Astronomical light variation curve characterization and classification method based on comparative learning

Abstract

The invention provides an astronomical light change curve characterization and classification method based on comparative learning, which can be used for automatically analyzing time sequence light change data from different patrol days. The method comprises the following steps: firstly, carrying out normalization and other preprocessing on an original optical variable sequence; and then inputting the data into a Transform-based representation model, and obtaining universal features shared across the patrol through hierarchical comparison learning of a sample and a time level. In the unsupervised pre-training stage, the characterization model learns an internal rule of light change along with time change; in the fine tuning stage, the pre-training model is loaded to perform fine tuning on the classification head so as to realize classification. In terms of system implementation, a unified reasoning interface is provided, the weight can be automatically loaded, and the classification result and confidence can be output. Compared with a traditional method depending on artificial features or a single data source, the method has the advantages that cross-patrol migration can be achieved through one-time pre-training, feature consistency and good expandability under different observation conditions are enhanced, and a high-reliability scheme is provided for intelligent processing of large-scale time-domain astronomical data.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence algorithms and astronomical time-domain data processing technology, specifically involving a method for representing and classifying astronomical light curves based on contrastive learning. Background Technology

[0002] With the continuous operation of large-scale time-domain sky survey projects (such as the Transiting Exoplanet Survey Satellite (TESS), Zwicky Transient Facility (ZTF), and Gaia), astronomy has entered the era of "time-domain big data" with time as the core dimension. Different survey telescopes continuously generate massive amounts of light curves during long-term observations, recording the brightness information of celestial objects over time. These light curve data contain rich information on physical processes and can be used to identify various celestial object types, such as pulsating variable stars (Cepheid, RR Lyrae, Delta Scuti, etc.), eclipsing binaries, rotational variables, and non-variable stars. However, significant differences exist between observation systems in terms of sampling frequency, observation band, signal-to-noise ratio, and data completeness, leading to inconsistencies in the distribution characteristics and temporal structure of light curves, thus posing a significant challenge to automated classification and feature extraction.

[0003] Currently, most mainstream light curve classification models rely on manually constructed features (such as period, amplitude, Fourier parameters, etc.) combined with traditional machine learning algorithms (such as random forests, support vector machines, etc.) for training. These methods require astronomical experts to design prior features for each variable type, resulting in high manual costs and difficulty adapting to massive datasets. In recent years, deep learning models (especially convolutional and recurrent networks) have been introduced into light curve analysis, enabling them to automatically learn time-series features to some extent. However, these deep learning models typically require supervised training on large-scale labeled data, and the acquisition of variable star labels is heavily reliant on manual comparison of light curve features, which is costly. Furthermore, the inconsistency of different survey data significantly reduces the model's generalization performance across telescope scenarios: models trained on a single dataset often cannot be directly transferred to another survey mission.

[0004] Currently, research on transfer learning between multiple survey telescopes remains relatively limited. Different survey systems exhibit significant differences in their observation strategies, such as observation intervals, observation lengths, and signal-to-noise ratios. These factors directly influence the temporal structure and feature distribution of light curves. Due to the uneven temporal sampling and large differences in the time span among different surveys, directly transferring deep learning models between different data sources often leads to feature mismatch and performance degradation, becoming a major technical challenge in cross-scan learning. Existing work largely focuses on light curve analysis under a single telescope or a single sampling rate, lacking an effective model capable of achieving unified representation and stable classification under multiple observation conditions.

[0005] With the rapid development of contrastive learning techniques in the field of representation learning, feature alignment models based on unsupervised signals have achieved significant breakthroughs in image and speech processing. These models construct positive and negative sample pairs in the sample space, maximizing the consistency between similar samples and minimizing the differences between dissimilar samples, thereby achieving efficient data representation learning. However, most existing contrastive learning models originate from image and speech processing and are not designed for the long-term temporal structure, multi-source heterogeneity, and non-uniform sampling characteristics of astronomical light curves, making them difficult to directly apply to the unified representation of multi-survey light curve data.

