Vibration signal processing method based on adaptive frequency domain partition and morphological denoising

Abstract

This invention discloses a vibration signal processing method based on adaptive frequency domain partitioning and morphological denoising. The method involves performing a short-time Fourier transform on the original one-dimensional vibration signal to obtain an initial time-frequency energy map; dividing the initial time-frequency energy map into various frequency domain sub-regions based on a Bayesian optimization model and determining the corresponding energy thresholds; using a quantum particle swarm optimization model to adaptively optimize and denoise the morphological parameters of each frequency domain sub-region to determine a high-confidence energy point mask; constructing a multi-branch decoupled network model based on a physical information neural network to separate and complete the time-frequency trajectory energy points on the high-confidence energy point mask, obtaining a set of time-frequency trajectory energy points; performing multi-dimensional quantitative evaluation on the set of time-frequency trajectory energy points and calculating a comprehensive evaluation score function; and implementing cross-cycle asynchronous parameter closed-loop feedback adjustment based on the comprehensive evaluation score until the comprehensive evaluation score meets a preset engineering standard, outputting the target set of time-frequency trajectory energy points.

Description

Technical Field

[0001] This invention relates to the field of signal processing technology, and more particularly to a vibration signal processing method based on adaptive frequency domain partitioning and morphological denoising. Background Technology

[0002] In the operation of large machinery, time-frequency analysis is a commonly used signal processing technique for vibration signals generated under non-stationary conditions such as variable speed and load. Existing time-frequency analysis methods typically obtain the time-frequency distribution matrix (i.e., time-frequency graph) of the signal through algorithms such as short-time Fourier transform or wavelet transform. Based on this, they then combine global threshold segmentation, maximum value search, or simple morphological filtering to extract time-frequency ridges or time-frequency trajectory energy points to track the characteristics of non-stationary signals. However, in actual large machinery collaborative operation scenarios, equipment is often in an environment with strong electromagnetic interference and complex mechanical noise. Simultaneously, there is multi-source dynamic coupling between multiple components such as engines, pumps, and valves, and the frequencies of different excitation sources may overlap locally. Under such complex conditions, traditional extraction algorithms that rely on a globally unified threshold and manual experience often struggle to simultaneously account for the differences between low-frequency strong energy components and high-frequency weak energy components in the time-frequency graph, easily leading to high-frequency background noise leakage or the breakage or loss of effective low-frequency energy points. In addition, existing methods mostly use simple mathematical and geometric interpolation when dealing with trajectory discontinuities, lacking constraints on the inherent dynamic physical laws of mechanical systems; the energy points of the reconstructed time-frequency trajectory obtained in this way are often physically unreasonable, further affecting the decoupling effect and parameter estimation performance of complex multi-source coupled signals.

[0003] The time-frequency energy distribution corresponding to the original vibration signal typically exhibits characteristics such as multi-source coupling, local crossover, strong noise coverage, and significant differences in the strength of frequency components. Therefore, using a single unified threshold, fixed parameters, or conventional geometric interpolation methods makes it difficult to simultaneously consider the preservation of effective energy points, noise suppression, local fracture repair, and the physical rationality constraints of the trajectory. Existing threshold setting methods based on human experience often rely on trial and error, resulting in poor parameter transferability. When low-frequency strong energy regions coexist with high-frequency weak energy regions, problems such as over-preservation of low frequencies and under-detection of high frequencies, or over-preservation of high frequencies and the incorporation of low-frequency noise, are prone to occur. Existing parameter optimization methods based on traditional particle swarm optimization, genetic algorithms, or enumeration search often suffer from insufficient search efficiency, premature convergence, or unstable parameter fit under conditions of multi-parameter coupling and significant regional differences. Existing interpolation, smoothing, or ordinary neural network fitting methods for fracture trajectories rely more on data-driven approaches or geometric continuity assumptions, making it difficult to ensure that the completion results satisfy the dynamic constraints and physical evolution laws of the mechanical system under conditions of strong interference and multi-source aliasing. Summary of the Invention

[0004] Purpose of the invention: In order to overcome the shortcomings of the existing technology, the present invention provides a vibration signal processing method based on adaptive frequency domain partitioning and morphological denoising. By performing adaptive frequency domain partitioning and morphological denoising on the initial time-frequency energy map of the vibration signal, the ability to extract time-frequency trajectory energy points is improved.

[0005] Technical Solution: To achieve the above objectives, the vibration signal processing method based on adaptive frequency domain partitioning and morphological denoising of the present invention includes the following steps:

[0006] Step 1: Obtain the original one-dimensional vibration signal of the large mechanical system, and perform a short-time Fourier transform on the original one-dimensional vibration signal to obtain the initial time-frequency energy map.

[0007] Step 2: Based on the Bayesian optimization model, perform adaptive frequency domain partitioning on the initial time-frequency energy map to obtain each frequency domain sub-region, and collaboratively determine the energy threshold corresponding to each frequency domain sub-region.

[0008] Step 3: Based on the energy threshold corresponding to each frequency domain sub-region, convert the initial time-frequency energy map into a binary mask. Use the quantum particle swarm optimization model to perform adaptive optimization and noise reduction processing on the morphological parameters of each frequency domain sub-region to determine the global high-confidence energy point mask.

[0009] Step 4: Construct a multi-branch decoupled network model based on physical information neural network to perform time-frequency trajectory energy point separation and completion processing on the global high confidence energy point mask, and finally obtain a continuous and decoupled time-frequency trajectory energy point set.

[0010] Step 5: Perform multidimensional quantitative evaluation on the time-frequency trajectory energy point set to obtain a multidimensional quantitative score, and use weighted calculation based on the multidimensional quantitative score to obtain a comprehensive evaluation score function.

[0011] Step 6: Implement cross-cycle asynchronous parameter closed-loop feedback adjustment based on the comprehensive evaluation score until the comprehensive evaluation score meets the preset engineering standard, and output the target time-frequency trajectory energy point set.

[0012] Furthermore, in step 2, the initial time-frequency energy map is adaptively partitioned into frequency domains based on a Bayesian optimization model to obtain each frequency domain sub-region, and the energy threshold corresponding to each frequency domain sub-region is determined collaboratively; this includes the following steps:

[0013] Step 1-1: Convert the initial time-frequency energy map from the decibel domain to the linear energy domain, and continuously integrate the linear energy domain along the time dimension to obtain the overall energy distribution curve in the frequency direction.

[0014] Steps 1-2: Generate a set of candidate segmentation locations based on the specific characteristics of the overall energy distribution curve, and construct a set of parameter vectors to be optimized; the set of parameter vectors to be optimized includes the specific location of the segmentation line and the energy threshold corresponding to each frequency domain sub-region.

[0015] Steps 1-3: Based on the set of parameter vectors to be optimized, construct a comprehensive objective function that combines the effective energy retention rate, noise leakage penalty term, and initial trajectory coherence term. Use Gaussian process regression to establish a probabilistic surrogate model for the comprehensive objective function.

[0016] Steps 1-4: Using the expected improvement index as the acquisition function, perform iterative search operations to determine the optimal segmentation line position and the optimal energy threshold of each frequency domain sub-region; based on the optimal segmentation line position, divide the initial time-frequency energy map into each frequency domain sub-region.

[0017] Furthermore, in steps 1-3, a comprehensive objective function is constructed based on the set of parameter vectors to be optimized, which is a combination of the effective energy retention rate, the noise leakage penalty term, and the initial trajectory coherence term; the calculation process is shown below:

[0018] ;

[0019] In the formula, For effective energy retention rate, As a penalty item for noise leakage, For the initial trajectory coherence term, , , These are the optimization weight coefficients for the effective energy retention rate, the noise leakage penalty term, and the initial trajectory coherence term, respectively. A probabilistic surrogate model for the comprehensive objective function is established using Gaussian process regression, as shown below:

[0020] ;

[0021] In the formula, It is a mean function. It is a covariance kernel function that measures the similarity and uncertainty among combinations of parameters.

