Fault analysis on deep groove ball bearing using ResNet50 and AlexNet50 algorithms

Abstract

Deep Groove Ball Bearings (DGBBs) serve multipurpose and are used for the propeller shaft movement and applications based on revolving. They have great applications in industry related to axial and radial loads. The major risk factors are faults in bearings. Data analyzed for faults in the DGBBs help us conclude that there are 4 types of bearing faults. For instance, Excluding HB- Healthy Bearing, there are CF- Case Fault, BF- Ball Fault, IRF- Inner Ring Fault, and ORF- Outer Ring Fault. The input parameters are represented by using 14 features in the evaluation. Next, a feature ranking method is established to classify the bearing fault and contribution of each of the features is used as input conditions. It displays the involvement value for each of the 14 parameters. Automatic fault classification has been done by Artificial Neural Networks (ANN). Training on various algorithms is performed, noting and storing the probability of correct prediction for comparison. The probability of correct predictions decreases as the number of samples representing faults increases. A high efficiency of around 97.9% has been achieved for the Resnet50 algorithm. The classifier learner achieved an accuracy of 97% using the neural network, followed by the decision tree and discriminant analysis.

Keywords: Artificial neural network; Ball bearing; Deep neural network; Fault classification; Machine learning; Support vector machine.

Fig. 1

Experimental Set-up. (A) Frequency control Unit: Applying 50 Hz of frequency as a standard given by the Indian electricity. (B) Dc motor: The motor is rotating the ball bearing at a standard of 4000 rpm at a power input of 1/3 HP. (C) Coupling (Flexible coupling): It minimises the occurrence of misalignment effect. It was utilised to absorb shock loads and damped vibration that’s produced between the connecting shafts. It reduced the unwanted disturbance produced by the experiment and increased the Accuracy of the experimental input values. (D) Support bearing (MBER10K) deep groove ball bearing: It was used for the purpose of input values. providing about 5 classes of different bearing. Each of the bearings has unique defects. (E) Balancing Disk: It improved imbalance that was induced by the vibration by controlling the vibration produced and reduced the noise. It increased the signal clarity for the fault bearing. (F) Test bearing: Provided the fault bearing by the company, a total of 5 classes for bearing were given or used for the respective analysis. (G) Triaxial Accelerometer: Its role was to measure vibration for all the three given Gaussian coordinate (X, Y, Z) Direction. The bearing was rotated at a high speed that resulted in vibration, that was transferred to the DAQ system after rectifying noise and transferring analog to digital signal. (H) Microphone: It captures airborne Sound that was produced due to faulty bearing. It then converts the signal into an electrical signal output. It captures the surrounding sound as well hence the experiment was conducted at a silent place without any disturbance. (I) DAQ System (Data Acquisition system): It helps in storing real-time vibration data from the sensor. It supports converting the analog signal vibration data into digital signal vibration data. The reading from the sensors was stored and graphs were plotted through transferring of the data to the Dewesoft software. (J) Then Apply (FFT)- Fourier Transform and Time-Domain Analysis which supports in monitoring the vibration signal and prevents from exceeding the pre-required frequency domain. The Domain was fixed due to Indian standard that is 50 HZ and it was maintained throughout the period of operation. (K) Storage device such as DeweSoft 7.1: It stores the large Set of data within the format while consuming less space. It utilises for measuring the vibration real-time and produces the vibration value into an excel format.

Fig. 2

ResNet-50 Architecture

Fig. 3

AlexNet-50 Architecture

Fig. 4

Functioning of a traditional CNN

Fig. 5

Feature rankings using the MRMR approach

Fig. 6

Training process of the ResNet50 model

Fig. 7

Confusion matrix for ResNet50

Fig. 8

Training process of the AlexNet50 model

Fig. 9

Confusion matrix for AlexNet50.

Fig. 10

Confusion matrix for Decision Tree

Fig. 11

Confusion matrix for discriminant analysis

Fig. 12

Confusion matrix for logistic regression classification

Fig. 13

Confusion matrix for Naive Bayes classification

Fig. 14

Confusion matrix for Support Vector Machine

Fig. 15

Confusion matrix for KNN

Fig. 16

Confusion matrix for Kernels-Random

Fig. 17

Confusion matrix for Ensemble Classification

Fig. 18

Confusion matrix for neural network classification

Fig. 19

Comparison of accuracy for the algorithm