Irregular multivariable time sequence prediction method and device based on patch block pre-alignment
By using patch block pre-alignment to efficiently map and fuse multi-scale features of irregular time series, the problem of unstable performance and high computational cost in irregular multivariate time series prediction is solved, and efficient and accurate prediction results are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHENZHEN UNIV
- Filing Date
- 2025-11-27
- Publication Date
- 2026-05-01
AI Technical Summary
Existing technologies struggle to effectively handle irregular, multivariate time-series data, leading to unstable predictive model performance and excessive computational overhead. They also fail to accurately capture the potential dynamics and cross-variable correlations of irregularly sampled data.
The method of patch block pre-alignment is adopted. The patch block aligner module efficiently maps irregular time series and constructs multi-scale time embedding features. The group mixer module fuses time period features and channel features. Finally, the prediction output module obtains the prediction results of irregular time series.
It significantly improves the accuracy and efficiency of irregular multivariate time series forecasting, reduces computational resource consumption, and ensures the accuracy and parallelism of time series forecasting.
Smart Images
Figure CN121959036A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of irregular multivariate time series prediction technology, and in particular to an irregular multivariate time series prediction method and apparatus based on patch block pre-alignment. Background Technology
[0002] Currently, existing technologies are mostly based on fixed sampling intervals and complete observation data, such as Figure 1 (a) However, in practical applications, data is often missing or irregular due to equipment failure, communication interruptions, or excessively high collection costs; at the same time, user behavior, social media activity, and other data are inherently irregular, resulting in sparse timestamps and difficulty in aligning across variables. Such irregular data, such as... Figure 1 (b) Not only does non-uniform sampling disrupt temporal consistency, but it also contains complex features, posing challenges to predictive modeling.
[0003] Irregular multivariate time series forecasting has significant practical implications. In the medical field, predicting physiological indicators from irregularly sampled records enables early risk warning and resource allocation; in astronomy, predicting irregular observations can be used for trajectory reconstruction and signal estimation; similarly, irregular time series forecasting is a core component in fields such as financial risk control and intelligent transportation, and its performance directly determines the reliability of the system and the effectiveness of decision-making. Therefore, how to conduct efficient and accurate forecasting for complex irregular multivariate data is a crucial problem that urgently needs to be solved.
[0004] However, directly applying regular time series prediction models to irregular time series data often results in poor performance. On one hand, strategies based on filler-style pre-alignment (zero-padding each variable's time series until their lengths match the longest series) distort the true temporal relationships between variables, such as... Figure 2 (c) On the other hand, irregular sampling intervals disrupt temporal consistency, making it difficult for existing standard regular sampling models to capture their potential dynamics and intervariate correlations. Meanwhile, most existing irregular time series forecasting methods rely on a pre-alignment strategy based on a unified axis (first constructing a unified time axis, then aligning each variable to the same set of timestamps, with missing observations typically filled with zeros), such as... Figure 2 (b) This leads to an explosive increase in storage and computational overhead, reducing training efficiency. Summary of the Invention
[0005] Therefore, the purpose of this invention is to at least partially address the shortcomings of the prior art, thereby proposing an irregular multivariate time series prediction method and apparatus based on patch block pre-alignment.
[0006] In a first aspect, the present invention provides an irregular multivariate time series prediction method based on patch block pre-alignment, applied to an irregular multivariate time series prediction model, the method comprising:
[0007] The patch block aligner module efficiently maps irregular time series to obtain aligned regular time series, which include fine-grained patch blocks, medium-grained patch blocks, and coarse-grained patch blocks.
[0008] The fine-grained patch block, the medium-grained patch block, and the coarse-grained patch block are aggregated layer by layer by the multi-scale temporal encoder module to construct multi-scale temporal embedding features;
[0009] The multi-scale temporal embedding features are fused using a group mixer module to obtain a global representation by combining time-segment features and channel features respectively.
[0010] The global representation is processed by the prediction output module to obtain the prediction results for irregular time series.
[0011] Secondly, the present invention provides an irregular multivariate time series prediction device based on patch block pre-alignment, applied to an irregular multivariate time series prediction model, the device comprising:
[0012] Mapping module: Used to efficiently map irregular time series through patch block aligner module to obtain aligned regular time series, the aligned regular time series including fine-grained patch blocks, medium-grained patch blocks and coarse-grained patch blocks;
[0013] Aggregation module: used to aggregate the fine-grained patch block, medium-grained patch block and coarse-grained patch block layer by layer through the multi-scale temporal encoder module to construct multi-scale temporal embedding features;
[0014] Hybridization module: used to fuse the multi-scale temporal embedding features with time period features and channel features respectively through the group mixer module to obtain a global representation;
[0015] Prediction module: Used to process the global representation through the prediction output module to obtain the prediction results of irregular time series.
[0016] Thirdly, the present invention also provides an irregular multivariate time series prediction device based on patch block pre-alignment, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the memory executes the computer program, it implements the various steps of the irregular multivariate time series prediction method based on patch block pre-alignment as described in the first aspect.
[0017] Fourthly, the present invention also provides a storage medium having a computer program stored thereon, which, when executed by a processor, implements the various steps of the irregular multivariate time series prediction method based on patch block pre-alignment as described in the first aspect.
[0018] This invention provides a method and apparatus for predicting irregular multivariate time series based on patch block pre-alignment, applied to irregular multivariate time series prediction models. The method includes: efficiently mapping irregular time series using a patch block aligner module to obtain aligned regular time series, wherein the aligned regular time series includes fine-grained, medium-grained, and coarse-grained patch blocks; a multi-scale temporal encoder module aggregating the fine-grained, medium-grained, and coarse-grained patch blocks layer by layer to construct multi-scale temporal embedding features; a group mixer module fusing time period features and channel features of the multi-scale temporal embedding features to obtain a global representation; and a prediction output module processing the global representation to obtain the prediction result of the irregular time series. The method provided by this invention introduces an efficient patch block pre-alignment mechanism in the data preprocessing stage, while preserving multi-scale temporal similarity during modeling, to address the problem that traditional prediction algorithms struggle to adapt to irregular sampled data. This method effectively overcomes the limitations of existing irregular prediction models, such as unstable performance and excessive computational overhead, aiming to significantly improve prediction accuracy while ensuring high efficiency. Attached Figure Description
[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.
