Cross-domain time series data quality scoring method and device based on large language model

Abstract

The invention discloses a cross-domain time series data quality scoring method and device based on a large language model, and solves the technical problem that an existing time series data quality scoring method faces extremely high calculation cost in practice. The method comprises the following steps: acquiring a cross-domain original time sequence sample, and dividing through a sliding window mechanism to generate a plurality of overlapped data blocks; secondly, on the basis of a majority voting mode, a large language model is combined with a predefined cue word template to make preference judgment on the overlapped data blocks, and a plurality of pairwise preference probabilities are output; then, mapping pairwise preference probabilities by using a Bradell-trie model to obtain continuous quality scores of the overlapped data blocks; training the initial scoring model by using a cross entropy loss function, and determining a target scoring model; enhancing the model through a symbol stochastic gradient descent algorithm and a back propagation method to obtain an enhanced scoring model; and finally, scoring each overlapped data block by the enhanced scoring model, and outputting a target scoring result.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a cross-domain time-series data quality scoring method and apparatus based on a large language model. Background Technology

[0002] With the development of artificial intelligence technology, time series data has become one of the key data types driving modern intelligent systems. In many high-value application scenarios such as healthcare, financial transactions, industrial manufacturing, and traffic management, time series data is widely used for tasks such as monitoring, prediction, control, and decision support. For example, physiological sensors monitor patients' vital signs to generate medical time series data; stock trading records constitute high-frequency financial sequences; power plant equipment sensors continuously report time series data such as temperature, voltage, and current; and intelligent transportation systems collect and update urban road network traffic status in real time. These data exhibit significant temporal structure, meaning they possess strong time dependencies, non-stationarity, trend evolution characteristics, and periodic changes, and are often long, complex sequences with multidimensional variables.

[0003] However, the quality of the actual time-series data collected varies greatly and is often affected by factors such as the reliability of data acquisition equipment, network transmission stability, environmental interference, and human operation, resulting in a variety of quality problems. Common problems include: (1) missing values ​​and outliers, such as loss or sudden changes in measurement values ​​caused by sensor failure or signal interruption; (2) inconsistent sampling, such as different equipment using different sampling frequencies, or intermittent sampling due to energy-saving strategies, which affects the model's ability to model trends; (3) signal drift and equipment aging effects, which cause long-term deviations in measurement values; (4) task-irrelevant fragments mixed in, such as equipment commissioning data mixed in with some industrial data that cannot represent stable operating conditions; (5) poor label quality or serious missing labels, which makes supervised modeling face difficulties.

[0004] During the training of deep models (such as Transformers or time-series foundational models), low-quality samples can lead to a series of problems, including noisy learning, overfitting, difficulty in model convergence, and unstable performance. Therefore, assessing data quality and selecting high-quality samples before training is widely recognized as a crucial step in improving model accuracy, robustness, and training efficiency. However, techniques for assessing the quality of time-series data are still in their early stages of development. Achieving an accurate, efficient, and scalable quality assessment mechanism in situations with complex time structures, heterogeneous domain contexts, and a lack of labeled data remains a significant technical challenge.

[0005] Existing time-series data quality scoring methods, such as Shapley Value and Influence Function, typically rely on resampling or perturbing each sample and then retraining the model to assess its marginal impact on overall performance. While this approach is theoretically fair, in scenarios where time-series data has high dimensionality, strong temporal dependencies, and complex structures, the required Hessian matrix approximation and backpropagation calculations cause training time to increase exponentially, resulting in extremely high computational costs in practice. Summary of the Invention

[0006] This invention provides a cross-domain time-series data quality scoring method and apparatus based on a large language model, which solves the technical problem that the Hessian matrix approximation and backpropagation calculation required by existing time-series data quality scoring methods cause the training time to increase exponentially, resulting in extremely high computational costs in practice.

[0007] The first aspect of this invention provides a cross-domain time-series data quality scoring method based on a large language model, comprising:

[0008] Obtain cross-domain original time series samples, and divide the cross-domain original time series samples based on a sliding window mechanism to generate multiple overlapping data blocks;

[0009] Based on the majority voting method, a large language model is used to perform pairwise preference discrimination according to the predefined prompt word template and each overlapping data block, and output multiple pairwise preference probabilities.

[0010] The Bradley-Terry model is used to map the pairwise preference probabilities of each pairwise data block, and the continuous quality score of each overlapping data block is output.

[0011] The initial scoring model is trained using the cross-entropy loss function based on the continuous quality scores of each overlapping data block to determine the target scoring model;

[0012] The target scoring model is augmented using the signed stochastic gradient descent algorithm and backpropagation method to determine the augmented scoring model.

[0013] The enhanced scoring model is used to score each of the overlapping data blocks, and the target score result is output.

[0014] Optionally, the method based on majority voting employs a large language model to perform pairwise preference discrimination based on predefined prompt word templates and each overlapping data block, outputting multiple pairwise preference probabilities, including:

[0015] Select any two overlapping data blocks from the plurality of overlapping data blocks and combine them to output a plurality of pairs of overlapping data blocks;

[0016] The large language model is used to perform quality comparison on each pair of overlapping data blocks according to the predefined prompt word template, and outputs pair preference labels for each pair of overlapping data blocks;

[0017] The proportion of pairwise preference labels for each of the overlapping data blocks is calculated using the majority voting method to determine the pairwise preference probability of each of the overlapping data blocks.

[0018] Optionally, the step of mapping the pairwise preference probabilities using the Bradley-Terry model to output continuous quality scores for each of the overlapping data blocks includes:

[0019] A mapping relationship is constructed based on the Bradley-Terry model;

[0020] With maximum likelihood estimation as the objective, an optimization objective function is constructed by combining the aforementioned mapping relationship;

[0021] The optimization objective function is used to solve for each pairwise preference probability, and the scores of each overlapping data block in multiple dimensions are output.

[0022] The scores of each overlapping data block are averaged across multiple dimensions to determine the continuous quality score of each overlapping data block.

[0023] Optionally, the step of training the initial scoring model using the cross-entropy loss function based on the continuous quality scores of each of the overlapping data blocks to determine the target scoring model includes:

[0024] The representation extractor in the initial scoring model is used to extract features from each of the overlapping data blocks to obtain the feature vectors corresponding to each of the overlapping data blocks;

[0025] The feature vectors corresponding to each overlapping data block are mapped using the quality mapper in the initial scoring model, and the scalar quality score corresponding to each overlapping data block is output.

