Machine learning in production: application fields and free-access data records

© Fraunhofer IPT

It is often due to a lack of experience in dealing with machine learning itself, when machine learning projects cannot be successfully completed even with sufficient amounts of data. Missing or unstructured data make it difficult to train models and thus slow down the gain of experience in dealing with machine learning algorithms.

Here, freely available data files can help to gather initial experience and test own machine learning approaches. At present there are not yet many such publicly accessible data sets in the area of production, often due to existing confidentiality obligations, which are also stored on different platforms such as kaggle, ucirvine or openml.

The Fraunhofer IPT has therefore compiled the publicly available data records collected from production in a clearly arranged table on the basis of extensive investigations: 38 data records are currently available in this way. Since the list is maintained and completed by the Aachen scientists, there will be an increasing number soon. The data sets are assigned to seven application areas, which were also worked out at the Fraunhofer IPT.

The first column names the name of the record. The following columns contain the information when the data set was created or last updated and which learning task (e.g. C = Classification or R = Regression) can be expected for the given application field. It also provides information about the number of instances and attributes of the data set. A record that consists of several files is marked accordingly.

When using machine learning in production, it is important to have a sufficient quantity of the parameter to be determined present in the historical data, e.g. the amount of defective parts if rejects are to be predicted. A trained model is then able to reliably predict the conditions it has learned from the historical data. Since an underrepresented class, such as the number of product defects, must occur in sufficient numbers, an additional column with this information has been added to the table.

The table was created to provide a background of data sets which is a powerful tool aiming the growth of machine learning algorithms utilized for companies to accomplish better results. The table has data sets from several fields, in this way, a broad number of option appear. Besides that, there are categories (A - G) to fit the data set into certain classes or modalities (see below) from Machine Learning. The Dataset can fit into more than one class, and it is represented by the circles below.

