CN109084186B - Pipeline leakage signal identification method based on improved ELMD (ensemble empirical mode decomposition) multi-scale entropy - Google Patents

Pipeline leakage signal identification method based on improved ELMD (ensemble empirical mode decomposition) multi-scale entropy Download PDF

Info

Publication number
CN109084186B
CN109084186B CN201810937852.1A CN201810937852A CN109084186B CN 109084186 B CN109084186 B CN 109084186B CN 201810937852 A CN201810937852 A CN 201810937852A CN 109084186 B CN109084186 B CN 109084186B
Authority
CN
China
Prior art keywords
signal
elmd
value
leakage signal
waveform
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN201810937852.1A
Other languages
Chinese (zh)
Other versions
CN109084186A (en
Inventor
郝永梅
杜璋昊
覃妮
吴雨佳
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Changzhou University
Original Assignee
Changzhou University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Changzhou University filed Critical Changzhou University
Priority to CN201810937852.1A priority Critical patent/CN109084186B/en
Publication of CN109084186A publication Critical patent/CN109084186A/en
Application granted granted Critical
Publication of CN109084186B publication Critical patent/CN109084186B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • F—MECHANICAL ENGINEERING; LIGHTING; HEATING; WEAPONS; BLASTING
    • F17—STORING OR DISTRIBUTING GASES OR LIQUIDS
    • F17D—PIPE-LINE SYSTEMS; PIPE-LINES
    • F17D5/00—Protection or supervision of installations
    • F17D5/02—Preventing, monitoring, or locating loss
    • F17D5/06—Preventing, monitoring, or locating loss using electric or acoustic means
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01M—TESTING STATIC OR DYNAMIC BALANCE OF MACHINES OR STRUCTURES; TESTING OF STRUCTURES OR APPARATUS, NOT OTHERWISE PROVIDED FOR
    • G01M3/00—Investigating fluid-tightness of structures
    • G01M3/02—Investigating fluid-tightness of structures by using fluid or vacuum
    • G01M3/04—Investigating fluid-tightness of structures by using fluid or vacuum by detecting the presence of fluid at the leakage point
    • G01M3/24—Investigating fluid-tightness of structures by using fluid or vacuum by detecting the presence of fluid at the leakage point using infrasonic, sonic, or ultrasonic vibrations
    • G01M3/243—Investigating fluid-tightness of structures by using fluid or vacuum by detecting the presence of fluid at the leakage point using infrasonic, sonic, or ultrasonic vibrations for pipes

Abstract

The invention provides a pipeline leakage signal identification method based on an improved ELMD (empirical mode decomposition) multi-scale entropy, which comprises the steps of firstly carrying out noise pretreatment on acquired experimental data to eliminate low-correlation components in a signal; then ELMD processing is carried out on the preprocessed signals to obtain each PF component; weakening the endpoint effect problem remained by ELMD decomposition through a peak waveform matching method; respectively calculating the multi-scale entropy of the PF component, and arranging and comparing the multi-scale entropy values of the leakage signals to eliminate background noise; selecting a main PF component according to the multi-scale entropy value to construct a feature vector; taking the feature vector as an input vector of a BP neural network, and training the network; and inputting the sample to be detected into the trained BP neural network to obtain a pipeline leakage identification result. The method provided by the invention can adapt to various conditions of the pipeline and has better detection precision.

