Patent Application
20170052060
The present invention applies to the field of defectology and damage prevention, more specifically in the automatic detection of faults in rotating shafts, such as a railroad car axle, to diagnose the condition of said shafts and assure maintenance thereof.
Methods for diagnosing and monitoring the condition of mechanical systems today are of enormous interest in industry. Predictive maintenance is fundamental to prevent fractures in rotating machines, and one of the most widely used diagnostic methods is vibration analysis, because useful patterns for defecting machinery defects can be extracted from acoustic and vibration signals.
Detecting defects in rotating mechanical elements is critical for the maintenance of said elements, and due to the fact that gears and bearings are the most widely used components in industrial applications, these monitoring techniques have been amply developed for these elements. Today, however, applied maintenance techniques are very different depending on the sector, and for mechanical elements such as railroad car axles, mainly ultrasonic non-destructive inspection techniques of a different type are currently used. In this case fixed intervals are established for performing inspections which, when being performed, make it necessary to leave the machine out of service during the process, wheels even being removed, with the expenses this entails. Ultrasounds furthermore have the drawback of not offering very high probabilities of success in the areas comprising changes in section (which are precisely those which have the highest probability of having a crack), and in the case of magnetic particles (another common technique), the low automation level of inspection equipment means that the technique is susceptible to human error.
In contrast, condition monitoring-based techniques have the advantages of eliminating the need to stop and dismantle the machine to check the condition of its elements, and furthermore the probability of detecting a fault before it becomes severe increases, thereby avoiding losses and increasing process safety. These two reasons have turned condition monitoring in industrial processes into the object of a number of research projects in recent decades.
Specifically referring to shaft monitoring, since the 1960s many studies have been conducted on the dynamic behavior of shafts using both defective and healthy shafts. However, such studies do not usually include the inverse process of defect identification. With respect to the problem of detecting defects in shafts, the diagnostic methods existing in the state of the art can be classified as vibration-based methods, modal analysis-based methods, and non-conventional techniques.
With respect to non-conventional methods, methods using intelligent classification systems, such as neural networks and fuzzy logic, in addition to sophisticated signal processing techniques, such as wavelet transform (WT), for example, are included in this field. In fact, wavelet transform (WT) is the tool that is most extensively used for signal treatment. Use thereof is replacing use of FFT because it is able to eliminate the limitations of this conventional technique by including information about the time domain. As a result, WT is a very suitable tool for working with non-stationary signals, making it particularly indicated for defectology in machines because when a fissure occurs in an element, transient effects that cannot be detected taking into account only the frequency domain are generated in the signal. Therefore, wavelet transform has become much more important in recent years both for defectology of shafts and other important mechanical elements such as bearings, gears, and including structural elements.
One of the latest progresses made in WT is the so-called wavelet packet transform (WPT), which has a large number of applications. This technique allows decomposition of all the frequency bands of the signal, which allows finding differences between healthy and defective elements, which furthermore allows increasing or reducing the resolution used to study the information. The coefficients obtained from WPT can be used directly as patterns for defect detection because they are very sensitive to faults, but there are a number of alternatives for applying this technique and using other types of information relating to the coefficients, such as statistical parameters of the signal or energy measurements.
The state of the art for diagnosing defects in machines is provided with different types of artificial classifiers. It has progressed from initial classical (deterministic and probabilistic) statistical approaches to decision making, syntactic approaches or decision trees, to other approaches that have recently been developed through software packages and are derived from earlier ones, such as fuzzy group theoretical approaches, genetic and connectionist algorithms (ANNs: artificial neural networks); including hybrid versions of all the aforementioned. Today, ANNs have become an intelligent classification system of enormous interest for industry.
Artificial neural networks are an example of automatic learning and processing inspired by how the nervous system in animals works. It is a system of interconnected neurons in a network that collaborates in producing an output stimulus. They are commonly referred to as neural networks in the field of artificial intelligence. The fundamental parts of a basic neural unit are:
Neurons are organized in layers, where they are connected through weights. Basic ANNs have at least three layers of neurons: an input layer, one or more hidden layers, and an output layer. Each layer of the ANN contains a group of processing units called nodes or neurons that send their own output to all the nodes of the next layer in a unidirectional manner.
