Methods, apparatus and electronic devices for time-series data processing based on generative adversarial networks

Abstract

This application discloses a time-series data processing method, apparatus, and electronic device based on generative adversarial networks (GANs), relating to the fields of medical information and artificial intelligence. The method includes: generating N candidate datasets based on previous time-series data and random noise within a first time period using N target generators; different candidate datasets correspond to different categories and different target generators, and each candidate dataset includes K predicted sequences; determining a target dataset from the N candidate datasets based on the similarity of the K predicted sequences in each candidate dataset; selecting a predicted sequence from the target dataset as the target predicted sequence for a second time period, and using the category corresponding to the target dataset as the target category. This application solves the technical problem in existing technologies where low prediction accuracy is caused by insufficient information in the early stages of time-series prediction.

Description

Technical Field

[0001] This application relates to the fields of medical information and artificial intelligence, and more specifically, to a method, apparatus, and electronic device for processing time-series data based on generative adversarial networks. Background Technology

[0002] With the development of computer technology, there is a large demand in the field of information technology for applications that require prediction based on historical time series data.

[0003] For time series modeling, as the length of the input sequence increases, the amount of information also increases, and the trend of data changes over time becomes more obvious, thus improving the model's prediction accuracy. However, in practical applications, it is often desirable to obtain high-accuracy prediction results with a shorter input sequence, which is beneficial for timely early warning and response. But in the early stages of time series forecasting, the length of high-quality time series data is limited, and the input information is insufficient. Existing single prediction models often struggle to make accurate predictions due to insufficient input information.

[0004] There is currently no effective solution to the above problems. Summary of the Invention

[0005] This application provides a time series data processing method, apparatus, and electronic device based on generative adversarial networks, to at least solve the technical problem in the prior art where the prediction accuracy is low due to insufficient information in the early stages of time series prediction.

[0006] According to one aspect of the embodiments of this application, a time-series data processing method based on adversarial generative networks is provided, comprising: generating N candidate datasets based on previous time-series data and random noise within a first time period using N pre-trained target generators; wherein different candidate datasets correspond to different categories and different target generators, each candidate dataset includes K predicted sequences, each predicted sequence being time-series data corresponding to a second time period predicted by a target generator based on previous time-series data and random noise, the second time period being after the first time period, and N and K being integers greater than 1; determining a target dataset from the N candidate datasets based on the similarity between the K predicted sequences in each candidate dataset; selecting a predicted sequence from the target dataset as the target predicted sequence for the second time period, and using the category corresponding to the target dataset as the target category.

[0007] Optionally, the target dataset is determined from the N candidate datasets based on the similarity between the K predicted sequences in each candidate dataset, including: determining the confidence level of the candidate dataset based on the similarity between the K predicted sequences in each candidate dataset, wherein the similarity between the K predicted sequences in each candidate dataset is positively correlated with the confidence level of the candidate dataset; and selecting the candidate dataset with the highest confidence level from the N candidate datasets as the target dataset.

[0008] Optionally, selecting a prediction sequence from the target dataset as the target prediction sequence for the second time period includes: selecting the median sequence of the K prediction sequences included in the target dataset as the target prediction sequence, or calculating the average sequence of the K prediction sequences included in the target dataset as the target prediction sequence.

[0009] Optionally, the training steps for the N target generators include: acquiring time-series data; dividing the time-series data into an earlier sequence and a later sequence according to time order; constructing N generative adversarial neural networks (GANs), wherein each GAN includes a generator and a discriminator, the generator of each GAN is used to generate a predicted sequence of a certain category, and the discriminator of the GAN constrains the generated predicted sequence to conform to the data characteristics of that category, and constrains the similarity between the generated predicted sequence and the later sequence to be greater than a preset threshold; using the earlier sequence as the input data of the N GANs and the later sequence as the training target of the N GANs, iteratively training the N GANs until all N GANs enter a convergent state, and using the N generators included in the N GANs as the N target generators.

[0010] Optionally, the generator's loss function includes: an adversarial loss subfunction, a reconstruction loss subfunction, and weights for balancing the adversarial loss subfunction and the reconstruction loss subfunction. The adversarial loss subfunction is used to optimize the fidelity and class representativeness of the generator's generated data by maximizing the probability that the predicted sequence generated by the generator is misclassified as a real subsequent sequence. The reconstruction loss subfunction is used to measure the difference between the predicted sequence generated by the generator and the real subsequent sequence.

[0011] Optionally, the known terms in the generator's loss function include the subsequent sequence and the noise distribution, where the noise distribution is the information distribution source of random mutations introduced by the generator in the process of generating the predicted sequence; the unknown terms in the generator's loss function are the neural network parameters of the generator.

[0012] Optionally, the discriminator's loss function includes: a real sample loss sub-function and a generated sample loss sub-function, wherein the real sample loss sub-function is used to maximize the discriminator's score on the authenticity of samples from real data; and the generated sample loss sub-function is used to minimize the discriminator's score on the authenticity of fictitious samples generated by the generator.

[0013] According to another aspect of the embodiments of this application, a time-series data processing apparatus based on adversarial generative networks is also provided, comprising: a first processing unit, configured to generate N candidate datasets based on previous time-series data and random noise within a first time period using N pre-trained target generators; wherein different candidate datasets correspond to different categories and different target generators, each candidate dataset includes K predicted sequences, each predicted sequence being time-series data corresponding to a second time period predicted by a target generator based on previous time-series data and random noise, the second time period being after the first time period, and N and K being integers greater than 1; a second processing unit, configured to determine a target dataset from the N candidate datasets based on the similarity between the K predicted sequences in each candidate dataset; and a third processing unit, configured to select one predicted sequence from the target dataset as the target predicted sequence for the second time period, and to use the category corresponding to the target dataset as the target category.

