Abstract: This study deals with the early detection of damages in rolling bearings of an asynchronous machine. It presents a neuronal network based application for motor bearing fault diagnosis. Vibration simulation is used to assist in the design of various motor rolling bearing fault strategies. These simulation results indicate that neural networks can be effective agents in the diagnosis of motor bearing fault through vibration signature.
INTRODUCTION
The problem of fault diagnosis in systems is of utmost importance. It constitutes an important issue in maintenance management strategies in order to reduce the failures and their consequences. That requires an important knowledge of system and also the use of methods for formalization of this knowledge. This problem is subject of many scientific works and three main methods of diagnosis are proposed.
The first method is the diagnosis by analytical model (Henao et al., 2005; Prashad, 2002). It is based on the use of analytical models of the process to be supervised. If one has the inputs of the model, it is able, by comparing the outputs of model and the measurement of the real the system, to determine if the system is healthy or abnormal. And by analyzing the errors, it is also possible to determine the causes of diseases and to deduce the failing behaviour from it. The improvement of the diagnosis needs a model of failure state. However, it is difficult to build this method because it results of a compromise between the precision of the model, its robustness in the operating conditions and the constraint to work in real-time. The techniques using observers as well as parameters identifications techniques have been largely employed on the academic case of the asynchronous machine and showed the difficulty of there implementation.
The second method is the diagnosis by signal analysis (Djebala et al., 2007; McInery and Dai, 2003). Referring to the diagnosis practiced in industries, the expertise associated with an adapted instrumentation allows to detect and to isolate the failures. In the context of automation of these techniques and reduction of instrumentation, the analysis of the harmonic of the measured values is used. Thus signal analysis techniques are largely used in detection of breakdown. The first difficulty of this technique is that the instrumentation sensors must be appropriated to the system in order to avoid signal distortion witch could affect the quality of the measures by the multiplication of the natural harmonics of the system and by the attenuation of the transfer function. The second difficulty deals with the choice of the signatures analysis method and the decision procedures. Once again that calls upon the expertise in each feed.
The third method is the diagnosis by evaluation of the state of a system. It consists to determine of the behaviour of the system when it is subject to certain constraints. Thus, we can evaluate the state of the system by knowing what it is endured. For that, it is necessary to know relations between causes and the potential failures. It calls upon the experiment either of the manufacturer or of the user (the experience feedback). Unfortunately this information is generally partial. That why monitoring systems are used. It is interesting to have a tool witch is able to analyze real-time state of the system in order to highlight the scenarios leading to a situation of failure. In this analysis there is at the same time the concept of operating condition (classification between nominal, overload and defect state for example), concept of distribution in the time of these successive states and concept of relation between a type of scenario and a situation of breakdown.
To carry out this analysis, it is necessary to organize the data acquisition on the supervised system in order to classify the operating conditions, to establish a history and to determine the evolution of the state of the system. The knowledge of the scenarios leading to failure can be built by capitalizing the data of maintenance in order to improve the training of the scenarios so that we can anticipate on the occurrence of the failures. The big size of the data and there treatment require to use data mining techniques. Nowadays, the improvement of the performances of the expert systems especially fuzzy logic and neuronal networks make this method of diagnosis available (Chow, 1997).
Concerning the predictive maintenance of turning machines, the monitoring of rolling bearings is of the utmost importance. Those components have indeed a strategic role to play in the good operation of turning machines. One of the processing methods often used for the detection of damages in rolling bearing, is based on the analysis of mechanical vibrations. The main objective of this paper is to improve this analysis by using multi-neuronal networks in order to help for fault classification and taking decision. Our work is a contribution to asynchronous motor incipient fault detection using artificial neural networks techniques.
MOTOR BEARING VIBRATION FREQUENCY FEATURES
Since most bearing vibrations are periodical movements, it is easy to extract vibration features from the frequency domain using the powerful and popular FFT technique. Many publications have studied the frequency features of rolling bearing (Kowalski and Orlowska-Kowalska, 2003). Generally, rolling bearings consist of two concentric rings, called the inner raceway and outer raceway, with a set of rolling elements running in their tracks. Standard shapes of rolling elements include the ball, cylindrical roller, tapered roller, needle roller and symmetrical and unsymmetrical barrel roller. Typically, the rolling elements in a bearing are guided in a cage that ensures uniform spacing and prevents mutual contact.
