An Improved Rotating Machinery Fault Diagnosis Method Based on H5N1-VMD-MFWA Noise Reduction

Abstract

This invention provides an improved H5N1-VMD-MFWA noise reduction method for rotating machinery fault diagnosis, belonging to the field of mechanical fault diagnosis technology. Its contents include: acquiring vibration signals of rotating machinery under normal and fault conditions; using the minimum envelope entropy as the fitness function, employing an improved avian influenza virus algorithm to optimize the variational mode decomposition parameters to obtain the optimal parameters. K , α Combination; variational mode decomposition of the vibration signal based on optimal parameters, and obtaining the results through decomposition. K This invention identifies several intrinsic mode function components. Based on the correlation coefficient, kurtosis, and envelope entropy of each component, a comprehensive evaluation index is constructed using a multi-feature weighted attention mechanism. Weights are assigned to each component, and the signal is reconstructed. The reconstructed signal is then segmented into a window and input into a convolutional neural network model to complete fault diagnosis. This invention effectively improves the accuracy of fault diagnosis and provides a guarantee for the safe operation of mechanical equipment.

Description

Technical Field

[0001] This invention relates to the field of rotating machinery fault diagnosis technology, and more specifically, to a rotating machinery fault diagnosis method with improved H5N1-VMD-MFWA noise reduction. Background Technology

[0002] When rotating machinery malfunctions, it can disrupt normal production processes and reduce product quality, or even cause huge economic losses and serious safety accidents resulting in casualties. Furthermore, because such equipment typically operates in harsh environments, fault signals are easily masked by strong background noise, making effective identification difficult. Therefore, achieving timely and accurate diagnosis of early-stage faults in rotating machinery under strong noise interference is of great practical significance.

[0003] Numerous scholars have conducted extensive research on signal denoising. Among them, Variational Mode Decomposition (VMD) has been widely applied in the field of mechanical fault diagnosis due to its ability to effectively suppress mode aliasing and avoid end-point effects. However, the VMD method requires the manual setting of multiple parameters during application, which introduces a degree of subjectivity, particularly regarding the number of decomposition layers. K and penalty factor α Especially crucial. The values ​​of these two parameters directly affect the decomposition effect of VMD: if K Setting the value too small will result in "under-decomposition"; if α If the value is too small, modal aliasing may still occur; conversely, if... K Setting the value too large will cause "over-decomposition"; while α If the value is too large, it is easy to generate false components.

[0004] Currently, fault diagnosis methods based on intelligent optimization algorithms to optimize VMD parameters and combined with Intrinsic Mode Function (IMF) component screening criteria have been widely adopted to suppress noise interference. However, such methods still have certain limitations: when screening IMF components, useful signals containing fault characteristics are easily mistakenly rejected, leading to the problem of "over-denoising," thus losing some effective information; and how to better preserve the integrity of the signal while denoising is also a challenge. Therefore, constructing an IMF reconstruction strategy that can fully preserve useful information remains an urgent problem to be solved. Summary of the Invention

[0005] The purpose of this invention is to address the problem that early, subtle fault characteristics of rotating machinery are easily masked and difficult to accurately identify under high-noise environments. This invention proposes a noise reduction fault diagnosis method based on an improved H5N1-VMD-MFWA. This method can effectively extract subtle fault characteristics, providing a reliable basis for early fault diagnosis of rotating machinery under high-noise environments, and offering technical support for ensuring the safe operation of equipment and the safety of people's lives and property.

[0006] The technical solution for achieving the present invention is: an improved method for diagnosing rotating machinery noise reduction using H5N1-VMD-MFWA, comprising the following steps:

[0007] Step S1: Acquire vibration signals of rotating machinery under normal conditions and with different types of faults;

[0008] Step S2: Using the minimum envelope entropy of the vibration signal as the fitness function, the improved H5N1 optimization algorithm is used to optimize the key parameters of variational mode decomposition. K , α We perform an optimization search to obtain the optimal parameter combination.

[0009] Step S3: Based on the optimal parameter combination [ K , α The vibration signal is subjected to variational mode decomposition to obtain K eigenmode function components;

[0010] Step S4: Calculate the correlation coefficient, kurtosis, and envelope entropy of each intrinsic mode function component and the original signal. Use the multi-feature weighted attention (MFWA) mechanism to determine the weight of each component, normalize and weightedly fuse the scoring index, and reconstruct the vibration signal.

[0011] Step S5: Perform window segmentation on the reconstructed vibration signal and perform fault diagnosis on the reconstructed signal based on the convolutional neural network model.

[0012] Furthermore, in step S2, the method for calculating the envelope entropy is as follows:

[0013] E p =-∑ M z=1 ε ( z ) lgε ( z )

[0014] ε ( z )= a ( z ) / ∑ Mz=1 a ( z )

[0015] in, a ( z The envelope of the signal is obtained by performing a Hilbert transform on the vibration signal of rotating machinery. ε ( z )yes a ( z The normalized envelope sequence, satisfying: ∑ M z=1 ε ( z )=1, ε ( z )≥0, M The number of sampling points. E p Let be the envelope entropy.