[0014] According to another aspect of the embodiments of this application, a computer-readable storage medium is also provided, which stores a computer program, wherein when the computer program is executed, the device where the computer-readable storage medium is located performs the above-described time-series data processing method based on generative adversarial networks.

[0015] According to another aspect of the embodiments of this application, an electronic device is also provided, including one or more processors and a memory, the memory being used to store one or more programs, wherein when the one or more programs are executed by one or more processors, the one or more processors cause the one or more processors to perform the above-described temporal data processing method based on generative adversarial networks.

[0016] In this embodiment, the medical data analysis system first generates N candidate datasets based on pre-trained N target generators and earlier time-series data and random noise within a first time period. Different candidate datasets correspond to different categories and different target generators. Each candidate dataset includes K predicted sequences, each predicted sequence being time-series data corresponding to a second time period predicted by a target generator based on earlier time-series data and random noise. The second time period follows the first time period. Then, based on the similarity between the K predicted sequences in each candidate dataset, a target dataset is determined from the N candidate datasets. Finally, a predicted sequence is selected from the target dataset as the target predicted sequence for the second time period, and the category corresponding to the target dataset is designated as the target category.

[0017] As described above, the medical data analysis system utilizes N target generators trained for different categories. Based on previous time-series data and random noise within the first time period, it generates N candidate datasets, enabling extensive sampling of N future development paths. Due to the introduction of random noise, each target generator outputs K distinct prediction sequences, simulating reasonable variations within the category and facilitating the transformation of prediction from point estimation to distribution estimation. Furthermore, by determining the target dataset based on the similarity between prediction sequences, this generational diversity can be used to assess the system's own determinism. Calculating the similarity of the K prediction sequences within each candidate dataset quantifies the system's predictive consistency for each category. Selecting the candidate dataset with the highest similarity as the target dataset allows the system to choose the most stable and certain prediction from multiple possible but uncertain ones, making the decision regarding the target category more dependent on the stability of the prediction process rather than a single, low-probability numerical output. Finally, selecting the target prediction sequence from the target dataset smooths out random fluctuations, contributing to a robust prediction value. By transforming prediction accuracy into deterministic assessment and utilizing internal consensus within the system as a basis for decision-making, it helps to screen and output relatively more reliable prediction results in the early stages of data processing, thereby solving the technical problem of low prediction accuracy due to insufficient information in the early stages of time series prediction in existing technologies. Attached Figure Description

[0018] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments of this application and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0019] Figure 1 This is a schematic diagram of an optional time-series data processing method based on a generative adversarial network according to an embodiment of this application;

[0020] Figure 2 This is a schematic diagram of an optional time-series data processing apparatus based on a generative adversarial network according to an embodiment of this application. Detailed Implementation

[0021] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0022] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0023] According to an embodiment of this application, a method embodiment of a time-series data processing method based on a generative adversarial network is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0024] It should be noted that the information collected in this application (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for display, data used for analysis, etc.) are information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure, and application of this data all comply with relevant laws, regulations, and standards, necessary confidentiality measures have been taken, and they do not violate public order and good morals. Corresponding access points are provided for users to choose to authorize or refuse. For example, interfaces are set up between this system and relevant users or organizations, providing users with corresponding access points to choose to agree to or refuse automated decision-making results; if the user chooses to refuse, the process proceeds to the expert decision-making stage.

[0025] According to the embodiments of this application, a data analysis system (hereinafter referred to as the system) can be used as the execution subject of the time-series data processing method based on adversarial generative networks in the embodiments of this application. The system can be a software system or an embedded system combining software and hardware. Of course, the execution subject of the method in the embodiments of this application can also be other forms of execution subject, such as devices, equipment, etc. It should be known by those skilled in the art that this application does not particularly limit the specific form of the execution subject.

[0026] Figure 1 This is a schematic diagram of a time-series data processing method based on a generative adversarial network according to an embodiment of this application. Figure 1 As shown, the method includes the following steps:

[0027] Step S101: Using N pre-trained target generators, N candidate datasets are generated based on the previous time series data and random noise within the first time period. Different candidate datasets correspond to different categories and different target generators. Each candidate dataset includes K predicted sequences. Each predicted sequence is the time series data corresponding to the second time period predicted by a target generator based on the previous time series data and random noise. The second time period is located after the first time period. N and K are both integers greater than 1.

[0028] Optionally, the target generator can refer to a trained machine learning model. The function of the target generator can be to receive historical time-series data such as previous time-series data and random inputs such as random noise, and output predictions for time-series data in subsequent time periods. In the embodiments of this application, each target generator can be trained to generate time-series data patterns of a specific category.

[0029] Optionally, the prior time series data includes, but is not limited to, time series data on vital signs, time series data on medical images, time series data on network traffic, time series data on network packet loss rate, and other data whose time series relationships can characterize the data change patterns.

[0030] Optionally, the N candidate datasets can be two candidate datasets, for example, one is a negative dataset and the other is a positive dataset. The N candidate datasets can also be more categories, not limited to just two, for example, negative, positive and invalid in three categories.

[0031] Optionally, during the prediction process, N candidate datasets can be generated. For example, the earlier time series data of the object to be predicted within the first time period can be obtained first. For example, data from the first 3 hours, then processed by a pre-trained first target generator corresponding to the negative category. Based on previous time series data With multiple different random noise samples ,generate The predicted sequences constitute the candidate dataset for the negative category. And through a pre-trained second target generator corresponding to the positive category. Based on previous time series data With multiple different random noise samples ,generate The predicted sequences constitute the candidate dataset for the positive category. .

[0032] By injecting different random noise samples into the generation process of each predicted sequence ,For example Having different noise samples helps the same target generator to output diverse prediction sequences, thereby simulating possible temporal change patterns within the corresponding category.

[0033] By configuring a corresponding generator for each candidate category and introducing diversity using random noise, it is beneficial to generate multiple possible prediction directions that conform to different category patterns for the same historical data. Each possible future category direction can be fully sampled, rather than just providing a single, low-probability point estimate. Generating multiple prediction sequences and forming a candidate dataset facilitates subsequent evaluation of the stability of the target generator's predictions by analyzing the consistency between sequences.

