Method, device and medium for constructing time series analysis system
By building a time series analysis system through federated learning, the issues of data security and privacy protection have been resolved, enabling more accurate time series analysis in the financial system and improving data utilization efficiency and analysis effectiveness.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- PING AN TECH (SHENZHEN) CO LTD
- Filing Date
- 2024-08-07
- Publication Date
- 2026-06-16
AI Technical Summary
Existing time series processing methods rely on high-quality centralized data, which cannot be effectively analyzed while considering data security and privacy protection. This results in inaccurate analysis results and makes it difficult for them to play a full role in the financial system.
A time series analysis system is constructed using federated learning. The model is optimized by using federated variable information from multiple analysis subsystems. The time series is decomposed and encoded using an Encoder-Decoder-Feder architecture to obtain global variable information and generate federated variable information, thereby enabling multi-party joint training.
While ensuring data privacy and security, it improves the accuracy and scalability of time series analysis, enabling better utilization of data resources from all parties, and is suitable for tasks such as prediction, classification, and anomaly detection in financial systems.
Smart Images
Figure CN119149603B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of artificial intelligence, and in particular to a method, apparatus, device and medium for constructing a time series analysis system. Background Technology
[0002] Time series data are data points arranged in chronological order, typically at consecutive time intervals such as daily, monthly, or yearly. This type of data is common in many fields, such as finance, meteorology, stock markets, and sales data. In financial systems, processing time series data for forecasting, classification, and anomaly detection can play a predictive and supportive role in risk control, sales, and management.
[0003] However, existing time series processing methods all use high-quality data collected by relevant institutions, without missing values. Furthermore, in real-world scenarios, different clients possess their own data, and for data security and privacy reasons, data is not aggregated in a centralized organization for reverse training. Additionally, the processing of time series data only addresses the time series itself, leading to inaccurate analysis results and limited applicability to the operation of financial systems. Summary of the Invention
[0004] This application discloses a method, apparatus, device, and medium for constructing a time series analysis system, aiming to address the problems of existing time series processing methods that rely on high-quality data collected by relevant institutions without missing values. Furthermore, in real-world scenarios, different clients possess their own data, and for data security and privacy reasons, data is not aggregated in a centralized organization for reverse training. Additionally, processing of time series data only addresses the time series itself, leading to inaccurate analysis results and limited applicability to the financial system.
[0005] Firstly, this application provides a method for constructing a time series analysis system. The time series analysis system to be constructed includes multiple analysis subsystems, each of which stores a time series prediction model to be trained. The time series prediction model is used to predict the time series corresponding to the analysis subsystem. The method includes:
[0006] Obtain the original time series corresponding to each analysis subsystem;
[0007] The original time series is input into the time series prediction model of the corresponding analysis subsystem. The time series prediction model decomposes the original time series and obtains global variable information.
[0008] Obtain global variable information for each analysis subsystem, and generate federated variable information based on multiple global variable information;
[0009] Each time series prediction model is optimized based on the information of the federal variables and the global variables corresponding to each analysis subsystem. The training of multiple time series prediction models is completed, and the constructed time series analysis system is obtained.
[0010] In some embodiments, the time series forecasting model includes a decomposition layer that inputs the original time series data into the time series forecasting model of the corresponding analysis subsystem, including:
[0011] The original time series is input into the decomposition layer of each analysis subsystem. The decomposition layer decomposes the original time series and obtains global variable information.
[0012] For example, the decomposition layer decomposes the original time series and also obtains first period information and first trend information; the time series prediction model also includes an encoding layer and a decoding layer; the method further includes: inputting the original time series into the encoding layer, the encoding layer encodes the original time series to obtain second period information and second trend information; inputting the first period information into the decoding layer, the decoding layer decodes the first period information to obtain third period information; obtaining period difference information between the third period information and the first period information, and trend difference information between the first trend information and the second trend information; optimizing the decomposition layer and updating global variable information based on the period difference information and trend difference information, so that the first period information approaches the third period information and the first trend information approaches the second trend information.
[0013] It should be noted that, in some embodiments, the encoding layer includes at least one periodic trend decomposition module; the original time series is input into the encoding layer, and the encoding layer encodes the original time series to obtain second periodic information and second trend information, including: inputting the original time series into the first periodic trend decomposition module to obtain sub-period information and second trend information; transmitting the sub-period information layer by layer to the remaining periodic trend decomposition modules, and outputting the second periodic information.
