Power transmission tower bolt state diagnosis method

Abstract

This invention discloses a method for diagnosing the condition of bolts in power transmission towers, relating to the field of structural health monitoring technology. The method includes the following steps: transmitting a swept-frequency acoustic wave to the bolt under test and acquiring the reflected acoustic signal, separating the linear fundamental frequency response component and the nonlinear harmonic response component; performing time-frequency domain transformation on the linear fundamental frequency response component and the nonlinear harmonic response component to generate a multidimensional time-frequency feature matrix to be tested; calculating the current reconstruction error based on the multidimensional time-frequency feature matrix to be tested, comparing the error with a warning boundary to determine the loosening state; the warning boundary is determined by calculating the reconstruction offset vector using the multidimensional time-frequency feature matrix under the loosening state, and directionally correcting the initial boundary; the initial boundary is generated by calculating the reconstruction error of the multidimensional time-frequency feature matrix under the healthy state. This invention requires only a very small number of loosening samples to achieve highly sensitive and robust non-contact online diagnosis, suitable for early loosening detection of bolts in power transmission towers.

Description

Technical Field

[0001] This invention relates to the field of structural health monitoring technology, and more specifically, to a method for diagnosing the condition of bolts on power transmission towers. Background Technology

[0002] The tightness of transmission tower bolts directly affects the safety and stability of power grid operation. Under long-term alternating loads such as wind loads and icing, early, hidden loosening caused by the decay of bolt preload can lead to structural failure or even tower collapse if not detected in time. Therefore, early non-contact diagnosis of bolt loosening is of significant engineering importance.

[0003] Current acoustic detection methods typically process the acquired mixed audio signals as a whole before extracting features for diagnostic judgment. However, this holistic processing approach has the following technical drawbacks: Firstly, the acoustic response distortion energy caused by early slight loosening is extremely low, and it is easily submerged by the structural fundamental frequency vibration and outdoor broadband background noise in the mixed signal, resulting in insufficient signal-to-noise ratio for feature extraction and low sensitivity for detecting concealed loosening.

[0004] Secondly, in real power grid operation and maintenance scenarios, samples of normal, secure states can be collected in massive quantities, while samples of actual loosening faults are extremely scarce. Existing methods, when establishing diagnostic thresholds or classification boundaries, either rely on large-scale fault label data for training, or can only establish symmetrical fixed boundaries when only healthy samples are available. It is difficult to utilize the fault offset direction information carried by the extremely small number of loose samples to specifically optimize the judgment boundaries, resulting in limited diagnostic accuracy under conditions of extremely imbalanced data categories.

[0005] Therefore, there is an urgent need for a method that can target and separate weak loosening features from mixed audio signals and utilize their directional information to optimize the diagnostic boundary under conditions where loosening samples are extremely scarce. Summary of the Invention

[0006] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a method for diagnosing the condition of transmission tower bolts. This method accurately extracts nonlinear weak fault features through alternating dual-phase acoustic excitation and time-domain targeted decoupling. It also constructs an asymmetric early warning boundary using an anisotropic topological stretching technique guided by a reconstructed offset vector calculated from a multidimensional time-frequency feature matrix under a small number of loose states. This addresses the problems in existing acoustic detection methods where early slight loosening features are easily submerged by structural fundamental frequency and environmental noise, resulting in low detection sensitivity. Furthermore, it addresses the difficulty in optimizing and fixing the judgment boundary using fault direction information under conditions of extremely unbalanced data categories, leading to limited diagnostic accuracy.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A method for diagnosing the condition of transmission tower bolts includes the following steps: transmitting a swept-frequency acoustic wave to the bolt under test and acquiring the reflected acoustic signal, separating the linear fundamental frequency response component and the nonlinear harmonic response component; performing time-frequency domain transformation on the linear fundamental frequency response component and the nonlinear harmonic response component to generate a multidimensional time-frequency feature matrix to be tested; calculating the current reconstruction error based on the multidimensional time-frequency feature matrix to be tested, comparing the error with the warning boundary, and determining the loosening state; the warning boundary is determined by calculating the reconstruction offset vector using the multidimensional time-frequency feature matrix under the loosening state, and directionally correcting the initial boundary; the initial boundary is generated by calculating the reconstruction error of the multidimensional time-frequency feature matrix under the healthy state.

[0008] In a preferred embodiment, the step of transmitting a swept frequency sound wave to the bolt under test and acquiring the reflected sound frequency signal includes: setting the frequency band of the swept frequency sound wave to cover the reference natural frequency and its lower offset frequency band of the bolt under test under standard preload; sequentially transmitting a first swept frequency sound wave signal and a second swept frequency sound wave signal with the same frequency band and an initial phase difference of 180 degrees to the bolt under test, and simultaneously acquiring the first reflected sound frequency signal and the second reflected sound frequency signal generated by the bolt under test after being excited.

[0009] In a preferred embodiment, separating the linear fundamental frequency response component and the nonlinear harmonic response component includes: performing time-domain difference and time-domain summation operations on the first reflected audio signal and the second reflected audio signal, respectively, to separate the linear fundamental frequency response component and the nonlinear harmonic response component.

