Method for estimating a target variable

By aggregating and clustering time series data, the method addresses the challenge of estimating rarely measured target variables, resulting in a precise and interpretable regression model.

WO2026234241A1PCT designated stage Publication Date: 2026-11-12SIEMENS MOBILITY GMBH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
PCT/EP2026/063424
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Priority Date
2025-05-09
Filing Date
2026-05-08
Publication Date
2026-11-12

AI Technical Summary

Technical Problem

Conventional regression models struggle to accurately estimate rarely measured target variables due to the mismatch between the large number of time series data and the small number of target variables, leading to inaccurate and less interpretable models.

Method used

A method involving data processing to aggregate time series data into sections, cluster these sections based on event characteristics, and train a regression model using both time series and rare target variables, enabling accurate estimation of target variables.

Benefits of technology

This approach creates a powerful and interpretable regression model that accurately predicts target variables, reconciling the disparity between abundant time series data and scarce target variables.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure EP2026063424_12112026_PF_FP_ABST
    Figure EP2026063424_12112026_PF_FP_ABST
Patent Text Reader

Abstract

The invention relates to a method for estimating a target variable using a regression model. In this method, measured values are used as time series data (ST1), with the time series data (ST1) influencing a state of an object. During data processing (ST2-ST7), the time series data for use in the regression model (ST8, ST10) are prepared as follows: In a first step (ST1), the time series data are regularly determined and are transferred as data to a subsequent step (ST2). In the subsequent step (ST2), sections are identified on the basis of the data and the time series data (ST3) divided into sections are transferred to a subsequent step (ST4). In the subsequent step (ST4), the time series data are aggregated per section to form event features, preferably on the basis of statistical methods. These (ST5) are transferred to a subsequent step (ST6). In the subsequent step (ST6), statistical clustering is carried out on the basis of the event features in order to determine a grouping of the sections. In a seventh step, a resulting absolute frequency of the sections in the formed groups (ST7) is determined and is used in a subsequent step (ST8) to train a selected regression model. During the training of the selected regression model, additionally measured target variables (ST9) are used, which describe the state of the object. In a final step (ST10), the trained regression model is used to determine a target variable by means of estimation.
Need to check novelty before this filing date? Find Prior Art

Description

[0001] 202504234

[0002] 1

[0003] Description

[0004] Methods for estimating a target variable

[0005] The invention relates to a method for estimating a rarely measured target variable using a regression model, which specifically addresses the characteristic mismatch between continuously recorded time series data and rarely measurable target variables.

[0006] Introduction and State of the Art

[0007] It is known to use a regression model as a statistical method for estimating the wear and tear of a rail vehicle or a rail vehicle component.

[0008] This can involve an assessment of the wear and tear of a pantograph, a brake pad, a wheelset, a complete rail vehicle, etc.

[0009] Creating such a regression model becomes difficult if the regression model is based on the following specifications:

[0010] on target variables that are measured very rarely and are therefore available in significantly smaller numbers compared to time series data, and

[0011] based on time series data that are constantly measured during the operational use of the rail vehicle and are therefore available in a large to very large number.

[0012] For understanding

[0013] The target variable describes a specific condition or wear state of a considered object (e.g., the pantograph, the brake pad, the wheelset, the rail vehicle, etc.), and

[0014] The time series data, as measured values, describe a specific operation that influences the condition or wear state of the object under consideration (e.g., vehicle speed, track condition, environmental influences (rain, snow, ice), overhead line condition, number of shipments carried out, number of braking maneuvers, etc.).

[0015] The described discrepancy between the number of target variables and the number of time series data points poses a challenge for conventional regression approaches, which aim to create a regression model that fits the situation.

[0016] 2

[0017] The basic steps of a regression approach are:

[0018] - the collection of relevant data, where the data includes independent and dependent variables,

[0019] - preparing the collected data, for example, cleaning or transforming it for a planned analysis,

[0020] - selecting a regression model (e.g., for linear, logistic, multiple, polynomial regression),

[0021] - estimating parameters of the regression model, e.g. using statistical techniques,

[0022] - the evaluation of the regression model by checking its model fit and validity using tests and metrics, and

[0023] - the interpretation of the results obtained in order to understand relationships between the data.

