paper implements a 1 Dimension-Convolution Neural Networks-Long Short-Term Memory (1D-CNN-LSTM)

Auto-Encoder model for fault diagnosis of bearing data. The 1D-CNN-LSTM Auto-Encoder model showed

high accuracy of 58 to 100 percent for eccentric bearing data that are difficult to visually diagnose as faults. In

the future, we would like to extend this to a real-time failure diagnosis system that can remotely monitor the

condition of the bearing equipment through real-time communication with the cloud server and test bed.

Key-Words: - 1D-CNN, LSTM, Auto-Encoder, Unsupervised Anomaly Detection, Bearing Data, Smart Factory

Received: March 29, 2022. Revised: October 25, 2022. Accepted: November 27, 2022. Published: January 9, 2023.

increasing. If the bearing device fails, the product

manufacturing process will be disrupted, which can

lead to huge losses for the company. If the bearing

equipment is predicted to fail and the equipment is

replaced in advance, it will be a great advantage in

industrial and economic aspects by increasing the

production efficiency of the company.

whether to replace the product by listening to the

sound of the equipment, or by judging whether the

changes in industrial sites, including an increase in

non-face-to-face work and a decrease in field

technicians. It is time for an AI algorithm model to

diagnose failures with objective criteria, not failure

diagnosis of equipment using the know-how of

skilled technicians.

Encoder model. Section 4 details the test bed used,

the bearing data experimental environment, data

extraction and processing methods, and

experimental results. In Section 5, the conclusion is

made based on the results obtained in Section 4.

deep learning algorithm, specialized in processing

data arranged in a grid shape, and is an effective

neural network for identifying patterns of data.

Therefore, CNN utilizes several filters that can be

used as shared parameters to maintain spatial

information of images in two dimensions and

effectively extract and learn features with adjacent

images. CNN has the advantage of enabling simpler

learning through minimal parameters and

preprocessing. Among them, 1D-CNN (1

LSTM, [6], stands for Long Short-Term Memory

and is a model generated to address the long-term

dependence of Recurrent Neural Networks (RNNs)

used for learning time series data. LSTM has the

advantage of storing and utilizing information on all

input data, so it is widely applied to time series data

processing. LSTM is used to solve the vanishing

gradient problem and is advantageous for long time

preprocessing. The simultaneous use of 1D-CNN

Fig. 1: Undercomplete Auto-Encoder model

addition, by applying the 1D-CNN layer, learning

was configured to proceed while moving the

timestamp and feature information in detail. The

structure of the model is designed so that the

encoder and decoder are symmetrical with the 1D-

threshold value and compared with the loss value of

the test data, and then the data was determined

whether it was normal or abnormal. Experiments

kernel size of 1D-CNN and the unit values of Dense

and LSTM.

Auto-Encoder model are shown in Table 1.

Experiments were conducted with various

weighted, when rotating plate B is weighted, and

when rotating plate A and B are simultaneously

weighted.

with normal data, so the angle of the two screws is

180 degrees, that is, the case where there is no

eccentricity. The test data was set when the angles

of the two screws were 160 degrees and 170

degrees.

accurately evaluates the performance of the model

when the data labels are unbalanced.

The 1D-CNN-LSTM Auto-Encoder model was

applied to bearing data extracted from the Sewoo

Industrial System test bed, and the accuracy shown

in Table 2 was derived. Rotating plate A diagnosed

the failure for most of the situations except for 180

degrees data of the angles of the two screws. In

particular, when there are simultaneous eccentric

data of rotating body B and rotating body (A and B)

at the same time, failure was diagnosed in all

situations.

of the two screws on the rotating body A is set to

180 degrees data as train and 160 degrees data of the

two screws as a test on the rotating body A. The

result of rotating body A is not as good as that of

rotating body B or rotating body (A and B) at the

same time because the distance between the

acceleration sensor and rotating body A is relatively

too close to that of rotating body B to detect a

CNN-LSTM Auto-Encoder model proposed in this

paper showed 58 to 100 percent accuracy. It is not

easy to obtain fault data in the actual field, and the

model proposed in this paper is an Unsupervised

model, which has the advantage of being able to

learn only with a normal sample. Failure of bearing

data may occur in a misalignment-like manner,

except for eccentricity. It may be set as an additional

diagnostic failure evaluation element for the

misalignment. In addition, the current experimental

“This research was supported by the National

Research Foundation of Korea (NRF) grant funded

by the Korea government (MSIT) (No.

2021R1F1A1060054), the MSIT (Ministry of

Science and ICT), Korea, under the ITRC

(Information Technology Research Center) support

program (IITP-2022-2018-0-01417) and the ITC

Hyunseung Choo and Jongpil Jeong.

[1] M. Dix, A. Chouhan, S. Ganguly, S. Pradhan, D.

[4] B. Hou, J. Yang, P. Wang, and R. Yan, “LSTM

[5] Eren, L., Ince, T, and Kiranyaz, S, “A Generic