[0010] In a preferred embodiment, the step of performing time-frequency domain transformation on the linear fundamental frequency response component and the nonlinear harmonic response component to generate the multidimensional time-frequency feature matrix to be measured includes: performing continuous wavelet transform on the linear fundamental frequency response component and the nonlinear harmonic response component respectively to generate a two-dimensional linear time-frequency matrix and a two-dimensional nonlinear time-frequency matrix; and concatenating the two-dimensional linear time-frequency matrix and the two-dimensional nonlinear time-frequency matrix along the feature channel dimension to obtain the multidimensional time-frequency feature matrix to be measured.

[0011] In a preferred embodiment, generating the two-dimensional linear time-frequency matrix and the two-dimensional nonlinear time-frequency matrix includes: performing continuous wavelet transforms on the linear fundamental frequency response component and the nonlinear harmonic response component using first and second wavelet scale sequences, respectively, to generate corresponding first initial time-frequency matrices and second initial time-frequency matrices; establishing a target mapping grid with a unified time-scale dimension; mapping the first initial time-frequency matrix to the grid via linear interpolation; and mapping the second initial time-frequency matrix to the grid after extracting the local maximum modulus value.

[0012] In a preferred embodiment, the step of obtaining the initial boundary includes: performing compression dimensionality reduction and inverse reconstruction operations on the multidimensional time-frequency feature matrix in a healthy state with the constraint of minimizing the reconstruction error, calculating the baseline reconstruction error, and obtaining a baseline reconstruction error set; performing probability density estimation on the baseline reconstruction error set to obtain a probability density distribution; extracting the corresponding error quantile values ​​according to the probability density distribution and the preset one-sided confidence level, and establishing the error quantile values ​​as the initial boundary.

[0013] In a preferred embodiment, the step of calculating the reconstructed offset vector using the multidimensional time-frequency feature matrix in the loose state includes: performing the compression dimensionality reduction and inverse reconstruction operation on the multidimensional time-frequency feature matrix in the loose state, calculating the corresponding abnormal reconstruction error; calculating the out-of-bounds deviation value of each abnormal reconstruction error exceeding the initial boundary, and performing vector mean operation on the out-of-bounds deviation value to obtain the reconstructed offset vector.

[0014] In a preferred embodiment, the directional correction of the initial boundary includes: performing anisotropic topological stretching on the initial boundary along the vector direction of the reconstructed offset vector; and establishing the asymmetric closed boundary formed after stretching as the warning boundary.

[0015] In a preferred embodiment, performing anisotropic topological stretching on the initial boundary along the vector direction of the reconstruction offset vector includes: orthogonally decomposing the initial boundary into a first boundary component parallel to the reconstruction offset vector and a second boundary component orthogonal to the reconstruction offset vector; performing a shrinking mapping on the first boundary component and a maintaining or expanding mapping on the second boundary component; and vector superimposing the mapped first boundary component and the second boundary component to obtain an asymmetric closed boundary.

[0016] In a preferred embodiment, the step of calculating the current reconstruction error based on the multidimensional time-frequency feature matrix to be tested includes: performing compression dimensionality reduction and inverse reconstruction operations on the multidimensional time-frequency feature matrix to be tested to obtain the corresponding current reconstruction matrix; and calculating the feature difference value between the multidimensional time-frequency feature matrix to be tested and the current reconstruction matrix as the current reconstruction error.

[0017] This invention emits sweeping acoustic waves at the bolt under test and forcibly separates the reflected acoustic signal into a linear fundamental frequency response component and a nonlinear harmonic response component. This removes the dominant background vibration and extracts the nonlinear component containing information about bolt loosening interface friction and impact, converting it to the time-frequency domain. This significantly improves the ability to capture features and the signal-to-noise ratio of concealed early loosening. Furthermore, after establishing an initial boundary using massive amounts of normal historical data, this invention requires only a very small number of known loosening samples to calculate a high-dimensional reconstructed offset vector, which is then used to directionally correct the initial boundary and establish an early warning boundary. This mechanism transforms loosening samples from the large-scale training labels required in traditional methods into guiding signals for boundary direction correction, solving the long-tail problem of extremely imbalanced data categories. This allows the diagnostic model to establish accurate and interference-resistant early warning boundaries even in real-world conditions lacking large-scale fault samples, significantly improving its generalization capability for industrial applications. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the method for diagnosing the condition of transmission tower bolts according to the present invention. Figure 2 This is a schematic diagram illustrating the diagnostic effect of a conventional symmetrical fixed boundary provided in an embodiment of the present invention; Figure 3 A schematic diagram illustrating the diagnostic effect of the asymmetric early warning boundary provided in an embodiment of the present invention; Figure 4 The first reflected audio signal time-domain waveform diagram provided in the embodiment of the present invention; Figure 5 The second reflected audio signal time-domain waveform diagram provided in the embodiment of the present invention; Figure 6 The waveform diagram of the nonlinear harmonic response component output by time-domain summation is provided for an embodiment of the present invention; Figure 7 The waveform diagram of the linear fundamental frequency response component of the time-domain differential output provided in the embodiment of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0020] Example 1, Figure 1 The present invention provides a method for diagnosing the condition of bolts on power transmission towers, comprising the following steps: S1. It should be noted that, in this embodiment, the method for obtaining the multidimensional time-frequency feature matrix under the healthy state is as follows: For sample bolts whose physical state has been verified by a torque wrench or ultrasonic force measuring instrument to be properly tightened, that is, whose preload is within the range of 90% to 110% of the design standard value, under different working conditions such as temperature, wind speed, and tower location, steps S2 to S3 are repeated to generate and store the multidimensional time-frequency feature matrix, thus forming the set of multidimensional time-frequency feature matrices under the healthy state.