[0020] Figure 1 This is a schematic diagram illustrating the difference between regular multivariate time series prediction tasks and irregular multivariate time series prediction tasks in the prior art of this invention;
[0021] Figure 2 This is a schematic diagram illustrating the differences between the present invention and the prior art;
[0022] Figure 3 This is a flowchart illustrating the irregular multivariate time series prediction method based on patch block pre-alignment of the present invention.
[0023] Figure 4 This is a schematic diagram of the training process of the irregular multivariate time series prediction model based on the patch block pre-alignment method of the present invention.
[0024] Figure 5 This is a schematic diagram of the algorithm framework of the irregular multivariate time series prediction method based on patch block pre-alignment of the present invention.
[0025] Figure 6 This is a schematic diagram of a sub-process of the irregular multivariate time series prediction method based on patch block pre-alignment of the present invention;
[0026] Figure 7 This is a schematic diagram of another sub-process of the irregular multivariate time series prediction method based on patch block pre-alignment of the present invention;
[0027] Figure 8 This is a flowchart of the patch block aligner module of the irregular multivariate time series prediction method based on patch block pre-alignment of the present invention;
[0028] Figure 9 This is a schematic diagram of another sub-process of the irregular multivariate time series prediction method based on patch block pre-alignment of the present invention;
[0029] Figure 10 This is a schematic diagram of another sub-process of the irregular multivariate time series prediction method based on patch block pre-alignment of the present invention;
[0030] Figure 11 This is a schematic diagram of the multi-scale time encoder module of the irregular multivariate time series prediction method based on patch block pre-alignment of the present invention;
[0031] Figure 12 This is a schematic diagram of the learnable temporal embedding mapping module and the time-aware convolutional network in the multi-scale encoder module of the irregular multivariate temporal prediction method based on patch block pre-alignment of the present invention.
[0032] Figure 13 This is a schematic diagram of another sub-process of the irregular multivariate time series prediction method based on patch block pre-alignment of the present invention;
[0033] Figure 14 This is a schematic diagram of the group mixer module of the irregular multivariate time series prediction method based on patch block pre-alignment of the present invention;
[0034] Figure 15 This is a flowchart of the alternating attention mechanism in the group mixer module of the irregular multivariate temporal prediction method based on patch block pre-alignment of the present invention;
[0035] Figure 16 This is a schematic diagram of another sub-process of the irregular multivariate time series prediction method based on patch block pre-alignment of the present invention;
[0036] Figure 17 This is a schematic diagram illustrating the prediction effect of the irregular multivariate time series prediction method based on patch block pre-alignment of the present invention;
[0037] Figure 18 This is a schematic diagram of the program module of the irregular multivariate time series prediction device based on patch block pre-alignment of the present invention. Detailed Implementation
[0038] To make the objectives, features, and advantages of this invention more apparent and understandable, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.
[0039] Please refer to Figure 3 , Figure 4 and Figure 5 , Figure 3 This is a flowchart illustrating the irregular multivariate time series prediction method based on patch block pre-alignment in an embodiment of this application. Figure 4 This is a schematic diagram of the training process of the irregular multivariate time series prediction model in the irregular multivariate time series prediction method based on patch block pre-alignment in the embodiments of this application. Figure 5 This is a schematic diagram of the algorithm framework of the irregular multivariate time series prediction method based on patch block pre-alignment of the present invention. In this embodiment, it is applied to an irregular multivariate time series prediction model. The above-mentioned irregular multivariate time series prediction method based on patch block pre-alignment includes:
[0040] Step 101: The irregular time series is efficiently mapped by the patch block aligner module to obtain the aligned regular time series. The aligned regular time series includes fine-grained patch blocks, medium-grained patch blocks and coarse-grained patch blocks.
[0041] In this embodiment, the patch block aligner module achieves efficient mapping of irregular observation data in irregular time series through an efficient patch block pre-alignment strategy, thereby obtaining aligned regular time series. The aligned regular time series includes fine-grained patch blocks, medium-grained patch blocks, and coarse-grained patch blocks. Unlike previous methods that required "forcibly stretching the time axis," this efficient patch block pre-alignment strategy avoids errors caused by interpolation or time axis stretching, supports batch processing and GPU acceleration, eliminates the need to generate a complete unified time matrix, significantly reduces resource consumption, and can significantly improve computational parallelism and memory utilization while maintaining time series accuracy, thus achieving a precise, parallel, and memory-friendly time alignment process. Specifically, irregular time series refer to time series data with non-fixed observation intervals. The patch pre-aligner uses a "blocking + aggregation" strategy to transform the original irregular observation data into regular patch sequences with fixed time ranges. This generates aligned regular time series containing fine-grained, medium-grained, and coarse-grained patch blocks. Fine-grained patch blocks are small time windows, such as 10 minutes, used to capture high-frequency changes, short-term dependencies, and local patterns; medium-grained patch blocks are medium windows, such as 30 minutes, used to balance local details and global structure; and coarse-grained patch blocks are large time windows, such as 60 minutes, used to capture long-term trends, periodicity, and global dependencies. Each of the fine-grained, medium-grained, and coarse-grained patch blocks is a collection of multiple patch blocks. A single patch block refers to a single local time segment (a "small slice" of the time series), and the fine-grained, medium-grained, and coarse-grained patch blocks are complete collections composed of single patch blocks from multiple time windows.
[0042] Step 102: The fine-grained patch block, the medium-grained patch block and the coarse-grained patch block are aggregated layer by layer through the multi-scale temporal encoder module to construct multi-scale temporal embedding features.
[0043] In this embodiment, a multi-scale encoder module processes fine-grained, medium-grained, and coarse-grained patch blocks, enabling the multi-scale encoder to extract temporal feature representations of the patch blocks at different temporal resolutions. These representations are then aggregated layer by layer from fine-grained to medium-grained to coarse-grained patch blocks to construct multi-scale temporal embedding features. Through temporal feature fusion using the multi-scale encoder module, the irregular multivariate time series prediction model can simultaneously capture local short-term changes and global long-term trends, providing rich and hierarchical temporal semantic information for subsequent time series relationship modeling.
[0044] Step 103: The multi-scale temporal embedding features are fused using the group mixer module to obtain a global representation by combining time period features and channel features.
[0045] In this embodiment, a group mixer module based on the Transformer architecture is used to fuse time-period features and channel features of the multi-scale temporal embedding features generated in step 102, thereby obtaining a global representation. This module includes two features: time-period feature fusion and channel feature fusion. In the time-period feature fusion stage, features within the same scale or uniform variable are fused to enhance internal consistency and representational power. In the channel feature fusion stage, dependencies across variables and time scales are explicitly modeled, thereby capturing multivariate interactions and multi-scale dynamic associations. Through this group mixer module, the irregular multivariate temporal prediction model can ensure local feature aggregation while facilitating cross-level information flow, thus significantly improving the overall temporal modeling capability and prediction accuracy.