[0016] Furthermore, in step S2, the improved avian influenza virus optimization algorithm comprises the following steps:

[0017] The improved avian influenza virus optimization algorithm is in D In a dimensional search space, through a containing N The search was optimized by analyzing the population of individual virus individuals; among them, D A positive integer, representing the dimension of the optimization problem; N The integer is greater than zero, representing the size of the virus population; the position of each individual virus represents a candidate solution to the problem.

[0018] Step S21: "Initialization" phase, setting the virus population size. N With maximum number of iterations T A uniform random initialization method is adopted: X j i = LB +( UB - LB )⊙ R j i Initialize the virus population and calculate the fitness value of each virus to determine the current best individual location and fitness value;

[0019] Where ⊙ represents element-wise multiplication, and the superscript... i Represents an index of individual viruses. i = 1, 2, …, N subscript j Represents the dimension of the search question space. D index, j = 1, 2, …, D , X represent N OKD A matrix of columns, where X j i Representing the i The virus individual in the first j Positional components in each dimension UB and LB These are the upper and lower bounds of the search space. R This represents elements whose values ​​are in the range [0~1]. N OK D A random number matrix of columns, where R j i It is the first in the matrix i Line number j The numerical components of the column;

[0020] Step S22: In the "population position matrix" construction stage, based on the position update strategy of multiple random permutations, three randomly rearranged population position matrices are constructed. PMP 1, PMP 2 and PMP 3;

[0021] Step S23: The "selection" phase, the virus will select based on the attack probability. P attack Select an attack target, if P attack If the value is less than 0.8, the virus attacks poultry; otherwise, the virus attacks humans.

[0022] Step S24: "Judgment" phase, the virus adapts according to its probability. P adapt Determining whether the host is alive or whether the environment is conducive to viral survival: P adapt When the value is ≥0.85, it indicates that the host has died or the environment is unfavorable, and the virus will escape the current environment and enter the "escape" stage. P adapt When the value is less than 0.85, it indicates that the host is alive or the environment is suitable. The virus location will be fine-tuned to adapt to the host. The location update method is as follows:

[0023] when P attack <0.8 and P adapt When <0.85:

[0024] X j i = PMP 3 j i + R j i⊙( X g,j - X p,j i )

[0025] when P attack ≥0.8 and P adapt When <0.85:

[0026] X j i = X j i +0.5* R j i ⊙( PMP 1 j i - PMP 2 j i )+0.5* R j i ⊙( X g,j - X j i )

[0027] in, PMP 1 j i , PMP 2 j i and PMP 3 j i Represent PMP 1, PMP 2 and PMP 3 in the i Line number j The numerical components of the column;

[0028] Step S25: The "escape" phase, the virus will adjust its escape probability accordingly. P escape Choose one of two strategies: If P escape < P The virus then executes a mutation strategy; if P escape ≥ P Then the diffusion strategy is executed, where P The parameters for determining "escape" are generated as follows:

[0029] when d ≥0.4,P =0.7

[0030] When 0.4> d >0.15, P =0.5

[0031] When 0.15≥ d , P =0.3

[0032] in, d To integrate diversity, d =0.7× sr +0.3× fr ;

[0033] in, sr This refers to the relative dispersion of spatial diversity. fr The relative differences in solution quality among populations are calculated as follows:

[0034] sr =mean(std( X p,j )) / max( UB - LB )

[0035] fr =std( F p ) / (abs(mean( F p )))

[0036] Among them, mean() calculates the mean; std() calculates the standard deviation; and abs() calculates the absolute value. X p,j It is the first j The individual's historical optimal position; F p It is the individual's historical best fitness value;

[0037] During the "escape" phase, the virus location update formula is as follows:

[0038] when P attack <0.8 and P adapt ≥0.85 and P escape < P , X new,j i = X j i + R j i⊙( X g,j - X j i )

[0039] when P attack ≥0.8 and P adapt ≥0.85 and P escape < P , X new,j i = X j i + R j i ⊙( X g,j - X p,j i )

[0040] X mutate,j i =0.5* c *( X new,j i + X g,j )* r * ω

[0041] when P escape ≥ P , X spread,j i = X g,j + R j i ⊙( X new,j i - X g,j )

[0042] If a mutation strategy is executed, then let X j i =X mutate,j i ;

[0043] If the diffusion strategy is executed, then let X j i=X spread,j i ;

[0044] in, X g,j It is the first j The globally optimal individual position in dimension. X new,j i Individual virus i In the j The escape midpoint in dimensional space; X mutate,j i Individual virus i In the j New positions generated by mutations in a dimension; X spread,j i Individual virus i In the j The new position generated by propagation in dimensions F g It is the globally optimal fitness value. P attack , P adapt , P escape and r It is a random number in the range (0,1). c It is the exponential decay coefficient:

[0045] c = r × e -4×t / T

[0046] ω It is a weighting factor:

[0047] ω = ω × ω d

[0048] ω d = ω min +( ω max - ω min )× e -(t / T)

[0049] in, t Indicates the current iteration number. ω d The damping coefficient representing the weight.ω max and ω min These represent the maximum and minimum values ​​of the weights, respectively. ω The initial value is 1;

[0050] Step S26: After each iteration, calculate the fitness value of all new virus positions and compare the new fitness value with the individual's historical best value. If the new fitness value is better, update the individual's historical best position and historical best fitness value.