[0034] Step S102: Determine the target dataset from the N candidate datasets based on the similarity between the K predicted sequences in each candidate dataset.

[0035] Optionally, similarity can refer to a quantitative indicator used to measure the closeness or consistency between two or more predicted sequences. The higher the similarity, the more similar the sequences are. In the embodiments of this application, the pairwise similarity between all K predicted sequences in a candidate dataset can be calculated, and a metric that can represent the overall consistency of the dataset can be obtained by taking the average or other comprehensive methods. This metric can be positively correlated with the confidence level.

[0036] Optionally, if the K predicted sequences within a candidate dataset corresponding to a category are highly similar to each other, it indicates that the system has a stable and consistent prediction of the future pattern of that category based on the limited prior time-series data and other early information, meaning the system has high confidence in the prediction result. Conversely, if the sequences differ greatly, it indicates that the model prediction is uncertain and has low confidence. By comparing the similarity of all candidate datasets and selecting the dataset corresponding to the most consistent category as the target dataset, it helps to identify predictions with high internal consensus within the system, even if their absolute accuracy may be limited, in the early stages when information is insufficient. This allows for the selection of more reliable prediction directions and avoids blindly trusting single, potentially unstable prediction results.

[0037] Step S103: Select a prediction sequence from the target dataset as the target prediction sequence for the second time period, and use the category corresponding to the target dataset as the target category.

[0038] Optionally, after determining the most reliable target dataset and corresponding target category based on the similarity between predicted sequences, the system still needs to provide a specific, single future sequence value that can be used for subsequent decision-making or analysis, rather than retaining a set of K possibilities. Therefore, it is necessary to select the most representative predicted sequence from the K predicted sequences included in the target dataset as the final target predicted sequence. This helps eliminate random fluctuations introduced by random noise during the generation process, resulting in a more stable prediction that better reflects the central trend. Furthermore, explicitly defining the category corresponding to the target dataset as the target category ensures that the system's output not only includes specific time-series predicted values ​​but also clearly indicates the attribute attribution of this prediction to the target category, thus providing users with complete and relevant prediction information. This embodiment of the application can converge the prediction process into a clear output that combines category labels and time-series values, helping to improve the practicality and interpretability of the prediction results.

[0039] Of course, this application embodiment also supports randomly selecting a prediction sequence from the target dataset as the target prediction sequence for the second time period, or having the user specify a prediction sequence from the target dataset as the target prediction sequence for the second time period.

[0040] In one optional embodiment, a target dataset is determined from N candidate datasets based on the similarity between K predicted sequences in each candidate dataset. This includes: the medical data analysis system can determine the confidence level of a candidate dataset based on the similarity between the K predicted sequences in each candidate dataset, wherein the similarity between the K predicted sequences in each candidate dataset is positively correlated with the confidence level of the candidate dataset. Then, the candidate dataset with the highest confidence level is selected as the target dataset from the N candidate datasets.

[0041] This application's embodiments calculate the similarity between K predicted sequences in each candidate dataset and establish a positive correlation between similarity and confidence. This transforms the predictive stability of the target generator for the future development of the same category into a confidence score. Higher similarity between predicted sequences indicates a more concentrated and less divergent prediction result from the target generator based on current prior time-series data; therefore, the confidence of the candidate dataset is correspondingly higher. Selecting the candidate dataset with the highest confidence from N candidate datasets as the target dataset allows for the selection of the category with the strongest internal consensus and highest predictive certainty as the final prediction category.

[0042] Optionally, during the data preparation phase, the early time series data within the first time period (e.g., the patient's vital signs time series data for the first 3 hours) can be acquired, and the subsequent time series data within the second time period corresponding to the early time series data (e.g., the time series data for the next 6 hours) and the target category to which the time series data belongs, such as whether sepsis will occur, can be predicted. At the same time, the confidence level of the prediction can be calculated.

[0043] For example, consider the time-series data of vital signs (such as heart rate, blood pressure, and blood oxygen saturation) of patients in the Intensive Care Unit (ICU). In the first few hours (i.e., the first time period), it may not be possible to clearly determine whether a patient will develop sepsis (i.e., whether they belong to the negative or positive category). However, in the subsequent hours (i.e., the second time period), the differences in data characteristics between different categories will become significant. This embodiment of the application facilitates the simultaneous prediction of categories and subsequent time-series data based on this earlier time-series data.

[0044] It should be noted that prior time series data can be not only vital signs time series data, but also other types of time series data, such as time series data about medical images, time series data about network traffic, time series data about network packet loss rate, and other types of data whose change patterns can be characterized by time series relationships.

[0045] This application integrates the prediction of subsequent sequences of time-series data, the prediction of time-series data categories, and the calculation of prediction confidence into a single processing flow. This not only predicts the time-series data values ​​(i.e., the target prediction sequence) for a future time period (i.e., the second time period) based on previous time-series data, but also simultaneously predicts the category to which the time-series data belongs (i.e., the target category) and outputs the confidence level of the prediction result. Users can evaluate the reliability of the model's output based on this confidence level. If the model outputs a low confidence level in the early stages when there is limited sequence information, predictions and evaluations can be performed again after collecting longer time-series data, thereby improving the model's practicality and safety in critical scenarios such as early diagnosis and treatment.

[0046] For example, after generating N candidate datasets, intra-class dissimilarity can be calculated. For instance, the similarity between the K predicted sequences contained in each candidate dataset can be calculated. If the predicted sequences within a candidate dataset corresponding to a certain class have low dissimilarity, it indicates that the target generator has high consistency in predicting that class, and thus suggests high confidence in the predictions for that class.