Description

Pipeline leakage signal identification method based on improved ELMD (ensemble empirical mode decomposition) multi-scale entropy
Technical Field
The invention relates to the technical field of pipeline leakage detection, in particular to a pipeline leakage signal identification method based on improved ELMD (empirical mode decomposition) multi-scale entropy.
Background
Urban pipelines become indispensable tools for modern urban development, along with the continuous expansion of the scale of the urban pipelines, pipeline fault events tend to rise due to the influences of natural aging of equipment, climatic environment, artificial damage and the like, and particularly once a gas pipeline leaks, serious accidents such as fire, explosion, poisoning, environmental pollution and the like are easily caused. Therefore, an effective pipeline leakage detection method is found, the hidden leakage danger is accurately identified, and the method has good economic value and social significance.
In recent years, with the development of computer technology, pipeline leakage detection technology is developing towards the combination of software and hardware, and at present, various new pipeline leakage detection methods are still leading research directions in various countries. The Local Mean Decomposition (LMD) method is a new self-adaptive time-frequency analysis method, which can self-adaptively decompose a multi-component signal into a plurality of amplitude modulation and frequency modulation component signals, and has better time-frequency feature extraction capability, but the LMD decomposition result also has modal aliasing phenomenon. The improved algorithm comprises the following steps: the Ensemble Local Mean Decomposition (ELMD) overcomes the modal aliasing phenomenon, is more suitable for analyzing complex leakage signals, and greatly improves the measurement accuracy.
Because it is difficult to obtain pressure data of all pipe sections of the whole pipeline, especially pressure data can not be obtained almost at pipe sections where the pipeline can not be close to, and under the condition of lacking pressure data of all pipe sections, the existing method for detecting pipeline leakage according to pressure data calculates according to few pressure sampling points, which results in that the detection effect is not very ideal. In recent years, with the development of detection technology, methods for acquiring pressure data of the whole pipeline appear, but a method for detecting leakage based on pressure data of a whole pipeline section is not proposed, while a BP neural network is a method capable of detecting leakage according to pressure data of the whole pipeline section, and has unique excellent performances of parallel distribution processing, self-organization, self-adaptation, self-learning, good fault tolerance and the like.
Disclosure of Invention
The technical problem to be solved by the invention is as follows: aiming at the problems in the existing pipeline leakage detection technology, the invention provides a pipeline leakage signal identification method based on improved ELMD multi-scale entropy. The method is based on pressure data of a whole pipe section in a specific pipeline, and low-correlation components in signals are eliminated according to an autocorrelation function; then, decomposing the signal by adopting an ELMD method, effectively overcoming the modal aliasing phenomenon and obtaining an accurate PF component; the extreme point continuation method is adopted to solve the problem of the end effect generated by decomposition; processing background noise by calculating a multi-scale entropy value of each PF component; finally, whether the pipeline leaks can be accurately and quickly judged through the BP neural network, and the safe operation of the urban pipeline is guaranteed, so that the property and life safety of people is guaranteed.
The technical scheme adopted for solving the technical problems is as follows: because modal aliasing can be generated after LMD decomposition is carried out on the acquired signals, and the conventional mathematical model is difficult to describe, the invention provides a pipeline leakage signal identification method based on improved ELMD multi-scale entropy from the aspects of artificial intelligence and signal processing, analyzes the pressure data of the whole pipeline, and has better superiority in signal processing, feature extraction and fault analysis, and comprises the following steps:
s1: experimental data are acquired, and a phenomenon caused by rapid release of strain energy due to crack propagation, plastic deformation or phase change and the like in the material is called acoustic emission. I.e., a non-destructive testing method that assesses the performance or structural integrity of a material by receiving and analyzing acoustic emission signals from the material. Respectively acquiring signal data by using an acoustic emission leakage system to obtain an original leakage signal x (b);
s2: performing autocorrelation preprocessing on the original leakage signal x (b), and calculating autocorrelation function rxPerforming time domain analysis to eliminate low-correlation components in the signal to obtain a preprocessed leakage signal x (t); the signal acquired by acoustic emission contains essential background noise and some low-correlation components, and the low-correlation components in the signal, although the proportion of the low-correlation components is small, have important influence on signal processing and reconstruction. The autocorrelation function of a random signal reflects the degree to which the signal correlates with itself at various points in time. The specific method comprises the following steps:
Figure BDA0001768279930000031
wherein, x (b) is an original leakage signal, x (b + a) is a leakage signal of a delay time a, and a and b are positive integers; the autocorrelation function reflects the degree of similarity between the signal x (b) and x (b + a) after a delay, and the signal is preprocessed by autocorrelation of the signal to eliminate some low-correlation components of the signal itself.