There are five basic motions that are used to describe the dynamics of bearing elements, with each movement having a corresponding frequency (Pan et al., 2006; McInerny and Dai, 2003). These five frequencies are denoted as the shaft rotational frequency (FS), the fundamental cage frequency (FC), the ball pass inner raceway frequency (FBPI), the ball pass outer raceway frequency (FBPO) and the ball rotational frequency (FB). These frequencies are shown in Fig. 1.
| Fig. 1: | Basic frequencies in bearing |
| Fig. 2: | Structure of a ball bearing and definition of each variable |
Figure 2 describes several important variables that will are Vi, Vc and Vo represent the linear velocities of the inner raceway, ball center and outer raceway, respectively. Db is the ball diameter, Dc is the bearing cage diameter measured from one ball center to the opposite ball center and θ is the contact angle of the hearing.
We can summarise the main relation concerning the vibration frequencies by next equations:
(1) |
(2) |
(3) |
(4) |
where, Fs is rotor speed.
Frequency-domain studies show that, when defects exist in a bearing, the defects will generate some of the above frequencies in the vibration signals. Many publications have discussed the use of these five frequencies to identify defects in a bearing assembly.
| Table 1: | Bearing vibration features |
For defects on the raceway of a rolling bearing each time a roller hits the defective raceway, the corresponding ball pass inner raceway frequency FBPI or ball pass outer raceway frequency FBPO will be excited. If the defective area is large, harmonics of FBPI or FBPO will also be present as an indication of the severity of the defects. For defects existing on a bearing roller, usually, two times the ball rotational frequency 2FB will be generated. This is because the roller hits both the inner and outer raceways each time it spins on its own axis. In most cases, this frequency will be modulated with other existing frequencies, such as FBPI and FBPO, resulting in a more complicated spectrum. Table 1 summarizes the possible motor bearing vibration features in the frequency domain.
DESCRIPTION OF THE DEVICE
The general synopsis of the device as showed in Fig. 3 consists of two modules:
| • | FFT Transformer |
| • | Neural Network Module |
FFT transformer: It allows to definite representative vectors from information provided by the measured acceleration. We determine the spectrum of amplitude of the signal by using the FFT Transformer.
Neural network module: This module is a hybrid neural structure. Its consists of a Learning Vector Quantization (LVQ) neural network used as classifier and a back-propagation (BP) neural network used as decider. The size of the input layer of the LVQ is definite automatically by the network; it corresponds to the size of the input vectors. In the same way, the output layer is consisted of the various categories or class. The performances of the network depend on the number of neurons of the hidden layer and training bases.
| Fig. 3: | General synopsis of the device |
| Table 2: | Bearing vibration simulated signals distribution |
| Table 3: | Results |
RESULTS
To validate our program we used data obtained by a computer simulation of the induction motor (1.5 kW/25 Hz). Although motor system dynamics simulation software cannot completely model all real situations, a computer simulation can assist in many aspects of motor system operation and control. It allows us to define the general base of training and the various bases of test.
Two kinds of bearing faults have been simulated: 1) defect on bearing ball; 2) defect on inner raceway. In addition normal bearing condition has also been simulated. Our data base contains 150 signals corresponding different scenarios of rolling bearing as shown in Table 2.
Each signal consists of 2084 sampling points with a sampling time of 0.001 seconds. An FFT is performed to obtain power spectrum of these simulated vibration signals. That will be used as input of the neural network. It allows to train and to optimise neural network. Table 3 presents the main results we have obtained.
| Table 4: | Neural network performances |
Two parameters, the number training data sets and the number of hidden neurons are varied to find the optimal design. Table 4 shows the performance of the neural network. The result demonstrates that with data and possible training procedure, the neural network motor bearing fault diagnosis schema can diagnose bearing faults with desired accuracy.
CONCLUSION
This study has discussed several popular rolling bearing vibration features in both the time and frequency domain and the use of signal processing to provide features to be used for bearing fault diagnosis. Neural networks have been used in this paper to perform motor bearing fault diagnosis. Computer simulated data have been used to study and design the neural network motor bearing fault diagnosis algorithm.
Actually we plan to collect bearing vibration data in real-time in order to perform initial testing and to validate of the approach. The results show that neural networks can be effectively used in the diagnosis of various motor bearing faults through appropriate measurement and interpretation of motor bearing vibration signals.
ACKNOWLEDGMENTS
The authors would like to thank all the contributors to the database used in this study. Special thank to Prof K. Madani our research director and M. Barret who provides all simulated data.