[0046] Step 104: Process the global representation through the prediction output module to obtain the irregular time series prediction result.
[0047] In this embodiment, the prediction output module processes the global representation obtained in step 103 to obtain the prediction result of irregular time series.
[0048] This application provides a method for predicting irregular multivariate time series based on patch block pre-alignment, applied to an irregular multivariate time series prediction model. The method includes: efficiently mapping the irregular time series using a patch block aligner module to obtain aligned regular time series, where the aligned regular time series includes fine-grained, medium-grained, and coarse-grained patch blocks; a multi-scale temporal encoder module aggregating the fine-grained, medium-grained, and coarse-grained patch blocks layer by layer to construct multi-scale temporal embedding features; a group mixer module fusing time period features and channel features of the multi-scale temporal embedding features to obtain a global representation; and a prediction output module processing the global representation to obtain the irregular time series prediction result. The method provided by this invention introduces an efficient patch block pre-alignment mechanism in the data preprocessing stage while preserving multi-scale temporal similarity during modeling, addressing the problem that traditional prediction algorithms struggle to adapt to irregular sampled data. This method effectively overcomes the limitations of existing irregular prediction models, such as unstable performance and excessive computational overhead, aiming to significantly improve prediction accuracy while ensuring high efficiency.
[0049] Further, please refer to Figure 6 , Figure 6 This is a schematic diagram of a sub-process of the irregular multivariate time series prediction method based on patch block pre-alignment in this embodiment of the application. In this embodiment, after processing the global representation through the prediction output module to obtain the irregular time series prediction result, the method further includes:
[0050] Step 201: Calculate the supervised loss by performing mean square error calculation on the irregular time series prediction results and the preset output results;
[0051] Step 202: Iteratively optimize the irregular multivariate time series prediction model using the supervised loss.
[0052] In this embodiment, the mean square error of the irregular time series prediction results actually obtained by the irregular multivariate time series prediction model and the preset output results that need to be obtained by the irregular multivariate time series prediction model are calculated to obtain the supervised loss. That is, the mean square error of the actual prediction results and the expected output results are calculated to obtain the actual error. The obtained supervised loss is then input into the irregular multivariate time series prediction model for iterative optimization, so that the irregular time series prediction results actually obtained by the irregular multivariate time series prediction model can be close to the expected prediction output results.
[0053] Further, please refer to Figure 7 and Figure 8 , Figure 7 This is a schematic diagram of another sub-process of the irregular multivariate time series prediction method based on patch block pre-alignment in this application embodiment. Figure 8 This is a flowchart of the patch block aligner module of the irregular multivariate time series prediction method based on patch block pre-alignment in this application embodiment. In this embodiment, the step of efficiently mapping the irregular time series through the patch block aligner module to obtain the aligned regular time series includes:
[0054] Step 301: Predefine the time interval of the multi-scale patch block, and pre-calculate and store the boundary information of the patch block at each scale, wherein the multi-scale patch block includes the fine-grained patch block, the medium-grained patch block and the coarse-grained patch block, which each contain multiple patch blocks.
[0055] Step 302: Generate a patch block allocation mask based on the irregular time series and the time interval of the multi-scale patch block. The shape of the patch block allocation mask is [input time point, interval].
[0056] Step 303: Observe the irregular time series to obtain multiple observation masks of different lengths, fill the observation masks of different lengths to the maximum number of valid values of all data in the irregular time series to obtain the filled mask, and obtain the original valid observation values through the dual filtering of the observation mask and the filled mask;
[0057] Step 304: Expand and broadcast the observation mask, and simultaneously perform one-to-one dot multiplication of the filling mask and the patch block allocation mask at the corresponding positions to obtain a three-dimensional variable-aware mask, wherein the shape of the variable-aware mask is [input time point, interval, variable];
[0058] Step 305: Calculate the maximum number of observations based on the index mask mechanism and generate the final patch block mapping mask. Efficiently map the original valid observations to the final patch block mapping mask through a one-to-one correspondence to obtain the aligned rule time series.
[0059] In this embodiment, due to the lack of a unified timestamp for irregular time series, the data acquisition intervals vary greatly (e.g., the sensor sometimes collects data once per minute, and sometimes once every 10 minutes). Predefined time intervals for multi-scale patches involve artificially setting multiple analysis dimensions with different time granularities. Fine-grained patches correspond to local details, while coarse-grained patches correspond to overall trends, allowing the model to analyze data from different levels. The essence of predefinition is to set a "fixed time window size" and "division strategy" for fine-grained, medium-grained, and coarse-grained patches respectively, ensuring that the time interval for each scale is "regular, continuous, and reusable," rather than dynamically generated based on the irregular timestamps of the original data. For each set time scale, the "start and end markers" of each interval are determined and saved. During subsequent data processing, there is no need to temporarily calculate the interval range; the stored boundary information can be directly called to quickly classify scattered time series data into corresponding intervals, providing a unified reference for patch division and alignment, and significantly improving processing efficiency. For example, between 8:00 and 9:00, the time window for a fine-grained patch block is 10 minutes, and the interval for a fine-grained patch block is [08:00-08:10, 08:10-08:20, ..., 08:50-09:00]. The number of patch blocks in a fine-grained patch block is 6. Taking the interval 8:10-8:20 as an example, the start and end time boundary information to be stored is 8:10 and 8:20. The time window for a medium-grained patch block is 30 minutes, and the interval for a medium-grained patch block is [08:00-08:30, 08:30-09:00]. The number of patch blocks in a medium-grained patch block is 2. Taking the interval 8:30-9:00 as an example, the start and end time boundary information to be stored is 8:30 and 9:00. The time window for a coarse-grained patch block is 60 minutes, and the range of the coarse-grained patch block is [8.00-9.00]. The number of patches in a coarse-grained patch block is 1. Taking the range of 8.00-9.00 in the coarse-grained patch block as an example, the boundary information of the start time and the end time that needs to be stored is 8.00 and 9.00.
[0060] First, initialize the tensor patch allocation mask with all zeros, its shape being [input time point, interval]. The core definition of the shape is: the "number of dimensions" and "number of elements in each dimension" of the data (usually existing in the form of a tensor)—it describes how the data is organized in memory. All shape descriptions are essentially combinations of the number of elements in each dimension of the tensor; that is, the shape is the data structure specification, explaining how the data is arranged. The shape [input time point, interval] of the tensor patch allocation mask is a two-dimensional binary matrix, its function being to establish a mapping relationship: for each time point in the original time series, it should be assigned to which one or more predefined time intervals (i.e., patch blocks).
