CN101799367A - Electromechanical device neural network failure trend prediction method - Google Patents

Abstract

The invention relates to an electromechanical device neural network failure trend prediction method, comprising the following steps: (1) obtain a section continuous vibration signal which is sensitive to the failure and is output by a measuring point sensor; (2) respectively carry out exceptional value elimination and missing data filling to the vibration data by a 3 sigma method and an interpolation method; (3) carry out a normalization process to a vibration data sequence; (4) calculate a vibration data sequence which is entropy-weighted according to the sequence which is carried out the normalization process; (5) carry out a time-weighted calculation to the vibration data sequence which is entropy-weighted by utilizing time weight due to the influence of time factor; (6) build a nonlinear dynamic recurrent neural network prediction model by utilizing the data sequence which is obtained by step (5) and determine a hidden layer optimal node number by utilizing a golden section method; (7) carry out normalization process to a trend prediction result and obtain a actual prediction result. A dynamic recurrent neural network model is adopted to carry out prediction in the invention, therefore, the failure prediction reliability is increased. The electromechanical device neural network failure trend prediction method can be widely applied to the failure prediction and analysis of all kinds of electromechanical devices.

Description

A kind of electromechanical device neural network failure trend prediction method

Technical field

The present invention relates to a kind of mechanical failure prediction method, particularly about a kind of electromechanical device neural network failure trend prediction method based on the weighting of information entropy weighted sum time factor.

Background technology

In the face of non-linear, the non-stationary dynamic problem of electromechanical equipment failure prediction, traditional Linearization Method effect is not good enough, and some essence to be non-linear dynamic system analysis disposal routes have application prospect aspect failure prediction.Neural net prediction method has self-learning function, and characteristics such as non-linear, non-locality, non-stationarity, can be by appropriate network layer and the hidden layer unit number selected, can approach arbitrarily the characteristic of nonlinear function and all-order derivative thereof continuously with arbitrary accuracy, thereby quilt is extensive in failure prediction.At present, the method that adopts neural network to carry out failure prediction is substantially: the time series data of equipment running status is input to the input layer of neural network successively, adopts general neural network to train and predict then.In this method, the network input value is equal on probability basically to the percentage contribution of network prediction of output value, and used neural network is static network mostly, is not suitable for the real-time identification of dynamic system.In actual applications, the hidden layer node number generally is the way that relies on tentative calculation in the network structure, but this method calculated amount is bigger, and is not easy to determine the quality of gained forecast model structure.

Summary of the invention

At the problems referred to above, the purpose of this invention is to provide that a kind of failure prediction reliability is higher, calculated amount is less and can be applicable to the electromechanical device neural network failure trend prediction method of dynamic system real-time identification.

For achieving the above object, the present invention takes following technical scheme: a kind of electromechanical device neural network failure trend prediction method, and its step is as follows: (1) obtains one section continuous shaking signal that certain measuring point sensor of fault sensitivity is exported by the remote monitoring diagnostic center; (2) utilize 3 σ methods and interpolation method that the vibration data that obtains is carried out abnormality value removing and missing data respectively and fill up, obtain vibration data sequence { X ₁..., X _n; (4) to vibration data sequence { X ₁..., X _nCarry out obtaining sequence { x after the normalized ₁..., x _n; (5) according to the sequence { x after the normalized ₁..., x _n, calculate based on information entropy weighting coefficient w _Si, and then calculate vibration data sequence { y after the entropy weighting ₁..., y _n; (6) because time factor influence, the vibration data sequence { y after utilizing time weight to the entropy weighting ₁..., y _nCarry out time weight and calculate after, obtain data sequence { z ₁..., z _n; (7) utilize data sequence { z ₁..., z _nMake up nonlinear Dynamical Recurrent Neural Networks forecast model, and adopt Fibonacci method to determine the optimum node number of hidden layer, so that definite neural network optimum prediction model structure, carrying out failure trend prediction, { u is predicted the outcome ₁..., u _m, m is the number of future position; (8) the trend prediction result is carried out anti-normalized, obtain the actual prediction result.

