments of this application, the goal of discretization is to convert continuous numerical values ​​into discrete symbols, which will become the basic input units of a large language model.

[0045] A feature label sequence refers to an ordered combination of discrete labels output after a single initial channel feature value has undergone complete discretization. This sequence serves as a symbolic representation of the feature value and will be used for subsequent sequence construction.

[0046] Residual Vector Quantization (RVQ) is a hierarchical vector quantization technique that uses multiple sequentially connected quantizers to progressively refine and approximate the input vector and its residuals after each quantization stage.

[0047] In this embodiment, residual vector quantization (RVQ) maps continuous features to discrete labels. RVQ progressively quantizes the input vector through multiple stages (multiple codebooks). Each stage processes the residual (i.e., quantization error) of the previous stage, thereby progressively refining the quantization and improving accuracy.

[0048] S140. Based on the feature tag sequence and preset instruction prompts, construct the input sequence for large language model recognition.

[0049] For example, the instruction prompt is used to instruct the large language model on the evaluation task to be performed. The instruction prompt can be a carefully crafted piece of natural language text that clearly informs the large language model of the specific evaluation task to be performed, the expected output format, and the evaluation dimensions to focus on. For example, the instruction prompt could be "Please evaluate the impact of this channel on the 5G NR system".

[0050] An input sequence refers to a data sequence organized according to a specific structure and input into a large language model as a whole. This sequence integrates the objective data to be analyzed with the subjective task instructions that the model needs to execute.

[0051] Constructing the input sequence using the above-described structured approach provides a well-organized and task-oriented input paradigm for large language models. This method forces the model to first understand objective channel feature data and then execute the generation task according to explicit instructions, significantly reducing the randomness and bias of the model output and ensuring that the final generated channel evaluation text meets practical requirements in terms of accuracy and relevance.

[0052] S150. Input the input sequence into the pre-trained large language model to obtain the evaluation results of the initial wireless channel data.

[0053] For example, a Large Language Model (LLM) is a general natural language processing model based on a deep learning architecture that has been pre-trained on massive amounts of text data to master the rules of human language and rich world knowledge. It can understand the semantics of input sequences and generate coherent and relevant natural language text.

[0054] The large language model, based on the input sequence, outputs an evaluation result after undergoing complex semantic understanding and reasoning calculations. This evaluation result is a complete natural language text containing a multi-dimensional analysis of channel quality, typically including descriptions of various feature states, explanations of causal relationships between features, and assessments of the potential impact on the system.

[0055] The channel data evaluation method provided in this application, after acquiring initial wireless channel data, first extracts key channel features from the initial wireless channel data to obtain multiple channel features and corresponding initial channel feature values. Then, for each initial channel feature value, a residual vector quantization algorithm is used to discretize the initial channel feature value, resulting in a feature label sequence. Next, based on the feature label sequence and preset instruction prompts, an input sequence for large language model recognition is constructed. Finally, the input sequence is input into a pre-trained large language model to obtain the evaluation result of the initial wireless channel data. On the one hand, this method uses a residual vector quantization algorithm to discretize continuous features into a label sequence with semantic information. Through progressive refinement, the residual vector quantization algorithm improves the accuracy of channel data quantization, solves the codebook collapse problem caused by traditional vector quantization methods, and improves label utilization, thereby ensuring the effective transmission of feature information and improving the accuracy of channel data evaluation results. On the other hand, this method constructs an input sequence for large language model recognition based on the feature label sequence and preset instruction prompts, inputs the input sequence into the large language model, and finally generates a natural language evaluation result, solving the problem of the difficulty in understanding traditional numerical results and improving practicality.

[0056] In some embodiments, such as Figure 2 As shown, S130 may include: S1301. Determine the initial residual vector based on the initial channel feature values; S1302. Based on the initial residual vector, select the first codebook from multiple pre-trained codebooks; the codebook includes multiple code vectors and the feature label corresponding to each code vector; S1303. Select the first code vector corresponding to the initial residual vector and the first feature tag corresponding to the first code vector from the first codebook; S1304. Based on the initial residual vector and the first code vector, the second residual vector is obtained; S1305. Select the second code vector corresponding to the second residual vector and the second feature tag corresponding to the second code vector from the second codebook; S1306. Based on the second residual vector and the second code vector, the third residual vector is obtained; S1307. Select the third code vector corresponding to the third residual vector and the third feature tag corresponding to the third code vector from the third codebook; S1308. Update the third residual vector to the initial residual vector, update the third codebook to the first codebook, and return the first code vector corresponding to the initial residual vector and the first feature label corresponding to the first code vector selected from the first codebook until the residual vector converges, to obtain the feature label sequence; the feature label sequence includes the first feature label, the second feature label, the third feature label, and the feature labels obtained in each subsequent iteration.

[0057] It should be noted that when using residual vector quantization to map continuous features (such as delay spread values) to discrete labels, the feature space can be divided into multiple subspaces, and each subspace can independently train a codebook. That is, multiple codebooks are pre-trained, each codebook corresponds to a quantization stage, and each codebook contains multiple code vectors and the feature label corresponding to each code vector.