There are two phases in using a conventional ANN: the learning or training phase and the testing and/or use phase. In the case of an ANN with supervised learning, training consists of presenting a varied set of input patterns, the output of which is known. The network adjusts connections between layers (weights wj) in an iterative error minimization process, until getting said network to comply with the established requirements. There are different algorithms for adjusting weights that make up different network configurations.
According to the foregoing, inventions known up until now through the state of the art perform the function of the object of the invention, but in a rather unreliable and inefficient manner because they do not adapt in the most optimal manner to being applied to rotating shafts such as railroad car axles. The state of the art therefore requires developing these techniques to allow performing a reliable and automatic diagnosis of the condition of a wheel set (detecting if there are faults/defects) along with the remaining elements forming part of a rolling system, without the drawback of having to dismantle the shaft and stopping the operation thereof.
The present invention relates to a method for automatically detecting faults in a rotating shaft. The method comprises the steps of:
According to one embodiment of the present invention, the processing of step b) can be performed by applying a wavelet packet transform to the acquired signal. By applying the wavelet packet transform to the acquired signal, the method of the invention contemplates the possibility of choosing a decomposition level for the acquired signal and a mother wavelet.
According to one embodiment of the present invention, the energy patterns are distributed in packets corresponding to different frequency bands of the acquired signal.
Optionally, one of the embodiments of the present invention contemplates detecting changes in energy in one and the same frequency band between different signals previously obtained from the rotating shaft to establish energy patterns.
In one of the embodiments of the present invention, the comparison of the energy measurements with previously established patterns is performed by means of a neural network. Depending on the different embodiments of the invention, the neural network is selected from a unidirectional neural network, a neural network for pattern recognition, a multilayer perceptron neural network, a radial basis function neural network or a probabilistic neural network. The possibility of adding an additional neural network training step where signals the output result of which is known are used as input of the neural network is contemplated. Optionally, the method can comprise increasing the decomposition level of the acquired signal in the case of obtaining a hit rate less than a certain threshold.
One of the embodiments of the present invention contemplates the possibility that the rotating shaft is a railroad axle mounted on a railroad car, where said railroad car is in motion.
In other words, the railroad car or rolling system comprises a rotating shaft of a railroad car or a rotating element assembled together with the remaining elements normally comprised in said railroad car or rolling system.
Faults or defects detected in a rotating shaft by the method of the present invention can be classified in different levels of severity according to one of the embodiments of the invention. For example, the levels of severity can correspond to the depth of cracks or fissures detected in a railroad car axle.
A second aspect of the invention relates to a system for automatically detecting faults in a rotating shaft. The system comprises:
A possible embodiment of the system of the invention contemplates the possibility that the rotating shaft is a railroad car axle.
According to a particular embodiment of the invention, the sensor used for acquiring vibration signals from the rotating shaft is an accelerometer. Said accelerometer can be installed, for example, in the grease box linked to the railroad car axle under study.
A final aspect of the invention relates to a computer program comprising program code means suitable for performing the steps of the method of the invention, when said program is run in a general purpose processor, a digital signal processor, an FPGA, an ASIC, a microprocessor, a microcontroller, or any other form of programmable hardware.
Therefore, according to the described invention, the method and system proposed by the invention is an advancement in methods for diagnosing and detecting faults in rotating shafts, primarily applied to railroad car axles, used up until now. The drawbacks described above are therefore resolved in a fully satisfactory manner by way of predicting the wear of shafts by means of detecting faults at different levels of severity without having to dismantle the shafts or leave machines out of service while they are being inspected, this inspection being done automatically with the subsequent improvement in operations associated with both operation and maintenance of the machine, and the subsequent cost reduction due to process optimization and efficiency, and all this without requiring any special operator skills and without any difficulty in use.
To complement the description being made and for the purpose of aiding to better understand the features of the invention according to a preferred practical embodiment thereof, a set of drawings is attached as an integral part of said description in which the following has been depicted with an illustrative and non-limiting character:
In view of the mentioned drawings, it can be seen how one of the possible embodiments of the invention proposes a non-destructive method and system for detecting faults in shafts rotating at low revolutions, being fundamentally applied to the wheel set of a railroad car, including an axle. It is based on condition monitoring, and the advantage of applying the invention to railroad car axles is that such axles do not need to be dismantled for inspection.
The vibrations given off are continuously monitored and processed in the time domain and in the frequency domain, selecting optimal parameters that allow being used as input information in an intelligent classification system.
