Abstract

Reading the present and future from sensor data

With the spread of IoT devices, sensor data are collected across many fields—weather, space, ocean, infrastructure, manufacturing, healthcare, and more. Sensor data carry information about what is happening now and what is about to happen next in the observed target. We study sensor information processing that extracts this present and future information from large-scale data in real time. Present information helps with the early detection of problems, and future information helps with their prevention.

Concept: sensor information as present and future information
Fig. 1: Extracting present (early detection) and future (prevention) information from the sensor data of observation targets
Core Technology

General-purpose analysis of sensor time series — decomposition, forecasting, and anomaly detection

As challenges common to sensor information processing, we work on time-series decomposition, time-series forecasting, and anomaly detection (Term 1). Recently, we have focused on condensing large-scale stream data (Term 2) into a general-purpose summary (Term 3)—an ultra-lightweight in-memory structure—and performing fast exhaustive search on it. Low-frequency patterns can be regarded as signs of novelty, anomaly, or trend change. Using general-purpose summaries, we aim to estimate signs of present and future change in real time, even in remote monitoring and observation settings where cloud–edge communication latency is a bottleneck.

Concept of the general-purpose summary, an ultra-lightweight data structure
Fig. 2: Condensing large-scale streams into a general-purpose summary for real-time processing

References

  1. T. Phungtua-eng and Y. Yamamoto:
    Adaptive seasonal-trend decomposition for streaming time series data with transitions and fluctuations in seasonality. Proc of ECML-PKDD (2024)
  2. Y. Yamamoto and Y. Yamamoto:
    Constructing a sublinear summary answering support queries over transactional data streams. JSAI SIG-KBS (2023)
  3. Yoshitaka Yamamoto, Yasuo Tabei, Koji Iwanuma:
    Approximate-closed-itemset mining for streaming data under resource constraint (2019)

Funding

JSPS Grant-in-Aid for Scientific Research (C)
(PI) Development and feasibility study of fast, memory-efficient partial-order support summaries URL
(PI) Sublinear summarization of stream data based on partial-order relations URL

JST Feasibility Study (Startup)
(PI) Development of data-purification techniques to capture signs of anomalies from noisy time-series data URL

Application | Time-domain Astronomy

Detecting novel patterns through exhaustive search of light curves

We exhaustively search the light curves of numerous celestial objects and detect never-before-seen variation patterns in real time. We also work on detecting the light trails of nano-satellites (CubeSats). Targeting roughly 30 terabytes of data generated per night by Tomo-e Gozen, the wide-field high-speed camera at the Kiso Observatory (University of Tokyo), we aim to discover space debris and unknown transients.

Application to time-domain astronomy (Tomo-e Gozen, light curves, light-trail detection)
Fig. 3: Exhaustive search of light curves and real-time detection of novel variation patterns

References

  1. T. Phungtua-eng, N. Arima, Y. Yamamoto:
    TriLinear: Time series anomaly detection using tricube smoothing decomposition and a linear forecasting model. Proc of ADMA 2025, LNCS vol. 16198, pp. 19-33 (2026)
  2. T. Phungtua-eng, Y. Yamamoto, and S. Sako:
    Elastic data binning for transient pattern analysis in time-domain astrophysics. Proc of SIGSAC'23 (2023)

Funding

JSPS Grant-in-Aid for Scientific Research (A)
(Co-I) Survey of the Universe changing on a sec time scale by wide-field, high-cadence photometry and anomaly detection URL

Application | Smart Manufacturing

Work-process recognition and AI-driven skill-level feedback

For work on manufacturing sites, we study work-process sensing and process recognition (action segmentation) (Term 4). Based on the recognition results, we developed FIELDS (Feedback Integrated Expert Level Description System), a training-support system in which an AI model assesses proficiency and provides feedback. Through the AI model, the system helps workers become aware of their own current state and proactively improve their skill level. In collaboration with Yamaha Motor, we are validating its effectiveness on tasks that emulate vehicle-body assembly.

Application to smart manufacturing (process recognition and skill-level feedback with FIELDS)
Fig. 4: FIELDS recognizes work processes from work video and returns feedback on proficiency

References

  1. K. Nakamura, T. Ueyama, M. Nishimura, T. Nakano, T. Aoki, Y. Yamamoto:
    Cross-silo human training in operational assembly: Integrating machine feedback for enhanced efficiency. Computers & Industrial Engineering, Vol. 213, 111774 (2026)

Funding

Industry–Academia Collaboration
(PI) Shizuoka University–Yamaha Motor Collaborative Research Course, "Smart Factory Area" (2020–2025)
(PI) Quantification of skills requiring body motion and haptics for the transfer of expert techniques (joint research with Yamaha Motor)

JST A-STEP
Stage I (PI) Building a foundation for skill transfer through datafication of expert skills and AI feedback URL

Application | Wearable Health Monitoring

On-device monitoring of behavior and vitals with smart devices

We store biosignals from smart devices—accelerometer/gyroscope (IMU) (Term 5), heart rate, skin temperature, PPG, and others—as general-purpose summaries on the device (on-device). We predict regular daily behavior and vitals, detect deviations from them, and use these as a personal "behavioral fingerprint." On-device processing addresses the challenges of communication cost, privacy, and limited device resources.

Application to wearable health monitoring (on-device processing vs. the cloud)
Fig. 5: Condensing biosignals on-device to predict regular patterns and detect deviations

References

  1. J. Kutsuzawa and Y. Yamamoto:
    Towards On-Device Physiological Data Monitoring. Proc. of IEEE International Conference on Activity and Behavior Computing (2026)

Funding

Industry–Academia Collaboration
(PI) Study on physiological responses and intoxication in social settings involving eating and drinking (joint research with Sapporo Breweries)

Glossary

  1. Time-series decomposition, forecasting, and anomaly detection
    The three foundational tasks for handling sensor time series. Decomposition separates the data into components such as long-term trend and periodic (seasonal) variation. Forecasting estimates "what will happen next" from past observations, and anomaly detection captures "what unusual event is happening now."
  2. Stream data
    An unbounded sequence of data continuously generated by an observation system. Because the total volume grows rapidly over time, keeping all of the data is impractical, so "discarding data wisely" becomes essential.
  3. General-purpose summary (data summarization / sketching)
    An ultra-lightweight representation that characterizes the distribution of large-scale data. It enables real-time processing while keeping data in memory, and by searching over combinations of attribute conditions, low-frequency patterns can be captured as signs of novelty or anomaly.
  4. Process recognition (action segmentation)
    A technique that divides video or sensor time-series data into segments corresponding to meaningful actions (processes) and labels each one. It can automatically recognize multi-step tasks, such as assembly work, in terms of which process was performed and when.
  5. IMU (inertial measurement unit)
    A device that combines an accelerometer and a gyroscope to measure an object's motion (acceleration and angular velocity). It is built into wearables such as smartwatches and is used to estimate human behavior and posture.