[0051] At the same time, the global optimal solution is updated based on a probabilistic mechanism, that is, only when... ps Less than the global update probability G The global optimal position and global optimal fitness value will only be updated if there is an individual fitness value that is better than the global optimal fitness value; among which, ps Given a random number in the range (0,1), the global update probability is... G The initial value is 0, and the update process is as follows:

[0052] G =(1 / 1+e -10(t / T-0.5) )+ r* (1- G )

[0053] Among them, when the number of iterations t achieve T When the iteration count reaches zero, the algorithm terminates the iteration and outputs the global optimal position and its fitness value as the optimal solution; if the iteration count is zero... t Failed to achieve T If so, return to step S22 and repeat steps S22 to S26.

[0054] Furthermore, in step S4, the steps of determining the weight of each component using a multi-feature weighted attention mechanism, normalizing and weighted fusing the scoring index, and reconstructing the vibration signal are as follows:

[0055] Step S41: Based on the decomposition K The positive ideal reference benchmark for each index is determined by the maximum correlation coefficient, maximum kurtosis, and minimum envelope entropy of each intrinsic mode function component. K The minimum correlation coefficient, minimum kurtosis, and maximum envelope entropy of each IMF component determine the negative ideal reference benchmark for each indicator.

[0056] Step S42: Standardize the evaluation indicators by using a normalization method to map each indicator to the [0,1] interval, thus eliminating the influence of dimensionality.

[0057] Normalized value = (current value - negative ideal benchmark) / (positive ideal benchmark - negative ideal benchmark)

[0058] Step S43: Set three weights: correlation coefficient weight, kurtosis weight, and envelope entropy weight, and calculate the final score of each intrinsic mode function component based on the weighted comprehensive scoring method; wherein, correlation coefficient weight + kurtosis weight + envelope entropy weight = 1;

[0059] Step S44: Convert the overall score into attention weights using the softmax function to achieve a probability distribution transformation of the score;

[0060] Step S45: Weight all intrinsic mode function components based on attention weights to form a new signal.

[0061] Furthermore, in step S22, the construction of three randomly rearranged population position matrices... PMP 1, PMP 2 and PMP 3. Its generation process is as follows:

[0062] Step S221: Generate the basic index vector r o =[0,1,2,..., N -1], and generate a random permutation vector. σ = [ σ 1 , σ 2 , σ 3], used to control the cyclic offset of subsequent indices, where σ 1 , σ 2 , σ 3 is a random permutation of the numbers 1, 2, and 3;

[0063] Step S222: Generate initial random index vector a 1, a 1=[ j 1, j 2,..., j N ],in, j 1, j 2,..., j N It is the interval [1, N A random permutation of positive integers;

[0064] Step S223: Utilize the first element of σ σ 1 pair of basic index vectors r o Perform a cyclic offset to obtain the intermediate index vector. r t1 The cyclic offset is achieved by:r o Each index value in the database is related to... σ 1 after addition N Use the remainder to ensure the result index is between 0 and 1. N Cycle within the range of -1;

[0065] Step S224: Based on r t1 right a 1. Perform index remapping, that is, use the index vector [ r t1 +1] From vector a Extract the element at the corresponding position from step 1 to obtain the second random index vector. a 2;

[0066] Step S225: Use σ The second element σ 2. Again r o Perform a cyclic offset to obtain the intermediate index vector. r t2 and based on r t2 right a 2. Perform index remapping to obtain a third random index vector. a 3;

[0067] Step S226: Finally, calculate the optimal historical position matrix of the virus. X p Row index extraction is performed to construct three random permutation matrices:

[0068] PMP 1= X p ( a 1,:)

[0069] PMP 2= X p ( a 2,:)

[0070] PMP 3= X p ( a 3,:)

[0071] The beneficial effects of this invention are as follows: Based on traditional signal decomposition-based denoising and fault diagnosis methods, improvements have been made to significantly reduce the interference of strong noise environments on the fault diagnosis process. An adaptive synthesis diversity and local exploitation "escape" stage has been added to the H5N1 algorithm, improving the H5N1 optimization algorithm and enabling it to effectively avoid local optima in complex multi-mode optimization problems, maintain population diversity, and achieve high-quality global convergence. The vibration signal is reconstructed, considering the correlation coefficient, kurtosis, and envelope entropy of the IMF components. A multi-feature weighted attention mechanism is introduced, and normalization is performed based on the positive and negative ideal solutions of each IMF component to construct a comprehensive scoring index, thereby achieving adaptive weighting of each IMF component. Combined with multi-dimensional evaluation indicators, the fault characteristic representation ability of each component is more comprehensively assessed, effectively enhancing the significance of mechanical fault characteristics and improving the reliability of diagnostic results. Attached Figure Description

[0072] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0073] Figure 1 This is a flowchart of rotating machinery fault diagnosis provided by the present invention;

