Timing sequence mining method based on multilevel features
By using a multi-level feature mining method to slice, extract features, and quantify similarity in time series data, the problem of systematically summarizing massive heterogeneous time series data is solved, forming a typical pattern knowledge base, which improves the robustness and cross-scenario adaptability of the prediction algorithm.
Patent Information
- Application Number
- CN202511279341.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-09
- Publication Date
- 2025-12-30
AI Technical Summary
Existing technologies cannot effectively and systematically summarize massive heterogeneous time-series data, lack the ability to form a typical pattern knowledge base, and are difficult to identify frequently occurring state patterns and regular evolution trajectories.
A time series mining method based on multi-level features is adopted. By slicing the time series data, parametric statistical distribution features, differential statistical distribution features and expert-level features are extracted, weighted summation is performed, the similarity between segments is quantified, redundant segments are removed, and typical patterns are formed.
It significantly improves the ability to jointly represent long-term evolutionary patterns and short-term mutation characteristics, builds a standardized knowledge base, provides interpretable time-series pattern prior knowledge for subsequent business development, and enhances the robustness of multi-step prediction and cross-scenario transferability.
Smart Images
Figure CN121233641A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of time-series signal processing technology, and specifically to a time-series sequence mining method based on multi-level features. Background Technology
[0002] Time series sequences contain rich dynamic information, and their patterns of change are widely present in the physical world and human activities. Time series data, such as stock price fluctuations and electrocardiogram (ECG) signal rhythms, carry key short-term characteristics and long-term patterns. Time series signal processing, as a core technology for analyzing continuous or discrete-time signals, has significant application value in fields such as communication transmission, biomedical diagnostics, industrial process control, and financial quantitative analysis. The core objective of this technology is to extract effective information from time series data streams to achieve key tasks such as signal noise reduction and enhancement, feature pattern recognition, future trend prediction, and state classification.
[0003] The theoretical foundation of time-series signal processing can be traced back to classical frequency domain analysis methods. Fourier transform achieves global frequency analysis by decomposing a signal into spectral components, but it cannot reveal the frequency domain characteristics of a signal within a local time window. To overcome this limitation, wavelet transform introduces scaling and time-shifting mechanisms, constructing a multi-resolution analysis framework that significantly improves the ability to capture transient features. With the evolution of computing power, joint time-frequency analysis methods have become a research hotspot. Short-time Fourier transform uses a fixed window function for local spectral analysis, but it is limited by the Heisenberg uncertainty principle, making it difficult to simultaneously optimize time and frequency resolution. While quadratic time-frequency analysis tools such as the Wigner-Ville distribution can provide more accurate characterization of time-frequency energy concentration, their practical application is affected by cross-term interference generated by multi-component signals.
[0004] In recent years, machine learning models have significantly expanded the boundaries of time series modeling. Long Short-Term Memory (LSTM) networks, through gating mechanisms, effectively learn long-sequence dependencies, demonstrating outstanding performance in tasks such as speech recognition and financial time series forecasting. For example, their fusion framework with Kalman filtering can synergistically suppress observation noise and improve the robustness of multi-step predictions. The Transformer architecture utilizes self-attention mechanisms to establish global dependency modeling capabilities, achieving breakthroughs in scenarios such as power load forecasting and medical physiological signal analysis. These methods provide new technical pathways for mining complex time series patterns.
[0005] With the widespread adoption of high-precision sensing technology, the amount of time-series data collected across various industries has exploded. This data contains highly heterogeneous patterns, including abrupt jumps, periodic oscillations, high-frequency fluctuations, and sharp abrupt changes. While existing technologies can handle specific types of time-series tasks, they lack the ability to systematically summarize knowledge about massive amounts of heterogeneous time-series fragments. There is an urgent need to construct a cognitive system covering the entire spectrum of time-series patterns. This system should identify frequently occurring typical state patterns, regular evolutionary trajectories, and characteristic waveform structures to form a priori knowledge base that can support prediction algorithms, regression models, similarity retrieval, and historical data tracing. This requires establishing a deep cognitive framework for the inherent evolutionary laws of time-series data, thereby laying the theoretical foundation for advanced time-series intelligent applications. Summary of the Invention
[0006] The purpose of this invention is to address the limitation of massive heterogeneous time-series data lacking systematic summarization and unable to form a typical pattern knowledge base. Therefore, it proposes a time-series sequence mining method based on multi-level features. This invention can realize the generation of typical time-series patterns, thereby generating a dedicated pattern knowledge base for time-series data, providing basic theoretical data support for the development of algorithms based on this time-series data.
