Data-driven Exponential Framing for Pulsive Temporal Patterns without Repetition or Singularity

Yohei Kono, Yoshiyuki Tajima · 2025

Read Data-driven Exponential Framing for Pulsive Temporal Patterns without Repetition or Singularity by Yohei Kono, Yoshiyuki Tajima (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2510-22472.txt.

Introduction

Yohei Kono and co-author Yoshiyuki Tajima present a method for finding brief pulse like patterns in data too small for the patterns to repeat. Such one off patterns matter in manufacturing but draw little study, since most tools need repetition or a single sharp spike. Their method, Data driven Exponential Framing, measures how long a pattern lasts without leaning on either clue.

A linear model on time delayed data yields basic shapes, each fading at its own rate. Sliding windows are then scored against those shapes to find where each pattern rules. Tests on a toy model and on punching machine current data found several patterns.

Catalog record

CREATORYohei Kono, Yoshiyuki Tajima
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE62 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2510-22472

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Data-driven Exponential Framing for Pulsive Temporal Patterns without Repetition or Singularity Yohei Kono, Yoshiyuki Tajima Abstract Extracting pulsive temporal patterns from a small dataset without their repetition or singularity shows significant importance in manufacturing applications but does not sufficiently attract scientific attention. We propose to quantify how long temporal patterns appear without relying on their repetition or singularity, enabling to extract such temporal patterns from a small dataset. Inspired by the celebrated time delay embedding and data-driven Hankel matrix analysis, we introduce a linear dynamical system model on the time-delay coordinates behind the data to derive the discrete-time bases each of which has a distinct exponential decay constant. The derived bases are fitted onto subsequences that are extracted with a sliding window in order to quantify how long patterns are dominant in the set of subsequences. We call the quantification method Data-driven Exponential Framing (DEF). A toy model-based experiment shows that DEF can identify multiple patterns with distinct lengths. DEF is also applied to electric current measurement on a punching machine, showing its possibility to extract multiple patterns from real-world oscillatory data. 1 Introduction Temporal patterns, meaning subsequences which are in some sense representative of their original time-series data, play a crucial role in various research areas [ 1 , 2 , 3 , 4 , 5 , 6 ] . This paper suggests an alternate perspective on continuous-valued temporal patterns that is more suitable for manufacturing applications. In manufacturing processes, the development of sensing and information technologies has accelerated data-driven enhancement of their operation and maintenance. This trend motivates the use of temporal patterns for monitoring complex phenomena that cannot be described in closed-form models. For instance, in the case of injection-molding machines, Takatsugi, et al. [ 7 ] reported that temporal patterns in load torque data can characterize in-situ resin pressure, and proposed detecting the reverse flow of resin by monitoring the data. Similarly, for punching machines, Takahashi, et al. [ 8 ] analyzed a temporal pattern in die acceleration data to monitor the tool wear. For human manufacturing, Terada, et al. [ 9 ] developed an unsupervised method for detecting a series of repetitive actions by recognizing basic actions from human motion data. We also developed an anomaly detection method based on temporal patterns and applied it to manufacturing machines [ 10 ] . Motivated by the above studies, we address data-driven extraction of temporal patterns specific to manufacturing processes. The time-series data and patterns addressed here are characterized by the three assumptions. Firstly, as in manufacturing machines operated with the position or speed control of their tools (e.g., Ref. [ 11 ] ) and chemical processes controlled on different levels of temperature reference, reference values (setpoints) for control of manufacturing processes can jump discretely, leading to process data containing rapidly rising (or falling) curves or irregular oscillations. These patterns are referred to as hereafter pulsive temporal patterns in this paper. Secondly, the manufacturing process has multiple modes of operation where the frequency of occurrence of each mode is not necessarily high. This means that the number of occurrence of each temporal pattern can be few in number, including one-shot occurrence. Thirdly, pulsive temporal patterns relevant in engineering practice are not restricted to anomalous events or singular behaviors. Many relevant cases, such as transient responses to external inputs or one-off operational changes, occur without singularities, implying the necessity to develop a framework that can address one-shot patterns, with or without singularity, beyond the conventional scope of anomaly detection. Therefore, this paper aims to extract pulsive temporal patterns from a small dataset that may lack repeated instances. Research on extracting pulsive temporal patterns has largely been conducted using machine learning techniques, specifically time-series data mining [ 12 ] . Different definitions of temporal patterns exist based on the machine learning task at hand. Ref. [ 13 ] defined previously unknown, frequently occurring patterns as motifs and developed a discovery algorithm for them. In the last two decades, many research groups have developed sophisticated algorithms for finding motifs [ 14 , 15 ] . Ref. [ 16 ] introduced time series discords , which are subsequences of a longer time series that are maximally different from all the rest, making it highly compatible with anomaly detection; see, e.g., [ 17 , 18 ] . Ref. [ 19 ] addressed the classification problem of time series and formulated a concept of time series shapelet , which are subsequences maximally representative of a class. Jointly learning shapelets and a classifier has also been reported in Refs. [ 20 , 21 ] . These studies suggest that previous definitions of temporal patterns rely on repetition or singularity. However, to the best of our knowledge, not enough attention has been given to identifying pulsive patterns in small datasets. This suggests that additional criteria, beyond repetition or singularity, are necessary to characterize temporal patterns. To quantify how long patterns appear in pulsive time series in an unsupervised manner, we propose an alternative view of the temporal patterns based on their length. Inspired by the celebrated time delay embedding [ 22 , 23 , 24 ] and data-driven Hankel matrix analysis [ 25 , 26 , 27 ] , we introduce a linear dynamical system model on the time-delay coordinates behind the data to derive the discrete-time bases, each of which has a distinct exponential decay constant. These bases are then fitted onto subsequences that are extracted with a sliding window to determine how long patterns are dominant in the set of subsequences. We call this quantification method Data-driven Exponential Framing (DEF). We demonstrate the effectiveness of DEF using artificial data generated by a toy model based on the harmonic oscillator. Our results show that DEF can identify multiple patterns with distinct lengths. In addition, we apply DEF to electric current data measured in a punching machine and show its effectiveness in this real-world scenario. Note that this paper is a joint submission with our conference paper [ 28 ] , and the discrete-time bases with exponential decay constants have been proposed in our non-archival conference paper [ 10 ] . The contributions of this paper are twofold. The first capability is the development of a practical method for characterizing small timeseries datasets. The method extracts dominant temporal patterns from a timeseries dataset without their repetition or singularity, which can be search queries for the timeseries data mining. The second contribution is the construction of an alternate framework for decomposing timeseries based on the exponential decay rate. This DEF framework enables the identification of multiple (complex) frequencies that explain pulsive patterns in timeseries data. In simpler terms, DEF offers an alternative interpretation of the modal components calculated by the Singular Spectrum Analysis [ 25 , 29 ] or the extensions of Dynamic Mode Decomposition (DMD) [ 30 ] to Hankel matrices [ 26 , 27 ] . The rest of this paper is organized as follows. Sec. 2 focuses on explaining the mathematical concept of DEF and its computation algorithm. In Sec. 3 , we provide evidence that DEF can identify the lengths of dominant subsequences in numerical data obtained from a toy model. Additionally, in Sec. 4 , we demonstrate that DEF can identify

Read the complete text