In the described step (5), described oscillating sequence obtains sequence { y after described entropy weighting ₁..., y _nBe: y _i=x _iw _SiI=1,2 ... n, in the formula, the information entropy weighting coefficient $w_{\mathrm{Si}}=\frac{E_i}{\max \left\{E_i\right\}},$ E wherein _iFor reflecting that vibration data carries the information entropy of quantity of information, E _i=-p _iLog ₂p _i(i=1,2 ..., n), p _iFor each vibration data provides the probability of information, $p_i=\frac{x_i}{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{x}_i}(i=\mathrm{1,2},\·\·\·,n).$

In the described step (6), described oscillating sequence obtains data sequence { z after described time weight ₁..., z _nBe: z _i=y _iw _NiI=1,2 ... n, in the formula, w _NiBe the time weight coefficient, it is: $w_{\mathrm{Ni}}=\left\{\begin{array}{cc}\frac{{\mathrm{\&alpha;}}^{\frac{i-1}{k}}}{w}& 1\&le;i\&le;k\\ \frac{{\mathrm{\&beta;}}^{\frac{i-k}{n-k}}}{w}& k<i\&le;n\end{array}\right.,$ Wherein $w=\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}^{\frac{i-1}{k}}+\underset{i=k+1}{\overset{n}{\mathrm{\&Sigma;}}}{\mathrm{\&beta;}}^{\frac{i-k}{n-k}},\mathrm{\&beta;}>\mathrm{\&alpha;}>1.$

In the described step (7), definite method of the optimum node number of described hidden layer is as follows: 1. according to formula n ₁=log ₂N and $n_1=\sqrt{n+m}+\mathrm{\&alpha;}$ Definite respectively hidden layer node is counted the minimum value n_min and the maximal value n_max in the interval at place, and wherein, n is the input neuron number, and m is the output neuron number, and α is the constant between [1,10]; 2. calculate hidden layer node and count the error mean square E (n_min) and the E (n_max) at interval endpoint place, place; 3. the size of end points place error relatively when E (n_min)≤E (n_max), is counted the right side of minimum value between the location and is carried out golden search from hidden layer node, a search point i=n_min+0.618 (n_max-n_min) then is provided with n_max=i; Otherwise, counting between the location peaked left side from hidden layer node and carry out golden search, a search point i=n_max-0.618 (n_max-n_min) then is provided with n_min=i; 4. judge according to the square error lower limit of having set whether gold point satisfies error requirements, when satisfying error requirements, best hidden layer node number is search point i; Otherwise, enter step and 2. circulate.

The present invention is owing to take above technical scheme, it has the following advantages: 1, the present invention is owing to adopt dynamic neural network model to predict, the vibration signal that utilizes information entropy that the characterization device of being imported is moved, carry out information fusion to obtain consistance description to the electromechanical equipment running status, foundation is based on the weight matrix of information entropy, consider the influence of time factor simultaneously to the network input, set up the dynamic neural network forecast model of new breath weighting, in the prediction of dynamic neural network, adopt Fibonacci method to determine the number of hidden layer node number, therefore the predict device running status has improved the failure prediction reliability efficiently.2, the present invention is owing to merge carry out information entropy from the raw information of sensor, the information entropy weighting coefficient that obtains, and be weighted according to information percentage contribution of quantity of information on information entropy, therefore can fully effectively utilize sensor information and predict.3, the present invention is owing to consider the time factor influence, and then sets up new breath weighting matrix, therefore determined the percentage contribution of network input value to network prediction of output value.4, therefore the present invention has simplified calculated amount greatly owing to adopt the Fibonacci method search to determine the hidden layer node number, and determines the quality of gained forecast model structure easily.5, therefore the present invention makes forecast model have the dynamic self-adapting characteristics owing to utilize the dynamic neural network forecast model be in the nature non-linear topological structure, can the variation of adaptation condition conditions and environment etc., realized nonlinear failure prediction.The present invention can be widely used in the various electromechanical equipment Fault Forecast Analysis.

Description of drawings

Fig. 1 is an overall flow synoptic diagram of the present invention,

Fig. 2 of the present inventionly determines that based on Fibonacci method the optimum node of hidden layer counts schematic flow sheet.

Embodiment

Below in conjunction with drawings and Examples the present invention is described in detail.