[0007] The present invention employs the following technical solutions to achieve its objective: A time-series sequence mining method based on multi-level features includes the following steps: S1. Obtain the time-series data corresponding to the current business information, and slice the time-series data in the time domain dimension to obtain multiple time-series fragments; S2. For each time series segment, extract its multi-level features; these multi-level features include parametric statistical distribution features, differential statistical distribution features, and expert-level features. S3. Perform a weighted summation of the multi-level features of each time series segment to obtain the comprehensive features of the corresponding time series segment; S4. Based on the comprehensive characteristics of each time series segment, the representativeness of each time series segment as a typical pattern is quantified by measuring the similarity between segments. The representativeness of each time series segment is sorted, and redundant segments with similarity reaching a preset threshold are removed. The remaining time series segments are the typical patterns. After arranging the typical patterns according to a preset method, the typical time series pattern corresponding to the time series data is formed, thus completing the mining process.
[0008] Preferably, in step S1, outliers in the time series data are removed before slicing the time series data; when slicing the time series data, the time domain slicing process is performed according to the preset duration and overlap.
[0009] Specifically, in step S2, each time series segment has a composite data set consisting of multiple interrelated dimensional parameters. This composite data set is used to describe the state of the business system at a specific time point or within a specific time period from different dimensions, serving as the basis for the extraction of corresponding hierarchical features.
[0010] Specifically, for the extraction of parameter statistical distribution features, firstly, histogram statistical calculations are performed on the original observed values of multiple dimensions of parameters in the corresponding time series segment to obtain the statistical histogram sequence of each dimension parameter value; then, the statistical histogram sequence of each dimension parameter value is normalized to obtain the normalized histogram sequence representing the distribution of each dimension parameter in the time series segment. This completes the extraction of the statistical distribution characteristics of the parameters.
[0011] Specifically, for the extraction of difference statistical distribution features, firstly, histogram statistical calculations are performed on the first differences of the original observations of multiple dimensions of the corresponding time series segment to obtain the difference statistical histogram sequences of each dimension parameter; then, the difference statistical histogram sequences of each dimension parameter are normalized to obtain the normalized histogram sequences representing the changes of each dimension parameter in the time series segment. The differential statistical distribution characteristics are extracted; the first-order difference is performed using forward difference, backward difference, or central difference.
[0012] Preferably, expert-level features are divided based on the characteristics of each time-series segment, including conventional expert features and domain expert features; the sequence is formed by the combination of conventional expert features and domain expert features. To characterize the expert-level features of the corresponding time sequence segments.
[0013] Specifically, typical expert characteristics include any one or more of the following combinations: Fourier series information characterizes the periodicity of time series segments; interval error information characterizes the wave-like qualitative characteristics of time series segments; range information characterizes the extreme range of jitter in time series segments; interquartile range information characterizes the distribution of data in time series segments; autocorrelation information characterizes the correlation between adjacent points in time series segments.
[0014] Specifically, domain expert characteristics include any one or more of the following combinations: Volatility is used to capture the time-varying jitter characteristics of time series segments; jump detection is used to capture sudden and discontinuous jitter in time series segments; eye diagram is used to observe the impact of jitter on signal decision by superimposing multiple periodic signals, and the jitter and noise tolerance are characterized by eye width and eye height.
[0015] Preferably, in step S3, the parametric statistical distribution characteristics, differential statistical distribution characteristics, and expert-level characteristics of each time series sequence segment are respectively denoted as sequence. , and Then, preset weight parameters , and The comprehensive feature obtained after weighted summation is... As shown in the following formula:
[0016] The obtained comprehensive features Store it in the database for use in step S4.