[0024] Previous approaches to wear estimation or wear modeling either ignored sparsely given target values ​​in order to build the regression model solely on the available time series data, or used methods (such as interpolation or simulation) to supplement missing target variables needed for modeling.

[0025] Both approaches have disadvantages in terms of accuracy and interpretability of the regression models used.

[0026] Task

[0027] It is therefore the object of the invention described below to provide an improved method for estimating a rarely measured target variable, which is to be used in a regression model.

[0028] This problem is solved by the features of claim 1. Advantageous further developments are specified in the dependent claims.

[0029] Description of the invention 202504234

[0030] 3

[0031] The invention relates to a method for estimating a target variable using a regression model. In this method, measured values ​​are used as time series data, wherein the time series data influences the state of an object. Using data processing, the time series data are prepared for use in a selected regression model as follows:

[0032] In a first step, the time series data are regularly determined and passed on as data to a subsequent step.

[0033] In this step, sections are identified based on the data, and the time series data divided into sections is passed to a subsequent step.

[0034] In this step, the time series data for each section is aggregated into event characteristics. The resulting data is then passed to a subsequent step.

[0035] In this step, clustering is performed based on the event characteristics to determine a grouping of the sections. The resulting data is then passed to a subsequent step.

[0036] In this step, a resulting absolute frequency of the sections within the formed groups is determined. The data thus generated is then passed to a subsequent step.

[0037] In this step, the data is used to train a selected regression model. During the training of the selected regression model, additional measured target variables describing the state of the object will be used.

[0038] In a final step, the trained regression model is used to determine one or the target variable by estimation.

[0039] Advantages:

[0040] In the present invention, a grouping of sections is carried out in order to effectively integrate information from the time series data into the regression model.

[0041] The present invention makes it possible to create a powerful and interpretable regression model even with a small number of target values. 202504234

[0042] 4

[0043] The present invention implements a regression model that enables predictions with high accuracy, allowing conclusions to be drawn about the influence of various event characteristics on the target variable.

[0044] The method according to the invention reconciles the large number of available input data with the small number of target variables.

[0045] The method according to the invention reduces a large number of data points in the input data.

[0046] Character description:

[0047] The invention is explained in more detail below by way of example and with the aid of a drawing. The drawing shows:

[0048] FIG 1 shows a first flowchart for the method according to the invention, and

[0049] FIG 2 shows a second flowchart for the method according to the invention.

[0050] FIG 1 shows, as the first flowchart for the method according to the invention, a training of a regression model, which in this example is based on an “Artificial Intelligence, Cl”.

[0051] Inputs and outputs are represented as parallelograms, while process steps are represented as rectangles.

[0052] The parallelograms and rectangles are described below as blocks ST1 to ST9.

[0053] The first block ST1 contains time series data, which are regularly measured, for example, in the operational use of a rail vehicle and are available in a large to very large number.

[0054] For example, to analyze the wear of a brake pad, time series data are collected that describe the temporal progression and / or intensity, etc., of braking processes during the operational use of the rail vehicle (e.g., the speed of the rail vehicle, the brake pressure, etc.). 202504234

[0055] 5

[0056] In the second block ST2, sections for the time series data of block ST1 are identified according to the statistics.

[0057] In statistics, "identifying sections" means recognizing and delimiting specific sections or parts within a dataset. These sections can represent different groups, categories, or areas relevant to the analysis.

[0058] For example, to analyze the wear of the brake pad, the time series data is divided into sections, with the sections describing individual braking processes.

[0059] A braking process is identified, for example, by the features "brake continuously active" and "vehicle is moving".

[0060] The data thus obtained and divided into sections form the third block ST3 and are fed into the fourth block ST4.

[0061] In the fourth block, ST4, the data undergoes an aggregation of event characteristics within the sections, according to statistical principles.