[0061] For each input time point, determine whether it falls within the predefined time interval in step 301. If the timestamp belongs to the interval, mark it as valid (1) at the corresponding position in the mask; otherwise, mark it as invalid (0). The resulting patch block allocation mask can intuitively reflect the patch block information to which each input time point belongs. For example, original timestamp T: [00:05, 00:15, 00:25, 01:10] (M = 4), I_0: [00:00, 00:10) (fine-grained patch block); I_1: [00:10, 00:20) (fine-grained patch block); I_2: [00:20, 00:30) (fine-grained patch block); I_3: [00:00, 01:00) (medium-grained patch block). Patch block); I_4:[01:00,02:00)(medium-grained patch block), t_0(00:05) belongs to both fine-grained patch block I_0 and medium-grained patch block I_3, so the positions of I_0 and I_3 are marked as 1, and the rest are marked as 0. t_3(01:10) belongs only to medium-grained patch block I_4, so the position of I_4 is marked as 1, and the rest are marked as 0.
[0062] The observation mask originates from the irregular time series input data itself and clearly indicates the missing value of each variable at each time point in the original time series. Missing values are 0, and valid values are 1. The shape is [input time point, variable]. Here, the "variable" refers to the "different types of observation indicators" collected in the original irregular time series—such as "vibration value," "temperature value," and "pressure value" in equipment monitoring scenarios, or "humidity," "PM2.5 concentration," and "wind speed" in environmental monitoring. That is, the variable = the "specific indicator type" to be monitored / analyzed, with each variable corresponding to an independent type of observation data. The variable represents the variable dimension of the data (e.g., for temperature and humidity, which have two dimensions, the variable is 2). Simultaneously, observation masks of different lengths are zero-padded to the maximum number of valid values in all data, thus generating a padding mask. This padding mask distinguishes between the original valid data and the invalid positions generated by padding; valid positions are 1, and padding positions are 0. After double-filtering the initially input irregular time series using the observation mask and padding mask, the "true, valid, non-missing, and non-padded" original valid observation values are retained.
[0063] Building upon this, the observation mask is expanded and broadcast along the variable dimension. Simultaneously, the patch block assignment mask and filler mask generated in step 302 are multiplied one-to-one at corresponding positions to form a three-dimensional variable-aware mask with the shape [input time point, interval, variable]. This mask simultaneously indicates the validity of each time point and each variable within each patch block interval, while excluding the influence of missing values and filler data on model processing.
[0064] The maximum number of observations is calculated based on an index mask mechanism using variables, and a final patch block mapping mask is generated. The original valid observations are then efficiently mapped to the final patch block mapping mask through a one-to-one correspondence, resulting in the aligned regular time series. The index mask mechanism precisely locates the time index corresponding to each observation at the point-to-point level, achieving efficient mapping of irregular time sampling points to a fixed patch structure.
[0065] Further, please refer to Figure 9 , Figure 9 This is another sub-process diagram of the irregular multivariate time series prediction method based on patch block pre-alignment in this application embodiment. In this embodiment, the step of calculating the maximum number of observations based on the index mask mechanism and generating the final patch block mapping mask, and efficiently mapping the original valid observations to the final patch block mapping mask through a one-to-one correspondence, to obtain the aligned regular time series, includes:
[0066] Step 401: Sum and summarize the variable perception mask along the dimension of the input time point to obtain the original valid observation value and the number of valid observation values observed by each variable in each patch block. The shape of the variable perception mask after summing and summarizing is [interval, variable].
[0067] Step 402: Traverse all the patch blocks and the statistical results of the number of valid observations of all the variables, find the maximum number of observations and use it as the upper limit of the fixed-length patch block representation;
[0068] Step 403: Create and initialize a final patch block mapping mask with all zeros. For each patch block, the number of valid observations actually observed is marked as valid in the final patch block mapping mask, and the remaining positions that have not reached the maximum number of observations are marked as invalid. The shape of the final patch block mapping mask is [interval, variable, maximum number of observations].
[0069] Step 404: Create and initialize a final new aligned time series data set with all zeros. According to the valid positions marked in the final patch block mapping mask, fill the original valid observations into the final new aligned time series data set in sequence. For invalid positions, place them as placeholders to obtain the aligned rule time series. The shape of the final new aligned time series data set is [interval, variable, maximum number of observations].
[0070] In this embodiment, the variable-aware mask generated in step 304, with the shape [input time point, interval, variable], is summed and statistically analyzed along the input time point dimension to obtain the number of valid observations actually observed for each variable in each patch block, with the shape [interval, variable]. Subsequently, the statistical results of the number of valid observations for all patch blocks and all variables are traversed to find the maximum number of observations, which is then used as the upper limit for subsequent fixed-length patch blocks.
[0071] Create and initialize a final patch block mapping mask with all zeros, in the shape [interval, variable, maximum number of observations], with the total number of valid observations matching the number of valid observations in the original data. For each patch block, mark the corresponding position in the final patch block mapping mask as valid (1); for the remaining positions that have not reached the maximum number of observations, keep the invalid mark (0). Each valid mark in the mask corresponds to the position of a valid observation in the original time series. For example, suppose a variable in the original time series has three valid observations at irregular time points. The maximum number of observations for this variable in the corresponding patch block calculated in step 402 is five. Set the first, second, and third valid positions to 1 (valid), while keeping the remaining two invalid positions in the mask as 0 (invalid).
[0072] Create and initialize a new, fully zero-based aligned time series dataset (which can be used for subsequent feature aggregation, model training, or other time series analysis operations). Its shape is the same as the final patch block mapping mask: [interval, variable, maximum number of observations]. Fill the newly generated aligned time series dataset with the original input valid observations in sequence. For positions marked as invalid in the mask, leave them empty or fill them with placeholders to ensure that invalid positions do not affect subsequent processing. The final sequence is a regularized representation in both the patch block and variable dimensions, i.e., an aligned, regular time series. For example, suppose a variable in the original time series has three valid observations at irregular time points, namely A, B, and C. The final patch block mapping mask indicates that the maximum number of observations for this variable in the corresponding patch block is five. Fill A into the first valid position, B into the second, and C into the third, while keeping the remaining two invalid positions in the mask as 0.