[0017] Furthermore, in step S4, the quantification of the representativeness of each temporal sequence segment as a typical pattern by measuring the similarity between segments specifically involves: The total amount of time series data The time series sequence segments are denoted as follows: Each time series segment is assigned a corresponding typicality parameter value, denoted as follows: The initial value of the typicality parameter is 0. For the A time series sequence fragment Comprehensive characteristics Calculate its relationship with the remaining time series segments in the time series data. Corresponding comprehensive features The Euclidean distance between them yields the first... A time series sequence fragment Euclidean distance difference sequence , That is to represent the first A time series sequence fragment With the A time series sequence fragment The Euclidean distance between the comprehensive characteristics of the two sides, and so on; Arrange the elements in the obtained Euclidean distance difference sequence in ascending order, and then sort the first... The typicality parameter value of the time series segment corresponding to each ranking is incremented by one; after the same operation is performed on each time series segment, the typicality parameter value corresponding to each time series segment is the degree of representativeness of the time series segment as a typical pattern. The process of sorting the representativeness of each time series segment specifically involves arranging them in descending order of the specific numerical value of each typicality parameter to form a new typicality sequence. The corresponding time sequence segment is also denoted as .
[0018] Furthermore, the removal of redundant segments whose similarity reaches a preset threshold specifically involves: For the newly arranged time sequence fragments ,by For the initial selection of segments, calculate and The Euclidean distance between the integrated features corresponding to these two segments, if the Euclidean distance is less than a preset threshold, represents... and The differences are small. Can be replaced Thus The rearranged sequence Eliminate from the middle; continue the comparison in this way. and until found and Once the Euclidean distance between the integrated features corresponding to these two segments is greater than or equal to a preset threshold, then... As a new initial selection fragment, subsequent calculations and comparisons are performed in the same manner; Complete the entire sequence After traversal, the sequence The remaining time series segments are multiple typical patterns. After being sorted in descending order according to the representativeness of each time series segment as a typical pattern, the typical time series pattern corresponding to the time series data is formed.
[0019] In summary, due to the adoption of this technical solution, the beneficial effects of this invention are as follows: This invention effectively solves the problem of systematically extracting typical variation patterns from massive time-series data through a multi-level feature mining mechanism. It can automatically identify and summarize high-frequency typical patterns in heterogeneous forms such as cliff oscillations, regular fluctuations, and rapid fluctuations. By deeply analyzing the multi-scale feature correlations of time-series signals, it overcomes the limitations of traditional methods in understanding fragmented time-frequency features, significantly improving the joint representation ability of long-range evolutionary patterns and short-term abrupt changes. Based on this method, the mined typical time-series patterns can be constructed into a standardized knowledge base, providing interpretable prior knowledge of time-series patterns for subsequent business development prediction algorithms, greatly enhancing the robustness of multi-step predictions. Simultaneously, it establishes a unified knowledge support framework for tasks such as regression modeling, similarity retrieval, and data backtracking, significantly reducing algorithm development complexity and improving cross-scenario transferability. Attached Figure Description
[0020] The present invention is further described in detail with reference to the following figures, which include 7 figures as follows: Figure 1This is a schematic diagram illustrating the overall process of the time series mining method of the present invention; Figure 2 This is a detailed flowchart of steps S1 to S3 in the method of the present invention; Figure 3 This is a detailed flowchart of step S4 in the method of the present invention; Figure 4 This is a schematic diagram of the mining results of a typical pattern of fluctuation + burr in this invention. Figure 5 This is a schematic diagram illustrating the mining results of a typical pattern of regular, periodic increases, as exemplified by the present invention. Figure 6 This is a schematic diagram of the mining results of a typical pattern of flatness + mutation in this invention. Figure 7 This is a schematic diagram of the mining results for a typical pattern of flatness + abrupt change points, as exemplified by this invention. Detailed Implementation
[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.
[0022] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.
[0023] A time-series sequence mining method based on multi-level features. Figure 1 This document provides a brief overview of the overall process of the method, which can be viewed concurrently. The key steps of the method can be summarized as follows: S1. Obtain the time-series data corresponding to the current business information, and slice the time-series data in the time domain dimension to obtain multiple time-series fragments; S2. For each time series segment, extract its multi-level features; these multi-level features include parametric statistical distribution features, differential statistical distribution features, and expert-level features. S3. Perform a weighted summation of the multi-level features of each time series segment to obtain the comprehensive features of the corresponding time series segment; S4. Based on the comprehensive characteristics of each time series segment, the representativeness of each time series segment as a typical pattern is quantified by measuring the similarity between segments. The representativeness of each time series segment is sorted, and redundant segments with similarity reaching a preset threshold are removed. The remaining time series segments are the typical patterns. After arranging the typical patterns according to a preset method, the typical time series pattern corresponding to the time series data is formed, thus completing the mining process.