[0026] The cross-entropy loss function is used to train the quality mapper in the initial scoring model based on the scalar quality score and continuous quality score corresponding to each of the overlapping data blocks, thereby determining the target scoring model.

[0027] Optionally, the step of enhancing the target scoring model based on the signed stochastic gradient descent algorithm and backpropagation to determine the enhanced scoring model includes:

[0028] Build a set of tasks;

[0029] The target scoring model is trained using the symbolic stochastic gradient descent algorithm and the backpropagation method based on the task set to determine the enhanced scoring model.

[0030] Optionally, the target scoring result includes point-level scoring and sample scoring; the step of using the enhanced scoring model to score each of the overlapping data blocks and outputting the target scoring result includes:

[0031] The enhanced scoring model is used to score each of the overlapping data blocks, and the target score corresponding to each of the overlapping data blocks is output.

[0032] Based on the target score corresponding to each of the overlapping data blocks, point-level score mapping and sample-level score mapping are performed respectively, and point-level score and sample score are output.

[0033] The second aspect of this invention provides a cross-domain time-series data quality scoring device based on a large language model, comprising:

[0034] The acquisition module is used to acquire cross-domain original time series samples and divide the cross-domain original time series samples based on a sliding window mechanism to generate multiple overlapping data blocks;

[0035] The discrimination module is used to perform pairwise preference discrimination based on the majority voting method and a large language model according to the predefined prompt word template and each overlapping data block, and output multiple pairwise preference probabilities.

[0036] The mapping module is used to map the pairwise preference probabilities of each data block using the Bradley-Terry model and output the continuous quality score of each data block.

[0037] The training module is used to train the initial scoring model using the cross-entropy loss function based on the continuous quality scores of each of the overlapping data blocks, and to determine the target scoring model.

[0038] The enhancement module is used to enhance the target scoring model based on the symbolic stochastic gradient descent algorithm and the backpropagation method, and to determine the enhanced scoring model;

[0039] The scoring module is used to score each of the overlapping data blocks using the enhanced scoring model and output the target scoring result.

[0040] A third aspect of the present invention provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the cross-domain time-series data quality scoring method based on a large language model as described above.

[0041] The fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed, implements the cross-domain time-series data quality scoring method based on a large language model as described above.

[0042] The fifth aspect of the present invention provides a computer program product, the computer program product comprising a computer program stored on a non-transitory computer-readable storage medium, the computer program comprising program instructions, wherein, when the program instructions are executed by a computer, the computer performs the cross-domain time-series data quality scoring method based on a large language model as described above.

[0043] As can be seen from the above technical solutions, the present invention has the following advantages:

[0044] The technical solution of this invention provides a cross-domain time-series data quality scoring method based on a large language model. It acquires original cross-domain time-series samples and divides them using a sliding window mechanism to generate multiple overlapping data blocks. Based on majority voting, a large language model is used to perform pairwise preference discrimination according to predefined prompt word templates and each overlapping data block, outputting multiple pairwise preference probabilities. A Bradley-Terry model is used to map each pairwise preference probability, outputting a continuous quality score for each overlapping data block. A cross-entropy loss function is used to evaluate the initial score based on the continuous quality scores of each overlapping data block. The target scoring model is determined by training the sub-models. The target scoring model is then augmented using the signed stochastic gradient descent algorithm and backpropagation to determine the augmented scoring model. The augmented scoring model is used to score each overlapping data block and output the target scoring result. Based on the above scheme, this invention introduces a large language model (LLM) for pairwise preference discrimination, combines the lightweight analytical mapping of the Bradley-Terry model with the signed stochastic gradient descent (signSGD) algorithm, and achieves fast scoring without retraining and high-order gradient calculation, which greatly reduces the computational overhead and is suitable for deployment in large-scale data systems. Attached Figure Description

[0045] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 The flowchart illustrates the steps of a cross-domain time-series data quality scoring method based on a large language model, as provided in Embodiment 1 of the present invention.

[0047] Figure 2This is an overall framework diagram of a cross-domain time series data quality scoring method based on a large language model, provided in Embodiment 1 of the present invention.

[0048] Figure 3 This is a structural block diagram of a cross-domain time-series data quality scoring device based on a large language model, provided in Embodiment 2 of the present invention. Detailed Implementation

[0049] This invention provides a method and apparatus for cross-domain time-series data quality scoring based on a large language model. It solves the technical problem that the Hessian matrix approximation and backpropagation calculation required by existing time-series data quality scoring methods cause the training time to increase exponentially, resulting in extremely high computational costs in practice.

[0050] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. It should be noted that in the optional embodiments of the present invention, the object information and other related data involved require the permission or consent of the object when the embodiments of the present invention are applied to specific products or technologies, and the collection, use, and processing of related data must comply with the relevant laws, regulations, and standards of the relevant countries and regions. That is to say, if the embodiments of the present invention involve data related to the object, it needs to be obtained with the authorization and consent of the object, the authorization and consent of the relevant departments, and in compliance with the relevant laws, regulations, and standards of the country and region. If personal information is involved in the embodiments, the acquisition of all personal information requires the consent of the individual. If sensitive information is involved, the separate consent of the information subject is required, and the embodiments also need to be implemented with the authorization and consent of the object.

[0051] Terminology Explanation:

[0052] Large Language Models (LLMs) refer to pre-trained language models based on the Transformer architecture with billions or even hundreds of billions of parameters, such as GPT and LLaMA. After being pre-trained on large-scale corpora, they possess abilities such as understanding, generation, reasoning, and judgment, and are widely used in tasks such as question answering, writing, and code generation.

[0053] Meta-learning, also known as learning how to learn, is a learning method aimed at learning general policies from multiple tasks to achieve rapid adaptation to new tasks. Typical methods include MAML (Model Independent Meta-learning), which is often used for few-shot learning and task generalization.

[0054] Time series data refers to data sequences observed continuously over time, and is widely found in fields such as finance, healthcare, industry, and transportation. Its essential characteristics are time dependence, periodicity, and trend changes.

[0055] Pairwise comparison: A relative evaluation strategy that obtains the preference order by having the model judge which sample A or sample B is better, avoiding the subjectivity and bias of subjective scoring.

