CN106934126A - Component of machine health indicator building method based on Recognition with Recurrent Neural Network fusion - Google Patents

Abstract

A kind of component of machine health indicator building method based on Recognition with Recurrent Neural Network fusion, obtains component of machine vibration signal first, is calculated vibration signal temporal signatures sequence and frequency domain character sequence；Similarity feature is calculated according to temporal signatures sequence and frequency domain character sequence；Three layers of wavelet package transforms are carried out to vibration signal, frequency band energy is obtained and is compared feature；The sensitive features collection of component of machine degenerative process is filtered out using the comprehensive evaluation index of feature, is used to train Recognition with Recurrent Neural Network；New component of machine health indicator RNN HI can be obtained by sensitive features collection and the Recognition with Recurrent Neural Network for training, the present invention has fully excavated the degradation information in component of machine vibration signal, the determination for being not only convenient for failure threshold and the precision that improve life prediction using similarity feature and Recognition with Recurrent Neural Network.

Description

Mechanical part health index construction method based on recurrent neural network fusion

Technical Field

The invention belongs to the technical field of mechanical equipment state monitoring and service life prediction, and particularly relates to a mechanical part health index construction method based on recurrent neural network fusion.

Background

In modern industrial production, mechanical equipment is more and more complicated, the connection between mechanical parts is more and more compact, and because the operating environment is complicated and changeable, mechanical parts are easy to have faults of different degrees, so that the whole mechanical equipment cannot work, and even economic loss and serious accidents which are difficult to estimate are caused. Therefore, the working state is effectively identified through the state monitoring of the mechanical parts, the accurate prediction of the residual service life is further realized, and the optimal maintenance time of the mechanical equipment is determined according to the effective identification, so that the mechanical equipment is effectively prevented before the mechanical equipment breaks down, and the service life of the mechanical parts is prolonged. The structure of the health index is a crucial link in the state monitoring and life prediction of mechanical parts, and the quality of the health index directly influences the accuracy of the life prediction in the later period.

The traditional health index structure is mainly extracted from an original signal by using a modern signal processing method, and is closely related to the degradation process of mechanical parts in a physical sense, such as root mean square value, peak-to-peak value and the like. Generally, the value range of the health index is greatly influenced by the working condition, so that the difficulty of determining the failure threshold of the mechanical part under the variable working condition is increased; a single health indicator is often sensitive to a particular degradation stage of a mechanical part and does not exhibit a good trend over the life cycle. Meanwhile, as the vibration signals of the mechanical parts have complex variability, the recession process of the mechanical parts is difficult to effectively reflect by directly using the traditional health indexes. The above-mentioned disadvantages may cause early failure to be difficult to find in time during the monitoring process of the state of the mechanical component, and then affect the accuracy of the life prediction.

Disclosure of Invention

In order to overcome the defects of the prior art, the invention aims to provide a mechanical part health index construction method based on the fusion of a recurrent neural network, which fuses a plurality of health indexes into one through an information fusion method, fully excavates state information in a mechanical part vibration signal, realizes accurate expression of a degradation process, and can control the value of the health index within a constant range to facilitate the determination of a failure threshold value.

In order to achieve the purpose, the invention adopts the technical scheme that:

a mechanical part health index construction method based on recurrent neural network fusion comprises the following steps:

1) acquiring a vibration signal of the mechanical part, calculating a time domain characteristic of the vibration signal, and obtaining a time domain characteristic sequence f1 of the mechanical part (f 1)¹,f1²,...,f1¹¹) Wherein f1¹，f1²，…，f1¹¹Mean value, root mean square value, peak-to-peak value, variance, entropy, kurtosis index, skewness index, peak index, wave form factor, pulse index and margin index; fourier transform is carried out on the vibration signal to obtain a global spectrum characteristic sequence f2, and the global spectrum characteristic sequence is divided averagely to obtain four sub-spectrum characteristic sequences f3, f4, f5 and f6 from low frequency to high frequency;