Vibration measurement-based techniques include, among others, techniques in which signals are obtained in both an experimental and analytical manner. Phenomena observed in dynamic analyses of these elements when a defect occurs, such as certain instabilities observed in critical shaft speed subharmonics and rotational speed harmonics, are sought for defect detection. However, although these effects are theoretically transparent, in practice early detection of a defect is not a simple task, and the rate of detection is often below 25%.
The present invention replaces typical optimal coefficients of MRA analysis with the use of energies calculated by means of WPT. The proposed invention is therefore based on obtaining WPT energy from vibration signals obtained during normal operation of the rotating shaft under study without having to dismantle it. Optimal patterns are chosen to be classified by means of neural networks. Currently, the present invention has been tested successfully using an RTK-01 prototype laboratory bench on a ⅕ scale.
To apply methods for monitoring gears and bearings to the case of railroad car axles, which are extremely critical elements, studies of several functionals have been conducted according to different embodiments of the present invention in order to determine the most optimal indicators of the presence of a defect in a shaft, such as a crack or a fissure.
The most widely studied functionals in the state of the art are, on one hand, statistical parameters (such as the mean, median, standard deviation, bias, the kurtosis, root mean square, minimum, maximum, peak-to-peak value, peak value, shape factor or crest factor), and on the other hand, representations of the signal that allow studying it in both the time domain and frequency domain.
With respect to frequency domain representations, they have been conventionally used to examine frequency peaks in which amplitude is modified or which are displaced due to the presence of a defect. This is because the presence of a crack in a structural element reduces rigidity thereof, so natural frequencies are reduced and vibration modes change. The study of natural frequencies has always been the first step in finding indicators of damage in a mechanical element because it can be measured in a rapid and easy manner and it experiences very little contamination due to experimental noise. Differences in energy in certain bands indicating the presence of a defect are detected by conducting thorough analysis by frequency bands, but the main drawback of this technique is that when a defect occurs, the changes in natural frequencies are very minor and can be covered up due to experimental errors.
In the present invention, however, vibration modes are not observed, but rather response signals occurring during normal motion of the machine associated with the rotating shaft are, so the noise drawback will be included.
With respect to the functionals describing a signal in the time domain and frequency domain, wavelet transform stands out and it has been chosen for the present invention. Such transforms are a very suitable mathematical tool for detecting defects in mechanical elements. One of its greatest advantages is that it allows working in multiresolution mode, which allows efficiently locating local phenomena in non-stationary signals with rapid transients. Just as frequency coefficients of sinusoidal signal components are obtained with Fourier transform, correlation coefficients between the signal and a function referred to as mother wavelet are obtained with wavelet transform. Said correlation coefficients furthermore depend on the scale and position of the mother wavelet. The mother wavelet is chosen due to its morphology, as it more greatly resembles the effects produced in signals with the presence of a defect.
In the experiments conducted for the present invention, a mother wavelet set is selected in one of the embodiments to evaluate which has the highest correlation with the evaluated defects. It is determined that the Daubechies wavelet family, and particularly the Daubechies wavelet 6, is the one having the best fit.
Continuous wavelet transform (CWT) is a linear operator that associates each finite energy signal with a particular function, referred to as mother wavelet (21), suitably chosen according to the signal. Once the instant and the observation scale are selected, the CWT allows analyzing structures of the signal through correlation coefficients (22), instead of analyzing the signal (23) as a whole, as seen in
The advantage of the continuous wavelet transform is that it is a very useful tool for detecting, characterizing and classifying signals with complex spectral characteristics, transient contents and other properties related to non-stationary signals.
The mathematical formulation of continuous wavelet transform is as follows:
Where X(t) is the transient signal and ψ is a mother wavelet function. w(a) is a weight function. In equation 1, b is related to positioning and a is related to the scale. The weight function is used for energy conservation purposes (assuring that wavelets will have the same energy in each scale). T(a, b) are the wavelet coefficients that depend on each scale a and on displacement b. The product of the wavelet conjugated with the signal is integrated throughout the entire range of the signal, which is mathematically known as convolution.
