In this paper, we discuss the current issues in time-series anomaly detection using deep learning and propose directions for their resolution. Instead of focusing solely on anomaly detection methods, we emphasize the importance of clarifying the problem setting from four perspectives: "system," "failure modes," "available data," and "operations." Additionally, as a potential future development, we suggest research directions that utilize tools such as Large Language Models (LLMs) against reliability design documentation to support operations and problem definition surrounding anomaly detection.

時系列異常検知における重要なパラメータにウインドウ長がある。そのウインドウ長のチューニングが不要な異常検知方式MADLIDと、MADLIDを利用した振動データの異常検知の評価について紹介する。

To automate the maintenance of equipment that requires high-frequency data for diagnostics, such as bearings and motors, a high-performance and interpretable anomaly prediction method is essential. However, detecting slight changes in waveforms, which indicate early signs of anomalies, is challenging due to noise interference. This paper proposes a method that combines the Shapelets learning technique, known for its clear decision evidence, with a band-pass filter. This combination helps detect slight waveform changes, capturing early signs of anomalies. An experiment show that this method can learn from a small amount of data and automatically identify frequency bands associated with anomalies.

Predictive maintenance is a technique to conserve maintenance and it is the key effective maintenance operations and reduced downtime. Methods for predictive maintenance based on anomaly detection using deep learning have been actively studied, but the identification of anomalous sensors remains a challenging task. In this work, we use a graph neural network for anomaly detection and estimation. The vertices of the graph correspond to the sensors, so we can interpret the relevant weights as the relationship between the sensors. We specifically used sparse graph to improve graph interpretability and we confirmed the effectiveness of the method.

The author has been developed the active structural health monitoring method using ultrasonic natural vibrations. In this paper, the structural health monitoring method using self-excited ultrasonic vibrations is introduced. First, the self-excitation method at natural frequencies is explained. Next, detection of contact-type failure based on nonlinear wave modulation is introduced. Finally, the prospects for development of structural health monitoring using neural network system is described.

we propose a novel deep learning approach using Unsupervised Domain Adaptation (UDA) that considers physical operation conditions for improving RUL prediction across different domains. Our method modifies the traditional Deep Adversarial Neural Network (DANN) structure by replacing the domain classifier with multiple classifiers. The method was developed and validated based on the new NASA dataset simulated by the Commercial Modular Aero-Propulsion system simulation (N-CMAPSS) with run-to-failure data under realistic flight conditions. The proposed model architecture is compared with the traditional DANN model to demonstrate the goodness of the results and the improvements.

本稿では、NECの量子インスパイアド技術であるベクトルアニーリング（VA）を活用した生産計画の自動化の取り組みの具体例を紹介するとともに、コンピュータを用いた最適化ソリューションの価値について説明する。

Recently, the demand for data analysis technologies to process large volumes of sensor time series data has grown due to the need for predictive maintenance and productivity improvement in smartmanufacturing. This paper addresses the challenges of processing sensor time series data and introduces the Spikelet approach, which uses value fluctuation range as a hyperparameter instead of time window size,as a potential solution tailored to the unique characteristics of sensor time series.