[0058] For example, a label sequence can be generated from an input feature vector z through multi-stage quantization.

[0059]

[0060] in, It is the reconstructed feature vector after RVQ processing, and it is a high-precision approximation of the original input feature vector z. M represents the quantization stage. For feature labeling.

[0061] The specific quantification formula is as follows:

[0062] in, For the first Stage residual vector, It can be distance; the smaller the value, the more similar the two vectors are. For the first The codebook, represented as starting from the first... Find the code vector with the smallest distance in the codebook.

[0063] Taking a delay spread of 650ns as an example, the quantization process is as follows: First, based on the initial channel feature value (650ns), the initial residual vector (650ns) is determined as the input feature vector z.

[0064] Then, based on the initial residual vector (650ns), the first codebook corresponding to the initial residual vector quantization stage is selected from multiple pre-trained codebooks, and the first code vector (e.g., 600ns) corresponding to the initial residual vector and the first feature label (e.g., 15) corresponding to the first code vector are selected from the first codebook.

[0065] Then, based on the initial residual vector and the first code vector, the second residual vector is obtained. For example, the second residual vector = 650 - 600 = 50ns.

[0066] Then, a second code vector (e.g., 40ns) corresponding to the second residual vector (e.g., 50ns) and a second feature tag (e.g., 07) corresponding to the second code vector are selected from the second codebook.

[0067] Then, based on the second residual vector (e.g., 50ns) and the second code vector (e.g., 40ns), the third residual vector is obtained, for example, the third residual vector = 50ns - 40ns = 10ns.

[0068] Select a third code vector (e.g., 12ns) corresponding to the third residual vector (e.g., 10ns) and a third feature tag (e.g., 02) corresponding to the third code vector from the third codebook.

[0069] Then, the third residual vector is updated to the initial residual vector, and the third codebook is updated to the first codebook. The above process is repeated until the residual vector converges.

[0070] For example, residual vector convergence can be achieved when the residual vector is less than a preset threshold. For instance, if a new residual vector is obtained, such as 10-12=-2ns (a very small error), which is less than the preset threshold, then the iterative residual vector quantization process stops.

[0071] Finally, the reconstructed feature vector =600+40+12=652ns, which is very close to the initial input feature vector of 650ns. At the same time, the feature label sequence is obtained. The feature label sequence includes the first feature label, the second feature label, the third feature label, and the feature labels obtained in each subsequent iteration. For example, in the example above, the feature label sequence can be: [15, 07, 02].

[0072] It should be noted that the input feature vector can be a one-dimensional vector or a multi-dimensional vector. The codebook training process may include: collecting channel feature vectors for training; for the first codebook, training is performed using the channel feature vectors for training; for the i-th codebook, where i ranges from 2 to M, training is performed using the residual vector from the (i-1)-th stage.

[0073] By employing residual vector quantization, continuous channel feature values ​​that are difficult to process directly by large language models can be effectively transformed into discrete labeled sequences rich in semantic information. This multi-stage, stepwise approximation quantization method, compared to traditional single-codebook vector quantization, can represent the original features more precisely, significantly improve the utilization efficiency of labels, avoid codebook collapse, and lay a reliable foundation for the accurate understanding and generation of subsequent large language models.

[0074] In some embodiments, selecting the first code vector corresponding to the initial residual from the first codebook may include: Calculate the similarity between the initial residual vector and each code vector in the first codebook; If the similarity satisfies the preset similarity conditions, the code vector that satisfies the preset similarity conditions will be used as the first code vector corresponding to the initial residual vector.

[0075] For example, the similarity between the initial residual vector and each code vector in the first codebook can be represented by the distance between them; as a more concrete example, this could be Euclidean distance. It should be noted that during iterative quantization, the selection of similar code vectors from the codebook can be implemented in the same way.

[0076] By calculating the similarity between the initial residual vector and each code vector in the first codebook, and performing precise screening based on preset similarity conditions, the best code vector that best matches the input features can be determined from the codebook. This ensures the most accurate approximation of the original channel features in the first quantization stage, laying a reliable foundation for the progressive refinement in subsequent stages. This ensures that the final generated discrete label sequence can represent the original channel state with high fidelity.

[0077] In some embodiments, when constructing an input sequence for large language model recognition based on a feature-labeled sequence and preset instruction prompts, preset sequence key position markers can also be obtained; then, based on the feature-labeled sequence, preset instruction prompts, and sequence key position markers, an input sequence for large language model recognition can be constructed.

[0078] For example, the key position markers of the sequence may include: a sequence start marker [CLS] (ClassificationToken, abbreviated as CLS), a sequence separation token [SEP] (Separator Token, abbreviated as SEP), and a sequence end marker [EOS] (End of Sequence Token, abbreviated as EOS); the sequence start marker is used to mark the beginning of the input sequence, the sequence end marker is used to mark the end of the input sequence, and the sequence separation token is used to separate the feature marker sequence and the instruction prompt word.

[0079] It should be noted that when constructing input sequences for large language model recognition, introducing sequence key position markers is an important technical means to ensure that the model accurately understands the input structure and semantic relationships. Sequence key position markers are a set of special markers with specific grammatical functions. They play a key role in structural control and semantic segmentation in the input sequence and are control symbols specifically designed for model sequence processing.