[0073] Further, please refer to Figure 10 , 11 and Figure 12 , Figure 10 This is a schematic diagram of another sub-process of the irregular multivariate time series prediction method based on patch block pre-alignment in the embodiments of this application. Figure 11 This is a schematic diagram of the multi-scale temporal encoder module of the irregular multivariate temporal prediction method based on patch block pre-alignment in an embodiment of this application. Figure 12 This diagram illustrates the learnable temporal embedding mapping module and the time-aware convolutional network of the irregular multivariate temporal prediction method based on patch block pre-alignment in this application embodiment. In this embodiment, the multi-scale temporal embedding features are constructed by layer-by-layer aggregation of the fine-grained patch blocks, the medium-grained patch blocks, and the coarse-grained patch blocks through the multi-scale temporal encoder module, including:
[0074] Step 501: Perform learnable temporal embedding mapping on the timestamps of each patch block within the fine-grained patch block, the medium-grained patch block, and the coarse-grained patch block to generate temporal embeddings. Each of the fine-grained patch block, the medium-grained patch block, and the coarse-grained patch block includes at least the original valid observations, timestamps, and mask information. The shapes of the original valid observations, timestamps, and mask information are all [interval, variable, maximum number of observations];
[0075] Step 502: Concatenate the time embedding with the original valid observations to obtain enhanced features containing time semantics in each patch block;
[0076] Step 503: Extract the local temporal dependencies and semantic features within each of the fine-grained patch blocks, medium-grained patch blocks, and coarse-grained patch blocks using a time-aware convolutional network to obtain a fixed-length representation of each patch block.
[0077] Step 504: Concatenate the enhancement features and fixed-length representations of all the patch blocks of the fine-grained patch block, the medium-grained patch block, and the coarse-grained patch block in the order of fine-grained patch block, medium-grained patch block, and coarse-grained patch block to obtain the multi-scale temporal embedding feature. The shape of the multi-scale temporal embedding feature is [number of patch blocks, variable, embedding dimension].
[0078] In this embodiment, fine-grained, medium-grained, and coarse-grained patch blocks all include at least the observation value X, the timestamp T, and the mask information M. The shape of each type of information is [interval, variable, maximum number of observations], specifically: variable C * fine / medium / coarse-grained patch block length - P1 / P2 / P3, with each scale containing N1 / N2 / N3 patch blocks. "Variable C" represents the variable dimension of the data (e.g., temperature and humidity have 2 variables, so C = 2), "fine / medium / coarse-grained patch block length P1 / P2 / P3" represents the maximum number of observations in all patch blocks at this granularity, and N1 / N2 / N3 patch blocks correspond to the number of independent data units contained in X at each scale, i.e., the number of patch blocks, or the interval.
[0079] The observed value X is the core data value of the irregular time series (such as the vibration value, temperature value, humidity value, etc. of the equipment), which is the original material for generating patch features. For example, the "vibration value = 3.2g, temperature = 45℃" corresponding to the original timestamp 08:02 is the observed value of that timestamp. Therefore, the overall shape of X under fine-grained patch can be understood as C×P1×N1, C×P2×N2 at the mesoscale, and C×P3×N3 at the coarse scale. For example, suppose we are monitoring two variables, vibration and temperature, of a device (C=2). The fine-grained patch block P1=5 (the maximum number of observations within all fine-grained patches under this level is 5), and N1=6 (six 10-minute patches); the medium-grained patch block P2=8 (the maximum number of observations within all medium-grained patches under this level is 8), and N2=2 (two 30-minute patches); the coarse-grained patch block P3=10 (the maximum number of observations within all coarse-grained patches under this level is 10), and N3=1 (one 1-hour patch). Then, under the fine-grained patch block, X is a 2×5×6 feature matrix, where each 2×6 submatrix corresponds to the integrated vibration or temperature data of one 10-minute patch block, containing five maximum observation points; under the mesoscale, X is 2×8×2, where each 2×2 submatrix corresponds to the features of one 30-minute patch block, containing integrated vibration or temperature data of eight maximum observation points, and so on.
[0080] The timestamp T is used to mark the time position of each patch block at each scale. It is not the scattered timestamps of the original data, but a standardized time identifier corresponding to each patch block after the multi-scale time interval is defined, which is used to reflect the time sequence and interval belonging of the patch blocks. For example, the timestamp T1 corresponding to the 6 patch blocks (N1=6) in the fine-grained patch block can be marked as [08:00-08:10, 08:10-08:20,..., 08:50-09:00]; the timestamp T2 of the 2 patch blocks (N2=2) in the medium-grained patch block is [08:00-08:30, 08:30-09:00]; and the timestamp T3 of the 1 patch block (N3=1) in the coarse-grained patch block is [08:00-09:00]. These timestamps not only clarify the time range of each patch block, but also allow the encoder to quickly associate the inclusion relationship between the "08:00-08:10" fine-scale patch block and the "08:00-08:30" medium-scale patch block.
[0081] The mask information M is a binary identifier used to mark the validity of data in the patch block (usually 1 for valid and 0 for invalid). Because irregular time series may have some missing or abnormal patch block data, the role of M is to tell the encoder which data can be used for calculation, so as to avoid invalid data from interfering with the encoding results. For example, in a fine-grained patch block (one of N1=6), if the maximum number of observations within the patch block is 5 and the number of valid observations is 3, then the number of valid observations in this patch block is 3, which does not reach the maximum number of observations of 5 for the fine-grained patch block. In this case, M is marked as 0 at positions 4 and 5, and marked as 1 at positions 1, 2, and 3. This "1×5" mask vector corresponds to this patch block. If in a medium-grained patch block (one of N2=2), the maximum number of observations and the number of valid observations within the patch block are both 5, then the number of valid observations reaches the maximum number of observations for the medium-grained patch block. In this case, the corresponding P2=8 granularity patch block is 1×P2×N2, with all values being 1, and the medium-grained patch block is 1×P3×N3, which corresponds one-to-one with the structure of the observation X, allowing the encoder to identify the valid data region.
[0082] Learnable temporal embeddings are mapped to the timestamps of each patch block at each scale to generate continuous temporal representations, i.e., temporal embeddings, such as... Figure 12 The temporal embedding Dt consists of linear and periodic components, used to simultaneously capture aperiodic trends and periodic patterns. The generated temporal embedding is concatenated with the original valid observations X along the channel dimension, resulting in enhanced features C*(Dt+1) for each patch block, containing temporal semantics. Because the temporal embedding Dt is a "continuous vector representation," it can be concatenated with the original valid observations X along the "channel dimension"—which is essentially a "feature dimension." This concatenation effectively adds a "temporal semantic dimension" to the observed features of each patch block, making the features more complete.
