Device structure performance condition monitoring method based on layered depth dynamic potential

Abstract

The invention discloses a device structure performance condition monitoring method based on layered depth dynamic potential. The method comprises the following steps: determining a device easy-to-damage position based on discrete mechanics; obtaining key mechanical response characteristics of the vulnerable position based on structured compressed sensing; constructing a structural performance condition monitoring model based on global capture and local fusion; characteristic enhancement of real-time structure performance condition interactive monitoring poles is realized; and solving the structural performance resonance measurement of the device. Through hierarchical processing and depth potential modeling, the problems that a traditional method is insufficient in local anomaly recognition capability and limited in global prediction precision are effectively solved, and higher anomaly detection sensitivity, higher robustness and better prediction accuracy are shown in experimental comparison; the method is especially suitable for real-time structure performance condition monitoring under complex working conditions.

Description

Technical Field

[0001] The present invention relates to the field of equipment structure performance status monitoring, and in particular to a device structure performance status monitoring method based on layered depth dynamic potential. Background Art

[0002] Equipment often performs tasks in complex and ever-changing environments, making them prone to structural fatigue, component wear, and loose connections. To ensure the long-term, stable operation of equipment in its intended operating environment, structural performance monitoring technology has become a key means of improving its reliability and intelligence. Through modal analysis, fault identification, feature extraction, and machine learning methods, real-time perception of equipment operating status, anomaly detection, and lifespan prediction are achieved. Especially under dynamic operating conditions, the loads on equipment exhibit strong nonlinearity, sudden changes, and uncertainty, making it difficult to effectively detect early damage or potential failures through traditional periodic maintenance alone.

[0003] Developing a high-precision, high-stability structural performance monitoring system that integrates structural simulation, field measurements, and data-driven models has important theoretical significance and engineering application value for improving the intelligence level of equipment, reducing operation and maintenance costs, and extending service life. Therefore, it is necessary to strengthen the interaction between the overall and local mechanical characteristics of the device, taking into account the complex operating environment, and thus realize the monitoring of the device's structural performance status. Summary of the Invention

[0004] Purpose of the invention: The purpose of the present invention is to provide a method for monitoring the performance status of a device structure based on the dynamic potential of layered depth.

[0005] Technical solution: The device structure performance condition monitoring method based on layered depth dynamic potential of the present invention comprises the following steps:

[0006] (1) Determination of the vulnerable position of the device based on discrete mechanics;

[0007] (2) Acquisition of key mechanical response features of vulnerable locations based on structured compressed sensing;

[0008] (3) Construction of a structural performance monitoring model based on global capture and local integration;

[0009] (4) Feature enhancement of the extreme points for interactive monitoring of real-time structural performance conditions;

[0010] (5) Solution of the resonance metric of the structural performance of the device.

[0011] Furthermore, the step (1) includes establishing a three-dimensional model of the structure in the three-dimensional software SolidWorks based on the known device, accurately describing the shape, size and relative position of each component and structure.

[0012] Furthermore, the step (1) includes importing the established three-dimensional model into ABAQUS software, performing discretization processing, carrying out structural statics simulation, identifying stress concentration areas based on stress cloud maps and deformation fields, and determining primary vulnerable location clusters based on Birnbaum-Saunders distribution statistics;

[0013] Apply frequency domain loads to candidate clusters through the harmonic response analysis module, solve the structural dynamic strain-frequency response function, and use the peak extraction algorithm to obtain the extreme value of the amplitude spectrum of each node; define the amplitude sensitivity coefficient With η max The corresponding topological units are used as key vulnerable sites, and the precise positioning of the vulnerable positions of the device is achieved through static-dynamic joint simulation and discrete parameter quantification.

[0014] Furthermore, the step (2) includes, based on the analysis of step (1), arranging high-frequency dynamic sensors at the vulnerable positions determined in step (1), collecting vibration, strain and acoustic emission signals in real time, combining wavelet transform and empirical mode decomposition technology to extract transient impact response and micro-damage characteristic frequency under non-stationary working conditions, using deep generative models to construct potential mechanical state space, and mining early damage degradation characteristics, stress wave propagation path distortion and local stiffness attenuation nonlinear dynamic indicators implicit in the noise environment through adversarial training.