[0014] It should be noted that, in some embodiments, the time series prediction model further includes an attention layer; before obtaining the period difference information between the third period information and the first period information, the trend difference information between the first trend information and the second trend information, the model further includes: inputting the second period information and the third period information into the attention layer, the attention layer performing frequency domain correlation learning on the second period information and the third period information to obtain frequency domain correlation information; and updating the parameters of the decoding layer according to the frequency domain correlation information to update the third period information.
[0015] In some embodiments, optimizing each time series prediction model based on the federated variable information and the global variable information corresponding to each analysis subsystem further includes: obtaining the correlation information between the federated variable information and the global variable information corresponding to each analysis subsystem based on a self-attention mechanism; updating the global variable information corresponding to the analysis subsystem based on the correlation information, so as to optimize the parameters of the time series prediction model corresponding to the analysis subsystem based on the global variable information.
[0016] In some embodiments, obtaining the original time series corresponding to each analysis subsystem includes: obtaining multiple original time series collected by multiple analysis subsystems; constructing a local dataset corresponding to each analysis subsystem based on the original time series corresponding to each analysis subsystem; and obtaining original time series with the same time axis from the local dataset corresponding to each analysis subsystem.
[0017] Secondly, this application provides a device for constructing a time series analysis system. The time series analysis system to be constructed includes multiple analysis subsystems, each of which stores a time series prediction model to be trained. The time series prediction model is used to predict the time series corresponding to the analysis subsystem; including:
[0018] The sequence acquisition unit is used to acquire the original time series corresponding to each analysis subsystem;
[0019] The variable acquisition unit is used to input the original time series into the time series prediction model of the corresponding analysis subsystem. The time series prediction model decomposes the original time series and obtains global variable information.
[0020] The federated generation unit is used to obtain global variable information for each analysis subsystem and generate federated variable information based on multiple global variable information.
[0021] The completed unit is used to optimize each time series prediction model based on the information of the federated variables and the global variable information corresponding to each analysis subsystem, complete the training of multiple time series prediction models, and obtain the completed time series analysis system.
[0022] Thirdly, this application provides a computer device including a processor, a memory, and a computer program stored in the memory and executable by the processor. The memory stores a strategy model, and when the computer program is executed by the processor, it implements a method for constructing a time series analysis system as provided in any embodiment of this application.
[0023] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, causes the processor to implement a method for constructing a time series analysis system as provided in any embodiment of this application.
[0024] This application provides a method, apparatus, device, and medium for constructing a time series analysis system. The time series analysis system to be constructed includes multiple analysis subsystems. Each analysis subsystem stores a time series prediction model to be trained. The time series prediction model is used to predict the time series corresponding to the analysis subsystem. The provided method obtains the original time series corresponding to each analysis subsystem and inputs it into the time series prediction model of each analysis subsystem. The time series prediction model decomposes the original time series, obtains global variable information, and then generates federated variable information based on the multiple global variable information. Each time series prediction model is then optimized based on the federated variable information and the global variable information corresponding to each analysis subsystem, completing the training of multiple time series prediction models and obtaining the constructed time series analysis system.
[0025] Furthermore, the proposed method can be extended to existing encoder-decoder architectures, exhibiting good scalability. It only uploads global variable information, excluding periodic and trend information, thus offering better data privacy and security. Moreover, based on the federated approach, multiple participants can collaboratively train the model while considering global trends, better utilizing each party's data resources and knowledge. In summary, the system built using the method proposed in this application exhibits scalability, better privacy and security, and can learn more accurate latent variables from time-series data.
[0026] When applied to systems such as finance, the provided method can unite multiple sub-suppliers or data management parties in different regions for federated learning, enabling the trained time series analysis system to better perform time series prediction, classification, and anomaly detection in financial systems, playing a predictive and auxiliary role in risk control, sales, and management.
[0027] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description
[0028] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0029] Figure 1 This is a schematic flowchart illustrating the steps of a method for constructing a time series analysis system according to an embodiment of this application;
[0030] Figure 2 This is a schematic flowchart illustrating the steps of a method for obtaining raw time series data according to an embodiment of this application;
[0031] Figure 3 This is a schematic flowchart illustrating the steps of a parameter optimization method for a time series prediction model provided in an embodiment of this application;
[0032] Figure 4 This is a schematic diagram of the structure of a time series analysis system construction apparatus provided in an embodiment of this application;
[0033] Figure 5 This is a schematic block diagram of the structure of a computer device provided in an embodiment of this application.
[0034] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Detailed Implementation
[0035] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0036] The flowchart shown in the attached diagram is for illustrative purposes only and does not necessarily include all content and operations / steps, nor does it necessarily have to be performed in the order described. For example, some operations / steps can be broken down, combined, or partially merged, so the actual execution order may change depending on the actual situation.
[0037] It should be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the scope of the application. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.