[0083] By using a time-aware convolutional network, local temporal dependencies and semantic features within each patch block at each scale are extracted to obtain a fixed-length representation (C*Dout) for each patch block, i.e., the embedding dimension, such as... Figure 12 The core of the local temporal dependency within each patch block refers to the "sequential correlation" between observations arranged in the original timestamp order within the time window of a single patch block—for example, how an observation at a previous moment affects an observation at a later moment, or the short-term trends, abrupt changes, periodic fluctuations, and other time-series patterns formed by adjacent observations.
[0084] All patch features at different scales, with shapes [N1, variable C, embedding dimension Dout], [N2, C, Dout], and [N3, C, Dout], are concatenated sequentially along the patch dimensions N1 / N2 / N3 to obtain a global multi-scale feature representation with the shape [N = N1 + N2 + N3, C, Dout]. Here, N1, N2, and N3 represent the number of patches at different scales, and N is the total number of patches across all scales. This operation can simultaneously fuse fine-grained, medium-grained, and coarse-grained temporal information. Specifically, the shape of all patches without the multi-scale temporal encoder module is [interval, variable, maximum number of observations], while the shape of patches processed by the multi-scale temporal encoder module is [number of patches, variable, embedding dimension]. In this context, the "interval" of the patch block before processing is the same as the "number of patch blocks" of the patch block after processing. Essentially, it is the total number of predefined local time segments at a certain scale. The "maximum number of observations" is converted into the "embedding dimension" through a learnable temporal embedding map and a time-aware convolutional network.
[0085] Further, please refer to Figure 13 , 14 and Figure 15 , Figure 13 This is a schematic diagram of another sub-process of the irregular multivariate time series prediction method based on patch block pre-alignment in the embodiments of this application. Figure 14 This is a schematic diagram of the group mixer module of the irregular multivariate time series prediction method based on patch block pre-alignment in an embodiment of this application. Figure 15 This is a flowchart of the alternating attention mechanism in the group mixer module of the irregular multivariate temporal prediction method based on patch block pre-alignment in this application embodiment. In this embodiment, the group mixer module fuses the multi-scale temporal embedding features with time period features and channel features respectively to obtain a global representation, including:
[0086] Step 601: Divide the multi-scale temporal embedding features into several groups according to the channel dimension. Each group includes multiple variables and fine-grained patch blocks, medium-grained patch blocks, and coarse-grained patch blocks.
[0087] Step 602: An alternating stacking structure is used to enable odd-numbered layers to perform time-period attention and even-numbered layers to perform channel attention. After each layer is completed, it needs to be normalized, residual superimposed, and feedforward network fusion processed. Among them, several groups include the odd-numbered layers and the even-numbered layers.
[0088] Step 603: After repeatedly stacking the group mixer module, the global representation is obtained by fusion. The shape of the global representation is [number of patch blocks, variables, embedding dimension].
[0089] In this embodiment, the multi-scale embedding dimension is divided into several groups according to the channel dimension, and each group contains several related variables. Grouping helps reduce the computational complexity of attention and provides a structured foundation for subsequent "time period modeling" and "channel modeling".
[0090] The group mixer module employs an alternating stacked structure: odd-numbered layers perform "time-period attention," while even-numbered layers perform "channel attention." This alternating design separates time dependencies from variable dependencies, thus more effectively modeling multidimensional relationships in irregular time series. Specifically, time-period attention primarily models dependencies along the time dimension. It first captures local short-term temporal features within each group, and then aggregates global long-term temporal information between groups, enabling the model to simultaneously focus on short-term fluctuations and long-term trends. Channel attention primarily focuses on correlations between variables. It first captures tightly coupled channel dependencies within a group (e.g., coordinated changes between multiple physiological signals), and then models global intervariate relationships through inter-group attention, preventing a few strong signal variables from masking information from other weak signal channels.
[0091] After each odd or even layer completes its output, it must undergo normalization, residual stacking, and feedforward network fusion to enhance feature stability and representational ability, and avoid the gradient vanishing problem caused by deep structures.
[0092] After the group mixer modules are repeatedly stacked alternately, the irregular multivariate time series prediction model can fuse feature information from different time scales and variable groups. That is, the model iterates and optimizes repeatedly between the "time dimension" and the "variable dimension" to obtain a global representation that includes time dependence, variable interaction and multi-scale trends. Since the "fusion essence of the group mixer is feature interaction rather than changing the dimensional scale", the core role of the group mixer is to "achieve deep fusion of features without changing the original dimensional structure" - it will not increase / decrease the number of patch blocks (N) or the number of variables (C), nor will it change the embedding dimension (Dout). It will only make the feature content more expressive through "time period group interaction" and "channel group interaction". Therefore, the shape of the global representation is still (number of patch blocks, variables, embedding dimension).
[0093] Further, please refer to Figure 16 and Figure 17 , Figure 16 This is a schematic diagram of another sub-process of the irregular multivariate time series prediction method based on patch block pre-alignment in the embodiments of this application. Figure 17 This is a schematic diagram illustrating the prediction effect of the irregular multivariate time series prediction method based on patch block pre-alignment in this application embodiment. In this embodiment, the step of processing the global representation through the prediction output module to obtain the irregular time series prediction result includes:
[0094] Step 701: Rearrange the dimensions of the global representation so that the shape of the global representation becomes [variable, embedding dimension, number of patch blocks];
[0095] Step 702: Perform aggregation through the aggregation module to obtain the final aggregated features [variables, embedding dimensions];
[0096] Step 703: Restore the final aggregated features to the standard output format [variable, prediction time point] to obtain the irregular time series prediction result.
[0097] In this embodiment, the shape of the global representation is rearranged in dimensions to restore the shape to [variable C, embedding dimension Dout, number of patch blocks N]. The purpose of the dimensional rearrangement is to "restore the sequence structure of the patch blocks" so that the features can be adapted to subsequent multi-scale aggregation.
[0098] Aggregation is performed based on the aggregation module (an aggregation module instantiated at startup after reading the manually preset aggregation methods in the configuration file, including both convolutional and linear aggregation). Different temporal aggregation methods are selected. If a convolutional method is used, local receptive fields are used to compress the patch dimensions, extract patterns, and smooth the data to capture local temporal structure and short-term dynamics. If a linear method is used, weighted fusion of global temporal information is achieved. The final aggregated features are obtained [variable C, embedding dimension Dout].