  • A – Design
  • B – Optimization of Routing and Scheduling
  • C – Predictive Control
  • D – Self Learning Machines and Assets
  • E –  Anomaly Detection
  • F – Predictive Maintenance
  • G – Product Design
Completely fits: ++
Regurlarly fits: +
Does not fit: o
  Description
Donate Date Learning Task Number of Instances Number of Attributes Minor Class a b c d e f g Web link
3D Printer The aim of the study is to determine how much of the adjustment parameters in 3D printers affect the print quality, accuracy, and strength. There are nine setting parameters and three measured output parameters. 22.09.2018 Regression 50 12 25 + o ++ o o o o Kaggle /
3D-Printer
Mercedes-Benz Greener Manufacturing In this competition, Daimler challenged Kagglers to tackle the curse of dimensionality and reduce the time that cars spend on the test bench. This data set contains an anonymized set of variables, each representing a custom feature in a Mercedes car. 2016 Regression 4,210 378
23
+ +
+ o
+ o
o
Kaggle / Mercedes-benz-manufact.
APS Failure at Scania Trucks This set contains data from heavy Scania trucks in daily usage. The system in focus is the Air Pressure system (APS), which generates pressurized air used in various functions, such as braking and gear shifting. 01.02.2018 Classification 60,000 171
1,000 + o ++ o
+ ++ o Kaggle /
Scania-trucks
SECOM The data was collected from a semiconductor manufacturing process. It represents a selection of features, in which each example represents a single production entity with associated measured features. 19.11.2008 Classification 1,567 591
104 + o ++ o
++ o o Archive.ics.uci / SECOM
Cylinder Bands Process delays known as cylinder banding in rotogravure printing were substantially mitigated using control rules discovered by decision tree induction. ML shows to be promising for knowledge acquisition. 01.08.1995 Classification 512 40
200
+ o ++ o o o o Archive.ics.uci / Cylinder-Bands
Bosch Production Line Performance The data for this competition represents measurements of parts as they move through Bosch's production lines. Each part has a unique ID. The goal is to predict which parts will fail in quality control. 2016 Classification 1,183,747 2   + o ++ o + o o Kaggle /
Bosch-production-line
Quality Prediction Mining Process Data from a mining plant. The goal is to predict how much impurity is in the ore concentrate that is measured every hour. 2017 Regression 734,000 24 12,269 o o ++ o o o o Kaggle /
quality-prediction-in-a-mining-process
Energy Optimization This data was collected from a demonstrator of a high storage system, which transported one package between two spots. The high storage system consists of 4 short conveyor belts and 2 rails. 01.07.2018 Classification,
Regression
4 files;
à 20,000
20 10,200 o o ++ o ++ o o Kaggle / Energy-Optimization
Production Plant Data for Condition Monitoring Data for 8 run-to-failure experiments were provided and 8 features related to the component were selected. Training and prediction data were selected using the leave-one-out method: data under test were selected as the target for the prediction. 01.09.2018 Classification,
Regression
8 files;
à 20,000 inst.
26 15,800 o o ++ o + ++ o Kaggle / Monitoring
CNC Mill Tool Wear Machining data was collected from a CNC machine for variations of tool condition, feed rate, and clamping pressure. 01.04.2018 Classification 18 files;
à 500 inst.
48 2,304 o o ++ o ++ + o Kaggle /
CNC-mill
Bolts Data from an experiment, which analyzes the effects of machine adjustments on the time to count bolts. Bolts are dumped into a large metal dish. A plate that forms the bottom of the dish rotates counterclockwise. 04.10.2014 Classification,
Regression
40 8 14 o o ++ o + o o Openml.org /
857
Milling The data was collected from experiments on a milling machine for different speeds, feeds, and depth of cut. Additionally, data from the wear of the milling process is acquired. 2007 Regression 167 13 59 o o + o ++ + o Nasa.gov /
Prognostic-datarepository
Li-ion Battery Aging This data set has been collected from a custom built battery prognostics testbed. The aim is to be able to manage this uncertainty of actual usage and make reliable predictions of Remaining Useful Life. 01.10.2008 Regression 2,167 12 636 o o ++ o o ++ o Nasa.gov / Resources133
Airfoil Self-Noise The NASA data set comprises different size NACA 0012 airfoils at various wind tunnel speeds and angles of attack. The goal is to predict sound pressure levels. 04.03.2014 Regression 1,503 6 36 o o o
o o o ++ Archive.ics.uci/ Airfoil-self-noise
CFRP Composites Run-to-failure experiments were run on CFRP panels with periodic measurements to capture internal damage growth under tension-tension fatigue. 2008 Classification 3 files;
à 4 Layouts;
à 150 inst.
7 316 o o o
o o o ++
Nasa.gov /
Prognostic-datarepository
Mechanical Analysis Fault diagnosis problems of electromechanical devices. Each instance contains many components, each one has eight attributes. Different instances in this database have different numbers of components. 01.06.1990 Classification 209 8 5 o o + o ++ o o Archive.ics.uci / Mechanical-Analysis