[0038] It should be understood that, in order to clearly describe the technical solutions of the embodiments of this application, the terms "first" and "second" are used in the embodiments of this application to distinguish the same or similar items with substantially the same function and effect. Those skilled in the art will understand that the terms "first" and "second" do not limit the quantity or execution order, and the terms "first" and "second" are not necessarily different.
[0039] It should also be understood that the term "and / or" as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.
[0040] To facilitate understanding of the embodiments of this application, some terms involved in the embodiments of this application will be briefly explained below.
[0041] 1. Federated Learning: Also known as federated machine learning, federated learning, or consortium learning, federated learning is a machine learning framework that effectively helps multiple organizations use data and perform machine learning modeling while meeting user privacy, data security, and government regulations. As a distributed machine learning paradigm, federated learning effectively solves the data silo problem, allowing participants to jointly model data without sharing existing data. It technically breaks down data silos and enables AI collaboration.
[0042] Federated learning defines a machine learning framework that addresses the problem of collaboration among different data owners without exchanging data by designing virtual models. The virtual model is the optimal model that aggregates data from all parties, with each region using the model to serve its local goals. Federated learning requires that the modeling result closely approximate the traditional model, where data from multiple owners is aggregated in one place for modeling. Under a federated mechanism, all participants have equal status and can establish data-sharing strategies. Because data is not transferred, user privacy is not compromised, and data standards are not affected. This is done to protect data privacy and meet legal compliance requirements.
[0043] Federated learning has three main components: data sources, a federated learning system, and users. The relationship between these three is shown in the diagram. In a federated learning system, each data source preprocesses the data, collaboratively builds its learning model, and then feeds the output back to the user.
[0044] The following detailed description of some embodiments of this application is provided in conjunction with the accompanying drawings. Unless otherwise specified, the following embodiments and features can be combined with each other.
[0045] Time series analysis is a task involving modeling, forecasting, and analyzing time series data (i.e., the original time series in this application). A time series consists of data points arranged in chronological order, typically at consecutive time intervals, such as daily, monthly, or yearly. This type of data is common in many fields, such as finance, meteorology, stock markets, and sales data.
[0046] Here are some key aspects of time series tasks:
[0047] 1. Time Series Forecasting: This is the most common time series task. The goal is to predict future values based on past data points. This forecasting can be a single-step forecast (predicting the next data point) or a multi-step forecast (predicting multiple future data points).
[0048] 2. Anomaly Detection: In time series data, there may be anomalous or abrupt data points, which may indicate faults, abnormal events, or other important information. The task of anomaly detection is to detect and identify these anomalies.
[0049] 3. Time Series Classification: In some cases, it is necessary to classify time series data into different categories. For example, based on the patterns and characteristics of the time series, stock price fluctuations can be classified as either rising or falling.
[0050] 4. Seasonal Analysis: Time series data may exhibit seasonality, meaning that recurring patterns appear within specific time intervals. Seasonal analysis aims to identify these seasonal patterns.
[0051] 5. Rolling forecast: The model can be continuously updated and rolling forecasts can be made over time, which is a dynamic time series forecasting task.
[0052] 6. Time series generation: New time series samples can be generated based on historical data, which is useful in some synthetic data applications.
[0053] 7. Time Series Correlation Analysis: Time series data may be correlated with other data series or events. Time series correlation analysis helps to discover these correlations.
[0054] In the field of time series forecasting, current time series models are primarily based on Transformer methods, such as the Informer model (a novel transformer model for long-series time series forecasting). Transformer-based methods have significantly improved the results of time series forecasting.
[0055] Meanwhile, a novel decomposition architecture, Autoformer, has been designed in the existing technology. It features an automatic correlation mechanism, breaking away from the traditional preprocessing of series decomposition and transforming it into a fundamental internal block of deep models. This design endows Autoformer with the ability to progressively decompose complex time series. Furthermore, it combines Transformer with a seasonal trend decomposition method, where the decomposition method captures the global features of the time series, while Transformer captures a more detailed structure.
[0056] The above methods all employ the idea of decomposition, but they have the following problems:
[0057] 1. No consideration of real data: The datasets used in the article are all high-quality data collected by relevant institutions, with no missing values.
[0058] 2. Failure to consider real-world scenarios: In real-world scenarios, different clients possess their own data. For data security and privacy reasons, data would not be aggregated in a centralized organization or used for reverse training.
[0059] 3. The decomposition only yields seasonal trends, lacking other latent variables: Taking the real-world pork price curve as an example, in addition to seasonal trends, it is also affected by the Consumer Price Index (CPI) and the prices of other meats and vegetables.