[0047] For dissimilarity metrics, Dynamic Time Warping (DTW) distance can be used as a measure of the similarity between predicted sequences. DTW distance is suitable for cases where temporal alignment is not strict. It calculates the minimum cumulative distance between two sequences after time warping alignment on the time axis. The smaller the DTW distance, the more similar the two sequences are in terms of morphology.

[0048] Taking the candidate dataset for the positive category as an example, the confidence level of the candidate dataset for the positive category. The calculation can be referenced from formula (1):

[0049] Formula (1);

[0050] Formula (1) calculates the average dynamic time-warped distance between all pairs of different predicted sequences in the positive candidate dataset. If the K predicted sequences generated by the target generator for the positive category are highly similar to each other, i.e., the average dynamic time-warped distance is small, it indicates that the model has high consistency in predicting the positive category, i.e., high confidence. Conversely, if the predicted sequences are significantly different, i.e., the average dynamic time-warped distance is large, the consistency is low, i.e., low confidence. The dynamic time-warped distance is only an exemplary indicator. In practical applications, other indicators can also be used to calculate the similarity between predicted sequences, such as mean squared error (MSE).

[0051] In one optional embodiment, selecting a predicted sequence from the target dataset as the target predicted sequence for the second time period includes: selecting the median sequence of the K predicted sequences included in the target dataset as the target predicted sequence, or calculating the average sequence of the K predicted sequences included in the target dataset as the target predicted sequence.

[0052] Optionally, the target dataset itself includes K diverse prediction sequences introduced by random noise, while the final output requires a definite sequence for subsequent analysis or presentation. Choosing the median sequence or calculating the average sequence can converge multiple possibilities into a representative result. Choosing the median sequence is insensitive to extreme values ​​or outliers in the data, helping to obtain a robust result that better reflects the central tendency of the prediction sequences; while calculating the average sequence gives equal weight to all K prediction sequences, helping to smooth out random fluctuations and obtain a smooth prediction trajectory. Both methods effectively utilize all the information in the target dataset, reducing the random bias that may exist in a single generated result through aggregation operations, thereby outputting a numerically more stable and statistically representative target prediction sequence, which enhances the reliability and practicality of the final prediction result.

[0053] Optionally, after calculating intra-class dissimilarity, class decision and sequence prediction can be performed. For example, the rules for class decision include: selecting the candidate dataset with the highest confidence from N candidate datasets based on the confidence level of each candidate dataset. This candidate dataset with the highest confidence level is the determined target dataset, and the class corresponding to this target dataset is the predicted target class.

[0054] For example, for two candidate datasets, negative and positive, the decision rule can be expressed as formula (2):

[0055] Formula (2);

[0056] in, Used to indicate category The difference index values ​​corresponding to the candidate datasets, This is used to represent the selected target category. Formula (2) indicates that the selection has less discrepancy (i.e., confidence level). The category with the higher value is used as the target category.

[0057] Optionally, during sequence prediction, a predicted sequence can be selected from the defined target dataset corresponding to the target category as the target predicted sequence for the second time period. The selection method may include: calculating the median sequence of the K predicted sequences included in the target dataset and using this median sequence as the target predicted sequence; or, calculating the average sequence of the K predicted sequences included in the target dataset and using this average sequence as the target predicted sequence.

[0058] For example, if the determined target category is category c, then the target prediction sequence is... It can be obtained by calculating the median of all predicted sequences in the target dataset, which can be expressed as formula (3):

[0059] Formula (3);

[0060] in, Used to indicate a specific target category.

[0061] Furthermore, based on the K predicted sequences in the target dataset, the confidence interval of the target predicted sequence can be constructed by calculating the statistical quantiles (e.g., the 5th percentile and the 95th percentile). This confidence interval can characterize the range of uncertainty in the prediction and is helpful in intuitively showing the possible fluctuation range of the predicted value.

[0062] For example, the training phase may include:

[0063] Inputs include: a patient's early time-series data (vital signs in the first 3 hours), which show no clear signs of sepsis.

[0064] The generated results include:

[0065] 1. The negative category target generator, based on the previous time series data, can generate K=100 predicted sequences, forming a negative candidate dataset. The predicted sequences in this negative candidate dataset can exhibit a stationary vital sign pattern. Calculations show that the average dynamic time-normalized distance among the 100 predicted sequences in this negative candidate dataset is small, indicating that these predicted sequences are highly similar to each other, thus the confidence level of the negative category is high.

[0066] 2. The positive category target generator, based on the same prior time-series data, can generate K=100 predicted sequences, forming a positive candidate dataset. The predicted sequences in this positive candidate dataset exhibit diverse patterns; some sequences show a sudden drop in blood pressure, while others show no significant change. Calculations show that the average dynamic time-normalized distance between the predicted sequences in this positive candidate dataset is relatively large, indicating significant differences among these predicted sequences; therefore, the confidence level for the positive category is low.

[0067] Conclusion: The comparison revealed that the negative candidate dataset exhibited higher predictive consistency, i.e., a smaller DTW mean distance. Therefore, this negative candidate dataset was selected as the target dataset, and the predicted target category was negative. Ultimately, the median sequence of 100 predicted sequences from this negative candidate dataset can be used as the target predicted sequence for the patient's future vital signs.

[0068] It should be noted that the above examples use negative and positive categories. The embodiments of this application can also be applied to scenarios with more categories. For scenarios with N (N>1) categories, a Generative Adversarial Network (GAN) model can be trained independently for each category, and the generator in the GAN can be used as the target generator for that category.

[0069] For example, the training phase may include:

[0070] The input includes: the target object's previous time series data (medical image time series data).

[0071] The generated results include:

[0072] 1. The first target generator (corresponding to the first category) can generate K predicted sequences based on the previous time series data, forming the first candidate dataset. The predicted sequences in the first candidate dataset can display the image features of medical images. Calculations show that the average dynamic time-normalized distance between the K predicted sequences in the first candidate dataset is small, indicating that these predicted sequences are highly similar to each other, thus the confidence level of the first category is high.

