A Low Complexity Rolling Bearing Diagnosis Technique Based on Machine Learning and Smart Preprocessing

Abstract

In this work, we present a diagnosis system for rolling bearings that leverages simultaneous measurements of vibrations and machine rotation speed. Our approach combines the robustness of simple time domain methods for fault detection with the potential of machine learning techniques for fault location. This research is based on a neural network classifier, which exploits a simple and novel preprocessing algorithm specifically designed for minimizing the dependency of the classifier performance on the machine working conditions, on the bearing model and on the acquisition system set-up. The overall diagnosis system is based on light algorithms with reduced complexity and hardware resource demand and is designed to be deployed in embedded electronics. The fault diagnosis system was trained using emulated data, exploiting an ad-hoc test bench thus avoiding the problem of generating enough data, achieving an overall classifier accuracy larger than 98%. Its noteworthy ability to generalize was proven by using data emulating different working conditions and acquisition set-ups and noise levels, obtaining in all the cases accuracies greater than 97%, thereby proving in this way that the proposed system can be applied in a wide spectrum of different applications. Finally, real data from an on-line database containing vibration signals obtained in a completely different scenario are used to demonstrate the distinctive capability of the proposed system to generalize.

Keywords: condition monitoring; embedded systems; fault bearing fault diagnosis; machine learning.

Figure 1

Bearing structure and components. The main components where wear typically occur are the inner and the outer race as well as the rolling elements.

Figure 2

Roller bearing geometry, representation of the defects placement and characteristic frequencies related to them. Wear occurs in components that mechanically interact during operation: the outer race with rolling elements, and the inner race with rolling elements. FTF is the train/cage frequency, BPFI is the inner race failure to rolling element impact frequency, BPFO is the outer race failure to rolling element impact frequency, and BSF is the rolling element failure to two-race impact frequency.

Figure 3

Simulated signals as per Equation (1), considering the same rolling bearing, for the three defect classes. In the reported example, fs = 10 kHz, RPM = 3000 rpm, BPFO = 361 Hz, BPFI = 488 Hz, BSF = 161 Hz, FTF = 21 Hz.

Figure 4

Real vibration signals from an online database (CRWU) for the three different failure classes; top: outer race, mid: inner race, bottom: ball. The signal to noise ratio decreases in inner race defect and the ball defect signals, due to the decreased amplitude of the vibrations, caused also by larger distance between the sensor and the defect.

Figure 5

Example of accelerometer filtering. Left: vibration signal response, outer race, resonant frequency f_res = 7 kHz. Right: simulation of the sensor output considering the sensor to behave as a second order resonant system, with resonance at f_n = 1 kHz and quality factor Q = 10.

Figure 6

Diagnosis system block diagram. The system uses rotational speed and vibration data from an accelerometer. Faults are detected using conventional techniques; a neural network (NN) classifier is then employed for fault localization when a fault is identified.

Figure 7

Representation of the images related to F_fault matrices according to the type of signal and the delay used: (a) vibration signal is a train of transients, and the delay time is equal to the train period; (b) delay different from the train period; and (c) the vibration is not a transient train (Images are saturated for clarity’s sake).

Figure 8

Example of fault feature images, F, realized with the proposed method, using noisy signals (images are saturated for clarity’s sake). The example images were built exploiting signal windows made up of 6400 samples.

Figure 9

Neural Network Classifier block diagram. The classifier consists of five layers designed to match the characteristics of the input matrix. The complexity is intentionally low to minimize resource consumption, enabling deployment on resource-limited devices.

Figure 10

Example of emulated noisy signals. Ball fault, machine vibration at 10 kHz. Above: Gaussian noise standard deviation = 20% of the acceleration peak value, a₁ = 10% of the signal peak value. Below: Gaussian noise standard deviation = 10% of the acceleration peak value, a₁ = 10% of the signal peak value.

Figure 11

Results of different analysis types on data from the CRWU database. Analysis of time window of signals in the ‘O’ fault class: (a) feature image obtained with the preprocessing proposed in this paper; (b) amplitude of the spectral components (of the signal envelope) located at BPFO, BPFI and 2BFS normalized with respect to the spectral energy in the base-band (20 Hz–2 kHz). Analysis of signals form the ‘I’ fault; (c) feature image obtained with the preprocessing proposed in this paper; (d) amplitude of the spectral components located at BPFO, BPFI and 2BFS normalized with respect to the spectral energy in the base-band (20 Hz–2 kHz). Analysis of signals form the ‘B’ fault; (e) feature image obtained with the preprocessing proposed in this paper; (f) amplitude of the spectral components located at BPFO, BPFI and 2BFS normalized with respect to the spectral energy in the base-band (20 Hz–2 kHz); (g) envelope spectra obtained from three signals classified as ‘O’, ‘I’ and ‘B’ faults, respectively.

Figure 12

Training and validation results, in terms of accuracy and loss. Training was stopped after 3000 epochs: (a) final training and validation accuracy were 99.45% and 98.86%, respectively; (b) training and validation loss.

Figure 13

ML model test confusion matrix, signals with increased noise. Each class contains 147 samples; misclassification occurs mostly in the discrimination of ball defect and noise.

Figure 14

ML model test confusion matrix, signals from CWRU database. Each class has 129 elements. The dataset includes only I, O, and B classes due to the absence of Noise class in the database. The high misclassification in the B class is attributed to the lack of periodicity in the real signals, preventing accurate classification.