A bolt loosening diagnosis method, system, storage medium and device

Abstract

The present application relates to the technical field of data signal processing, and discloses a bolt loosening diagnosis method, system, storage medium and equipment, the method comprising: obtaining the original ultrasonic signal of the bolt; adopting ASSA adaptive sparse self-attention to carry out denoising processing, and obtaining a first denoised signal with high signal-to-noise ratio; adopting EWT empirical wavelet transform to carry out denoising and modal decomposition on the first denoised signal, and constructing a multi-channel time sequence feature tensor for representing the bolt state; inputting the same into a pre-constructed Mamba network, extracting long time sequence hysteresis features caused by bolt loosening, and obtaining a global dynamic feature vector; inputting the global dynamic feature vector into a pre-constructed KAN network, utilizing a learnable univariate activation function, fitting nonlinear judgment conditions of different loosening states of the bolt, and outputting a loosening state recognition result of the bolt; through the method, the missed detection rate and false positive rate of slight loosening defects are reduced, and the real-time reasoning efficiency and detection accuracy of ultrasonic nondestructive testing are improved.

Description

Technical Field

[0001] This invention relates to the field of data signal processing technology, and in particular to a method, system, storage medium, and device for diagnosing bolt loosening. Background Technology

[0002] Bolts are core connecting components in critical infrastructure such as power transmission towers, and early diagnosis of loosening defects is crucial for ensuring structural safety. Ultrasonic non-destructive testing (NDT) technology, due to its sensitivity to changes in internal stress, is widely used in bolt loosening diagnosis. However, in practical engineering applications, existing testing methods still face many technical bottlenecks.

[0003] Traditional bolt loosening detection mainly relies on manual tapping and listening or conventional ultrasonic testing. These methods are highly subjective, difficult to quantify, and unable to effectively identify early, subtle loosening. With the development of signal processing and artificial intelligence technologies, some studies have attempted to combine wavelet transform, empirical mode decomposition, and other algorithms with neural networks for automatic diagnosis. However, existing technologies struggle to capture the subtle time delay characteristics caused by early bolt loosening in complex noise environments. The main reasons include the following: First, the on-site inspection environment is complex. The background broadband thermal noise band in the ultrasonic signal overlaps with the weak defect characteristic band of the bolt. Traditional filtering methods struggle to suppress noise while preserving the characteristic information completely, resulting in a low signal-to-noise ratio. Second, the ultrasonic signal exhibits severe multipath scattering when propagating inside the bolt. Echoes from different paths overlap in both the time and frequency domains, causing mode aliasing. This makes it difficult for traditional time-frequency analysis methods to effectively separate the effective echo components containing loosening information.

[0004] Furthermore, early bolt loosening manifests as minute group time delay variations, which accumulate over long time sequences but with weak amplitudes. Traditional convolutional neural networks, limited by their local receptive fields, struggle to effectively capture such long-term time-lag features; while recurrent neural networks, though capable of processing sequential data, suffer from the vanishing gradient problem and have low computational efficiency. Simultaneously, existing deep learning diagnostic models often separate signal preprocessing from pattern recognition, lacking an end-to-end joint optimization mechanism. Moreover, the decision boundaries of multilayer perceptrons are highly black-boxed, resulting in poor interpretability and failing to meet the stringent reliability requirements of industrial applications.

[0005] Therefore, there is an urgent need for an intelligent diagnostic method that can accurately extract the weak time delay characteristics of early bolt loosening under strong noise interference and has high interpretability, so as to achieve accurate assessment of the bolt status of transmission towers and non-stop maintenance. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide a bolt loosening diagnosis method based on the ASSA-EWT joint denoising and Mamba-KAN fusion network. This method integrates ASSA adaptive sparse self-attention analysis denoising, EWT empirical wavelet transform modal aliasing avoidance mechanism, and Mamba selective state space model with complexity modeling capability, and introduces KAN network to judge bolt loosening.