[0099] The aggregated features are restored to a standard output format through the output layer, with the shape being [variable C, prediction time point], generating the final irregular time series prediction result.
[0100] In summary, the irregular multivariate time series prediction model comprises a patch block aligner module, a multi-scale encoder module, a group mixer module, and a prediction output module. These modules work collaboratively to achieve efficient alignment and accurate prediction of irregular sequences without requiring a unified timestamp. Specifically, the patch block aligner module is first introduced. Through an efficient patch block pre-alignment strategy, it achieves efficient mapping of irregular observation data. This module, based on an index mask mechanism, accurately determines the temporal position of observations at the point-to-point level, avoiding interpolation errors and redundant calculations inherent in traditional unified time axis methods. This results in a precise, parallel, and memory-friendly alignment process. In time series tasks, irregular multivariate time series prediction aims to learn and generate future sequences that conform to real dynamic patterns from multivariate observation data with uneven time intervals. A method for predicting irregular multivariate time series based on an efficient patch-block pre-alignment strategy introduces an efficient patch-block pre-alignment mechanism during data preprocessing, while preserving multi-scale temporal similarity during modeling. This addresses the problem that traditional prediction algorithms struggle to adapt to irregularly sampled data. The method effectively overcomes the limitations of existing irregular prediction models, such as unstable performance and excessive computational cost, aiming to significantly improve prediction accuracy while ensuring high efficiency. Figure 2 (d) The patch-based alignment provided in this application can achieve efficient alignment without time unification, while maintaining the temporal order of variables and capturing the similarity between multi-scale patches. Both the patch-based and unified axis pre-alignment strategies generally ignore the fact that "multi-scale similarity is possible, which is very helpful for understanding the relationship across variables and time levels, and through patch similarity," that is, the local patterns between variables at different time granularities. Figure 2 As can be seen, the model in this embodiment outperforms several similar methods in irregular multivariate time series prediction tasks, achieving a significant performance improvement with only a small increase in computational cost. Subsequently, the multi-scale temporal encoder module extracts feature representations of patches at different temporal resolutions, constructing multi-scale temporal embedding features layer by layer from fine-grained to coarse-grained, providing rich temporal semantic information for subsequent time series modeling. Finally, the group mixer module, based on the Transformer architecture's alternating attention mechanism (time-segment feature fusion and channel feature fusion), explicitly models the interaction relationships between variables and time scales. This module achieves feature fusion within groups while completing information interaction across groups, enabling the model to effectively capture cross-variable dependencies and multi-scale dynamic associations, thereby significantly enhancing the overall time series modeling capability. Through the organic combination of the above modules, features of different variables and different time granularities can fully interact and collaboratively predict. This method significantly improves the model's prediction accuracy and generalization performance while ensuring computational efficiency and resource utilization.
[0101] Furthermore, this application also provides an irregular multivariate time series prediction device 800 based on patch block pre-alignment, applied to irregular multivariate time series prediction models. Figure 18 This is a schematic diagram of the program modules of the irregular multivariate time series prediction device based on patch block pre-alignment in this application embodiment. In this embodiment, the aforementioned irregular multivariate time series prediction device 800 based on patch block pre-alignment includes:
[0102] Mapping module 801: used to efficiently map irregular time series through patch block aligner module to obtain aligned regular time series, wherein the aligned regular time series includes fine-grained patch blocks, medium-grained patch blocks and coarse-grained patch blocks;
[0103] Aggregation module 802: used to aggregate the fine-grained patch block, the medium-grained patch block and the coarse-grained patch block layer by layer through the multi-scale temporal encoder module to construct multi-scale temporal embedding features;
[0104] Hybrid module 803: Used to fuse the multi-scale temporal embedding features with time period features and channel features respectively through the group mixer module to obtain a global representation;
[0105] Prediction module 804: Used to process the global representation through the prediction output module to obtain the prediction result of the irregular time series.
[0106] This application provides an irregular multivariate time series prediction device 800 based on patch block pre-alignment, applied to an irregular multivariate time series prediction model. It can achieve the following: an efficient mapping of irregular time series using a patch block aligner module to obtain aligned regular time series, where the aligned regular time series includes fine-grained, medium-grained, and coarse-grained patch blocks; layer-by-layer aggregation of the fine-grained, medium-grained, and coarse-grained patch blocks using a multi-scale time encoder module to construct multi-scale time embedding features; fusion of time period features and channel features of the multi-scale time embedding features using a group mixer module to obtain a global representation; and processing of the global representation using a prediction output module to obtain the irregular time series prediction result. The method provided by this invention introduces an efficient patch block pre-alignment mechanism in the data preprocessing stage while preserving multi-scale time similarity during modeling, addressing the problem that traditional prediction algorithms struggle to adapt to irregular sampled data. This method effectively overcomes the limitations of existing irregular prediction models, such as unstable performance and excessive computational overhead, aiming to significantly improve prediction accuracy while ensuring high efficiency.
[0107] Furthermore, this application also provides an irregular multivariate time series prediction device based on patch block pre-alignment, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the memory executes the computer program, it implements the various steps in the irregular multivariate time series prediction method based on patch block pre-alignment as described above.
[0108] Furthermore, this application also provides a storage medium storing a computer program thereon, which, when executed by a processor, implements the various steps in the irregular multivariate time series prediction method based on patch block pre-alignment as described above.
[0109] In the various embodiments of this invention, the functional modules can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can be stored in a computer-readable storage medium.
[0110] Based on this understanding, the technical solutions of this invention, in essence, or the parts that contribute to the prior art, or all or part of the technical solutions, 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 invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0111] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the present invention is not limited to the described order of actions, as some steps can be performed in other orders or simultaneously according to the present invention. Secondly, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to the present invention. In the above embodiments, the descriptions of each embodiment have their own emphasis; for parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.
[0112] For those skilled in the art, based on the ideas of the embodiments of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.
Claims
1. A method for predicting irregular multivariate time series based on patch block pre-alignment, characterized in that, The method, applied to irregular multivariate time series prediction models, includes: The patch block aligner module efficiently maps irregular time series to obtain aligned regular time series, which include fine-grained patch blocks, medium-grained patch blocks, and coarse-grained patch blocks. The fine-grained patch block, the medium-grained patch block, and the coarse-grained patch block are aggregated layer by layer by the multi-scale temporal encoder module to construct multi-scale temporal embedding features; The multi-scale temporal embedding features are fused using a group mixer module to obtain a global representation by combining time-segment features and channel features respectively. The global representation is processed by the prediction output module to obtain the prediction results for irregular time series.