S3: the ELMD method utilizes the property that white noise signals have the function of uniformly polluting the whole time-frequency space of target signals, each time, a limited number of different white noise signals are doped in the target signals, then, the LMD method is utilized to decompose the mixed signals, all scales in the target signals are automatically decomposed into the related pass-band in a filter bank determined by the white noise signals, the signal-to-noise ratio of decomposition results is reduced, and therefore, the mode aliasing phenomenon is also reduced. However, since different white noises doped each time are uncorrelated and the average value of white noises tends to zero only when the white noises are sufficiently large, the mixed signals doped with different white noises can be decomposed for many times by using the LMD method, and then the PF component obtained after the multiple decomposition is subjected to overall averaging and is used as a final decomposition result, so that the white noises are counteracted in the overall averaging process, and the signal-to-noise ratio of the signals is improved.
ELMD (global local area mean decomposition) processing is carried out on the preprocessed leakage signals x (t) to obtain each PF component ynAnd a residual component epsilonr(ii) a Wherein the ELMD decomposition step comprises the following steps:
s3.1: and determining the overall average number of times M and the added white noise level, and setting the initial decomposition number of times M to be 1. The value of M and the white noise level are selected according to actual conditions, preferably, M is 100, and the added white noise level is 0.1.
S3.2: adding a determined level of white noise n to the preprocessed signal x (t)m(t), then the leakage signal can be expressed as:
xm(t)＝x(t)+nm(t) (2)
wherein m is the number of decompositions, t is time, nm(t) is a white noise signal, xmAnd (t) is the mixed leakage signal after white noise is added.
S3.3: for mixed leakage signal xm(t) performing an LMD decomposition m times to obtain a plurality of PF components epsilonn,m(N ═ 1,2, …,. N), N being a positive integer. Epsilonn,mIs the nth PF component obtained by the mth decomposition.
S3.4: if M is less than M, repeating steps S3.2 and S3.3, and adding 1 to the decomposition times.
S3.5: the overall average of the PF components for M decompositions is:
Figure BDA0001768279930000041
wherein N is 1,2, …, N, M is 1,2, …, M, N, M is a positive integer;
mean y of M decompositions of N PF componentsn(N-1, 2, …, N) as the final PF component. When M is 100, i.e. the average y of the N PF components decomposed 100 timesn(N-1, 2, …, N) as the final PF component.
S4: weakening the endpoint effect remained in ELMD decomposition on the preprocessed signal x (t) by peak waveform matching continuation method to obtain a continued signal x '(t), and performing ELMD decomposition again to obtain a PF component y'nAnd residual component ε'r。
Compared with the LMD decomposition, the ELMD decomposition overcomes the problem of modal aliasing to a certain extent, but large reconstruction errors still exist, and noise information doped in components is much, because in the LMD decomposition program, as a section of unknown false signals exist at endpoints of a local mean envelope function, if the unknown false signals are not processed or the processing method is not proper, false information is finally generated during the program operation, and the decomposition result is influenced. Aiming at the problem, a peak waveform matching continuation method is adopted to weaken the end effect problem:
the data at the end point is extended, and the key is to determine the variation trend of the original data at the end point. The peak waveform matching continuation method is based on an adaptive waveform matching method, namely, a waveform which best accords with the signal trend is found out from the interior of an original signal to continue the signal, the internal trend of the signal is maintained to the maximum extent, and the peak waveform matching continuation method is realized by setting a threshold value according to the peak coefficient of the waveform. The continuation of the left endpoint is described in detail below.
First, proceeding left end extension, and setting the signal after pre-processing as x (t), p1、q1The waveforms of signals x (t) respectively being at time tp、tqMaximum and minimum values of (x), (t), the left-hand endpoint of x (1), in x (1) -p1—q1Forming a triangular waveform by the three points, and calling the triangular waveform as a characteristic waveform, and then searching a triangular waveform which is most matched with the characteristic waveform along a signal x (t); taking the data before the matched waveform as the continuation waveform of x (t), the data can accord with the natural trend of the signal; the method comprises the following specific steps:
s4.1: finding a specialThe starting point x (i) of the triangular waveform outside the symbolic waveform corresponds to the time point tx(i)Then, then
Figure BDA0001768279930000051
S4.2: calculating the matching errors e (i) of all triangular waveforms and characteristic waveforms, wherein the error formula is as follows:
e(i)＝|pi-p1|+|qi-q1|+|xi-x(1)| (5)
s4.3: finding out the minimum error value e (i) marked as emin(i) And sets the threshold α if emin(i)<α, then emin(i) The corresponding triangular waveform is used as a matching waveform, the data before the matching waveform is extended to the front of the original signal, if emin(i) α, the next step is carried out.
S4.4: if emin(i) Not less than α, calculating the peak coefficient F in each triangular waveform,
Figure BDA0001768279930000052
and set a threshold value β if FminIf < β, repeat S4.3, if FminAnd β, directly setting an extreme value at the signal endpoint and carrying out the next step.
S4.5: if FminAnd β, the extreme values at the signal end points are directly set, namely the average values of two adjacent maximum values and minimum values which are closest to the left end of the signal are respectively obtained and are respectively used as the maximum value and the minimum value of the signal x (t).
S4.6: the same method is used for continuation of the right endpoint, and the completely prolonged signal is x'tELMD decomposition is performed to obtain a PF component y'nAnd residual component ε'rWherein the threshold α is derived from the error value e (i), β is derived from the waveform peak coefficient F, both of which can be adjusted according to the actual program running condition, the smaller α is, the stronger regularity is shown in the original signal, and the larger β is, the more the matched waveform is matched with the boundary data between the original signal.
