Rough Set Neural Network Feature Extraction and Pattern Recognition of Shaft Orbits Based on the Zernike Moment

,


Introduction
Because of the complex operating conditions involved, there are many factors that cause the instability of hydropower units, including hydraulic, mechanical, and electromagnetic factors. Among these three important factors, hydraulic vibration is the excitation source that has a crucial influence on the vibration of the unit. If the low-frequency oscillation is induced by the vibration of the power system, it will affect the stable operation of the power grid [1,2]. rough hydraulic excitation, each component of the unit exhibits different vibration characteristics. During the shafting monitoring, the shaft orbit directly and comprehensively reflects the operation state and vibration fault of the unit, and different shaft orbits usually reflect different faults. e hydroelectric unit condition monitoring system can obtain many operating characteristics of the unit from the axis orbit of the unit, so the recognition of the shaft orbit plays an important role in the fault diagnosis.
In recent years, many experts have performed theoretical research and simulation analysis on hydrogenerator shaft orbits, using different methods to research the recognition of shaft orbits [3][4][5][6][7]. At present, the recognition of shaft orbits mainly includes precession direction recognition and shape recognition. Shape recognition of shaft orbits mainly includes feature extraction and pattern recognition. e feature extraction methods include Fourier descriptors [8][9][10][11], Hu moments [12][13][14][15], and Zernike moments [16], as well as classification of the data with an advanced pattern recognition method [17]. Kauppinen et al. [18] tried to use autoregressive modeling and Fourier descriptors of closed contours to recognize the typical shape of shaft orbits. is showed better performance for Fourier-based methods, especially for images containing noise. Fu et al. [8] combined a BP neural network and modified Fourier descriptors to classify the shaft orbits for fault diagnosis of the turbine rotor. e simulation results were for the automatic identification of three typical shaft orbits. In order to obtain the characteristics of the axis orbit, the edge acquisition algorithm is typically used, and, by contrasting each edge extraction algorithm, the applicability of the corresponding algorithm is obtained. en the algorithm is improved according to the corresponding boundary features. An intelligent model, such as the neural network model, is established to achieve the intelligent classification of faults in the mechanical system.
Fourier descriptors have the advantage of feature extraction for completely closed curves, but the shaft orbits' coincidence performance for the hydropower unit is so poor that it is hardly closed, so this paper uses Hu moments and Zernike moments to extract features of the shaft orbit.
For intelligent pattern recognition, this paper adopts the common BP neural network and applies the rough set theory to simplify the neural network with overly large input samples and a complex network structure.

Formation and Characteristics of the Shaft Orbit
In the hydrogenerating unit stability test and application of the monitoring system, two eddy current displacement sensors are orthogonally installed in the guide bearing (upper and lower guide) horizontal section. is ensures that the vibration signal of two mutually perpendicular directions can be measured. en the high-frequency noise of the vibration signal is eliminated using a low pass filter, and finally a center orbital path is formed using the analogto-digital converter. As shown in Figure 1, two shaft swing sensors are installed in the unified plane of the rotating shaft. e position of the rotor center can be determined by the outputs x(t) and y(t) of the two sensors at the t moment, and then the position of the rotor center at different time can be transformed into a curve, which is the vibration track of the rotor in the bearing, called the shaft orbit. e shaft orbit of different shapes reflects the operation of different units, and its shape is an important basis for judging the running state and failure of the shaft, which implies the fault information of the unit. Table 1 shows the shaft orbit shapes and the potential failure types, and Figure 2 shows typical faults.