[0024] For detailed procedures from steps S1 to S3, please refer to [link / reference]. Figure 2 As illustrated, in step S1 of this embodiment, outliers in the time series data are pre-removed before slicing the time series data; when slicing the time series data, the time domain slicing process is performed according to the preset duration and overlap.
[0025] This step involves acquiring time-series data corresponding to the current business information and preprocessing it to improve the accuracy and efficiency of subsequent analysis. First, time-series data refers to a data stream that changes over time, typically originating from sensors, business systems, or other data acquisition devices, reflecting the dynamic changes of a specific business scenario. After acquiring this data, preliminary data cleaning is required to remove outliers or singular values. Outliers are isolated points that do not conform to the overall data distribution pattern and may be caused by equipment failure, external interference, or data input errors. Removing these values helps avoid noise interference with subsequent slicing and feature extraction, improving data quality. Specific methods may include calculating statistical indicators such as the mean or standard deviation, setting a reasonable threshold range, identifying and removing data points exceeding that range, thereby ensuring the stationarity and reliability of the time-series data.
[0026] After the time-series data is cleaned, it undergoes time-domain slicing. This process divides the continuous time series into multiple shorter time-series segments, each representing an independent time window. The slicing operation is based on preset duration and overlap parameters. The duration defines the duration of each segment, and its setting needs to be flexibly adjusted according to the specific needs of the business scenario. For example, in a biosignal monitoring scenario, the duration can be set to capture the time period of physiological homeostasis information, ensuring that the segment covers key change cycles; in a financial data analysis scenario, the duration may cover a complete market cycle to reflect long-term trends. The overlap parameter controls the proportion of time overlap between adjacent segments. By appropriately setting the overlap, the continuity between segments can be enhanced, information loss can be prevented, and subsequent feature extraction can capture the transition characteristics of the sequence.
[0027] The core principle of the entire slicing process lies in decomposing long sequences into manageable sub-units, facilitating the extraction and comparison of subsequent multi-level features. By dynamically adjusting the duration and overlap, this implementation can adapt to the heterogeneity of different business scenarios and optimize search efficiency. Ultimately, the resulting multiple time-series sequence fragments provide the basic input for subsequent steps, ensuring the robustness and adaptability of the entire mining method in complex time-series data.
[0028] In step S2 of this embodiment, each time series segment has a composite data set consisting of multiple interrelated dimensional parameters. This composite data set is used to describe the state of the business system at a specific time point or time period of the corresponding time series segment from different dimensions, serving as the basis for the extraction of corresponding hierarchical features.
[0029] This forms the basis for multi-level feature extraction, namely, analyzing the complex data structure inherent in each time series segment. Each segment contains multiple interrelated dimensional parameters (e.g., a bank's securities holding information includes the number of shares held, the proportion of shares held, and the cost of shares held; a person's heartbeat information includes heart rate, P wave, and stroke volume). These parameters together constitute a dataset describing the state of the business system within a specific time period. The core objective of extracting the statistical distribution characteristics of these parameters is to quantify the overall distribution pattern of the observed values of each dimensional parameter within the corresponding time series segment.
[0030] In this implementation, the extraction of parameter statistical distribution characteristics first involves performing histogram statistical calculations on the original observed values of multiple dimensions of parameters in the corresponding time series segment to obtain the statistical histogram sequence of each dimension parameter. Then, the statistical histogram sequence of each dimension parameter is normalized to obtain a normalized histogram sequence representing the distribution of each dimension parameter in the time series segment. This completes the extraction of the statistical distribution characteristics of the parameters.