[0060] To address the aforementioned issues and considering data security and privacy, this application proposes using federated learning to construct a time series analysis system that can better and more accurately learn the latent variables of time series data. This is a system architecture for time series learning that can solve problems related to multi-party participation, data privacy, and security, and can be termed Fed-former. This paper aims to propose a unified time series learning system that achieves the federation of transformer-based methods.
[0061] Please see Figure 1 , Figure 1 This is a schematic flowchart illustrating the steps of constructing a time series analysis system according to an embodiment of this application. The time series analysis system to be constructed includes multiple analysis subsystems, each of which stores a time series prediction model to be trained. The time series prediction model is used to predict the time series corresponding to the analysis subsystem.
[0062] It should be noted that the method for constructing the time series analysis system provided in this application can be used in financial systems, which involve large volumes of data, high time sensitivity, and multiple data providers (such as the financial system and multiple banks) and data management providers in multiple regions (such as subsidiaries in different regions). By constructing a time series analysis system, multiple subsystems can be combined to perform forecasting, classification, and anomaly detection on time series data, providing predictive and supportive functions in risk control, sales, and management. It can also be applied to any other system with large volumes of highly time-sensitive data and multiple subsystems, such as those in the medical and education fields. This application uses a financial scenario as an example for illustration.
[0063] like Figure 1 As shown, the proposed method for constructing a time series analysis system includes steps S101 to S104.
[0064] S101. Obtain the original time series corresponding to each analysis subsystem.
[0065] Specifically, when the method provided in this application is applied to a financial system, the multiple analysis subsystems can be multiple data providers of different types within the financial system, such as multiple banks acting as analysis subsystems. They can also be data providers corresponding to multiple regions within the financial system, such as data providers from different provinces or cities. This application does not impose any restrictions on the type or number of analysis subsystems. This application aims to use federated learning to enable multiple previously unconnected data providers to collaboratively construct a time series analysis system without compromising data privacy. By acquiring the time series data generated during the operation of each analysis subsystem—the raw time series—and processing it, training of each analysis subsystem can be completed without compromising data privacy.
[0066] In some embodiments, please refer to Figure 2 , Figure 2 This is a schematic flowchart illustrating the steps of a method for obtaining an original time series according to an embodiment of this application.
[0067] like Figure 2 As shown, the provided method includes steps S101a to S101c.
[0068] Step S101 a. Obtain multiple raw time series data collected by multiple analysis subsystems.
[0069] Step S101 b. Construct a local dataset for each analysis subsystem based on the original time series data for each analysis subsystem.
[0070] Step S101 c. Obtain the original time series with the same time axis from the local dataset corresponding to each analysis subsystem.
[0071] In financial systems, each analysis subsystem generates data continuously during its operation. To ensure standardized management of data from each subsystem, this application constructs a local dataset for each analysis subsystem based on its original time series data. Then, original time series with the same time axis are obtained from each local dataset and used in the construction method provided in this application. This ensures temporal consistency of training data across multiple systems, improving training effectiveness. Furthermore, due to differences in the devices used to acquire data from each subsystem, there are no restrictions on the data sampling frequency; therefore, this application only requires the time axis to be identical, i.e., the start and end times must be consistent.
[0072] S102. Input the original time series into the time series prediction model of the corresponding analysis subsystem. The time series prediction model decomposes the original time series and obtains global variable information.
[0073] Specifically, by inputting the original time series data from each analysis subsystem into the corresponding time series prediction model, the time series prediction model can decompose the input original time series and obtain the global variable information corresponding to the original time series. Therefore, the provided method extracts only the corresponding global variable information from the original time series, enabling the extraction of training samples for federated learning without compromising data privacy.
[0074] In some embodiments, the time series prediction model includes a decomposition layer, which inputs the original time series into the time series prediction model of the corresponding analysis subsystem, including: inputting the original time series into the decomposition layer of each analysis subsystem, the decomposition layer decomposing the original time series and obtaining global variable information.
[0075] By setting decomposition layers in the time series prediction model, such as the Feder layer, which includes a MoE (Mixed Experts) decomposition module, the original time series can be decomposed to obtain corresponding global variables G. Furthermore, by obtaining the global variables G of each analysis subsystem, federated learning can be performed on multiple analysis subsystems.
[0076] For example, the decomposition layer decomposes the original time series and also obtains first period information and first trend information (such as first period information S1 and first trend information T1); the time series prediction model also includes an encoding layer and a decoding layer (such as an Encoder layer and a Decoder layer); the method further includes: inputting the original time series into the encoding layer, the encoding layer encodes the original time series to obtain second period information and second trend information (such as second period information S2 and second trend information T2); inputting the first period information into the decoding layer, the decoding layer decodes the first period information to obtain third period information (such as third period information S3); obtaining period difference information between the third period information and the first period information, and trend difference information between the first trend information and the second trend information; optimizing the decomposition layer and updating the global variable information based on the period difference information and trend difference information, so that the first period information approaches the third period information and the first trend information approaches the second trend information.
