Overview of Publicly Available Degradation Data Sets for Tasks within Prognostics and Health Management

Technical Summary: Overview of Publicly Available Degradation Data Sets for Tasks within Prognostics and Health Management

Problem Statement
The efficacy of Prognostics and Health Management (PHM) methods relies heavily on the acquisition and analysis of degradation data, which encapsulates the evolving health condition of engineering systems (ESs). Despite the critical importance of such data for developing and benchmarking PHM methodologies, the availability of appropriate degradation data remains a significant challenge in both industrial practice and academic research. While publicly available datasets exist, the literature addressing them is fragmented; existing overviews often restrict their scope to specific platforms or applications, and analyses frequently prioritize data volume or quality over specific PHM-relevant aspects. Consequently, the search for suitable datasets is often time-consuming for researchers and practitioners.

Methodology
This work provides a comprehensive update and expansion of a previous overview of publicly available degradation datasets. The authors updated the scope defined in prior work (Mauthe et al., 2025; Hagmeyer et al., 2021) by considering an identical range of platforms and sources. The study excludes datasets focusing solely on process quality without degradation states.

The core of the methodology involves a detailed, PHM-specific analysis of 110 publicly available datasets. This analysis utilizes a taxonomy developed in previous work to classify datasets based on four specific dimensions:

1. Application: The specific object of consideration (e.g., a single component like a bearing or an entire system like a production line).
2. Domain: The originating field of the application (e.g., mechanical components, electrical components, drive technology).
3. Task: The specific PHM task the dataset addresses, categorized as:
   • Fault Detection/Anomaly Detection: Binary classification of fault vs. no fault.
   • Diagnosis: Assigning causes to detected faults.
   • Health Assessment: Quantifying current degradation or failure risk.
   • Prognosis: Predicting future degradation or Remaining Useful Life (RUL).
4. Signal: The types of signals contained in the dataset (e.g., vibration, current, temperature), distinguishing between directly measured signals and derived values.

Key Contributions and Results
The primary contribution is the extension of the publicly available dataset overview to 110 entries, accompanied by a structured classification that highlights the distribution of data across domains, applications, and tasks.

• Domain and Application Distribution: The 110 datasets are classified into 11 domains. The "Mechanical component" domain contains the most datasets (26), followed by "Electrical component" (25) and "Drive technology" (12). The remaining domains contain between two and eight datasets each. Five datasets are assigned to an "unknown" domain due to unspecified applications.

  • The most frequent applications are bearings (17 datasets) and batteries (16 datasets).
  • Other notable applications include aircraft engines, filtration systems, and production lines (5 datasets each).
  • The "Mechatronic system" domain contains the highest diversity of applications (eight), but each is represented by only one dataset. Overall, only nine applications are covered by three or more datasets.

• Task Distribution:

  • Prognosis is the most represented task, with 54 datasets. Within this category, the electrical component domain contributes 22 datasets (15 for batteries), while mechanical components and drive technology follow.
  • Diagnosis accounts for 35 datasets, with the mechanical component domain being the primary source (13 datasets). Bearings are the most frequent application for diagnosis.
  • Fault Detection is assigned to 18 datasets, primarily from mechanical component and production system domains.
  • Health Assessment is explicitly addressed by only three datasets, as the transition between diagnosis, health assessment, and prognosis is often not distinct in existing literature.

• Signal Analysis: The analysis identified 22 signals used in at least two datasets.

  • Vibration is the most prevalent signal (34 datasets), followed closely by current and temperature (33 datasets each), and voltage (25 datasets). These are dominant in electrical and mechanical component domains.
  • Speed (20) and pressure (18) are also common.
  • Signal availability varies by task: Prognosis datasets rely heavily on current and temperature, whereas diagnosis datasets show a dominance of vibration signals.
  • A limitation noted is that ten of the 22 signals appear in four or fewer datasets, mirroring the scarcity observed in specific applications.

Significance and Claims
The paper claims that its structured taxonomy and comprehensive overview make the current coverage of publicly available data explicit, highlighting where data is concentrated versus where coverage is sparse. The authors assert that while benchmarking is well-supported for a few frequently studied use cases (such as batteries and bearings) and signal modalities (vibration, current, temperature), broader generalization and task-specific evaluations remain limited by uneven coverage.

The work aims to serve as a resource to shorten the time-consuming search for suitable datasets for PHM users. The authors emphasize that maintaining and extending this overview is necessary to track the evolution of data coverage in the dynamic field of PHM and to identify where additional publicly available data would have the greatest impact. The paper does not propose new experimental methods or applications but rather structures existing knowledge to facilitate the development and empirical benchmarking of PHM methodologies.