2) sequentially carrying the six characteristic sequences obtained in the step 1) into a formula (1) to calculate and obtain the similarity characteristics of the mechanical parts:

wherein K is the length of the characteristic sequence,andrespectively, the mean values of the characteristic sequences at the initial time and at the time t,and f_t ⁱThe obtained similarity features are sequentially marked as F1, F2, F3, F4, F5 and F6, wherein the values are the ith component of the feature sequence at the initial time and the t time respectively;

3) three-layer wavelet packet transformation is carried out on the vibration signal, and the energy ratio characteristics of eight frequency bands are obtained through calculation according to the formula (2):

wherein E_i(i 1, 2.., 8) is the energy value of each frequency band of the wavelet packet,the obtained band energy bit characteristics are sequentially marked as F7, F8, F9, F10, F11, F12, F13 and F14 for the energy sum of all wavelet packet bands;

4) and (3) respectively calculating comprehensive evaluation indexes of six similarity characteristics and eight frequency band energy ratio characteristics F1-F14 of the mechanical parts according to the formula (3):

in the formula:

wherein, F_lAnd t_lRespectively the characteristic value and time of the ith observation point,andthe method comprises the steps that the average values of a characteristic sequence and a time sequence in a full life cycle are respectively, L is the length of the characteristic sequence in the full life cycle, (x) is a unit step function, Corr measures the linear correlation degree of the characteristic sequence and the time sequence, and Mon measures the monotonicity of the characteristic sequence;

5) standardizing the comprehensive evaluation index Cri of the fourteen calculated characteristics to obtain Cri ', and selecting the characteristics with Cri' being more than 0.5 as a sensitive characteristic set in the degradation process of the mechanical parts;

6) utilizing life-cycle data setsTraining a recurrent neural network model, where x_t∈R^N×1Set of sensitivity features for time t, y_t∈[0,1]Is the actual degradation rate of the mechanical parts at the corresponding moment, y_t0 represents a mechanical part in a fully healthy state, y_t1 represents that the cyclic neural network is in a complete fault state, and the cost function in the training process of the cyclic neural network is shown as the formula (6):

wherein,as an output result of the recurrent neural network model, y_tIs the true degradation rate;

7) and in the testing stage, the sensitivity characteristic set obtained by calculation is used as the input of the trained recurrent neural network to obtain a new mechanical part health index RNN-HI.

The invention has the beneficial effects that:

the method constructs a new similarity characteristic used for representing the degree of the mechanical parts deviating from the normal state, screens out the sensitive characteristic set of the mechanical parts in the degradation process through comprehensive evaluation indexes of the characteristic, and obtains a new mechanical parts health index RNN-HI by fully fusing and extracting state information in the sensitive characteristic set through a recurrent neural network, so that the degradation trend of the mechanical parts can be well reflected, the obtained health index is used for life prediction, and the accuracy of the prediction result is improved.

Drawings

FIG. 1 is a flow chart of the method of the present invention.

FIG. 2 is a schematic representation of similarity features derived from a sequence of spectral features.

FIG. 3 is a schematic structural diagram of a long-term and short-term memory model.

Fig. 4 shows life-cycle vibration signals of the rolling bearing1_1 and the rolling bearing2_ 6.

Fig. 5 shows the normalized overall evaluation index.

Fig. 6 shows the life prediction result of the rolling bearing2_ 6.

Detailed Description

The invention is further elucidated with reference to the figures and embodiments.

Referring to fig. 1, a method for constructing a mechanical part health index based on recurrent neural network fusion includes the following steps:

1) acquiring a vibration signal of the mechanical part, calculating the time domain characteristic of the vibration signal of the mechanical part, and obtaining a time domain characteristic sequence f1 of the mechanical part (f 1)¹,f1²,...,f1¹¹) Wherein f1¹，f1²，…，f1¹¹Mean value, root mean square value, peak-to-peak value, variance, entropy, kurtosis index, skewness index, waveform factor, pulse index and margin index; fourier transform is carried out on the vibration signal to obtain a global spectrum characteristic sequence f2, and the global spectrum characteristic sequence is divided averagely to obtain four sub-spectrum characteristic sequences f3, f4, f5 and f6 from low frequency to high frequency;

2) sequentially carrying the six characteristic sequences obtained in the step 1) into a formula (1) to calculate and obtain the similarity characteristics of the mechanical parts:

wherein K is the length of the characteristic sequence,andrespectively, the mean values of the characteristic sequences at the initial time and at the time t,and f_t ⁱOf the i-th component of the signature sequence at the initial and t-times, respectivelyThe obtained similarity characteristics are sequentially recorded as F1, F2, F3, F4, F5 and F6, if the states of the mechanical parts are close at two moments, the similarity characteristics are approximate to 1, otherwise, the similarity characteristics are approximate to 0, and FIG. 2 is a schematic diagram of the similarity characteristics obtained by a spectrum characteristic sequence;

3) three-layer wavelet packet transformation is carried out on the vibration signal, and the energy ratio characteristics of eight frequency bands are obtained through calculation according to the formula (2):

wherein E_i(i 1, 2.., 8) is the energy value of each frequency band of the wavelet packet,the obtained band energy bit characteristics are sequentially marked as F7, F8, F9, F10, F11, F12, F13 and F14 for the energy sum of all wavelet packet bands;

4) and (3) respectively calculating comprehensive evaluation indexes of six similarity characteristics and eight frequency band energy ratio characteristics F1-F14 of the mechanical parts according to the formula (3):

in the formula:

wherein, F_lAnd t_lRespectively the characteristic value and time of the ith observation point,andthe method comprises the steps that the average values of a characteristic sequence and a time sequence in a full life cycle are respectively, L is the length of the characteristic sequence in the full life cycle, (x) is a unit step function, Corr measures the linear correlation degree of the characteristic sequence and the time sequence, and Mon measures the monotonicity of the characteristic sequence;

5) standardizing the comprehensive evaluation index Cri of the fourteen calculated characteristics to obtain Cri ', and selecting the characteristics with Cri' being more than 0.5 as a sensitive characteristic set in the degradation process of the mechanical parts;

6) utilizing life-cycle data setsTraining a recurrent neural network model, where x_t∈R^N×1Set of sensitivity features for time t, y_t∈[0,1]Is the actual degradation rate of the mechanical parts at the corresponding moment, y_t0 represents a mechanical part in a fully healthy state, y_t1 represents that the cyclic neural network is in a complete fault state, and the cost function in the training process of the cyclic neural network is shown as the formula (6):

wherein,as an output result of the recurrent neural network model, y_tFor the real degradation rate, an improved recurrent neural network-long-short-term memory model is used, the problem of gradient explosion and disappearance in the recurrent neural network is solved, fig. 3 is a schematic structural diagram of the long-short-term memory model, and the biggest characteristic is that valve nodes of each layer are added on the basis of the recurrent neural network structure and are used for judging whether information in a memory unit is added into the calculation of the layer or not, and the information is given to the layer according to the calculation sequenceAnd (3) calculating a formula of each part:

g_t＝φ(w_gxx_t+w_ghh_t-1+b_g) (7)

i_t＝σ(w_ixx_t+w_ihh_t-1+b_i) (8)

f_t＝σ(w_fxx_t+w_fhh_t-1+b_f)(9)

o_t＝σ(w_oxx_t+w_ohh_t-1+b_o) (10)

wherein, w_gx、w_ix,、w_fxAnd w_oxInput layers x at time t_tAnd an output layer h_tWeight value of between, w_gh、w_ih、w_fhAnd w_ohIs the weight between t and t-1 times of the hidden layer, b_g、b_i、b_fAnd b_oIs a deviation, h_t-1Output value of hidden layer at previous time g_t,，i_t，f_tAnd o_tOutput values, s, of input node, input gate, forget gate and output gate, respectively_tAnd s_t-1The values of the internal states at time t and the previous time, respectively, σ (x) is the gate activation function, φ (x) is the input activation function, h (x) is the output activation function;

7) and in the testing stage, the sensitivity characteristic set obtained by calculation is used as the input of the trained recurrent neural network to obtain a new mechanical part health index RNN-HI.

In order to further prove the effectiveness of the mechanical part health index construction method based on the recurrent neural network fusion, the experimental data of the accelerated life of the rolling bearing of the PRONOSTIA test bed is used for verification.

The PRONOSTIA test bed provides life cycle data of the rolling bearing from a normal state to complete failure, the rolling bearing used in the test is not provided with an initial fault, pressure is provided for the outer ring of the rolling bearing through the air cylinder so that the whole degradation process can be accelerated and completed within a few hours, each degraded rolling bearing can contain any fault mode, and the acceleration sensor is used for collecting data in the vertical direction and the horizontal direction of the outer ring of the rolling bearing. In the experimental process, the sampling frequency is 25600Hz, the number of sampling points is 2560, the sampling time is 0.1s each time, and the time interval of two times of sampling is 10 s. The three groups of experimental data respectively correspond to different rotating speeds and loads, the rotating speeds of bearing1_ 1-bearing 1_7 are 1800rpm, and the loads are 4000N; the rotation speed of bearing2_ 1-bearing 2_7 is 1650rpm, and the load is 4200N; the rotation speed of bearing3_ 1-bearing 3_3 is 1650rpm, the load is 4200N, and each group of experimental data is divided into a training set and a test set. Fig. 4 shows life-cycle vibration signals of the rolling bearing1_1 for training and the rolling bearing2_6 for testing.

And (3) carrying out feature extraction on the rolling bearing vibration signals of the training set to obtain feature sets F1-F14, calculating corresponding comprehensive evaluation indexes, wherein the normalized comprehensive evaluation index Cri' is shown in FIG. 5. It can be seen that the overall evaluation index after F1, F2, F3, F9, F10, F11, F12 and F14 standardization is greater than 0.5, which should be taken as a sensitive feature set, wherein F1, F2 and F3 are the proposed similarity features, and prove the effectiveness of the similarity features in describing the degeneration process. Sensitive feature sets of training set data are extracted to train a recurrent neural network to obtain corresponding health indexes RNN-HI, and for comparison, health indexes SOM-HI based on a self-organizing neural network proposed by Huang are extracted, wherein the comprehensive evaluation index of RNN-HI is 0.7474, the comprehensive evaluation index of SOM-HI is 0.6877, and the two have good effects on describing the degradation trend of a rolling bearing, however, the SOM-HI has no constant failure threshold under three working conditions, the failure threshold of RNN-HI is approximate to 1, and the effect of RNN-HI is better than that of SOM-HI.

In order to verify the effectiveness of RNN-HI in the life prediction of the rolling bearing, the life prediction is carried out by using a dual-exponential life prediction model based on particle filtering, wherein the dual-exponential life prediction model is shown as the following formula:

Y＝ae^bt+ce^dt

in the formula, Y is a rolling bearing health index RNN-HI, t is time, a, b, c and d are model parameters, and a particle filter method is used for updating the parameters. As shown in fig. 6, taking the test rolling bearing2_6 as an example, the predicted remaining life median 1470s and the confidence interval of the confidence level of 0.95 are [720s,1690s ].

The life prediction results of all rolling bearings in the test set are shown in table 1, and the life prediction precision is measured by the average percentage error:

wherein, ActRIL_iAnd RUL_iThe actual remaining life and predicted remaining life of the ith test bearing, respectively. Meanwhile, by comparison, the SOM-HI is used for predicting the residual life of the rolling bearing, corresponding prediction results are given in the table 1, the average percentage error of the RNN-HI is 32.48%, the average percentage error of the SOM-HI is 53.24%, and the mechanical part health index construction method based on the recurrent neural network fusion has higher precision in the residual life prediction.