[0077] Furthermore, this application utilizes a time series analysis system framework consisting of an Encoder-Decoder (encoding and decoding layers) and a Federer (decomposition layer). It can extract the third periodic information S3 and the second trend information T2 from the Encoder-Decoder portion. The differences between these extracted information and the first periodic information S1 and the first trend information T1 from the decomposition layer are calculated, thereby optimizing the decomposition layer so that its output periodic and trend information approximates the outputs of the encoding and decoding layers.
[0078] In addition, the provided method can also update the model parameters of Encoder, Decoder, and Feder in reverse based on the difference between the local true values and predicted values of the trend information and periodic information corresponding to the original time series.
[0079] It should be noted that, in some embodiments, the encoding layer includes at least one periodic trend decomposition module; the original time series is input into the encoding layer, and the encoding layer encodes the original time series to obtain second periodic information and second trend information, including: inputting the original time series into the first periodic trend decomposition module to obtain sub-period information and second trend information; transmitting the sub-period information layer by layer to the remaining periodic trend decomposition modules, and outputting the second periodic information.
[0080] Each analysis subsystem's time series forecasting model employs the aforementioned Encoder-Decoder-Fender architecture. The Encoder input passes through at least one period-trend decomposition module, each layer decomposing the signal into sub-period information and second trend information T2. Multiple period-trend decomposition modules successively input the corresponding sub-period information, ultimately outputting the second period information S2, which is then passed to the Decoder. T2 is not input to the Decoder. Furthermore, in the Decoder section, the input S2 is decoded to output the third period information S3.
[0081] It should be noted that, in some embodiments, the time series prediction model further includes an attention layer; before obtaining the period difference information between the third period information and the first period information, and the trend difference information between the first trend information and the second trend information, the model further includes: inputting the second period information and the third period information into the attention layer; the attention layer performing frequency domain correlation learning on the second period information and the third period information to obtain frequency domain correlation information; and updating the parameters of the decoding layer based on the frequency domain correlation information to update the third period information. By setting the attention layer, frequency domain correlation learning can be performed on the second period information and the third period information to obtain frequency domain correlation information; and the parameters of the decoding layer can be updated based on the frequency domain correlation information to update the third period information.
[0082] S103. Obtain global variable information for each analysis subsystem, and generate federated variable information based on multiple global variable information.
[0083] Specifically, the provided method obtains global variable information G for each analysis subsystem, and can determine federated variable information based on this global variable information G. For example, it can input multiple global variable information Gs into a pre-defined federated learning model to output federated variable information G′, or it can obtain federated variable information G′ by averaging multiple global variable information Gs. Furthermore, the federated variable information enables the optimization of each analysis subsystem.
[0084] S104. Optimize each time series prediction model based on the information of the federated variables and the global variable information corresponding to each analysis subsystem, complete the training of multiple time series prediction models, and obtain the constructed time series analysis system.
[0085] Specifically, based on the federated variable information G′ and the global variable information G corresponding to each analysis subsystem, the corresponding time series prediction model can be optimized, thereby completing the training of multiple time series prediction models and obtaining the trained time series analysis system. Furthermore, the provided method can be extended to existing encoder-decoder architectures, exhibiting good scalability. Simultaneously, by uploading only global variable information and excluding periodic and trend information, it offers better data privacy and security. Moreover, based on the federated approach, considering global trends, multiple participants can collaboratively train the model, better utilizing their data resources and knowledge. In summary, the system learned and constructed by the method proposed in this application has better scalability, privacy, and security, while also learning more accurate latent variables from time series data.
[0086] In some embodiments, optimizing each time series prediction model based on the federated variable information and the global variable information corresponding to each analysis subsystem further includes: obtaining correlation information between the federated variable information and the global variable information corresponding to each analysis subsystem based on a self-attention mechanism; updating the global variable information corresponding to the analysis subsystem based on the correlation information, so as to optimize the parameters of the time series prediction model corresponding to the analysis subsystem based on the global variable information. By obtaining correlation information between the federated variable information G′ and the global variable information G corresponding to each analysis subsystem through a self-attention mechanism, the parameters of the time series prediction model corresponding to the analysis subsystem can be optimized based on the global variable information.
