Fast Mining and Dynamic Time-to-Event Prediction over Multi-sensor Data Streams

Kota Nakamura, Koki Kawabata, Yasuko Matsubara, Yasushi Sakurai

Abstract

Given real-time sensor data streams obtained from machines, how can we continuously predict when a machine failure will occur? This work aims to continuously forecast the timing of future events by analyzing multi-sensor data streams. A key characteristic of real-world data streams is their dynamic nature, where the underlying patterns evolve over time. To address this, we present TimeCast, a dynamic prediction framework designed to adapt to these changes and provide accurate, real-time predictions of future event time. Our proposed method is dynamic, practical, and scalable, scaling linearly with input size and enabling online model updates.

Introduction

In predictive maintenance and clinical monitoring, the critical challenge is determining the remaining time before a significant event occurs. Traditional survival analysis often assumes a static relationship between sensor inputs and event timing. However, real-world systems are non-stationary; for instance, a jet engine progresses through distinct wear stages before failure. Existing methods often fail to capture these "stage shifts" or are too computationally intensive for true real-time streaming applications. TimeCast addresses these limitations by identifying time-evolving patterns and learning adaptive models for each stage.

Methodology: The TimeCast Framework

The core of TimeCast is a sequential multi-model structure. Rather than employing a single monolithic model, it decomposes the data stream into a sequence of "stage models."

The TimeCast sequential multi-model structure, illustrating the mapping from labeled collections to stage models consisting of descriptors and predictors.

As shown in the figure above, each stage model $\theta^{(k)}$ consists of two primary components:

  1. The Descriptor: This uses Gaussian Graphical Models (GGMs) to capture the interdependencies between different sensors. By analyzing sensor correlations, the model can identify which operational "stage" the system is currently in.
  2. The Predictor: This treats the path toward an event as a stochastic process (specifically a Wiener process). It calculates the "First Hitting Time" to estimate when the process is expected to cross a failure threshold.

To handle streaming data, TimeCast employs a dual-algorithm approach. A Learning Algorithm alternates between updating stage models via Graphical Lasso and optimizing stage assignments using Dynamic Programming. A Streaming Algorithm then uses a "generate-and-validate" strategy to predict the current stage and event probability in real-time, adding new stage models as novel patterns are detected.

Experimental Evaluation

The researchers evaluated TimeCast on five diverse, real-world datasets, including NASA’s C-MAPSS turbofan engine data, industrial machine logs, and three subsets of the MIMIC-III clinical database.

Key Results and Insights

  • Superior Accuracy: TimeCast consistently outperformed popular baselines like DeepSurv and DeepHit, achieving significantly lower Mean Absolute Percentage Error (MAPE) across all datasets.
  • Massive Speedup: In terms of computational efficiency, TimeCast outperformed competitors by up to four orders of magnitude in prediction time, making it uniquely suited for high-frequency IoT sensor streams.
  • Dynamic Adaptation: The model successfully adapts its predictions as the system evolves through different stages.
Snapshots of streaming time-to-event prediction at different time cycles ($t_c=135$ and $t_c=210$), showing how the predicted event probability (PDF) shifts as the system progresses through stages.

As illustrated in the snapshots above, the model continuously updates the Probability Density Function (PDF) of the event time. As the system moves from Stage #4 toward Stage #10, the prediction becomes increasingly certain and converges toward the actual event time.

Conclusion

TimeCast provides a scalable and interpretable solution for continuous time-to-event prediction. By combining graphical descriptors with stochastic predictors, it captures the non-stationary nature of multi-sensor streams while maintaining high computational efficiency. Future work includes exploring non-linear link functions and adapting the framework for systems that may oscillate between states rather than progressing linearly.

Summary updated

Abstract

Given real-time sensor data streams obtained from machines, how can we continuously predict when a machine failure will occur? This work aims to continuously forecast the timing of future events by analyzing multi-sensor data streams. A key characteristic of real-world data streams is their dynamic nature, where the underlying patterns evolve over time. To address this, we present TimeCast, a dynamic prediction framework designed to adapt to these changes and provide accurate, real-time predictions of future event time. Our proposed method has the following properties: (a) Dynamic: it identifies the distinct time-evolving patterns (i.e., stages) and learns individual models for each, enabling us to make adaptive predictions based on pattern shifts. (b) Practical: it finds meaningful stages that capture time-varying interdependencies between multiple sensors and improve prediction performance; (c) Scalable: our algorithm scales linearly with the input size and enables online model updates on data streams. Extensive experiments on real datasets demonstrate that TimeCast provides higher prediction accuracy than state-of-the-art methods while finding dynamic changes in data streams with a great reduction in computational time.
BibTeX
@inproceedings{nakamura2026fast,
  author = {Kota Nakamura and Koki Kawabata and Yasuko Matsubara and Yasushi Sakurai},
  title = {{Fast Mining and Dynamic Time-to-Event Prediction over Multi-sensor Data Streams}},
  booktitle = {ACM SIGKDD Conference on Knowledge Discovery and Data Mining},
  year = {2026},
  doi = {10.1145/3770854.3780164},
}
Figure