[0073] 2. The second target generator (corresponding to the second category) generates K predicted sequences based on the same prior time-series data, forming the second candidate dataset. The predicted sequences in the second candidate dataset exhibit diversity and significant differences in image features. Calculations show that the average dynamic temporal regularization distance between the predicted sequences in the second candidate dataset is relatively large, indicating significant differences among these predicted sequences; therefore, the confidence level of the second category is low.

[0074] Conclusion: The comparison revealed that the first candidate dataset exhibits higher prediction consistency, i.e., a smaller DTW average distance. Therefore, the first candidate dataset is selected as the target dataset, and the predicted target class is the first class. Ultimately, the median or average sequence of the K predicted sequences from the first candidate dataset can be used as the target prediction sequence for the target object.

[0075] For example, the training phase may include:

[0076] The input includes: the early timing data of a certain device (network traffic timing data within the first time period).

[0077] The generated results include:

[0078] 1. The first target generator (corresponding to the category of normal network traffic) generates K predicted sequences based on the previous time series data, forming the first candidate dataset. The predicted sequences in this candidate dataset can show that the device's network traffic is in a normal state. Calculations show that the average dynamic time warp distance between the K predicted sequences in this candidate dataset is small, indicating that these predicted sequences are highly similar to each other, and therefore have a high confidence level.

[0079] 2. The second target generator (corresponding to the category of network traffic anomalies, such as network traffic overload) generates K predicted sequences based on the same previous time-series data, forming the second candidate dataset. The predicted sequences in this candidate dataset exhibit diverse patterns, with significant differences in traffic characteristics between each sequence. Calculations show that the average dynamic time-warped distance between the predicted sequences in this candidate dataset is large, indicating significant differences among these sequences, and therefore, a low confidence level.

[0080] Conclusion: The comparison revealed that the candidate dataset corresponding to the first target generator exhibits higher prediction consistency, i.e., a smaller DTW average distance. Therefore, the candidate dataset corresponding to the first target generator is selected as the target dataset, and the predicted target category is the category indicating normal network traffic. Finally, a prediction sequence can be selected from the candidate dataset corresponding to the first target generator as the target prediction sequence.

[0081] In one optional embodiment, the training steps for the N target generators include: the medical data analysis system acquires time-series data and divides the time-series data into early and later sequences according to time order. Then, N generative adversarial neural networks (GANs) are constructed, each GAN including a generator and a discriminator. The generator of each GAN generates a predicted sequence for a specific category, while the discriminator constrains the generated predicted sequence to conform to the data characteristics of that category and ensures that the similarity between the generated predicted sequence and the later sequence is greater than a preset threshold. Finally, the early sequences are used as input data for the N GANs, and the later sequences are used as training targets for the N GANs. The N GANs are iteratively trained until they all converge, and the N generators included in the N GANs are used as the N target generators.

[0082] Optionally, dividing time series data into early and later sequences according to time order is beneficial for simulating scenarios where data from the first time period is predicted for the second time period. This division allows the training process to focus on learning the mapping relationship from historical segments to future segments, which helps the trained target generator to have temporal extrapolation capabilities.

[0083] Optionally, embodiments of this application can construct an independent generative adversarial neural network (GAN) for each candidate category, enabling the generator of each GAN to focus on learning and generating time-series data patterns that conform to the characteristics of the corresponding category. For example, two independent GANs can be trained, one for the negative category and the other for the positive category. The generator of each GAN is responsible for generating subsequent sequences for a specific category, while the discriminator ensures that the generated sequences are both realistic and conform to the category characteristics.

[0084] Optionally, the system may include a generator, which is a specific implementation of the target generator during the training phase. It is one of the core components of each generative adversarial neural network and is responsible for generating multiple possibilities of future sequences based on historical data.

[0085] The generator can receive two inputs, including: early timing data. and random noise Among them, the early time series data Characterizes historical sequences observed within the first time period, such as patient vital signs data from the first 3 hours. Random noise. The representation is sampled from a predefined distribution to control the diversity of the generated data. For example, the predefined distribution could be a standard normal distribution N(0, 1).

[0086] The generator can output a predicted sequence. Predicted sequence Used to represent time-series data that the generator model predicts may occur within a second time period, such as vital signs data for the next few hours.

[0087] Network structure: The specific neural network architecture of the generator can be a variety of networks suitable for modeling time series data, such as Long Short-Term Memory (LSTM), Transformer, or Temporal Convolutional Network (TCN), etc. This application does not limit this.

[0088] The generator's function is to learn the mapping from earlier time-series data and random noise to subsequent sequences. Here, random noise... The introduction of this method injects controllable randomness into the generation process, enabling the generator to produce multiple different but consistent prediction sequences that conform to the distribution characteristics of the data category by combining different random noise samples, given the same prior time series data input. This helps to simulate the temporal change patterns that different individuals in the same category may exhibit in the real world. For example, the specific changes in vital signs of different patients when they develop sepsis may vary from person to person.

[0089] Optionally, the system may include a discriminator, which is another core component paired with the generator in each generative adversarial neural network. The discriminator can act as an evaluator to assess the authenticity and class attributes of the input sequence.

[0090] The discriminator can receive a complete time sequence. As input, this time series can be derived from previous time series data. and the actual subsequent sequence The spliced-together real training data can also be the predicted sequences generated by the corresponding generator. Compared with previous time series data A complete sequence pieced together.

[0091] The discriminator can output two results: a sequence authenticity score and a sequence category score. These results can be output through a dual-task output layer. The sequence authenticity score indicates the probability that the complete input sequence comes from the real data distribution. The sequence category score indicates the probability that the complete input sequence belongs to the corresponding category of the discriminator; for example, a positive discriminator determines whether the input belongs to the positive category.