[0087] When the provided method is applied to systems such as financial systems, it can unite multiple sub-suppliers or data management parties in different regions for federated learning. This enables the trained time series analysis system to better perform time series prediction, classification, and anomaly detection in financial systems, playing a predictive and auxiliary role in risk control, sales, and management.
[0088] like Figure 4 As shown, Figure 4This is a schematic diagram of a time series analysis system construction apparatus provided in an embodiment of this application. This apparatus is used to execute the aforementioned time series analysis system construction method. The time series analysis system to be constructed includes multiple analysis subsystems. Each analysis subsystem stores a time series prediction model to be trained. The time series prediction model is used to predict the time series corresponding to the analysis subsystem. The time series analysis system construction apparatus can be configured on a terminal or a server.
[0089] like Figure 4 As shown, the construction device 200 of the time series analysis system includes a sequence acquisition unit 201, a variable acquisition unit 202, a federated generation unit 203, and a construction completion unit 204.
[0090] The sequence acquisition unit 201 is used to acquire the original time series corresponding to each analysis subsystem;
[0091] The variable acquisition unit 202 is used to input the original time series into the time series prediction model of the corresponding analysis subsystem. The time series prediction model decomposes the original time series and obtains global variable information.
[0092] The federation generation unit 203 is used to obtain global variable information for each analysis subsystem and generate federation variable information based on multiple global variable information.
[0093] Unit 204 is constructed to optimize each time series prediction model based on the information of the federated variables and the global variable information corresponding to each analysis subsystem, complete the training of multiple time series prediction models, and obtain the constructed time series analysis system.
[0094] In some embodiments, the time series prediction model includes a decomposition layer and a variable acquisition unit 202, which is further configured to: input the original time series into the decomposition layer of each analysis subsystem, the decomposition layer decomposes the original time series, and acquires global variable information.
[0095] For example, the decomposition layer decomposes the original time series and also obtains first period information and first trend information; the time series prediction model also includes an encoding layer and a decoding layer; the variable acquisition unit 202 is further configured to: input the original time series into the encoding layer, the encoding layer encodes the original time series to obtain second period information and second trend information; input the first period information into the decoding layer, the decoding layer decodes the first period information to obtain third period information; obtain period difference information between the third period information and the first period information, and trend difference information between the first trend information and the second trend information; optimize the decomposition layer and update the global variable information based on the period difference information and trend difference information, so that the first period information approaches the third period information and the first trend information approaches the second trend information.
[0096] It should be noted that, in some embodiments, the encoding layer includes at least one periodic trend decomposition module; the variable acquisition unit 202 is further configured to: input the original time series into the first periodic trend decomposition module to obtain sub-period information and second trend information; transmit the sub-period information layer by layer to the remaining periodic trend decomposition modules and output the second periodic information.
[0097] It should be noted that, in some embodiments, the time series prediction model further includes an attention layer; before obtaining the period difference information between the third period information and the first period information, the trend difference information between the first trend information and the second trend information, the variable acquisition unit 202 is further configured to: input the second period information and the third period information into the attention layer, the attention layer performs frequency domain correlation learning on the second period information and the third period information to obtain frequency domain correlation information; and update the parameters of the decoding layer according to the frequency domain correlation information to update the third period information.
[0098] In some embodiments, the construction completion unit 204 is further configured to: obtain correlation information between federated variable information and global variable information corresponding to each analysis subsystem based on a self-attention mechanism; update the global variable information corresponding to the analysis subsystem according to the correlation information, so as to optimize the parameters of the time series prediction model corresponding to the analysis subsystem based on the global variable information.
[0099] In some embodiments, the sequence acquisition unit 201 is further configured to: acquire multiple raw time series collected by multiple analysis subsystems; construct a local dataset corresponding to each analysis subsystem based on the raw time series corresponding to each analysis subsystem; and acquire raw time series with the same time axis from the local dataset corresponding to each analysis subsystem.
[0100] It should be noted that those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the above-described apparatus and modules can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.
[0101] The aforementioned device can be implemented as a computer program, which can be used in, for example... Figure 5 It runs on the computer device shown.
[0102] Please see Figure 5 , Figure 5 This is a schematic block diagram illustrating the structure of a computer device according to an embodiment of this application. The computer device may be a server. See also... Figure 5 The computer device includes a processor, memory, and network interface connected via a system bus, wherein the memory may include non-volatile storage media and internal memory.
[0103] Non-volatile storage media can store operating systems and computer programs. These computer programs include program instructions that, when executed, cause the processor to perform any method for constructing a time series analysis system.
[0104] The processor provides computing and control capabilities, supporting the operation of the entire computer device.