S5: calculating each PF component y 'after continuation'nMulti-scale entropy value MSE ofnThe multi-scale entropy (MSE) is proposed to coarsely granulate the original data on the basis of the sample entropy and to group the sample entropy values on each scale into a set of number series, i.e. the sample entropy of the time series under different scales. If one sequence and the other sequence are under the same scale, the entropy value of the former is higher than that of the latter, which shows that the time sequence of the former has higher complexity than that of the latter. At present, learners apply the multi-scale entropy to fault diagnosis of a rolling bearing and a rotor system, and results show that the multi-scale entropy can effectively distinguish various faults. Compared with the traditional sample entropy based on a single scale, the MSE can better reflect the complexity characteristics of a time sequence from a plurality of scales, has the advantages of short data required by calculation, good stability, strong anti-noise capability and the like, and is widely applied to analysis of signals of biology, electromyography, electroencephalogram, mechanical faults and the like.
The method comprises the following steps:
s5.1: setting a discrete original time sequence as { x1, x2, …, xn }, and performing coarse-grained transformation on the original time sequence to obtain a new time sequence:
Figure BDA0001768279930000061
wherein the content of the first and second substances,
Figure BDA0001768279930000062
μ,
Figure BDA0001768279930000063
is a positive integer and
Figure BDA0001768279930000064
s is the length of the discrete time sequence, tau is a scale factor, the original sequence is divided into tau segments, each segment is a coarse-grained sequence with the length of s/tau, and when tau is 1, the new time sequence is the original sequence; preferably τ -3.
S5.2: given a mode dimension k and a similarity tolerance r (r > 0), a time series of k-dimensional vectors is constructed:
xk(μ)＝{xμ,xμ+1,…,xμ+k-1}i＝1,2,…,S-k (8)
s5.3: calculating the vector xk(mu) with
Figure BDA0001768279930000071
The distance between:
Figure BDA0001768279930000072
wherein L ═ 0,1, …, k-1; mu, the ratio of the measured value to the measured value,
Figure BDA0001768279930000073
and is
Figure BDA0001768279930000074
S5.4: for each μ, calculate xk(mu) with
Figure BDA0001768279930000075
Counting the number of distances smaller than r and taking the ratio of the total number of distances S-k +1 as Ckμ(r) that
Figure BDA0001768279930000076
S5.5: c is to bekμThe average value of (r) is recorded as
Figure BDA0001768279930000077
S5.6: adding 1 to the dimension to k +1, repeating steps S5.1 and S5.5, and calculating
Figure BDA0001768279930000078
When S is a finite value, obtaining a sample entropy estimated value when the sequence length is S according to the steps:
Figure BDA0001768279930000079
s5.7: and repeating the process to obtain sample entropy values SampEn under different scales, and selecting the optimal sample entropy under each scale as the multi-scale entropy. The multi-scale entropy is derived from the sample entropy and can be used to represent the complexity of the signal, i.e. the larger the entropy value, the more complex the signal. And finally, calculating to obtain the multi-scale entropy value of each PF component after continuation, and recording as: MSE1，MSE2，MSE3…MSEn。
S6: and selecting a main PF component according to the multi-scale entropy to construct a feature vector E, and training the network by taking the feature vector E as an input vector of the BP neural network. Artificial neural networks (ans) are a mathematical model that simulates biological neural networks for information processing. The BP (back propagation) neural network is a neural network based on an error back propagation training algorithm, and has the advantages that under the condition that the number of hidden layers and nodes is enough, the BP (back propagation) neural network can approximate any nonlinear mapping relation and has better generalization capability. And inputting the sample to be detected into the trained BP neural network to obtain a pipeline leakage result. The method for classifying and identifying the leakage signal based on the BP neural network mainly comprises the following steps:
s6.1: each acoustic leakage signal to be analyzed is subjected to characteristic analysis by calculating each PF component y'nMulti-scale entropy value MSE ofnThe feature vector E ═ MSE (MSE) is constructed by comparing the magnitudes of the multiscale entropy values1，MSE2…MSEn)；
S6.2: from different feature vectors E1、E2、EnTo construct a training input matrix a and a test output matrix B, where n is a positive integer, and the number of samples of matrix A, B is determined according to the ratio of training samples to test samples in the BP neural network (e.g., the ratio is 9: 1). Further determining system parameters of the BP neural network according to the input matrix A and the expected output matrix B; the data consisted of multiscale entropy values for PF components of the pipeline leakage at different pressures in the experiment.
S6.3: inputting the input characteristic parameter vector matrix of the training sample into a BP neural network classifier for learning to obtain a neural network classification model;
s6.4: and inputting the input characteristic parameter vector matrix of the sample to be identified into a trained neural network classifier, performing classification and identification on the signal, and outputting a final identification real leakage signal. After BP neural network training, the error between the predicted value and the true value is obtained, and the error is expressed by the root mean square value. And finally obtaining the identification accuracy of the pipeline leakage according to the error.
The invention has the beneficial effects that: the invention provides a pipeline leakage signal identification method based on an improved ELMD multi-scale entropy. The signal is decomposed by global local mean decomposition, resulting in several PF components. And weakening the endpoint effect problem remained by ELMD decomposition through a peak waveform matching continuation method. And then calculating the multi-scale entropy value of the PF component, and arranging and comparing the multi-scale entropy of the signal to judge the complexity of the signal, thereby eliminating the residual component and the background noise existing after ELMD decomposition. And finally, selecting a PF component containing main leakage characteristics to construct a characteristic vector, taking the characteristic vector as an input vector of the BP neural network, training the network, and finally inputting a sample to be detected into the trained BP neural network to obtain a pipeline leakage result. The method can accurately identify the pipeline leakage, and basically eliminates background noise, low-correlation components and residual components in signals, so that the leakage identification result is more accurate.
