Improved single-layer forward neural network for fault diagnosis of nuclear power million-kilowatt turbine generator

With the rapid development of China's economic construction and nuclear energy utilization technology, it is very necessary to develop nuclear power turbine generators with more reliable performance and larger single-unit capacity, which is the main research direction to cater to this development.

On the other hand, in recent years, large-scale steam turbine generators at home and abroad often have accidents during operation. One of the main reasons for these accidents is the lack of reasonable and effective methods for dynamic analysis and condition monitoring of these power generation equipment. For example, most of the conventional analysis and monitoring methods are based on the analysis of stationary random vibration signals. This is obviously not enough, because the rotational speed of large steam turbine generator sets is fluctuating during operation, especially in the accidents that occur. Many of them are in the stage of speed up and down speed of the unit. Therefore, it is necessary to use the method of non-stationary random vibration signal analysis to study the dynamic characteristics of nuclear power million-kilowatt turbine generators, and develop such turbine generators for China. Lay the foundation.

For non-stationary random vibration signals, so far, there is no unified and perfect analysis method, and the commonly used sometimes variable AR model, Wigner-Ville distribution, segmentation balance processing based on segmentation criterion and artificial neural network. The author uses KL decomposition to orthogonally decompose the non-stationary random vibration signal, so that the non-stationary random signal is effectively decomposed in the high-dimensional orthogonal space in the original low-dimensional space, and then decomposed. The high-dimensional orthogonal space signal is used as the input mode of the single-layer forward neural network. Finally, a recursive association learning is performed by using recursive association in the single-layer forward network, thus overcoming the conventional single-layer forward network is only linearly separable. In the process of learning, it is necessary to repeatedly submit a series of input modes to be memorized to adjust the deficiency of weights. It also speeds up the classification.

On the unique, self-designed nuclear power million-kilowatt turbo-generator shafting simulation test bench, this paper collected non-stationary random vibration signals in various states and applied these signals to the C language by applying the above method. The improved single-layer forward neural network is classified in computer programs. The results show that the proposed method is very effective.

1 Improved single-layer forward neural network based on the above, here, K-L decomposition is performed on the non-stationary random vibration signal.

The basic idea of ​​L decomposition is to a non-stationary random vibration signal x((), sampled discrete sample vector whose component is x(k), k=1−K and its correlation function is Rx(h, K2), where k, k2 take values ​​in 1-K. The x(k) is orthogonally decomposed into: m-1 and then output by the output layer. A threshold is taken as zero in the following calculation.

For a stationary random vibration system, we collect the non-stationary random vibration signal, and decompose the eigenvalues ​​from K-L to transform the eigenvalues ​​into m-space, which can establish the input and output of the improved single-layer forward neural network. The pattern pair can be used to associate the input with the output of a Lenovo matrix Y. That is, equation (6) is equivalent to decomposition. There are many methods for decomposition. In this paper, the QR full rank decomposition method is used: upper R is the upper triangular matrix Q is an orthogonal matrix. Using the GranrSchmidt orthogonalization method, the calculation formulas of R and Q can be derived. They are: the is column vector of 0 (=1(1)m), which is the largest specification orthogonal group of the A column space, and A+1 represents A / +1 column vector of A.

2 The example is a multi-functional shafting simulation test bench developed for the study of the dynamic characteristics of nuclear power million-kilowatt steam turbine generators (has applied for national patents and accepted). The test bench has the following characteristics: it can be rubbed, not right. In the middle, the change of the degree of the curve and the looseness of the support; the support form can be simply supported, semi-simplified and fixed; the horizontal support stiffness and the vertical support stiffness can be steplessly adjusted and combined within a certain range; YA The output of the network is obtained by inputting a pattern to the neural network. To be built in the test system, the pick-up meter with the Danish-made bk4366-type accelerating stand-up Lenovo matrix only needs to be operated as follows. The sensor C is fixed on the generator excitation support. X-signal amplification uses =XA+ '(8) BK2635 type charge Amplifier, data acquisition and analysis using CRAS soft (, A's inverse field generalized inverse to represent the modified single-layer forward neural network table 1 improved single-layer forward direction written by Gan and the author in C language Neural network classification result working condition input sample expected output actual output normal XC1) Table 2 classification time iteration number comparison running condition iteration number conventional single layer forward neural network improved single layer forward neural network normal 51202959 rubbing 61763778 Misaligned 68503031 Yangdu curve change 63223802 loose 70404095 network" fault diagnosis software.

In the test system, the rotor system is a three-axis six-turn six-support system that simulates a system of exciters, generators, and three rotors of a steam turbine (low-pressure stage) in a nuclear-power million-kilowatt-class steam turbine generator set. When the system is speeding up, the five conditions of normal simulation, rubbing, misalignment, change of the curve and looseness of the support (bolt) are in the same time course (take 10s in the test) and the same initial and final speed ( In the test, the non-stationary random vibration signal of the rotor system under various working conditions was repeatedly measured from 500 rpm to 3000 rpm. After the dispersion, 6 samples were taken for various working conditions, and 80 components were taken in each sample vector. Learning and classification is then performed by the improved single layer forward neural network shown. All results are shown in Table 1, where the expected output of the normal condition sample is 1, and the expected output of the fault condition sample is zero. It can be seen from Table 1 that the actual output of the neural network is close to 1 under normal operating conditions, and the output of various fault conditions is close to zero, so the "improved single-layer forward neural network" proposed in this paper has a very non-stationary running state. Effective classification function.

In Table 2, the number of times when a conventional single-layer forward neural network is classified with the improved single-layer forward neural network herein is compared, and the latter is significantly less than the number of iterations at the time of classification.

3 Conclusions According to the nuclear power million kilowatt-class turbo generator rotor system, a multi-function shafting simulation test bench is designed. On this test bench, various fault conditions such as rubbing, misalignment, and change of the curve and looseness are simulated. And the non-stationary random vibration signal collected under the fault condition is input as the input parameter to identify the "improved single-layer forward neural network", so that the single-layer forward network does not fall into the local minimum point. The advantages are effectively improved the learning efficiency of the conventional single-layer forward network, and the number of iterations in the classification is greatly reduced. The test and analysis of the rotor system and the calculation results show that the practicality of the "improved single-layer forward neural network" is satisfactory.

Combined Rice Milling Machine

Combined Rice Milling Machine,Mini Rice Mill Machine,Portable Rice Milling Machine,Automatic Rice Mill Machine

HUNAN JINSONG MACHINERY CO., LTD , https://www.sheller.nl