[0007] On the one hand, the present invention provides a method for diagnosing bolt loosening, comprising the following steps: S1: Acquire the raw ultrasonic signal of the bolt; S2: The original ultrasound signal is denoised using ASSA adaptive sparse self-attention to obtain a first denoised signal with a high signal-to-noise ratio. S3: The first denoised signal is further denoised and modally decomposed using EWT empirical wavelet transform to construct a multi-channel temporal feature tensor for characterizing the bolt state; S4: Input the multi-channel temporal feature tensor into the pre-constructed Mamba network, and use its selective state-space model to extract the long temporal hysteresis features caused by bolt loosening to obtain the global dynamic feature vector; S5: Input the global dynamic feature vector into the pre-constructed KAN network, and use the learnable univariate activation function deployed on the connection weights to fit the nonlinear judgment conditions for different loosening states of the bolt, and finally output the bolt loosening state identification result. Further, in step S2, the denoising process of the original ultrasound signal using ASSA adaptive sparse self-attention includes: S21: Obtain the length as One-dimensional discrete ultrasound raw time-series signal Select window length Let the number of columns of the matrix be... By using a sliding window mechanism, a one-dimensional signal is mapped to a Hankel trajectory matrix. : , in, Representing discrete ultrasound signals in the first... The amplitude of each sampling point, matrix Each column corresponds to a length of Local window sequence; S22: For the Hankel trajectory matrix Perform singular value decomposition to obtain The singular values ​​of rank 1 and their corresponding elementary matrices are represented as follows: , in, Let be a singular value diagonal matrix, satisfying , Let be the rank of the matrix, and each singular value... This reflects the weights of the corresponding components when reconstructing the original trajectory matrix. It is a left singular matrix. This is the transpose of the right singular matrix. Indicates the first A left singular vector Indicates the first Transpose of a right singular vector; S23: Calculate the cumulative information entropy based on the energy contribution rate of each singular value. The calculation formula is as follows: , , in, Indicates the first The weight of each singular value Indicates the preceding The cumulative information entropy of singular values; based on the difference of singular information entropy. The adaptive change determines the optimal rank reduction cutoff order for signal reconstruction, when... When it approaches zero, the corresponding value is... The value is the number of singular values ​​retained. ; S24: The previous ones will be retained The reconstructed matrix after filtering out noise components is obtained by summing the dominant elementary matrices. , represented as: ; S25: Perform a diagonal averaging operation on the reconstructed matrix to restore it to a one-dimensional time-domain signal. The high signal-to-noise ratio first denoised signal is obtained, for the first... One sampling point: , in, The set of all elements that satisfy the diagonal index condition. The number of elements in the set.

[0008] Furthermore, in step S3, the EWT empirical wavelet transform is used to further denoise and decompose the first denoised signal, including: S31: Perform a Fast Fourier Transform on the first denoised signal to obtain its Fourier spectrum. , This represents the angular frequency variable, which is the independent variable in the frequency domain; S32: In the spectrum Local maxima are detected, and the spectrum is adaptively segmented using the midpoint between adjacent maxima as the boundary. A number of non-overlapping frequency bands are represented as follows: ,in, Each frequency band corresponds to a specific oscillation mode in the signal; S33: Based on Meyer wavelet construction theory, for each frequency band... Design a corresponding empirical wavelet filter bank, the filter bank including an empirical scaling function. and empirical wavelet bandpass filter Select appropriate wavelet basis functions to construct the filter; S34: Utilize the filter bank to analyze the spectrum of the first denoised signal. Frequency domain filtering is performed, followed by inverse Fourier transform to extract the non-overlapping intrinsic mode functions (IMFs) corresponding to each frequency band. IMF components Calculated by the following formula: , in, Represents the inverse discrete Fourier transform, for Using empirical scaling functions Alternative ; S35: Filter the intrinsic mode function (IMF) components of the features and the IMF components of the noise, retain the key information of the bolt loosening state, align and splice them along the channel dimension, and construct the multi-channel temporal feature tensor.

[0009] Furthermore, in step S4, the extraction of long-term hysteresis features caused by bolt loosening using its selective state-space model includes: S41: Use the multi-channel temporal feature tensor as input to the Mamba network. Long-term time-series modeling is performed using a discretized state-space model. This discretized state-space model is used to propagate and accumulate the temporal phase delay features caused by bolt loosening across time steps; that is, through high-dimensional hidden states... Mapping to output response Its mapping formula is as follows: , , in, This represents the ultrasonic modal characteristics input at the current moment. express The system's implicit state memory vector at time t. express Dynamic feature output after time step extraction , and These represent the discretized state transition matrix, input matrix, and observation matrix, respectively. S42: Discretize the state-space model into a recursive form using a zero-order hold; S43: Introducing a selective mechanism based on data dependencies, by inputting the matrix... With time step The parameter is converted into a nonlinear function of the input signal to achieve adaptive evaluation of the importance of the ultrasonic wave shape at different time steps; S44: Based on the aforementioned selective state-space model, with linear time complexity The global spatiotemporal dynamic features caused by bolt loosening are extracted to obtain the global dynamic feature vector.

[0010] Furthermore, step S5 specifically includes: S51: Construct a KAN network topology, wherein the KAN network decomposes the high-dimensional decision function into a nested combination of single-variable functions according to the Kolmogorov-Arnold representation theorem; S52: Deploy learnable univariate activation functions on the connection weights of a KAN network; S53: The learnable univariate activation function is parameterized as a linear combination of the residual activation function and a cubic B-spline curve, as shown in the following equation: , in, This indicates that basic nonlinear residual activation support is provided. The core is a cubic B-spline basis function. These are the coordinates of the control points of the spline curve. and These are the weight coefficients for adaptive adjustment by the network; S54: Input the global dynamic feature vector into the KAN network, and dynamically adjust the control points of the B-spline curve through end-to-end joint backpropagation. To fit the discrimination boundary of ultrasound features in high-dimensional space; S55: In the classification layer of the KAN network, the probability distribution of the bolt health status is output based on the fitted discriminant boundary to complete the classification diagnosis.