[0015] Furthermore, the step (2) includes introducing an equivalent stress criterion based on the stress concentration effect of weak parts in the device structure and using Von Mises equivalent stress to identify high-risk areas;

[0016] Introducing structured sparsity to capture the physical structural characteristics of vulnerable locations:

[0017]

[0018] Among them, x g For the g-th structure group, w g is the weight of each group, G is the total number of groups, and φ is the observation matrix;

[0019] When processing dynamic strain and acoustic emission signals, continuous wavelet transform is used to extract the transient characteristics of the signals:

[0020]

[0021] Among them, ψ is the mother wavelet, a is the scale parameter, and b is the translation parameter.

[0022] Furthermore, the step (3) includes establishing a model capable of capturing global and local features and monitoring structural performance based on the analysis in step (2);

[0023] For the kernel K m The output of position (i, j) in the mth output feature map is:

[0024]

[0025] Among them, K m，c (u, v) is the weight of the mth core to the cth channel, b m is the bias term of the mth kernel.

[0026] Furthermore, step (3) includes introducing nonlinearity so that the network can approximate any function:

[0027] f(x)=max(0,x)

[0028] Perform average pooling:

[0029]

[0030] Where Ω(i, j) is the set of positions covered by the pooling window; after vectorizing the input, a linear transformation is used to add activation:

[0031] y=f(Wx+b)

[0032] Among them, x is the input vector, W is the weight matrix, and f is the activation function; if the network output is The true label is y, then the loss function is:

[0033]

[0034] Furthermore, step (4) includes enhancing the global search capability based on the analysis in step (3), and randomly determining the initial position in the search space using the following formula:

[0035] I i =I min +rand(1,D)e(I max -I min ), i=1,...,Npop

[0036] Among them, the vector of dimension D of uniformly distributed random numbers in the interval [0, 1] is represented by rand(1, D); max and I min are the upper and lower bounds of the search space respectively. The Hadamard product of two vectors is represented by the operation “⊙”;

[0037] Let G v is the growth rate, φ is the growth rate, is the correction coefficient for deviation growth. In the algorithm proposed in this paper, based on data-intensive experiments and simulation processes, the following formula is modeled as member I i The growth rate G vi The difference equation for (t):

[0038] ΔGv i (t+1)=rand 2 e(N(1,D)e ΔGv i (t))

[0039] Among them, the vector ΔG vi (t) and ΔG vi (t+1) represents the growth rate of the discrete time system (time t and time t+1), rand is a random real number in the interval [0, 1] (i.e. rand∈U[0, 1]), rand 2 is a random number of a random variable whose probability density function is equal to N(1, D) represents a random vector of dimension D, which is a random number in a standard Gaussian distribution.

[0040] Furthermore, the step (4) includes describing member I i How to use Member I ii Climb and move logically in the direction of the light source as follows:

[0041]

[0042] where |N(1, D)| is a vector whose components are the absolute values ​​of the components of vector N(1, D), and the operation "" is the Hadamard division of vector u by vector v;

[0043] In Member I i By roaming globally through the search space to the nearest and most important neighbor I ii After the stage, there is a stage where member I i Try to directly follow the best member of the entire population I Best , which is equivalent to member I Best Searching for a better optimal solution around; this stage is expressed in the following mathematical formula:

[0044]

[0045] Subsequently, the current member I new Growth rate ΔG vnew The new value of is calculated by the following formula:

[0046]

[0047] Furthermore, step (5) includes measuring the frequency shift caused by the change in structural stiffness based on the analysis of steps (1) to (4), and the resonance frequency change measurement formula is:

[0048]

[0049] Where f0 is the initial resonance frequency, f d is the currently measured resonance frequency;

[0050] The resonance amplitude change measurement formula is:

[0051]

[0052] Among them, A0 is the resonance peak amplitude of the frequency response function under the structural performance condition, A d is the resonance peak amplitude in the current state; the total energy change in the resonance area is evaluated by frequency domain energy integration, and the formula is:

[0053]

[0054] Among them, H d (f) and H0(f) are the frequency response functions of the performance status and the current state, respectively, and [f1, f2] is the resonance frequency range;

[0055] Construct a normalized weighted comprehensive index, the formula is:

[0056] RHI=ɑ·RFS+β·|RAV|+γ·(1-RER)

[0057] Among them, ɑ, β, and γ are empirical weight coefficients, which can be determined through experimental calibration or machine learning optimization.