2. The method according to claim 1, characterized in that, The step of processing the global representation through the prediction output module to obtain the irregular time series prediction result further includes: The supervised loss is obtained by calculating the mean square error between the irregular time series prediction results and the preset output results. The irregular multivariate time series prediction model is iteratively optimized using the supervised loss.
3. The method according to claim 1, characterized in that, The process of efficiently mapping irregular time series using the patch block aligner module to obtain aligned regular time series includes: The time interval of the multi-scale patch block is predefined, and the boundary information of the patch block at each scale is pre-calculated and stored. The multi-scale patch block includes the fine-grained patch block, the medium-grained patch block and the coarse-grained patch block, which each contain multiple patch blocks. A patch block allocation mask is generated based on the irregular time series and the time interval of the multi-scale patch block, and the shape of the patch block allocation mask is [input time point, interval]; Multiple observation masks of different lengths are obtained by observing the irregular time series. The observation masks of different lengths are filled with the maximum number of valid values of all data in the irregular time series to obtain the filled mask. The original valid observation values are obtained by double filtering through the observation mask and the filled mask. The observation mask is expanded and broadcast, and the padding mask and the patch block allocation mask are multiplied one-to-one at the corresponding positions to obtain a three-dimensional variable-aware mask, wherein the shape of the variable-aware mask is [input time point, interval, variable]; The maximum number of observations is calculated based on the index mask mechanism, and the final patch block mapping mask is generated. The original valid observations are efficiently mapped to the final patch block mapping mask through a one-to-one correspondence, resulting in the aligned rule time series.
4. The method according to claim 3, characterized in that, The process involves calculating the maximum number of observations based on an index mask mechanism and generating a final patch block mapping mask. The original valid observations are then efficiently mapped to the final patch block mapping mask through a one-to-one correspondence, resulting in the aligned rule-based time series. This includes: The variable perception mask is summed and statistically summarized along the dimension of the input time point to obtain the number of valid observations of each variable in each patch block. The shape of the variable perception mask after summing and summing is [interval, variable]. Iterate through all the patch blocks and the statistical results of the number of valid observations for all the variables, find the maximum number of observations and use it as the upper limit for the fixed-length patch block representation; Create and initialize a final patch block mapping mask that is all zeros. For each patch block, the number of valid observations actually observed is marked as valid in the final patch block mapping mask, and the remaining positions that have not reached the maximum number of observations are marked as invalid. The shape of the final patch block mapping mask is [interval, variable, maximum number of observations]. Create and initialize a final new aligned time series data set with all zeros. According to the valid positions marked in the final patch block mapping mask, fill the original valid observations into the final new aligned time series data set one by one in sequence. For invalid positions, place them as placeholders to obtain the aligned regular time series. The shape of the final new aligned time series data set is [interval, variable, maximum number of observations].
5. The method according to claim 4, characterized in that, The process involves layer-by-layer aggregation of the fine-grained patch blocks, medium-grained patch blocks, and coarse-grained patch blocks using a multi-scale temporal encoder module to construct multi-scale temporal embedding features, including: Learnable temporal embedding mapping is performed on the timestamps of each patch block within the fine-grained patch block, the medium-grained patch block, and the coarse-grained patch block to generate temporal embeddings. Each of the fine-grained patch block, the medium-grained patch block, and the coarse-grained patch block includes at least the original valid observations, timestamps, and mask information. The shapes of the original valid observations, timestamps, and mask information are all [interval, variable, maximum number of observations]; The time embedding is concatenated with the original valid observations to obtain enhanced features containing temporal semantics in each patch block; The local temporal dependencies and semantic features within each of the fine-grained, medium-grained, and coarse-grained patch blocks are extracted using a time-aware convolutional network to obtain a fixed-length representation of each patch block. The enhancement features and fixed-length representations of all the patch blocks of the fine-grained patch block, the medium-grained patch block, and the coarse-grained patch block are concatenated in the order of the fine-grained patch block, the medium-grained patch block, and the coarse-grained patch block to obtain the multi-scale temporal embedding feature. The shape of the multi-scale temporal embedding feature is [number of patch blocks, variable, embedding dimension].
6. The method according to claim 5, characterized in that, The multi-scale temporal embedding features are fused using a group mixer module to obtain a global representation, including: The multi-scale temporal embedding features are divided into several groups according to the channel dimension. Each group includes multiple variables and fine-grained patch blocks, medium-grained patch blocks, and coarse-grained patch blocks. An alternating stacking structure is used to enable odd-numbered layers to perform time-period attention and even-numbered layers to perform channel attention. After each layer is completed, it needs to undergo normalization, residual superposition, and feedforward network fusion processing. Among them, several groups include the odd-numbered layers and the even-numbered layers. After repeated alternating stacking of the group mixer module, the global representation is obtained by fusion, and the shape of the global representation is [number of patch blocks, variables, embedding dimension].
7. The method according to claim 6, characterized in that, The step of processing the global representation through the prediction output module to obtain the irregular time series prediction result includes: The global representation is rearranged in dimensions so that its shape becomes [variable, embedding dimension, number of patch blocks]; The aggregation module performs aggregation to obtain the final aggregated features [variables, embedding dimensions]. The final aggregated features are restored to the standard output format [variable, prediction time point] to obtain the prediction result of the irregular time series.
8. An irregular multivariate time series prediction device based on patch block pre-alignment, characterized in that, The device, applied to irregular multivariate time series prediction models, includes: Mapping module: Used to efficiently map irregular time series through patch block aligner module to obtain aligned regular time series, the aligned regular time series including fine-grained patch blocks, medium-grained patch blocks and coarse-grained patch blocks; Aggregation module: used to aggregate the fine-grained patch block, medium-grained patch block and coarse-grained patch block layer by layer through the multi-scale temporal encoder module to construct multi-scale temporal embedding features; Hybridization module: used to fuse the multi-scale temporal embedding features with time period features and channel features respectively through the group mixer module to obtain a global representation; Prediction module: Used to process the global representation through the prediction output module to obtain the prediction results of irregular time series.
9. A device for predicting irregular multivariate time series based on patch block pre-alignment, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the memory executes the computer program, it implements each step of the irregular multivariate time series prediction method based on patch block pre-alignment as described in any one of claims 1-7.
10. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements each step of the irregular multivariate time series prediction method based on patch block pre-alignment as described in any one of claims 1-7.