[0011] Preferably, the method further includes result visualization and storage, specifically including: The model output, including bolt condition identification results, corresponding signal feature vectors, model confidence probability, and timestamps, is automatically stored in a structured database or a data file in a specified format. At the same time, the model version and parameters used for each diagnosis are recorded to ensure data traceability. Based on the stored diagnostic results, a two-dimensional or three-dimensional layout diagram of the bolt group is drawn in the visualization interface. The real-time or historical status of each bolt is intuitively marked in the layout diagram using different colors and number labels. By selecting a specific bolt, the system can plot its original ultrasonic signal, optimized key IMF component waveforms, and time-domain and frequency-domain feature maps during the feature extraction process, providing a detailed analysis view of a single bolt in a visualization interface.

[0012] More preferably, the method further includes trend analysis and alerts: Automatically generate and display the historical trend chart of the status of key bolts for predictive maintenance. When the system detects a loose state or an unknown fault, it automatically triggers an audible and visual alarm and records the event time, location, specific description and suggested measures in a dedicated alarm log list on the interface. It provides a one-click diagnostic report generation function, allowing users to export the current visualization interface, raw data, feature data, and diagnostic report to a specified local path in a common format for easy archiving or reporting later.

[0013] On the other hand, the present invention provides a bolt loosening diagnosis system, comprising: The signal acquisition module is used to acquire the raw ultrasonic signal of the bolt; The ASSA denoising module is used to denoise the original ultrasound signal using ASSA adaptive sparse self-attention to obtain a first denoised signal with a high signal-to-noise ratio. The EWT decomposition module is used to further denoise and decompose the first denoised signal using EWT empirical wavelet transform, and to construct a multi-channel temporal feature tensor for characterizing the bolt state. The Mamba feature extraction module is used to input the multi-channel temporal feature tensor into a pre-constructed Mamba network and use its selective state-space model to extract long-term temporal hysteresis features caused by bolt loosening to obtain a global dynamic feature vector. The KAN recognition module is used to input the global dynamic feature vector into a pre-constructed KAN network, and use the learnable univariate activation function deployed on the connection weights to fit the nonlinear judgment conditions of different loosening states of the bolt, and finally output the bolt loosening state recognition result.

[0014] In addition, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the bolt loosening diagnosis method described in any of the preceding claims.

[0015] Meanwhile, an electronic device is provided, comprising: one or more processors; and a storage device for storing one or more programs, which, when executed by the one or more processors, cause the one or more processors to implement the bolt loosening diagnosis method described above.

[0016] Compared with the prior art, the beneficial effects of the present invention are: At the signal preprocessing level, this invention achieves effective purification of weak loose features in a strong noise background by using ASSA and EWT for joint denoising. ASSA introduces singular information entropy to adaptively determine the order of rank reduction, which suppresses broadband noise while retaining the main echo. EWT adaptively divides the frequency band based on the extreme points of the spectrum to eliminate mode mixing caused by multipath scattering, providing high signal-to-noise ratio input data for subsequent feature extraction. At the temporal feature extraction level, this invention introduces the Mamba selective state space model to achieve efficient capture of weak temporal delay features of early bolt loosening; At the state recognition level, this invention constructs a KAN network to replace the traditional multilayer perceptron, which improves model interpretability while maintaining high accuracy, making it easier for engineers to understand the judgment criteria. In terms of overall architecture, this invention constructs an end-to-end integrated process from raw signals to diagnostic results, avoiding information loss caused by the separation of preprocessing and recognition. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a bolt loosening diagnosis method according to the present invention; Figure 2 This is a schematic diagram of a simulated original ultrasonic noisy signal generated under actual working conditions according to the present invention; Figure 3 This is a schematic diagram of a signal waveform after denoising using the ASSA-EWT joint algorithm according to the present invention; Figure 4 This invention presents a scatter distribution of different samples in a two-dimensional feature space. Figure 5 This invention provides a data detection and import visualization interface; Figure 6 This invention provides a data display visualization interface. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0019] This invention addresses the problems of strong background broadband thermal noise interference and severe multipath scattering in ultrasonic testing of transmission towers. It adaptively decouples multipath scattering interference and reconstructs a clean physical mode with a high signal-to-noise ratio based on the ASSA-EWT joint denoising technique. It introduces the Mamba selective state-space model, utilizing its unique data-dependent selectivity mechanism to achieve linear complexity lossless tracking of the time delay of weak echo groups inside bolts in ultra-long discrete ultrasonic sequences. A global spatiotemporal dynamic feature extraction framework based on Mamba is constructed to extract key features of cumulative hysteresis caused by bolt stiffness attenuation. Finally, an end-to-end diagnostic model is built based on the KAN network. By introducing cubic B-spline curves and learnable activation functions placed at the network edge to enhance the fitting of high-dimensional complex nonlinear discrimination boundaries, it achieves high-precision classification and intelligent identification of early, weak loosening defects in transmission towers that are easily missed.