[0058] Beneficial effects: Compared with the existing technology, the present invention has the following significant advantages: Through hierarchical processing and deep potential modeling, the present invention effectively overcomes the problems of insufficient local anomaly recognition ability and limited global prediction accuracy of traditional methods, and shows higher anomaly detection sensitivity, stronger robustness and better prediction accuracy in experimental comparison, which is particularly suitable for real-time structural performance status monitoring under complex working conditions; among them, "hierarchical deep dynamic potential" extracts key dynamic features of each level by hierarchically dividing different physical areas or response levels in the structure, and uses deep networks to construct potential feature space, so as to comprehensively characterize the performance status of the structure. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 Flowchart of the present invention. DETAILED DESCRIPTION

[0060] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0061] like Figure 1 As shown, the present invention includes the following steps:

[0062] (1) Determination of the vulnerable position of the device based on discrete mechanics:

[0063] Based on the known device, a 3D model of the structure was created in SolidWorks, accurately describing the shape, size, and relative position of each component and structure. The completed 3D model was imported into ABAQUS for discretization and structural statics simulation. Stress concentration areas were identified based on stress contours and deformation fields, and primary vulnerable location clusters were determined using Bimbaum-Saunders distribution statistics.

[0064] Apply frequency domain loads to candidate clusters through the harmonic response analysis module, solve the structural dynamic strain-frequency response function, and use the peak extraction algorithm to obtain the extreme value of the amplitude spectrum of each node. Define the amplitude sensitivity coefficient With η max The corresponding topological units are used as key vulnerable points. Through static-dynamic joint simulation and discrete parameter quantification, the vulnerable locations of the device are accurately located.

[0065] (2) Acquisition of key mechanical response features of vulnerable locations based on structured compressed sensing:

[0066] Based on the analysis in step (1), high-frequency dynamic sensors are placed at the vulnerable locations determined in step (1) to collect vibration, strain, and acoustic emission signals in real time. Wavelet transform and empirical mode decomposition techniques are combined to extract transient impact responses and micro-damage characteristic frequencies under non-stationary working conditions. A deep generative model is used to construct a potential mechanical state space, and adversarial training is used to mine the early damage degradation characteristics, stress wave propagation path distortion, and local stiffness attenuation nonlinear dynamic indicators hidden in the noise environment.

[0067] Aiming at the stress concentration effect of weak parts in the device structure, the equivalent stress criterion is introduced, and the Von Mises equivalent stress is used to identify high-risk areas.

[0068] To capture the physical structure characteristics of vulnerable locations, we introduce structured sparsity:

[0069]

[0070] Among them, x g is the g-th structural group (corresponding to physical partition, layer or component block), w g is the weight of each group (which can be set based on the vulnerability), G is the total number of groups, and φ is the observation matrix.

[0071] When processing dynamic strain and acoustic emission signals, continuous wavelet transform is used to extract the transient characteristics of the signals:

[0072]

[0073] Among them, ψ is the mother wavelet, a is the scale parameter, and b is the translation parameter.

[0074] (3) Construction of a structural performance monitoring model based on global capture and local integration:

[0075] Based on the analysis in step (2), a model is established that can capture global and local characteristics and monitor the structural performance status.

[0076] For the convolution kernel K m The output of position (i, j) in the mth output feature map is:

[0077]

[0078] Among them, K m，c (u, v) is the weight of the mth core to the cth channel, b m is the bias term of the mth kernel.

[0079] Introducing nonlinearity allows the network to approximate any function:

[0080] f(x)=max(0,x)

[0081] Then perform average pooling:

[0082]

[0083] Among them, Ω(i, j) is the set of positions covered by the pooling window.

[0084] After vectorizing the input, use linear transformation plus activation:

[0085] y=f(Wx+b)

[0086] Among them, x is the input vector, W is the weight matrix, and f is the activation function.

[0087] If the network output is The true label is y, then the loss function is:

[0088]

[0089] (4) Feature enhancement of the extreme points for interactive monitoring of real-time structural performance conditions;

[0090] Based on the analysis in step (3), the global search capability is enhanced to effectively explore the solution space, avoid falling into the local optimal solution, and improve the stability and reliability of structural performance monitoring. The initial position in the search space is randomly determined using the following formula:

[0091] I i =I min +rand(1,D)e(I max -I min ), i=1,...,Npop

[0092] Among them, the vector of dimension D of uniformly distributed random numbers in the interval [0, 1] is represented by rand(1, D). max and I min are the upper and lower bounds of the search space respectively, and the Hadamard product of two vectors (also called element-wise product, represented as “*” in Matlab) is represented by the operation “⊙”. Let G v is the growth rate, φ is the growth rate, is the correction coefficient for deviation growth. In the algorithm proposed in this paper, based on data-intensive experiments and simulation processes, the following formula is modeled as member I i The growth rate G vi (t) is the difference equation.