[0074] Figure 2 This is the flowchart of the improved H5N1 optimization algorithm;

[0075] Figure 3 These are convergence curves of the parameter optimization process for the original H5N1 optimization algorithm and the improved H5N1 optimization algorithm;

[0076] Figure 4 This is a confusion matrix diagram of the fault diagnosis results of H5N1-VMD+CNN;

[0077] Figure 5 It is a confusion matrix diagram of the fault diagnosis results of the improved H5N1-VMD+CNN;

[0078] Figure 6 It is an improved H5N1-VMD-MFWA+CNN fault diagnosis result confusion matrix diagram. Detailed Implementation

[0079] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the implementation examples and accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0080] like Figure 1 An improved method for fault diagnosis of rotating machinery using H5N1-VMD-MFWA noise reduction includes the following steps:

[0081] Step S1: Acquire vibration signals of rotating machinery under normal conditions and with different types of faults;

[0082] Step S2: Using the minimum envelope entropy of the vibration signal as the fitness function, the improved H5N1 optimization algorithm is used to optimize the key parameters of variational mode decomposition. K , α We perform an optimization search to obtain the optimal parameter combination.

[0083] Step S3: Based on the optimal parameter combination [ K , α The vibration signal was subjected to variational mode decomposition to obtain... K One eigenmode function component;

[0084] Step S4: Calculate the correlation coefficient, kurtosis, and envelope entropy of each intrinsic mode function component with the original signal. Use a multi-feature weighted attention mechanism to determine the weight of each component. Normalize and weightedly fuse the scoring index to reconstruct the vibration signal.

[0085] Step S5: Perform window segmentation on the reconstructed vibration signal and perform fault diagnosis on the reconstructed signal based on the convolutional neural network model.

[0086] Furthermore, in step S2, the method for calculating the envelope entropy is as follows:

[0087] E p =-∑ M z=1 ε ( z ) lgε (z )

[0088] ε ( z )= a ( z ) / ∑ M z=1 a ( z )

[0089] in, a ( z The envelope of the signal is obtained by performing a Hilbert transform on the vibration signal of rotating machinery. ε ( z )yes a ( z The normalized envelope sequence, satisfying: ∑ M z=1 ε ( z )=1, ε ( z )≥0, M The number of sampling points. E p Let be the envelope entropy.

[0090] Furthermore, in step S2, the improved avian influenza virus optimization algorithm comprises the following steps:

[0091] The improved avian influenza virus optimization algorithm is in D In a dimensional search space, through a containing N The search was optimized by analyzing the population of individual virus individuals; among them, D A positive integer, representing the dimension of the optimization problem; N The integer is greater than zero, representing the size of the virus population; the position of each individual virus represents a candidate solution to the problem.

[0092] Step S21: "Initialization" phase, setting the virus population size. N With maximum number of iterations T A uniform random initialization method is adopted: X j i = LB +( UB - LB )⊙ R j i Initialize the virus population and calculate the fitness value of each virus to determine the current best individual location and fitness value;

[0093] Where ⊙ represents element-wise multiplication, and the superscript... iRepresents an index of individual viruses. i = 1, 2, …, N subscript j Represents the dimension of the search question space. D index, j = 1, 2, …, D , X represent N OK D A matrix of columns, where X j i Representing the i The virus individual in the first j Positional components in each dimension UB and LB These are the upper and lower bounds of the search space. R This represents elements whose values ​​are in the range [0~1]. N OK D A random number matrix of columns, where R j i It is the first in the matrix i Line number j The numerical components of the column;

[0094] Step S22: In the "population position matrix" construction stage, based on the position update strategy of multiple random permutations, three randomly rearranged population position matrices are constructed. PMP 1, PMP 2 and PMP 3;

[0095] Step S23: The "selection" phase, the virus will select based on the attack probability. P attack Select an attack target, if P attack If the value is less than 0.8, the virus attacks poultry; otherwise, the virus attacks humans.

[0096] Step S24: "Judgment" phase, the virus adapts according to its probability. P adapt Determining whether the host is alive or whether the environment is conducive to viral survival: P adapt When the value is ≥0.85, it indicates that the host has died or the environment is unfavorable, and the virus will escape the current environment and enter the "escape" stage. P adapt When the value is less than 0.85, it indicates that the host is alive or the environment is suitable. The virus location will be fine-tuned to adapt to the host. The location update method is as follows:

[0097] when P attack <0.8 and P adapt When <0.85:

[0098] X j i = PMP 3 j i + R j i ⊙( X g,j - X p,j i )

[0099] when P attack ≥0.8 and P adapt When <0.85:

[0100] X j i = X j i +0.5* R j i ⊙( PMP 1 j i - PMP 2 j i )+0.5* R j i ⊙( X g,j - X j i )

[0101] in, PMP 1 j i , PMP 2 j i and PMP 3 j i Represent PMP 1, PMP 2 and PMP 3 in the i Line number j The numerical components of the column;

[0102] Step S25: The "escape" phase, the virus will adjust its escape probability accordingly. P escape Choose one of two strategies: If P escape < P The virus then executes a mutation strategy; ifP escape ≥ P Then the diffusion strategy is executed, where P The parameters for determining "escape" are generated as follows:

[0103] when d ≥0.4, P =0.7

[0104] When 0.4> d >0.15, P =0.5

[0105] When 0.15≥ d , P =0.3

[0106] in, d To integrate diversity, d =0.7× sr +0.3× fr ;

[0107] in, sr This refers to the relative dispersion of spatial diversity. fr The relative differences in solution quality among populations are calculated as follows:

[0108] sr =mean(std( X p,j )) / max( UB - LB )

[0109] fr =std( F p ) / (abs(mean( F p )))

[0110] Among them, mean() calculates the mean; std() calculates the standard deviation; and abs() calculates the absolute value. X p,j It is the first j The individual's historical optimal position; F p It is the individual's historical best fitness value;

[0111] During the "escape" phase, the virus location update formula is as follows:

[0112] when P attack <0.8 and P adapt ≥0.85 and P escape <P , X new,j i = X j i + R j i ⊙( X g,j - X j i )

[0113] when P attack ≥0.8 and P adapt ≥0.85 and P escape < P , X new,j i = X j i + R j i ⊙( X g,j - X p,j i )

[0114] X mutate,j i =0.5* c *( X new,j i + X g,j )* r * ω

[0115] when P escape ≥ P , X spread,j i = X g,j + R j i ⊙( X new,j i - X g,j )

[0116] If a mutation strategy is executed, then let Xj i =X mutate,j i ;

[0117] If the diffusion strategy is executed, then let X j i =X spread,j i ;

[0118] in, X g,j It is the first j The globally optimal individual position in dimension. X new,j i Individual virus i In the j The escape midpoint in dimensional space; X mutate,j i Individual virus i In the j New positions generated by mutations in a dimension; X spread,j i Individual virus i In the j The new position generated by propagation in dimensions F g It is the globally optimal fitness value. P attack , P adapt , P escape and r It is a random number in the range (0,1). c It is the exponential decay coefficient:

[0119] c = r × e -4×t / T

[0120] ω It is a weighting factor:

[0121] ω = ω × ω d

[0122] ω d = ω min +( ω max - ωmin )× e -(t / T)

[0123] in, t Indicates the current iteration number. ω d The damping coefficient representing the weight. ω max and ω min These represent the maximum and minimum values ​​of the weights, respectively. ω The initial value is 1;

[0124] Step S26: After each iteration, calculate the fitness value of all new virus positions and compare the new fitness value with the individual's historical best value. If the new fitness value is better, update the individual's historical best position and historical best fitness value.

[0125] At the same time, the global optimal solution is updated based on a probabilistic mechanism, that is, only when... ps Less than the global update probability G The global optimal position and global optimal fitness value will only be updated if there is an individual fitness value that is better than the global optimal fitness value; among which, ps Given a random number in the range (0,1), the global update probability is... G The initial value is 0, and the update process is as follows:

[0126] G =(1 / 1+e -10(t / T-0.5) )+ r* (1- G )

[0127] Among them, when the number of iterations t achieve T When the iteration count reaches zero, the algorithm terminates the iteration and outputs the global optimal position and its fitness value as the optimal solution; if the iteration count is zero... t Failed to achieve T If so, return to step S22 and repeat steps S22 to S26.

[0128] Furthermore, in step S4, the steps of determining the weight of each component using a multi-feature weighted attention mechanism, normalizing and weighted fusing the scoring index, and reconstructing the vibration signal are as follows:

[0129] Step S41: Based on the decomposition K The positive ideal reference benchmark for each index is determined by the maximum correlation coefficient, maximum kurtosis, and minimum envelope entropy of each intrinsic mode function component. K The minimum correlation coefficient, minimum kurtosis, and maximum envelope entropy of each IMF component determine the negative ideal reference benchmark for each indicator.

[0130] Step S42: Standardize the evaluation indicators by using a normalization method to map each indicator to the [0,1] interval, thus eliminating the influence of dimensionality.

[0131] Normalized value = (current value - negative ideal benchmark) / (positive ideal benchmark - negative ideal benchmark)

[0132] Step S43: Set three weights: correlation coefficient weight, kurtosis weight, and envelope entropy weight, and calculate the final score of each intrinsic mode function component based on the weighted comprehensive scoring method; wherein, correlation coefficient weight + kurtosis weight + envelope entropy weight = 1;

[0133] Step S44: Convert the overall score into attention weights using the softmax function to achieve a probability distribution transformation of the score;

[0134] Step S45: Weight all intrinsic mode function components based on attention weights to form a new signal.