[0020] The specific embodiments of the present invention will be described below with reference to the accompanying drawings and examples. Example 1

[0021] Please see Figure 1 The technical solution for diagnosing bolt loosening provided in this embodiment includes the following steps: S1: Acquire the raw ultrasonic signal of the bolt; S2: The original ultrasound signal is denoised using ASSA adaptive sparse self-attention to obtain a first denoised signal with a high signal-to-noise ratio. S3: The first denoised signal is further denoised and modally decomposed using EWT empirical wavelet transform to construct a multi-channel temporal feature tensor for characterizing the bolt state; S4: Input the multi-channel temporal feature tensor into the pre-constructed Mamba network, and use its selective state-space model to extract the long temporal hysteresis features caused by bolt loosening to obtain the global dynamic feature vector; S5: Input the global dynamic feature vector into the pre-constructed KAN network, and use the learnable univariate activation function deployed on the connection weights to fit the nonlinear judgment conditions for different loosening states of the bolt, and finally output the bolt loosening state identification result.

[0022] First, receive the raw ultrasound signal from step S1, such as... Figure 2 The image shows the original noisy ultrasonic signal generated under simulated actual working conditions. The time-domain waveform reveals a significant amount of noise interference in the original signal, severely masking the true ultrasonic echo waveform and making it difficult to directly extract effective time delay and amplitude characteristics. Further observation of the Fourier spectrum shows that the noise results in a wide frequency band distribution, with aliasing occurring in the mid-to-high frequency regions.

[0023] Then, the signal undergoes denoising in step S2. To address the overlap between the background noise frequency band and the weak defect feature frequency band of the bolt in the ultrasonic signal, ASSA adaptive sparse self-attention analysis is used for denoising. First, a one-dimensional discrete ultrasonic signal is acquired and mapped to a high-dimensional phase space using the delay embedding theorem, constructing a Hankel trajectory matrix. Then, singular value decomposition is performed on the trajectory matrix, and a singular information entropy estimator is introduced to calculate the energy distribution gradient of the feature space in real time, determining the optimal rank reduction cutoff order. By removing the small singular values ​​corresponding to the background noise, and then reconstructing the signal through diagonal averaging, a high signal-to-noise ratio signal is obtained, providing reliable input data for subsequent frequency band decoupling and feature extraction. Specifically, this includes: S21: Obtain the length as One-dimensional discrete ultrasound raw time-series signal Select window length Let the number of columns of the matrix be... By using a sliding window mechanism, a one-dimensional signal is mapped to a Hankel trajectory matrix. : , in, Representing discrete ultrasound signals in the first... The amplitude of each sampling point, matrix Each column corresponds to a length of The local window sequence, through continuous sliding delay, matrix It fully encompasses the high-dimensional dynamic evolution characteristics of ultrasonic waves in the time domain; S22: For the Hankel trajectory matrix Perform singular value decomposition to obtain The singular values ​​of rank 1 and their corresponding elementary matrices are represented as follows: , in, Let be a singular value diagonal matrix, satisfying , Let be the rank of the matrix, and each singular value... This reflects the weights of the corresponding components when reconstructing the original trajectory matrix. It is a left singular matrix. This is the transpose of the right singular matrix. Indicates the first A left singular vector Indicates the first Transpose of a right singular vector; S23: Calculate the cumulative information entropy based on the energy contribution rate of each singular value. The calculation formula is as follows: , , in, Indicates the first The weight of each singular value Indicates the preceding The cumulative information entropy of singular values; based on the difference of singular information entropy. The adaptive change determines the optimal rank reduction cutoff order for signal reconstruction, when... When it approaches zero, the corresponding value is... The value is the number of singular values ​​retained. ,in, The energy percentage of a single characteristic component was quantified. The algorithm uses the concept of entropy from information theory to measure the degree of disorder in a signal. It continuously calculates the entropy difference, and when the difference value decays to a minimum threshold, it determines that the structured features characterizing the physical main echo have been fully extracted, thus dynamically determining the optimal truncation order. .

[0024] S24: The previous ones will be retained The reconstructed matrix after filtering out noise components is obtained by summing the dominant elementary matrices. , represented as: ; S25: Perform a diagonal averaging operation on the reconstructed matrix to restore it to a one-dimensional time-domain signal. The high signal-to-noise ratio first denoised signal is obtained, for the first... One sampling point: , in, The set of all elements that satisfy the diagonal index condition. The number of elements in the set.