TABLE 1

Claims (1)

1. A mechanical part health index construction method based on recurrent neural network fusion is characterized by comprising the following steps:

1) acquiring a vibration signal of the mechanical part, calculating a time domain characteristic of the vibration signal, and obtaining a time domain characteristic sequence f1 of the mechanical part (f 1)¹,f1²,...,f1¹¹) Wherein f1¹，f1²，…，f1¹¹Mean value, root mean square value, peak-to-peak value, variance, entropy, kurtosis index, skewness index, peak index, wave form factor, pulse index and margin index; to pairFourier transformation is carried out on the vibration signals to obtain a global spectrum characteristic sequence f2, and the global spectrum characteristic sequence is divided averagely to obtain four sub-spectrum characteristic sequences f3, f4, f5 and f6 from low frequency to high frequency;

2) sequentially carrying the six characteristic sequences obtained in the step 1) into a formula (1) to calculate and obtain the similarity characteristics of the mechanical parts:

$RS=\frac{\left|\underset{i=1}{\overset{K}{\Σ}}(f_0^i-{\tilde{f}}_0)(f_t^i-{\tilde{f}}_t)\right|}{\sqrt{\underset{i=1}{\overset{K}{\Σ}}{(f_0^i-{\tilde{f}}_0)}^2\underset{i=1}{\overset{K}{\Σ}}{(f_t^i-{\tilde{f}}_t)}^2}}---\left(1\right)$

wherein K is the length of the characteristic sequence,andrespectively, the mean values of the characteristic sequences at the initial time and at the time t,andthe obtained similarity features are sequentially marked as F1, F2, F3, F4, F5 and F6, wherein the values are the ith component of the feature sequence at the initial time and the t time respectively;

3) three-layer wavelet packet transformation is carried out on the vibration signal, and the energy ratio characteristics of eight frequency bands are obtained through calculation according to the formula (2):

${\mathrm{ER}}_i=\frac{E_i}{E}---\left(2\right)$

wherein E_i(i 1, 2.., 8) is the energy value of each frequency band of the wavelet packet,the obtained band energy bit characteristics are sequentially marked as F7, F8, F9, F10, F11, F12, F13 and F14 for the energy sum of all wavelet packet bands;

4) and (3) respectively calculating comprehensive evaluation indexes of six similarity characteristics and eight frequency band energy ratio characteristics F1-F14 of the mechanical parts according to the formula (3):

$Cri=\frac{Corr+Mon}{2}---\left(3\right)$

in the formula:

$Corr=\frac{\left|\underset{l=1}{\overset{L}{\Σ}}\left(F_l-\tilde{F}\right)\left(t_l-\tilde{t}\right)\right|}{\sqrt{\underset{l=1}{\overset{L}{\Σ}}{(F_l-\tilde{F})}^2\underset{l=1}{\overset{L}{\Σ}}{(t_l-\tilde{t})}^2}}---\left(4\right)$

$Mon=\frac{|\underset{l=1}{\overset{L-1}{\Σ}}\mathrm{\ϵ}(F_{l+1}-F_l)-\underset{l=1}{\overset{L-1}{\Σ}}\mathrm{\ϵ}(F_l-F_{l+1})|}{L-1}---\left(5\right)$

wherein, F_lAnd t_lRespectively the characteristic value and time of the ith observation point,andthe method comprises the steps that the average values of a characteristic sequence and a time sequence in a full life cycle are respectively, L is the length of the characteristic sequence in the full life cycle, (x) is a unit step function, Corr measures the linear correlation degree of the characteristic sequence and the time sequence, and Mon measures the monotonicity of the characteristic sequence;

5) standardizing the comprehensive evaluation index Cri of the fourteen calculated characteristics to obtain Cri ', and selecting the characteristics with Cri' being more than 0.5 as a sensitive characteristic set in the degradation process of the mechanical parts;

6) utilizing life-cycle data setsTraining a recurrent neural network model, where x_t∈R^N×1Set of sensitivity features for time t, y_t∈[0,1]Is the actual degradation rate of the mechanical parts at the corresponding moment, y_t0 represents a mechanical part in a fully healthy state, y_t1 represents that the cyclic neural network is in a complete fault state, and the cost function in the training process of the cyclic neural network is shown as the formula (6):

$C=\frac{1}{2}\underset{t=1}{\overset{T}{\Σ}}\left|\right|y_t-{\hat{y}}_t|{|}_2^2---\left(6\right)$

wherein,is to circulateOutput of the Ring neural network model, y_tIs the true degradation rate;

7) and in the testing stage, the sensitivity characteristic set obtained by calculation is used as the input of the trained recurrent neural network to obtain a new mechanical part health index RNN-HI.