[0135] Furthermore, in step S22, the construction of three randomly rearranged population position matrices... PMP 1, PMP 2 and PMP 3. Its generation process is as follows:

[0136] Step S221: Generate the basic index vector r o =[0,1,2,..., N -1], and generate a random permutation vector. σ = [ σ 1 , σ 2 , σ 3], used to control the cyclic offset of subsequent indices, where σ 1 , σ 2 , σ 3 is a random permutation of the numbers 1, 2, and 3;

[0137] Step S222: Generate initial random index vector a 1, a 1=[ j 1, j 2,..., j N ],in, j 1, j 2,..., j N It is the interval [1, N A random permutation of positive integers;

[0138] Step S223: Utilize the first element of σσ 1 pair of basic index vectors r o Perform a cyclic offset to obtain the intermediate index vector. r t1 The cyclic offset is achieved by: r o Each index value in the database is related to... σ 1 after addition N Use the remainder to ensure the result index is between 0 and 1. N Cycle within the range of -1;

[0139] Step S224: Based on r t1 right a 1. Perform index remapping, that is, use the index vector [ r t1 +1] From vector a Extract the element at the corresponding position from step 1 to obtain the second random index vector. a 2;

[0140] Step S225: Use σ The second element ​ 2. Again r o Perform a cyclic offset to obtain the intermediate index vector. r t2 and based on r t2 right a 2. Perform index remapping to obtain a third random index vector. a 3;

[0141] Step S226: Finally, calculate the optimal historical position matrix of the virus. X p Row index extraction is performed to construct three random permutation matrices:

[0142] ​ 1= X p ( a 1,:)

[0143] ​ 2= X p ( a 2,:)

[0144] ​ 3= X p ( a 3,:)

[0145] The method of the present invention will be explained below using a specific case of rotating machinery failure as an example.

[0146] Example 1: Taking the Case Western Reserve University bearing failure (CWRU) dataset as an example. In the experiment, drive-end (DE) data with a sampling frequency of 12kHz and a rotational speed of 1797RPM was selected, and its data information is shown in Table 1.

[0147] Table 1 CWRU Dataset Parameter Table

[0148]

[0149] The specific implementation process is as follows:

[0150] Step S1: Acquire vibration signals and introduce -12 dB Gaussian white noise into the vibration signals to simulate strong background noise interference in actual industrial scenarios, thereby verifying the robustness of the proposed method under low signal-to-noise ratio conditions.

[0151] Step S2: Set VMD parameters K and α The scope will K The value range is set to (3, 10). α The value range is set to (100, 2500) in VMD. K The value represents the number of intrinsic mode functions (IMFs) to be decomposed; when K When the value is 2, it is decomposed into two IMF components named IMF1 and IMF2; when K When the value is 3, it is decomposed into three IMF components named IMF1, IMF2, and IMF3, and so on; using the minimum envelope entropy as the fitness function, the improved H5N1 optimization algorithm is used to optimize the parameters of VMD to obtain the optimal parameters. K , α Combinations, where the population size of the algorithm is... N Set to 50, maximum number of iterations T Set to 30, ​ max =1 and ​ min =0.001;

[0152] The flowchart of the improved avian influenza virus optimization algorithm is as follows: ​ As shown in the figure, the convergence curve of the parameter optimization process of the original avian influenza virus optimization algorithm is as follows: ​ As shown;

[0153] Step S3: Based on optimal parameters [ K , α The vibration signal is combined and decomposed using VMD to obtain... K One IMF component;

[0154] Step S4: Assign weights to each IMF component and reconstruct the signal based on the correlation coefficient, kurtosis value, and envelope entropy of each IMF component using the MFWA mechanism; where the correlation coefficient weight is 0.4, the kurtosis weight is 0.3, and the envelope entropy weight is 0.3; the signal index evaluation results are shown in Tables 2 and 3.

[0155] Step S5: The reconstructed vibration signal is windowed and input into a one-dimensional CNN model for fault diagnosis. The main body of the model consists of four stacked convolutional layers (each with a ReLU activation function) and three pooling layers forming a feature extractor. The hierarchical connection relationship is as follows: the first two levels are a "convolution-convolution-pooling" structure, and the last two levels are a cascade of "convolution-pooling". The extracted features are flattened and then fed into a classifier consisting of two fully connected layers, with a Dropout layer introduced in between to prevent overfitting. Finally, the output layer provides the diagnostic results. The evaluation results are shown in Table 4. ​ , ​ and ​ The confusion matrices corresponding to the original avian influenza virus optimization algorithm, the improved avian influenza virus optimization algorithm, and the improved avian influenza virus optimization algorithm combined with a multi-feature weighted attention mechanism are shown respectively.

[0156] Table 2 Comparison Results of Negative Indicators

[0157]

[0158] Table 3 Comparison Results of Positive Indicators

[0159]

[0160] Table 4. Model Diagnostic Accuracy and F1 Score

[0161]

[0162] As shown in Tables 2 to 4, the diagnostic results of the present invention are significantly better than those of the comparative method, which fully verifies the advantages of the present invention in diagnosing rotating machinery faults.

[0163] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the invention by those skilled in the art. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the invention should be included within the scope of protection of the invention.