[0025] Next, to address the multipath scattering interference and its resulting mode aliasing problem still present in the signal after ASSA processing, the EWT empirical wavelet transform from step S2 is used for further processing. First, a Fourier transform is performed on the ASSA output signal to obtain its spectral distribution. Local maxima are detected in the spectrum, and these extrema are used as boundaries for adaptive frequency band division. A bandpass filter bank is constructed based on Meyer wavelet theory to decompose the original signal into several eigenmode functions without frequency band overlap, each component corresponding to the echo components of different propagation paths. Effective mode components containing the main echo and secondary reflections are selected, and time-axis alignment and splicing are performed along the channel dimension to construct a multi-channel temporal feature tensor. This processing effectively suppresses multipath scattering interference, providing reliable input features for subsequent state recognition. Specifically, this includes: S31: For the first denoised signal Perform a Fast Fourier Transform to obtain its Fourier spectrum. , represented as: , in, Let be the angular frequency variable, and be the independent variable in the frequency domain. The imaginary unit, For time variables, that is, the time-domain independent variables of the signal, in the Fourier transform, through... Introducing phase rotation allows the integral to extract the frequency-related information from the signal. The ingredients; S32: In the spectrum Local maxima are detected, and the spectrum is adaptively segmented using the midpoint between adjacent maxima as the boundary. A number of non-overlapping frequency bands are represented as follows: ,in, Each frequency band corresponds to a specific oscillation mode in the signal; S33: Based on Meyer wavelet construction theory, for each frequency band... Design a corresponding empirical wavelet filter bank, the filter bank including an empirical scaling function. and empirical wavelet bandpass filter Select appropriate wavelet basis functions to construct the filter; S34: Utilize the filter bank to analyze the spectrum of the first denoised signal. Frequency domain filtering is performed, followed by inverse Fourier transform to extract the non-overlapping intrinsic mode functions (IMFs) corresponding to each frequency band. IMF components Calculated by the following formula: , in, Represents the inverse discrete Fourier transform, for Using empirical scaling functions Alternative This formula represents the ultrasonic spectrum. Constructed bandpass filter Perform an inner product in the frequency domain, then inversely transform back to the time domain to obtain the strictly isolated... Narrowband physical echo components within the frequency band ; Low-frequency trend term of the signal Extracted from the scaling function: , Broadband residual noise that does not conform to the main echo energy distribution is removed, and the purified signal is reconstructed from each mode, as shown in the following equation: ; S35: Selecting the Feature Intrinsic Mode Function (IMF) components and the Noise Intrinsic Mode Function (IMF) components, retaining key information about the bolt loosening state, aligning and stitching them along the channel dimension to construct the multi-channel temporal feature tensor. In this embodiment, we extract the preceding... One IMF component containing the core main echo and secondary reflection waves, a single time-domain waveform input, and a high-dimensional modal feature tensor output. It provides data with a high signal-to-noise ratio.

[0026] like Figure 3 The image shows the signal waveform after denoising using the ASSA-EWT joint algorithm. The denoised signal restores the phase and characteristics of the main ultrasonic echo, and the overall waveform becomes smoother. This demonstrates that the algorithm achieves effective denoising.

[0027] Next, addressing the issue that early-stage weak loosening of bolts manifests as small group time delay variations, and that traditional convolutional neural networks struggle to effectively capture such long-term time lag features, the multi-channel modal data obtained through ASSA-EWT processing is... As input to the Mamba module ,in The number of valid IMF components selected. This refers to the number of time sampling points. Mamba, based on a selective state-space model, discretizes the continuous state equations using a zero-order hold and introduces a data-dependent step-size adjustment mechanism. This mechanism enables effective tracking of long-term temporal hysteresis characteristics caused by bolt loosening. Specifically, it includes: S41: Use the multi-channel temporal feature tensor as input to the Mamba network. Long-term time-series modeling is performed using a discretized state-space model. This discretized state-space model is used to propagate and accumulate the temporal phase delay features caused by bolt loosening across time steps; that is, through high-dimensional hidden states... Mapping to output response Its mapping formula is as follows: , , in, This represents the ultrasonic modal characteristics input at the current moment. express The system's implicit state memory vector at time t. express Dynamic feature output after time step extraction , and These represent the discretized state transition matrix, input matrix, and observation matrix, respectively. S42: Discretize the state-space model into a recursive form using a zero-order hold; S43: Introducing a selective mechanism based on data dependencies, by inputting the matrix... With time step The parameter is converted into a nonlinear function of the input signal to achieve adaptive evaluation of the importance of the ultrasonic wave shape at different time steps; S44: Based on the aforementioned selective state-space model, with linear time complexity The global spatiotemporal dynamic features caused by bolt loosening are extracted to obtain the global dynamic feature vector.

[0028] Next, addressing the issue of highly black-box decision boundaries in traditional multilayer perceptrons, the global dynamic feature vector output from Mamba is fed into the KAN decision head, as in step S5. Based on the Kolmogorov-Arnold representation theorem, learnable univariate activation functions are deployed on the network's connection weights. A combination of residual activation and cubic B-spline curves is used to fit nonlinear judgment conditions of normal-slightly loosened-severely loosened states in the high-dimensional feature space. Finally, spline control points are adjusted through joint backpropagation to output a three-state probability distribution vector. Specifically, this includes: S51: Construct a KAN network topology, wherein the KAN network decomposes the high-dimensional decision function into a nested combination of single-variable functions according to the Kolmogorov-Arnold representation theorem; S52: Deploy learnable univariate activation functions on the connection weights of a KAN network; S53: The learnable univariate activation function is parameterized as a linear combination of the residual activation function and a cubic B-spline curve, as shown in the following equation: , in, This indicates that basic nonlinear residual activation support is provided. The core is a cubic B-spline basis function. These are the coordinates of the control points of the spline curve. and These are the weight coefficients for adaptive adjustment by the network; S54: Input the global dynamic feature vector into the KAN network, and dynamically adjust the control points of the B-spline curve through end-to-end joint backpropagation. To fit the discrimination boundary of ultrasound features in high-dimensional space; S55: In the classification layer of the KAN network, the probability distribution of the bolt health status is output based on the fitted discriminant boundary to complete the classification diagnosis.