[0021] The method for obtaining the multidimensional time-frequency feature matrix under the loose state is as follows: For sample bolts whose preload is artificially reduced to 30%~70% of the standard value and which are verified to be in a known loose state, repeat steps S2 to S3 to generate and store a small set of multidimensional time-frequency feature matrices under the loose state.

[0022] To verify the long-tail distribution characteristics of massive healthy samples and extremely scarce loose samples in real industrial scenarios, and to establish the offline training sample benchmark for this embodiment, the inventors conducted offline acquisition of audio signals in a real power transmission tower test environment. The test covered two typical tower types: a 220kV straight-line goblet tower and a 500kV tension tower. The specific physical parameters of the experimental towers and the distribution of the multi-dimensional time-frequency feature matrix samples of the acquired healthy and loose states are shown in Table 1.

[0023] Table 1. Long-tail distribution information of experimental tower parameters and historical multidimensional feature samples.

[0024] As shown in Table 1, the sample size of the multidimensional time-frequency feature matrix under healthy conditions is several thousand or even tens of thousands of sets, while the sample size of the multidimensional time-frequency feature matrix under loose conditions is only tens to hundreds of sets. The two show an order of magnitude difference, which truly reflects the long-tail distribution characteristics of the industrial field, where there are massive normal samples and extremely scarce fault samples.

[0025] S2, emit sweeping sound waves to the bolt under test and collect the reflected sound signals to separate the linear fundamental frequency response component and the nonlinear harmonic response component; In this embodiment, the step of emitting a swept-frequency sound wave to the bolt under test and acquiring the reflected sound frequency signal includes: S21. Based on the structural parameters of the bolt under standard preload, its reference natural frequency is obtained through finite element modal analysis or hammer impact method. The lower limit of the swept frequency band is determined to cover at least the lower offset frequency band offset downwards by 10% to 20% from the reference natural frequency. An original swept frequency signal is generated using this frequency band. This original swept frequency signal is then divided into two paths, one of which is used as the first swept frequency acoustic signal without phase shifting. Another path introduces a 180-degree phase delay through a phase shifter to generate a second swept-frequency acoustic signal with the same frequency band as the first swept-frequency acoustic signal but a starting phase difference of 180 degrees. .

[0026] By using two sweeping sound waves with the same transmission frequency band and an initial phase difference of 180°, the corresponding reflected sound frequency signals are collected respectively, providing a physical basis for the subsequent accurate separation of the linear fundamental frequency response component and the nonlinear harmonic response component through time-domain difference and summation operations. At the same time, the sweeping frequency band is deliberately covered to cover the reference natural frequency of the bolt under standard preload and its lower bias band, ensuring effective excitation of the bolt's linear vibration mode and nonlinear energy dissipation characteristics.

[0027] Furthermore, a directional sound source (such as a parametric array loudspeaker or a focusing horn array) is aimed at the bolt head or connecting plate area of ​​the bolt to be tested, and a first sweep frequency sound wave signal and a second sweep frequency sound wave signal are emitted sequentially within a preset time window, with sufficient interval between the two emissions to avoid signal aliasing. Simultaneously with the emission, the first reflected sound frequency signals generated after excitation are synchronously acquired by microphones or vibration sensors placed near the bolt connection area. Second reflected audio signal The reflected acoustic signal from the excited bolt contains a linear response reflecting the overall macroscopic vibration (proportional to the amplitude of the excitation signal) and a nonlinear response caused by microscopic friction and impact at the loosened interface (mainly manifested as second harmonics, proportional to the square of the excitation signal amplitude). Let the system's linear transfer function be... The nonlinear transmission coefficient is Then, the first reflected audio signal and the second reflected audio signal can be approximately expressed as: (1) (2) In this embodiment, the separation of the linear fundamental frequency response component and the nonlinear harmonic response component specifically refers to: S22, because the linear acoustic response satisfies the superposition principle and is sensitive to the excitation phase, when two excitations with opposite phases are applied to the same structure, the linear fundamental frequency response components have opposite signs. However, the nonlinear harmonic response components (especially even harmonics) are related to the square of the excitation amplitude, are not sensitive to phase reversal, and their signs remain unchanged. Therefore: Perform a time-domain summation operation on the first and second reflected audio signals to separate the nonlinear harmonic response components. : (3) At this point, the linear fundamental frequency response components cancel each other out because they have opposite signs, while the nonlinear harmonic response components are preserved and enhanced.

[0028] Perform time-domain difference operations on the first and second reflected audio signals to separate the linear fundamental frequency response component. : (4) At this point, the nonlinear harmonic response components are canceled out, and the linear fundamental frequency response components are purified.