[0080] The sequence start marker [CLS] is placed at the very beginning of the input sequence and clearly identifies the start boundary of a new input sequence. This marker has special significance in the internal processing of the model, and its corresponding hidden state vector is usually designed to gather global semantic information of the entire sequence, providing a comprehensive representation for subsequent sequence-level tasks.

[0081] Sequence segmentation markers (SEPs) are positioned between feature marker sequences and instruction prompts, fulfilling a strict semantic separation function. They establish a clear boundary, isolating objective feature data from subjective task instructions. This mandatory separation effectively prevents large language models from confusing two different types of information sources, ensuring that the model fully understands the channel characteristics before executing the generation task according to the instruction requirements.

[0082] It should be noted that, to support interactive evaluation (such as asking for details), [SEP] can be used to separate historical dialogue from new input, thereby enabling multi-turn dialogue. By forcing LLM to first understand the channel characteristics (left side) and then generate targeted evaluations based on instructions (right side), [SEP] significantly reduces output bias.

[0083] The sequence end marker [EOS] is appended to the end of the input sequence, and its main function is to mark the end position of the input sequence. When the model generates text, this marker also serves as a termination signal for the generation process, ensuring the integrity and appropriate length of the output content and preventing the model from generating redundant or off-topic content.

[0084] In some embodiments, when constructing an input sequence for large language model recognition based on a feature marker sequence, a preset instruction prompt word, and key position markers of the sequence, an input sequence template can be constructed first based on a sequence start marker, a feature marker sequence, a sequence segmentation marker, an instruction prompt word, and a sequence end marker; then, an input sequence for large language model recognition can be constructed based on the input sequence template.

[0085] For example, the input sequence template is arranged in the following order: sequence start marker, feature marker sequence, sequence segmentation marker, instruction prompt word, and sequence end marker. The input sequence template format is as follows: [CLS] Feature marker sequence [SEP] Instruction prompt word [EOS].

[0086] By introducing key positional markers, the input sequence exhibits clear structural features, greatly enhancing the accuracy of the large language model's understanding of the input content's organization. This significantly improves the stability and relevance of the large language model's output, providing a reliable guarantee for generating high-quality, compliant channel evaluation text. Simultaneously, this design also lays a solid structural foundation for possible interactive evaluation extensions, such as multi-turn dialogue scenarios.

[0087] It should be noted that in the example above, the feature label sequence is a string of discrete numbers (indices) obtained by residual vector quantization, such as [15, 07, 02], but its semantics are implicit and unintuitive to the model. The model needs to rely on the knowledge it learns during the fine-tuning phase to understand what 15 represents.

[0088] To further facilitate the recognition of large language models, the feature tag sequence can be transformed into multiple channel feature relationship pairs containing channel features and index attributes based on the relationship between feature tags and preset channel feature index attributes.

[0089] It should be noted that during the training of the residual vector quantization codebook, each code vector (and its index) corresponds to a specific feature value range. Through mapping relationships, the labeled sequence can be mapped back to its represented physical feature name and qualitative attribute. Predefined channel feature index attributes refer to semantic labels that are defined during system design and are user-friendly for both humans and models. These mainly include: channel features such as "delay spread" and "Doppler spread"; and index attributes such as "high," "medium," and "low," or more specific range descriptions. For example, the feature label sequence [15, 07, 02] (representing delay spread) corresponds to the channel feature relationship pair "delay spread: high."

[0090] For example, when constructing an input sequence for large language model recognition based on an input sequence template, the input sequence for large language model recognition can be constructed based on a sequence start marker, channel feature relationship pairs, sequence segmentation markers, instruction prompts, and sequence end markers.

[0091] When constructing the input sequence for the large language model, the original [15, 07, 02] is no longer used directly. Instead, a more semantically meaningful channel feature pair (such as delay spread: high) is used. As an example, the final input sequence is constructed as follows: [Sequence start marker] Delay spread: high, Doppler spread: medium [Sequence segmentation marker] [Instruction prompt] [Sequence end marker].

[0092] Because the semantics of the input signals are more explicit and standardized, the evaluation texts generated by the model at different times and under similar channel conditions will be more consistent and reliable, reducing the randomness of the output results. At the same time, using these semantic "relationship pairs" as input makes the association between training data (feature-evaluation pairs) more direct, and the model can more easily learn the mapping relationship between channel features and natural language evaluations, thereby accelerating convergence and improving the final performance.

[0093] In some embodiments, S150 may include: Obtain the input sequence corresponding to each channel feature; Based on each input sequence, the evaluation result of each channel feature is obtained; The evaluation results of the initial wireless channel data are obtained by combining the evaluation results of multiple channel characteristics.

[0094] For example, each channel feature corresponds to an input sequence. For instance, input sequence A is constructed for delay spread: [CLS] [Delay Spread Marker Sequence] [SEP] Please evaluate the delay spread characteristics of this channel. [EOS]. Input sequence B is constructed for Doppler spread: [CLS] [Doppler Spread Marker Sequence] [SEP] Please evaluate the Doppler spread characteristics of this channel. [EOS]. Input sequence C is constructed for path loss: [CLS] [Path Loss Marker Sequence] [SEP] Please evaluate the path loss characteristics of this channel. [EOS].