[0029] In this embodiment, we input the features along the time dimension into a Mamba network and use a state-space model for long-term time series modeling. The network adaptively captures the weak group delay features caused by bolt loosening in ultrasonic echoes through data-dependent dynamic parameter adjustments, outputting a hidden state sequence containing global time series information. Then, global average pooling is performed on the hidden state sequence output by the Mamba network to obtain a fixed-dimensional global feature vector. This vector contains all the temporal information related to the bolt loosening state in the ultrasonic signal; the global feature vector The input is processed by a KAN network, which performs a non-linear mapping using a learnable activation function. KAN utilizes the local fitting ability of B-spline curves to construct classification boundaries in the feature space, mapping feature vectors to probabilities for each category. Finally, the KAN output is converted into a probability distribution of three health states through a Softmax function: normal, slightly loose, and severely loose. The category corresponding to the highest probability is selected as the final diagnosis.

[0030] like Figure 4 As shown, the scatter distribution of different samples in the two-dimensional feature space reveals that the wave mode features exhibit obvious clustering characteristics for bolt loosening at different locations, indicating that the wave mode features contain information about bolt loosening.

[0031] Furthermore, the method of the present invention also includes result visualization and storage, specifically including: The model output, including bolt condition identification results (including tightness, looseness level, unknown fault markers, etc.), corresponding signal feature vectors, model confidence probability, and timestamps, is automatically stored in a structured database or a data file in a specified format. At the same time, the model version and parameters used for each diagnosis are recorded to ensure data traceability. Based on the stored diagnostic results, draw a two-dimensional or three-dimensional layout diagram of the bolt group in a visualization interface (such as PYTHON GUI), and intuitively mark the real-time or historical status of each bolt in the layout diagram with different colors and number labels. By selecting a specific bolt, the system can plot its original ultrasonic signal, optimized key IMF component waveforms, and time-domain and frequency-domain feature maps during the feature extraction process, providing a detailed analysis view of a single bolt in a visualization interface.

[0032] The method also includes trend analysis and alerts: Automatically generate and display the historical trend chart of the status of key bolts for predictive maintenance. When the system detects a loose state or an unknown fault, it automatically triggers an audible and visual alarm and records the event time, location, specific description and suggested measures in a dedicated alarm log list on the interface. It provides a one-click diagnostic report generation function, allowing users to export the current visualization interface, raw data, feature data, and diagnostic report to a specified local path in a common format for easy archiving or reporting later.

[0033] In this embodiment, the development platform used is PYTHON GUI, a graphical user interface design tool. This software can display images of bolt loosening and analyze the bolt loosening state through image recognition algorithms. The interface design diagram is shown below. Figure 5 This is the interface for data detection and import. Figure 6 This is the data display interface.

[0034] In summary, the bolt loosening diagnosis method based on the ASSA-EWT joint denoising and Mamba-KAN fusion network of this invention features an end-to-end algorithm design with strong long-range temporal feature tracking capabilities. It can overcome strong noise interference caused by complex outdoor electromagnetic environments and multipath reflections at the underlying signal level. In the complex working conditions of transmission towers, it effectively distinguishes background clutter from early relaxation features using a selective state-space mechanism. This method facilitates the timely detection of problems such as early preload decay of tower bolts, significantly reducing the missed detection rate and false alarm rate of minor loosening defects, and improving the real-time inference efficiency and detection accuracy of ultrasonic non-destructive testing. It is of great significance for ensuring the long-term safe and stable operation of core power load-bearing components and transmission lines.

[0035] Based on the above method, the present invention provides a bolt loosening diagnosis system, comprising: The signal acquisition module is used to acquire the raw ultrasonic signal of the bolt; The ASSA denoising module is used to denoise the original ultrasound signal using ASSA adaptive sparse self-attention to obtain a first denoised signal with a high signal-to-noise ratio. The EWT decomposition module is used to further denoise and decompose the first denoised signal using EWT empirical wavelet transform, and to construct a multi-channel temporal feature tensor for characterizing the bolt state. The Mamba feature extraction module is used to input the multi-channel temporal feature tensor into a pre-constructed Mamba network and use its selective state-space model to extract long-term temporal hysteresis features caused by bolt loosening to obtain a global dynamic feature vector. The KAN recognition module is used to input the global dynamic feature vector into a pre-constructed KAN network, and use the learnable univariate activation function deployed on the connection weights to fit the nonlinear judgment conditions of different loosening states of the bolt, and finally output the bolt loosening state recognition result.