Hu Moments and Improvement
3.1.1. Hu Moments. Hu moment theory has been widely used in the field of image recognition. It is a method of extracting image features based on a statistical feature. Two-dimensional moments are often used in digital image recognition. In fact, it extends the order moments of onedimensional signals to two-dimensional space. By using normalization, we obtain the moment's characteristic quantities that are sensitive to the change of shape, while the translation, rotation, and scaling of two-dimensional images remain unchanged. Hu [19] constructed 7 moments with rotation, translation, and scale invariance in 1962. In the field of image recognition, a statistical feature extraction method was established using moment theory, and it was widely used. In 1978, Wong and Hall [20] proved that the invariance of the Hu moment could be guaranteed when the scaling factor ρ ≤ 2 and the rotation angle of the trajectory graph θ ≤ 45°. For a two-dimensional function f(x, y) ∈ (R 2 ), which is defined on the plane o − xy, the p + q order mixed origin moment is defined according to the following formula: e p + q order mixed central moment is defined according to the following formula: where (x, y) is the gray center of gravity for the image f(x, y).
Ming-kuei Hu proposed the construction method of moments in 1962. By using the above central moments, 7 moment functions are constructed. ese seven expressions correspond to translation, scaling, and rotation. e p + q order origin moment and central moment of a two-dimensional image function f(x, y) in a discrete state are defined according to the following formulas: Under the constraints, Normalized central moments in discrete states, which are defined by formulas (8) and (9), can clearly show that the moment functions in discrete states are also translationindependent. By observing the definition of the central moment, it is found that the central moment may be negative. Because the change range of moment s is large, data compression can be achieved by the logarithmic processing of moments. Since the moments may appear negative, we need to take the absolute value before doing logarithmic operations. e actual moments can be obtained by means of a logarithm:

Improvement of Moments.
If the scaling factor of the trajectory graph is ρ, the relationship between the moment ϕ i ′ in discrete states after scaling and the moment ϕ i of the original graph is shown with the following formula: In order to make these moment functions independent of scaling, we can consider the elimination of the scaling factor and use ϕ 2 ′ as the benchmark to find the final expression of the rest moment functions obtained by this moment function.

Zernike Moment.
Zernike moment is one kind of special complex moments, which are based on orthogonal functions called Zernike polynomials. Compared with HU moment, the computation of Zernike moments is more complex, but they have greater advantages in terms of their expressive power and noise sensitivity. Zernike moments have the characteristics of rotation invariance and have been widely used in the field of target recognition.
Radial polynomials R nm ( ρ) satisfy the following relationship: where δ nm is the Kronecker delta. e orthogonality conditions of Zernike polynomials are as follows: (16) where * denotes the complex conjugate. e following formula can be obtained from formula (16): Some important properties of R nm (ρ) can be obtained from the following formula: An n order Zernike polynomial with a repetition rate of M is defined according to the following formula: For a digital image, the integral is replaced by a summation, namely, To calculate the Zernike moments of an axis shaft orbit, it is necessary to move the center of the axis trajectory image to the coordinate circle and map the pixel points of the trajectory into the unit circle. rough the formula, it can be seen that V * nm (ρ, θ) can extract the features of images, and the high-frequency characteristics are obtained by V * nm (ρ, θ) for larger n values. e lowfrequency characteristics are obtained by V * nm (ρ, θ) for smaller n values. Zernike moments can be used to construct high moments. In theory, higher moments can contain more image information, so pattern recognition will be better.

Rotation of Zernike Moments.
When the rotation angle of an image is a, the expression of its Zernike moments is as follows: From the upper formula, it is shown that the Zernike moments' phases change, but the modulus remains unchanged. So, |A nm | can be regarded as the target rotation invariance feature.

Rough Set Neural Network.
Rough set theory (RS theory) [22] was proposed by Pawlak, a mathematician of Poland in 1982. It is an effective method for dealing with uncertain, inconsistent, and imprecise problems. It can discover hidden knowledge from uncertainty and reveal its potential law. It defines knowledge from a new point of view, regards knowledge as a division of the universe, and holds that knowledge is granular. In recent years, the research on rough set theory in the field of artificial intelligence has been a hot topic, and this theory has been widely used. Knowledge reduction [23][24][25] is one of the core concepts of rough set theory.  table is meant to simplify  the conditional attributes in the decision table.  e simplified decision table also has the function of simplifying the  decision table before, and the simplified decision table has less conditional attributes. e rough set is used to describe the knowledge, and the reduction of redundant attributes can be used to simplify the knowledge. Finally, a complex decision table is reduced to the minimum conditional attribute decision table with no redundant attributes and the correct classification is ensured.

Rough Set Neural Network (RS-BP Hybrid Model).
e RS-BP hybrid model is a hybrid model combining artificial neural network and rough set theory. e training time is too long when the neural network and the sample are both large. is inherent disadvantage is one of the main factors which restricts the further practicability of the neural network. Although various algorithms for improving training speed have emerged, the problem is far from being completely solved. Simplifying the training sample set and eliminating redundant data are another way to improve training speed. e rough set is used to simplify the neural network in order to train the sample set, and the redundant data is eliminated under the premise of retaining important information. Simulation results show that the training speed is 4.77 times better, and good results are achieved [26]. e characteristic parameters of the vibration are extracted through the condition monitoring system and a field test in the fault diagnosis process of the hydroelectric generating set. In order to describe the vibration mode of the hydrogenerator set as much as possible, more units are set up to extract the characteristic quantities of these units, some of which are related and some of which are independent. ese feature parameters are often incomplete and redundant, which makes the modeling of subsequent fault diagnosis complex. e rough set theory can be used to reduce the knowledge of the training samples of neural network and get fewer attribute values while ensuring the goal of not missing important information.