[0095] Each independent input sequence constructed in the previous step is fed into the large language model sequentially or in parallel, allowing the model to generate a targeted evaluation text for each individual feature.

[0096] For example, for sequence A (delay spread), LLM might output: "The channel delay spread is extremely high, far exceeding the typical security threshold." For sequence B (Doppler spread), LLM might output: "Moderate Doppler spread was observed, indicating some terminal mobility." For sequence C (path loss), LLM might output: "Path loss is within the normal range and requires no further attention." After obtaining the evaluation results for each channel feature, the evaluation results of multiple channel features are combined, and the sub-evaluation texts of all features are used as new inputs. An instruction such as "Based on the above analyses, please generate a comprehensive channel quality evaluation report" is attached, allowing LLM to extract and integrate information.

[0097] For example, in the comprehensive judgment, it can be carried out by preset logical rules. For instance, as long as the evaluation of a key feature (such as delay extension) is "extremely poor", the tone of the final evaluation will be "poor", and relevant sub-evaluations will be cited as evidence.

[0098] As an example, the final output could be: "Comprehensive evaluation shows that the channel quality is severely degraded. The main problem is extremely high latency spread (far exceeding the safety threshold), which will lead to severe inter-symbol interference and is the primary factor limiting performance. At the same time, there is moderate Doppler spread, which needs to be considered in link adaptation. Path loss is good and is not the current bottleneck. It is recommended to prioritize measures (such as equalizer optimization) to overcome the impact of high latency spread." Large language models analyze each feature, avoiding the feature overload phenomenon that may occur during comprehensive analysis (i.e., an important but easily overlooked feature is masked by other features), ensuring in-depth analysis of each dimension and improving the accuracy and depth of the evaluation.

[0099] To improve the credibility of the evaluation results, in this embodiment, after obtaining the evaluation results of the channel data, feature tags corresponding to the channel features that are related to the evaluation results can be obtained based on the evaluation results of the initial wireless channel data; based on the mapping relationship between the feature tags and the channel feature data interval, the channel feature data interval corresponding to the feature tags is obtained; if the initial channel feature value is within the channel feature data interval, the evaluation result is determined to be verified.

[0100] After obtaining the channel data evaluation results, the natural language evaluation results generated by the large language model are further analyzed to find the key channel features mentioned and their qualitative descriptions. Then, the process is traced back to the discretization stage to find the original feature labels corresponding to these descriptions.

[0101] For example, the evaluation result is that the model outputs "the channel delay spread is extremely high...". Based on the above result, the key information "delay spread: extremely high" is extracted. Then, it is found which feature tag sequence(s) was used to make the model arrive at the conclusion of "delay spread is extremely high". Assuming that this conclusion was triggered by the tag sequence [15, 07, 02], the output is: associated feature tag [15, 07, 02].

[0102] After obtaining the feature labels, the mapping dictionary established in the "residual vector quantization" stage is used to translate the feature labels obtained in the previous step back to their original numerical range. For example, the feature label [15, 07, 02] corresponds to the feature "delay spread," and its quantized numerical range is (500ns, +∞), meaning "greater than 500ns" is defined as "extremely high." The original delay spread value is 650ns. It is determined whether 650ns falls within the range (500ns, +∞). If it does, the evaluation result is verified. If the original delay spread value is 300ns, then 300ns is not within the range (500ns, +∞), and the verification fails. In the case of verification failure, the process can return to the feature discretization step to adjust the codebook partitioning strategy.

[0103] By introducing this automated verification method, discrepancies between the model's textual description and the objective facts of the input data can be prevented, ensuring numerical consistency and guaranteeing the objective authenticity of the evaluation. This makes the output of the entire system credible and reliable. Simultaneously, when verification fails, it provides a trigger signal for an important feedback loop, which can identify potential problems (e.g., unreasonable discretization codebook partitioning or insufficient LLM fine-tuning), thereby allowing for reverse adjustments to previous modules (e.g., "returning to the feature discretization step to adjust the codebook partitioning strategy"), driving the entire system to self-iterate and optimize.

[0104] To further improve the reliability of the evaluation results, in this embodiment, after obtaining the evaluation results, causal relationship pairs in the evaluation results are obtained based on the evaluation results of the initial wireless channel data; the causal relationship pairs are matched with a preset knowledge base to obtain the verification results of the causal relationships; the knowledge base stores the causal relationships between channel characteristics and communication results of different indicator attributes in advance; if the causal relationship represented by the causal relationship pair is consistent with the causal relationship in the knowledge base, it is determined that the evaluation result verification is successful.

[0105] For example, a causal pair refers to a combination of features and results that are causally related, extracted from the evaluation result text. This combination contains a cause item and an effect item. The cause item is usually a certain state of channel features, and the effect item is usually a communication phenomenon or system impact that the state may cause.