[0036] It should be noted that the steps in the bolt loosening diagnosis method provided in this embodiment can be implemented based on the corresponding modules in the bolt loosening diagnosis system. Those skilled in the art can refer to the technical solution of the system to implement the steps of the method. That is, the embodiment in the system can be understood as a preferred example of implementing the method, and will not be elaborated here.

[0037] In addition to the methods and systems described above, this embodiment also provides a computer device applicable to the DIA algebraic simplification quality assessment method for improving space target positioning data, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the DIA algebraic simplification quality assessment method for improving space target positioning data as proposed in the above embodiments.

[0038] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0039] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the DIA algebraic simplification quality assessment method for spatial target positioning data as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0040] Besides implementing the system and its various devices provided by this invention in purely computer-readable program code, the same functions can be achieved by logically programming the method steps, making the system and its various devices of this invention appear as logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, the system and its various devices provided by this invention can be considered as a hardware component, and the devices included therein for implementing various functions can also be considered as structures within the hardware component; alternatively, the devices for implementing various functions can be considered as both software modules implementing the method and structures within the hardware component.

[0041] Finally, it should be noted that the above description is only a preferred embodiment of the present invention, and the scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be pointed out that for those skilled in the art, any improvements and modifications made without departing from the principle of the present invention should also be considered within the scope of protection of the present invention.

[0042] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A method for diagnosing bolt loosening, characterized in that, Includes the following steps: S1: Acquire the raw ultrasonic signal of the bolt; S2: The original ultrasound signal is denoised using ASSA adaptive sparse self-attention to obtain a first denoised signal with a high signal-to-noise ratio. S3: The first denoised signal is further denoised and modally decomposed using EWT empirical wavelet transform to construct a multi-channel temporal feature tensor for characterizing the bolt state; S4: Input the multi-channel temporal feature tensor into the pre-constructed Mamba network, and use its selective state-space model to extract the long temporal hysteresis features caused by bolt loosening to obtain the global dynamic feature vector; S5: Input the global dynamic feature vector into the pre-constructed KAN network, and use the learnable univariate activation function deployed on the connection weights to fit the nonlinear judgment conditions for different loosening states of the bolt, and finally output the bolt loosening state identification result.

2. The bolt loosening diagnosis method according to claim 1, characterized in that, In step S2, the denoising process of the original ultrasound signal using ASSA adaptive sparse self-attention includes: S21: Obtain the length as One-dimensional discrete ultrasound raw time-series signal Select window length Let the number of columns of the matrix be... By using a sliding window mechanism, a one-dimensional signal is mapped to a Hankel trajectory matrix. : ， in, Representing discrete ultrasound signals in the first... The amplitude of each sampling point, matrix Each column corresponds to a length of Local window sequence; S22: For the Hankel trajectory matrix Perform singular value decomposition to obtain The singular values ​​of rank 1 and their corresponding elementary matrices are represented as follows: ， in, Let be a singular value diagonal matrix, satisfying , Let be the rank of the matrix, and each singular value... This reflects the weights of the corresponding components when reconstructing the original trajectory matrix. It is a left singular matrix. This is the transpose of the right singular matrix. Indicates the first A left singular vector, Indicates the first Transpose of a right singular vector; S23: Calculate the cumulative information entropy based on the energy contribution rate of each singular value. The calculation formula is as follows: ， ， in, Indicates the first The weight of each singular value Indicates the preceding The cumulative information entropy of each singular value; based on the difference of singular information entropy. The adaptive change determines the optimal rank reduction cutoff order for signal reconstruction, when... When it approaches zero, the corresponding value is... The value is the number of singular values ​​retained. ; S24: The previous ones will be retained The reconstructed matrix after filtering out noise components is obtained by summing the dominant elementary matrices. , represented as: ； S25: Perform a diagonal averaging operation on the reconstructed matrix to restore it to a one-dimensional time-domain signal. The high signal-to-noise ratio first denoised signal is obtained, for the first... One sampling point: ， in, The set of all elements that satisfy the diagonal index condition. The number of elements in the set.

3. The bolt loosening diagnosis method according to claim 1, characterized in that, In step S3, the EWT empirical wavelet transform is used to further denoise and decompose the first denoised signal, including: S31: Perform a Fast Fourier Transform on the first denoised signal to obtain its Fourier spectrum. , This represents the angular frequency variable, which is the independent variable in the frequency domain; S32: In the spectrum Local maxima are detected, and the spectrum is adaptively segmented using the midpoint between adjacent maxima as the boundary. A number of non-overlapping frequency bands are represented as follows: ,in, Each frequency band corresponds to a specific oscillation mode in the signal; S33: Based on Meyer wavelet construction theory, for each frequency band... Design a corresponding empirical wavelet filter bank, the filter bank including an empirical scaling function. and empirical wavelet bandpass filter Select appropriate wavelet basis functions to construct the filter; S34: Utilize the filter bank to analyze the spectrum of the first denoised signal. Frequency domain filtering is performed, followed by inverse Fourier transform to extract the non-overlapping intrinsic mode functions (IMFs) corresponding to each frequency band. IMF components Calculated by the following formula: ， in, Represents the inverse discrete Fourier transform, for Using empirical scaling functions Alternative ; S35: Filter the intrinsic mode function (IMF) components of the features and the IMF components of the noise, retain the key information of the bolt loosening state, align and splice them along the channel dimension, and construct the multi-channel temporal feature tensor.

