A Summary on “Vibration analysis for bearing fault detection and classification using an intelligent filter”
In this story, I’m going to summarize a paper on using machine learning techniques in fault detection for bearings. Paper’s title is “Vibration analysis for bearing fault detection and classification using an intelligent filter” and it has been published in Mechatronics journal in 2014. This paper proposes an intelligent method based on artificial neural networks (ANNs) to detect bearing defects of induction motors.
Importance of the subject
Induction motors with many important advantages such as high
reliability and performance have a critical role in many industries. In spite of their reliability, they are subject to some failures. Fault detection techniques can be classified into three groups of signal-based, knowledge-based and model-based techniques. Knowledge-based techniques use intelligent methods such as fuzzy systems and neural networks, for instance, to diagnose the faults in induction motors. This paper focuses on knowledge-based techniques. Based on the analysis of the origin of induction motor failures, the bearing fault is the major source of most mechanical faults.
Therefore, the bearing fault detection and troubleshooting in the early stages will decrease the cost of the unwanted shutdown.
Locate The Problem
The bearing condition can be monitored very well via machine vibration (among many condition monitoring techniques). This is because bearing faults will typically produce consecutive and periodic impulse terms in machine vibration caused by passing the ball bearing through the defect points. When we record the machine’s vibration, we have domain-time data which is a super complicated diagram to be analyzed by the human. to solve this problem, engineers convert this domain-time diagram to frequency-domain diagram which is much simpler to observe and detect abnormal behaviors. this is also called as “spectra”. The simplest frequency-domain conversion method used for bearing fault detection is Fast-Fourier Transform (FFT).
Problem #1) The impact of vibration generated by a bearing fault has relatively low energy and it is often accompanied by high energy noise and vibration generated by simultaneously-active equipment. Therefore, it is difficult to identify the bearing fault in the spectra using conventional FFT methods.
Problem #2) Using FFT to detect faults needs rely on bearing parameters such as number of balls and dimension, it means that these solutions are specific for each machine, based on its characteristics.
Propose The Solution
In the proposed method, first, healthy components of the vibration signal are estimated by an intelligent filter that is designed based on NN. Note that this network is trained under the normal condition. Therefore, when a measured vibration signal passes through this filter, components that exist in a healthy condition are removed and the filter output contains the components that are relevant to faulty conditions (addresses problem #1). In the next stage, time-domain features of the filtered signal are extracted to use as inputs of a classifier in order to distinguish fault localization (addresses problem #2).
It is noting that this approach does not need any bearing parameters
for calculation of defect frequencies as it uses only time-domain
features (and not frequency-domain) to detect the faults.
Implementation of Solution
As mentioned in the previous section, at first we need to remove unwanted part of sampled data which is healthy part or parts which are related to other defects (not bearing ones).
v(n): Motor vibration signal.
y(n): Estimated irrelevant part of the vibration signal (nonbearing
fault components).
e(n): Faulty part of the vibration signal.
n0: Number of data samples.
Several algorithms can be used for designing an RNFC filter. In this paper, a NN is chosen due to its advantages. The considered RNFC filter is an Adaptive Linear Neuron (ADALINE) neural network with purelin activation functions. A supervised training method is applied to achieve the healthy part as the target.
It is worth mentioning that this training process for RNFC filter is just needed for one time. More severe bearing fault results in more considerable RNFC filter output amplitude. Therefore, this signal can be used as an indicator for fault detection in the motor. Note that analysis of this information only enables us to detect the fault, in other words, it is not able to determine where the damage is. Motivated by this consideration in the next section fault classification using RNFC filter output is proposed to determine the fault location.
The aim of this section is classification of healthy and defective bearings in four categories, including healthy condition, inner race defect, outer race defect and double holes in outer race. Proper features from the RNFC filter output have to be selected in order to train the neural network based classifier. Four features are selected as follows:
Among different methods to implement a NN, It is shown that the MLP (Multi-Layer Perceptron) is superior to others in terms of speed. Therefore, where speed is an important factor, the MLP may be more appropriate.
Using Solution in Action
A three-phase, 1.2 kW, 380 V, 1500 rpm, fourpole induction motor is used to collect experimental data. Both shaft-end and fan-end bearings are 6205–2Z. The vibration signal is sampled by Advantech PCI-1711 data acquisition card with 32 kHz sampling frequency using B&K 4395 accelerometer.
After training is completed for RNFC, the test input must be extracted
from the sampled vibration signal in the form of input patterns. In order to gather input prototype for the training of the second neural network, different periods of faulty vibration signals are considered as inputs for the RNFC filter. A period of RNFC filter output amplitude for four different fault categories is shown in the following picture.
Evaluation
The ability of the proposed method is obviously verified by comparing the results of Table 1 with the results of a fault detection system without RNFC filter (classifier only) that are listed in Table 2.
The robustness of the proposed method in the presence of low-quality signals is shown in Table 4 which compares the fault detection in presence of low quality sampled signals using RNFC filter with a system without filter.
Therefore, this method can be used in noisy industrial environments for fault detection with high performance.
Conclusion
In this paper a new algorithm for fault detection and classification, in two steps, is proposed. In the first step, a filter is designed by neural network to omit the non-bearing fault component of the sampled signal. This filter called removing non-bearing fault component (RNFC) filter. In the next step, time domain features are extracted from RNFC filter output and applied to another neural network in the form of prototype features to detect faults using a pattern recognition technique.
References:
https://www.sciencedirect.com/science/article/abs/pii/S095741581400004X