[0029] To visually demonstrate the above time-domain targeted separation process, Figures 4 to 7 This is a schematic diagram of the time-domain waveforms before and after targeted separation of the audio signal, provided in an embodiment of the present invention. Figure 3 This is a schematic diagram comparing the time-domain waveforms of the audio signal before and after targeted separation, as provided in an embodiment of the present invention. Wherein, as... Figure 4 and Figure 5 As shown, the first and second reflected audio signals were acquired after being excited by acoustic waves with phases of 0° and 180°, respectively. During this stage, the weak transient characteristics of loosening were completely overwhelmed by macroscopic vibration and broadband noise; as... Figure 6 As shown, the output after time-domain summation is displayed. The linear fundamental frequency is completely canceled out due to phase inversion, and the weak nonlinear harmonic response (i.e., the interface impact pulse caused by loosening) is purified and amplified, successfully achieving high signal-to-noise ratio extraction of early hidden faults; as shown... Figure 7 As shown, the pure macroscopic structural linear response extracted after time-domain difference operation is presented.

[0030] Existing active acoustic excitation detection methods typically transmit a single swept-frequency signal and collect mixed reflected signals for overall processing. The weak nonlinear harmonic response caused by early loosening is severely aliased with the strong linear fundamental frequency response in the time-frequency domain. When relying on algorithms such as blind source separation or adaptive filtering for signal separation, inherent limitations exist, including reliance on statistical assumptions, high computational complexity, and significant noise-dependent separation accuracy. This embodiment transmits dual-sweep-frequency acoustic signals with an initial phase difference of 180° and simultaneously collects two reflected signals. Utilizing the different physical characteristics of linear responses satisfying the superposition principle and being sensitive to phase reversal, while nonlinear even-order harmonic responses are insensitive to phase reversal, time-domain summation and difference operations are directly performed on the two signals: during summation, linear components are precisely canceled, and nonlinear components are enhanced in phase; during difference, nonlinear components are canceled, and linear components are purified. This scheme transforms signal separation from "algorithmic separation" relying on statistical estimation into deterministic analytical separation based on physical algebraic relationships. With near-zero computational overhead, it achieves principle-level precise separation of linear and nonlinear components, fundamentally solving the signal-to-noise ratio problem where early loosening features are submerged by a strong linear background.

[0031] S3, performs time-frequency domain transformation on the linear fundamental frequency response component and the nonlinear harmonic response component to generate the multidimensional time-frequency characteristic matrix to be measured; In this embodiment, the step of performing time-frequency domain transformation on the linear fundamental frequency response component and the nonlinear harmonic response component to generate the multidimensional time-frequency characteristic matrix to be measured is specifically as follows: S31, In this embodiment, the complex Morlet wavelet, which has excellent joint resolution for transient impact characteristics, is selected as the mother wavelet function. The mathematical expression for the continuous wavelet transform is: (5) in, This indicates that after continuous wavelet transform, at scale Translation position wavelet coefficients at the location, For the input audio frequency response components, The wavelet scale (corresponding frequency) is used. For time shift parameters, Indicates complex conjugation. For the mother wavelet function.

[0032] Due to wavelet scale With actual physical frequency The two satisfy a mapping relationship: (6) in, The center frequency of the mother wavelet. The sampling frequency.

[0033] S32, the generation of the two-dimensional linear time-frequency matrix and the two-dimensional nonlinear time-frequency matrix includes: S32-1, the first wavelet scaling sequence is used to perform continuous wavelet transform on the linear fundamental frequency response component to generate the first initial time-frequency matrix. Specifically: The linear fundamental frequency response (representing macroscopic structural vibration) is concentrated in a narrow low-frequency band. First, a low-frequency target analysis band covering the reference natural frequency (e.g., ,in This represents the reference natural frequency of the bolt under standard preload, and discretization is performed within this frequency band using linear equidistant steps. The first wavelet scale sequence is then generated by substituting the above mapping relationship. ,in, This represents the total number of discrete scales contained in the first wavelet scale sequence.

[0034] Furthermore, the first wavelet scale sequence Substituting the linear fundamental frequency response component into the continuous wavelet transform formula yields the wavelet coefficients in complex form. The formula for calculating the modulus is: (7) in, and These represent the real and imaginary parts of the corresponding wavelet coefficients, respectively.

[0035] Arrange the modulus sequences at different scales by row and the time shift step by column to generate a dimension of [dimensionality]. The first initial time-frequency matrix. Wherein, This indicates the total number of time sampling points for the signal.

[0036] S32-2, the second wavelet scaling sequence is used to perform continuous wavelet transform on the nonlinear harmonic response components to generate a second initial time-frequency matrix. Specifically: Determine the target analysis frequency band covering high-frequency harmonics (e.g., set to...). The second wavelet scale sequence is generated by discretizing the sample within the wide bandwidth using logarithmic equidistant steps. .in, This represents the total number of discrete scales contained in the second wavelet scale sequence. Similarly, performing a transformation and extracting the complex modulus generates a sequence with a dimension of... The second initial time-frequency matrix. Using a logarithmic step size can control the number of rows in the matrix. Under the premise of achieving high-fidelity capture of high-frequency transient isolated spikes.

[0037] S32-3, due to the dimensions of the two-dimensional linear time-frequency matrix and the two-dimensional nonlinear time-frequency matrix ( and If the scale direction is inconsistent, a target mapping grid with a unified time-scale dimension must be established, and the dimension of the target mapping grid is set to [value missing]. .in, This represents the total number of time steps defined for the target grid. This represents the total number of scale frequency points defined in the target mesh.