[0056] The Bradley-Terry model is a statistical model that transforms pairwise comparisons (e.g., A is better than B) into continuous scores. It is widely used in preference modeling, ranking, and athletic ranking scenarios.

[0057] Signed stochastic gradient descent (signSGD) is a lightweight stochastic gradient descent method that uses gradient signs for optimization. It has low computational cost, is suitable for inner loop optimization in meta-learning, and avoids the high complexity of second-order gradient calculation.

[0058] Please see Figure 1 , Figure 1 The flowchart illustrates the steps of a cross-domain time-series data quality scoring method based on a large language model, as provided in Embodiment 1 of the present invention.

[0059] This invention provides a cross-domain time-series data quality scoring method based on a large language model, comprising:

[0060] Step 101: Obtain cross-domain original time series samples and divide the cross-domain original time series samples based on the sliding window mechanism to generate multiple overlapping data blocks.

[0061] Cross-domain raw time series samples refer to raw data sequences obtained from different business fields such as finance, healthcare, industry, and transportation, which are continuously observed in the time dimension. They are usually time series data sets containing multi-channel measurements and corresponding labels, and have typical time series characteristics such as time dependence, periodicity, and trend changes.

[0062] The sliding window mechanism refers to a time series data segmentation method. It is a processing strategy that splits a complete long time series sequence into multiple local data segments by setting a fixed window length and sliding step size and performing sliding truncation on the original time series samples.

[0063] Overlapping data blocks refer to local time-series data segments obtained after being segmented using a sliding window mechanism. Because the window sliding step size is usually smaller than the window length, there are overlapping parts in time segments between different data blocks, and the local time dependency structure of the original data is preserved.

[0064] It should be noted that the form of the original time series sample across domains is defined as follows:

[0065] ;

[0066] in, Let t be the t-th cross-domain original time series sample; T represents the time step of the t-th cross-domain original time series sample; This represents the measurement value at the l-th time point in the original cross-domain time series sample. These are the corresponding labels; D represents the number of measurement channels, and C represents the dimension of the label. This invention performs sliding window segmentation on each cross-domain original time-series sample by setting the window length and sliding step size, generating multiple local data blocks. (i.e., overlapping data blocks). This step enhances the model's ability to identify local quality differences in temporal structures, while effectively avoiding the problem of limited input context length in Large Language Models (LLMs), ensuring that prompt words and data content can be encoded and understood simultaneously, thereby improving the accuracy and stability of quality judgment.

[0067] In this embodiment, a sliding window length and sliding step size are set for the cross-domain original time series samples to adapt to the input length of the large language model. Then, the time series samples are truncated by sliding the window length according to the step size. Each data block obtained contains continuous time series segments. Since the step size is smaller than the window length, there is overlap of time segments between the data blocks. This preserves the local time dependency structure of the original samples. These overlapping data blocks will serve as the basic units for subsequent pairwise quality comparison.

[0068] Step 102: Based on the majority voting method, a large language model is used to perform pairwise preference discrimination according to the predefined prompt word template and each overlapping data block, and output multiple pairwise preference probabilities.

[0069] It should be noted that all pairwise combinations are extracted from multiple overlapping data blocks. For each pair of data block combinations (i.e., paired overlapping data blocks), four types of predefined prompt word templates (trend, frequency, amplitude, and pattern) are used to guide the large language model to judge the quality. In order to improve the robustness of the judgment, each pair of data blocks is judged M times in different order and with random seeds. Then, the proportion of the same preference tendency in the M judgment results is counted by majority voting. The pairwise preference probability corresponding to each pair of overlapping data block combinations is calculated in this way.

[0070] Furthermore, step 102 may include the following sub-steps:

[0071] S21. Select any two overlapping data blocks from multiple overlapping data blocks and combine them to output multiple pairs of overlapping data blocks.

[0072] S22. Using a large language model, the quality of each pair of overlapping data blocks is compared based on a predefined prompt word template, and the pair preference labels of each pair of overlapping data blocks are output.

[0073] S23. Calculate the proportion of pairwise preference labels for each overlapping data block using a majority voting method to determine the pairwise preference probability for each pairwise overlapping data block.

[0074] Predefined prompt word templates refer to natural language instruction templates designed in advance to guide large language models to focus on specific quality dimensions of time series data for discrimination. Specifically, they include four categories: trend, frequency, amplitude, and pattern, which correspond to quality assessment dimensions such as long-term structural stability, periodicity, volatility, and typical interpretable form of data blocks.

[0075] It should be noted that by constructing four types of cue word templates (trend, frequency, amplitude, and pattern), a large language model (such as GPT-4o-mini) is invoked to analyze any two data blocks. and Perform pairwise comparisons and output pairwise preference labels for overlapping data blocks. :

[0076] ;

[0077] To improve the robustness of the judgment, each pair of overlapping data blocks will be repeatedly queried with a random seed in multiple orders, and the confidence preference score (i.e., pairwise preference probability) will be obtained through majority voting, providing highly reliable score pairs for subsequent numerical modeling.

[0078] ;

[0079] in, Let represent the i-th data block in a pair of overlapping data blocks (denoted as ). ) relative to the j-th data block (denoted as A higher-quality confidence preference score (i.e., pairwise preference probability) is obtained after multiple evaluations of the pair of data blocks. The confidence quantification results of the tendency to "be better"; The binary preference label represents the output of the k-th repeated discrimination between the i-th data block and the j-th data block; The number of comparisons represents the total number of times the large language model is repeatedly queried for quality judgment on the same set of overlapping data blocks in different orders and with different random seeds. It is the number of repeated operations to improve the robustness of the judgment.

[0080] It is worth mentioning that the detailed design of the four quality scoring criteria is as follows:

[0081] Trend: Determines whether a data block exhibits a stable upward, downward, or flat trend, reflecting long-term structural information.

[0082] Frequency: Assess whether periodicity and repetitive patterns are obvious, and measure the rhythm and regularity of a time series.

[0083] Amplitude: Analyze whether the data fluctuations are drastic or abnormal, and pay attention to quality risks such as abnormal peaks and sharp jumps.

[0084] Pattern: Determine whether the data has a typical and interpretable structural form, such as a mixture of cycles and trends, or a clear local pattern.

[0085] Each criterion is accompanied by a corresponding natural language prompt template, guiding the LLM to focus on the specified quality dimension for judgment. Finally, the preference results of the four criteria are integrated into an overall preference label, which is used to train the downstream scoring model (initial scoring model).