Versatile Production Data from Versatile Production System (VPS) for a wide variety of tasks, including model learning, anomaly detection, and alarm management. 01.09.2018 Classification 8 files;
à 10,000 inst.
6 65 o o + o
++ o o Kaggle / Versatile-Production
Steel Plates Faults A data set of steel plates faults, classified into seven different types. The goal was to train machine learning for automatic pattern recognition. 01.11.2017 Classification 1,941 34 55 o o + o ++ o o Kaggle /
Steel-plates
Bearing Four bearings were installed on a shaft. The rotation speed was kept constant at 2,000 RPM by an AC motor coupled to the shaft via rub belts. Three data sets are included in the data packet. Each data set describes a test-to-failure experiment. 2007 Regression 3 files;
à 2,156 / 984 / 4,448 inst.
8 / 4 / 4 984 o o + o + o o Nasa.gov / Prognostic-data-repository
Plant Fault Detection PHM Data Challenge 2015: Fault detection and prognostics, a common problem in industrial plant monitoring. The final aim is the ability to detect plant faults. 05.06.2015 Regression 70 files;
à 127,691 inst.
10 700 o o + o ++ + o Phmsociety / Competition15
Robot Execution Failures This data set contains force and torque measurements on a robot after failure detection. All features are numeric although they are integer valued only. 23.04.1999 Classification 5 files;
à 88 / 47 / 47 / 117 / 164 inst
90 3 o o o + ++ + o Archive.ics.uci / Robot-Execution-Failures
Turbofan Engine Degradation Simulation The data was extracted from an engine, which is operating normally at the start of each time series until a fault occurs. The objective of the competition is to predict the number of remaining operational cycles before failure. 22.09.2010 Regression 4 files;
à 20,000 inst.
26 76 o o o o ++ + o Nasa.gov /
Resources139
Gearbox Fault Detection PHM Data Challenge 2009: Fault detection and magnitude estimation for a generic gearbox using accelerometer data and information about bearing geometry. 02.11.2017 Regression 560 files;
à 133,000 inst.
3 65,000 o o o o ++ + o Nasa.gov / Resources997
Anemometer Fault Detection PHM Data Challenge 2011: Anemometer fault detection, a critical problem for the wind power industry, strongly affecting among other things the financing of a potential site. 03.05.2011 Regression 420 files;
à 720 inst.
16 63,000 o o o o ++ o o Phmsociety / Competition11
Maintenance Action Recommendation PHM Data Challenge 2013: Maintenance action recommendation, which is a common problem in industrial remote monitoring and diagnostics. 2013 Regression 1,200,000 32 10,461 o o o o ++ o o Phmsociety / Competition13
Asset Health Condition PHM Data Challenge 2014: Asset health calculation that is a common problem in industrial remote monitoring and diagnostics. 05.10.2014 Regression 270,831 4 9,200 o o o o ++ + o Phmsociety / Competition14
Genesis Demonstrator The Genesis Demonstrator is a portable pick-and-place demonstrator, which uses an air tank to supply gripping and storage units. The data from the whole process is acquired. 01.07.2018 Regression 5 files;
à (3x) 7,500 inst.
(2x) 16,000 inst.
24 424 o o o o ++ o
o Kaggle /
Genesis-Demonstrator
Maintenance of Naval Propulsion Plants Data has been generated from a sophisticated simulator of Gas Turbines (GT), mounted on a Frigate characterized by a Combined Diesel Electric and Gas (CODLAG) propulsion plant. 11.09.2014 Regression 11,934 18
460
o o o o o ++ o Archive.ics.uci /
Naval-Plants
Azure Blob Each machine includes a device, which stores data such as warnings, problems and errors generated by the machine over time. 13.06.2017 Classification 2,000,000 172 159,150 o o o o o ++ o github / Azure-Predictive-Maintenance
Predictive Maintenance The data set is in kind of time series, consisting of the log message and failure records of 984 days. The goal is to predict machine failure in advance. 01.09.2018 Classification,
Regression
984 2 98 o o o o o ++ o DeepLearning / Predictive-Maintenance
Aircraft Engine The engine is operating normally at the start of each time series, and starts to degrade at some point during the series. 2008 Classification,
Regression
3 files;
à 45,000 inst.
26 105 o o o o o ++ o Nasa / PHM 08 Aircraft Engine
Semiconductor CMP PHM Data Challenge 2016: the challenge is focused on tracking the health state of components within a wafer chemical-mechanical planarization (polishing) system. 2016 Regression 2 folders; à 184 files;
à 1,300 inst.
26 815 o o o o o ++ o PHM / Semiconductor CMP
Condition monitoring of hydraulic systems The data set addresses the condition assessment of a hydraulic test rig based on multi-sensor data. Four fault types are superimposed with several severity grades impeding selective quantification. 2018 Classification,
Regression
2205 43680 756 o o o o o ++ o UCI / Hydraulic Monitoring

Contact Press / Media

Jonathan Krauß M.Sc.

Gruppenleitung

Fraunhofer-Institut für Produktionstechnologie IPT
Steinbachstr. 17
52074 Aachen

Telefon +49 241 8904-475

Fax +49 241 8904-6475