[0106] A knowledge base is a specially constructed set of rules that systematically stores causal relationship rules between different channel characteristics and their potential communication outcomes, verified by communication theory. These rules constitute objective standards for verifying the correctness of causal relationships.

[0107] When verifying the correctness of the physical relationship description between features (e.g., "delay spread → inter-symbol interference"), the system first automatically identifies and extracts implicit causal relationship pairs from the natural language evaluation results generated by the large language model. This process typically employs natural language processing techniques, analyzing syntactic structure and specific causal conjunctions to extract structured causal relationship pairs. For example, from the evaluation text "due to high delay spread, significant inter-symbol interference may occur," the causal relationship pairs "high delay spread" and "inter-symbol interference" can be extracted. Secondly, the extracted causal relationship pairs are matched against a pre-set knowledge base. The knowledge base stores professional knowledge in the field of communications, such as verified causal rules like "high delay spread → inter-symbol interference" and "Doppler spread → frequency-selective fading." The system queries the knowledge base to determine if the causal relationship extracted from the text exists within these verified rules. Finally, a verification judgment is made based on the matching results. When the causal logic expressed by the extracted causal relationship pair is completely consistent with a rule stored in the knowledge base, the system determines that the causal relationship verification is successful. If all causal relationships in the evaluation results are verified, then the entire evaluation result is determined to have passed the causal relationship verification.

[0108] Causal relationship verification provides crucial reliability assurance for the application of large language models in specialized fields. By introducing a professional knowledge base as an objective standard, potential technical errors in the model can be effectively identified and filtered, ensuring that the final evaluation results are not only linguistically fluent but, more importantly, technically accurate and reliable. This significantly enhances the professional credibility of the entire channel data evaluation system, laying a solid foundation for its widespread application in practical communication system optimization.

[0109] As an example, a large language model can be trained in the following way: Acquire training samples, which include channel feature label data samples and evaluation result samples corresponding to the channel feature label data samples; The training samples are input into the initial large language model to generate initial evaluation results; Based on the initial evaluation results and the evaluation result samples, determine the loss function value of the initial large language model; If the loss function value meets the preset training stopping condition, the initial large language model is determined as the trained large language model. If the loss function value does not meet the preset training stopping condition, adjust the model parameters of the initial large language model, update the initial large language model, return to input the training samples into the initial large language model, and generate the initial evaluation result.

[0110] By repeatedly showing the standard answer to the large language model and continuously correcting the error between its output and the standard answer, the massive parameters inside the model are gradually adjusted, enabling the model to learn the professional terminology, textual logic, causal relationships, and evaluation scales of channel evaluation. By continuously narrowing the gap between the model output and the expert standard, the accuracy of the model's output is improved.

[0111] As another example, a graph neural network or causal discovery algorithm module can be added between feature extraction and discretization. After extracting key channel features from the initial wireless channel data to obtain multiple channel features and corresponding initial channel feature values, a graph neural network (GNN) is first used to process multiple time-series or spatially related channel features to extract the implicit correlation graph structure between them. The topological information of the correlation graph (such as node importance and edge weights) is also discretized as a type of feature and input into the LLM. When generating evaluations, the LLM can more deeply understand the interactions between features (e.g., "delay spread and Doppler spread are strongly coupled in this scenario, jointly deteriorating frequency selectivity").

[0112] As another example, channel data can be multimodally fused with environmentally perceived data (such as images, point clouds, and maps). During the data acquisition phase, not only is the channel impulse response acquired, but visual or LiDAR data from the surrounding area of ​​the base station / terminal is also simultaneously obtained. A visual encoder (such as ViT) is used to process the environmental images, extracting environmental semantic markers such as "dense buildings," "open areas," and "moving obstacles." These environmental semantic markers are then combined with channel feature markers to construct an input sequence, enabling LLM to generate more insightful assessments such as, "The sudden increase in non-line-of-sight path due to the terminal moving out from behind a building (based on visual markers) causes a sharp rise in latency spread."

[0113] The channel data evaluation method provided in this application, after acquiring initial wireless channel data, first extracts key channel features from the initial wireless channel data to obtain multiple channel features and corresponding initial channel feature values. Then, for each initial channel feature value, a residual vector quantization algorithm is used to discretize the initial channel feature value, resulting in a feature label sequence. Next, based on the feature label sequence and preset instruction prompts, an input sequence for large language model recognition is constructed. Finally, the input sequence is input into a pre-trained large language model to obtain the evaluation result of the initial wireless channel data. On the one hand, this method uses a residual vector quantization algorithm to discretize continuous features into a label sequence with semantic information. Through progressive refinement, the residual vector quantization algorithm improves the accuracy of channel data quantization, solves the codebook collapse problem caused by traditional vector quantization methods, and improves label utilization, thereby ensuring the effective transmission of feature information and improving the accuracy of channel data evaluation results. On the other hand, this method constructs an input sequence for large language model recognition based on the feature label sequence and preset instruction prompts, inputs the input sequence into the large language model, and finally generates a natural language evaluation result, solving the problem of the difficulty in understanding traditional numerical results and improving practicality.

[0114] Based on the channel data evaluation method provided in the above embodiments, this application also provides specific implementations of the channel data evaluation apparatus. Please refer to the following embodiments.