[0022] Furthermore, in step 3, the initial time-frequency energy map is converted into a binary mask based on the energy threshold corresponding to each frequency domain sub-region. A quantum particle swarm optimization model is then used to adaptively optimize and denoise the morphological parameters of each frequency domain sub-region to determine the global high-confidence energy point mask. This includes the following steps:

[0023] Step 2-1: For each divided frequency domain sub-region, the morphological denoising parameters are defined as the particle position vectors inside the quantum particle swarm optimization model. The morphological denoising parameters include the area screening threshold and the length, width, and orientation tilt of the structuring element.

[0024] Step 2-2: Use the ratio of the degree of retention of effective energy points after denoising to the degree of noise suppression as the fitness function.

[0025] Steps 2-3: By calculating the average optimal position of individuals and local attractor points of the entire population, the particle positions are continuously updated and iterated to determine the optimal morphological denoising parameters.

[0026] Steps 2-4: Use the obtained optimal area screening threshold to remove isolated noise patches in the local area, and use the obtained optimal structuring element to perform morphological closing operation to repair small gaps in the trajectory, and finally generate a global high-confidence energy point mask.

[0027] Furthermore, in step 4, a multi-branch decoupled network model based on a physical information neural network is constructed to perform time-frequency trajectory energy point separation and completion processing on the global high-confidence energy point mask, ultimately obtaining a continuous and decoupled set of time-frequency trajectory energy points; including the following steps:

[0028] Step 3-1: Establish a set of generalized multi-source coupled dynamic differential equations with unknown physical parameters.

[0029] Step 3-2: Construct a multi-branch decoupling network model based on physical information neural network, using the received time variable as input, and each branch of the decoupling network model directly predicts the time-domain latent variable prediction result of a single vibration source.

[0030] Step 3-3: Use a differentiable short-time Fourier transform layer to map the time-domain latent variable prediction results of each decoupled network model branch to the time-frequency domain, obtain the prediction time-frequency response matrix of each decoupled network model branch, and fuse the prediction time-frequency responses of each decoupled network model branch to output the overall prediction result.

[0031] Steps 3-4: Differentiate the temporal latent variable prediction results output by each branch of the decoupled network model and substitute them into the generalized multi-source coupled dynamic differential equation system to construct the dynamic residual as the physical residual loss. Construct the data reconstruction loss based on the overall prediction results and construct the multi-source decoupling constraint loss based on the temporal latent variable prediction results corresponding to each branch of the decoupled network model. Combine the physical residual loss, data reconstruction loss and multi-source decoupling constraint loss to construct the joint loss function.

[0032] Steps 3-5: Adaptive weight training is performed on the multi-branch decoupled network model based on physical information neural network using the joint loss function. This forces the branches of the decoupled network model to perform inverse physical smoothing in the trajectory breakage area, thereby determining the set of continuous and decoupled time-frequency trajectory energy points.

[0033] Furthermore, in step 3-1, a set of generalized multi-source coupled dynamic differential equations with unknown physical parameters is established; the large mechanical system is assumed to consist of M coupled vibration sources, and the calculation process is as follows:

[0034] ;

[0035] In the formula, Let be the latent variable of the vibration response of the i-th independent vibration source in the time domain. , , Let be the equivalent mass parameter, damping parameter, and stiffness parameter of the i-th vibration source, respectively. It is represented as the coupling transmission coefficient between different vibration sources.

[0036] Furthermore, in step 5, a multidimensional quantitative evaluation is performed on the time-frequency trajectory energy point set to obtain a multidimensional quantitative score. Based on the multidimensional quantitative score, a weighted calculation is used to obtain a comprehensive evaluation score function; this includes the following steps:

[0037] Step 4-1: Based on the obtained time-frequency trajectory energy point set, calculate the effective energy point coverage index and out-of-band noise leakage index respectively, and at the same time calculate the time continuity index for measuring trajectory gaps.

[0038] Step 4-2: Perform ridge extraction on the time axis for the set of energy points of the time-frequency trajectory to obtain the center frequency evolution curve of the trajectory, and introduce physical evolution mapping to calculate physical rationality index.

[0039] Step 4-3: Weight the effective energy point coverage index, out-of-band noise leakage index, time continuity index, and physical rationality index to construct a comprehensive evaluation score function.

[0040] Furthermore, in step 6, asynchronous parameter closed-loop feedback adjustment across cycles is implemented based on the comprehensive evaluation score until the comprehensive evaluation score meets the preset engineering standard, and the target time-frequency trajectory energy point set is output; the comprehensive evaluation score of the current analysis cycle is calculated based on the comprehensive evaluation score function, and the comprehensive evaluation score of the current analysis cycle is compared with the preset engineering standard threshold; if the comprehensive evaluation score of the current analysis cycle is lower than the preset engineering standard threshold, the comprehensive evaluation score of the current analysis cycle is used as an environmental reward signal and transmitted back to the Bayesian optimization model and the quantum particle swarm optimization model. By adjusting the kernel function hyperparameter of the Bayesian model and expanding the search boundary of the quantum particle swarm algorithm in the next analysis cycle, the macroscopic parameter correction process is triggered; until the comprehensive evaluation score of any analysis cycle is higher than or equal to the preset engineering standard threshold.

[0041] Beneficial Effects: The vibration signal processing method based on adaptive frequency domain partitioning and morphological denoising of this invention performs adaptive frequency domain partitioning on the initial time-frequency energy map based on a Bayesian optimization model to obtain various frequency domain sub-regions, and collaboratively determines the energy threshold corresponding to each frequency domain sub-region, thereby achieving stable extraction of effective time-frequency energy points in complex mechanical signals; it utilizes its small sample size, high efficiency, and global optimization capabilities to complete adaptive partitioning and threshold search. A quantum particle swarm optimization model is used to perform adaptive optimization denoising of morphological parameters on each frequency domain sub-region, which can automatically determine the area screening threshold and morphological structural element parameters according to the time-frequency energy point distribution characteristics of different frequency domain sub-regions, thereby achieving effective suppression of noise points and enhanced preservation of continuous trajectory structures, and realizing adaptive optimization of denoising parameters for different frequency domain sub-regions. A multi-branch decoupling network model based on a physical information neural network is constructed to perform time-frequency trajectory energy point separation and completion processing on a global high-confidence energy point mask. This model addresses the issues of energy point aliasing, breakage, and missing points under conditions of multi-source coupling and strong interference in large machinery. Under multi-source dynamic physical constraints, it achieves high-fidelity continuous reconstruction of missing energy points and aliasing decoupling. Furthermore, dynamic physical constraints are applied beyond data consistency constraints to achieve high-fidelity completion and decoupled reconstruction of time-frequency trajectory energy points. A multi-dimensional quantitative evaluation of the time-frequency trajectory energy point set yields a multi-dimensional quantitative score. Based on this score, a weighted calculation is used to obtain a comprehensive evaluation score function, which objectively evaluates the extraction results from four aspects: effective point retention, noise leakage, temporal continuity, and physical rationality. The evaluation results are then used inversely in the aforementioned intelligent parameter optimization process. Attached Figure Description

[0042] Figure 1 This is a flowchart of a vibration signal processing method based on adaptive frequency domain partitioning and morphological denoising.

[0043] Figure 2 To display the original time-frequency energy map of complex background noise.

[0044] Figure 3 Comparison chart showing signal missed detection or false detection due to background noise caused by improper global fixed threshold setting.

[0045] Figure 4 This is a time-frequency energy point map extracted based on frequency-domain adaptive threshold segmentation.

[0046] Figure 5 This is a visualization of the time-frequency energy point trajectory set after processing by a multi-branch decoupled network model based on a physical information neural network. Detailed Implementation

[0047] The invention will now be further described with reference to the accompanying drawings.