[0086] In this embodiment, any two overlapping data blocks are selected from multiple overlapping data blocks and combined to output multiple pairs of overlapping data blocks. A large language model is used to compare the quality of each pair of overlapping data blocks according to a predefined prompt word template, and outputs the pair preference labels of each pair of overlapping data blocks. The predefined prompt word template specifically covers four time series data quality assessment dimensions: trend, frequency, amplitude, and pattern. In order to improve the robustness of the judgment, for each pair of overlapping data blocks, the large language model is queried M times with different data block orders and different random seeds, and the corresponding pair preference labels are output each time. The proportion of these M pair preference labels corresponding to each pair of overlapping data blocks is calculated by majority voting, that is, the number of labels representing the same preference tendency is counted and divided by the total number of times M, so as to determine the pair preference probability corresponding to each pair of overlapping data blocks.

[0087] Step 103: Use the Bradley-Terry model to map the pairwise preference probabilities and output the continuous quality scores of each overlapping data block.

[0088] It should be noted that the mapping relationship between the pairwise preference probability and the difference in data block quality score is constructed based on the Bradley-Terry model, i.e., the pairwise preference probability. The result is equal to the difference in quality scores between the i-th and j-th overlapping data blocks when the Sigmoid function is applied. Then, the corresponding optimization objective function is constructed with maximum likelihood estimation as the objective. The optimization objective function is solved according to the pairwise preference probabilities to obtain the quality score of each overlapping data block under each measurement channel. Finally, the average of the scores of multiple channels is taken to determine and output the continuous quality score of each overlapping data block.

[0089] Furthermore, step 103 may include the following sub-steps:

[0090] S31. Construct mapping relationships based on the Bradley-Terry model;

[0091] S32. With maximum likelihood estimation as the objective, construct an optimization objective function by combining mapping relationships;

[0092] S33. The objective function is optimized and solved according to the probability of each pair of preferences, and the scores of each overlapping data block in multiple dimensions are output.

[0093] S34. Average the scores of each overlapping data block across multiple dimensions to determine the continuous quality score of each overlapping data block.

[0094] The mapping relationship refers to the association between the pairwise preference probability and the quality score difference of the pairwise overlapping data blocks, which is constructed based on the Bradley-Terry model. Specifically, the pairwise preference probability is equal to the result of the Sigmoid function acting on the quality score difference of the pair of data blocks, realizing the transformation from discrete preference probability to continuous score difference.

[0095] Multiple dimensions refer to multiple measurement channels that actually correspond to the original time series samples across different domains. Each channel corresponds to a type of observation index of the time series data, and each overlapping data block will have its score calculated separately under each channel.

[0096] Continuous quality score refers to a continuous value obtained by averaging the scores of overlapping data blocks across multiple measurement channels. It is used to quantify the overall quality level of a single overlapping data block and serves as a supervisory signal for subsequent training of the scoring model.

[0097] It should be noted that this invention introduces the Bradley-Terry model to measure pairwise preference probabilities. This is mapped to a scalar quality score (i.e., a continuous quality score). The model assumes blocks rating and rating The difference between them can be represented by the Sigmoid function to indicate the preference probability:

[0098] ;

[0099] With maximum likelihood estimation as the objective, the following optimization objective function is constructed:

[0100] ;

[0101] in, The objective function for optimization, which aims at maximum likelihood estimation, is a likelihood function constructed based on the preference probabilities of all pairs of overlapping data blocks, used to solve for the continuous quality score of each overlapping data block. It is the set of all “paired overlapping data blocks + corresponding paired preference probabilities”.

[0102] For multi-channel data, the scoring process is performed independently on each dimension and then averaged to obtain a global score (continuous quality score):

[0103] ;

[0104] in, The global score (i.e., continuous quality score) for the i-th overlapping data block is the comprehensive quality quantification result obtained by averaging the scores of the data block across all measurement channels. The independent quality score of the i-th overlapping data block in the d-th measurement channel (i.e., the score of overlapping data block i in multiple dimensions d) is the result obtained after performing the scoring process separately in that channel dimension; This refers to the total number of measurement channels in the original time-series samples across different domains, i.e., the number of observation indicator categories included in the multi-channel data. This mechanism maintains sensitivity to quality differences among channels and enhances the expressive power of the scoring.

[0105] In this embodiment, a mapping relationship is constructed based on the Bradley-Terry model. Specifically, the pairwise preference probability is associated with the quality score difference of the corresponding pairwise overlapping data blocks. That is, the pairwise preference probability is equal to the result of the Sigmoid function acting on the quality score difference of the pairwise data blocks. With maximum likelihood estimation as the objective, an optimization objective function is constructed in combination with the mapping relationship. This function sums the likelihood terms of all pairwise overlapping data blocks. Each likelihood term includes the log-likelihood corresponding to the pairwise preference probability and the log-likelihood corresponding to the reverse preference. The optimization objective function is used to solve for each pairwise preference probability. Here, the multiple dimensions actually correspond to multiple measurement channels of the cross-domain original time series samples. Therefore, the score of each overlapping data block under each measurement channel will be output. The arithmetic mean of the scores of each overlapping data block under multiple measurement channels is calculated to determine the continuous quality score of each overlapping data block.

[0106] Step 104: Use the cross-entropy loss function to train the initial scoring model based on the continuous quality scores of each overlapping data block, and determine the target scoring model.

[0107] It should be noted that the initial scoring model consists of a pre-trained temporal feature extractor (such as MOMENT) and a three-layer MLP network, and the parameters of the pre-trained feature extractor are kept frozen to control computational costs. First, each overlapping data block is input into the initial scoring model. The pre-trained feature extractor extracts its temporal features, and then the MLP network outputs the corresponding predicted quality scores. Then, using the continuous quality scores of each overlapping data block as supervision, a loss term corresponding to the pair preference is constructed. The cross-entropy loss function is used to calculate the deviation between the predicted quality score and the continuous quality score. The model parameters are iteratively optimized to gradually converge the loss, and finally the target scoring model is determined.

[0108] Furthermore, step 104 may include the following sub-steps:

[0109] S41. Use the representation extractor in the initial scoring model to extract features from each overlapping data block to obtain the feature vector corresponding to each overlapping data block.