Based on the definition of continuous wavelet transform, working with discrete wavelet transforms is common practice. This is due to their computational efficacy and to the fact that discrete data signals are usually used. The most common aspect of making CWT discrete is using the dyadic grid, where a=2i. In this case, the transform is determined by:
Each i is referred to as octave and consists of each of the levels in which the signal is decomposed. Low scales take into account low frequencies, and high scales take into account higher frequencies. When orthonormal wavelets are used, a method referred to as the decimation method is commonly used. This method consists of decomposing the signal into a number of coefficients that is proportional to the analyzed scale. This means that the signal will have a different number of coefficients on each scale. Physically speaking, it reflects the fact that lower frequencies of a signal need fewer coefficients in order to be represented.
Both multiresolution analysis (MRA) and wavelet packet transform (WPT) analysis are an extension of the discrete wavelet transform, where a discrete signal X is processed at different frequency bands and with different resolutions, when the signal is decomposed into approximation information (32) and detail information (33) with high-pass recursive filters (34) and low-pass recursive filters (35), as seen in
The approximation contains information about the high-scale and low-frequency components, whereas the details contain low-scale and high-frequency information.
The main difference between MRA methods and analysis by means of WPT is fundamentally that in MRA analysis, only approximation parameters are decomposed, losing detail information and information about approximation details, whereas in analysis by means of WPT, both detail and approximation are decomposed in each of the levels.
In this case, the decomposition algorithm is identical to the multiresolution analysis, except in this case both the approximation and detail coefficients are decomposed. Therefore, if each correlation coefficient vector, W(k,j), is in the following form:
W(k,j)={w1(k,j), . . . , wN(k,j)}={wi(k,j)} (Eq. 3)
signal recomposition in this case must take into account the eight coefficient vectors (packets), W(3,j) corresponding to level 3, according to the following equation:
W(0,0)=Σj=07|wi(k,j)|2 (Eq. 4)
The concept of energy used in the wavelet packet analysis is closely linked to the known notions derived from Fourier theory. Wavelet packet energy is obtained from the sum of the squares of the coefficients of each packet:
E
k,j=Σi|wi(k,j)|2 (Eq. 5)
To obtain the relative energy, total energy of the signal before the decomposition wavelet is calculated:
Etotal=ΣjEk,j (Eq. 6)
Finally, the relative energy of each wavelet packet j in decomposition level k is defined by:
According to different tests of the present invention, the signals are obtained from railroad car axles on a ⅛ scale, tested in a test bench called ROTOKIT. The studied signals, obtained with accelerometers, are from both healthy shafts and from shafts with several defect levels (for example 16%, 33% and 50% of the diameter), and at different test speeds (20 Hz, 40 Hz and 60 Hz). The first approximation is taken into account as the optimal functional for the wavelet packet energy classification system because it has been determined that it is the most relevant for studying defects in mechanical elements and it allows reducing the number of pieces of data to be entered in the classification system without losing relevant information.
System output (64) generates a diagnosis that corresponds to the actual operating condition of the shaft. Conditions of healthy shafts/defective shafts are used in the first approximation, and as more information is obtained at the input by means of system training, discrimination of fault conditions is performed going from minor to serious, according to the different previously established levels of severity.
As previously explained, the optimal functional chosen to perform comparisons and/or classifications is the wavelet packet energy. Visually, it is very easy to compare differences between energy patterns, as shown in
For the tasks of classification and comparison with patterns, the present invention uses artificial neural network (ANN) architecture. Different tests from among all the possibilities have been performed to see which is the most suitable for the particular case of fault detection in shafts by means of running different tests.
According to one embodiment of the invention, the energy of the wavelet packets obtained from a Daubechies mother wavelet 6 and having a decomposition level equal to 3 have been used as patterns in test cases. In other words, eight energy levels corresponding to the eight packets (23) are presented to the neural network as input, as depicted in
Tests have been conducted with two types of multilayer ANNs integrated in the easy-to-apply MATLAB™ toolbox, which will allow studying the applicability of such classification system to the particular case of railroad car axles. These are the FeedForward type network and pattern recognition network:
The good results obtained with the training of the two simple ANN configurations described above make applying the same concepts to other more complex and flexible networks that are widely applied in the field of classification, such as multilayer perceptron (MLP) networks, radial basis function (RBF) networks and probabilistic neural networks (PNN), an evident step.
| Number | Date | Country | Kind |
|---|---|---|---|
| P201430606 | Apr 2014 | ES | national |
| Filing Document | Filing Date | Country | Kind |
|---|---|---|---|
| PCT/ES2015/070348 | 4/24/2015 | WO | 00 |