[0038] S32-4, the method of mapping the first initial time-frequency matrix to the target mapping grid using linear interpolation specifically involves: for any coordinate in the target mapping grid... Determine its corresponding floating-point mapping coordinates in the first initial time-frequency matrix. Determine four integer neighboring points that enclose the floating-point mapped coordinates. Calculate interpolation weights based on the inverse ratio of the distance between each neighboring point and the floating-point mapped coordinates. Use the weighted sum of the moduli of the four neighboring points as the target coordinates. The modulus value at the specified point. After traversing all coordinate points of the target mapping grid, a two-dimensional linear time-frequency matrix aligned with the target mapping grid is generated. Linear interpolation can smoothly preserve the slowly varying characteristics of the linear fundamental frequency response components.

[0039] S32-5, the step of extracting the maximum local modulus of the second initial time-frequency matrix and mapping it to the target mapping mesh specifically involves: dynamically determining the size and sliding step of the sliding window based on the dimensional ratio of the original matrix and the target mesh. The row-oriented (scale-oriented) window size is set. Column-oriented (time-oriented) window size .

[0040] Subsequently, a matrix of size is constructed in the second initial time-frequency matrix. A sliding window, with a row step size of 1. The column step size is Perform a non-overlapping sliding traversal (Stride operation). Extract the... The maximum local modulus values, in accordance with their spatial topological order in the second initial time-frequency matrix, are filled into the target mapping grid in a one-to-one correspondence. In coordinates, thus generating dimensions of A two-dimensional nonlinear time-frequency matrix.

[0041] S32-6 concatenates the generated two-dimensional linear time-frequency matrix with the generated two-dimensional nonlinear time-frequency matrix along the feature channel dimension to obtain the multi-dimensional time-frequency feature matrix to be tested. The final generated multi-dimensional time-frequency feature matrix is ​​a three-dimensional tensor, whose shape and size are strictly defined as... , where 2 indicates that the matrix has two characteristic channels: linear and nonlinear.

[0042] S4, calculate the current reconstruction error based on the multi-dimensional time-frequency feature matrix to be tested, compare the error with the warning boundary, and determine the loosening status; The warning boundary is determined by calculating and reconstructing the offset vector using the multidimensional time-frequency feature matrix under loose conditions, and then directionally correcting the initial boundary; the initial boundary is generated by calculating the reconstruction error of the multidimensional time-frequency feature matrix under healthy conditions.

[0043] S41, the step of obtaining the initial boundary includes: S41-1, Let the set of multi-dimensional time-frequency feature matrices under the healthy state obtained in step S1 be... ,in, This represents the total number of normal samples, and each sample... All are of dimension size Three-dimensional tensor data.

[0044] Due to the extreme scarcity of bolt loosening fault data in actual industrial scenarios, this embodiment employs unsupervised learning. In specific implementation, a deep neural network based on an autoencoder or variational autoencoder is constructed as the actuator.

[0045] The compression and dimensionality reduction operation will convert the input normal matrix... Spatial features are downsampled using multi-layer two-dimensional convolutional kernels (2D-CNN) and mapped to a low-dimensional latent space to extract the core latent variables of the bolt tightening state. .

[0046] The reverse reconstruction operation will remove hidden variables. Upsampling and reconstruction are performed using deconvolution layers, resulting in a reconstructed matrix with the exact same dimensions as the input. .

[0047] Minimizing the reconstruction error is the constraint, i.e., a loss function is defined. The network weights are continuously updated through backpropagation until the network converges. The specific loss function is: (8) Using the trained network, the above compression and dimensionality reduction operations and inverse reconstruction operations are performed on all normal historical multidimensional time-frequency feature matrices to obtain the reconstructed matrix of each normal matrix. .

[0048] S41-2, In this embodiment, the mean square error (MSE) is used as the metric for calculating the baseline reconstruction error, and the calculation formula is as follows: (9) in, Indicates the first The baseline reconstruction error scalar value corresponding to each normal sample. Represents the coordinate index of the feature channel dimension. Represents the original matrix In three-dimensional coordinates The actual element value at that location. Represents the reconstruction matrix In three-dimensional coordinates The value of the reconstructed element at that location.

[0049] Calculate the errors of all normal samples to obtain the baseline reconstruction error set. .

[0050] S41-3 performs probability density estimation on the baseline reconstruction error set to obtain the probability density distribution. Specifically, a nonparametric kernel density estimation algorithm is used, and its probability density distribution function is derived as follows: (10) in, This represents the probability density distribution function of the reconstruction error obtained through fitting. The continuous independent variable representing the reconstruction error, Represents a set The known baseline reconstruction error sample values ​​in the data. This represents the smoothing bandwidth hyperparameter in kernel density estimation, used to control the smoothness of the fitted curve. This indicates the specified Gaussian kernel function.

[0051] Based on the above calculations, the probability density distribution of the reference reconstruction error is obtained. .

[0052] S41-4, In the physical scenario of bolt monitoring, a smaller reconstruction error indicates that the feature is closer to normal; a surge in error indicates a loosening anomaly. Therefore, only the right-side boundary where the error increases needs to be monitored. The preset one-sided confidence level is set to... (For example, based on industrial engineering experience, the percentage is set at 95%), by solving the following definite integral equation: (11) Calculate the upper limit of the integral that satisfies the confidence interval. and the error quantile values ​​(i.e. ) is established as the initial boundary.