[0105] Internal memory provides an environment for the execution of computer programs stored in non-volatile storage media. When executed by a processor, the computer program enables the processor to execute any method for constructing a time series analysis system.
[0106] This network interface is used for network communication, such as sending assigned tasks. Those skilled in the art will understand that... Figure 5 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.
[0107] It should be understood that the processor can be a Central Processing Unit (CPU), but it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among these, a general-purpose processor can be a microprocessor or any conventional processor.
[0108] In one embodiment, the processor is configured to run a computer program stored in memory to perform the following steps:
[0109] Obtain the original time series corresponding to each analysis subsystem;
[0110] The original time series is input into the time series prediction model of the corresponding analysis subsystem. The time series prediction model decomposes the original time series and obtains global variable information.
[0111] Obtain global variable information for each analysis subsystem, and generate federated variable information based on multiple global variable information;
[0112] Each time series prediction model is optimized based on the information of the federal variables and the global variables corresponding to each analysis subsystem. The training of multiple time series prediction models is completed, and the constructed time series analysis system is obtained.
[0113] In some embodiments, the time series forecasting model includes a decomposition layer that inputs the original time series data into the time series forecasting model of the corresponding analysis subsystem, including:
[0114] The original time series is input into the decomposition layer of each analysis subsystem. The decomposition layer decomposes the original time series and obtains global variable information.
[0115] For example, the decomposition layer decomposes the original time series and also obtains first period information and first trend information; the time series prediction model also includes an encoding layer and a decoding layer; the method further includes: inputting the original time series into the encoding layer, the encoding layer encodes the original time series to obtain second period information and second trend information; inputting the first period information into the decoding layer, the decoding layer decodes the first period information to obtain third period information; obtaining period difference information between the third period information and the first period information, and trend difference information between the first trend information and the second trend information; optimizing the decomposition layer and updating global variable information based on the period difference information and trend difference information, so that the first period information approaches the third period information and the first trend information approaches the second trend information.
[0116] It should be noted that, in some embodiments, the encoding layer includes at least one periodic trend decomposition module; the original time series is input into the encoding layer, and the encoding layer encodes the original time series to obtain second periodic information and second trend information, including: inputting the original time series into the first periodic trend decomposition module to obtain sub-period information and second trend information; transmitting the sub-period information layer by layer to the remaining periodic trend decomposition modules, and outputting the second periodic information.
[0117] It should be noted that, in some embodiments, the time series prediction model further includes an attention layer; before obtaining the period difference information between the third period information and the first period information, the trend difference information between the first trend information and the second trend information, the model further includes: inputting the second period information and the third period information into the attention layer, the attention layer performing frequency domain correlation learning on the second period information and the third period information to obtain frequency domain correlation information; and updating the parameters of the decoding layer according to the frequency domain correlation information to update the third period information.
[0118] In some embodiments, optimizing each time series prediction model based on the federated variable information and the global variable information corresponding to each analysis subsystem further includes: obtaining the correlation information between the federated variable information and the global variable information corresponding to each analysis subsystem based on a self-attention mechanism; updating the global variable information corresponding to the analysis subsystem based on the correlation information, so as to optimize the parameters of the time series prediction model corresponding to the analysis subsystem based on the global variable information.
[0119] In some embodiments, obtaining the original time series corresponding to each analysis subsystem includes: obtaining multiple original time series collected by multiple analysis subsystems; constructing a local dataset corresponding to each analysis subsystem based on the original time series corresponding to each analysis subsystem; and obtaining original time series with the same time axis from the local dataset corresponding to each analysis subsystem.
[0120] The embodiments of this application also provide a computer-readable storage medium storing a computer program, the computer program including program instructions, and the processor executing the program instructions to implement any of the time series analysis system construction methods provided in the embodiments of this application.
[0121] The computer-readable storage medium may be an internal storage unit of the computer device described in the foregoing embodiments, such as the hard disk or memory of the computer device. The computer-readable storage medium may also be an external storage device of the computer device, such as a plug-in hard disk, SmartMedia Card (SMC), Secure Digital (SD) card, or Flash Card equipped on the computer device.