[0110] S42. Use the quality mapper in the initial scoring model to map the feature vectors corresponding to each overlapping data block, and output the scalar quality score corresponding to each overlapping data block.

[0111] S43. The cross-entropy loss function is used to train the quality mapper in the initial scoring model based on the scalar quality score and continuous quality score corresponding to each overlapping data block, and the target scoring model is determined.

[0112] MLP networks are multi-layer perceptrons.

[0113] It should be noted that the initial scoring model consists of two parts: 1) Representation extractor: based on the frozen pre-trained model MOMENT (with approximately 109 million parameters), it extracts the feature vector of each data block; 2) Quality mapper: it inputs the features into a three-layer MLP network (each layer is 256 wide and includes residual connections, ReLU and LayerNorm), and outputs a scalar quality score.

[0114] The training objective is to fit the LLM preference relationship, using cross-entropy as the loss function:

[0115] ;

[0116] in, To use model parameters The cross-entropy loss function is used to quantify the deviation between the prediction results of the initial rating model and the LLM (Large Language Model) preference relationship. It is the objective function that needs to be optimized during model training. This is the operator that takes the expected value from all elements in set J; When the model parameters are At that time, the initial scoring model output The predicted quality score (i.e., scalar quality score) is the model's prediction of the quality of the data block. When the model parameters are At that time, the initial scoring model output The predicted quality score is the model's prediction of the quality of the data block; this step realizes the transformation of weak supervision signals into a generalizable scoring model, which facilitates efficient evaluation of new data blocks.

[0117] In this embodiment, the representation extractor in the initial scoring model is used to extract features from each overlapping data block, obtaining the feature vector corresponding to each overlapping data block. Specifically, the representation extractor is a pre-trained temporal feature extractor (such as MOMENT), whose parameters are frozen to control computational cost. The extracted feature vectors contain key information such as temporal dependencies, trends, and fluctuations of the overlapping data blocks. The quality mapper in the initial scoring model is used to map the feature vectors corresponding to each overlapping data block, outputting a scalar quality score for each overlapping data block. This quality mapper is a three-layer MLP network that transforms the feature vectors into single-value scalar quality scores through fully connected layers. The quality mapper in the initial scoring model is trained using a cross-entropy loss function based on the scalar quality scores and continuous quality scores corresponding to each overlapping data block. This cross-entropy loss function is constructed based on the difference in scalar quality scores between pairs of overlapping data blocks and the pairwise preference probabilities corresponding to continuous quality scores. During training, only the parameters of the quality mapper are updated. The loss function is minimized iteratively, and after the loss converges to a preset threshold, the optimized model is determined as the target scoring model.

[0118] Step 105: Enhance the target scoring model based on the symbolic stochastic gradient descent algorithm and backpropagation method to determine the enhanced scoring model.

[0119] It should be noted that the target scoring model is augmented based on the symbolic stochastic gradient descent algorithm and backpropagation method to determine the augmented scoring model. First, the symbolic stochastic gradient descent algorithm is used to symbolize the gradient of the parameters of the target scoring model, retaining only the positive and negative directions of the gradient to reduce the complexity of calculating higher-order gradients. Then, the backpropagation method is used to backpropagate the model's predicted values ​​and the loss gradient of the actual continuous quality scores to each layer of the model. During the augmentation process, only the top-level parameters of the representation extractor and all parameters of the quality mapper are unfrozen and jointly updated. The training process is controlled by a preset number of iterations and a loss convergence threshold. Once the scoring error of the model on the validation set stabilizes within a preset range, training is stopped and the final augmented scoring model is determined.

[0120] Furthermore, step 105 may include the following sub-steps:

[0121] S51. Construct a task set;

[0122] S52. The target scoring model is trained using the symbolic stochastic gradient descent algorithm and backpropagation method based on the task set to determine the enhanced scoring model.

[0123] A task set refers to a task set that integrates multiple time-series data quality scoring sub-tasks from different domains. Each sub-task corresponds to a scoring requirement in a specific business domain and includes two data subsets: a support set and a query set. It is the core training carrier for achieving cross-domain generalization enhancement of the model.

[0124] Signed stochastic gradient descent is an optimization algorithm that retains only the positive and negative signs of the gradient. It can significantly reduce the complexity of gradient storage and computation, avoid the exponential training time caused by high-order gradients in traditional gradient descent, and adapt to the rapid iteration of models in multi-task scenarios.

[0125] Backpropagation is a parameter update algorithm that propagates the model's predicted loss gradient back from the output layer to the input layer. It can guide the precise adjustment of parameters in each layer of the model based on the loss gradient and is the core method for optimizing model parameters.

[0126] It should be noted that, to enhance the transferability of the rating regressor (i.e., the target rating model) across multi-domain data, this invention employs a model-agnostic meta-learning (MAML) strategy. The specific steps are as follows:

[0127] 1. Construct a task set T = {T1, T2, ..., T}n Each task corresponds to a separate domain of data (such as finance, electricity, or healthcare).

[0128] 2. Divide each task into a support set and a query set;

[0129] 3. Perform several inner loops of the Signed Stochastic Gradient Descent (signSGD) algorithm on the support set to quickly fit the task;

[0130] 4. Evaluate the loss on the query set and accumulate backpropagation to update the global parameters.

[0131] The ultimate optimization objective is to minimize the sum of query losses across all tasks:

[0132] ;

[0133] in, For the global parameters of the model To optimize variables, we need to find parameter values ​​that minimize the objective function. These are the global parameters of the enhanced scoring model, which are the model parameters that need to be updated and adjusted during this optimization process. For all tasks sampled from the task set T Perform a summation operation; T represents the set of tasks for cross-domain time series data quality scoring, with each task corresponding to a specific domain time series data scoring scenario; Let be the i-th specific task sampled from the task set T, corresponding to the time series data quality scoring task in a single domain; For the i-th task The loss function value on its query set is obtained by taking the parameters as... After applying the model to the query set of this task, the resulting loss is evaluated to measure the model's prediction bias on that query set. This mechanism enables the scoring model to achieve rapid transfer and efficient scoring when facing new domains.