Drawings
The invention is further illustrated by the following figures and examples.
FIG. 1 is a schematic flow chart of a preferred embodiment of the present invention;
FIG. 2 is a diagram of a typical BP neural network topology;
FIG. 3 is a time domain frequency domain waveform diagram of an original signal;
FIG. 4 is a graph of leakage signal autocorrelation analysis;
FIG. 5 is a result of ELMD decomposition of the leakage signal;
FIG. 6 is a graph of PF component variation before peak waveform matching extension;
FIG. 7 is a graph of PF component variation after peak waveform matching continuation;
FIG. 8 is a BP neural network algorithm flow chart;
FIG. 9 illustrates a typical BP neural network topology;
FIG. 10 is a diagram of a BP neuron model;
FIG. 11 is a graph of the variation of the training error of the BP neural network;
FIG. 12 is a plot of predicted versus true values;
FIG. 13 is a graph of error variation;
FIG. 14 is a neural network target-to-output fit curve.
Detailed Description
The present invention will now be described in detail with reference to the accompanying drawings. This figure is a simplified schematic diagram, and merely illustrates the basic structure of the present invention in a schematic manner, and therefore it shows only the constitution related to the present invention.
The invention discloses a pipeline leakage signal identification method based on improved ELMD multi-scale entropy, and fig. 1 is a specific flow chart of the invention. The method comprehensively applies a global local mean decomposition algorithm, a multi-scale entropy and a BP neural network, and comprises the following specific steps:
as shown in FIG. 2, the length of the simulated pipeline in the experiment is 42m, the material of the pipeline is steel, the specification of the pipeline is DN90, the medium is compressed air, and the medium in the pipeline is in a flowing state. The position of the No. 1 upstream sensor is 0m, and a leakage hole with the leakage aperture of 1mm is arranged at a position 18m away from the No. 1 sensor. And a fixed No. 2 downstream sensor is arranged at a distance of 42m from the No. 1 upstream sensor, namely, a distance of 42m between the two sensors.
In step S1, the acoustic emission leakage detection system is used to collect the original signal waveform of the pipeline leakage, and the time domain and frequency domain distributions of the original signal waveform are shown in fig. 3.
In step S2, signal preprocessing is performed, and the signal acquired by acoustic emission contains essential background noise and some low-correlation components, and the autocorrelation function of the random signal reflects the degree of correlation between the signal and itself at different time points. The low correlation component in the signal, although not in a large proportion, has a significant effect on signal processing and reconstruction. The autocorrelation analysis of the original signal was performed using matlab, the analysis results are shown in fig. 4.
ELMD decomposition is performed on the preprocessed signal in step S3. The ELMD method is characterized in that a plurality of groups of different white noise signals are added into a signal to be decomposed before LMD decomposition, and the white noise can be uniformly distributed in the whole time-frequency space by utilizing the characteristics that the average value of the white noise is zero and the distribution of the spectral energy is uniform, and the signals with different time scales can be automatically distributed on a proper scale related to the background noise. The fixed white noise level was 0.1 and the ensemble average number was 100, and the ELMD decomposition results are shown in FIG. 5. As can be seen from fig. 5, the preprocessed signal is decomposed by ELMD to obtain a series of PF components and a residual component, where only the first 4 PF components containing main information are shown.
In step S4, compared with LMD decomposition, ELMD decomposition overcomes the problem of modal aliasing to some extent, but still has the problems of large reconstruction error and more noise information doped in the component. This is because in the ELMD decomposition program, since there is a section of unknown false signal at the end points of the local mean envelope function, if it is not done well or the processing method is not proper, false information will be generated during the program running, which will affect the decomposition result. Aiming at the problem, a peak waveform matching continuation method is adopted to weaken the end effect problem. The peak waveform matching continuation method mainly calculates the matching error e of two parameter waveforms(i)The crest factor F, the prolongation effect are shown in fig. 6 and 7. Fig. 6 and 7 are graphs of the PF components before and after solving the endpoint effect by using the peak waveform matching prolongation, respectively, and it is apparent from the graphs that the waveform of the PF component in fig. 7 is better at the endpoint and the endpoint effect is significantly reduced after the peak waveform matching prolongation. Therefore, the peak waveform matching prolongation method can weaken the endpoint effect brought by ELMD decomposition to a certain extent.
In step S5, in order to further eliminate the background noise and residual components existing after ELMD decomposition, the accuracy of pipeline leakage identification is improved. And calculating the multi-scale entropy of the PF component, and arranging the multi-scale entropy values of the comparison signals, wherein the multi-scale entropy values reflect the complexity of the leakage signals.