[0115] First see Figure 3 The channel data evaluation apparatus 300 provided in this application embodiment includes: The acquisition module 310 is used to acquire initial wireless channel data; the initial wireless channel data is used to measure the quality of wireless communication. The extraction module 320 is used to extract key channel features from the initial wireless channel data to obtain multiple channel features and corresponding initial channel feature values; the channel features include at least one of delay spread, Doppler spread, path loss, and coherence bandwidth. Discretization module 330 is used to discretize the initial channel feature value for each initial channel feature value using a residual vector quantization algorithm to obtain a feature label sequence; Module 340 is used to construct an input sequence for large language model recognition based on feature label sequences and preset instruction prompts; the instruction prompts are used to indicate the evaluation task to be performed by the large language model. The input module 350 is used to input the input sequence into a pre-trained large language model to obtain the evaluation results of the initial wireless channel data.

[0116] In some possible implementations, the discretization module 330 is also used for: Determine the initial residual vector based on the initial channel feature values; Based on the initial residual vector, the first codebook is selected from multiple pre-trained codebooks; the codebook includes multiple code vectors and the feature label corresponding to each code vector. Select the first code vector corresponding to the initial residual vector and the first feature label corresponding to the first code vector from the first codebook; Based on the initial residual vector and the first code vector, the second residual vector is obtained; Select the second code vector corresponding to the second residual vector and the second feature tag corresponding to the second code vector from the second codebook; Based on the second residual vector and the second code vector, the third residual vector is obtained; Select the third code vector corresponding to the third residual vector and the third feature tag corresponding to the third code vector from the third codebook; The third residual vector is updated to the initial residual vector, the third codebook is updated to the first codebook, and the first code vector corresponding to the initial residual vector and the first feature label corresponding to the first code vector are selected from the first codebook until the residual vector converges, thus obtaining the feature label sequence. The feature label sequence includes the first feature label, the second feature label, the third feature label, and the feature labels obtained in each subsequent iteration.

[0117] In some possible implementations, a first code vector corresponding to the initial residual vector is selected from the first codebook, including: Calculate the similarity between the initial residual vector and each code vector in the first codebook; If the similarity satisfies the preset similarity conditions, the code vector that satisfies the preset similarity conditions will be used as the first code vector corresponding to the initial residual vector.

[0118] In some possible implementations, module 310 is also used for: Obtain the preset key position markers in the sequence; Module 340 is also used for: Based on feature-labeled sequences, pre-defined instruction prompts, and key position markers in the sequence, an input sequence for large language model recognition is constructed.

[0119] In some possible implementations, key position markers for the sequence include: a sequence start marker, a sequence segmentation marker, and a sequence end marker; the sequence start marker is used to mark the beginning of the input sequence, the sequence end marker is used to mark the end of the input sequence, and the sequence segmentation marker is used to segment the feature marker sequence and the instruction prompt word; Module 340 is also used for: An input sequence template is constructed based on the sequence start marker, feature marker sequence, sequence segmentation marker, instruction prompt word, and sequence end marker. Based on the input sequence template, an input sequence for large language model recognition is constructed.

[0120] In some possible implementations, the channel data evaluation device 300 further includes: a conversion module; The conversion module is used to convert a sequence of feature tags into multiple pairs of channel feature relationships containing channel features and index attributes, based on the relationship between feature tags and preset channel feature index attributes. Module 340 is also used for: An input sequence for large language model recognition is constructed based on sequence start markers, channel feature relationship pairs, sequence segmentation markers, instruction prompts, and sequence end markers.

[0121] In some possible implementations, the input sequence template is arranged in the following order: sequence start marker, feature marker sequence, sequence segmentation marker, instruction prompt word, and sequence end marker.

[0122] In some possible implementations, the input module 350 is also used for: Obtain the input sequence corresponding to each channel feature; Based on each input sequence, the evaluation result of each channel feature is obtained; The evaluation results of the initial wireless channel data are obtained by combining the evaluation results of multiple channel characteristics.

[0123] In some possible implementations, module 310 is also used for: Based on the evaluation results of the initial wireless channel data, feature labels corresponding to channel features that are related to the evaluation results are obtained; Based on the mapping relationship between feature markers and channel feature data intervals, the channel feature data intervals corresponding to the feature markers are obtained; If the initial channel characteristic values ​​are within the channel characteristic data range, the evaluation result is verified.

[0124] In some possible implementations, module 310 is also used for: Based on the evaluation results of the initial wireless channel data, obtain the causal relationship pairs in the evaluation results; The causal relationship pairs are matched with a pre-set knowledge base to obtain the verification results of the causal relationship; the knowledge base pre-stores the causal relationships between channel characteristics and communication results of different indicator attributes. If the causal relationship expressed by the causal relationship pair is consistent with the causal relationship in the knowledge base, the evaluation result is deemed to have passed verification.