[0134] In this embodiment, a task set is constructed by dividing the overlapping data blocks and their real-time quality scores corresponding to the original time-series samples from different business domains into domains. Each domain's time-series data quality scoring requirement corresponds to an independent sub-task. Each sub-task includes a support set for rapid model adaptation and a query set for effect verification, thereby integrating to form a complete task set covering multiple domains. The target scoring model is trained using the signed stochastic gradient descent algorithm and backpropagation method based on the task set to determine the enhanced scoring model. First, for each sub-task in the task set, the parameter gradient of the target scoring model is symbolized using the signed stochastic gradient descent algorithm to simplify the calculation. Then, the loss gradient on the support set is backpropagated to each layer of the model using the backpropagation method. Only the top-level parameters of the representation extractor and all parameters of the quality mapper are unfrozen for joint update. After completing the rapid adaptation of a single sub-task, the adaptation effect is verified through the query set and the global loss is calculated. After multiple rounds of task iterative training, the final enhanced scoring model is determined when the scoring error of the model on all sub-task query sets converges to a preset threshold.

[0135] Step 106: Use an enhanced scoring model to score each overlapping data block and output the target score result.

[0136] The target scoring results include point-level scores and sample scores.

[0137] Furthermore, step 106 may include the following sub-steps:

[0138] S61. Use an enhanced scoring model to score each overlapping data block and output the target score corresponding to each overlapping data block.

[0139] S62. Based on the target score corresponding to each overlapping data block, perform point-level score mapping and sample-level score mapping respectively, and output point-level score and sample score.

[0140] It should be noted that an enhanced scoring model is used to score each overlapping data block, outputting the target score corresponding to each overlapping data block. Based on the target scores corresponding to each overlapping data block, point-level score mapping and sample-level score mapping are performed respectively, outputting point-level scores and sample scores. The point-level score mapping is applied to each time point in the original cross-domain time series samples. First, determine the set of all overlapping data blocks involved at that point in time. Then, calculate the arithmetic mean of the target scores of all overlapping data blocks within the set to obtain the point-level score corresponding to that time point. The sample-level rating mapping involves counting the total number of time points T in the entire sample, summing the point-level ratings of all time points in the entire sample, and dividing by T to obtain the sample rating corresponding to the entire original time series sample. Through this hierarchical clustering mechanism, the mapping from the rating of locally overlapping data blocks to the rating of a single time point, and then to the rating of the entire sample is completed.

[0141] To achieve the scoring mapping from local data blocks to the overall sample, this invention designs a hierarchical aggregation mechanism:

[0142] 1) Point-level scoring: The score at a given point in time is the average of the scores of all the blocks it participated in.

[0143] ;

[0144] in, The point-level score for the i-th time point in the cross-domain original time series sample is the average score of all overlapping data blocks involved at that time point; For time points in the original time series samples across domains The set consisting of all overlapping data blocks involved; For the kth overlapping data block The corresponding target score is output by the enhanced scoring model; For including time points One of the overlapping data blocks is a set The elements in.

[0145] 2) Sample rating: The rating of the entire sample is the average of its scores at all time points.

[0146] ;

[0147] in, For the entire original time series sample The sample score is the average of the point-level scores at all time points of the entire sample; Original time series samples The total number of time points included. This step enables a natural transition from local judgment to global evaluation, helping to support data selection and filtering at different granularities.

[0148] For comparison of technical effectiveness, existing technologies can be used as a reference. With the rapid development of industries such as industry, finance, healthcare, and transportation, the scale of time-series data has exploded. However, in the actual data acquisition process, time-series data inevitably encounters problems such as noise pollution, missing values, sensor malfunctions, and inconsistent sampling frequencies, leading to a significant decline in data quality. Low-quality data not only weakens the model's learning effect but may also introduce biases and amplify noise, ultimately affecting the model's stability and generalization performance. Therefore, before training deep learning models, quality scoring of the raw data and selection of high-quality samples have become important means to improve overall system performance and reduce training costs.

[0149] To address this need, the academic community has explored various data quality assessment schemes, but limitations remain. Firstly, one approach attempts to transfer general data quality metrics to time-series tasks, such as Influence Function and ShapleyValue, judging quality by calculating the marginal impact of samples on model performance. These methods offer theoretical impartiality and model independence, but are computationally extremely expensive, especially when considering time-series dependencies and high-dimensional data, facing costly challenges such as gradient backpropagation and Hessian matrix approximation. Furthermore, their initial assumptions are often based on independent and identically distributed (ICD) data, making it difficult to effectively model time-dependent structures. Although subsequent methods such as TimeInf and TimeShap have attempted to introduce time-series-aware mechanisms to improve performance in specific tasks, they still face bottlenecks such as weak cross-task generalization ability and high adaptation costs.

[0150] Another emerging approach leverages the powerful reasoning and understanding capabilities of Large Language Models (LLMs) to explore their potential in data quality assessment. Inspired by text quality evaluation work, existing research has guided LLMs to score and rank data samples based on human standards using Prompt templates. Methods such as Qurating and GPT Judge have demonstrated excellent performance in dimensions such as readability and accuracy. These methods showcase the strong generalization ability of LLMs under zero-sample or few-sample conditions, but their main applications remain focused on symbolic data such as text or code. Although a few works have attempted to convert time-series data into text input for LLM task processing, a systematic exploration of LLM's capabilities in time-series data quality understanding, pairwise preference judgment, and score generation remains lacking. Existing methods have not fully explored the potential of LLMs in processing non-natural language structure data; their feasibility and advantages in understanding complex time-series patterns, identifying anomalous behavior, and performing generalized scoring tasks require further investigation.

[0151] Existing mainstream methods, such as Shapley Value and Influence Function, primarily rely on resampling or perturbing samples and then retraining the model to estimate their impact on performance. While these methods have some theoretical support, they are computationally extremely expensive, especially when dealing with high-dimensional time-series data and complex tasks, making them difficult to apply in large-scale data scenarios. Second, most scoring methods are often trained and used in fixed tasks or single domains, lacking adaptability to new domains. Once deployed across domains, their scorers often fail or misjudge, exhibiting severely insufficient generalization. Third, existing methods generally rely on high-quality labeled data or expert-defined features as sources of supervision information. They lack scalability in unlabeled, weakly labeled, or cross-task settings, making them difficult to use in general-purpose scoring systems for practical deployment.

[0152] To address the aforementioned issues, this invention provides a cross-domain time series data quality scoring method based on a large language model. This method utilizes LLM to generate pairwise quality preference judgments and combines preference modeling and meta-learning mechanisms to achieve accurate, reliable, low-cost, and highly adaptable time series data scoring across different domains.