When the pipeline leaks, the vibration energy can be changed due to the changes of different pore diameters, different pressures and different carrier frequency distributions, and the effective value of the vibration signal can also be changed. The effective value of the vibration signal describes the magnitude of the vibration of the instantaneous value of the amplitude of the signal over the sampling time, which is an average over a time series, adapted to analyze a continuous vibration signal. The pipeline leakage signal has the characteristics of uncertainty, complexity and nonlinearity, and the method selects the multi-scale entropy to judge the complexity of the pipeline leakage signal and calculates the sample entropy value for comparison. And (3) randomly taking 30 preprocessed signals, and respectively calculating the multi-scale entropy value and the sample entropy value of the preprocessed signals, wherein the multi-scale entropy value and the sample entropy value are shown in a table 1 and a table 2.
Table 1 is the multi-scale entropy values for the PF components;
Figure BDA0001768279930000111
Figure BDA0001768279930000121
table 2 is the sample entropy values for the PF components;
Figure BDA0001768279930000122
Figure BDA0001768279930000131
from Table 1, it can be seen that the PF component contains leakage information from the PF1To the PF4Gradually decrease. Comparing the multiscale entropy values of the PF components, it can be seen that the PF1—PF3Multiscale entropy of components relative to PF4Larger, indicates that the former contains more leakage information. Therefore, the first 3 PF components are selected to construct a feature vector, so that the influence of background noise and residual components on a leakage signal can be greatly reduced. And inputting the feature vector into a BP neural network for training and identification.
Comparing the multi-scale entropy and the sample entropy in table 1 and table 2, it can be clearly found that, under the same PF component, the multi-scale entropy is slightly larger than the sample entropy, which is related to the principle of multi-scale entropy and sample entropy, and the larger the entropy, the more the effective information is contained. Therefore, the input feature vector of the BP neural network is constructed by selecting the PF component multi-scale entropy value.
The BP neural network algorithm flow is shown in figure 8, and comprises initialization parameters, input values, expected values, normalized data, adjustment weights and the like, the specific structure of the BP neural network is shown in figure 9 and figure 10, the initialization parameters of the BP neural network are set to be a 3-layer topological structure, transfer functions of an implied layer and an output layer are both selected from logarithmic sigmoids, a training algorithm selects an adaptive learning rate momentum gradient descent method, the initialization parameters of the BP neural network are set to be that the training times k are 2000, the target error coefficient e is 0.005, the learning rate η is 0.01, the momentum factor is 0.9, and the number of nodes of the implied layer is 20.
And calculating multi-scale entropy according to the PF components subjected to ELMD decomposition in the step S5, and selecting multi-scale entropy values of the first 3 PF components to construct a feature vector.
BP neural network identification is performed according to step S6. 30 sets of pre-processed samples were taken, each from the leakage signal of the pipeline at different pressures (0.1mpa, 0.2mpa, 0.3mpa), each set of samples containing a multi-scale entropy of 3 PF components. Wherein 27 groups are selected as training samples, 3 groups are selected as testing samples, and the variation of mean square error performance in the training process is shown in fig. 11.
As can be seen from fig. 11, as the number of iteration steps increases, the mean square error between the network output value and the target value becomes smaller. In the 505 th iteration, the mean square error between the output value and the target value is 0.0049726 and is less than the set value by 0.005, and the set training effect is achieved.
The classification results of the test samples are shown in fig. 12. Through comparison of the predicted value and the true value in fig. 12 and the error change diagram in fig. 13, it can be seen that the multi-scale entropy value of the pipeline leakage signal after the BP neural network training is very close to the actual value, the error is small, the accuracy rate of the pipeline leakage identification is high, and the requirement of the leakage signal identification is met. Meanwhile, the training time of the BP neural network is 0.2s, and the training process is very quick.
The BP neural network training fit curve is shown in fig. 14. The abscissa in the figure is the target value and the ordinate is the output value of the network. To prevent overfitting, the data was divided into thirds, training, validation, test. As can be seen from FIG. 14, the training fitting results are distributed uniformly and in diagonal distribution, and the fitting effect is good.
Finally, the pipeline leakage signal is decomposed by an improved ELMD method, the end effect is weakened, a feature vector is constructed by calculating a multi-scale entropy value and is input to a BP neural network for training and recognition, and low related components, residual components and background noise in the test signal are eliminated. The BP neural network identification results are shown in table 3:
table 3 shows the BP neural network training recognition results
Figure BDA0001768279930000151
As can be seen from table 3, the input vector of the BP neural network is constructed by calculating the multi-scale entropy of the PF component after ELMD decomposition, and the final recognition result is higher than the calculated sample entropy. Therefore, the pipeline leakage identification method based on the ELMD multi-scale entropy and BP neural network firstly carries out autocorrelation analysis on a pipeline leakage signal to remove low-correlation components, then carries out improved ELMD decomposition to weaken the end effect, calculates the multi-scale entropy value, selects PF components containing main leakage information to form a feature vector, and finally carries out training identification by the BP neural network. As can be seen from fig. 12-14, the method of the present invention is feasible and has good recognition effect.
In light of the foregoing description of preferred embodiments in accordance with the invention, it is to be understood that numerous changes and modifications may be made by those skilled in the art without departing from the scope of the invention. The technical scope of the present invention is not limited to the contents of the specification, and must be determined according to the scope of the claims.