As shown in Figure 1, the present invention adopts dynamic neural network model to predict, utilizes the information entropy weighting that the vibration signal of representative equipment operation of input is carried out information fusion, and then obtains the consistance of electromechanical equipment running status is described.Foundation is based on the information entropy weighting time, consider the influence of time factor, set up the dynamic neural network forecast model of new breath weighting, in the prediction of dynamic neural network the network input, adopt Fibonacci method to determine the optimum node number of hidden layer, to realize predict device running status efficiently.Its concrete steps are as follows:

(1) obtains one section continuous shaking signal data that certain measuring point sensor of fault sensitivity is exported by existing remote monitoring diagnostic center;

(2) utilize 3 σ methods and interpolation method that the vibration data that obtains is carried out abnormality value removing and missing data respectively and fill up, obtain vibration data sequence { X ₁..., X _n;

(4) to vibration data sequence { X ₁..., X _nCarry out obtaining sequence { x after the normalized ₁..., x _n, that is: $x_i=\frac{X_i-\stackrel{\&OverBar;}{X}}{\stackrel{\&OverBar;}{X}},i=\mathrm{1,2},\&CenterDot;\&CenterDot;\&CenterDot;,n,$ Wherein $\stackrel{\&OverBar;}{X}=\frac{1}{n}\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{X}_i,---\left(1\right)$

(5) according to the sequence { x after the normalized ₁..., x _n, calculate based on information entropy weighting coefficient w _Si, through obtaining vibration data sequence { y after the entropy weighting ₁..., y _nBe:

y _i＝x _iw _Si?i＝1，2，…n， (2)

In the following formula, the information entropy weighting coefficient $w_{\mathrm{Si}}=\frac{E_i}{\max \left\{E_i\right\}},$ E wherein _iFor reflecting that vibration data carries the information entropy of quantity of information, E _i=-p _iLog ₂p _i(i=1,2 ..., n), p _iFor each vibration data provides the probability of information,

$p_i=\frac{x_i}{\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{x}_i}(i=\mathrm{1,2},\&CenterDot;\&CenterDot;\&CenterDot;,n);$

(6) owing to the newness degree difference of time series data with vibration data, can be also inequality to the contribution of prediction of output value, so the vibration data sequence { y after using the time weighting to the entropy weighting ₁..., y _nCarry out time weight and calculate, and then can embody that new legacy data obtains data sequence { z to the contribution of predicted value in the sample data after time weight ₁..., z _nBe:

z _i＝y _iw _Ni?i＝1，2，…n (3)

In the following formula, w _NiBe the time weight coefficient, it is:

$w_{\mathrm{Ni}}=\left\{\begin{array}{cc}\frac{{\mathrm{\&alpha;}}^{\frac{i-1}{k}}}{w}& 1\&le;i\&le;k\\ \frac{{\mathrm{\&beta;}}^{\frac{i-k}{n-k}}}{w}& k<i\&le;n\end{array}\right.,$ Wherein $w=\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}^{\frac{i-1}{k}}+\underset{i=k+1}{\overset{n}{\mathrm{\&Sigma;}}}{\mathrm{\&beta;}}^{\frac{i-k}{n-k}},\mathrm{\&beta;}>\mathrm{\&alpha;}>1;$

(7) utilize data sequence { z ₁..., z _nThe nonlinear Dynamical Recurrent Neural Networks forecast model of structure, and in the dynamic neural network forecast model, adopt Fibonacci method to determine the optimum node number of hidden layer, and then definite neural network optimum prediction model structure, carry out failure trend prediction, { u is predicted the outcome ₁..., u _m, m is the number of future position;

(8) the trend prediction result is carried out anti-normalized, obtain actual prediction result { v ₁..., v _m, v predicts the outcome _iFor:

$v_i=\stackrel{\&OverBar;}{X}+u_i\&CenterDot;\stackrel{\&OverBar;}{X}.---\left(4\right)$

As shown in Figure 2, in the above-mentioned steps (7), definite method of the optimum node number of hidden layer is as follows:

1. formula n rule of thumb ₁=log ₂N determines that hidden layer node counts the minimum value n_min in the interval at place, rule of thumb formula $n_1=\sqrt{n+m}+\mathrm{\&alpha;}$ Determine that hidden layer node counts the maximal value n_max in the interval at place, and then determined that hidden layer node is counted and be [n_min, n_max] between the location; Wherein, n is the input neuron number, and m is the output neuron number, and α is the constant between [1,10];

2. calculate hidden layer node and count square error E at interval endpoint place, place (n_min) and E (n_max);

3. the size of end points place error relatively when E (n_min)≤E (n_max), is counted the right side of minimum value between the location and is carried out golden search from hidden layer node, a search point i=n_min+0.618 (n_max-n_min) then is provided with n_max=i; Otherwise, counting between the location peaked left side from hidden layer node and carry out golden search, a search point i=n_max-0.618 (n_max-n_min) then is provided with n_min=i;

4. judge according to the square error lower limit of having set whether gold point satisfies error requirements, when satisfying error requirements, best hidden layer node number is search point i; Otherwise, enter step and 2. circulate.