[0093] ΔGv i (t+1)=rand 2 e(N(1,D)e ΔGv i (t))

[0094] Among them, the vector ΔG vi (t) and ΔG vi (t+1) represents the growth rate of the discrete time system (time t and time t+1), rand is a random real number in the interval [0, 1] (i.e. rand∈U[0, 1]), rand 2 is a random number of a random variable whose probability density function is equal to N(1, D) represents a random vector of dimension D, which is a random number in a standard Gaussian (normal) distribution.

[0095] The following formula describes member I i How to use Member I ii Climb and move logically in the direction of the light source.

[0096]

[0097] Where |N(1, D)| is a vector whose components are the absolute values ​​of the components of vector N(1, D), and the operation "" is the Hadamard division of vector u by vector v (also known as element-wise division, represented as ". / " in Matlab). In member I i By roaming globally through the search space to the nearest and most important neighbor I ii After the stage, there is a stage where member I i Try to directly follow the best member of the entire population I Best , which is equivalent to member I Best Searching for a better optimal solution around. This stage is expressed in the following mathematical formula:

[0098]

[0099] Subsequently, the current member I new Growth rate ΔG vnew The new value of is calculated by the following formula (this is the same as the one used in the initialization step to calculate ΔG vi The formula is exactly the same)

[0100]

[0101] (5) Solution of the structural performance resonance metric of the device:

[0102] Based on the analysis of steps (1) to (4), the frequency shift caused by the change in structural stiffness is measured. The resonance frequency change measurement formula is:

[0103]

[0104] Where f0 is the initial resonance frequency, f d is the currently measured resonant frequency.

[0105] The resonance amplitude change measurement formula is:

[0106]

[0107] Among them, A0 is the resonance peak amplitude of the frequency response function under the structural performance condition, A d is the resonance peak amplitude in the current state. The total energy change in the resonance area is evaluated by frequency domain energy integration, and the formula is:

[0108]

[0109] Among them, H d (f) and H0(f) are the frequency response functions of the performance status and the current state respectively, and [f1, f2] is the resonance frequency band range.

[0110] Construct a normalized weighted comprehensive index, the formula is:

[0111] RHI=ɑ·RFS+β·|RAV|+γ·(1-RER)

[0112] Among them, ɑ, β, and γ are empirical weight coefficients, which can be determined through experimental calibration or machine learning optimization.

Claims

1. A method for monitoring the performance of a device structure based on layered depth dynamic potential, characterized in that: The steps include: (1) Determination of the vulnerable position of the device based on discrete mechanics; (2) Acquisition of key mechanical response features of vulnerable locations based on structured compressed sensing; (3) Construction of a structural performance monitoring model based on global capture and local integration; (4) Feature enhancement at the extreme points of interactive monitoring of real-time structural performance conditions; (5) Solution of the resonance metric of the structural performance of the device.

2. The device structure performance condition monitoring method based on layered depth dynamic potential according to claim 1 is characterized in that: The step (1) includes establishing a three-dimensional model of the structure in the three-dimensional software SolidWorks based on the known device, accurately describing the shape, size and relative position of each component and structure.

3. The device structure performance condition monitoring method based on layered depth dynamic potential according to claim 1 is characterized in that: The step (1) includes importing the established three-dimensional model into ABAQUS software, performing discretization processing, carrying out structural statics simulation, identifying stress concentration areas based on stress cloud maps and deformation fields, and determining primary vulnerable location clusters based on Bimbaum-Saunders distribution statistics; Frequency domain loads are applied to candidate clusters through the harmonic response analysis module to solve the structural dynamic strain-frequency response function, and the peak extraction algorithm is used to obtain the extreme value of the amplitude spectrum of each node; Define the amplitude sensitivity coefficient With η max The corresponding topological units are used as key vulnerable sites, and the precise positioning of the vulnerable positions of the device is achieved through static-dynamic joint simulation and discrete parameter quantification.

4. The device structure performance condition monitoring method based on layered depth dynamic potential according to claim 1 is characterized in that: The step (2) includes, based on the analysis of step (1), arranging high-frequency dynamic sensors at the vulnerable positions determined in step (1), collecting vibration, strain and acoustic emission signals in real time, combining wavelet transform and empirical mode decomposition technology to extract transient impact response and micro-damage characteristic frequency under non-stationary working conditions, using deep generative model to construct potential mechanical state space, and mining early damage degradation characteristics, stress wave propagation path distortion and local stiffness attenuation nonlinear dynamic indicators implicit in the noise environment through adversarial training.