4. The bolt loosening diagnosis method according to claim 1, characterized in that, In step S4, the long-term hysteresis features caused by bolt loosening are extracted using its selective state-space model, including: S41: Use the multi-channel temporal feature tensor as input to the Mamba network. Long-term time-series modeling is performed using a discretized state-space model. This discretized state-space model is used to propagate and accumulate the temporal phase delay features caused by bolt loosening across time steps; that is, through high-dimensional hidden states... Mapping to output response Its mapping formula is as follows: ， ， in, This represents the ultrasonic modal characteristics input at the current moment. express The system's implicit state memory vector at time t. express Dynamic feature output after time step extraction , and These represent the discretized state transition matrix, input matrix, and observation matrix, respectively. S42: Discretize the state-space model into a recursive form using a zero-order hold; S43: Introducing a selective mechanism based on data dependencies, by inputting the matrix... With time step The parameter is converted into a nonlinear function of the input signal to achieve adaptive evaluation of the importance of the ultrasonic wave shape at different time steps; S44: Based on the aforementioned selective state-space model, with linear time complexity The global spatiotemporal dynamic features caused by bolt loosening are extracted to obtain the global dynamic feature vector.

5. The bolt loosening diagnosis method according to claim 1, characterized in that, Step S5 specifically includes: S51: Construct a KAN network topology, wherein the KAN network decomposes the high-dimensional decision function into a nested combination of single-variable functions according to the Kolmogorov-Arnold representation theorem; S52: Deploy learnable univariate activation functions on the connection weights of a KAN network; S53: The learnable univariate activation function is parameterized as a linear combination of the residual activation function and a cubic B-spline curve, as shown in the following equation: ， in, This indicates that basic nonlinear residual activation support is provided. The core is a cubic B-spline basis function. These are the coordinates of the control points of the spline curve. and These are the weight coefficients for adaptive adjustment by the network; S54: Input the global dynamic feature vector into the KAN network, and dynamically adjust the control points of the B-spline curve through end-to-end joint backpropagation. To fit the discrimination boundary of ultrasound features in high-dimensional space; S55: In the classification layer of the KAN network, the probability distribution of the bolt health status is output based on the fitted discriminant boundary to complete the classification diagnosis.

6. The bolt loosening diagnosis method according to claim 1, characterized in that, The method also includes result visualization and storage, specifically including: The model output, including bolt condition identification results, corresponding signal feature vectors, model confidence probability, and timestamps, is automatically stored in a structured database or a data file in a specified format. At the same time, the model version and parameters used for each diagnosis are recorded to ensure data traceability. Based on the stored diagnostic results, a two-dimensional or three-dimensional layout diagram of the bolt group is drawn in the visualization interface. The real-time or historical status of each bolt is intuitively marked in the layout diagram using different colors and number labels. By selecting a specific bolt, the system can plot its original ultrasonic signal, optimized key IMF component waveforms, and time-domain and frequency-domain feature maps during the feature extraction process, providing a detailed analysis view of a single bolt in a visualization interface.

7. The bolt loosening diagnosis method according to claim 6, characterized in that, The method also includes trend analysis and alerts: Automatically generate and display the historical trend chart of the status of key bolts for predictive maintenance. When the system detects a loose state or an unknown fault, it automatically triggers an audible and visual alarm and records the event time, location, specific description and suggested measures in a dedicated alarm log list on the interface. It provides a one-click diagnostic report generation function, allowing users to export the current visualization interface, raw data, feature data, and diagnostic report to a specified local path in a common format for easy archiving or reporting later.

8. A bolt loosening diagnostic system, characterized in that, include: The signal acquisition module is used to acquire the raw ultrasonic signal of the bolt; The ASSA denoising module is used to denoise the original ultrasound signal using ASSA adaptive sparse self-attention to obtain a first denoised signal with a high signal-to-noise ratio. The EWT decomposition module is used to further denoise and decompose the first denoised signal using EWT empirical wavelet transform, and to construct a multi-channel temporal feature tensor for characterizing the bolt state. The Mamba feature extraction module is used to input the multi-channel temporal feature tensor into a pre-constructed Mamba network and use its selective state-space model to extract long-term temporal hysteresis features caused by bolt loosening to obtain a global dynamic feature vector. The KAN recognition module is used to input the global dynamic feature vector into a pre-constructed KAN network, and use the learnable univariate activation function deployed on the connection weights to fit the nonlinear judgment conditions of different loosening states of the bolt, and finally output the bolt loosening state recognition result.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the bolt loosening diagnosis method as described in any one of claims 1-7.

10. An electronic device, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the bolt loosening diagnosis method as described in any one of claims 1-7.

Images