[0092] The specific neural network architecture of the discriminator is not limited to a particular form, but it needs to include network layers that can effectively extract temporal features (such as convolutional layers, recurrent neural network layers, etc.), as well as task layers that can output two judgment results: sequence authenticity score and sequence category score.

[0093] The discriminator's function is to rigorously evaluate the predicted sequences generated by the generator, thereby forcing the generator to continuously improve. During training, the discriminator attempts to maximize its ability to distinguish between real and generated sequences, and to determine the sequence category. The generator, on the other hand, attempts to generate sequences that can deceive the discriminator, i.e., to induce the discriminator to give the generator's generated sequences high realism scores and high corresponding category scores. This adversarial training process enables the generator to learn to generate predicted sequences that are not only numerically close to real data, but also highly realistic in terms of statistical distribution and category features.

[0094] Optionally, the system may include an independent generative adversarial network (GAN) training module. In this embodiment, an independent GAN can be trained for each category, forming N GAN neural networks. The network structure of the GAN training module includes: corresponding to the negative category, a GAN can be constructed, including a negative generator. A negative discriminant Corresponding to the positive category, another independent generative adversarial network is constructed, including a positive generator. A positive discriminant The decoupled architecture allows each category's generator to focus on learning the time-series data evolution patterns specific to its category, without being influenced by data patterns from other categories. For example, the positive generator... It can focus on learning patterns of vital sign changes associated with the occurrence and development of sepsis. After training, each generator in the generative adversarial network will serve as a target generator for a different category, used to generate candidate sequences during the prediction phase.

[0095] In one optional embodiment, the generator's loss function includes: an adversarial loss subfunction, a reconstruction loss subfunction, and weights for balancing the adversarial loss subfunction and the reconstruction loss subfunction. The adversarial loss subfunction is used to optimize the fidelity and class representativeness of the generator's generated data by maximizing the probability that the predicted sequence generated by the generator is misclassified as a true subsequent sequence. The reconstruction loss subfunction is used to measure the difference between the predicted sequence generated by the generator and the true subsequent sequence.

[0096] Alternatively, a single optimization objective often fails to balance the diversity of generated data with accurate fitting of real data. By introducing two sub-functions—adversarial loss and reconstruction loss—and balancing them with a single weight parameter, the generator's training process can be finely controlled, helping the generator achieve the desired balance between pursuing data realism and maintaining similarity to real samples.

[0097] In one alternative embodiment, the known terms in the generator's loss function include the subsequent sequence and the noise distribution, wherein the noise distribution is the information distribution source of random variations introduced by the generator in the process of generating the predicted sequence; the unknown terms in the generator's loss function are the neural network parameters of the generator.

[0098] Optionally, the subsequent sequence, as the true target output, can provide a benchmark for calculating the differences in the reconstruction loss function, which is beneficial for guiding the generator to learn accurate mapping. The noise distribution, as a pre-defined source of information for introducing random variations during the generation of prediction sequences, can inject controllable randomness into the generation process. Each time a generation occurs, a specific random noise vector is sampled from this noise distribution and input into the generator along with the previous sequence. This allows the same generator, faced with the same previous sequence, to output multiple different but reasonable prediction sequences due to different input random noise, thus simulating the inherent diversity or uncertainty of the data. Explicitly treating the noise distribution as a known term means that its probability density function is predefined; the training process does not change the distribution itself, but rather learns how to use samples sampled from this distribution to generate useful data.

[0099] Optionally, minimizing the loss function using optimization algorithms such as stochastic gradient descent is beneficial for finding the optimal set of neural network parameters that minimizes the loss function value. These optimal neural network parameters determine the complex nonlinear mapping relationship within the generator from inputs such as the initial sequence and random noise to outputs such as the predicted sequence.

[0100] Optionally, during the training phase of N target generators, the generator loss function can be determined first. This includes generators targeting the positive category. For example, the training objective of the generator is to minimize a comprehensive loss function. ,Should The function-guided generator learns to generate realistic sequences that conform to specific category features, and the generated sequences are numerically as close as possible to the real data, taking into account the overall loss function. Please refer to formula (4):

[0101] Formula (4);

[0102] in, Used to represent the distribution of random noise z from its distribution The mathematical expectation during sampling. In formula (4) Part of it is used to represent the loss against adversarial losses. Among them, This indicates the positive discriminator versus the generator. Based on previous time series data and noise The generated predicted sequence is scored. The closer the discriminator score is to 1, the more likely the discriminator considers the generated sequence to be a true positive sequence. The generator's goal is to minimize this loss term by optimizing its own parameters, striving to make the generated sequence able to deceive the corresponding discriminator, causing the discriminator to give a high score (i.e., a score close to 1), thereby forcing the generator to learn to generate sequences that are difficult for the discriminator to distinguish, have a high degree of realism, and have obvious class characteristics. In formula (4) Partially used to represent the reconstruction loss, which can be expressed as Mean Squared Error (MSE). This term directly measures the predicted sequence output by the generator. With the actual subsequent sequence Numerical differences. The introduction of reconstruction loss helps constrain the generator, ensuring that the generated sequences not only resemble real data in statistical properties but also maintain a high degree of similarity to real samples in specific numerical values, thus avoiding the generation of sequences that deviate too much from actual observations. The weight parameter, a positive scalar parameter, is used to balance the adversarial loss and reconstruction loss in the overall loss. The relative importance of [something]. For example, It can be set to 10.

[0103] Among the known terms during the training process are: complete real time-series data. (Can be divided into early time series data) and subsequent sequences ), and the distribution of random noise. Unknowns during training (i.e., parameters that need to be solved through optimization) include: the generator. The neural network parameters.

[0104] In one optional embodiment, the discriminator's loss function includes: a real sample loss sub-function and a generated sample loss sub-function, wherein the real sample loss sub-function is used to maximize the discriminator's score on the authenticity of samples from actual data; and the generated sample loss sub-function is used to minimize the discriminator's score on the authenticity of fictitious samples generated by the generator.