Claims (5)

1. A pipeline leakage signal identification method based on improved ELMD multi-scale entropy is characterized in that: comprises the following steps of (a) carrying out,
s1: acquiring experimental data, and acquiring a leakage signal x (b) of a pipeline by using an acoustic emission leakage detection system;
s2: calculating an autocorrelation function r for the leakage signal x (b)xPerforming time domain analysis to eliminate low-correlation components in the signal to obtain a preprocessed leakage signal x (t);
s3: determining the total average times M and the added white noise level, and carrying out ELMD decomposition on the preprocessed leakage signal x (t) containing the background noise to obtain a series of PF components ynAnd a residual component epsilonrWherein y isn＝εn，m(n＝1，2，...，.N)；
S4: attenuating endpoint effect remained in ELMD decomposition on the preprocessed leakage signal x (t) by peak waveform matching continuation method to obtain a continued signal x '(t), and performing ELMD decomposition to obtain a PF component y'nAnd residual component ε'r；
S5: calculating each PF component y 'after continuation'nMulti-scale entropy value MSE ofnN is a positive integer, and PF components containing a large amount of leakage information are selected according to the size of the multi-scale entropy value to construct an eigenvector E (MSE)1，MSE2…MSEn) Further eliminating background noise and residual components in the original leakage signal;
s6: taking the characteristic vector E as an input vector of a BP neural network, setting the neural network into a 3-layer topological structure, wherein transfer functions of a hidden layer and an output layer are both selected from a logarithmic S shape, selecting a self-adaptive learning rate momentum gradient descent method by a training algorithm, training and testing the network, and finally identifying a pipeline leakage signal;
wherein, step S4 specifically includes the following steps:
first, proceeding left end extension, and setting the signal after pre-processing as x (t), p1、q1The waveforms of signals x (t), respectively, being at time tp、tqThe maximum value and the minimum value of the position, the left end point value of the waveform of the signal x (t) is set as x (1), andx(1)-p1-q1forming a triangular waveform by the three points, and calling the triangular waveform as a characteristic waveform, and then searching a triangular waveform which is most matched with the characteristic waveform along the waveform of the signal x (t); taking the data before the matched waveform as the continuation waveform of x (t), the data can accord with the natural trend of the signal; the method comprises the following specific steps:
s4.1: finding out the starting point x (i) of the triangular waveform except the characteristic waveform, wherein the corresponding time point is tx(i)Then, then
Figure FDA0002301739640000021
S4.2: calculating the matching errors e (i) of all triangular waveforms and characteristic waveforms, wherein the error formula is as follows:
e(i)＝|pi-p1|+|qi-q1|+|xi-x(1)| (5)
s4.3: finding out the minimum error value e (i) marked as emin(i) And sets the threshold α if emin(i) < α, then emin(i) The corresponding triangular waveform is used as a matching waveform, the data before the matching waveform is extended to the front of the original signal, if emin(i) Not less than α, and carrying out the next step;
s4.4: if emin(i) Not less than α, calculating the peak coefficient F in each triangular waveform,
Figure FDA0002301739640000022
and set a threshold value β if FminIf < β, repeat S4.3, if FminAt least β, directly setting an extreme value at the signal endpoint, and carrying out the next step;
s4.5: if Fminβ, setting extremum at signal end point directly, i.e. calculating average value of two adjacent maximum and minimum values nearest to left end of signal, and using the average value as maximum and minimum values of signal x (t);
s4.6: the same method is used for continuation of the right endpoint, and the completely prolonged signal is x'tPerforming ELMD decompositionLine ELMD decomposition to obtain PF component y'nAnd residual component ε'r。
2. The improved ELMD multi-scale entropy based pipeline leakage signal identification method of claim 1, wherein: the autocorrelation function of the leakage signal x (b) in step S2 is represented as:
Figure FDA0002301739640000023
wherein, x (b) is a pressure signal converted from an original leakage signal of the pipeline by an acoustic emission instrument, x (b + a) is a leakage signal of the delay time a, and a and b are positive integers; the autocorrelation function reflects how similar the signal x (b) is to x (b + a) after some delay from itself.
3. The improved ELMD multi-scale entropy based pipeline leakage signal identification method of claim 2, wherein: the ELMD decomposition in step S3 specifically includes the following steps:
s3.1: determining the total average times M and the added white noise level, and setting the initial decomposition times M to be 1;
s3.2: adding a determined level of white noise n to the preprocessed signal x (t)m(t), then the leakage signal is expressed as:
xm(t)＝x(t)+nm(t) (2)
wherein m is the number of decompositions, t is time, nm(t) is a white noise signal, xm(t) is the mixed leakage signal after white noise is added;
s3.3: for mixed leakage signal xm(t) performing an LMD decomposition m times to obtain a plurality of PF components epsilonn，m(N ═ 1,2, ·,. N), N being a positive integer; epsilonn，mThe nth PF component obtained by the mth decomposition;
s3.4: if M is less than M, repeating the steps S3.2 and S3.3, and adding 1 to the decomposition times M;
s3.5: the overall average of the PF components for M decompositions is:
Figure FDA0002301739640000031
wherein N is 1,2,., N, M is 1,2,.., M, N, M are positive integers;
mean y of M decompositions of N PF componentsn(N ═ 1, 2.., N) as the final PF component.
4. The improved ELMD multi-scale entropy based pipeline leakage signal identification method of claim 1, wherein: the specific step of calculating the multi-scale entropy of the PF component in step S5 includes:
s5.1: let PF component y'nHas an original time sequence of { x }1，x2，...，xnConverting data of the time sequence into pressure signals obtained by an acoustic emission instrument, and carrying out coarse graining transformation on the original time sequence to obtain a new time sequence:
Figure FDA0002301739640000032
wherein the content of the first and second substances,
Figure FDA0002301739640000033
μ，
Figure FDA0002301739640000034
is a positive integer and
Figure FDA0002301739640000035
s is the length of the discrete time sequence, tau is a scale factor, the original sequence is divided into tau segments, each segment is a coarse-grained sequence with the length of s/tau, and when tau is 1, the new time sequence is the original sequence;
s5.2: given a mode dimension k and a similarity tolerance r (r > 0), a time series of k-dimensional vectors is constructed:
xk(μ)＝{xμ，xμ+1，...，xμ+k-1} (8)
wherein, i ═ 1, 2.., s-k;
s5.3: computing vectors
Figure FDA0002301739640000041
And
Figure FDA0002301739640000042
the distance between:
Figure FDA0002301739640000043
wherein L ═ 0,1,. ·, k-1; mu, the ratio of the measured value to the measured value,
Figure FDA0002301739640000044
and is
Figure FDA0002301739640000045
S5.4: for each μ, calculate xk(mu) with
Figure FDA0002301739640000046
Counting the number of distances smaller than r and taking the ratio of the total number of distances s-k +1 as Ckμ(r) then
Figure FDA0002301739640000047
S5.5: c is to bekμThe average value of (r) is recorded as
Figure FDA0002301739640000048
Then
Figure FDA0002301739640000049
S5.6: adding 1 to the dimension to k +1, repeating steps S5.1 to S5.5, and calculating
Figure FDA00023017396400000410
When S is a finite value, the sample entropy estimated value obtained according to the steps when the sequence length is S is as follows:
Figure FDA00023017396400000411
s5.7: repeating the process to obtain sample entropy values SampEn under different scales, and selecting the optimal sample entropy value under each scale as multi-scale entropy MSEnFinally, the extended multi-scale entropy value of each PF component is calculated and recorded as: MSE1，MSE2，MSE3…MSEn。
5. The improved ELMD multi-scale entropy based pipeline leakage signal identification method of claim 4, wherein: the step of constructing the feature vector in step S6 specifically includes:
s6.1: calculating each PF component y'nMulti-scale entropy value MSE ofnSelecting the characteristic vector E (MSE) with larger multi-scale entropy value to construct1，MSE2…MSEn)；
S6.2: from different feature vectors E1、E2…EnConstructing a training input matrix A and a testing output matrix B, wherein n is a positive integer, determining the sample number of a matrix A, B according to the proportion of training samples and testing samples in the BP neural network, and determining the system parameters of the BP neural network according to the input matrix A and the expected output matrix B;
s6.3: inputting the input feature vector matrix of the training sample into a BP neural network classifier for learning to obtain a neural network classification model;
s6.4: and inputting the input characteristic parameter vector matrix of the sample to be identified into a trained neural network classifier, performing classification and identification on the signal, and outputting a final identification real leakage signal.
CN201810937852.1A 2018-08-17 2018-08-17 Pipeline leakage signal identification method based on improved ELMD (ensemble empirical mode decomposition) multi-scale entropy Active CN109084186B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201810937852.1A CN109084186B (en) 2018-08-17 2018-08-17 Pipeline leakage signal identification method based on improved ELMD (ensemble empirical mode decomposition) multi-scale entropy