[0006] Therefore, there is an urgent need for a light curve representation model that can simultaneously handle differences in observations from multiple sky surveys, possess unsupervised feature learning capabilities, and achieve universal feature transfer across different data sources. This model should be able to: Under unlabeled conditions, temporal structural features of light curves are extracted through a two-level comparative learning mechanism of time and samples. By utilizing a unified model architecture and normalization strategy, feature alignment and representation sharing of different survey data can be achieved. By using a pre-trained model from a single sky survey, efficient variable star classification can be achieved by transferring the model across telescopes to other datasets. Summary of the Invention

[0007] The purpose of this invention is to overcome the limitations of existing models in processing astronomical time-domain data, including inconsistent feature distributions due to observational differences between different survey telescopes, difficulties in model transfer, and insufficient modeling of temporal structures by existing deep learning models. To address these issues, a contrastive learning-based method for representing and classifying astronomical light curves is proposed to achieve high-precision, low-cost, and automated feature learning and intelligent classification of time-domain astronomical data across surveys.

[0008] The training method provided in this invention performs feature alignment simultaneously on both the time and sample scales through an unsupervised contrastive learning mechanism, obtaining a general representation of light curves that can be transferred across surveys. This method has the advantages of unified modeling and cross-domain generalization, making it suitable for automated analysis and classification tasks of large-scale temporal astronomical survey data.

[0009] To achieve the above objectives, the present invention adopts the following technical solution: This invention provides a method for characterizing and classifying astronomical light curves based on contrastive learning, comprising the following steps: (1) Data preprocessing process (optical curve standardization and mask filling) The raw light curves acquired by different survey telescopes were uniformly converted into structured time-series inputs that could be directly received by subsequent depth characterization models. Due to significant differences in observation intervals and lengths among the surveys, the raw observation sequences were processed sequentially as follows: Based on the overall statistical characteristics of each light curve, this invention employs the 3igma criterion to identify and eliminate outlier observation points. Specifically, the global mean is calculated for each light curve. and standard deviation and for each observation in the sequence Make a judgment: if its deviation satisfies The observation point is then considered an outlier and removed to eliminate non-physical outliers caused by instantaneous observation errors, instrument jitter, or telescope attitude adjustments. To eliminate differences in timescale, exposure rhythm, and photometric measurement range among different survey telescopes, this invention performs minimum-maximum normalization processing on both the time series and flux series of each light curve. Specifically, a minimum value can be found for both the time series and the flux series. and the maximum value Minimum-maximum normalization is based on Map each time point and flow point to an interval according to the following rules. This step ensures that the input sequence maintains a consistent numerical range across different data sources; To address the uneven sampling across different sky surveys, a masking and padding process is applied to the time series. Specifically, for each light curve, a fixed sequence length required for the model input is first determined. If the original sequence length is less than the fixed sequence length, zero values ​​are added to the end to achieve a uniform length. If the original sequence length exceeds the fixed sequence length, the preceding observation points are truncated in chronological order, and a mask vector of the same length as the sequence is generated. This mask vector is used in the Transformer's self-attention calculation to mask the padding positions, ensuring that the model models temporal relationships based solely on real observation points. This step allows the model to maintain structured input even under sparse observation conditions.

[0010] (2) Representation learning process (hierarchical contrastive learning based on Transformer) The preprocessed light curve is input into a deep representation model based on the Transformer encoder architecture. The model captures the long-term dependency features of the light curve sequence through a self-attention mechanism and optimizes the representation at two levels using a hierarchical contrastive learning mechanism: The learning steps for temporal-level contrast are as follows: (a) Randomly select two partially overlapping subsequences from the same light curve; (b) Input the two subsequences into the deep representation model respectively to obtain the corresponding time series feature vectors; (c) In the overlapping region of two feature sequences, based on the alignment relationship of time indices, the feature vectors at positions with consistent time indices are defined as positive sample pairs, and the feature vectors at positions with inconsistent time indices are defined as negative sample pairs. (d) Train a deep representation model based on the positive and negative sample pairs.

[0011] The learning steps for instance-level contrast are as follows: (a) Input multiple optical curves simultaneously in a training batch; (b) For each light curve, generate two enhanced sequences with partially overlapping regions, and input all enhanced sequences into the depth representation model to obtain its time series feature vector; (c) On the alignment time index of the feature sequence, the feature vectors of two enhancement sequences from the same light curve with the same time index are defined as positive sample pairs; the feature vectors from different light curves but with the same time index are defined as negative sample pairs. A deep representation model is trained using the positive and negative sample pairs. By jointly optimizing the instance-level and temporal-level contrastive losses, the model learns a general light variation representation that can be transferred across surveys under unsupervised conditions.

[0012] (3) Fine-tuning and classification (cross-survey migration and downstream mission adaptation) After unsupervised pre-training of the model using the aforementioned representation learning combined with light curve data from any sky survey (such as TESS), a shared Transformer backbone network can be obtained. When transferring the model to other sky survey projects (such as ZTF, Gaia, or any target sky survey dataset), the parameters of this backbone network are kept fixed, and the classification fully connected layer is fine-tuned using only the labeled samples from the target sky survey to complete the adaptation to the sky survey data.

[0013] (4) Reasoning and output (unified interface and classification result generation) The system provides a unified inference interface for model loading, data input, and result output.

[0014] When the light curve to be classified is input, the model automatically outputs the following results: Predict categories (such as Cepheid, Mira, RR Lyrae, Eclipsing Binary, Non-variable, etc.); The classification confidence score is used to measure the reliability of the model's prediction results. Optional visualization outputs include light curves and UMAP plots.

[0015] It supports batch inference and distributed execution, making it suitable for deployment in astronomical data centers and high-performance computing platforms.

[0016] (5) Continuous expansion and model update By periodically introducing unlabeled samples and optimizing the contrast loss, the model can adaptively expand its feature space, maintaining stable performance and cross-domain consistency in long-term sky survey missions.

[0017] The present invention also provides an electronic device, including a memory and a processor, wherein the memory is coupled to the processor; wherein the memory is used to store program data, and the processor is used to execute the program data to implement the aforementioned method for characterizing and classifying astronomical light curves based on contrastive learning.

[0018] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the aforementioned method for characterizing and classifying astronomical light curves based on contrastive learning.

[0019] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the aforementioned method for characterizing and classifying astronomical light curves based on contrastive learning.

[0020] Compared with existing classification methods that rely on manual features or a single data source, the present invention has the following advantages: One pre-training can enable transfer and reuse of data from multiple surveys, significantly reducing the cost of manual annotation and training; Hierarchical contrastive learning mechanisms can simultaneously capture local temporal features and global sample distribution, improving the model's discriminativeness and robustness. It supports distributed inference and scalable deployment, making it suitable for intelligent processing of massive amounts of time-domain data. This invention provides an efficient, unified, and transferable intelligent learning scheme for the analysis of light curves in time-domain astronomy, laying the foundation for the automatic classification and scientific mining of large-scale sky survey data in the future. Attached Figure Description

[0021] Figure 1 This is a complete process flowchart of an embodiment of this application; Figure 2 This is a flowchart illustrating the deployment method of the unified inference interface in this application embodiment; Figure 3 This is a comparison diagram of the method described in this application with other methods. In this diagram, Model represents the model trained using the method described in this paper, LSTM represents the Long Short-Term Memory network, Transformer (from scratch) represents the Transformer model trained from scratch, epoch represents the number of training epochs corresponding to the validation set inference stage, Loss represents the loss value of validation set inference, and Macro F1 represents the Macro F1 metric of validation set inference. Figure 4 This is a schematic diagram of a computer device according to an embodiment of this application. Detailed Implementation

[0022] To further illustrate the technical solution of this invention, a specific implementation of the invention is given below in conjunction with a typical astronomical light curve classification task. This embodiment is only used to illustrate the core principles, execution flow, and key steps of this invention, to facilitate understanding and implementation by those skilled in the art, and should not be construed as a limitation on the scope of protection of this invention.

[0023] The following will combine Figure 1 The workflow of the present invention will be described in detail, and specific embodiments of the invention will be explained in detail.

[0024] I. Implementation Environment and Data Sources The light curve data used in this embodiment comes from two mainstream astronomical time-domain survey projects: TESS (Transiting Exoplanet Survey Satellite) and ZTF (Zwicky Transient Facility). Since different surveys have significant differences in observation intervals and lengths, to ensure the model can achieve a unified representation across multiple data sources, this invention first selects light curve samples from each survey dataset, covering various typical categories, including Cepheid, Mira, RR Lyrae, eclipsing binaries, and non-variable samples.

[0025] Each light curve consists of observation time and corresponding luminous flux. To eliminate differences in data structure and photometric scale among different sky surveys, this invention performs format conversion and normalization operations on the original light curves through unified preprocessing, thereby ensuring that the model can be trained based on a consistent input dimension and data structure in the subsequent representation learning stage.

[0026] II. Data Preprocessing Flow 1. Outlier removal In this embodiment, the 3 sigma criterion is used to identify outliers in both the TESS and ZTF photometric curves. Specifically, the mean and standard deviation of each photometric curve are calculated, and deviation detection is performed on each observation point: if the deviation exceeds 3 sigma, the point is identified as an outlier and removed. Simultaneously, NaN values ​​in the photometric data are removed or marked to avoid invalid data interfering with subsequent characterization.

[0027] 2. Time and flow normalization To eliminate the differences in observation timescales and photometric measurement ranges among different sky surveys, a min-max normalization operation was performed on the time series and flux series of each light curve, mapping them uniformly to the interval [0,1]. This normalization ensures that all input sequences are consistent in dimensions, providing a stable input basis for the model's subsequent cross-source comparative learning.

[0028] 3. Mask padding Due to significant differences in sampling density between different sky surveys (e.g., TESS uses high sampling, while ZTF uses sparse sampling), this invention employs a fixed-length sequence format and constructs a mask matrix to preserve the original sampling sparsity. Specifically, TESS light curves are uniformly aligned to a length of 8192, and ZTF light curves are uniformly aligned to a length of 1024. Sequences shorter than the required length are padded with placeholders, and missing positions are marked in the mask vector, ensuring that the model performs correlation calculations based solely on real observation points in the self-attention mechanism.

[0029] After the above processing steps, the light curves from different sky surveys were transformed into tensor data with a consistent structure, achieving unified formatting of multi-source light curve sequences and laying a data foundation for subsequent deep representation learning and cross-domain transfer.

[0030] III. Representation Learning Process (Unsupervised Comparative Learning Phase) 1. Model Structure The preprocessed light curves are input into a Transformer-based representation learning model. This model utilizes a self-attention mechanism to model the long-term dependencies and irregular sampling patterns of the light curve sequences, thereby extracting consistent temporal features across surveys without assuming fixed time intervals. The global attention structure of the Transformer enables the model to adapt to varying light curve inputs, providing a stable model foundation for subsequent contrastive learning.

[0031] 2. Hierarchical contrastive learning mechanism The hierarchical contrastive learning mechanism proposed in this invention comprises two complementary levels: "temporal-level contrast" and "sample-level contrast." These levels are used to constrain the local temporal structure consistency of the light curve and the global sample representation distribution, respectively, thereby obtaining a unified feature space that can be shared across surveys. The two types of contrastive losses work together to form the training objective of the Transformer backbone network in the unsupervised pre-training stage.

[0032] (1) Temporal-level Contrast: Two partially overlapping subsequences are randomly selected from the same optical curve and input into the deep representation model to obtain the corresponding time-series feature vectors. Subsequently, within the overlapping interval of the two feature sequences, based on the alignment relationship of the time indices, the feature vectors at positions with consistent time indices are defined as positive sample pairs, and the feature vectors at positions with inconsistent time indices are defined as negative sample pairs. The deep representation model minimizes the cosine distance between positive sample pairs and increases their feature difference with negative sample pairs, enabling the feature space to accurately represent the local structural consistency of the optical curve under time alignment conditions.

[0033] (2) Instance-level Contrast: Multiple light curves are input simultaneously in the same training batch. Two enhanced sequences with partially overlapping regions are generated for each light curve, and all enhanced sequences are input into the deep representation model to obtain their time-series feature vectors. Subsequently, based on the aligned time index of the feature sequences, the feature vectors at the positions where the two enhanced sequences from the same light curve have the same time index are defined as positive sample pairs, and the feature vectors at the positions where the time index is the same from different light curves are defined as negative sample pairs. By minimizing the cosine distance of positive sample pairs and increasing their feature differences with negative sample pairs, the model enables the representation space to reflect the global structural differences of different variable star samples, thereby enhancing the discrimination capability across survey scenarios.

[0034] 3. Pooling strategy and hierarchical structure: In the unsupervised pre-training phase, this invention performs max pooling operations on the time-series features output by the Transformer encoder in a progressively larger time dimension to construct a multi-layered temporal representation. Assume the size of the encoder output feature tensor is... ,in For batch size, The length of the time series. If the feature dimension is used, then the first max pooling will reduce the sequence length from [previous value] in the time dimension. downsampling The resulting size is The model extracts the feature sequences and calculates the similarity between positive and negative samples based on the corresponding time-level and sample-level contrast losses at that scale. Then, it performs max pooling again on the downsampled feature sequences in the time dimension, further halving the sequence length, and repeats the contrast loss calculation at the new time scale. This process iterates until downsampling in the time dimension is no longer possible. By applying contrast constraints at each downsampling level, the model can simultaneously learn local change patterns and long-term structural information across multiple time scales, from fine to coarse, thus forming a hierarchical temporal feature representation and enhancing its ability to model the multi-scale structure of light curves.

[0035] 4. Unsupervised pre-training Without labels, the model continuously optimizes the feature space through contrastive loss, causing similar light curves to cluster in close regions, forming a transferable unified representation space. After training, the model obtains a universal light feature vector that can be reused across surveys.

[0036] IV. Fine-tuning and Classification Process After completing one unsupervised pre-training, this invention can perform multiple supervised fine-tuning operations for different sky survey missions. In each fine-tuning, the model is trained using labeled variable star samples from the corresponding sky survey (such as TESS, ZTF, or Gaia) to adapt to the specific observational characteristics and noise distribution of that data source. During the fine-tuning stage, the global feature aggregation method is switched to mean pooling to obtain a smoother and more robust global temporal representation, thereby improving the stability and classification consistency of the model when transferring between different sky surveys.

[0037] The fine-tuning process employs the standard cross-entropy loss function for optimization, outputting the final classification probability distribution. Since the model has already acquired a unified light curve representation space through contrastive learning, the fine-tuning of each survey mission can converge quickly with fewer labeled samples, achieving efficient transfer learning and accurate classification. Through this "one-time pre-training, multiple fine-tuning" strategy, this invention significantly reduces the cost of manual annotation and training, while maintaining excellent generalization ability and classification stability under different telescope observation conditions.

[0038] V. Reasoning and Output of Results 1. Unified Inference Interface This invention designs a unified inference interface, such as Figure 2 As shown, the system supports multi-format data input and automatic model loading. When a new optical curve is input, the system automatically performs data preprocessing and classification operations.

[0039] 2. Output Results The model output includes: Predict categories (e.g., Cepheid, Mira, RR Lyrae, Eclipsing Binary, Non-variable); Classification confidence score (in the range of 0–1).

[0040] 3. Deployment Method The inference interface of this invention consists of a main control module and several parallel inference sub-modules, as shown in the schematic diagram. Figure 2 As shown, the main control module receives the light curve data to be processed, segments the input data, and distributes it to multiple inference sub-modules. Each inference sub-module loads the same pre-trained model backbone and fine-tuned classification layer on an independent GPU and performs forward inference on the allocated data batch. After inference is completed, each sub-module returns its local inference results to the main control module, which then aggregates and sorts the results, ultimately outputting the target category and its confidence score. The entire process can be executed in parallel on multi-GPU servers or high-performance computing cluster (HPC) nodes, thereby achieving high-throughput inference of large-scale survey light curve data.

[0041] VI. Continuous Learning and Model Update Mechanism As new survey data is continuously generated, this invention enables incremental comparative learning under unsupervised conditions to continuously update the model's representational capabilities. When new unlabeled light curves are added, the system can extract features and update the comparative sample library, allowing the model to maintain stable performance despite changes in data distribution. This continuous learning mechanism not only ensures the model's robustness in long-term survey missions but also provides higher-quality initialization features for subsequent supervised fine-tuning, thereby improving adaptability and classification accuracy on new telescope data.

[0042] Through the above steps, this invention has created a unified representation and cross-survey transfer learning model for multi-source astronomical light curves, significantly improving the accuracy, stability, and scalability of automatic variable star classification. This method can be widely applied in time-domain astronomy fields such as variable star identification, anomalous light variation detection, and survey data archiving, providing an efficient intelligent analysis tool for future next-generation survey missions such as LSST and CSST.

[0043] VII. Method Results Presentation This section presents the comparison results between the proposed method and other training methods, such as... Figure 3 As shown in the figure, Model represents the model trained using the method presented in this paper. The specific process is as follows: first, pre-training is performed using TESS unlabeled survey data, and then fine-tuning is performed on labeled data from TESS, ZTF, and Gaia, respectively. As a comparison, Long Short-Term Memory (LSTM) networks and Transformer models trained from scratch are directly trained under supervision based on their corresponding labeled data, without introducing a pre-training process.

[0044] Figure 3 The top subplot shows the training results using TESS labeled data, the middle subplot shows the training results using ZTF labeled data, and the bottom subplot shows the training results using Gaia labeled data. The horizontal axis of each subplot represents the number of training epochs corresponding to the validation set inference stage, the left vertical axis represents the loss value of validation set inference, and the right vertical axis represents the Macro F1 metric of validation set inference.

[0045] Macro F1 is a comprehensive performance evaluation metric for multi-class classification tasks, calculated based on precision and recall for each class. Precision measures the accuracy of the model's predictions, while recall reflects the model's ability to cover real samples. This metric first combines precision and recall at the class level to obtain the corresponding F1 value, and then averages the results across all classes. Because each class has the same weight in the calculation, Macro F1 is unaffected by differences in the number of samples per class, and can more comprehensively reflect the model's overall classification performance in multi-class scenarios.

[0046] from Figure 3 As can be seen, the Model (trained using the method presented in this paper) outperforms LSTM and the Transformer model trained from scratch in terms of Macro F1 scores across different datasets. Furthermore, since the Model is pre-trained on unlabeled survey data, only a limited number of fine-tuning epochs are needed during the supervised training phase to achieve superior performance. Compared to LSTM and Transformer models, which rely entirely on labeled data for training, this significantly reduces the computational resources and data costs required for training, demonstrating the comprehensive advantages of this method in terms of both performance and efficiency.

[0047] Figure 4 This illustration shows a more specific hardware structure diagram of an electronic device provided in an embodiment of this application. The device may include: a processor 302, a memory 304, an input / output interface 306, a communication interface 308, and a bus 310. The processor 302, memory 304, input / output interface 306, and communication interface 308 are interconnected internally via the bus 310.

[0048] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the methods described in this application.

[0049] This application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the methods described in the embodiments.

[0050] Those skilled in the art should understand that the embodiments described herein are for the purpose of helping readers understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. A method for representing and classifying astronomical light curves based on contrastive learning, characterized in that, Includes the following steps: Obtain a dataset of light curves from different astronomical survey telescopes; After preprocessing the optical curve dataset, optical curve data in a unified format is generated. The light curve data is input into a deep representation model based on the Transformer encoder architecture. This model uses a self-attention mechanism to extract consistent temporal features from the light curve data and performs feature alignment on the temporal features under unsupervised conditions through a hierarchical contrastive learning mechanism, thereby obtaining general features of light curves that can be transferred across surveys and completing the pre-training of the deep representation model. The hierarchical contrastive learning mechanism includes temporal-level contrastive learning and sample-level contrastive learning. The deep representation model is supervised fine-tuned using labeled variable star sample data to achieve variable star classification task; The system automatically reads the light curve data file to be processed through a unified inference interface and outputs the target category and confidence score.

2. The method for characterizing and classifying astronomical light curves based on contrastive learning according to claim 1, characterized in that, The preprocessing method for the light curve dataset is as follows: first, outliers are removed from the observation points according to the 3sigma criterion; then, the time series and the flow series are subjected to minimum-maximum normalization respectively; and finally, masking is used to align the differences in the length of light curve sequences from different survey sources.

3. The method for characterizing and classifying astronomical light curves based on contrastive learning according to claim 1, characterized in that, The deep representation model is a deep contrastive learning model based on the Transformer encoder structure.

4. The method for characterizing and classifying astronomical light curves based on contrastive learning according to claim 1, characterized in that, The time-level contrastive learning steps include: (a) Randomly select two partially overlapping subsequences from the same light curve; (b) Input the two subsequences into the deep representation model respectively to obtain the corresponding time series feature vectors; (c) In the overlapping region of two feature sequences, based on the alignment relationship of time indices, the feature vectors at positions with the same time index are defined as positive sample pairs, and the feature vectors at positions with different time indices are defined as negative sample pairs. (d) Train a deep representation model based on the positive and negative sample pairs.

5. The method for characterizing and classifying astronomical light curves based on contrastive learning according to claim 1, characterized in that, The sample-level contrastive learning steps include: (a) Input multiple optical curves simultaneously in a training batch; (b) For each light curve, generate two enhanced sequences with partially overlapping regions, and input all enhanced sequences into the depth representation model to obtain its time series feature vector; (c) On the aligned time index of the feature sequence, the feature vectors of two enhancement sequences from the same light curve with the same time index are defined as positive sample pairs; the feature vectors from different light curves but with the same time index are defined as negative sample pairs. (d) Train a deep representation model based on the positive and negative sample pairs.

6. The method for characterizing and classifying astronomical light curves based on contrastive learning according to claim 1, characterized in that, The unified inference interface supports multi-format data input and automatic model loading. When a new light curve is input, the system automatically performs data preprocessing and classification operations, and the predicted category and confidence score are output simultaneously in the model inference results.

7. The method for characterizing and classifying astronomical light curves based on contrastive learning according to claim 1, characterized in that, The method is applied to the light curves of different sky surveys to achieve transfer learning across telescope data. The migration process is achieved by maintaining the consistency of feature space distribution and using contrast loss constraints, making it suitable for variable star identification under different time sampling rates and observation conditions.

8. An electronic device comprising a memory and a processor, characterized in that, The memory is coupled to the processor; wherein the memory is used to store program data, and the processor is used to execute the program data to implement the astronomical light curve characterization and classification method based on contrastive learning as described in any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the astronomical light curve characterization and classification method based on contrastive learning as described in any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the astronomical light curve characterization and classification method based on contrastive learning as described in any one of claims 1 to 6.

Images