[0105] Optionally, in actual training, optimizing the real sample loss function can encourage the discriminator to output high scores for the real subsequent sequences of the input, which is beneficial to enhancing the discriminator's ability to identify and affirm real data features, enabling the discriminator to accurately internalize the patterns of the real data distribution.

[0106] Optionally, the generated sample loss function can require the discriminator to give low scores to the predicted sequences generated by the current generator. Optimizing the generated sample loss function can help train the discriminator to effectively identify defects or false features in the data synthesized by the generator, thereby improving its sensitivity in distinguishing between true and false data.

[0107] For example, the training objective of a discriminator is to maximize its ability to distinguish between real and generated sequences and to correctly determine the category attribute of a sequence.

[0108] The discriminator loss function can be found in formula (5):

[0109] Formula (5);

[0110] in, It can represent real time series data From its distribution The mathematical expectation during sampling. The true sample loss can be expressed as in formula (5). This section indicates that the discriminator is encouraged. Distribution of data from the true positive category sequence Give a high rating Minimizing the negative value of the true sample loss term is equivalent to maximizing the discriminator's ability to identify true positive sequences. The sample loss can be generated using formula (5). This section indicates that the discriminator is encouraged. For generator The generated predicted sequence gives a low score. Minimizing the negative value of the generated sample loss term is equivalent to maximizing the discriminator's ability to recognize the generated sequence.

[0111] Among them, the discriminator The overall training objective is to minimize the loss function. This requires the discriminator to simultaneously maximize the score for real samples and minimize the score for generated samples, thereby improving the discriminator's discrimination ability.

[0112] Known factors during the training process include: the data distribution of true positive samples. And when the generator parameters are fixed, the generator... The generated samples. Unknowns during training include: the discriminator. Neural network parameters .

[0113] For example, the training example is as follows:

[0114] Sample description: A positive training sample includes: the patient presents normal vital signs in the first 3 hours (first time period) (which can be used as the initial time series data), and in the following 6 hours (second time period) presents vital signs data with sepsis-related characteristics such as a sudden drop in blood pressure and an increase in heart rate (which can be used as the subsequent sequence).

[0115] Generator task: During the training process of this positive training sample, the generator... Receive the patient's data from the first 3 hours (i.e., the initial time series data). Incorporating random noise z, the generator attempts to generate a predicted sequence for the next 6 hours. Optimization of the generator requires satisfying two conditions: the first condition is that the generated predicted sequence satisfies the corresponding positive discriminator. Misclassified positive sequences can counteract loss-driven factors. The second condition is that the generated predicted sequence should be as close as possible in numerical value to the patient's actual vital signs data in the last 6 hours, including sepsis characteristics, which can reconstruct loss-driven factors.

[0116] See Figure 2 According to another aspect of the embodiments of this application, a time-series data processing apparatus based on generative adversarial networks is also provided, including: a first processing unit 201, a second processing unit 202 and a third processing unit 203.

[0117] The first processing unit 201 is used to generate N candidate datasets based on the previous time series data and random noise within the first time period using N pre-trained target generators. Different candidate datasets correspond to different categories and different target generators. Each candidate dataset includes K predicted sequences, and each predicted sequence is the time series data corresponding to the second time period predicted by a target generator based on the previous time series data and random noise. The second time period is located after the first time period, and N and K are both integers greater than 1. The second processing unit 202 is used to determine the target dataset from the N candidate datasets based on the similarity between the K predicted sequences in each candidate dataset. The third processing unit 203 is used to select one predicted sequence from the target dataset as the target predicted sequence for the second time period and to use the category corresponding to the target dataset as the target category.

[0118] Optionally, the second processing unit 202 includes: a first processing subunit, configured to determine the confidence level of the candidate dataset based on the similarity between K predicted sequences in each candidate dataset, wherein the similarity between K predicted sequences in each candidate dataset is positively correlated with the confidence level of the candidate dataset; and a second processing subunit, configured to select the candidate dataset with the highest confidence level from N candidate datasets as the target dataset.

[0119] Optionally, the third processing unit 203 includes: a third processing subunit, used to select the median sequence of the K predicted sequences included in the target dataset as the target predicted sequence, or to calculate the average sequence of the K predicted sequences included in the target dataset as the target predicted sequence.

[0120] Optionally, the first processing unit 201 includes: a first acquisition subunit for acquiring time series data; a fourth processing subunit for dividing the time series data into an earlier sequence and a later sequence according to time order; a fifth processing subunit for constructing N generative adversarial neural networks, wherein each generative adversarial neural network includes a generator and a discriminator, the generator of each generative adversarial network is used to generate a predicted sequence of a certain category, and the discriminator of the generative adversarial network constrains the generated predicted sequence to conform to the data characteristics of that category, and constrains the similarity between the generated predicted sequence and the later sequence to be greater than a preset threshold; and a sixth processing subunit for using the earlier sequence as input data for the N generative adversarial neural networks and the later sequence as training targets for the N generative adversarial neural networks, iteratively training the N generative adversarial neural networks until all N generative adversarial neural networks enter a convergent state, and using the N generators included in the N generative adversarial neural networks as N target generators.

[0121] Optionally, the fifth processing subunit includes a first processing module for processing the generator's loss function. The generator's loss function includes an adversarial loss subfunction, a reconstruction loss subfunction, and weights for balancing the adversarial and reconstruction loss subfunctions. The adversarial loss subfunction optimizes the fidelity and class representativeness of the generator's generated data by maximizing the probability that the predicted sequence generated by the generator is misclassified as a true subsequent sequence. The reconstruction loss subfunction measures the difference between the predicted sequence generated by the generator and the true subsequent sequence. The known terms in the generator's loss function include the subsequent sequence and the noise distribution, where the noise distribution is the information distribution source of random mutations introduced by the generator during the generation of the predicted sequence. The unknown terms in the generator's loss function are the neural network parameters of the generator.