[0125] In some possible implementations, large language models are trained in the following ways: Acquire training samples, which include channel feature label data samples and evaluation result samples corresponding to the channel feature label data samples; The training samples are input into the initial large language model to generate initial evaluation results; Based on the initial evaluation results and the evaluation result samples, determine the loss function value of the initial large language model; If the loss function value meets the preset training stopping condition, the initial large language model is determined as the trained large language model. If the loss function value does not meet the preset training stopping condition, adjust the model parameters of the initial large language model, update the initial large language model, return to input the training samples into the initial large language model, and generate the initial evaluation result.

[0126] The various modules of the channel data evaluation device provided in this application embodiment can achieve Figure 1 It provides the functionality for each step of the channel data evaluation method and achieves the corresponding technical effects. For the sake of brevity, it will not be elaborated here.

[0127] See Figure 4 The channel data evaluation method described in the above embodiments can be further described in this embodiment of the invention as a channel data evaluation device 400, which includes a processor 410 and a memory 420 storing computer program instructions; the processor 410 executes the computer program instructions to implement any of the channel data evaluation methods described in the above embodiments.

[0128] The channel data evaluation method in the above embodiments can be implemented using a computer storage medium. This computer storage medium stores computer program instructions; when these instructions are executed by a processor, they implement any of the channel data evaluation methods described in the above embodiments.

[0129] This application also provides a computer program product, including a computer program that, when executed, implements any of the channel data evaluation methods described in the above embodiments.

[0130] It should be clarified that this application is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of this application is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of this application.

[0131] The functional blocks shown in the above block diagram can be implemented as hardware, software, firmware, or a combination thereof. When implemented in hardware, they can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this application are programs or code segments used to perform the required tasks. Programs or code segments can be stored on a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried on a carrier wave. "Machine-readable medium" can include any medium capable of storing or transmitting information. Examples of machine-readable media include electronic circuits, semiconductor memory devices, ROM, flash memory, erasable ROM (EROM), floppy disks, CD-ROMs, optical disks, hard disks, fiber optic media, radio frequency (RF) links, etc. Code segments can be downloaded via computer networks such as the Internet, intranets, etc.

[0132] It should also be noted that the exemplary embodiments mentioned in this application describe methods or systems based on a series of steps or apparatus. However, this application is not limited to the order of the above steps; that is, the steps can be performed in the order mentioned in the embodiments, or in a different order, or several steps can be performed simultaneously.

[0133] The foregoing flowcharts and / or block diagrams describing the method for determining the open-circuit voltage of a battery, the battery management system, and the power-consuming device according to embodiments of this application have described various aspects of this application. It should be understood that each block in the flowcharts and / or block diagrams, and combinations of blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to create a machine such that these instructions, executable by the processor of the computer or other programmable data processing apparatus, enable the implementation of the functions / actions specified in one or more blocks of the flowcharts and / or block diagrams. Such a processor can be, but is not limited to, a general-purpose processor, a special-purpose processor, a special application processor, or a field-programmable logic circuit. It is also understood that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can also be implemented by dedicated hardware performing the specified functions or actions, or can be implemented by a combination of dedicated hardware and computer instructions.

[0134] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and not to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. These modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application, and they should all be covered within the scope of the claims and specification of this application. In particular, as long as there is no structural conflict, the various technical features mentioned in the embodiments can be combined in any way. This application is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A channel data evaluation method, characterized in that, include: Acquire initial wireless channel data; The initial wireless channel data is used to measure the quality of wireless communication; Key channel features are extracted from the initial wireless channel data to obtain multiple channel features and corresponding initial channel feature values; the channel features include at least one of delay spread, Doppler spread, path loss, and coherence bandwidth. For each initial channel feature value, the residual vector quantization algorithm is used to discretize the initial channel feature value to obtain a feature label sequence; Based on the feature tag sequence and preset instruction prompts, an input sequence for large language model recognition is constructed; The instruction prompts are used to indicate the evaluation task to be performed by the large language model; The input sequence is fed into a pre-trained large language model to obtain the evaluation results of the initial wireless channel data.

2. The method according to claim 1, characterized in that, The residual vector quantization algorithm is used to discretize the channel feature values ​​to obtain a feature label sequence, including: Based on the initial channel feature values, determine the initial residual vector; Based on the initial residual vector, a first codebook is selected from a plurality of pre-trained codebooks; the codebook includes a plurality of code vectors and a feature label corresponding to each code vector. Select a first code vector corresponding to the initial residual vector and a first feature tag corresponding to the first code vector from the first codebook; Based on the initial residual vector and the first code vector, a second residual vector is obtained; Select the second code vector corresponding to the second residual vector and the second feature tag corresponding to the second code vector from the second codebook; Based on the second residual vector and the second code vector, the third residual vector is obtained; Select the third code vector corresponding to the third residual vector and the third feature tag corresponding to the third code vector from the third codebook; The third residual vector is updated to the initial residual vector, the third codebook is updated to the first codebook, and the process of selecting the first code vector corresponding to the initial residual vector and the first feature tag corresponding to the first code vector from the first codebook is repeated until the residual vector converges, resulting in a feature tag sequence. The feature tag sequence includes the first feature tag, the second feature tag, the third feature tag, and the feature tags obtained in each subsequent iteration.