Claims

1. A method for improved H5N1-VMD-MFWA denoising of rotating machinery fault diagnosis, characterized in that, The method comprises the following steps: Step S1: obtaining rotating machinery vibration signals of normal and different types of faults; Step S2: The minimum envelope entropy of the vibration signal is taken as the fitness function, and the improved avian influenza virus optimization algorithm is used to optimize and search the key parameters of the variational mode decomposition K , α ]. Step S3: based on the optimal parameter combination K , α ], the vibration signal is decomposed by variational modal decomposition to obtain K IMF components; Step S4: calculating correlation coefficients, kurtosis and envelope entropy of each intrinsic mode function component and the original signal, determining the weight of each component by using a multi-feature weighted attention mechanism, normalizing and weighting the score indicators, and reconstructing the vibration signal; Step S5: performing window segmentation on the reconstructed vibration signal, and performing fault diagnosis on the reconstructed signal based on a convolutional neural network model; In step S2, the improved avian influenza virus optimization algorithm has the following steps: The improved avian influenza virus optimization algorithm is in D In a dimensional search space, through a containing N The search was optimized by analyzing the population of individual virus individuals; among them, D A positive integer, representing the dimension of the optimization problem; N The integer is greater than zero, representing the size of the virus population; the position of each individual virus represents a candidate solution to the problem. Step S21: "Initialization" stage, set the virus population size N With the maximum number of iterations T , using a uniform random initialization method: X j i = LB +( UB - LB )⊙ R j i , initialize the virus population and calculate the fitness value of each virus, determine the current best individual position and fitness value; Where ⊙ represents element-wise multiplication, and the superscript... i Represents an index of individual viruses. i=1, 2, …, N subscript j Represents the search question space dimension D index, j=1, 2, …, D , X represent N OK D A matrix of columns, where X j i Representing the i The virus individual in the first j Positional components in each dimension UB and LB These are the upper and lower bounds of the search space. R This represents elements whose values ​​are in the range [0~1]. N OK D A random number matrix of columns, where R j i It is the first in the matrix i Line number j The numerical components of the column; Step S22: "population position matrix" construction phase, based on the position update strategy of multiple random permutation, three randomly rearranged population position matrices are constructed PMP 1, PMP 2 and PMP 3; Step S23: "selection" phase, the virus will be based on the probability of attack P attack select the target of attack, if P attack <0.8, the virus attacks poultry; otherwise, the virus attacks humans; Step S24: "decision" stage, the virus according to the adaptive probability P adapt Determine whether the host is alive or the environment is conducive to the survival of the virus: P adapt ≥ 0.85, indicating that the host is dead or the environment is not conducive, and the virus will escape the current environment and enter the "escape" stage; P adapt <0.85, indicating that the host is alive or the environment is suitable, and the virus position will be fine-tuned to adapt to the host. The position update method is as follows: When P attack <0.8 and P adapt <0.85: X j i = PMP 3 j i + R j i ⊙( X g,j - X p,j i ) When P attack ≥ 0.8 and P adapt < 0.85: X j i = X j i +0.5* R j i ⊙( PMP 1 j i - PMP 2 j i )+0.5* R j i ⊙( X g,j - X j i ) wherein PMP 1 j i , PMP 2 j i and PMP 3 j i represent the value components of the number in column PMP 1, PMP 2 and PMP 3 in row i 1 j 2. Step S25: "Escape" phase, the virus will be according to the escape probability P escape Two strategies are selected: if P escape P The virus executes the mutation strategy; if P escape ≥ P The diffusion strategy is executed, wherein P is the "escape" decision parameter, and the generation method is:​ When d ≥ 0.4, P = 0.7 when 0.4 d > 0.15, P = 0.5 When 0.15 ≥ d , P = 0.3 wherein, d for the synthesis of diversity, d = 0.7 x sr + 0.3 x fr ; wherein, sr for the relative degree of spatial diversity; fr for the relative degree of difference in the solution quality level of the population, the calculation method is: sr = mean(std( X p,j )) / max( UB - LB ) fr =std( F p ) / (abs(mean( F p ))) where mean() is the mean, std() is the standard deviation, and abs() is the absolute value. X p,j is the j dimensional individual historical best position; F p is the individual historical best fitness value; In the "escape" stage, the virus position update formula is as follows: When P attack <0.8 and P adapt ≥ 0.85 and P escape < P , X new,j i = X j i + R j i ⊙( X g,j - X j i ) When P attack ≥ 0.8 and P adapt ≥ 0.85 and P escape P , X new,j i X j i R j i ⊙( X g,j X p,j i )​​​​ X mutate,j i =0.5* c *( X new,j i + X g,j )* r * ω When P escape ≥ P , X spread,j i = X g,j + R j i ⊙( X new,j i - X g,j ) If the mutation strategy is executed, then X j i =X mutate,j i ; If the diffusion strategy is executed, then X j i =X spread,j i ; in, X g,j It is the first j The globally optimal individual position in dimension. X new,j i Individual virus i In the j The escape midpoint in dimensional space; X mutate,j i Individual virus i In the j New positions generated by dimensional mutations; X spread,j i Individual virus i In the j The new position generated by propagation in dimensions F g It is the globally optimal fitness value. P attack , P adapt , P escape and r It is a random number in the range (0,1). c It is the exponential decay coefficient: c = r × e -4×t / T ω is a weight factor: ω = ω × ω d ω d = ω min +( ω max - ω min )× e -(t / T) wherein, t denotes the current iteration number, ω d denotes a damping coefficient for the weights, ω max and ω min denote the maximum and minimum values for the weights, respectively, ω the initial value of is 1 ; Step S26: After each iteration, the fitness value of the new position of all viruses is calculated, and the new fitness value is compared with the historical optimal value of the individual. If the new fitness value is better, the historical optimal position and the historical optimal fitness value of the individual are updated. At the same time, the global optimal solution is updated based on a probabilistic mechanism, i.e. only when ps the global update probability is less than 1 G and there is an individual fitness value better than the global optimal fitness value, the global optimal position and the global optimal fitness value are updated; wherein, ps is a random number in the range of (0, 1), the global update probability G is initially 0, and the updating process is as follows: G =(1 / 1+e -10(t / T-0.5) )+ r* (1- G ) wherein the algorithm terminates the iteration and outputs the global optimal position and its fitness value as the optimal solution when the iteration number t reaches T ; if the iteration number t fails to reach T , the method returns to step S22 and repeats steps S22-S26. In step S4, the weight of each component is determined by using a multi-feature weighted attention mechanism, the score indicators are normalized and weighted, and the vibration signal is reconstructed. The steps are as follows: Step S41: determining the positive ideal reference benchmark of each index according to the correlation coefficient maximum value, the kurtosis maximum value, and the envelope entropy minimum value of each intrinsic mode function component decomposed in step S40; and determining the negative ideal reference benchmark of each index according to the correlation coefficient minimum value, the kurtosis minimum value, and the envelope entropy maximum value of each intrinsic mode function component decomposed in step S40. K Step S41: determining the positive ideal reference benchmark of each index according to the correlation coefficient maximum value, the kurtosis maximum value, and the envelope entropy minimum value of each intrinsic mode function component decomposed in step S40; and determining the negative ideal reference benchmark of each index according to the correlation coefficient minimum value, the kurtosis minimum value, and the envelope entropy maximum value of each intrinsic mode function component decomposed in step S40. K Step S41: determining the positive ideal reference benchmark of each index according to the correlation coefficient maximum value, the kurtosis maximum value, and the Step S42: standardize the evaluation indicators, and use a normalization method to map each indicator to the [0, 1] interval to eliminate the influence of the dimension: Normalized value=(current value-negative ideal reference) / (positive ideal reference-negative ideal reference) Step S43: set three weights: correlation coefficient weight, kurtosis weight and envelope entropy weight, and calculate the final score of each intrinsic mode function component based on the three weights using a weighted comprehensive scoring method; wherein the correlation coefficient weight+the kurtosis weight+the envelope entropy weight=1; Step S44: convert the comprehensive score into an attention weight by using a softmax function to realize the probability distribution conversion of the score; Step S45: based on the attention weight, all intrinsic mode function components are weighted to form a new signal.