[0053] It should be noted that existing bolt diagnostic schemes based on active acoustic excitation either require supervised training with a large number of loosening fault samples or rely solely on healthy data to establish fixed thresholds when setting judgment boundaries. In real industrial scenarios where loosening samples are extremely scarce, it is difficult to effectively establish diagnostic boundaries. This embodiment introduces an autoencoder network, using a massive multidimensional time-frequency feature matrix under healthy conditions as input. Through compression dimensionality reduction and inverse reconstruction operations, it learns the data distribution pattern of normal states and establishes initial safety boundaries based on the probability density estimation of reconstruction errors and one-sided confidence levels. This process is completely independent of any loosening fault samples, utilizing only the massive amounts of normally tightened data that can be collected in industrial scenarios, thus solving the problem of not being able to establish diagnostic models when loosening samples are scarce.

[0054] S42, the step of calculating and reconstructing the offset vector using the multi-dimensional time-frequency feature matrix under loosened state includes: S42-1, the set of multi-dimensional time-frequency feature matrices of a small number of loose states obtained in step S1 is as follows: ,in, This represents the set of feature matrices that clearly have loosening fault labels. This indicates the total number of loose samples contained in the set. Indicates the first One loose sample.

[0055] S42-2, the multidimensional time-frequency feature matrix under a small amount of loosening conditions The compression and dimensionality reduction operations and the inverse reconstruction operations that have been trained in S4 are input sequentially to obtain the reconstructed output. To preserve the directionality of fault evolution, the anomaly reconstruction error vector is first calculated in the multidimensional feature space. And calculate its scalar modulus. .

[0056] because To characterize the safety radius as a scalar, this embodiment extracts the unit vector of the error direction and assigns the out-of-bounds deviation value as a vector attribute. The calculation formula is as follows: (12) in, Indicates the first The vector of out-of-bounds deviation values ​​for each loose sample. This indicates that the deviation is counted as 0 when the module length does not exceed the boundary. This represents the unit direction vector of the error evolution of the loosened sample in the feature space.

[0057] The vector mean of the out-of-bounds deviation values ​​is used to obtain the reconstructed offset vector: (13) in, The final reconstructed offset vector indicates the dominant offset direction in the high-dimensional reconstruction error space when the bolt evolves from a healthy state to a loose state, and its magnitude reflects the average degree of loosening.

[0058] S43, the directional correction of the initial boundary includes: It should be noted that the scalar initial boundary In the multidimensional error space, a radius is essentially defined. The initial symmetric boundary is given. Let the boundary position vector of any point on this initial symmetric boundary be y. (and satisfy) ).

[0059] S43-1, along the vector direction of the reconstructed offset vector, perform anisotropic topological stretching on the initial boundary, specifically: First, calculate the unit direction vector of the reconstructed offset vector. Based on spatial analytic geometry, points Orthogonal decomposition into first boundary components that are parallel to and orthogonal to the reconstructed offset vector. With the second boundary component : (14) Furthermore, a shrinking mapping is performed on the first boundary component to compress the safety depth, and a maintaining or expanding mapping is performed on the second boundary component to ensure error tolerance. Finally, vector superposition is performed on the mapped first and second boundary components to construct an asymmetric closed boundary. (15) in, This represents the position vector of the new boundary point after spatial topological stretching. Let represent the shrinkage mapping coefficients, and satisfy . , Denotes the extended mapping coefficients, and satisfies .

[0060] S43-2, Traverse all points on the original symmetric initial boundary. Perform the above topological mapping calculation, and generate all the new points. The resulting geometric topological set is the asymmetric closed boundary formed after spatial topological stretching, which is the final established warning boundary for online real-time judgment.

[0061] To further clarify the technical suppression effect of the above-mentioned anisotropic topology stretching operation on missed and false alarms, comparison diagrams were drawn to show the anti-missed and anti-false alarm effects of the traditional symmetric fixed boundary and the asymmetric early warning boundary of the present invention. Figure 2 This is a schematic diagram illustrating the diagnostic effect of a traditional symmetrical fixed boundary provided in an embodiment of the present invention, such as... Figure 2 As shown, due to the lack of perception of the direction of fault evolution by the symmetrical fixed boundary, the real early loosening features (red dots) that are closer to the origin are incorrectly included in the safe zone (false alarms occur), while the wind noise interference features (gray crosses) that are farther away are incorrectly classified into the danger zone (false alarms occur). Figure 3 This is a schematic diagram illustrating the diagnostic effect of the asymmetric early warning boundary provided in an embodiment of the present invention, as shown below. Figure 3 As shown, this invention extracts the reconstructed offset vector. To characterize the spatial evolution direction of physical loosening. Along this In terms of direction, the asymmetric warning boundary performs an extreme contraction mapping (such as the minor axis of the blue solid line ellipse), forcibly stripping early red dot loosening features out of the safe zone for precise warning; while orthogonal to In the direction of the asymmetric warning boundary, an expansion mapping (such as the major axis of the blue solid-line ellipse) is performed, successfully containing the gray cross wind noise interference features within a safe range to achieve noise resistance. This scheme completely breaks the technical contradiction of being unable to balance false alarms and false alarms from the underlying spatial geometric logic. It achieves a leap from "symmetric fixed" to "directional sensitive" diagnostic boundary with very few loosened samples, balancing high detection rate and low false alarm rate.