[0031] The feature extraction process begins with processing the original observation sequence for each dimension parameter. By statistically analyzing all observation points for a specific dimension within a time series segment, the frequency of each observation falling into multiple pre-defined consecutive numerical intervals is calculated, thus forming a histogram sequence characterizing the distribution of values for that dimension. This histogram sequence intuitively reflects the probability density of different value levels for that dimension parameter within the time window. Subsequently, to eliminate the influence of differences in units or numerical ranges between different dimensions and to highlight the distribution pattern itself, the original histogram sequence generated for each dimension needs to be normalized. Normalization typically converts the frequency of each interval into the proportion of that interval's frequency to the total number of observation points, or further into a probability density form. The normalized histogram sequence obtained after normalization has a sum of one element, stably and consistently characterizing the statistical distribution characteristics of that dimension parameter within this time series segment, such as the central tendency, dispersion, and skewness of the distribution. Finally, the normalized histogram sequences for all dimensions... Together, these constitute the parametric statistical distribution feature vector of this time series segment, providing the basic dimensions for describing the static distribution characteristics of the data for subsequent comprehensive feature fusion. If the dimensional parameters of two different time series segments are almost identical, then their normalized histogram sequences... It will be very close.
[0032] This implementation method extracts the difference statistical distribution characteristics by first performing histogram statistical calculations on the first differences of the original observations of multiple dimensional parameters of the corresponding time series segment, obtaining the difference statistical histogram sequence of each dimensional parameter; then, normalizing the difference statistical histogram sequence of each dimensional parameter to obtain the normalized histogram sequence representing the changes of each dimensional parameter in the time series segment. The differential statistical distribution characteristics are extracted; the first-order difference is performed using forward difference, backward difference, or central difference.
[0033] The purpose of extracting differential statistical distribution features is to capture the dynamic evolution characteristics of parameters in each dimension within a time series segment. This feature reveals the implicit changing patterns of the business system's state by analyzing the changing trends and fluctuation patterns of the original observation sequence. This implementation performs a difference operation on the original observation sequence for each dimension parameter, calculating the numerical changes between adjacent observation points to generate a first-order difference sequence. This process strips away the absolute levels of the original data, focusing instead on the rate and direction of change.
[0034] Differential operations can employ different implementation strategies to adapt to the needs of various business scenarios. Forward differencing reflects the instantaneous trend of future changes by calculating the difference between the current point and the next observation point; backward differencing reflects the cumulative effect of historical changes by using the difference between the current point and the previous point; while center differencing combines forward and backward differencing, taking the average of the differences between two adjacent points, effectively smoothing random noise and highlighting the direction of continuous change. The choice of differencing strategy depends on the trade-off between real-time performance and noise resistance in the specific application scenario. For example, forward differencing may be preferred in real-time monitoring scenarios that require rapid response, while center differencing tends to be used in high-noise environments.
[0035] The generated difference sequence is essentially a time-varying signal characterizing the rate of change of the system state. By performing statistical histogram calculations and normalization on this sequence, the resulting difference statistical distribution feature vector can quantify the typical intensity of change, frequent fluctuation ranges, and stability of the change trend of parameters in each dimension within a time segment. For example, a difference distribution concentrated near zero suggests a stable system state, while a wide-range or skewed distribution indicates severe fluctuations or persistent drift. This feature provides a crucial supplementary perspective for understanding the dynamic behavior patterns of business systems, forming a static-dynamic complement to the parameter statistical distribution features.
[0036] In this embodiment, expert-level features are divided according to the characteristics of each time-series segment, including conventional expert features and domain expert features; the sequence is formed by the combination of conventional expert features and domain expert features. To characterize the expert-level features of the corresponding time sequence segments.
[0037] Typical expert characteristics include any one or more of the following combinations: Fourier series information characterizes the periodicity of time series segments; interval error information characterizes the wave-like qualitative characteristics of time series segments; range information characterizes the extreme range of jitter in time series segments; interquartile range information characterizes the distribution of data in time series segments; autocorrelation information characterizes the correlation between adjacent points in time series segments.
[0038] Domain expert characteristics include any one or more of the following: Volatility is used to capture the time-varying jitter characteristics of time series segments; jump detection is used to capture sudden and discontinuous jitter in time series segments; eye diagram is used to observe the impact of jitter on signal decision by superimposing multiple periodic signals, and the jitter and noise tolerance are characterized by eye width and eye height.