5. The device structure performance condition monitoring method based on layered depth dynamic potential according to claim 1 is characterized in that: The step (2) includes introducing an equivalent stress criterion based on the stress concentration effect of weak parts in the device structure and using Von Mises equivalent stress to identify high-risk areas; Introducing structured sparsity to capture the physical structural characteristics of vulnerable locations: Among them, x g For the g-th structure group, w g is the weight of each group, G is the total number of groups, and φ is the observation matrix; When processing dynamic strain and acoustic emission signals, continuous wavelet transform is used to extract the transient characteristics of the signals: Among them, ψ is the mother wavelet, a is the scale parameter, and b is the translation parameter.

6. The device structure performance condition monitoring method based on layered depth dynamic potential according to claim 1 is characterized in that: The step (3) includes establishing a model capable of capturing global and local features and monitoring structural performance based on the analysis in step (2); For the kernel K m The output of position (i, j) in the mth output feature map is: Among them, K m，c (u, v) is the weight of the mth core to the cth channel, b m is the bias term of the mth kernel.

7. The device structure performance condition monitoring method based on layered depth dynamic potential according to claim 1 is characterized in that: The step (3) includes introducing nonlinearity so that the network can approximate any function: f(x)=max(0,x) Perform average pooling: Where Ω(i, j) is the set of positions covered by the pooling window; after vectorizing the input, a linear transformation is used to add activation: y=f(Wx+b) Among them, x is the input vector, W is the weight matrix, and f is the activation function; if the network output is The true label is y, then the loss function is:

8. The device structure performance condition monitoring method based on layered depth dynamic potential according to claim 1 is characterized in that: The step (4) includes strengthening the global search capability based on the analysis of step (3), and randomly determining the initial position in the search space using the following formula: I i =I min +rand(1,D)e(I max -I min ),i=1,...,Npop Among them, the vector of dimension D of uniformly distributed random numbers in the interval [0, 1] is represented by rand(1, D); max and I min are the upper and lower bounds of the search space respectively. The Hadamard product of two vectors is represented by the operation "⊙"; Let G v is the growth rate, φ is the growth rate, is the correction coefficient for deviation growth. In the algorithm proposed in this paper, based on data-intensive experiments and simulation processes, the following formula is modeled as member I i The growth rate G vi The difference equation for (t): ΔGv i (t+1)=rand 2 e(N(1,D)eΔGv i (t)) Among them, the vector ΔG vi (t) and ΔG vi (t+1) represents the growth rate of the discrete time system (time t and time t+1), rand is a random real number in the interval [0, 1] (i.e. rand∈U[0, 1]), rand 2 is a random number of a random variable whose probability density function is equal to N(1, D) represents a random vector of dimension D, which is a random number in a standard Gaussian distribution.

9. The device structure performance condition monitoring method based on layered depth dynamic potential according to claim 1 is characterized in that: The step (4) includes describing member I i How to use Member I ii Climb and move logically in the direction of the light source as follows: where |N(1, D)| is a vector whose components are the absolute values ​​of the components of vector N(1, D), and the operation "" is the Hadamard division of vector u by vector v; In Member I i By roaming globally through the search space to the nearest and most important neighbor I ii After the stage, there is a stage where member I i Try to directly follow the best member of the entire population I Best , which is equivalent to member I Best Searching for a better optimal solution around; this stage is expressed in the following mathematical formula: Subsequently, the current member I new Growth rate ΔG vnew The new value of is calculated by the following formula:

10. The device structure performance condition monitoring method based on layered depth dynamic potential according to claim 1 is characterized in that: The step (5) includes measuring the frequency shift caused by the change in structural stiffness based on the analysis of steps (1) to (4), and the resonance frequency change measurement formula is: Where f0 is the initial resonance frequency, f d is the currently measured resonance frequency; The resonance amplitude change measurement formula is: Among them, A0 is the resonance peak amplitude of the frequency response function under the structural performance condition, A d is the resonance peak amplitude in the current state; the total energy change in the resonance area is evaluated by frequency domain energy integration, and the formula is: Among them, H d (f) and H0(f) are the frequency response functions of the performance status and the current state, respectively, and [f1, f2] is the resonance frequency range; Construct a normalized weighted comprehensive index, the formula is: RHI=ɑ·RFS+β·|RAV|+γ·(1-RER) Among them, ɑ, β, and γ are empirical weight coefficients, which can be determined through experimental calibration or machine learning optimization.

Images