[0062] S44, the calculation of the current reconstruction error based on the multi-dimensional time-frequency feature matrix to be measured includes: S44-1, the multidimensional time-frequency feature matrix generated from the bolt under test by S2 and S3. The compression and dimensionality reduction operation and the inverse reconstruction operation are performed in input S41-1 to obtain the corresponding current reconstruction matrix. .

[0063] S44-2, calculate the feature difference value between the multidimensional time-frequency feature matrix to be measured and the current reconstruction matrix, and establish the feature difference value as the current reconstruction error. In specific implementation, calculate the current reconstruction error vector. Calculate the multidimensional time-frequency feature matrix to be measured. With the current reconstruction matrix The multidimensional Euclidean distance (i.e., the Frobenius norm) between them is calculated, and the resulting scalar value is defined as the feature difference value. The calculation formula is as follows: (16) in, The calculated feature difference value is then used as the current reconstruction error for subsequent determination.

[0064] To perform accurate spatial determination, the difference matrix between the multidimensional time-frequency feature matrix to be measured and the currently reconstructed matrix is ​​calculated. The unit direction vector of the difference matrix is ​​extracted and denoted as the current evolution direction vector. .

[0065] S44-3, in order to determine scalar Whether the asymmetric warning boundary has been exceeded in multidimensional space must be dynamically calculated in conjunction with the current direction of error evolution.

[0066] First, the evolution direction of the current reconstruction error is extracted, that is, the evolution direction of the current error is calculated. The unit direction vector of the known reconstructed offset vector The spatial angle between It satisfies the algebraic relation: .

[0067] Secondly, based on the initial boundary established in step S41 and the shrinkage mapping coefficient set in step S43-1 With the extended mapping coefficient Calculate the dynamic boundary threshold radius of the warning boundary along the current evolution direction. The following equation for the intersection of polar coordinates in analytical geometry is rigorously calculated: (17) Finally, the established current reconstruction error will be... With the dynamic boundary threshold in this direction Compare and determine: like Less than or equal to This indicates that the current reconstruction error has not exceeded the dynamic safety defense line of the early warning boundary in this evolution direction, and the bolt under test is determined to be in a normal tightening state; if Greater than If the endpoint of the current reconstruction error has been breached or exceeded the warning boundary, it is determined that the bolt to be tested has become loose.

[0068] Existing technologies typically compare current signal characteristics with fixed thresholds or symmetrical boundaries using scalar methods, without considering the spatial relationship between the abnormal evolution direction and the boundary shape. This can easily lead to false alarms in complex noise environments due to inconsistencies between the direction of fluctuation under normal operating conditions and the direction of loosening. This embodiment extracts the evolution direction of the current reconstruction error during the online determination phase, calculates the spatial angle between it and the reconstruction offset vector, and performs direction-aware determination based on the dynamic boundary threshold radius along this direction. Along the offset direction, boundary contraction makes it easier to trigger alarms, while along the orthogonal direction, boundary expansion makes it more tolerant of fluctuations. This achieves direction-adaptive dynamic diagnosis, avoiding false alarms caused by inconsistencies between the direction of fluctuation under normal operating conditions and the direction of loosening, and significantly improving diagnostic accuracy in high-noise backgrounds and extremely unbalanced data distributions.

[0069] Example 2: In order to further objectively demonstrate the effectiveness and advancement of the power transmission tower bolt condition diagnosis method described in Example 1 in solving technical pain points in real industrial scenarios, this example provides a set of comparative test experiments under complex physical environments.

[0070] The verification tests in this embodiment were conducted using a physical test suite that incorporated strong wind noise (simulating random low-frequency vibrations of the tower caused by ambient wind speeds > 8 m / s) and electromagnetic interference. The existing "model based on Euclidean distance metric and symmetric dynamic threshold determination" (i.e., the existing conventional acoustic detection scheme) was selected as the benchmark. Simultaneously, to verify the independent contribution of the core feature "anisotropic topological stretching" in this invention, a "variant of this embodiment" employing only a semi-supervised architecture but without stretching action was set up as the ablation experimental group.

[0071] The same multidimensional time-frequency feature matrix to be tested was input into the three diagnostic models / judgment boundary models mentioned above, and the false alarm rate, false alarm rate, and overall diagnostic accuracy were statistically analyzed when dealing with early and slight loosening. The specific performance comparison test results are shown in Table 2. Table 2. Comparison of diagnostic accuracy of different decision boundary models in complex and noisy environments.

[0072] The experimental comparison in Table 2 shows that the traditional symmetric boundary model, due to its inability to distinguish the physical causes of feature offset, falls into a technical contradiction of balancing the false alarm rate and the missed alarm rate. In contrast, the anisotropic topology stretching algorithm used in this embodiment, by limiting contraction in the direction parallel to the fault evolution to prevent false alarms and by maintaining identity or expanding in the direction orthogonal to the fault evolution to prevent false alarms, simultaneously suppresses the false alarm rate and the false alarm rate to an extremely low level, achieving highly robust and accurate diagnosis of early minor loosening of transmission tower bolts.