[0039] This implementation method integrates general analysis rules and domain-specific prior knowledge in the construction of expert-level features, forming a quantitative representation of the deep-seated business meaning of time-series segment. Conventional expert features, based on mathematical statistics and signal processing principles, characterize the fundamental patterns of the sequence from a universal perspective. For example, Fourier series information reveals the periodicity and frequency cost of the sequence through spectral analysis; interval error information quantifies the cumulative deviation of the sequence waveform from the ideal clock edge; range and interquartile range describe the extreme fluctuation range and distribution concentration of the data, respectively; and autocorrelation captures the linear dependence between adjacent points within the sequence. These features collectively construct a general behavioral profile of the time-series segment.
[0040] Domain expert features further incorporate specialized models for specific business scenarios, transforming abstract domain knowledge into computable indicators. In financial time series analysis, volatility features use a sliding window to statistically analyze the variance of historical price changes, dynamically capturing the evolution of market risk; jump detection uses statistical hypothesis testing to identify discontinuous abrupt changes in price series, pinpointing the impact of sudden events. In communication signal quality assessment, eye diagram features overlay periodic signal waveforms to generate a two-dimensional image reflecting signal integrity; its eye width directly characterizes the tolerance boundary of time series jitter, while eye height indicates signal stability under noise interference. These features transform the domain expert's interpretation logic into machine-interpretable quantitative parameters, enabling the search process to possess business semantic understanding capabilities. The synergistic effect of these two types of features makes the expert-level feature layer both universal and scenario-adaptable, providing high-level decision-making basis for comprehensive feature fusion.
[0041] Subsequently, in step S3 of this embodiment, the parametric statistical distribution characteristics, differential statistical distribution characteristics, and expert-level characteristics of each time series sequence segment are respectively denoted as sequences. , and Then, preset weight parameters , and The comprehensive feature obtained after weighted summation is... As shown in the following formula:
[0042] The obtained comprehensive features Store it in the database for use in step S4.
[0043] Step S3 fuses multi-level features using preset weight parameters to generate a unified representation of time-series sequence fragments. After mapping parametric statistical distribution features, differential statistical distribution features, and expert-level features into feature sequences, corresponding weights are assigned based on their business importance. The weighted summation process uses a linear combination method to superimpose the three feature sequences according to their weights, forming a single comprehensive feature vector. This fusion mechanism takes into account static distribution characteristics, dynamic evolution laws, and domain knowledge representation, while flexibly adapting to the core needs of different business scenarios through weight adjustment. The final generated comprehensive feature... This information is stored in the database as a global descriptor for the time-series segment, providing a standardized comparison basis for subsequent similarity searches.
[0044] like Figure 3 As shown, in step S4 of this embodiment, the representativeness of each time-series sequence segment as a typical pattern is quantified by measuring the similarity between segments. Specifically: The total amount of time series data The time series sequence segments are denoted as follows: Each time series segment is assigned a corresponding typicality parameter value, denoted as follows: The initial value of the typicality parameter is 0. For the A time series sequence fragment Comprehensive characteristics Calculate its relationship with the remaining time series segments in the time series data. Corresponding comprehensive features The Euclidean distance between them yields the first... A time series sequence fragment Euclidean distance difference sequence , That is to represent the first A time series sequence fragment With the A time series sequence fragment The Euclidean distance between the comprehensive characteristics of the two sides, and so on; Arrange the elements in the obtained Euclidean distance difference sequence in ascending order, and then sort the first... The typicality parameter value of the time series segment corresponding to each ranking is incremented by one; after the same operation is performed on each time series segment, the typicality parameter value corresponding to each time series segment is the degree of representativeness of the time series segment as a typical pattern. The process of sorting the representativeness of each time series segment specifically involves arranging them in descending order of the specific numerical value of each typicality parameter to form a new typicality sequence. The corresponding time sequence segment is also denoted as .
[0045] In this embodiment, Euclidean distance is used to characterize the similarity between the comprehensive features of two segments. This distance metric calculates the sum of the squares of the differences in each dimension of the two comprehensive feature vectors, and then takes the square root to obtain the numerical result. The smaller the Euclidean distance value, the closer the geometric positions of the two feature vectors are in the multidimensional space, that is, the more similar the time-series segments they represent are in terms of statistical characteristics, dynamic evolution, and business semantics. This serves as the basis for accumulating the typicality parameter value.