[0153] Specifically, such as Figure 2 As shown, the overall framework of the cross-domain time-series data quality scoring method based on a large language model proposed in this invention includes six key modules: a data block partitioning module, a large language model discrimination module, a quality score modeling module, a score regression model training module, a meta-learning optimization module, and a hierarchical score aggregation module. These modules work collaboratively to construct a quality assessment system capable of extracting multi-granularity scoring information from the sample structure and possessing good cross-domain adaptability. The system first divides the input complete time-series samples using a sliding window mechanism, generating multiple overlapping data blocks with time-dependent structures. Then, by constructing specific prompt word templates, it guides a large language model to perform pairwise quality judgments on any two data blocks, outputting preference tendencies. Next, these preference labels are input into the Bradley-Terry model, mapping them to continuous quality scores for each data block. Subsequently, using the scores as supervision signals, a regression model is trained to predict quality scores from the original features. To improve its generalization ability, a MAML-based meta-learning mechanism is further introduced, enabling the model to quickly adapt to multi-source data. Finally, a hierarchical aggregation mechanism elevates the block-level scores to the point and sample levels, achieving refined scoring of the entire sample.

[0154] Based on the above, this invention constructs four types of natural language prompt templates—trend, frequency, amplitude, and pattern—to guide a Large Language Model (LLM) to make pairwise preference judgments on time-series data blocks, thereby achieving high-quality unsupervised scoring without the need for manual scoring. The scoring model's structure consists of a frozen temporal encoder and a lightweight regressor, maintaining the utilization of pre-trained temporal features while possessing good interpretability, transferability, and inference efficiency. Furthermore, this invention utilizes MAML for task generalization training of the scoring model, significantly enhancing the scorer's adaptability to new domains, and employs signSGD to avoid the high-overhead second-order gradient calculation in meta-learning, reducing training complexity. In addition, this invention uses the quality score output by the scorer to guide data pruning, data augmentation, and model fine-tuning tasks, thereby achieving better data utilization efficiency and stronger model performance in practical applications.

[0155] Compared with the prior art, the present invention has the following advantages:

[0156] (1) Significantly improved computational efficiency: Existing methods based on Shapley values ​​and Influence Functions require repeated training or multiple model evaluations to estimate the "influence" of each data point, resulting in high computational costs on large-scale datasets and making them difficult to apply in practice. This invention, however, requires only a limited number of LLM pairwise judgments and one scorer training, thus significantly reducing time and resource overhead. Furthermore, by introducing a Large Language Model (LLM) to generate quality preference judgments for paired samples, combined with a lightweight preference modeler and the Signed Stochastic Gradient Descent (signSGD) algorithm, this invention achieves rapid scoring without retraining or high-order gradient calculations, making it suitable for deployment in large-scale data systems. In practical tests, TSRating's scoring speed outperforms methods based on Shapley values ​​and Influence Functions.

[0157] (2) Achieving true "cross-domain scoring" capability: Traditional methods fit scoring functions on a single task. Once the task or data distribution changes, re-evaluation or even retraining is required, resulting in poor generalization. This invention trains the scorer through a meta-learning framework, enabling it to "quickly adapt" to multiple domains (finance, power, healthcare, etc.), thus achieving one-time training and multi-domain deployment, greatly improving the versatility of the scoring system. Simultaneously, this invention incorporates a meta-learning mechanism, learning a scorer that can quickly adapt to new tasks during multi-source, multi-task training, allowing for rapid adaptation using only a small amount of new task data. Furthermore, relying on the powerful generalization capability of LLM in multi-domain semantic understanding, the scoring system possesses excellent cross-domain transfer capability, fundamentally improving its practicality and flexibility.

[0158] (3) Unsupervised scoring without relying on manual labels: Existing methods often rely on information such as model prediction probabilities, training losses, or task labels, which limits their application in unsupervised scenarios. This invention constructs a Prompt template to guide the LLM to use its knowledge base to judge the quality preferences of sample pairs, thereby generating paired pseudo-labels as training signals. This bypasses the dependence on real labels and achieves "unsupervised scoring," breaking through the limitations of supervised learning. This method not only greatly reduces the deployment threshold of the scoring system but also endows the model with the ability to score on unlabeled data, significantly enhancing its practicality and adaptability.

[0159] (4) The scoring process is interpretable and the results are reliable and controllable: LLM preference judgments come from clear comparison dimensions and input prompts. Users can specify different dimensions to generate preference data. The source of the scores is clear and traceable. As a classic preference modeling method, Bradley-Terry's modeling results are stable and statistically interpretable. Compared with black-box scoring, it is more transparent and credible.

[0160] In this embodiment of the invention, a method for cross-domain time-series data quality scoring based on a large language model is provided. The method acquires original cross-domain time-series samples and divides them using a sliding window mechanism to generate multiple overlapping data blocks. Based on majority voting, a large language model is used to perform pairwise preference discrimination according to predefined prompt word templates and each overlapping data block, outputting multiple pairwise preference probabilities. A Bradley-Terry model is used to map each pairwise preference probability, outputting a continuous quality score for each overlapping data block. A cross-entropy loss function is used to adjust the initial... The scoring model is trained to determine the target scoring model; the target scoring model is augmented using the signed stochastic gradient descent algorithm and backpropagation to determine the augmented scoring model; the augmented scoring model is used to score each overlapping data block and output the target scoring result; based on the above scheme, this invention introduces a large language model (LLM) for pairwise preference discrimination, combines the lightweight analytical mapping of the Bradley-Terry model and the signed stochastic gradient descent (signSGD) algorithm to achieve fast scoring without retraining and high-order gradient calculation, greatly reducing computational overhead and making it suitable for deployment in large-scale data systems.

[0161] Please see Figure 3 , Figure 3 This is a structural block diagram of a cross-domain time-series data quality scoring device based on a large language model, provided in Embodiment 2 of the present invention.

[0162] This invention provides a cross-domain time-series data quality scoring device based on a large language model, comprising:

[0163] The acquisition module 301 is used to acquire cross-domain original time series samples and divide the cross-domain original time series samples based on the sliding window mechanism to generate multiple overlapping data blocks;

[0164] The discrimination module 302 is used to perform pairwise preference discrimination based on the majority voting method and the large language model according to the predefined prompt word template and each overlapping data block, and output multiple pairwise preference probabilities.