The various embodiments described above only are preferred implementations of the present invention, and are every based on the changes and improvements on the technical solution of the present invention in the present technique field, should not get rid of outside protection scope of the present invention.

Claims (4)

1. electromechanical device neural network failure trend prediction method, its step is as follows:

(1) obtains one section continuous shaking signal that certain measuring point sensor of fault sensitivity is exported by the remote monitoring diagnostic center;

(2) utilize 3 σ methods and interpolation method that the vibration data that obtains is carried out abnormality value removing and missing data respectively and fill up, obtain vibration data sequence { X ₁..., X _n;

(4) to vibration data sequence { X ₁..., X _nCarry out obtaining sequence { x after the normalized ₁..., x _n;

(5) according to the sequence { x after the normalized ₁..., x _n, calculate based on information entropy weighting coefficient w _Si, and then calculate vibration data sequence { y after the entropy weighting ₁..., y _n;

(6) because time factor influence, the vibration data sequence { y after utilizing time weight to the entropy weighting ₁..., y _nCarry out time weight and calculate after, obtain data sequence { z ₁..., z _n;

(7) utilize data sequence { z ₁..., z _nMake up nonlinear Dynamical Recurrent Neural Networks forecast model, and adopt Fibonacci method to determine the optimum node number of hidden layer, so that definite neural network optimum prediction model structure, carrying out failure trend prediction, { u is predicted the outcome ₁..., u _m, m is the number of future position;

(8) the trend prediction result is carried out anti-normalized, obtain the actual prediction result.

2. a kind of electromechanical device neural network failure trend prediction method as claimed in claim 1 is characterized in that: in the described step (5), described oscillating sequence obtains sequence { y after described entropy weighting ₁..., y _nBe:

y _i＝x _iw _Si?i＝1，2，…n，

In the formula, the information entropy weighting coefficient $w_{\mathrm{Si}}=\frac{E_i}{\max \left\{E_i\right\}},$ E wherein _iFor reflecting that vibration data carries the information entropy of quantity of information, E _i=-p _iLog ₂p _i(i=1,2 ..., n), p _iFor each vibration data provides the probability of information,

$p_i=\frac{x_i}{\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{x}_i}(i=\mathrm{1,2},\&CenterDot;\&CenterDot;\&CenterDot;,n).$

3. a kind of electromechanical device neural network failure trend prediction method as claimed in claim 1 is characterized in that: in the described step (6), described oscillating sequence obtains data sequence { z after described time weight ₁..., z _nBe:

z _i＝y _iw _Ni?i＝1，2，…n，

In the formula, w _NiBe the time weight coefficient, it is:

$w_{\mathrm{Ni}}=\left\{\begin{array}{cc}\frac{{\mathrm{\&alpha;}}^{\frac{i-1}{k}}}{w}& 1\&le;i\&le;k\\ \frac{{\mathrm{\&beta;}}^{\frac{i-k}{n-k}}}{w}& k<i\&le;n\end{array}\right.,$ Wherein $w=\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}^{\frac{i-1}{k}}+\underset{i=k+1}{\overset{n}{\mathrm{\&Sigma;}}}{\mathrm{\&beta;}}^{\frac{i-k}{n-k}},\mathrm{\&beta;}>\mathrm{\&alpha;}>1.$

4. a kind of electromechanical device neural network failure trend prediction method as claimed in claim 1 is characterized in that: in the described step (7), definite method of the optimum node number of described hidden layer is as follows:

1. according to formula n ₁=log ₂N and $n_1=\sqrt{n+m}+\mathrm{\&alpha;}$ Definite respectively hidden layer node is counted the minimum value n_min and the maximal value n_max in the interval at place, and wherein, n is the input neuron number, and m is the output neuron number, and α is the constant between [1,10];

2. calculate hidden layer node and count the error mean square E (n_min) and the E (n_max) at interval endpoint place, place;

3. the size of end points place error relatively when E (n_min)≤E (n_max), is counted the right side of minimum value between the location and is carried out golden search from hidden layer node, a search point i=n_min+0.618 (n_max-n_min) then is provided with n_max=i; Otherwise, counting between the location peaked left side from hidden layer node and carry out golden search, a search point i=n_max-0.618 (n_max-n_min) then is provided with n_min=i;

4. judge according to the square error lower limit of having set whether gold point satisfies error requirements, when satisfying error requirements, best hidden layer node number is search point i; Otherwise, enter step and 2. circulate.

Images