[0046] Finally, redundant segments with similarity reaching a preset threshold are removed, specifically: For the newly arranged time sequence fragments ,by For the initial selection of segments, calculate and The Euclidean distance between the integrated features corresponding to these two segments, if the Euclidean distance is less than a preset threshold, represents... and The differences are small. Can be replaced Thus The rearranged sequence Eliminate from the middle; continue the comparison in this way. and until found and Once the Euclidean distance between the integrated features corresponding to these two segments is greater than or equal to a preset threshold, then... This serves as a new initial selection fragment, and subsequent calculations and comparisons are performed in the same manner.
[0047] Complete the entire sequence After traversal, the sequence The remaining time series segments are multiple typical patterns. After being sorted in descending order according to the representativeness of each time series segment as a typical pattern, the typical time series pattern corresponding to the time series data is formed.
[0048] The following describes the process of mining an example time series sequence in this embodiment.
[0049] First, the time-series data is sliced in the time domain. In this implementation, a time length of 10 seconds is used for slicing. Each time series sequence segment is analyzed; then, the parametric statistical distribution characteristics of each time series segment are extracted. Difference statistical distribution characteristics and expert-level features Set certain weight parameters , and The comprehensive features are obtained after weighted summation. .
[0050] Set a typicality sequence with all initial values of 0. This is used to characterize the representativeness of each time series segment; a higher typicality value indicates that the segment is more representative. Then, starting with the first time series segment, the similarity between it and the combined features of all time series segments excluding itself is calculated to find the most similar segment. A segment; assuming that after arrangement and When the first time series segment has the highest similarity to the 100th, 200th, and 300th time series segments respectively, then the typicality parameters corresponding to these three segments are... , and The values are all incremented by 1. Then the second, third, and so on are calculated sequentially. The relevant information about each time series segment, in Under the given conditions, the corresponding typicality parameter is incremented by 1. This yields the final typicality value sequence for all time series fragments. .
[0051] The sequence that already has specific values The values are rearranged in descending order to obtain a new typicality sequence. The corresponding newly arranged time sequence fragment Then with Corresponding fragments Beginning, with As a reference segment, the calculation and sorting are performed on the first segment that follows. If the feature distance is less than a threshold, it is discarded. and Continue calculating and sorting the first segment that follows. If the feature distance is less than a threshold, it is discarded and the calculation continues. And so on; until the fragment is calculated. hour, and If the difference in features between them is greater than or equal to a threshold, then... As a new reference segment, continue calculating and sorting the first segment that follows it. The feature distance, and so on; until the entire sequence Once the traversal is complete, the deduplication of the segments is finished, and the segments that are not removed are typical temporal patterns. Figures 4 to 7 The diagram illustrates the mining results for four typical time series patterns. The method and process used in this implementation are shown. The typical patterns in the mined and determined time series data have good results and can be stored in the database for use by other business tasks.
Claims
1. A multi-level feature-based time series sequence mining method, characterized in that, The method comprises the following steps: S1, obtaining time sequence data corresponding to current service information, slicing the time sequence data in time domain dimension to obtain a plurality of time sequence segments; S2, for each time sequence segment, extracting multi-level features thereof; The multi-level features comprise parameter statistical distribution features, differential statistical distribution features and expert-level features; S3, performing weighted summation on the multi-level features of each time sequence segment to obtain comprehensive features of the corresponding time sequence segment; S4, according to the comprehensive features of each time sequence segment, quantifying the representative degree of each time sequence segment as a typical pattern by measuring the similarity between segments, sorting the representative degrees of each time sequence segment, eliminating redundant segments with a similarity reaching a preset threshold, and the remaining plurality of time sequence segments being the plurality of typical patterns, which are arranged in a preset manner to form the typical time sequence pattern corresponding to the time sequence data, thereby completing the mining process.
2. The method of claim 1, wherein: In step S1, before slicing the time sequence data, the abnormal values in the time sequence data are eliminated in advance; when slicing the time sequence data, the time domain slicing process is performed according to a preset time length and overlap degree.
3. The method of claim 1, wherein: In step S2, each time sequence segment has a composite data set composed of a plurality of interrelated dimension parameters, which is used to describe the state of the business system at the time point or time period of the corresponding time sequence segment from different dimensions as the basis for extracting corresponding level features.