Simulation and Pattern Recognition of the Shaft Orbit
Using Matlab to simulate the shaft orbits of the generator set when the fault occurs, formula (22) can be obtained: where ω is the angular frequency; A 1 , A 2 , α 1 ,α 2 , A 1 , A 2 , α 1 ,α 2 are the amplitudes and initial phases of the first and second frequency components of x(t) and y(t). By changing the 8 parameters A 1 , A 2 , α 1 ,α 2 , B 1 , B 2 , β 1 , β 2 , the shaft orbits of the hydroelectric unit are obtained, as shown in Figure 3. e 30 simulated shaft orbits on the left are training samples, and the 10 orbits on the right are test samples.
In this paper, Hu moments and Zernike moments are used to extract the features of shaft orbits. 7 features of the Hu moments are extracted and there are 6 improved Hu moments. Zernike moments can be used to construct high moments, and higher moments contain more image information.
In this paper, 49 features of Zernike moments can be obtained by choosing 0∼12 order Zernike moments. e Hu moments, improved Hu moments, and the data of the Zernike moments, after the rough set knowledge reduction is extracted from Figure 3, are shown in Tables 2-4. Tables 3  and 4 show the average values of the Hu moments and the improved Hu moments of typical shaft orbits. e Hu moments and the Zernike moments are chosen as input variables, and the type of shaft orbit is used as the output in order to establish a neural network. e shapes of the shaft orbits can be identified by training the samples. For 49 Zernike moments, the input vector node is 49, and the network structure is relatively complex, which will cause the training time to become longer. e rough set feature selection algorithm is used to reduce the input and remove redundant information. Finally, 13  In order to fully reflect the influence of different feature quantities on pattern recognition, different feature input parameters are used for the pattern recognition of shaft orbits. e first method uses 7 Hu moments as input variables. e second method uses 6 improved Hu moments as input variables. e third method takes the 49 moments as input variables. e fourth method uses the 13 moments eigenvalues reserved by the rough set feature selection algorithm as input variables. e neural network models of the four methods all adopt a three-layer BP neural network structure. e number of neurons in each input layer is as follows: 6, 7, 49, and 13. e hidden layer uses 10 neurons, and the output layer is 5 neurons, namely, five shaft orbits. e training step � 20000, the descent gradient � 0.1, and the error target � 0.001. e flow chart of the fourth method is shown in Figure 4. Input samples and test samples are shown in Figure 4. e data for the sample training are shown in Tables 1, 5, and 6 . e results of training and recognition for the four neural networks are shown in Table 7.
It can be seen from Table 7 that Zernike moments are better than Hu moments for pattern recognition, and the former have the advantages of a small calculation error and a high recognition rate. Calculating the effects of Model 3 and Model 4, both have the same recognition accuracies and convergence rates, but the latter significantly reduces the calculation step. We can see that using the rough set method to advance the information reduction pretreatment is effective, and it can increase the convergence speed of the effect. e combination of Zernike moments, the rough set, and three neural networks can efficiently and accurately identify the shaft orbits, while the neural network with the Zernike moments and the rough set preprocessing can achieve good

Conclusions
In this paper, the feature extraction and pattern recognition of the shaft orbit of a hydroelectric generating unit are studied. e main content and conclusions are as follows: (1) In terms of shaft orbits feature extraction and pattern recognition, the Zernike moment is better than the Hu moment and it has the advantages of a small calculation error and a high recognition rate. (2) A rough set neural network (RS-BP hybrid model) of shaft orbit recognition is established, which uses just 13 moment eigenvalues reserved by the rough set feature selection algorithm as input variables, has the same calculation error and recognition rate, and reduces the calculation time step. e simulation of the recognition of shaft orbits shows that the hybrid model has achieved good results in the identification of shaft orbits.

Data Availability
No data were used to support this study.

Conflicts of Interest
e authors declare that they have no conflicts of interest.