[0122] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
Claims
1. A method for constructing a time series analysis system, characterized in that, The time series analysis system to be constructed includes multiple analysis subsystems, each of which stores a time series prediction model to be trained. The time series prediction model is used to predict the time series corresponding to that analysis subsystem. The method includes: Obtain the original time series corresponding to each of the analysis subsystems; The original time series is input into the time series prediction model of the corresponding analysis subsystem. The time series prediction model decomposes the original time series to obtain global variable information. The time series prediction model includes a decomposition layer. The step of inputting the original time series into the time series prediction model of the corresponding analysis subsystem includes: inputting the original time series into the decomposition layer of each analysis subsystem. The decomposition layer decomposes the original time series to obtain the global variable information. Obtain the global variable information of each of the analysis subsystems, and generate federated variable information based on the multiple global variable information; The time series prediction model is optimized based on the federated variable information and the global variable information corresponding to each analysis subsystem, and the training of multiple time series prediction models is completed to obtain a constructed time series analysis system. The decomposition layer decomposes the original time series and obtains first period information and first trend information. The time series prediction model also includes an encoding layer and a decoding layer. The method further includes: inputting the original time series into the encoding layer, the encoding layer encodes the original time series to obtain second period information and second trend information; inputting the first period information into the decoding layer, the decoding layer decodes the first period information to obtain third period information; obtaining period difference information between the third period information and the first period information, and trend difference information between the first trend information and the second trend information; optimizing the decomposition layer and updating the global variable information based on the period difference information and the trend difference information, so that the first period information approaches the third period information and the first trend information approaches the second trend information.
2. The method according to claim 1, characterized in that, The encoding layer includes at least one periodic trend decomposition module; the step of inputting the original time series into the encoding layer, and the encoding layer encoding the original time series to obtain second periodic information and second trend information, includes: The original time series is input into the first cycle trend decomposition module to obtain sub-cycle information and second trend information; The sub-period information is transmitted layer by layer to the remaining periodic trend decomposition modules, and the second period information is output.
3. The method according to claim 1, characterized in that, The time series prediction model further includes an attention layer; before acquiring the period difference information between the third period information and the first period information, the trend difference information between the first trend information and the second trend information, it also includes: The second periodic information and the third periodic information are input into the attention layer, and the attention layer performs frequency domain correlation learning on the second periodic information and the third periodic information to obtain frequency domain correlation information; The parameters of the decoding layer are updated based on the frequency domain correlation information to update the third period information.
4. The method according to claim 1, characterized in that, The optimization of each time series prediction model based on the federated variable information and the global variable information corresponding to each analysis subsystem further includes: The correlation information between the federated variable information and the global variable information corresponding to each of the analysis subsystems is obtained based on the self-attention mechanism; The global variable information corresponding to the analysis subsystem is updated based on the correlation information, so as to optimize the parameters of the time series prediction model corresponding to the analysis subsystem based on the global variable information.
5. The method according to claim 1, characterized in that, The step of obtaining the original time series corresponding to each of the analysis subsystems includes: Acquire multiple raw time series data collected by the multiple analysis subsystems; Construct a local dataset for each analysis subsystem based on the original time series data corresponding to each analysis subsystem; The original time series with the same time axis is obtained from the local dataset corresponding to each of the analysis subsystems.
6. A device for constructing a time series analysis system, characterized in that, The time series analysis system to be constructed includes multiple analysis subsystems, each of which stores a time series prediction model to be trained. The time series prediction model is used to predict the time series corresponding to that analysis subsystem; including: A sequence acquisition unit is used to acquire the original time series corresponding to each of the analysis subsystems; A variable acquisition unit is used to input the original time series into the time series prediction model of the corresponding analysis subsystem. The time series prediction model decomposes the original time series to obtain global variable information. The time series prediction model includes a decomposition layer. Inputting the original time series into the time series prediction model of the corresponding analysis subsystem includes: inputting the original time series into the decomposition layer of each analysis subsystem, whereby the decomposition layer decomposes the original time series to obtain the global variable information. The federation generation unit is used to obtain the global variable information of each of the analysis subsystems and generate federation variable information based on the multiple global variable information. The system is constructed by optimizing each time series prediction model based on the federated variable information and the global variable information corresponding to each analysis subsystem, training multiple time series prediction models, and obtaining a constructed time series analysis system. The decomposition layer decomposes the original time series and obtains first period information and first trend information. The time series prediction model further includes an encoding layer and a decoding layer. The device further includes: inputting the original time series into the encoding layer, where the encoding layer encodes the original time series to obtain second period information and second trend information; inputting the first period information into the decoding layer, where the decoding layer decodes the first period information to obtain third period information; obtaining period difference information between the third period information and the first period information, and trend difference information between the first trend information and the second trend information; optimizing the decomposition layer and updating the global variable information based on the period difference information and the trend difference information, so that the first period information approaches the third period information, and the first trend information approaches the second trend information.
7. A computer device, characterized in that, The computer device includes a memory and a processor; The memory is used to store computer programs; The processor is configured to execute the computer program and, in executing the computer program, implement the method as described in any one of claims 1 to 5.
8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, causes the processor to implement the method as described in any one of claims 1 to 5.