4. The method of claim 3, wherein: For the extraction of the parameter statistical distribution feature, firstly, the original observation values of the multiple dimension parameters of the corresponding time sequence segment are subjected to histogram statistical calculation to obtain the value statistical histogram sequence of each dimension parameter; subsequently, the value statistical histogram sequence of each dimension parameter is subjected to normalization operation to obtain the normalized histogram sequence representing the distribution of each dimension parameter in the time sequence segment, and the extraction of the parameter statistical distribution feature is completed. , the extraction of the parameter statistical distribution feature is completed.
5. The method of claim 3, wherein: For the extraction of the differential statistical distribution feature, first, the first-order difference of the original observation values of the multiple dimension parameters of the corresponding time sequence segment is subjected to histogram statistical calculation to obtain the differential statistical histogram sequence of each dimension parameter; subsequently, the differential statistical histogram sequence of each dimension parameter is subjected to normalization operation to obtain the normalized histogram sequence representing the change of each dimension parameter in the time sequence segment , the extraction of the differential statistical distribution feature is completed; wherein the first-order difference adopts forward difference, backward difference or central difference.
6. The method of claim 3, wherein: The expert-level features are divided according to the characteristics of each time sequence segment, including conventional expert features and domain expert features; The sequence is formed by the conventional expert features and the domain expert features together to characterize the corresponding timing sequence segment.
7. The method of claim 6, wherein: The conventional expert features include any one or a combination of the following: Fourier series information representing the periodicity of the time sequence segment; interval error information representing the wave nature of the time sequence segment; range information representing the extreme range of jitter in the time sequence segment; quartile distance information representing the distribution information of data in the time sequence segment; autocorrelation information representing the correlation information of adjacent points in the time sequence segment; The domain expert features include any one or a combination of the following: Volatility for capturing the time-varying jitter characteristics of the time sequence segment; jump detection for capturing sudden and discontinuous jitter in the time sequence segment; eye diagram for observing the influence of jitter on signal decision by superimposing a plurality of periodic signals, and representing jitter and noise tolerance by eye width and eye height.
8. The method of claim 1, wherein: In step S3, the parameter statistical distribution features, the difference statistical distribution features and the expert-level features of each time sequence segment are respectively recorded as sequence , and respectively. After that, preset weight parameters , and are obtained, and the comprehensive features after weighted summation are as follows: The resulting overall feature is stored in the database for use in step S4.
9. The method of claim 8, wherein, In step S4, the representative degree of each time sequence segment as a typical pattern is quantified by measuring the similarity between segments, specifically: a total of the time series segments are respectively denoted as a corresponding typicality parameter value is configured for each time series segment, respectively denoted as ; The initial value of the typical degree parameter is 0; For the A time series sequence fragment Comprehensive characteristics Calculate its relationship with the remaining time series segments in the time series data. Corresponding comprehensive features The Euclidean distance between them yields the first... A time series sequence fragment Euclidean distance difference sequence , That is to represent the first A time series sequence fragment With the A time series sequence fragment The Euclidean distance between the comprehensive characteristics of the two sides, and so on; Arrange the elements in the obtained Euclidean distance difference sequence in ascending order, and then sort the first... The typicality parameter value of the time series segment corresponding to each ranking is incremented by one; after the same operation is performed on each time series segment, the typicality parameter value corresponding to each time series segment is the degree of representativeness of the time series segment as a typical pattern. The representative degrees of the respective time sequence segments are sorted, specifically, the each representative degree parameter value is arranged in descending order according to the specific numerical value, and a new representative degree sequence is formed , and the corresponding time sequence segment is also recorded as .
10. The method of claim 9, wherein, The redundant segments with a similarity reaching a preset threshold are eliminated, specifically: For the newly arranged time sequence segment , the sequence is calculated , and the Euclidean distance between the comprehensive features corresponding to the two segments is calculated. If the Euclidean distance is less than a preset threshold value, it represents that the difference between and is small, can replace , so that is eliminated from the newly arranged sequence . In this way, the comparison between and is continued until the Euclidean distance between the comprehensive features corresponding to the two segments and is greater than or equal to the preset threshold value. Then, is taken as a new initial screening segment and the same subsequent calculation comparison is performed. after the traversal of the entire sequence The remaining time sequence fragments in the sequence are the typical patterns. After arranging the time sequence fragments in descending order according to the representative degree of each time sequence fragment as a typical pattern, the typical time sequence patterns corresponding to the time sequence data are formed.