This is the code and dataset repository for Smart-Manufacturing (SM) Anomaly Detection (AD) that supports a paper in preparation:

"Anomaly detection in Smart-manufacturing era: A review"

Faults and unscheduled stops resulting from machine breakdowns in manufacturing processes have a substantial impact on both global direct and indirect production costs, as well as compliance with established delivery deadlines. With the continuous expansion of smart manufacturing environments, where shared data are readily available in real time, the implementation of anomaly detection strategies has become increasingly necessary. However, this task has also become more complex due to the multitude of diverse scenarios and methods that need to be considered. The paper presents a comprehensive state-of-the-art review, supported by experiments conducted on a repository of real manufacturing datasets included here. We provide a practical classification framework specifically tailored to smart manufacturing environments, which describes recent and successful methods and algorithms that have been applied to real scenarios. Furthermore, we introduce an experimental evaluation by executing state-of-the-art anomaly detection algorithms, aimed at gaining practical insights into the most suitable methods for various manufacturing environments.

SmartManuAD is an effort of the Department of Statistics, Informatics and Mathematics of the Public University of Navarra (UPNA), in a collaborative work between Miguel Pagola, PhD professor in Computer Science and Artificial Intelligence at the UPNA, and Iñaki Elía, PhD student in Artificial Intelligence as well as Robotics & Automation Engineer for different Aerospace companies.

SM is a specific application of the Industrial Internet of Things (IIoT) that utilizes embedded sensors to collect operational data and integrate it with physical and digital processes within factories. Through the analysis of data streams across entire factories or multiple facilities, manufacturing engineers and data analysts can, among other purposes, identify faults or anomalies in different processes, diagnose their potential causes, assess the health of systems or prognose the future state of health based on their current condition.

As shown in the following image, we established 5 categories related to SM concepts and 6 categories associated with common manufacturing processes or systems. Our intention was to map publications and datasets to a reformulated classification that is more focused on the emerging SM scenarios.

Our research evaluates 16 state-of-the-art AD algorithms on 29 publicly available real-world datasets in the field of Smart-Manufacturing. These datasets were selected from a pool of 100 publicly available ones, covering various industrial domains, such as additive and subtractive manufacturing, machinery and equipment, automotive, aerospace, rail, energy, logistics, chemicals, electronics, semiconductors, and multi-stage mass manufacturing, among others. The 29 chosen datasets contain real anomalies, which were extrapolated from the information provided by the sources, processed, and curated to create balanced datasets for the analysis.

As an example of the results obtained, see the following figure that shows the AUC-ROC and AUC-PR metrics for the Motors & Drivetrains category and for the 16 selected algorithms.

We use important anomaly detection libraries for our study, including anomaly detection for tabular (PyOD) and time-series data (TODS). We also use ADBench as the basis for our work with the intention of extending the benchmark of AD algorithms to the SM domain. Please refer to the links above for more information.

The code used from ADBench is written in Python 3 (see related dependencies). We simply tuned a notebook ("SmartManuAD_Test_Notebook.ipynb") to help with the specifics of Smart-Manufacturing environments.

You can find below the list of 29 publicly SM datasets used in our research. They include real anomalies. In the datasets folder you can find the expanded list of 100 collected datasets, though most of them do not include real anomalies, but we injected 10% synthetic ones. Both lists include links to the original dataset sources so that you can get the original metadata for the datasets.

The datasets are processed and converted to the Numpy .npz file format, which is a compressed archive of files. This also allows you to work with some of the utilities and code introduced in ADBench.