[0073] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0074] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0075] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0076] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0077] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0078] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for diagnosing the condition of bolts on power transmission towers, characterized in that, Includes the following steps: A sweeping frequency sound wave is emitted to the bolt under test and the reflected sound frequency signal is collected to separate the linear fundamental frequency response component and the nonlinear harmonic response component. The linear fundamental frequency response component and the nonlinear harmonic response component are transformed in the time-frequency domain to generate the multidimensional time-frequency characteristic matrix to be measured. The current reconstruction error is calculated based on the multi-dimensional time-frequency feature matrix to be tested, and the error is compared with the warning boundary to determine the loosening status; The warning boundary is determined by calculating and reconstructing the offset vector using the multidimensional time-frequency feature matrix under loose conditions, and then directionally correcting the initial boundary; the initial boundary is generated by calculating the reconstruction error of the multidimensional time-frequency feature matrix under healthy conditions.

2. The method according to claim 1, characterized in that, The process of emitting a swept-frequency sound wave to the bolt under test and acquiring the reflected sound frequency signal includes: The frequency band of the sweeping sound wave is set to cover the reference natural frequency and its lower offset frequency band of the bolt under standard preload. The first and second sweep frequency acoustic signals with the same radio frequency band and an initial phase difference of 180 degrees are sequentially sent to the bolt under test, and the first and second reflected audio signals generated by the bolt under test after being excited are simultaneously collected.

3. The method according to claim 2, characterized in that, The separation of the linear fundamental frequency response component and the nonlinear harmonic response component includes: Time-domain difference and time-domain summation operations are performed on the first reflected audio signal and the second reflected audio signal, respectively, to separate the linear fundamental frequency response component and the nonlinear harmonic response component.

4. The method according to claim 3, characterized in that, The step of performing time-frequency domain transformation on the linear fundamental frequency response component and the nonlinear harmonic response component to generate the multidimensional time-frequency characteristic matrix to be measured includes: Continuous wavelet transforms are performed on the linear fundamental frequency response component and the nonlinear harmonic response component respectively to generate a two-dimensional linear time-frequency matrix and a two-dimensional nonlinear time-frequency matrix. By concatenating the two-dimensional linear time-frequency matrix and the two-dimensional nonlinear time-frequency matrix along the feature channel dimension, the multi-dimensional time-frequency feature matrix to be tested is obtained.

5. The method according to claim 4, characterized in that, The generation of the two-dimensional linear time-frequency matrix and the two-dimensional nonlinear time-frequency matrix includes: The linear fundamental frequency response component and the nonlinear harmonic response component are subjected to continuous wavelet transform using the first and second wavelet scale sequences, respectively, to generate the corresponding first and second initial time-frequency matrices. Establish a target mapping grid with a unified time-scale dimension, map the first initial time-frequency matrix to the grid through linear interpolation, and map the second initial time-frequency matrix to the grid after extracting the local maximum value of the modulus.

6. The method according to claim 5, characterized in that, The steps for obtaining the initial boundary include: With minimizing the reconstruction error as a constraint, compression, dimensionality reduction, and inverse reconstruction operations are performed on the multidimensional time-frequency feature matrix under healthy conditions to calculate the baseline reconstruction error and obtain the baseline reconstruction error set; Probability density estimation is performed on the baseline reconstruction error set to obtain the probability density distribution; Based on the probability density distribution and the preset one-sided confidence level, the corresponding error quantile values ​​are extracted and established as the initial boundary.

7. The method according to claim 6, characterized in that, The calculation and reconstruction of the offset vector using the multi-dimensional time-frequency feature matrix under loosened state includes: Perform the compression, dimensionality reduction, and inverse reconstruction operations on the multidimensional time-frequency feature matrix in the loosened state, and calculate the corresponding abnormal reconstruction error; Calculate the out-of-bounds deviation value of each abnormal reconstruction error that exceeds the initial boundary, and perform vector mean operation on the out-of-bounds deviation value to obtain the reconstruction offset vector.

8. The method according to claim 7, characterized in that, The directional correction of the initial boundary includes: Anisotropic topological stretching is performed on the initial boundary along the vector direction of the reconstructed offset vector; The asymmetric closed boundary formed after stretching is established as the warning boundary.

9. The method according to claim 8, characterized in that, The anisotropic topological stretching of the initial boundary along the vector direction of the reconstructed offset vector includes: The initial boundary is orthogonally decomposed into a first boundary component that is parallel to the reconstructed offset vector and a second boundary component that is orthogonal to the reconstructed offset vector. Perform a shrinking mapping on the first boundary component and a maintaining or expanding mapping on the second boundary component; The first boundary component and the second boundary component after mapping are vector superimposed to obtain an asymmetric closed boundary.

10. The method according to claim 9, characterized in that, The calculation of the current reconstruction error based on the multidimensional time-frequency feature matrix to be tested includes: Perform compression, dimensionality reduction, and inverse reconstruction operations on the multidimensional time-frequency feature matrix to be tested to obtain the corresponding current reconstruction matrix; The feature difference between the multidimensional time-frequency feature matrix to be tested and the current reconstruction matrix is ​​calculated as the current reconstruction error.

Images