[0122] Optionally, the fifth processing subunit includes a second processing module, which is used to process the discriminator's loss function. The discriminator's loss function includes: a real sample loss subfunction and a generated sample loss subfunction, wherein the real sample loss subfunction is used to maximize the discriminator's score on the authenticity of samples from real data; and the generated sample loss subfunction is used to minimize the discriminator's score on the authenticity of fictitious samples generated by the generator.

[0123] According to another aspect of the embodiments of this application, a computer-readable storage medium is also provided, which stores a computer program, wherein when the computer program is executed, the device where the computer-readable storage medium is located performs the above-described time-series data processing method based on generative adversarial networks.

[0124] According to another aspect of the embodiments of this application, an electronic device is also provided, including one or more processors and a memory, the memory being used to store one or more programs, wherein when the one or more programs are executed by one or more processors, the one or more processors cause the one or more processors to perform the above-described temporal data processing method based on generative adversarial networks.

[0125] According to another aspect of the embodiments of this application, a computer program product is also provided, including a computer program or instructions, which, when executed by a processor, implement the above-described temporal data processing method based on generative adversarial networks.

[0126] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0127] In the above embodiments of this application, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0128] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For instance, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0129] 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 units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0130] Furthermore, the functional units in the various embodiments of this application 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.

[0131] 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 this application, 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 described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), portable hard drive, magnetic disk, or optical disk.

[0132] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method for processing time series data based on a generative adversarial network, characterized in that, include: N pre-trained target generators are used to generate N candidate datasets based on the preceding time-series data and random noise within a first time period. Different candidate datasets correspond to different categories and different target generators. Each candidate dataset includes K predicted sequences, each predicted sequence being the time-series data corresponding to a second time period predicted by a target generator based on the preceding time-series data and random noise. The second time period is located after the first time period. N and K are both integers greater than 1. The preceding time-series data includes at least one of time-series data related to medical images and time-series data related to vital signs. Determining a target dataset from N candidate datasets based on the similarity between K predicted sequences in each candidate dataset includes: determining the confidence level of the candidate dataset based on the similarity between K predicted sequences in each candidate dataset, wherein the similarity between K predicted sequences in each candidate dataset is positively correlated with the confidence level of the candidate dataset; and selecting the candidate dataset with the highest confidence level from the N candidate datasets as the target dataset. Selecting a predicted sequence from the target dataset as the target predicted sequence for the second time period, and using the category corresponding to the target dataset as the target category, includes: selecting the median sequence of the K predicted sequences included in the target dataset as the target predicted sequence, or calculating the average sequence of the K predicted sequences included in the target dataset as the target predicted sequence. 2.The method of claim 1, wherein, The training steps for the N target generators include: Obtain time series data; The time series data is divided into an earlier sequence and a later sequence according to the time order. Construct N generative adversarial neural networks, where each generative adversarial neural network includes a generator and a discriminator. The generator of each generative adversarial network is used to generate a prediction sequence of a certain category, and the discriminator of the generative adversarial network constrains the generated prediction sequence to conform to the data characteristics of that category, and constrains the similarity between the generated prediction sequence and the subsequent sequence to be greater than a preset threshold. The preceding sequence is used as the input data of the N generative adversarial networks (GANs), and the subsequent sequence is used as the training target of the N GANs. The N GANs are iteratively trained until they all converge. The N generators included in the N GANs are then used as the N target generators.

3. The adversarial generative network-based time series data processing method according to claim 2, characterized in that, The generator's loss function includes: an adversarial loss subfunction, a reconstruction loss subfunction, and weights for balancing the adversarial loss subfunction and the reconstruction loss subfunction. The adversarial loss subfunction is used to optimize the fidelity and class representativeness of the generator's generated data by maximizing the probability that the predicted sequence generated by the generator is misclassified as the real subsequent sequence. The reconstruction loss subfunction is used to measure the difference between the predicted sequence generated by the generator and the real subsequent sequence. 4.The method of claim 2, wherein, The known terms in the loss function of the generator include the subsequent sequence and the noise distribution, wherein the noise distribution is the information distribution source of random mutations introduced by the generator in the process of generating the predicted sequence; the unknown terms in the loss function of the generator are the neural network parameters of the generator.

5. The adversarial generative network-based time series data processing method according to claim 2, characterized in that, The loss function of the discriminator includes: a real sample loss sub-function and a generated sample loss sub-function, wherein the real sample loss sub-function is used to maximize the discriminator's realism score for samples from actual data; and the generated sample loss sub-function is used to minimize the discriminator's realism score for fictitious samples generated by the generator.

6. An apparatus for processing time series data based on a generative adversarial network, used to implement the method for processing time series data based on a generative adversarial network in any one of claims 1 to 5, characterized in that, include: The first processing unit is used to generate N candidate datasets based on the previous time series data and random noise within the first time period using N pre-trained target generators. Different candidate datasets correspond to different categories and different target generators. Each candidate dataset includes K prediction sequences. Each prediction sequence is a time series data corresponding to the second time period predicted by a target generator based on the previous time series data and random noise. The second time period is located after the first time period. N and K are both integers greater than 1. The second processing unit is used to determine the target dataset from the N candidate datasets based on the similarity between the K predicted sequences in each candidate dataset; The third processing unit is configured to select a prediction sequence from the target dataset as the target prediction sequence for the second time period, and to use the category corresponding to the target dataset as the target category.

7. A computer readable storage medium characterized in that, The computer-readable storage medium stores a computer program, wherein when the computer program is executed, the device containing the computer-readable storage medium performs the time-series data processing method based on generative adversarial networks as described in any one of claims 1 to 5.

8. An electronic device, characterized in that, It includes one or more processors and a memory, the memory being used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to perform the time-series data processing method based on adversarial generative networks as described in any one of claims 1 to 5.

Images