2. The improved H5N1-VMD-MFWA rotating machinery fault diagnosis method of claim 1, wherein, In step S2, the envelope entropy calculation method is as follows: E p =-∑ M z=1 ε ( z ) lgε ( z ) ε ( z )= a ( z ) / ∑ M z=1 a ( z ) Wherein, a z is the envelope line of the signal, and the Hilbert transform is performed on the rotating machinery vibration signal to obtain; ε z a z is the normalized envelope sequence, and satisfies: ∑ M z=1 ε z =1, ε z ≥0, M is the number of sampling points, E p is the envelope entropy, z represents the sampling point sequence number of the signal.​​​​​​ 3. The improved H5N1-VMD-MFWA rotating machinery fault diagnosis method of claim 1, wherein, In step S22, the three randomly rearranged population position matrices are constructed PMP 1, PMP 2 and PMP 3 are generated as follows: Step S221: Generating the base index vector r o =[0,1,2,..., N -1] and a random permutation vector σ= [ σ 1 , σ 2 ,σ 3] is generated for controlling the cyclic shift of the subsequent indices, where σ 1 ,σ 2 ,σ 3 is a random permutation of the numbers 1,2,3; Step S222: generating an initial random index vector a 1, a 1=[ j 1, j 2,..., j N ], wherein, j 1, j 2,..., j N is a random permutation of positive integers in the interval [1, N ] ; Step S223: Utilize the first element of σ σ 1 pair of basic index vectors r o Perform a cyclic offset to obtain the intermediate index vector. r t1 The cyclic offset is achieved by: r o Each index value in the database is related to... σ 1 after addition N Use the remainder to ensure the result index is between 0 and 1. N Cycle within the range of -1; Step S224: According to r t1 To a 1, that is, the elements of the corresponding positions in the vector r t1 +1] are extracted from the vector a 1 to obtain the second random index vector a 2; Step S225: Use σ The second element σ 2. Again r o Perform a cyclic offset to obtain the intermediate index vector. r t2 and based on r t2 right a 2. Perform index remapping to obtain a third random index vector. a 3; Step S226: Finally, the virus history optimal position matrix is extracted respectively X p The row index extraction is performed to construct three random permutation matrices: PMP 1= X p ( a 1,:) PMP 2= X p ( a 2,:） PMP 3= X p ( a 3,:)。

Images