[0062] In this process, specific event characteristics within the sections are summarized. This is done, for example, by summing, integrating, averaging, or similar methods to obtain properties and characteristics as event characteristics for the individual sections.

[0063] In the present example, the braking duration, the average brake pressure, the braking time, etc., are determined as event characteristics for the wear analysis of the brake pad.

[0064] The data obtained in this way form the fifth block ST5 and are fed into the sixth block ST6.

[0065] In the sixth block, ST6, the sections are clustered according to statistics in order to form groups.

[0066] In statistics, clustering is a method of dividing data into groups that share similar properties or characteristics. This method is used to identify and analyze patterns in the data.

[0067] 6

[0068] This involves grouping different sections or parts of the data so that elements within a group are as similar as possible, while the groups themselves are as different as possible from each other. This helps to better understand the structure of the data based on an assigned absolute frequency and to extract relevant information.

[0069] In the present example, the braking events are grouped according to their braking duration and / or average brake pressure and / or braking time, etc.

[0070] The following characteristics could be considered as examples for the groups: Group 1: High braking duration + high average brake pressure + high braking time; Group 2: High braking duration + high average brake pressure + low braking time; Group 3: High braking duration + low average brake pressure + high braking time; Group 4: High braking duration + low average brake pressure + low braking time; Group 5: Low braking duration + high average brake pressure + high braking time; Group 6: Low braking duration + high average brake pressure + low braking time; Group 7: Low braking duration + low average brake pressure + high braking time; Group 8: Low braking duration + low average brake pressure + low braking time

[0071] The symbol “+” here represents an “and” combination of the mentioned characteristics.

[0072] In summary, for all time series data belonging to the same target value, a distribution of sections into determined groups results.

[0073] The data obtained in this way represent an absolute frequency of the sections in the individual groups and form the seventh block ST7. The data are then added to the eighth block ST8.

[0074] In the eighth block, ST8, a selected regression model is trained using artificial intelligence (AI). For this purpose, data from the ninth block, ST9, is also fed into the eighth block, ST8.

[0075] The ninth block, ST9, contains target variables that are rarely measured in rail vehicles and are therefore present in relatively small numbers in the time series data. 202504234

[0076] 7

[0077] For example, the thickness of the brake pad is measured to assess its wear. This is usually done during routine maintenance of the rail vehicle in a workshop.

[0078] In summary, for the regression model training (see block ST8), a distribution of data across sections in groups is used as input data (see block ST7) - for example, in the form of "Group A occurred 400 times".

[0079] FIG 2, with reference to FIG 1, shows as a second flowchart for the method according to the invention an application of the selected regression model, which is also based on an “Artificial Intelligence, AI”.

[0080] In a tenth block ST10, the trained regression model is applied and in an eleventh block ST11 delivers a prediction for the target variable - here for the thickness of the brake pad or for the associated wear.

Claims

202504234 8 Patent claims 1. Methods for estimating a target variable using a regression model, - where measured values ​​are used as time series data (ST1) that influence the state of an object, - where, with the help of data processing (ST2 - ST7), the time series data are prepared for use in a selected regression model (ST8, ST10) as follows: - in which, in a first step (ST1), the time series data are regularly determined and passed as data to a subsequent step (ST2), - in which, in this step (ST2), sections are identified based on the data and the time series data divided into sections (ST3) are passed to a subsequent step (ST4), - in which, in this step (ST4), the time series data for each section is aggregated to create event characteristics, and the data thus generated is passed to a subsequent step (ST6), - in which, in this step (ST6), clustering is carried out based on the event characteristics to determine a grouping of the sections, and the data thus formed are passed to a subsequent step (ST7), - in which, in this step (ST7), a resulting absolute frequency of the sections in the formed groups is determined and the data thus formed are passed to a subsequent step (ST8), - in which, in this step (ST8), the data are used to train a selected regression model, - in which additionally measured target variables (ST9) are used during the training of the selected regression model, which describe the state of the object, and - in which, in a final step (ST10), the trained regression model is used to determine a target variable by estimation.