3. The method according to claim 2, characterized in that, The step of selecting the first code vector corresponding to the initial residual vector from the first codebook includes: Calculate the similarity between the initial residual vector and each code vector in the first codebook; If the similarity satisfies the preset similarity conditions, the code vector that satisfies the preset similarity conditions will be used as the first code vector corresponding to the initial residual vector.

4. The method according to claim 1, characterized in that, The method further includes: Obtain the preset key position markers in the sequence; The step of constructing an input sequence for large language model recognition based on the feature marker sequence and preset instruction prompts includes: Based on the feature tag sequence, preset instruction prompts, and key position tags of the sequence, an input sequence for large language model recognition is constructed.

5. The method according to claim 4, characterized in that, The sequence key position markers include: a sequence start marker, a sequence segmentation marker, and a sequence end marker; the sequence start marker is used to mark the beginning of the input sequence, the sequence end marker is used to mark the end of the input sequence, and the sequence segmentation marker is used to segment the feature marker sequence and the instruction prompt word; The step of constructing an input sequence for large language model recognition based on the feature-labeled sequence, preset instruction prompts, and key position markers of the sequence includes: An input sequence template is constructed based on the sequence start marker, the feature marker sequence, the sequence segmentation marker, the instruction prompt word, and the sequence end marker; Based on the input sequence template, an input sequence for large language model recognition is constructed.

6. The method according to claim 5, characterized in that, The method further includes: Based on the relationship between feature tags and preset channel feature index attributes, the feature tag sequence is converted into multiple channel feature relationship pairs containing channel features and index attributes; The step of constructing an input sequence for large language model recognition based on the input sequence template includes: Based on the sequence start marker, the channel feature relationship pair, the sequence segmentation marker, the instruction prompt word, and the sequence end marker, an input sequence for large language model recognition is constructed.

7. The method according to claim 5, characterized in that, The input sequence template is arranged in the following order: sequence start marker, feature marker sequence, sequence segmentation marker, instruction prompt word, and sequence end marker.

8. The method according to claim 1, characterized in that, The step of inputting the input sequence into a pre-trained large language model to obtain the evaluation result of the initial wireless channel data includes: Obtain the input sequence corresponding to each channel feature; Based on each input sequence, the evaluation result of each channel feature is obtained; The evaluation results of the initial wireless channel data are obtained by combining the evaluation results of multiple channel characteristics.

9. The method according to any one of claims 1-8, characterized in that, The method further includes: Based on the evaluation results of the initial wireless channel data, feature tags corresponding to channel features that are related to the evaluation results are obtained; Based on the mapping relationship between the feature marker and the channel feature data interval, the channel feature data interval corresponding to the feature marker is obtained; Given that the initial channel characteristic values ​​are within the channel characteristic data range, the evaluation result is deemed to have passed verification.

10. The method according to any one of claims 1-8, characterized in that, The method further includes: Based on the evaluation results of the initial wireless channel data, obtain the causal relationship pairs in the evaluation results; The causal relationship pairs are matched with a preset knowledge base to obtain the verification result of the causal relationship; the knowledge base stores causal relationships between channel characteristics and communication results of different indicator attributes in advance. If the causal relationship represented by the causal relationship pair is consistent with the causal relationship in the knowledge base, the evaluation result is determined to be verified.

11. The method according to any one of claims 1-8, characterized in that, The large language model is trained in the following manner: Acquire training samples, which include channel feature label data samples and evaluation result samples corresponding to the channel feature label data samples; The training samples are input into the initial large language model to generate initial evaluation results; Based on the initial evaluation results and the evaluation result samples, determine the loss function value of the initial large language model; If the loss function value satisfies the preset training stopping condition, the initial large language model is determined as the trained large language model. If the loss function value does not meet the preset training stopping condition, adjust the model parameters of the initial large language model, update the initial large language model, and return to the step of inputting the training samples into the initial large language model to generate the initial evaluation result.

12. A channel data evaluation device, characterized in that, include: The acquisition module is used to acquire initial wireless channel data; The initial wireless channel data is used to measure the quality of wireless communication; The extraction module is used to extract key channel features from the initial wireless channel data to obtain multiple channel features and corresponding initial channel feature values; the channel features include at least one of delay spread, Doppler spread, path loss, and coherence bandwidth. The discretization module is used to discretize the initial channel feature value for each initial channel feature value using a residual vector quantization algorithm to obtain a feature label sequence. The construction module is used to construct an input sequence for large language model recognition based on the feature tag sequence and preset instruction prompts; The instruction prompts are used to indicate the evaluation task to be performed by the large language model; The input module is used to input the input sequence into a pre-trained large language model to obtain the evaluation result of the initial wireless channel data.

13. A channel data evaluation device, characterized in that, The device includes: a processor and a memory storing computer program instructions; the processor, when executing the computer program instructions, implements the channel data evaluation method as described in any one of claims 1-11.

14. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer program instructions, which, when executed by a processor, implement the channel data evaluation method as described in any one of claims 1-11.

15. A computer program product, characterized in that, When the instructions in the computer program product are executed by the processor of the electronic device, the electronic device is able to perform the channel data evaluation method as described in any one of claims 1-11.

Images