[0048] like Figure 1 As shown, the vibration signal processing method based on adaptive frequency domain partitioning and morphological denoising includes the following steps:

[0049] Step 1: Obtain the original one-dimensional vibration signal of the large mechanical system, and perform a short-time Fourier transform on the original one-dimensional vibration signal to obtain the initial time-frequency energy map.

[0050] Step 2: Based on the Bayesian optimization model, perform adaptive frequency domain partitioning on the initial time-frequency energy map to obtain each frequency domain sub-region, and collaboratively determine the energy threshold corresponding to each frequency domain sub-region.

[0051] Step 3: Based on the energy threshold corresponding to each frequency domain sub-region, convert the initial time-frequency energy map into a binary mask. Use the quantum particle swarm optimization model to perform adaptive optimization and noise reduction processing on the morphological parameters of each frequency domain sub-region to determine the global high-confidence energy point mask.

[0052] Step 4: Construct a multi-branch decoupled network model based on physical information neural network to perform time-frequency trajectory energy point separation and completion processing on the global high confidence energy point mask, and finally obtain a continuous and decoupled time-frequency trajectory energy point set.

[0053] Step 5: Perform multidimensional quantitative evaluation on the time-frequency trajectory energy point set to obtain a multidimensional quantitative score, and use weighted calculation based on the multidimensional quantitative score to obtain a comprehensive evaluation score function.

[0054] Step 6: Implement asynchronous parameter closed-loop feedback adjustment across cycles based on the comprehensive evaluation score until the comprehensive evaluation score meets the preset engineering standards, output the target time-frequency trajectory energy point set, and visualize the final target time-frequency trajectory energy point set.

[0055] Acquiring the original one-dimensional vibration signal of a large mechanical system For the original one-dimensional vibration signal Preprocessing, including DC removal, trend removal, and normalization, is performed to reduce the impact of the bias term and the slowly varying trend on the subsequent time-frequency energy distribution; the calculations are shown below:

[0056] ;

[0057] In the formula, The mean of the original one-dimensional vibration signal. This represents the standard deviation of the original one-dimensional vibration signal.

[0058] The preprocessed original one-dimensional vibration signal Short-time Fourier transform (SFT) is used for time-frequency transformation to obtain the initial time-frequency energy map. The SFT converts the non-stationary signal into a series of locally stationary short-time signals through time-domain windowing, and performs a Fourier transform on each short-time signal. The calculation process is shown below:

[0059] ;

[0060] In the formula, t is the center time of the short-time analysis window; f is the frequency variable; τ is the integration variable; and j is the imaginary unit. For window functions, This is the preprocessed original one-dimensional vibration signal at the integral variable τ.

[0061] like Figure 2-3 As shown, Figure 2 It can be clearly observed that the time-frequency energy map has significant background noise. Under such complex background noise, the traditional global fixed threshold setting method is difficult to achieve the ideal extraction effect. Figure 3 The image shows a comparison of signal missed detection or background noise false detection caused by improper global fixed threshold settings. Figure 3 (a) shows that many valid signal components are mistakenly deleted due to an excessively high threshold, resulting in serious missed detections; Figure 3 (b) shows that due to the excessively small threshold, a large amount of background noise is retained, making it difficult to distinguish the effective energy point trajectory. Under the complex working conditions of large machinery, the energy span between strong vibration and weak fault signals in the original one-dimensional vibration signal is extremely large, exhibiting significant wide dynamic range characteristics. To enhance the contrast of effective feature energy points in the time-frequency diagram, the linear time-frequency energy spectrum is first converted to the decibel (dB) domain. Utilizing the nonlinear compression characteristics of the logarithmic scale, not only can the amplitude of the strong interference background energy be effectively converged, but the weak high-frequency feature energy induced by small local defects such as cracks can also be significantly amplified and highlighted. Therefore, the initial time-frequency energy diagram obtained is represented by a decibel domain time-frequency energy diagram, expressed by the following formula:

[0062] ;

[0063] In the formula, ε is a minimal positive constant to prevent zero values ​​from occurring in logarithmic calculations. Unless otherwise specified, the energy representation of the time-frequency graph is assumed to be in the decibel domain.

[0064] In step 2, the initial time-frequency energy map is adaptively partitioned into frequency domains based on a Bayesian optimization model to obtain each frequency domain sub-region, and the energy threshold corresponding to each frequency domain sub-region is determined collaboratively; this includes the following steps:

[0065] Step 1-1: Transform the initial time-frequency energy map from the decibel domain to the linear energy domain, as shown in the calculation below:

[0066] ;

[0067] And linear energy domain By performing continuous integration along the time dimension t, the overall energy distribution curve in the frequency direction is obtained; the calculation is shown below:

[0068] ;

[0069] In the formula, This is the overall energy distribution curve along the frequency direction.

[0070] Steps 1-2: Generate a set of candidate segmentation locations based on the specific characteristics of the overall energy distribution curve, and construct a set of parameter vectors to be optimized. The set of parameter vectors to be optimized includes the specific location of the segmentation line and the energy threshold corresponding to each frequency domain sub-region. Based on the monotonicity, local extrema, inflection points, and gradient changes of the overall energy distribution curve in the frequency direction, generate a set of candidate segmentation locations. A single segmentation line and dual partitioning method can be used to construct the set of candidate segmentation locations into a set of parameter vectors to be optimized. The set of parameter vectors to be optimized is defined as follows:

[0071]

[0072] In the formula, θ is the set of parameter vectors to be optimized. This is the location of the dividing line between the low-frequency and high-frequency regions. The energy threshold in the low-frequency region, This represents the energy threshold in the high-frequency region.

[0073] For cases where the frequency domain energy distribution has multiple significant inflection points, the set of parameter vectors to be optimized can be expanded into a multi-segment, multi-partition form, as shown in the following formula:

[0074]

[0075] In the formula, R represents the number of dividing lines. The position of the R-th dividing line. R+1 represents the energy threshold corresponding to the (R+1)th frequency domain sub-region, where R+1 is the number of frequency domain sub-regions.

[0076] Steps 1-3: Based on the set of parameter vectors to be optimized, construct a comprehensive objective function composed of the effective energy retention rate, noise leakage penalty term, and initial trajectory coherence term. Use Gaussian process regression to establish a probabilistic surrogate model for the comprehensive objective function. For the optimized parameter vector set θ, define the comprehensive objective function of the Bayesian optimization model. Since the Bayesian optimization model requires frequent iterations, in order to balance computational efficiency and quality search, a comprehensive objective function is adopted that combines the effective energy retention rate, noise leakage penalty term, and initial trajectory coherence term.

[0077] In steps 1-3, a comprehensive objective function is constructed based on the set of parameter vectors to be optimized, which is a combination of the effective energy retention rate, the noise leakage penalty term, and the initial trajectory coherence term; the calculation process is shown below:

[0078] ;

[0079] In the formula, The effective energy retention rate is used to characterize the degree of effective energy retention. It can be calculated by statistically comparing the total amplitude of the energy points retained under the current parameter θ cutoff with the total amplitude before cutoff. This is a noise leakage penalty term used to characterize the degree of noise leakage, which can be quantified by statistically analyzing the proportion of residual energy points in the unexpected frequency band. The initial trajectory coherence term is used to characterize the coherence of the initial energy points in the time direction, and can be evaluated by calculating the length of the continuous time window between adjacent time frames for energy points. , , The optimization weight coefficients for the effective energy retention rate, noise leakage penalty term, and initial trajectory coherence term are all custom-defined; the optimization is achieved by maximizing this comprehensive objective function. This guides the Bayesian optimization model towards a parameter space with high signal-to-noise ratio and high continuity. A probabilistic surrogate model is established using Gaussian process regression to represent the black-box comprehensive objective function, as shown below:

[0080] ;

[0081] In the formula, It is a mean function, usually set to 0; A covariance kernel function, such as the Matern kernel, is used to measure the similarity and uncertainty among parameter combinations. The covariance kernel function employs the Matern kernel function with automatic correlation scaling. Its hyperparameters include length scale, signal variance, and noise variance. The comprehensive evaluation score of the current analysis period is used as an environmental reward signal and backpropagated to the Bayesian optimization model to adjust the kernel function hyperparameters of the Bayesian model, performing cross-period adaptive updates based on the environmental reward signal.

[0082] Steps 1-4: Using the expected improvement index (ACPEI) as the acquisition function, perform iterative search operations to determine the optimal segmentation line position and the optimal energy threshold for each frequency domain sub-region; based on the optimal segmentation line position, divide the initial time-frequency energy map into each frequency domain sub-region; in each optimization iteration, use the expected improvement index (ACPEI) as the acquisition function to determine the optimal parameter position for the next exploration, as shown in the calculation process below:

[0083] ;

[0084] In the formula, Let E be the best overall objective function observed so far, and E be the expected value. By maximizing this acquisition function, the Bayesian optimization model can lock the optimal dividing line and the energy threshold of each frequency domain sub-region with very few calculations. Based on the position of the optimal dividing line, the initial time-frequency energy map is truncated to obtain each frequency domain sub-region, which is used to extract the initial set of effective time-frequency trajectory energy points.

[0085] In step 3, the initial time-frequency energy map is converted into a binary mask based on the energy threshold corresponding to each frequency domain sub-region. A quantum particle swarm optimization model is then used to adaptively optimize and denoise each frequency domain sub-region using morphological parameters to determine the global high-confidence energy point mask. This includes the following steps:

[0086] Each frequency domain sub-region is defined as Ω. r For any time-frequency point (t, f), if its frequency f falls into the r-th frequency domain sub-region and satisfies the energy threshold condition of that frequency domain sub-region, then the point is retained; otherwise, the point is discarded; the corresponding initial binary mask Defined as:

[0087] ;

[0088] In the formula, "other" represents other time-frequency points. Let be the energy threshold for the r-th frequency domain sub-region. For the r-th frequency domain sub-region, a binary mask for that sub-region can be further defined. The calculation is as follows:

[0089] .

[0090] This transforms the entire initial time-frequency energy map into a set of binary masks composed of several frequency domain sub-regions. This binary mask set not only preserves the significant energy points in the initial time-frequency energy map, but also provides a clear input for subsequent partitioned morphological denoising.

[0091] Step 2-1: For each predefined frequency domain sub-region, the morphological denoising parameters are defined as the particle position vector within the quantum particle swarm optimization model. These morphological denoising parameters include an area screening threshold and the length, width, and tilt angle of the structuring element. For the r-th frequency domain sub-region, the morphological denoising parameters are defined as the particle position vector of the quantum particle swarm optimization algorithm, specifically including the area screening threshold A. min and the length L of the structuring element x Width L y and direction tilt angle The expression is as follows:

[0092] ;

[0093] In the formula, Here, T represents the morphological denoising parameter, indicating a transpose operation used to construct a column vector of particle position vectors from the area filtering threshold, structuring element length, width, and orientation tilt angle. The area filtering threshold A... min and the length L of the structuring element x Width L y Both the direction and tilt angle ϕ are set as initial search boundaries. The comprehensive evaluation score of the current analysis period is used as an environmental reward signal and then transmitted back to the quantum particle swarm optimization model to expand the search boundary of the quantum particle swarm algorithm in the next analysis period. Cross-period adaptive expansion is performed based on the environmental reward signal.

[0094] Step 2-2: Use the ratio of the degree of preservation of effective energy points after denoising to the degree of noise suppression as the fitness function. The calculation is as follows:

[0095] ;

[0096] In the formula, The number of effective energy points retained after morphological processing of the particle is the estimated number of points. The morphological processing process is the subsequent area screening and closing operation. The number of noise points to be filtered out, ε is the minimum normal number. The fitness function... In the quantum particle swarm optimization model, all particles are driven to converge toward the optimal parameter region that preserves the effective signal while suppressing noise.

[0097] Steps 2-3: By calculating the average optimal position of each individual in the entire population and the local attractor points in the quantum particle swarm optimization model, the particle positions are continuously updated and iterated to determine the optimal morphological denoising parameters. The quantum particle swarm optimization model achieves global search through position updates, and in the k-th iteration, the average optimal position of each individual in the entire population is calculated. The calculation is as follows:

[0098] ;

[0099] In the formula, For population size, This represents the historical optimal position of the individual particle. Next, stochastic simulations are used to determine the current local attractor point. The calculation is as follows:

[0100] ;

[0101] In the formula, The optimal position for the entire group. These are uniformly distributed random numbers.

[0102] Therefore, the particle position update in the (k+1)th iteration uses an analytical formula, as shown below:

[0103] ;

[0104] In the formula, β is the contraction-expansion coefficient that controls the convergence rate. These are uniformly distributed random numbers.

[0105] Steps 2-4: Use the obtained optimal area screening threshold to remove isolated noise patches in the local area, and use the obtained optimal structuring element to perform morphological closing operation to repair small gaps in the trajectory, and finally generate a global high-confidence energy point mask.

[0106] For each frequency domain sub-region, the binary sub-mask corresponding to the r-th frequency domain sub-region By performing connected component labeling, we obtain the set of all connected components within this frequency domain sub-region. ;in Let represent the total number of connected components within the r-th frequency domain sub-region; the area of ​​the q-th connected component is defined as follows:

[0107] .

[0108] In the formula, For connected components The number of pixels contained; then filtered according to the area threshold A. minThe connected components are evaluated, and isolated noise patches with excessively small areas are removed; the corresponding connected component area filtering operator is used. The definition is as follows:

[0109] ;

[0110] ;

[0111] Therefore, the result of the r-th frequency domain sub-region after one processing area filtering is denoted as: The calculation is as follows:

[0112] .

[0113] After performing connected domain area filtering on each frequency domain sub-region using the aforementioned connected domain area filtering operator, further steps are taken to repair local discontinuities and enhance the structural connectivity between adjacent effective energy points. Perform morphological closing operation;

[0114] The morphological closing operation consists of two parts: dilation followed by erosion. The expression for the two-stage processing is as follows:

[0115] ;

[0116] In the formula, This is represented as a morphological closing operation. This is represented as an expansion operation. This is represented as an erosion operation. This is represented as the structuring element corresponding to the r-th frequency domain sub-region. The closing operation is performed in the order of expansion followed by erosion, used to connect adjacent effective energy points and repair local fractures.

[0117] The structural element The length L of the optimal parameter structuring element output by the quantum particle swarm optimization algorithm x Width L y and direction tilt angle The structuring element can be a rotated rectangle, defined as follows:

[0118]

[0119] In the formula, For local coordinates relative to the center of the structural element, an directional tilt angle is introduced. The structural elements can perform closed operations along the main extension direction of the time-frequency energy points, thereby improving the targeting of local fracture repair.

[0120] like Figure 4 As shown, this is a time-frequency energy point map extracted based on frequency-domain adaptive threshold segmentation, and... Figure 3The comparison can yield the following results Figure 4 It retains most of the effective signal components while suppressing a large amount of out-of-band and in-band noise.

[0121] In the above closing operation, the dilation operation expands the energy point region and crosses local small gaps, connecting previously close but unconnected effective energy points; the erosion operation, while maintaining the repaired connectivity, shrinks the over-expanded boundary after dilation, thereby restoring the true thickness of the trajectory structure. Through a combination of pre-screening and post-closing processing, optimal morphological denoising parameters are output for each frequency domain sub-region, ultimately obtaining high-confidence energy point masks for each cleaned frequency domain sub-region; after processing all frequency domain sub-regions, a global high-confidence energy point mask is obtained, calculated as follows:

[0122] .

[0123] Therefore, the connected region area filtering operator can be seen to be effective. The morphological closing operation is responsible for removing isolated noise patches that are too small, while the morphological closing operation is responsible for repairing local breaks and enhancing structural connectivity. The two respectively perform the different functions of "denoising" and "connectivity enhancement".

[0124] In step 4, a multi-branch decoupled network model based on a physical information neural network is constructed to perform time-frequency trajectory energy point separation and completion processing on the global high-confidence energy point mask, ultimately obtaining a continuous and decoupled set of time-frequency trajectory energy points; including the following steps:

[0125] Step 3-1: Establish a set of generalized multi-source coupled dynamic differential equations with unknown physical parameters. These equations describe the vibration evolution of multiple coupled vibration sources in a mechanical system over time. In Step 3-1, a set of generalized multi-source coupled dynamic differential equations with unknown physical parameters is established. The large mechanical system is assumed to consist of M coupled vibration sources. The calculation process is as follows:

[0126] ;

[0127] In the formula, Let be the latent variable of the vibration response of the i-th independent vibration source in the time domain. , , Let be the equivalent mass parameter, damping parameter, and stiffness parameter of the i-th vibration source, respectively. The above parameters are represented as the coupling transmission coefficients between different vibration sources; all of these parameters are used as learnable physical parameters during network training.

[0128] Step 3-2: Construct a multi-branch decoupling network model based on a physical information neural network. The received time variable is used as input, and each branch of the decoupling network model directly predicts the time-domain latent variable prediction result of a single vibration source. Using the time variable t as input, M parallel decoupling network model branches output the time-domain latent variable prediction results of each physical vibration source. Each branch of the decoupled network model corresponds to an independent vibration source. Each branch can be constructed using a multi-layer fully connected feedforward neural network. The multi-layer fully connected feedforward neural network includes an input layer, several hidden layers, and an output layer. The hidden layers use the tanh activation function, the sinusoidal activation function, or other nonlinear activation functions that can characterize continuous vibration response. The output layer outputs the time-domain latent variable prediction results corresponding to the vibration source. Through the aforementioned parallel branching structure, different branches of the decoupled network model learn the temporal evolution characteristics of different vibration sources.

[0129] Step 3-3: Utilize a differentiable short-time Fourier transform layer to map the time-domain latent variable prediction results of each decoupled network model branch to the time-frequency domain, obtaining the predicted time-frequency response matrix for each decoupled network model branch. Then, fuse the predicted time-frequency responses of each decoupled network model branch to output the overall prediction result. Introducing a differentiable short-time Fourier transform layer transforms the time-domain latent variable prediction results... Mapping to the time-frequency domain, we obtain the predicted time-frequency response matrix for each branch. The time-frequency responses of each decoupled network model branch are fused to construct the overall prediction result. The calculation process is as follows:

[0130] .

[0131] Steps 3-4: Differentiate the temporal latent variable prediction results output by each branch of the decoupled network model and substitute them into the generalized multi-source coupled dynamic differential equation system to construct the dynamic residual as the physical residual loss. Construct the data reconstruction loss based on the overall prediction results and construct the multi-source decoupling constraint loss based on the temporal latent variable prediction results corresponding to each branch of the decoupled network model. Combine the physical residual loss, data reconstruction loss and multi-source decoupling constraint loss to construct the joint loss function.

[0132] The data reconstruction loss Used to measure the network's overall prediction performance in the time-frequency domain. Global high-confidence energy point mask extracted by the preceding quantum particle swarm optimization algorithm The mean square error between them is calculated as follows:

[0133] ;

[0134] In the formula, The total number of valid data points used in this loss calculation. The squared term representing the difference between the overall prediction result's time-frequency response and the global high-confidence energy point mask is used to construct the data reconstruction loss in the form of mean square error. The squaring operation can prevent positive and negative errors from canceling each other out and impose a higher penalty on larger deviations.

[0135] The physical residual loss By utilizing the automatic differentiation mechanism in deep learning frameworks, the prediction results of the time-domain latent variables output by the network are directly processed. Differentiating the time variable and substituting it into the generalized multi-source coupled dynamic differential equations to construct dynamic residuals is used to penalize predictions that do not conform to actual vibration patterns; the calculation is shown below:

[0136] .

[0137] An adaptive weight scheduling strategy is used during training to mitigate data reconstruction losses caused by signal masking. When supervision fails, the physical residual loss is amplified. The constraint effect enhances the degree to which each branch of the decoupled network model follows the damping, inertial and coupling terms in the coupled dynamic differential equations, thereby performing reverse physical smoothing deduction in the trajectory break region and finally outputting the high-fidelity continuous time-frequency trajectory energy points after decoupling.

[0138] The multi-source decoupling constraint loss This is used to penalize the energy overlap of different sub-networks at the same time-frequency coordinates, forcing each sub-network to extract independent physical vibration source trajectories; its calculation formula is shown below:

[0139] ;

[0140] In the formula, and These represent the time-domain latent variable prediction results corresponding to the i-th and j-th decoupled network model branches, respectively. The smaller this loss term is, the higher the degree of separation of each vibration source in the time-frequency domain, thereby avoiding multiple sub-networks learning repetitive coupling features.

[0141] A joint loss function is constructed by combining physical residual loss, data reconstruction loss, and multi-source decoupling constraint loss. The calculation process is as follows:

[0142] ;

[0143] In the formula, For data reconstruction loss, For physical residual loss, For multi-source decoupling constraint loss; and These are the weighting coefficients for physical residual loss and multi-source decoupling constraint loss, respectively.

[0144] Steps 3-5: Adaptive weight training is performed on the multi-branch decoupled network model based on the physical information neural network using a joint loss function. This forces each branch of the decoupled network model to perform inverse physical smoothing in the trajectory breakage region, determining the set of continuous and decoupled time-frequency trajectory energy points. In the trained multi-branch decoupled network model based on the physical information neural network, M parallel decoupled network model branches predict the latent variable prediction results of M coupled vibration sources in the time domain. The time-domain latent variable prediction results output by each decoupled network model branch are mapped to the time-frequency domain. The predicted time-frequency response matrix corresponding to each decoupled network model branch is obtained, and the predicted time-frequency responses of each decoupled network model branch are superimposed and fused to output the overall prediction result. The overall prediction result is the set of continuous and decoupled time-frequency trajectory energy points.

[0145] The described multi-branch decoupling network model based on physical information neural networks is a physical information neural network model that combines conventional feedforward neural networks with dynamic constraints of mechanical systems. The overall multi-branch decoupling network model based on physical information neural networks includes an input layer, M parallel decoupling network model branches, a differentiable short-time Fourier transform layer, a fusion output layer, and a loss constraint layer. The M decoupling network model branches are M decoupling subnetworks. The input layer receives the time variable t; the M parallel decoupling network model branches correspond to M coupled vibration sources and are used to predict the latent variables of each vibration source in the time domain; the differentiable short-time Fourier transform layer maps the time-domain latent variable prediction results output by each decoupling network model branch to the time-frequency domain; the fusion output layer superimposes the predicted time-frequency responses of each decoupling network model branch to obtain the overall prediction result; the loss constraint layer jointly calculates the data reconstruction loss, dynamic physical residual loss, and multi-source decoupling constraint loss to construct a joint loss function, which is then used to adaptively train the multi-branch decoupling network model based on physical information neural networks.

[0146] In step 5, a multidimensional quantitative evaluation is performed on the time-frequency trajectory energy point set to obtain a multidimensional quantitative score. Based on this score, a weighted calculation is used to obtain a comprehensive evaluation score function. The multidimensional quantitative evaluation includes indicators for effective energy point coverage, out-of-band noise leakage, temporal continuity, and physical rationality. The effective energy point coverage indicator determines whether truly high-value energy points are completely preserved. The out-of-band noise leakage indicator measures the ability to suppress background noise. The temporal continuity indicator measures the breakage gaps in the trajectory along the time axis. The physical rationality indicator rigorously evaluates whether the rate of change of the trajectory frequency falls within the allowable physical dynamic limits of the mechanical equipment. This multidimensional quantitative evaluation of the time-frequency trajectory energy point set not only breaks the limitations of previous reliance on subjective human judgment, enabling objective comparison of effects under different working conditions, but also uses the comprehensive evaluation score as a reward signal to feed back into Bayesian optimization and quantum particle swarm optimization models to correct thresholds and denoising parameters, forming an adaptive parameter closed-loop adjustment mechanism. The process includes the following steps:

[0147] Step 4-1: Based on the obtained time-frequency trajectory energy point set, calculate the effective energy point coverage index and the out-of-band noise leakage index, and simultaneously calculate the time continuity index for measuring trajectory gaps; the effective energy point coverage index... This is the ratio of the energy of the extracted point set to the total effective energy within the corresponding frequency band; the calculation is shown below:

[0148] ;

[0149] In the formula, This is the set of time-frequency energy point trajectories obtained after partitioning, denoising, and completion processing using the preceding algorithm. It is the set of points in the theoretically effective frequency band region calculated based on mechanical and physical properties.

[0150] out-of-band noise leakage indicators This involves calculating the proportion of background noise energy that was incorrectly extracted; the calculation is shown below:

[0151] ;

[0152] In the formula, It is a set of points in a pure noise frequency band that is completely outside the physical frequency band of mechanical operation and is theoretically composed only of background interference.

[0153] The time continuity category of indicators This refers to the severity of breaks and gaps in the trajectory over time; the calculation process is shown below:

[0154] ;

[0155] In the formula, This refers to the number of times the energy point set has a discontinuity on the time axis that exceeds the system's allowed time window. This represents the total number of energy point time series points.

[0156] Step 4-2: Perform ridge extraction on the time-frequency trajectory energy point set along the time axis to obtain the center frequency evolution curve of the trajectory, and introduce physical evolution mapping to calculate physical rationality indexes; the physical rationality indexes To quantify the rate of change of energy point frequency, i.e., the degree to which the converted acceleration violates the physical limits; for the set of time-frequency trajectory energy points Ridge extraction is performed on the time axis to obtain a continuous function f(t) of the center frequency of the dominant feature as a function of time, which is the center frequency evolution curve of the obtained trajectory; the calculation is shown below:

[0157] ;

[0158] In the formula, This refers to the maximum permissible physical acceleration limit determined based on the operating parameters, structural parameters, design constraints, control limits, historical calibration results, or prior physical knowledge of engineering machinery and equipment; γ is the penalty sensitivity coefficient; K is the kinematic conversion coefficient; the dimension of the kinematic conversion coefficient K is determined by the specific frequency-physical quantity mapping relationship, used to map the time derivative of frequency to physical acceleration; for example, when processing non-contact detection signals containing Doppler frequency shift, K is taken as half the detection wavelength; when processing vibration signals of rotating machinery, the value of K is derived from the relationship between the fault characteristic frequency coefficient and the rotational frequency. Through this conversion mechanism, a reasonable mapping from time-frequency domain characteristics to the real spatial physical limits is ensured.

[0159] Step 4-3: Weight and sum the effective energy point coverage index, out-of-band noise leakage index, time continuity index, and physical rationality index to construct a comprehensive evaluation score function. The calculation is as follows:

[0160] ;

[0161] In the formula, , , , These are the weighting coefficients for effective energy point coverage indicators, out-of-band noise leakage indicators, time continuity indicators, and physical rationality indicators, respectively.

[0162] In step 6, asynchronous closed-loop feedback adjustment of parameters across cycles is implemented based on the comprehensive evaluation score until the comprehensive evaluation score meets the preset engineering standard, and the target time-frequency trajectory energy point set is output. The comprehensive evaluation score for the current analysis cycle is calculated based on the comprehensive evaluation score function, and compared with the preset engineering standard threshold. If the comprehensive evaluation score for the current analysis cycle is lower than the preset engineering standard threshold, the comprehensive evaluation score for the current analysis cycle is used as an environmental reward signal and transmitted back to the Bayesian optimization model and the quantum particle swarm optimization model. By adjusting the kernel function hyperparameters of the Bayesian model and expanding the search boundary of the quantum particle swarm algorithm in the next analysis cycle, the macroscopic parameter correction process is triggered. This continues until the comprehensive evaluation score (Score) of any analysis cycle is higher than or equal to the preset engineering standard threshold. .

[0163] The comprehensive evaluation score of the current analysis period is used as an environmental reward signal and passed back to the Bayesian optimization model and the quantum particle swarm optimization model. By adjusting the kernel function hyperparameter of the Bayesian model and expanding the search boundary of the quantum particle swarm algorithm in the next analysis period, the macroscopic parameter correction process is triggered. First, the environmental reward signal is normalized, that is, the comprehensive evaluation score of the current c-th analysis period is normalized to obtain the normalized environmental reward signal R. c ; and based on the normalized environmental reward signal R c Building a reward gap The calculation process is as follows:

[0164] , ;

[0165] In the formula, This is the comprehensive evaluation score for the current c-th analysis period. This represents the normalized environmental reward signal for the current c-th analysis period. The reward gap between the current c-th analysis cycle and the preset engineering standard; when A large deviation indicates a significant discrepancy between the time-frequency trajectory energy point set obtained in the current analysis cycle and the engineering standard, necessitating enhanced global exploration capabilities for parameter searching in the next analysis cycle. When the value is small, it indicates that the set of time-frequency trajectory energy points obtained in the current analysis cycle is close to the preset engineering standard, and the search expansion intensity needs to be reduced to avoid excessive parameter divergence.

[0166] For adjusting the kernel hyperparameters of the Bayesian optimization model, since a probabilistic surrogate model is established using a covariance kernel function with automatic correlation scaling, the kernel hyperparameters of the Bayesian optimization model are defined as follows: The calculation process is as follows:

[0167]

[0168] In the formula, Let represent the set of kernel function hyperparameters for the current c-th analysis period. This is represented as the length scale vector of the kernel function in the current c-th analysis period. Let the variance of the signal in the current c-th analysis period be denoted as . This is expressed as the noise variance for the current c-th analysis period.

[0169] Based on the reward gap in the current c-th analysis period The hyperparameters of the kernel function in the Bayesian optimization model are updated across periods, and the update formula for the length scaling vector of the kernel function is shown below:

[0170] ;

[0171] The update formula for the signal variance is as follows:

[0172] ;

[0173] The update formula for the noise variance is shown below:

[0174] ;

[0175] In the formula, This is expressed as a length scale adjustment coefficient. This is expressed as the signal variance adjustment coefficient. This represents the noise variance adjustment coefficient; and These represent the lower and upper limits of the length scale, respectively. and These represent the lower and upper limits of the signal variance, respectively. and These represent the lower and upper limits of the noise variance, respectively. This means limiting the updated parameters to a preset range.

[0176] Using the above update method, when the overall evaluation score for the current analysis period is low, the reward gap... Increasing the length scale vector of the kernel function of the Bayesian optimization model makes the surrogate model more sensitive to changes in the position of the dividing line and the energy threshold. At the same time, increasing the signal variance and noise variance of the Bayesian optimization model enables the surrogate model to accommodate stronger operating condition fluctuations and noise disturbances. Thus, the hyperparameters of the kernel function of the Bayesian optimization model are adaptively adjusted based on the environmental reward signal.

[0177] To further improve the Bayesian optimization model's ability to explore regions of insufficiently sampled parameters in the next analysis cycle, the exploration factor of the acquisition function can be adjusted simultaneously. The calculation process is as follows:

[0178] ;

[0179] In the formula, Let the acquisition function exploration factor be represented as the current c-th analysis period. This is represented as the adjustment coefficient of the acquisition function exploration factor. and These represent the lower and upper bounds of the acquisition function's exploration factor, respectively. The acquisition function's exploration factor is not a kernel function hyperparameter, but it can be updated synchronously with the kernel function hyperparameter to enhance the global exploration capability of the Bayesian optimization model in the next analysis cycle.

[0180] In quantum particle swarm optimization (QPSO) models, dynamically widening the local attractor search boundary in the next analysis cycle is defined as follows: the search boundary of the d-th morphological denoising parameter within the r-th frequency sub-region in the current c-th analysis cycle is defined as... The calculation process is as follows:

[0181] ;

[0182] In the formula, This represents the search boundary of the d-th morphological denoising parameter within the r-th frequency domain sub-region in the current c-th analysis period. The d-th morphological denoising parameter includes the area screening threshold and the length, width, and tilt angle of the structuring element.

[0183] Based on the reward gap in the current c-th analysis period Boundary spread coefficients for the next analysis cycle in the quantum particle swarm optimization model The calculation process is as follows:

[0184] ;

[0185] In the formula, Based on the expansion factor, This represents the reward gap in the current c-th analysis period. The corresponding boundary spread gain, and The table shows the minimum and maximum values ​​of the boundary expansion coefficient.

[0186] To avoid ambiguity in the search boundary formula, we first define the search width of the d-th morphological denoising parameter in the current c-th analysis period. The calculation process is as follows:

[0187] ;

[0188] In the formula, This represents the upper limit of the search in the current c-th analysis period. The lower bound for the current c-th analysis period; the globally optimal particle position obtained by the quantum particle swarm optimization model in the current c-th analysis period. As the search center for the next analysis cycle, the search boundary for the d-th morphological denoising parameter is updated, which means updating the upper and lower search boundaries for the next analysis cycle; the calculation is as follows:

[0189] ;

[0190] ;

[0191] In the formula, This represents the globally optimal particle position corresponding to the d-th morphological denoising parameter in the current c-th analysis period. This represents the global allowable lower bound for the d-th morphological denoising parameter. It represents the global allowable upper limit for the d-th morphological denoising parameter; the global allowable lower limit and global allowable upper limit are jointly determined by image scale, sampling frequency, frequency resolution, structuring element size constraints, and engineering experience.

[0192] If the overall evaluation score for the current c-th analysis period is low, the reward gap... Increasing the boundary expansion coefficient will increase the boundary expansion factor. As the threshold increases, the quantum particle swarm optimization model expands its search range for area screening thresholds, structural element lengths, structural element widths, and structural element directional tilt angles in the next analysis cycle. If the comprehensive evaluation score in the current analysis cycle gradually approaches the preset engineering standard threshold, the reward gap... Decrease it, and the boundary spread coefficient will be reduced. As a result, the quantum particle swarm optimization model reduces the invalid search expansion in the next analysis cycle, avoiding excessive divergence in the parameter optimization process. In this way, the quantum particle swarm optimization model can enhance its adaptability to strong noise, trajectory breakage, local orientation changes, and frequency domain sub-region differences in the next analysis cycle.

[0193] Using the comprehensive evaluation score as the objective function value, an asynchronous parameter closed-loop feedback adjustment mechanism is constructed. In actual continuous monitoring or batch processing tasks, the comprehensive evaluation score (Score) output by the current vibration signal analysis window is recorded as a reward signal for the environmental state. If the comprehensive evaluation score is lower than the preset engineering standard threshold... This triggers cross-cycle optimization of macroscopic parameters: the comprehensive evaluation score is passed back to the Bayesian optimization model and the quantum particle swarm optimization model. Through operations such as adjusting the length scale to expand the frequency domain exploration space, the hyperparameters of the surrogate model kernel function in the Bayesian optimization model are updated, and the local attractor search boundary of the quantum particle swarm optimization model in the next analysis cycle is dynamically widened. Through this asynchronous feedback, the Bayesian optimization model and the quantum particle swarm optimization model can adaptively track the time-varying characteristics of the operating conditions without interrupting the current forward inference flow, achieving continuous iterative improvement in the accuracy of the next batch of vibration signal processing.

[0194] Until the overall evaluation score for any analysis period is higher than or equal to the preset engineering standard threshold. If the comprehensive evaluation score meets the preset engineering standards, then the iteration stops and the target time-frequency trajectory energy point set is output. The calculation process is as follows:

[0195]

[0196] In the formula, This is a preset engineering standard threshold.

[0197] like Figure 5 As shown, the final set of time-frequency trajectory energy points will be obtained. Superimposed on the initial time-frequency energy map The above forms a visualization result of background time-frequency map superimposed with foreground trajectory energy points; where time t is the horizontal axis, frequency f is the vertical axis, the magnitude of time-frequency energy is the background color intensity, and the extracted energy point trajectory is used as the foreground highlight. Figure 5 The image shows a visualization of the time-frequency energy point trajectory set after processing by a multi-branch decoupled network model based on the physical information neural network. That is, after the multi-branch decoupled network model based on the physical information neural network performs trajectory separation and completion processing on the global high-confidence energy point mask, the image shows a continuous and decoupled time-frequency trajectory energy point set. Figure 5 The signal in the data is characterized by multiple clear, continuous, and decoupled independent trajectories, which conforms to the real physical evolution laws of mechanics and has physical rationality.

[0198] The system also includes a data acquisition module, an adaptive partitioning module, a morphological denoising module, a trajectory decoupling module, and a feedback evaluation module. The data acquisition module acquires the original one-dimensional vibration signal of the large mechanical system and performs a short-time Fourier transform on the original one-dimensional vibration signal to obtain an initial time-frequency energy map. The adaptive partitioning module performs adaptive frequency domain partitioning on the initial time-frequency energy map based on a Bayesian optimization model to obtain various frequency domain sub-regions, and collaboratively determines the energy threshold corresponding to each frequency domain sub-region. The morphological denoising module converts the initial time-frequency energy map into a binary mask based on the energy threshold corresponding to each frequency domain sub-region, and uses a quantum particle swarm optimization model to perform adaptive optimization denoising processing of morphological parameters on each frequency domain sub-region to determine a global high-confidence energy point mask. The trajectory decoupling module constructs a multi-branch decoupling network model based on a physical information neural network to perform time-frequency trajectory energy point separation and completion processing on the global high-confidence energy point mask, ultimately obtaining a continuous and decoupled set of time-frequency trajectory energy points. In the feedback evaluation module, a multi-dimensional quantitative evaluation is performed on the time-frequency trajectory energy point set to obtain a multi-dimensional quantitative score. Based on the multi-dimensional quantitative score, a weighted calculation is used to obtain a comprehensive evaluation score function. According to the comprehensive evaluation score, cross-cycle asynchronous parameter closed-loop feedback adjustment is implemented until the comprehensive evaluation score meets the preset engineering standard, and the target time-frequency trajectory energy point set is output.

[0199] The above description is merely a preferred embodiment of the present invention. Those skilled in the art can make several modifications and optimizations based on the above disclosure without departing from the basic principles described above. These modifications and optimizations should be considered within the scope of protection as understood by the present invention.

Claims

1. A vibration signal processing method based on adaptive frequency domain partitioning and morphological denoising, characterized in that: Includes the following steps: Step 1: Obtain the original one-dimensional vibration signal of the large mechanical system, and perform a short-time Fourier transform on the original one-dimensional vibration signal to obtain the initial time-frequency energy map; Step 2: Based on the Bayesian optimization model, perform adaptive frequency domain partitioning on the initial time-frequency energy map to obtain each frequency domain sub-region, and collaboratively determine the energy threshold corresponding to each frequency domain sub-region; Step 3: Based on the energy threshold corresponding to each frequency domain sub-region, convert the initial time-frequency energy map into a binary mask. Use the quantum particle swarm optimization model to perform adaptive optimization and noise reduction of morphological parameters for each frequency domain sub-region to determine the global high-confidence energy point mask. Step 4: Construct a multi-branch decoupled network model based on physical information neural network to perform time-frequency trajectory energy point separation and completion processing on the global high confidence energy point mask, and finally obtain a continuous and decoupled time-frequency trajectory energy point set; Step 5: Perform multidimensional quantitative evaluation on the time-frequency trajectory energy point set to obtain a multidimensional quantitative score, and use weighted calculation based on the multidimensional quantitative score to obtain a comprehensive evaluation score function; Step 6: Implement cross-cycle asynchronous parameter closed-loop feedback adjustment based on the comprehensive evaluation score until the comprehensive evaluation score meets the preset engineering standard, and output the target time-frequency trajectory energy point set.

2. The vibration signal processing method based on adaptive frequency domain partitioning and morphological denoising according to claim 1, characterized in that: In step 2, the initial time-frequency energy map is adaptively partitioned into frequency domains based on a Bayesian optimization model to obtain each frequency domain sub-region, and the energy threshold corresponding to each frequency domain sub-region is determined collaboratively; this includes the following steps: Step 1-1: Convert the initial time-frequency energy map from the decibel domain to the linear energy domain, and continuously integrate the linear energy domain along the time dimension to obtain the overall energy distribution curve in the frequency direction; Step 1-2: Generate a set of candidate segmentation locations based on the specific characteristics of the overall energy distribution curve, and construct a set of parameter vectors to be optimized; the set of parameter vectors to be optimized includes the specific location of the segmentation line and the energy threshold corresponding to each frequency domain sub-region. Steps 1-3: Based on the set of parameter vectors to be optimized, construct a comprehensive objective function that combines the effective energy retention rate, noise leakage penalty term, and initial trajectory coherence term. Use Gaussian process regression to establish a probabilistic proxy model for the comprehensive objective function. Steps 1-4: Using the expected improvement index as the acquisition function, perform iterative search operations to determine the optimal segmentation line position and the optimal energy threshold of each frequency domain sub-region; based on the optimal segmentation line position, divide the initial time-frequency energy map into each frequency domain sub-region.

3. The vibration signal processing method based on adaptive frequency domain partitioning and morphological denoising according to claim 2, characterized in that: In steps 1-3, a comprehensive objective function is constructed based on the set of parameter vectors to be optimized, which is a combination of the effective energy retention rate, the noise leakage penalty term, and the initial trajectory coherence term; the calculation process is shown below: ； In the formula, For effective energy retention rate, As a penalty item for noise leakage, For the initial trajectory coherence term, , , These are the optimization weight coefficients for the effective energy retention rate, the noise leakage penalty term, and the initial trajectory coherence term, respectively. A probabilistic surrogate model for the comprehensive objective function is established using Gaussian process regression, as shown below: ； In the formula, It is a mean function. It is a covariance kernel function that measures the similarity and uncertainty among combinations of parameters.

4. The vibration signal processing method based on adaptive frequency domain partitioning and morphological denoising according to claim 1, characterized in that: In step 3, the initial time-frequency energy map is converted into a binary mask based on the energy threshold corresponding to each frequency domain sub-region. A quantum particle swarm optimization model is then used to adaptively optimize and denoise each frequency domain sub-region using morphological parameters to determine the global high-confidence energy point mask. This includes the following steps: Step 2-1: For each divided frequency domain sub-region, the morphological denoising parameters are defined as the particle position vectors inside the quantum particle swarm optimization model. The morphological denoising parameters include the area screening threshold and the length, width and orientation tilt of the structuring element. Step 2-2: Use the ratio of the degree of retention of effective energy points after denoising to the degree of noise suppression as the fitness function; Steps 2-3: By calculating the average optimal position of individuals and local attractor points of the entire population, the particle positions are continuously updated and iterated to determine the optimal morphological denoising parameters. Steps 2-4: Use the obtained optimal area screening threshold to remove isolated noise patches in the local area, and use the obtained optimal structuring element to perform morphological closing operation to repair small gaps in the trajectory, and finally generate a global high-confidence energy point mask.

5. The vibration signal processing method based on adaptive frequency domain partitioning and morphological denoising according to claim 1, characterized in that: In step 4, a multi-branch decoupled network model based on a physical information neural network is constructed to perform time-frequency trajectory energy point separation and completion processing on the global high-confidence energy point mask, ultimately obtaining a continuous and decoupled set of time-frequency trajectory energy points; including the following steps: Step 3-1: Establish a set of generalized multi-source coupled dynamic differential equations with unknown physical parameters; Step 3-2: Construct a multi-branch decoupling network model based on physical information neural network, using the received time variable as input, and each branch of the decoupling network model directly predicts the time-domain latent variable prediction result of a single vibration source. Step 3-3: Use a differentiable short-time Fourier transform layer to map the time-domain latent variable prediction results of each decoupled network model branch to the time-frequency domain, obtain the prediction time-frequency response matrix of each decoupled network model branch, and fuse the prediction time-frequency responses of each decoupled network model branch to output the overall prediction result. Steps 3-4: Differentiate the temporal latent variable prediction results output by each branch of the decoupled network model and substitute them into the generalized multi-source coupling dynamic differential equation system to construct the dynamic residual as the physical residual loss. Construct the data reconstruction loss based on the overall prediction results and construct the multi-source decoupling constraint loss based on the temporal latent variable prediction results corresponding to each branch of the decoupled network model. Combine the physical residual loss, data reconstruction loss, and multi-source decoupling constraint loss to construct the joint loss function. Steps 3-5: Adaptive weight training is performed on the multi-branch decoupled network model based on physical information neural network using the joint loss function. This forces the branches of the decoupled network model to perform inverse physical smoothing in the trajectory breakage area, thereby determining the set of continuous and decoupled time-frequency trajectory energy points.

6. The vibration signal processing method based on adaptive frequency domain partitioning and morphological denoising according to claim 1, characterized in that: In step 3-1, a set of generalized multi-source coupled dynamic differential equations with unknown physical parameters is established; the large mechanical system is assumed to consist of M coupled vibration sources, and the calculation process is as follows: ； In the formula, Let be the latent variable of the vibration response of the i-th independent vibration source in the time domain. , , Let be the equivalent mass parameter, damping parameter, and stiffness parameter of the i-th vibration source, respectively. It is represented as the coupling transmission coefficient between different vibration sources.

7. The vibration signal processing method based on adaptive frequency domain partitioning and morphological denoising according to claim 1, characterized in that: In step 5, a multidimensional quantitative evaluation is performed on the time-frequency trajectory energy point set to obtain a multidimensional quantitative score. Based on the multidimensional quantitative score, a weighted calculation is used to obtain a comprehensive evaluation score function; this includes the following steps: Step 4-1: Based on the obtained time-frequency trajectory energy point set, calculate the effective energy point coverage index and out-of-band noise leakage index respectively, and at the same time calculate the time continuity index for measuring trajectory gaps. Step 4-2: Perform ridge extraction on the time axis for the set of energy points of the time-frequency trajectory to obtain the center frequency evolution curve of the trajectory, and introduce physical evolution mapping to calculate physical rationality index. Step 4-3: Weight the effective energy point coverage index, out-of-band noise leakage index, time continuity index, and physical rationality index to construct a comprehensive evaluation score function.

8. The vibration signal processing method based on adaptive frequency domain partitioning and morphological denoising according to claim 1, characterized in that: In step 6, asynchronous parameter closed-loop feedback adjustment is implemented across cycles based on the comprehensive evaluation score until the comprehensive evaluation score meets the preset engineering standard, and the target time-frequency trajectory energy point set is output. The comprehensive evaluation score for the current analysis period is calculated based on the comprehensive evaluation score function. This score is then compared with a preset engineering standard threshold. If the comprehensive evaluation score for the current analysis period is lower than the preset engineering standard threshold, it is used as an environmental reward signal and passed back to the Bayesian optimization model and the quantum particle swarm optimization model. By adjusting the kernel function hyperparameters of the Bayesian model and expanding the search boundary of the quantum particle swarm algorithm in the next analysis period, a correction process for macroscopic parameters is triggered. This process continues until the comprehensive evaluation score for any analysis period is higher than or equal to the preset engineering standard threshold.

Images