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201810937852.1A CN109084186B (en) 2018-08-17 2018-08-17 Pipeline leakage signal identification method based on improved ELMD (ensemble empirical mode decomposition) multi-scale entropy

Publications (2)

Publication Number Publication Date
CN109084186A CN109084186A (en) 2018-12-25
CN109084186B true CN109084186B (en) 2020-05-26

Family

ID=64793731

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201810937852.1A Active CN109084186B (en) 2018-08-17 2018-08-17 Pipeline leakage signal identification method based on improved ELMD (ensemble empirical mode decomposition) multi-scale entropy

Country Status (1)

Country Link
CN (1) CN109084186B (en)

Families Citing this family (13)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109630908B (en) * 2019-01-23 2020-10-09 常州大学 Pipeline leakage positioning method capable of reducing noise for multiple times
CN110146292B (en) * 2019-06-04 2021-08-31 昆明理工大学 Rolling bearing fault detection method based on total local mean decomposition of integrated noise reconstruction
CN110263832B (en) * 2019-06-11 2023-01-06 哈尔滨工程大学 AUV navigation system fault diagnosis method based on multi-scale analysis
CN110454687B (en) * 2019-07-22 2020-09-22 常州大学 Pipeline multipoint leakage positioning method based on improved VMD
CN110730037B (en) * 2019-10-21 2021-02-26 苏州大学 Optical signal-to-noise ratio monitoring method of coherent optical communication system based on momentum gradient descent method
CN111397700B (en) * 2020-03-02 2021-12-10 西北工业大学 Wall-mounted fault detection method of Coriolis mass flow meter
CN111444805B (en) * 2020-03-19 2023-03-17 哈尔滨工程大学 Improved multi-scale wavelet entropy digital signal modulation identification method
CN111735583B (en) * 2020-06-24 2022-01-28 东北石油大学 Pipeline working condition identification method based on LCD-EE pipeline sound wave signal characteristic extraction
CN113722977B (en) * 2021-05-26 2023-08-18 上海华电奉贤热电有限公司 Gas turbine rotor fault early warning method based on hybrid prediction
CN114234061A (en) * 2021-12-20 2022-03-25 北京工业大学 Neural network-based intelligent judgment method for water leakage sound of pressurized operation water supply pipeline
CN114576568B (en) * 2022-02-25 2023-08-29 辽宁石油化工大学 Pipeline leakage detection method and device based on infrasonic wave
CN115111537B (en) * 2022-08-24 2022-11-18 北京云庐科技有限公司 Method, device and medium for determining the position of a leak in a gas pipeline network
CN115204243B (en) * 2022-09-15 2023-02-07 西南交通大学 LMD endpoint effect improvement method based on similar triangular waveform matching continuation

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
KR20150041564A (en) * 2013-10-08 2015-04-16 (주)유민에쓰티 Alkali solution leak detection apparatus
CN106598910A (en) * 2016-12-28 2017-04-26 四川中烟工业有限责任公司 EMD (Empirical Mode Decomposition) end effect inhibiting method and system
CN106778594A (en) * 2016-12-12 2017-05-31 燕山大学 Mental imagery EEG signal identification method based on LMD entropys feature and LVQ neutral nets
CN107228282A (en) * 2017-07-06 2017-10-03 东北石油大学 A kind of gas pipeline leakage localization method and device
CN108152363A (en) * 2017-12-21 2018-06-12 北京工业大学 A kind of defect of pipeline recognition methods for pressing down end intrinsic time Scale Decomposition

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
DE102013102616A1 (en) * 2013-03-14 2014-09-18 Khs Gmbh Method and filling system for filling containers

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
KR20150041564A (en) * 2013-10-08 2015-04-16 (주)유민에쓰티 Alkali solution leak detection apparatus
CN106778594A (en) * 2016-12-12 2017-05-31 燕山大学 Mental imagery EEG signal identification method based on LMD entropys feature and LVQ neutral nets
CN106598910A (en) * 2016-12-28 2017-04-26 四川中烟工业有限责任公司 EMD (Empirical Mode Decomposition) end effect inhibiting method and system
CN107228282A (en) * 2017-07-06 2017-10-03 东北石油大学 A kind of gas pipeline leakage localization method and device
CN108152363A (en) * 2017-12-21 2018-06-12 北京工业大学 A kind of defect of pipeline recognition methods for pressing down end intrinsic time Scale Decomposition

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
基于ELMD多尺度模糊熵和概率神经网络的暂态电能质量识别;董国新;《中国优秀硕士学位论文全文数据库工程科技Ⅱ辑》;20170215(第2期);C042-1776页 *
改进LMD及高阶模糊度函数的管道泄漏定位;孙洁娣等;《仪器仪表学报》;20151031;第36卷(第10期);第2215-2223页 *

Also Published As

Publication number Publication date
CN109084186A (en) 2018-12-25

Similar Documents

Publication Publication Date Title
CN109084186B (en) Pipeline leakage signal identification method based on improved ELMD (ensemble empirical mode decomposition) multi-scale entropy
CN110454687A (en) A kind of pipeline multipoint leakage localization method based on improvement VMD
CN111664365A (en) Oil and gas pipeline leakage detection method based on improved VMD and 1DCNN
CN112528564A (en) Bridge structure damage identification method based on sparse Bayesian learning
KR100795227B1 (en) Method and apparatus for analyzing signal pattern of sensor array
CN109409271B (en) Ferromagnetic material hardness prediction algorithm based on BP neural network improved algorithm
CN113822201B (en) Deep learning method for underwater object shape recognition based on flow field velocity component time course
CN112861066B (en) Machine learning and FFT (fast Fourier transform) -based blind source separation information source number parallel estimation method
CN111985533A (en) Incremental underwater sound signal identification method based on multi-scale information fusion
CN113472390B (en) Frequency hopping signal parameter estimation method based on deep learning
CN103852525B (en) Acoustic emission signal recognition methods based on AR-HMM
CN111695461B (en) Radar clutter intelligent classification method based on image frequency characteristics
CN112327293A (en) Sea surface target detection method based on fractal feature intelligent learning
CN110057918B (en) Method and system for quantitatively identifying damage of composite material under strong noise background
Wang et al. Identification of engine faults based on acoustic emission signals using a 1DCNN-ViT ensemble model
CN110458071A (en) A kind of fiber-optic vibration signal characteristic abstraction and classification method based on DWT-DFPA-GBDT
CN109840386B (en) Damage identification method based on factor analysis
Heng et al. Acoustic emission source localization on a pipeline using convolutional neural network
Ye et al. Multi-variable classification model for valve internal leakage based on acoustic emission time–frequency domain characteristics and random forest
CN110555457A (en) Engine lubricating oil wear signal characteristic processing system and method
Babu et al. Misfire detection in a multi-cylinder diesel engine: A machine learning approach
Gao et al. Acoustic emission-based small leak detection of propulsion system pipeline of sounding rocket
Shin et al. Markov chain-based mass estimation method for loose part monitoring system and its performance
CN115346514A (en) Intelligent driving evaluation-oriented audio early warning intelligent identification method
CN110532635A (en) A kind of pipeline leakage testing algorithm based on time domain

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant