A deconvolution feature extraction method by minimizing the amplitude ratio of periodic noise

By minimizing the periodic noise amplitude ratio (PNAR) sparsity metric and optimizing the FIR filter using the RMSProp algorithm, the problem of blind deconvolution algorithms being unable to distinguish between strong and weak impacts in rotating machinery is solved, enabling more accurate fault diagnosis and health monitoring.

CN114564986BActive Publication Date: 2024-11-08SOUTHEAST UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210159676.X
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-02-21
Publication Date
2024-11-08
Estimated Expiration
2042-02-21

AI Technical Summary

Technical Problem

Existing blind deconvolution algorithms struggle to effectively distinguish the intensity of impacts when processing periodic fault impact signals from rotating machinery. Their performance is particularly limited in cases with multiple resonance bands and low signal-to-noise ratio signals, leading to difficulties in fault diagnosis.

Method used

Using the minimized periodic noise amplitude ratio (PNAR) as a sparsity metric, a blind deconvolution algorithm based on PNAR minimization is proposed. This algorithm optimizes the filtered signal to enhance the impact characteristics of periodic faults by constructing an FIR filter and updating the filter parameters using the RMSProp algorithm.

Benefits of technology

It effectively enhances the extraction of periodic fault impact characteristics in noisy environments, provides more accurate fault diagnosis and health status monitoring basis, and adapts to various extreme environments.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114564986B_ABST
    Figure CN114564986B_ABST
Patent Text Reader

Abstract

The present invention relates to a deconvolution feature extraction method for minimizing the periodic noise amplitude ratio, comprising the following steps: S1, collecting the acceleration signal of a rotating machine, determining the fault cycle and the maximum number of iterations of the fault type to be detected; S2, initializing an FIR filter using a random number that obeys a normal distribution; S3, filtering the acceleration signal using the FIR filter and calculating the periodic noise amplitude ratio of the filtered signal; S4, calculating the gradient of the periodic noise amplitude ratio; S5, updating the value of the FIR filter by the RMSProp algorithm so that the periodic noise amplitude ratio of the filtered signal decreases; S6, repeating S3-S5, and outputting the filtered signal when the maximum number of iterations is reached. The present invention can enhance the periodic impact characteristics excited by the rotating machine fault, and provide an accurate calculation basis for rotating machine fault diagnosis and health status monitoring.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to the technical field of rotating machinery fault diagnosis, in particular to a method for extracting deconvolution features by minimizing the periodic noise amplitude ratio. Background Art

[0002] When a rotating machinery component is defective, the defective area collides with the surface of other components to produce periodic impacts. Detecting and analyzing periodic impacts can provide an important basis for fault identification and location of rotating machinery, and is of great significance for the assessment of the health status of rotating machinery. However, these periodic fault impacts are interfered by the transmission path of the mechanical system and environmental noise, and they are usually very weak or even completely invisible in the measured signal. The blind deconvolution algorithm constructs an FIR filter to maximize a certain statistical indicator (maximization criterion) of the filtered signal to restore the characteristics of periodic fault impacts.

[0003] Different blind deconvolution algorithms are distinguished by maximization criteria. Existing blind deconvolution algorithms, including blind deconvolution based on kurtosis maximization-MED, blind deconvolution based on correlation kurtosis maximization-MCKD, blind deconvolution based on multi-point optimal norm maximization-MOMEDA, blind deconvolution based on second-order cyclostationarity maximization-CYCBD, etc., are greatly limited in performance because their maximization criteria cannot well characterize the strength of periodic shocks. When dealing with multiple resonance band signals and low signal-to-noise ratio signals, the existing blind deconvolution algorithms are very likely to fail completely. This restricts the industrial application of existing blind deconvolution algorithms.

[0004] Therefore, a criterion is designed to well characterize the impact intensity of periodic faults of rotating machinery. Based on the maximization (or minimization) of this criterion, a blind deconvolution algorithm that can adapt to various extreme environments is proposed, which is of great significance for the early diagnosis of rotating machinery faults and health status monitoring. Summary of the invention

[0005] In view of the defects of the prior art, the present invention provides a deconvolution feature extraction method for minimizing the periodic noise amplitude ratio, which solves the problem that the existing algorithm has insufficient performance and is difficult to process weak fault signals.

[0006] The technical solution adopted by the present invention is as follows:

[0007] A method for extracting features by deconvolution by minimizing the periodic noise amplitude ratio comprises the following steps:

[0008] S1: Collect the acceleration signal x of the rotating machinery, determine the fault period T and the maximum number of iterations N of the fault type to be detected;

[0009] S2: Initialize the FIR filter f using random numbers that obey the normal distribution;

[0010] S3: Filter the acceleration signal x using the FIR filter f to obtain a filtered signal y=x*f, where * represents a convolution operation, and calculate the periodic noise amplitude ratio of the filtered signal

[0011] Among them, abs(·) represents the absolute value, rms(·) represents the root mean square, · represents the inner product of two vectors, and M represents t noise The number of noise points, PNAR represents the periodic noise amplitude ratio of the filtered signal;

[0012] Periodic noise distribution vector t noise The vector has the same length as y, where 0 indicates that the position is a non-noise point, and 1 indicates that the position is a noise point. The noise points and non-noise points are continuous within a fault cycle T, and the proportion of the number of noise points is [0.5, 0.8];

[0013] S4: Solve the gradient of the periodic noise amplitude ratio PNAR to the FIR filter f

[0014] S5: updating the value of the FIR filter f by the RMSProp algorithm so that the periodic noise amplitude ratio of the filtered signal decreases;

[0015] S6: Repeat S3 to S5, and when the maximum number of iterations N is reached, output the filtered signal y.

[0016] Its further technical solution is:

[0017] Step S4, solving the gradient of the periodic noise amplitude ratio PNAR to the FIR filter f, comprises the following steps:

[0018] Calculate the gradient of PNAR with respect to the filtered signal y:

[0019]

[0020] Compute the gradient of y with respect to the FIR filter f:

[0021]

[0022] The gradient of the periodic noise amplitude ratio PNAR to the FIR filter f is calculated according to the chain rule:

[0023]

[0024] Where g represents the gradient vector.

[0025] Step S5, updating the value of the FIR filter f by the RMSProp algorithm, comprises the following steps:

[0026] Compute the square of the cumulative gradient:

[0027] r t =αr t-1 +(1-α)·g t ⊙g t

[0028] Among them, g t The vector represents the gradient vector of the tth iteration, r t The vector represents the cumulative squared gradient at the tth iteration, t∈[1,N], r t Vector and g t The number of elements in the vectors is the same; r t-1 The vector represents the cumulative gradient square of the t-1th iteration; all elements of the initialization r0 vector are 0, α represents the decay rate, α∈(0,1); the ⊙ operation represents the multiplication of the corresponding elements of two vectors;

[0029] Update the value of the filter f:

[0030]

[0031] Among them, l r represents the learning rate, ε is 0.000001; Represents r t +The vector obtained by performing the "square root" operation on the elements in the ε vector, Indicates g t Vector and The vectors obtained by performing corresponding "division" operations on the elements in the vector respectively. vector, g t Vector and r t The number of vector elements is the same; f t-1 represents the value of the filter f at the t-1th iteration, f t is the value of the filter f at the tth iteration.

[0032] In step S2, the length of the FIR filter f is in the range of [1.5T·f s ]~[3T·f s ], [] is the rounding symbol, f s is the sampling frequency, and T is the fault period.

[0033] The maximum number of iterations N in step S1 is in the range of 50≤N≤300.

[0034] The beneficial effects of the present invention are as follows:

[0035] By constructing the sparsity measurement indicator of periodic noise amplitude ratio (PNAR), the present invention proposes a blind deconvolution algorithm based on PNAR minimization. The algorithm can well enhance the periodic fault impact characteristics hidden in the noise environment, and provide more basis for fault diagnosis and health status monitoring of rotating machinery. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is a schematic diagram of the process of the present invention.

[0037] Figure 2 The fault signal of the inner ring of the rolling bearing of the rotating machinery collected in the specific embodiment of the present invention.

[0038] Figure 3 It is a filtered signal calculated in a specific embodiment of the present invention.

[0039] Figure 4 The spectrum of the filtered signal calculated according to the specific embodiment of the present invention.

[0040] Figure 5 It is the envelope spectrum of the filtered signal calculated according to the specific embodiment of the present invention. DETAILED DESCRIPTION

[0041] The solution of the present application is further described below through specific examples:

[0042] For reference Figure 1 A method for extracting features by deconvolution by minimizing the periodic noise amplitude ratio of the present application comprises the following steps:

[0043] S1: with f s The acceleration signal x of the rotating machinery is collected at a sampling frequency, the type of fault to be detected is determined, the fault cycle T is calculated by the fault type and the parameters of the rotating machinery parts, and the maximum number of iterations N is determined;

[0044] S2: construct a suitable FIR filter f;

[0045] The length L of f has a value range of [1.5T·f s ]~[3T·f s ], [] is the rounding symbol, f s is the sampling frequency.

[0046] Initialize the FIR filter f using random numbers that satisfy the standard normal distribution;

[0047] S3: Use FIR filter f to filter the original acceleration signal x. The filtering process is expressed as y=x*f, where * represents convolution operation:

[0048] Calculate the periodic noise amplitude ratio PNAR of the filtered signal y, including:

[0049] First, calculate the periodic noise distribution vector t noise :

[0050]

[0051] Among them, t noise The vector has the same length as y, where 0 indicates that the position is a non-noise point and 1 indicates that the position is a noise point. Within a fault cycle T, both the noise point and the non-noise point are continuous, and the proportion of the number of noise points is [0.5, 0.8], that is, the proportion is 50% to 80%, and the brackets represent the boundary value.

[0052] Then, the periodic noise amplitude ratio PNAR of the filtered signal y is calculated:

[0053]

[0054] Among them, abs(·) represents the absolute value, rms(·) represents the root mean square, · represents the inner product of two vectors, and M represents t noise PNAR stands for the periodic noise amplitude ratio, which physically means the ratio of the average amplitude of the periodic noise to the root mean square of the filtered signal.

[0055] The PNAR sparsity metric proposed in this application is a good criterion for characterizing the impact strength of periodic faults in rotating machinery. The blind deconvolution algorithm proposed by maximizing (or minimizing) this criterion can adapt to various extreme environments.

[0056] S4: Solve the gradient g of the periodic noise amplitude ratio PNAR to the FIR filter, specifically including:

[0057] Calculate the gradient of PNAR with respect to the filtered signal y:

[0058]

[0059] Compute the gradient of y with respect to the FIR filter f:

[0060]

[0061] The gradient of the periodic noise amplitude ratio PNAR to the FIR filter f is calculated according to the chain rule:

[0062]

[0063] Among them, g t represents the gradient vector.

[0064] The differential is automatically calculated based on the backward automatic differentiation algorithm, and there is no need to manually derive the differential formula.

[0065] S5: Updating the value of the FIR filter f by the RMSProp algorithm, including the following steps:

[0066] Compute the square of the cumulative gradient:

[0067] r t =αr t-1 +(1-α)·g t ⊙g t

[0068] where g t The vector represents the gradient vector of the tth iteration, r t The vector represents the cumulative squared gradient at the tth iteration, t∈[1,N], r t Vector and g t The number of elements in the vectors is the same; r t-1 The vector represents the cumulative gradient square of the t-1th iteration; all elements of the initialization r0 vector are 0, α represents the decay rate, α∈(0,1); the ⊙ operation represents the multiplication of the corresponding elements of two vectors;

[0069] Update the value of the filter f:

[0070]

[0071] Among them, l r represents the learning rate, ε is 0.000001; Represents r t +The vector obtained by performing the "square root" operation on the elements in the ε vector, Indicates g t Vector and The vectors obtained by performing corresponding "division" operations on the elements in the vector respectively. vector, g t Vector and r t The number of vector elements is the same; f t-1 represents the value of the filter f at the t-1th iteration, f t is the value of the filter f at the tth iteration.

[0072] S6: Repeat S3 to S5. After reaching the maximum number of iterations N, the filtered signal y is output for subsequent analysis.

[0073] The technical solution of the present application is further illustrated below with specific embodiments.

[0074] The 6205 rolling bearing inner ring fault signal is collected by the acceleration sensor, and the sampling frequency is f sThe frequency of the inner ring fault is 12000Hz, the shaft speed is 1800r / min, and the inner ring fault characteristic frequency BPFI can be calculated to be 161Hz through the bearing size parameters, and then the fault period can be calculated:

[0075]

[0076] The original vibration signal collected is as follows: Figure 2 As shown in the figure, the periodic fault pulse characteristics in the original signal are submerged by the noise component, and the health status of the rolling bearing cannot be determined.

[0077] Set the maximum number of iterations N = 300, and the length of the filter f L = 200. Use random numbers that satisfy the normal distribution to initialize the filter f.

[0078] The original fault signal is filtered by filter f and the periodic noise amplitude ratio of the filtered signal is calculated. The ratio of noise points ρ = 0.6;

[0079] Solve the gradient of the periodic noise amplitude comparison filter f;

[0080] The filter f is updated by the RMSProp algorithm, where the learning rate of the RMSProp algorithm is 0.01.

[0081] After 300 iterations, the periodic noise amplitude ratio of the filtered signal reaches the minimum, and the output filtered signal is as follows: Figure 3 As shown, the periodic fault impact characteristics of the filtered signal are very obvious.

[0082] The spectrum of the filtered signal is as follows Figure 4 As shown, the present invention adaptively detects that there are two resonance bands in the original signal and filters the frequency components of the non-resonance bands.

[0083] The envelope spectrum of the filtered signal is as follows: Figure 5 As shown, the fault characteristic frequency BPFI and its frequency multiple components are obvious, so it can be determined that the inner ring fault of the rolling bearing of the rotating machinery has occurred.

[0084] The algorithm of the present application can enhance the periodic impact characteristics of rotating machinery fault excitation, and provide an accurate calculation basis for rotating machinery fault identification and diagnosis, and health status monitoring.

Claims

1. A deconvolution feature extraction method for minimizing the periodic noise amplitude ratio, characterized in that: The following steps are involved: S1: Collect the acceleration signal x of the rotating machinery and determine the fault period T and the maximum number of iterations N of the fault type to be detected; S2: Initialize the FIR filter f using random numbers that obey the normal distribution; S3: Filter the acceleration signal x using the FIR filter f to obtain a filtered signal y = x*f, where * represents a convolution operation, and calculate the periodic noise amplitude ratio of the filtered signal Among them, abs(·) represents the absolute value, rms(·) represents the root mean square calculation, · represents the inner product of two vectors, and M represents t noise The number of noise points, PNAR represents the periodic noise amplitude ratio of the filtered signal; Periodic noise distribution vector t noise The vector has the same length as y, where 0 indicates that the location is a non-noise point and 1 indicates that the location is a noise point. Within a fault cycle T, the noise points and non-noise points are continuous, and the proportion of the number of noise points is [0.5, 0.8]. S4: Calculate the gradient of the periodic noise amplitude ratio PNAR to the FIR filter f S5: updating the value of the FIR filter f by the RMSProp algorithm so that the periodic noise amplitude ratio of the filtered signal decreases; S6: Repeat S3 to S5, and output the filtered signal y when the maximum number of iterations N is reached.

2. The deconvolution feature extraction method for minimizing the periodic noise amplitude ratio according to claim 1, characterized in that: Step S4, solving the gradient of the periodic noise amplitude ratio PNAR with respect to the FIR filter f, comprises the following steps: Calculate the gradient of PNAR with respect to the filtered signal y: Compute the gradient of y with respect to the FIR filter f: The gradient of the periodic noise amplitude ratio PNAR to the FIR filter f is calculated according to the chain rule: Where g represents the gradient vector.

3. The deconvolution feature extraction method for minimizing the periodic noise amplitude ratio according to claim 1, characterized in that: Step S5, updating the value of the FIR filter f by the RMSProp algorithm, comprising the following steps: Compute the square of the cumulative gradient: r t =αr t-1 +(1-a)·g t ⊙g t Among them, g t The vector represents the gradient vector of the t-th iteration, r t The vector represents the cumulative squared gradient of the tth iteration, t∈[1,N], r t Vector and g t The number of elements in the vectors is the same; r t-1 The vector represents the cumulative squared gradient of the t-1th iteration; all elements of the r0 vector are initialized to 0, α represents the decay rate, α∈(0,1); the ⊙ operation represents the multiplication of the corresponding elements of the two vectors; Update the value of the filter f: Among them, l r represents the learning rate, ε is 0.000001; Represents r t +The vector obtained by performing "square root" operation on the elements in the ε vector, Indicates g t Vector and The vectors obtained by performing corresponding "division" operations on the elements in the vector respectively. vector, g t Vector and r t The number of vector elements is the same; f t-1 represents the value of the filter f at the t-1th iteration, f t is the value of the filter f at the tth iteration.

4. The deconvolution feature extraction method for minimizing the periodic noise amplitude ratio according to claim 1, wherein: The length L of the FIR filter f in step S2 is in the range of [1.5T·f s ]~[3T·f s ], [] is the rounding symbol, f s is the sampling frequency, and T is the fault period.

5. The deconvolution feature extraction method for minimizing the periodic noise amplitude ratio according to claim 1, characterized in that: The maximum number of iterations N in step S1 is in the range of 50≤N≤300.