Bolt fault identification method, system and device based on ResMLP and medium

Abstract

The invention relates to a bolt fault identification method, system and device based on ResMLP and a medium. The method comprises the following steps: acquiring a transient impact sound signal; decomposing the transient impact sound signal by using an SVMD method to obtain a primary screening mode component; performing singular value decomposition, filtering and reconstruction on the preliminary screening mode component to obtain a de-noised signal; performing feature extraction according to the de-noised signal to obtain a target feature; inputting the target feature into a bolt fault recognition model to obtain a bolt fault recognition result; wherein the bolt fault identification model is a pre-trained residual multi-layer perceptron model. According to the method, after successive variational mode decomposition optimized by the crown porcupine optimization algorithm is adopted and singular value decomposition is used for denoising, most of noise in the signals can be effectively removed, main signals are reserved, the diagnosis precision is effectively improved, and the diagnosis accuracy is improved. Compared with a traditional neural network, the residual multi-layer perceptron can effectively improve the accuracy of taking the impact sound signal as a fault diagnosis basis.

Description

Technical Field

[0001] The present application relates to the field of fault monitoring and diagnosis, and in particular to a bolt fault identification method, system, device and medium based on ResMLP. Background Art

[0002] Bolt connections are widely used in engineering equipment and mechanical structures, with advantages such as simple structure, easy installation, and good economy. However, the preload of bolts is affected by many factors in actual applications, which can easily lead to problems such as bolt loosening and failure, thereby affecting the normal operation of the equipment. Traditional fault identification methods rely on manual inspection, which consumes a lot of manpower and material resources, and it is difficult to achieve early fault identification. In recent years, intelligent identification methods have gradually emerged, but the commonly used vibration signal acquisition methods are often limited by spatial conditions in industrial sites. In contrast, although impact sound signals are easy to collect, they are easily interfered by environmental noise, and new processing methods are needed to reduce the noise of impact sound signals. At the same time, the existing methods of fault identification using acoustic signals are too dependent on the basic parameters of the algorithm or model itself. Therefore, it is necessary to provide a bolt fault identification method, system, equipment and medium based on ResMLP. Summary of the Invention

[0003] In view of the shortcomings of the prior art described above, the purpose of the present invention is to provide a bolt fault identification method, system, device and medium based on ResMLP, which is used to solve the problems in the prior art that impact sound signals are easily interfered with by environmental noise during the acquisition process, traditional noise reduction methods are difficult to effectively remove noise, and the denoising algorithm is overly dependent on basic parameters. In addition, traditional neural network models have problems of information loss and slow convergence in fault identification, which affect the identification accuracy and efficiency.

[0004] To achieve the above-mentioned objectives and other related objectives, the present invention provides a bolt fault identification method, system, device and medium based on ResMLP, which are applied to the field of fault monitoring and diagnosis. The method includes: obtaining a transient impact sound signal; decomposing the transient impact sound signal using the SVMD (Sequential Variational Mode Decomposition) method to obtain a primary screening modal component; filtering and reconstructing the primary screening modal component through singular value decomposition (SVD) to obtain a denoised signal; performing feature extraction based on the denoised signal to obtain a target feature; inputting the target feature into a bolt fault identification model to obtain a bolt fault identification result; wherein the bolt fault identification model is a pre-trained residual multilayer perceptron (Residual Multilayer Perceptron, ResMLP) model.

[0005] In one embodiment of the present invention, the step of decomposing the transient impact sound signal using the SVMD method to obtain preliminary screening modal components includes: based on the transient impact sound signal, using the Crested Porcupine Optimizer (CPO) algorithm to obtain the optimal penalty coefficient of the SVMD method; standardizing the transient impact sound signal to obtain a standard sound signal; decomposing the standard sound signal according to the optimal penalty coefficient using the SVMD decomposition method to obtain multiple modal components, calculating the modal energy corresponding to each modal component, and taking the modal component with the largest modal energy as the principal component mode; subtracting the principal component mode from the standard sound signal to obtain and judge the residual signal: when the residual signal meets the preset threshold condition, all the principal component modes are used as preliminary screening modal components; otherwise, the residual signal is used as the standard sound signal, and the SVMD analysis method is continued to be used for decomposition to determine the principal component mode, and the residual signal is updated until the maximum number of iterations is reached or the residual signal meets the preset threshold condition, and all the principal component modes are used as preliminary screening modal components.

[0006] In one embodiment of the present invention, the step of obtaining the optimal penalty coefficient of the SVMD method based on the transient impact sound signal using the crested porcupine optimization algorithm includes: randomly generating multiple individuals and taking all individuals as a population; wherein one individual represents one penalty coefficient; decomposing the transient impact sound signal based on each penalty coefficient of the population using the SVMD method to obtain corresponding multiple modal components, and calculating the envelope entropy of the modal component corresponding to each penalty coefficient; optimizing the individuals of the population using the crested porcupine optimization algorithm, and again decomposing the transient impact sound signal based on the optimized population and calculating the envelope entropy until the number of optimization iterations reaches a preset maximum number; selecting a preset number of envelope entropies in order from small to large, and using the individuals corresponding thereto as the optimal penalty coefficient.

[0007] In one embodiment of the present invention, the step of decomposing, filtering and reconstructing the primary screening modal components through singular value decomposition to obtain a denoised signal includes: calculating the energy entropy and kurtosis of each of the primary screening modal components, and determining the target modal component from the primary screening modal components based on the energy entropy and kurtosis; decomposing the target modal component using singular value decomposition to obtain multiple singular values, and truncating the singular values ​​through a preset threshold to obtain high-order screening singular values ​​and low-order shrinkage singular values; performing singular value reconstruction on the low-order shrinkage singular values ​​and the high-order screening singular values ​​to obtain a denoised signal.

[0008] In one embodiment of the present invention, the step of calculating the energy entropy and kurtosis of each of the preliminary screening modal components and determining the target modal component based on the energy entropy and kurtosis includes: calculating the energy entropy and kurtosis of each modal component in the preliminary screening modal components, and calculating the weighted sum of the energy entropy and the kurtosis; and taking the preliminary screening modal component whose weighted sum is less than a preset screening threshold as the target modal component.

[0009] In one embodiment of the present invention, the step of decomposing the target modal component using a singular value to obtain a plurality of singular values, and truncating the singular values ​​by a preset threshold to obtain high-order screening singular values ​​and low-order shrinkage singular values ​​includes: decomposing the target modal component using a singular value to obtain a plurality of singular values; arranging the singular values ​​in descending order and truncating them according to a preset threshold; processing the singular values ​​greater than or equal to the preset threshold using a hard threshold method to obtain high-order screening singular values; and processing the singular values ​​less than the preset threshold using a soft threshold method to obtain low-order shrinkage singular values.

[0010] In one embodiment of the present invention, the process of inputting the target feature into the bolt fault recognition model to obtain the bolt fault recognition result includes: inputting the target feature into the residual processing module of the bolt fault recognition model, performing residual processing on two residual blocks in succession to obtain shallow features, and performing feature fusion according to the target feature to output the first fused feature; inputting the first fused feature into the classification output module of the bolt fault recognition model, extracting deep features and outputting the bolt fault recognition result.

[0011] In one embodiment of the present invention, a bolt fault identification system based on ResMLP is also provided, which includes: a signal sampling module for obtaining a transient impact sound signal; a decomposition and preliminary screening module for decomposing the transient impact sound signal using the SVMD method to obtain a preliminary screening modal component; a denoising module for decomposing, filtering and reconstructing the preliminary screening modal component through singular value decomposition to obtain a denoised signal; a feature extraction module for performing feature extraction based on the denoised signal to obtain a target feature; a detection output module for inputting the target feature into a bolt fault identification model to obtain a bolt fault identification result; wherein the bolt fault identification model is a pre-trained residual multi-layer perceptron model.

[0012] In one embodiment of the present invention, an electronic device is also provided, comprising: one or more processors; a storage device for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the electronic device implements any of the above-mentioned ResMLP-based bolt fault identification methods.

[0013] In one embodiment of the present invention, a computer-readable storage medium is further provided, on which a computer program is stored. When the computer program is executed by a processor of a computer, the computer is caused to execute any of the above-mentioned bolt fault identification methods based on ResMLP.

[0014] As described above, the present invention provides a method, system, device and medium for bolt fault identification based on ResMLP, which has the following beneficial effects: first, a transient impact sound signal is acquired, and then the successive variational modal decomposition method is used to decompose it into preliminary screening modal components, without presetting the number of modal components, thereby avoiding the problem of over-decomposition or under-decomposition caused by improper setting of the number of modal components in the traditional VMD method. The crown porcupine optimization algorithm is used to obtain the optimal penalty coefficient of the successive variational modal decomposition method, which can adaptively find the optimal decomposition parameters and improve the accuracy and reliability of the decomposition. In the denoising process, important features in the signal can be retained by selecting high-order singular values ​​and low-order singular values ​​for reconstruction. The target features are input into the pre-trained residual multi-layer perceptron model. The addition of residual blocks can alleviate the gradient problem, enhance feature transfer, and improve the neural network structure, thereby improving the accuracy and efficiency of fault identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 A schematic diagram of a flow chart of a bolt fault identification method based on ResMLP provided in an embodiment of the present invention;

[0016] Figure 2 A schematic diagram of a flow chart of a bolt fault identification method based on ResMLP provided in an embodiment of the present invention;

[0017] Figure 3 Confusion matrix diagram of the residual multilayer perceptron neural network test set provided by the embodiment of the present invention;

[0018] Figure 4 Confusion matrix diagram of the MLP neural network test set provided by an embodiment of the present invention;

[0019] Figure 5 Confusion matrix diagram of the CNN neural network test set provided by the embodiment of the present invention;

[0020] Figure 6 Shown is a structural block diagram of a bolt fault identification system based on ResMLP provided by one embodiment of the present invention;

[0021] Figure 7 Shown is a structural schematic diagram of an electronic device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0022] The following describes the embodiments of the present invention through specific examples. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments. The details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the following embodiments and features in the embodiments can be combined with each other unless they conflict.

[0023] It should be noted that the illustrations provided in the following embodiments are merely schematic illustrations of the basic concept of the present invention. Therefore, the illustrations only show components related to the present invention and are not drawn according to the number, shape, and size of components in actual implementation. In actual implementation, the type, quantity, and proportion of each component may be changed arbitrarily, and the component layout may also be more complex.

[0024] In the following description, numerous details are discussed to provide a more thorough explanation of the embodiments of the present invention. However, it will be apparent to those skilled in the art that the embodiments of the present invention may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring the embodiments of the present invention.

[0025] See Figure 1 , shows a flowchart of a bolt fault identification method based on ResMLP in an exemplary embodiment of the present application, comprising the following steps:

[0026] S100: Acquire a transient impact sound signal.

[0027] Currently, most commonly used fault feature extraction methods are based on vibration signals. When equipment fails, vibration signals can reflect the periodic impacts on the equipment. Accelerometers are typically used to collect these vibration signals and extract and analyze their characteristic components. While vibration signals offer advantages such as clear signatures, minimal noise interference during transmission, and a high fault identification success rate, collecting vibration signals often requires installing vibration sensors on the equipment and wiring them, which is not feasible in most industrial sites.

[0028] Therefore, the present invention selects an acoustic signal that also carries fault characteristic information, obtains the transient impact acoustic signal of the bolt under working conditions at a preset sampling frequency, and obtains sample data of a fixed length. The preferred value is: the preset sampling frequency is 25600 Hz, and the sample data length of the signal is 8192 sampling points.

[0029] S200: Decompose the transient impact sound signal using the SVMD method to obtain the primary screening modal components.

[0030] Transient impact acoustic signals acquired under working conditions contain significant background noise, requiring denoising to accurately identify bolt faults. Existing denoising methods include adaptive wavelet threshold denoising and mutual information entropy denoising. To avoid the over-reliance on fundamental parameters of these algorithms, this paper employs the SVMD method for preliminary denoising of transient acoustic signals.

[0031] Specifically, in one embodiment of the present invention, the step of decomposing the transient impact sound signal using the SVMD method to obtain the primary screening modal components includes:

[0032] S210 , based on the transient impact sound signal, using a crested porcupine optimization algorithm to obtain an optimal penalty coefficient of the SVMD method.

[0033] When applying traditional VMD decomposition methods, an inappropriate number of decomposition levels can lead to over-decomposition (redundant modes) or under-decomposition (feature loss). Inappropriate penalty factors can also cause signal distortion or introduce noise. Therefore, to ensure that the penalty factor selection does not affect the decomposition performance of the SVMD method, a preferred embodiment of the present invention uses the crested porcupine optimization algorithm to optimize the penalty factor for the SVMD method.

[0034] Furthermore, in a specific embodiment of the present invention, the step of obtaining the optimal penalty coefficient of the SVMD method using the crested porcupine optimization algorithm based on the transient impact sound signal includes:

[0035] S211, randomly generate multiple individuals and regard all individuals as a population; where one individual represents one penalty coefficient;

[0036] The purpose of this step is to initialize the population of the crested porcupine optimization algorithm. One individual corresponds to a penalty coefficient and a number of individuals are randomly generated as the population of the crested porcupine optimization algorithm. The formula is:

[0037] in, is the position of the i-th individual, α min and α max are the maximum and minimum values ​​of the penalty coefficient respectively, and rand(0,1) is a random number uniformly distributed in the range of [0,1].

[0038] S212 , decomposing the transient impact sound signal based on each penalty coefficient of the population using the SVMD method to obtain corresponding multiple modal components, and calculating the envelope entropy of the modal component corresponding to each penalty coefficient.

[0039] The SVMD method is based on the VMD method to iteratively extract the main frequency components. However, SVMD does not require the pre-specified number of modal components. The VMD decomposition process is as follows: in each iteration, VMD decomposition is performed on the residual signal to solve the variational optimization problem to obtain multiple modal components of the transient impact sound signal. The VMD decomposition expression is as follows:

[0040] Among them, μ k (t) is the kth modal component obtained by decomposition, K is the total number of modal components, ω k is the center frequency of the kth modal component, represents the time derivative, δ(t) is the Dirac function, r n (t) is the residual signal of the nth iteration. The residual signal is initially the input transient impulse signal f(t), t is the time series value, and j is the imaginary unit.

[0041] Correspondingly, the decomposition method of SVMD is:

[0042] Among them, α is the penalty coefficient, λ is the additional regularization coefficient, r k-1 (t), is the residual signal of the previous step, r0 = f(t).

[0043] After obtaining the modal components using the SVMD method according to each penalty coefficient, the envelope entropy of the modal components is calculated as follows:

[0044] H i is the total envelope entropy corresponding to the i-th penalty coefficient, K is the number of modal components, H i,k is the envelope entropy of the k modal components corresponding to the i-th penalty coefficient, and its calculation formula is: H i,k =-∫p i,k (e)logp i,k (e)de, where e is the envelope value, p i,k (e) is the normalized probability density function of the envelope of the k modal components corresponding to the i-th penalty coefficient, which is calculated by the following formula:

[0045] E i,k is the total energy of the envelope of the k modal components corresponding to the i-th penalty coefficient, μ i,k (t) is the k modal components corresponding to the i-th penalty coefficient.

[0046] S213, optimizing the individuals of the population by using the crested porcupine optimization algorithm, and again decomposing the transient impact sound signal and calculating the envelope entropy based on the optimized population, until the number of optimizations reaches a preset maximum number of iterations.

[0047] The specific calculation process of optimizing the individuals of the population by the crested porcupine optimization algorithm is as follows:

[0048] First, in the exploration phase, a global random search is performed, and the population size is dynamically adjusted to balance the convergence speed and diversity. The formula is as follows:

[0049] N is the current population size, N min is the minimum population size, T is the number of cycles, and n is the current number of iterations.

[0050] Secondly, through the first defense strategy, visual intimidation, random perturbation is performed in the direction. When the predator moves away, the individual adjusts its position by a normally distributed random number:

[0051] is the position of the i-th individual at the t-th iteration, is the optimal solution for the current crested porcupine position, that is, the current optimal penalty coefficient, That is, the middle position between the i-th individual and the random individual r, r is a random individual, τ1 and τ1 are random numbers in the range [0,1].

[0052] The second defense strategy is to conduct a localized and refined search using sound intimidation. When a predator approaches, the sound intimidation is simulated by random vectors:

[0053] is a binary vector of 0 or 1, τ3 is a random number in [0,1], and r1 and r1 are random individual indices, that is, the positions of two different individuals selected randomly.

[0054] Then, during development:

[0055] The third defense strategy, namely the odor attack, is to converge strongly to the optimal solution and control the local search range through the odor diffusion factor:

[0056] is the odor diffusion factor, is the total envelope entropy calculated at the position of the i-th individual at the t-th iteration, that is, the total envelope entropy corresponding to the penalty coefficient represented by the individual. ε is the minimum value set to prevent the denominator from being zero, for example: e -10 , γ t =2 rand(1-t / t max ) is the defense attenuation factor, Control the search direction, the elements are ±1, is the third randomly selected individual position, Based on Calculate the envelope entropy.

[0057] The fourth defense strategy, physical attack, simulates a brute force attack of inelastic collision:

[0058] α is the convergence speed factor, the default value is a constant 0.2, is the collision force, τ4 and τ5 are uniformly distributed random scalars in the range [0,1], and δ is the direction control scalar: ±1.

[0059] In summary, the initially generated population individuals will achieve a dynamic balance between exploration and exploitation after going through four defense strategy stages, namely, visual intimidation, sound intimidation, odor attack, and physical attack, and gradually converge to the vicinity of the global optimal solution. If the termination condition is not met, the cyclic population reduction technique of the crested porcupine optimization algorithm is used to adjust the population size and continue to iterate until the termination condition is met. The penalty coefficient represented by each individual is obtained, and the envelope entropy corresponding to each penalty coefficient is calculated.

[0060] S214: Select a preset number of envelope entropies in order from small to large, and use the individuals corresponding to the envelope entropies as optimal penalty coefficients.

[0061] S220 , normalizing the transient impact sound signal to obtain a standard sound signal.

[0062] S230, decomposing the standard acoustic signal according to the optimal penalty coefficient using the SVMD decomposition method to obtain multiple modal components, calculating the modal energy corresponding to each modal component, and taking the modal component with the largest modal energy as the principal component mode.

[0063] Calculate the modal energy corresponding to each modal component separately, and select the modal component with the largest modal energy as the principal component mode. The modal energy calculation expression is as follows:

[0064] Among them, E k is the energy of the kth modal component, k* is the index corresponding to the modal component with the largest energy. After removing the principal component, the residual signal is updated by the following formula:

[0065] r n+1 (t) = r n (t)-μ k* (t), where r n+1 (t) is the residual signal at the n+1th iteration, r n (t) is the residual signal at the nth iteration, μ k* (t) is the modal component with the largest energy in this decomposition.

[0066] S250, subtracting the principal component mode from the standard sound signal to obtain and determine a residual signal:

[0067] When the residual signal meets the preset threshold condition, all the principal component modes are used as the primary screening modal components;

[0068] The calculation formula of the preset threshold condition is as follows:

[0069] ∫|r n (t)| 2 dt<εor n≥N max , where ε is the preset convergence threshold, n is the current number of iterations, N max is the maximum number of iterations, r n (t) is the residual signal at the nth iteration.

[0070] Otherwise, the residual signal is used as the standard sound signal, and the SVMD analysis method is continued to be used for decomposition to determine the principal component mode, and the residual signal is updated until the maximum number of iterations is reached or the residual signal meets the preset threshold condition, and all the principal component modes are used as the initial screening modal components.

[0071] S300: Decompose, filter and reconstruct the pre-screened modal components through singular value decomposition to obtain a denoised signal.

[0072] The initial screened modal components are further denoised through singular value decomposition and filtering to obtain a denoised signal suitable for input into the ResMLP-based bolt fault identification system.

[0073] Specifically, in a specific embodiment of the present invention, the steps of performing singular value decomposition, filtering, and reconstructing the pre-screened modal components to obtain a denoised signal include:

[0074] S310, calculating the energy entropy and kurtosis of each pre-screened modal component, and determining the target modal component from the pre-screened modal components based on the energy entropy and kurtosis.

[0075] Using optimized SVMD parameters, the one-dimensional impact acoustic signal is decomposed into modal components. The energy entropy and kurtosis of each modal component are calculated. Energy entropy quantifies the uniformity of the signal's energy distribution and effectively identifies the energy concentration characteristics of the dominant component in the signal. Kurtosis is highly sensitive to transient impulses in the signal. Transient pulses generated by faults can significantly increase the kurtosis value, which is in stark contrast to the smooth characteristics of normal signals.

[0076] Furthermore, in one embodiment of the present invention, the step of calculating the energy entropy and kurtosis of each of the pre-screened modal components and determining the target modal component based on the energy entropy and kurtosis includes:

[0077] S311, calculating the energy entropy and kurtosis of each modal component in the primary screening modal components, and calculating the weighted sum of the energy entropy and the kurtosis.

[0078] S312: The pre-screened modal component whose weighted sum is less than a preset screening threshold is used as the target modal component.

[0079] S320 , decomposing the target modal component using singular value decomposition to obtain multiple singular values, and truncating the singular values ​​by a preset threshold to obtain high-order screening singular values ​​and low-order contraction singular values.

[0080] First, singular value decomposition is performed on the target modal component to obtain multiple singular values;

[0081] The target modal components are converted into matrix form through the Hankel matrix, and the matrix is ​​decomposed to obtain the left orthogonal matrix U, whose column vectors are called left singular vectors, the diagonal matrix ∑, whose non-zero real numbers on the diagonal are called singular values ​​and are arranged in descending order, and the right orthogonal matrix V, whose column vectors are called right singular vectors.

[0082] For any real matrix A corresponding to a target modal component, there exists a left orthogonal matrix U and a right orthogonal matrix V, as well as a diagonal matrix ∑, such that: A=U∑V T , where V T is an orthogonal matrix of the right orthogonal matrix V. The non-zero values ​​in the diagonal matrix ∑ are the singular values ​​corresponding to the target modal components. The corresponding singular values ​​are obtained by decomposing all target modal components through singular value decomposition.

[0083] Secondly, the singular values ​​are arranged in descending order and truncated according to a preset threshold;

[0084] After sorting all singular values ​​from large to small, the singular values ​​are truncated according to the preset threshold:

[0085] The singular values ​​greater than or equal to the preset threshold are regarded as high-order singular values, and the hard threshold method is used to process them to obtain high-order screening singular values. The hard threshold method processing expression is as follows, that is, the first k singular values ​​are retained and the rest are set to zero: ∑ k =diag(σ1,σ2,...,σ k ,0,...,0), where σ k is the kth singular value in the matrix signal, and σ1≥σ2≥…≥σ k ,∑ k A matrix that retains the first k singular values ​​and the rest are 0.

[0086] The singular values ​​that are smaller than the preset threshold are regarded as low-order singular values, and are processed by a soft threshold method to obtain low-order contraction singular values.

[0087] The soft threshold method processes the expression as follows, which shrinks the singular value:

[0088] σ' k =max(σk -τ,0), where σ' k is the low-order contracted singular value after the k-th singular value is contracted, σ k is the kth singular value, and τ is the threshold.

[0089] S330 , reconstructing the low-order contracted singular values ​​and the high-order screened singular values ​​to obtain a denoised signal.

[0090] Using a hybrid soft and hard thresholding method, for low-order preferred modal signals, where energy is relatively concentrated and stable, soft thresholding can avoid signal distortion and preserve smoothness. For high-order preferred modal components, which contain impulse characteristics, hard thresholding can preserve transient characteristics. Using a hybrid soft and hard thresholding method to denoise the target component can more thoroughly remove noise while preserving high-frequency details.

[0091] The reconstruction matrix is ​​constructed based on high-order screening singular values ​​and low-order shrinkage singular values. The construction process of the reconstruction matrix is:

[0092] A denoised =U∑ k V T , where A denoised is the reconstructed matrix, U and V are the left and right orthogonal matrices corresponding to the decomposition of the target modal component, respectively.

[0093] After obtaining the reconstruction matrix, the denoised signal is recovered from the reconstruction matrix by anti-diagonal averaging to obtain the denoised signal.

[0094] Furthermore, the present invention further verifies the method of denoising the signal by utilizing the above-mentioned singular value decomposition to obtain multiple singular values, and truncating the singular values ​​by a preset threshold to obtain high-order screening singular values ​​and low-order contraction singular values, and reconstructing the low-order contraction singular values ​​and high-order screening singular values ​​into singular values:

[0095] The generated simulation signal is added with Gaussian white noise to obtain a noisy signal, and the noise reduction method of the present invention is used to reduce the noise. The results are summarized in Table 1 below. By comparing the signal-to-noise ratio and root mean square error of the noisy signal and the noise-reduced signal, it can be seen that after the noise reduction method of the present invention is used, the signal-to-noise ratio and root mean square error of the noise-reduced signal are significantly different from those of the noisy signal. The noise reduction method can effectively improve the signal-to-noise ratio of the signal and reduce the root mean square error of the signal.

[0096] Table 1 Comparison of indicators before and after noise reduction

[0097]

[0098] S400: extracting features based on the denoised signal to obtain target features.

[0099] S500, inputting the target feature into a bolt fault recognition model to obtain a bolt fault recognition result; wherein the bolt fault recognition model is a pre-trained residual multi-layer perceptron model.

[0100] Specifically, in one embodiment of the present invention, the process of inputting the target feature into the bolt fault identification model to obtain the bolt fault identification result includes:

[0101] S510: The target feature is input into the residual processing module of the bolt fault recognition model, and is processed by two residual blocks to obtain shallow features. Feature fusion is performed based on the target feature to output a first fused feature.

[0102] S520: The first fusion feature is input into a classification output module of a bolt fault recognition model to extract deep features and output a bolt fault recognition result.

[0103] S510: The target feature is input into the residual processing module of the bolt fault recognition model, and is processed by two residual blocks to obtain a shallow feature. Feature fusion is performed based on the target feature to output a first fused feature.

[0104] The multilayer perceptron is a classic feedforward artificial neural network consisting of an input layer, at least one hidden layer, and an output layer. Its core feature is full connectivity between layers, enabling complex mapping through nonlinear activation functions, thus overcoming the linear limitations of single-layer perceptrons. To improve the stability and robustness of recognition network training, this paper makes the following improvements to the original multilayer perceptron model:

[0105] The purely sequential structure of the multilayer perceptron is prone to information attenuation. Therefore, a residual processing module is added after the input layer of the bolt fault recognition model. The residual processing module consists of a first residual block and a second residual block connected in sequence. Residual processing is performed to obtain shallow features, and feature fusion is performed according to the target features to output the first fused features, namely the residual multilayer perceptron (REMLP).

[0106] Among them, the first residual block includes: a 64-dimensional fully connected layer, a batch normalization layer, a dropout layer with a dropout rate of 0.2, and an addition layer. After the target feature is input, the feature residual processing is first implemented through the first main path of the first residual block to obtain the first residual feature, wherein the first main path is: a 64-dimensional fully connected layer, a batch normalization layer, and a dropout layer with a dropout rate of 0.2 connected in sequence. At the same time, the target feature is transformed in dimension through the first jump path to obtain the first jump feature. In order to ensure that the first residual feature output by the target feature through the main path is consistent with the first jump feature output by the jump path in dimension, and to avoid dimensional conflicts when the traditional residual network is directly added, a 128-dimensional fully connected layer and a batch normalization layer are set in sequence in the first jump path. The obtained first residual feature and the first jump feature are fused through the addition layer to output the fused feature.

[0107] The second residual block consists of a 128-dimensional fully connected layer, a batch normalization layer, a dropout layer with a dropout rate of 0.15, and an addition layer. After the fused features are input into the second residual block, they are first processed through the second main path of the second residual block to obtain the second residual features. The second main path consists of a 128-dimensional fully connected layer, a batch normalization layer, and a dropout layer with a dropout rate of 0.15. Simultaneously, the fused features undergo dimensionality transformation through the second skip path to obtain the second skip features. The obtained second residual features and the second skip features are fused through the addition layer to output the first fused features.

[0108] S520: The first fusion feature is input into a classification output module of a bolt fault recognition model to extract deep features and output a bolt fault recognition result.

[0109] After the first fusion feature is input into the classification output module, it passes through a 256-dimensional fully connected layer, a batch normalization layer, and a dropout layer in sequence, and the dropout rate is further reduced to 0.1 to enhance the model's ability to extract high-order abstract features. It then passes through a 4-dimensional fully connected layer, a Softmax layer, and a classification output layer to output the bolt fault recognition results.

[0110] Preferably, during model training, 80% of the target features are divided into a training set and 20% into a test set. A dynamic learning rate is used, initially at 0.01 and decaying by 50% every five iterations. This allows for rapid convergence in the early stages of training and fine-tuning of parameters in the later stages. An early stopping mechanism is implemented during training, using the cross-entropy loss function. If the cross-entropy loss on the test set does not decrease after multiple iterations, the adaptive optimizer Adam is used to reduce the learning rate. Iterations are stopped when the learning rate drops to a certain value to prevent overfitting.

[0111] Furthermore, the present invention verifies the effect through experimental data. The experimental subjects select 8.8-grade M36 studs and nuts. According to the national standard, the pre-tightening force is 340KN and the pre-tightening torque is 1450Nm. In order to simulate the loosening of the bolts, the bolts are tightened to four different levels of 40MPa, 45MPa, 50MPa, and 55MPa, and the corresponding pre-tightening forces are 247KN, 278KN, 309KN, and 340KN, respectively. The impact on the surface of the bolt generates an acoustic signal; the impact acoustic signal is recorded using an acquisition device. The sampling frequency of all signals in the experiment is set to 25.6kHz, and 200 sets of data are selected for each working condition, for a total of 800 sets of data.

[0112] There is still a lot of noise in the original signal. If the data set is directly constructed to extract features, it will be affected by the noise, which will affect the accuracy of the final diagnosis. First, the original signal is decomposed using SVMD after CPO optimization of the penalty coefficient. After decomposition, IMF components will be generated. Then, the energy entropy and kurtosis of each IMF component are calculated in turn, and the preferred component is selected. Then, SVD denoising is performed on the preferred component. Finally, the denoised preferred component is added to the remaining components to reconstruct a new denoised impact sound signal. The signal spectrum diagram is compared. Figure 2 As shown, the horizontal axis is frequency and the vertical axis is amplitude. It can be observed that the quality of the processed signal is significantly improved and is purer than the original signal.

[0113] The corresponding features were extracted to obtain a data set. 80% of the data set was used as the training set, and 20% was used as the test set. The features were then input into the bolt fault recognition model. The neural network hyperparameters were set as follows: the maximum number of training rounds was 30 rounds, 32 groups of data were randomly divided into a batch for iterative processing in each round of iteration, the initial learning rate was 0.01, and the learning rate was decayed by 50% after every 5 rounds of training. After each round of iteration, the data set was randomly shuffled and retrained. The test results are shown below. Figure 3 As shown, the results are presented in the form of a confusion matrix: the accuracy of the residual multilayer perceptron neural network verification is 98.75%.

[0114] Using the same test data, the traditional MLP neural network algorithm is used for recognition. The confusion matrix of the test set results is as follows Figure 4 As shown, the accuracy is 92.50%. The same test data is used to identify the CNN convolutional neural network algorithm. The convolution kernel is set to one dimension. The confusion matrix of the test set result is as follows Figure 5 As shown, the test set accuracy is 91.25%.

[0115] The prediction results of the MLP, CNN, and REMLP models under the four types of preload conditions are summarized in Table 2. It can be clearly seen that the test set accuracy of REMLP is significantly improved compared with the other two neural network algorithms, proving that the performance of this neural network is superior.

[0116] Table 2 Comparison of classification accuracy of each model

[0117]

[0118] See Figure 6 The ResMLP-based bolt fault identification system 200 includes: a signal sampling module 210, a decomposition and screening module 220, a denoising module 230, a feature extraction module 240, and a detection and output module 250, wherein the signal sampling module 210 is used to obtain a transient impact sound signal; the decomposition and screening module 220 is used to decompose the transient impact sound signal using the SVMD method to obtain a primary screening modal component; the denoising module 230 is used to decompose, filter and reconstruct the primary screening modal component through singular value decomposition to obtain a denoised signal; the feature extraction module 240 is used to extract features based on the denoised signal to obtain target features; the detection and output module 250 is used to input the target features into a bolt fault identification model to obtain a bolt fault identification result; wherein the bolt fault identification model is a pre-trained residual multi-layer perceptron model.

[0119] The specific limitations of the ResMLP-based bolt fault identification system can be found in the limitations of the ResMLP-based bolt fault identification method above and will not be repeated here. Each module in the aforementioned ResMLP-based bolt fault identification system can be implemented in whole or in part through software, hardware, or a combination thereof.

[0120] The above modules may be embedded in or independent of the processor in the computer device in hardware format, or may be stored in the memory of the computer device in software format, so that the processor can call the corresponding operations of the above modules.

[0121] It should be noted that, in order to highlight the innovative part of the present invention, this embodiment does not introduce modules that are not closely related to solving the technical problem proposed by the present invention, but this does not mean that there are no other modules in this embodiment.

[0122] like Figure 7 As shown, the electronic device 3 may include a memory 32, a processor 31 and a bus, and may also include a computer program stored in the memory 32 and executable on the processor 31, such as an energy entropy calculation program.

[0123] The memory 32 includes at least one type of readable storage medium, including a flash memory, a mobile hard disk, a multimedia card, a card-type memory (e.g., SD or DX memory), a magnetic memory, a magnetic disk, an optical disk, etc. In some embodiments, the memory 32 may be an internal storage unit of the electronic device 3, such as a mobile hard disk of the electronic device 3. In other embodiments, the memory 32 may also be an external storage device of the electronic device 3, such as a plug-in mobile hard disk, a smart memory card (SMC), a secure digital (SD) card, a flash card, etc. equipped on the electronic device 3. Furthermore, the memory 32 may include both an internal storage unit and an external storage device of the electronic device 3. The memory 32 may be used not only to store application software installed in the electronic device 3 and various types of data, such as the code for calculating and reconstructing the singular value decomposition of modal components, but also to temporarily store data that has been output or is about to be output.

[0124] In some embodiments, the processor 31 may be composed of an integrated circuit, such as a single packaged integrated circuit, or a plurality of packaged integrated circuits with the same or different functions, including one or more central processing units (CPUs), microprocessors, digital processing chips, graphics processors, and a combination of various control chips. The processor 31 is the control core (Control Unit) of the electronic device 3. It utilizes various interfaces and circuits to connect the various components of the entire electronic device 3. It executes or runs programs or modules stored in the memory 32 (such as a correction program for singular value decomposition calculation and reconstruction of modal components) and calls data stored in the memory 32 to perform various functions of the electronic device 3 and process data.

[0125] The processor 31 executes the operating system and various installed application programs of the electronic device 3. The processor 31 executes the application programs to implement the steps in the above-mentioned bolt fault identification method based on ResMLP.

[0126] Exemplarily, the computer program may be divided into one or more modules, which are stored in the memory 32 and executed by the processor 31 to complete the present application. The one or more modules may be a series of computer program instruction segments capable of completing specific functions, and the instruction segments are used to describe the execution process of the computer program in the electronic device 3.

[0127] The above-mentioned integrated unit implemented in the form of a software functional module can be stored in a computer-readable storage medium, which can be non-volatile or volatile. The above-mentioned software functional module is stored in a storage medium and includes a number of instructions for causing a computer device (which can be a personal computer, computer device, or network device, etc.) or a processor to perform part of the functions of the ResMLP-based bolt fault identification method described in various embodiments of this application.

[0128] In summary, the present invention discloses a bolt fault identification method, system, device, and medium based on ResMLP. These methods utilize the crown porcupine optimization algorithm to adaptively optimize the hyperparameters of successive variational modal decomposition, combine it with singular value decomposition for noise reduction, and perform fault diagnosis based on a bolt fault identification system based on ResMLP. This method effectively solves the problems of signal noise reduction, feature extraction, and high-precision diagnosis in bolt tightness fault diagnosis. This method not only improves the degree of automation in signal processing and reduces manual intervention, but also significantly improves the accuracy of fault diagnosis. Therefore, the present invention effectively overcomes the various shortcomings of the prior art and has a high industrial utilization value.

[0129] The above embodiments are merely illustrative of the principles and effects of the present invention and are not intended to limit the present invention. Anyone skilled in the art may modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by one of ordinary skill in the art without departing from the spirit and technical principles disclosed herein are intended to be covered by the claims of the present invention.

Claims

1. A bolt fault identification method based on ResMLP, characterized in that: The method comprises: Acquiring transient impact sound signals; Decomposing the transient impact sound signal using the SVMD method to obtain the primary screening modal components; Decomposing, filtering and reconstructing the pre-screened modal components through singular value decomposition to obtain a denoised signal; Performing feature extraction based on the denoised signal to obtain target features; The target feature is input into a bolt fault recognition model to obtain a bolt fault recognition result; wherein the bolt fault recognition model is a pre-trained residual multi-layer perceptron model.

2. The bolt fault identification method based on ResMLP according to claim 1 is characterized in that: The step of decomposing the transient impact sound signal using the SVMD method to obtain the primary screening modal components includes: Based on the transient impact sound signal, the optimal penalty coefficient of the SVMD method is obtained using the crested porcupine optimization algorithm; Standardizing the transient impact sound signal to obtain a standard sound signal; Decomposing the standard acoustic signal using the SVMD decomposition method according to the optimal penalty coefficient to obtain multiple modal components, calculating the modal energy corresponding to each modal component, and taking the modal component with the largest modal energy as the principal component mode; Subtract the principal component mode from the standard sound signal to obtain and determine the residual signal: When the residual signal meets the preset threshold condition, all the principal component modes are used as the primary screening modal components; Otherwise, the residual signal is used as the standard sound signal, and the SVMD analysis method is continued to be used for decomposition to determine the principal component mode, and the residual signal is updated until the maximum number of iterations is reached or the residual signal meets the preset threshold condition, and all the principal component modes are used as the initial screening modal components.

3. The bolt fault identification method based on ResMLP according to claim 2 is characterized in that: The step of obtaining the optimal penalty coefficient of the SVMD method based on the transient impact sound signal using the crested porcupine optimization algorithm includes: Randomly generate multiple individuals and regard all individuals as a population; where one individual represents a penalty coefficient; Decomposing the transient impact sound signal based on each penalty coefficient of the population using the SVMD method to obtain corresponding multiple modal components, and calculating the envelope entropy of the modal component corresponding to each penalty coefficient; Optimizing the individuals of the population using the crested porcupine optimization algorithm, and again decomposing the transient impact sound signal and calculating the envelope entropy based on the optimized population, until the number of optimizations reaches a preset maximum number of iterations; A preset number of envelope entropies are selected in order from small to large, and the corresponding individuals are used as the optimal penalty coefficients.

4. The bolt fault identification method based on ResMLP according to claim 1 is characterized in that: The step of performing singular value decomposition, filtering and reconstruction of the pre-screened modal components to obtain a denoised signal comprises: Calculating the energy entropy and kurtosis of each of the pre-screened modal components, and determining the target modal component from the pre-screened modal components based on the energy entropy and kurtosis; Decomposing the target modal component using singular value decomposition to obtain multiple singular values, and truncating the singular values ​​by a preset threshold to obtain high-order screening singular values ​​and low-order contraction singular values; The low-order contracted singular values ​​and the high-order screened singular values ​​are subjected to singular value reconstruction to obtain a denoised signal.

5. The bolt fault identification method based on ResMLP according to claim 4 is characterized in that: The step of calculating the energy entropy and kurtosis of each of the pre-screened modal components and determining the target modal component according to the energy entropy and kurtosis comprises: Calculating the energy entropy and kurtosis of each modal component in the primary screening modal component, and calculating a weighted sum of the energy entropy and the kurtosis; The primary screening modal component whose weighted sum is less than a preset screening threshold is used as the target modal component.

6. The bolt fault identification method based on ResMLP according to claim 5 is characterized in that: The step of decomposing the target modal component using singular value decomposition to obtain multiple singular values, and truncating the singular values ​​by a preset threshold to obtain high-order screening singular values ​​and low-order shrinkage singular values ​​includes: Performing singular value decomposition on the target modal component to obtain multiple singular values; Arrange the singular values ​​in descending order and cut them off according to a preset threshold; The singular values ​​greater than or equal to the preset threshold are processed using a hard threshold method to obtain high-order filtered singular values; The singular values ​​that are smaller than the preset threshold are processed using a soft threshold method to obtain low-order contraction singular values.

7. The bolt fault identification method based on ResMLP according to claim 1 is characterized in that: The process of inputting the target feature into the bolt fault identification model to obtain the bolt fault identification result includes: The target feature is input into the residual processing module of the bolt fault recognition model, and is processed by two residual blocks to obtain shallow features, and feature fusion is performed according to the target feature to output the first fused feature; The first fusion feature is input into the classification output module of the bolt fault recognition model to extract deep features and output the bolt fault recognition result.

8. A bolt fault identification system based on ResMLP, characterized in that: The system comprises: A signal sampling module is used to obtain transient impact sound signals; A decomposition and preliminary screening module is used to decompose the transient impact sound signal using the SVMD method to obtain preliminary screening modal components; A denoising module is used to decompose, filter and reconstruct the pre-screened modal components through singular value decomposition to obtain a denoised signal; A feature extraction module, configured to extract features based on the denoised signal to obtain target features; The detection output module is used to input the target feature into a bolt fault recognition model to obtain a bolt fault recognition result; wherein the bolt fault recognition model is a pre-trained residual multi-layer perceptron model.

9. An electronic device, characterized in that: The electronic device comprises: one or more processors; A storage device for storing one or more programs, which, when executed by the one or more processors, enables the electronic device to implement the bolt fault identification method based on ResMLP as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that A computer program is stored thereon, and when the computer program is executed by a processor of a computer, the computer is caused to execute the bolt fault identification method based on ResMLP according to any one of claims 1 to 7.

Images