Search

EP-4307633-B1 - DETERMINING TRENDS OF A TIME SERIES

EP4307633B1EP 4307633 B1EP4307633 B1EP 4307633B1EP-4307633-B1

Inventors

  • Ledoux, Veerle
  • Dreesen, Dennis

Dates

Publication Date
20260513
Application Date
20220713

Claims (15)

  1. A computer implemented method (200) for predicting future alarm threshold breaches in a network based on a time series of a network metric indicative for a state of a node within the network by determining trends of the time series (150) at different time resolutions; wherein the time series has consecutive values (213) at an initial time resolution, and wherein the initial time resolution defines a smallest granularity period (221) of a set of increasing granularity periods (221, 222, 223, 224, 225) within the time series; the method comprising, upon ending of a respective granularity period of the set of increasing granularity periods: - determining (201) an aggregated value (211) indicative for a statistical representation of consecutive values during the respective granularity period; - determining (202) a difference value (230) between the aggregated value and a preceding aggregated value; and - updating (203) a slope estimation (240) of the time series for the respective granularity period based on the difference value, a preceding slope estimation, and a first smoothing factor shared amongst slope estimations (241, 242) for the respective granularity periods.
  2. The computer implemented method according to claim 1, further comprising, for respective granularity periods, extrapolating (241, 242) the aggregated value (211) a number of granularity periods in time based on the slope estimation (240).
  3. The computer implemented method according to claim 2, further comprising detecting an alarm candidate for a respective granularity period if the slope estimation (240) and/or the extrapolated aggregated value (241, 242) violates an alarm threshold (243).
  4. The computer implemented method according to claim 2, wherein extrapolating the aggregated value for respective granularity periods further comprises determining (303) a confidence band (325) of the extrapolation (242) based on the slope estimation (240) and a slope deviation (321) indicative for the stability of the time series.
  5. The computer implemented method according to claim 4, further comprising determining (302) the slope deviation based on a preceding slope deviation (318), a second smoothing factor (316) derived from the first smoothing factor, the slope estimation (240), and an adaptive slope estimation (314) indicative for the contribution of the recent time series values to the trend of the time series.
  6. The computer implemented method according to claim 5, further comprising determining (301) the adaptive slope estimation (314) based on the difference value (230), a preceding adaptive slope estimation (315), and a third smoothing factor (313) derived from the first smoothing factor.
  7. The computer implemented method according to any of claims 4 - 6, further comprising detecting an alarm candidate for a respective granularity period if the confidence band (325) violates an alarm threshold (326).
  8. The computer implemented method according to any of the preceding claims, further comprising initializing (404) the slope estimation by, during an initialization period, performing the updating (409) of the slope estimation with less smoothening than with the first smoothing factor.
  9. The computer implemented method according to any of the preceding claims, further comprising performing the updating of the slope estimation for a respective granularity period with more smoothening than with the first smoothing factor if the difference value is an outlier.
  10. The computer implemented method according to claim 3 or 7, further comprising, upon detecting (501) an alarm candidate for a respective granularity period, training (504) one or more prediction models 515 based on previous stored time series values (511).
  11. The computer implemented method according to claim 10, wherein training the one or more prediction models comprises: - retrieving (502) stored time series values (511) having a time resolution comparable to the respective granularity period of the detected alarm candidate; - dividing (503) the retrieved time series values in a training set (513) and a test set (514); - training (504) one or more prediction models based on the training set; and - selecting (505) a prediction model from the one or more trained prediction models (515) that predicts the test set with the smallest error.
  12. The computer implemented method according to claim 10 or 11, further comprising verifying (506) the detected alarm candidate by predicting the time series values in time (516) based on the one or more prediction models.
  13. A data processing system configured to perform the computer implemented method according to any one of claims 1 to 12.
  14. A computer program comprising instructions which, when the program is executed by a computer, cause the computer to perform the computer implemented method according to any one of claims 1 to 12.
  15. A computer-readable medium comprising instructions which, when executed by a computer, cause the computer to perform the computer implemented method according to any one of claims 1 to 12.

Description

Field of the Invention The present invention generally relates to, amongst others, forecasting time series. Background of the Invention Time series forecasting aims to predict future values of a time series, e.g. in monitoring network devices in a telecommunication network. Hereto, a prediction model can be trained on a substantial amount of previous time series values, e.g. by means of machine learning. These prediction models are often trained for a specific time series and a single time resolution, which limits the range of observable temporal patterns in the forecast. Additionally, training a prediction model is typically performed offline and is resource intensive, e.g. uses substantial processing power, a large memory footprint, and long training cycles. As such, when forecasting a plurality of time series, in particular time series with different behaviours, it is expensive and inefficient to train a prediction model for every time series at different time resolutions. It is thus a problem to forecast different temporal patterns for a plurality of time series with limited computational resources, i.e. the scalability of existing solutions is a problem. EP 3401789 A1 discloses a method for detection of anomalies in a time series comprising: a) obtaining a next value of the time series; b) updating a plurality of derived time series thereby obtaining a next derived value; c) predicting a future derived value of the derived time series and a confidence indicator; d) retrieving, a previously predicted derived value and a confidence indicator of the previously predicted derived value as a prediction of the next derived value; e) detecting a change point when the difference between the next derived value and the previously predicted derived value is larger than indicated by the confidence indicator of the previously predicted derived value; f) detecting an anomaly in the time series based on the detection of one or more change points at the one or more resolutions of the derived time series. Summary of the Invention It is an object of the present invention, amongst others, to solve or alleviate the above identified challenges and problems by improving online forecasting of time series. According to a first aspect, this object is achieved by a computer implemented method as set out in claim 1. The respective granularity periods in the set of increasing granularity periods comprise a predetermined number of consecutive values, i.e. have a distinct length. The distinct length of the respective granularity periods may be a multiple of the length of the smallest granularity period. The consecutive values of the time series may be raw data, e.g. a polled metric in a telecommunications network. Alternatively, the consecutive values may be pre-processed or aggregated raw data. An aggregated value of the consecutive values in a respective granularity period is determined upon ending of the respective granularity period, i.e. when a next value of the time series is obtained thereby completing the period. The aggregated value may for example be a mean, a cumulative sum, a median, or any other statistical representation of consecutive values in the respective granularity period. The aggregated value for a larger granularity period may be determined based on aggregated values for a smaller granularity period, as the larger granularity period includes the consecutive values of the smaller granularity period. This allows to keep a limited amount of values in memory, i.e. the preceding aggregated value for the respective granularity periods. The slope estimations may express a change in the consecutive values over a respective granularity period. The slope estimation for a respective granularity period may be updated based on a weighted sum of the difference value, i.e. a recent change in the time series, and the preceding slope estimation, i.e. previous behaviour of the time series. Herein, the weights may be based on the first smoothing factor. The first smoothing factor may have a fixed or variable value that results in a relatively stable slope estimation by assigning more weight to the most recent slope estimation. The slope estimations may be updated for each granularity period in the set of granularity periods. Alternatively, the slope estimations may be updated for a portion of the granularity periods in the set of granularity periods thereby allowing to reduce the consumed computational resources. Updating slope estimations for the respective granularity periods allows to forecast or predict different temporal patterns of the time series at different prediction horizons without keeping a large amount of consecutive values in memory. The aggregated values for the respective granularity periods allow to use a single smoothing factor for updating the slope estimations for the respective granularity periods, further limiting memory usage. It is an advantage that the determining of trends of a time series can be perfo