TCM: Benchmark Datasets for Predictive Maintenance in Steel Manufacturing

Anomaly-TCM

Predictive Maintenance (PdM) is a strategy that uses advanced data analytics to predict equipment failures and maintain industrial machinery in good condition. Its goals are to minimize downtime, reduce operational costs, and ensure product quality. PdM methods are applicable across various industries, including steel manufacturing.

In steel production, cold rolling is a critical process that reduces the thickness of hot-rolled steel. Developing PdM methods for tandem cold mills (TCM) can significantly improve production efficiency. However, researchers often rely on real manufacturing data, which is typically unavailable, unlabeled, and noisy, making it difficult to validate and compare methods.

To overcome this, we created synthetic datasets for the cold rolling process to identify anomalies based on physical principles. These datasets were generated using a mathematical model of a 5-stand TCM, calculating key process parameters like rolling force, torque, speed, tension, gap, thickness reduction, and motor power. We introduced anomalies related to specific failures in the process.

We produced six diverse datasets, each with varying complexity, to enable benchmarking of machine learning-based PdM methods for the cold rolling process. Four different types of anomalies were introduced, which are related to a physics-based deviations in the process:

Anomaly in reduction scheme

Anomaly in work roll (increased work roll friction)

Anomaly in bearing (increased motor torque)

Anomaly in electric motor (decrease efficiency)

 The details of the datasets are provided below.

Dataset
Observations
Anomalies
Share of Anomalies
Features
Anomaly Types
Products
Data Drift

tcm5_dataset_1
20009
1045
5.2%
51
1
4
FALSE

tcm5_dataset_2
20001
1035
5.2%
51
1
20
FALSE

tcm5_dataset_3
20003
981
4.9%
51
4 (16)
4
FALSE

tcm5_dataset_4
20001
925
4.6%
51
4 (16)
20
FALSE

tcm5_dataset_5
20005
1031
5.2%
51
4 (16)
5
TRUE

tcm5_dataset_6
20008
954
4.8%
51
4 (16)
25
TRUE

Each dataset is generated as a data stream, meaning the observations follow a chronological order, represented by increasing work roll mileage (which is reset after a predefined threshold). The table below provides details about the features and labels present in the datasets. Several features are recorded for each rolling stand, totaling 51 features. Apart from the anomaly related to reduction, the other anomalies are specific to individual stands, resulting in 16 anomaly labels in total.

Feature
Suffixes
Unit
Description

thickness_entry
-
mm
steel entry thickness

thickness_exit
-
mm
steel exit thickness

width
-
mm
steel width

ys_entry
-
MPa
steel entry yield strength

ys_exit
-
MPa
steel exit yield strength

work_roll_diam
1 to 5
mm
work roll diamaeter (stands 1 to 5)

work_roll_mileage
1 to 5
km
work roll mileage (stands 1 to 5)

reduction
1 to 5
-
thickness reduction (stands 1 to 5)

tension
0 to 5
N
interstand tension (0 is tension before stand 1, 1-5 refer to tension after stands 1-5)

roll_speed
1 to 5
NaN
linear work roll speed (stands 1 to 5)

force
1 to 5
N
rolling force (stands 1 to 5)

torque
1 to 5
Nm
rolling torque (stands 1 to 5)

gap
1 to 5
mm
stand gap (stands 1 to 5)

motor_power
1 to 5
kW
electric motor power (stands 1 to 5)

Anomaly_Reduction
-
-
(label) anomaly in reduction scheme

Anomaly_Electric
1 to 5
-
(label) anomaly in electric motor (stands 1 to 5)

Anomaly_Bearing
1 to 5
-
(label) anomaly in stand bearing (stands 1 to 5)

Anomaly_WorkRoll
1 to 5
-
(label) anomaly in work roll friction (stands 1 to 5)