[0165] The mapping module 303 is used to map the pairwise preference probabilities based on the Bradley-Terry model and output the continuous quality scores of each overlapping data block.

[0166] Training module 304 is used to train the initial scoring model based on the continuous quality scores of each overlapping data block using the cross-entropy loss function, and to determine the target scoring model.

[0167] Enhancement module 305 is used to enhance the target scoring model based on the symbolic stochastic gradient descent algorithm and backpropagation method, and to determine the enhanced scoring model;

[0168] The scoring module 306 is used to score each overlapping data block using an enhanced scoring model and output the target scoring result.

[0169] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the above-described device and module can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0170] This invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program; when the computer program is executed by the processor, the processor performs the steps of the cross-domain time-series data quality scoring method based on a large language model as described in the above embodiments.

[0171] This invention also provides a computer-readable storage medium storing a computer program / instructions thereon, which, when executed by a processor, implements the steps of the cross-domain time-series data quality scoring method based on a large language model as described in the above embodiments.

[0172] This invention also provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of the cross-domain time-series data quality scoring method based on a large language model as described in the above embodiments.

[0173] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0174] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0175] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0176] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0177] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention 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 of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A cross-domain time-series data quality scoring method based on a large language model, characterized in that, include: Obtain cross-domain original time series samples, and divide the cross-domain original time series samples based on a sliding window mechanism to generate multiple overlapping data blocks; Based on the majority voting method, a large language model is used to perform pairwise preference discrimination according to the predefined prompt word template and each overlapping data block, and output multiple pairwise preference probabilities. The Bradley-Terry model is used to map the pairwise preference probabilities of each pairwise data block, and the continuous quality score of each overlapping data block is output. The initial scoring model is trained using the cross-entropy loss function based on the continuous quality scores of each overlapping data block to determine the target scoring model; The target scoring model is augmented using the signed stochastic gradient descent algorithm and backpropagation method to determine the augmented scoring model. The enhanced scoring model is used to score each of the overlapping data blocks, and the target score result is output.

2. The method for cross-domain time series data quality scoring based on a large language model according to claim 1, characterized in that, The majority voting method employs a large language model to perform pairwise preference discrimination based on predefined prompt word templates and overlapping data blocks, outputting multiple pairwise preference probabilities, including: Select any two overlapping data blocks from the plurality of overlapping data blocks and combine them to output a plurality of pairs of overlapping data blocks; The large language model is used to perform quality comparison on each pair of overlapping data blocks according to the predefined prompt word template, and outputs pair preference labels for each pair of overlapping data blocks; The proportion of pairwise preference labels for each of the overlapping data blocks is calculated using the majority voting method to determine the pairwise preference probability of each of the overlapping data blocks.

3. The method for cross-domain time series data quality scoring based on a large language model according to claim 1, characterized in that, The method of mapping the pairwise preference probabilities using the Bradley-Terry model to output continuous quality scores for each overlapping data block includes: A mapping relationship is constructed based on the Bradley-Terry model; With maximum likelihood estimation as the objective, an optimization objective function is constructed by combining the aforementioned mapping relationship; The optimization objective function is used to solve for each pairwise preference probability, and the scores of each overlapping data block in multiple dimensions are output. The scores of each overlapping data block are averaged across multiple dimensions to determine the continuous quality score of each overlapping data block.

4. The method for cross-domain time-series data quality scoring based on a large language model according to claim 1, characterized in that, The step of training the initial scoring model using the cross-entropy loss function based on the continuous quality scores of each of the overlapping data blocks to determine the target scoring model includes: The representation extractor in the initial scoring model is used to extract features from each of the overlapping data blocks to obtain the feature vectors corresponding to each of the overlapping data blocks; The feature vectors corresponding to each overlapping data block are mapped using the quality mapper in the initial scoring model, and the scalar quality score corresponding to each overlapping data block is output. The cross-entropy loss function is used to train the quality mapper in the initial scoring model based on the scalar quality score and continuous quality score corresponding to each of the overlapping data blocks, thereby determining the target scoring model.

5. The method for cross-domain time series data quality scoring based on a large language model according to claim 1, characterized in that, The method of enhancing the target scoring model based on the symbolic stochastic gradient descent algorithm and backpropagation to determine the enhanced scoring model includes: Build a set of tasks; The target scoring model is trained using the symbolic stochastic gradient descent algorithm and the backpropagation method based on the task set to determine the enhanced scoring model.

6. The method for cross-domain time series data quality scoring based on a large language model according to claim 1, characterized in that, The target scoring result includes point-level scoring and sample scoring; the step of using the enhanced scoring model to score each of the overlapping data blocks and outputting the target scoring result includes: The enhanced scoring model is used to score each of the overlapping data blocks, and the target score corresponding to each of the overlapping data blocks is output. Based on the target score corresponding to each of the overlapping data blocks, point-level score mapping and sample-level score mapping are performed respectively, and point-level score and sample score are output.

7. A cross-domain time-series data quality scoring device based on a large language model, characterized in that, include: The acquisition module is used to acquire cross-domain original time series samples and divide the cross-domain original time series samples based on a sliding window mechanism to generate multiple overlapping data blocks; The discrimination module is used to perform pairwise preference discrimination based on the majority voting method and a large language model according to the predefined prompt word template and each overlapping data block, and output multiple pairwise preference probabilities. The mapping module is used to map the pairwise preference probabilities of each data block using the Bradley-Terry model and output the continuous quality score of each data block. The training module is used to train the initial scoring model using the cross-entropy loss function based on the continuous quality scores of each of the overlapping data blocks, and to determine the target scoring model. The enhancement module is used to enhance the target scoring model based on the symbolic stochastic gradient descent algorithm and the backpropagation method, and to determine the enhanced scoring model; The scoring module is used to score each of the overlapping data blocks using the enhanced scoring model and output the target scoring result.

8. An electronic device, characterized in that, The system includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor causes the processor to perform the steps of the cross-domain time-series data quality scoring method based on a large language model as described in any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed, it implements the cross-domain time-series data quality scoring method based on a large language model as described in any one of claims 1-6.

10. A computer program product, characterized in that, The computer program product includes a computer program stored on a non-transitory computer-readable storage medium, the computer program including program instructions, wherein when the program instructions are executed by a computer, the computer performs the cross-domain time-series data quality scoring method based on a large language model as described in any one of claims 1-6.

Images