Insulator defect detection method, device, equipment and storage medium
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHANXI ZHONGSHI ELECTRICITY TECH CO LTD
- Filing Date
- 2026-02-10
- Publication Date
- 2026-06-02
Smart Images
Figure CN122134658A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of defect detection, and more particularly to a method, apparatus, equipment, and storage medium for detecting defects in insulators. Background Technology
[0002] During long-term operation, insulators are susceptible to various structural defects, such as cracks, porosity, delamination, and skirt damage, due to factors such as material aging, manufacturing defects, mechanical vibration and impact, and external environmental corrosion. These defects are often well-hidden and difficult to detect in their early stages, but they expand over time, leading to decreased insulation performance, enhanced partial discharge, and even flashover and breakage accidents, seriously threatening the safe operation of transmission lines and power equipment. Existing insulator defect detection methods mainly include manual inspection, visual inspection, infrared thermography, and ultrasonic or partial discharge detection. These methods have certain limitations in practical applications. For example, manual inspection relies on experience, is inefficient, and highly subjective; infrared and ultrasonic detection are sensitive to environmental conditions and easily affected by temperature, humidity, and noise interference; partial discharge detection equipment is complex, costly, and cannot comprehensively reflect the internal structural defects of the insulator. Furthermore, most detection methods rely on post-event analysis or periodic inspections, making it difficult to achieve real-time perception and early warning of insulator defects. Summary of the Invention
[0003] To address the aforementioned technical problems, this invention proposes a method, apparatus, device, and storage medium for detecting defects in insulators, thereby resolving at least one of the aforementioned technical problems.
[0004] To achieve the above objectives, the present invention provides a defect detection method for insulators, comprising the following steps: Step S1: Apply a controlled random excitation signal to the insulator under test and simultaneously acquire response signals at different positions of the insulator; Step S2: Perform vibration feature analysis on the response signal to obtain spectral shift feature sets at different locations; Step S3: Perform anomaly deviation analysis on the spectral offset feature set and mark it as anomaly acoustic vibration vector; Step S4: Perform multi-scale feature decomposition on the abnormal acoustic vibration vector to construct a defect feature vector matrix; Step S5: Input the defect feature vector matrix into a preset deep learning architecture for unsupervised learning and output an insulator inspection report.
[0005] This specification provides an insulator defect detection device for performing the insulator defect detection method described above, comprising: The application unit is used to apply a controlled random excitation signal to the insulator under test and simultaneously acquire response signals at different positions of the insulator. The vibration analysis unit is used to perform vibration characteristic analysis on the response signal to obtain a spectral shift feature set at different locations; The deviation analysis unit is used to perform abnormal deviation analysis on the spectral offset feature set and mark it as an abnormal acoustic vibration vector. The decomposition unit is used to perform multi-scale feature decomposition on the abnormal acoustic vibration vector to construct a defect feature vector matrix. The learning unit is used to input the defect feature vector matrix into a preset deep learning architecture for unsupervised learning and output an insulator inspection report.
[0006] The present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the defect detection method for insulators described in any of the above claims.
[0007] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the defect detection method for insulators described in any of the preceding claims.
[0008] The beneficial effects of this invention are as follows: By using controlled random excitation signals, the insulator can be forced to vibrate over a wide frequency range, simultaneously exciting both the overall structural modes and local microstructural responses, avoiding the mode omission problem caused by single-frequency excitation. Random excitation has rich energy distribution in both the time and frequency domains, effectively amplifying the dynamic response differences caused by early defects such as internal microcracks, loosening, and internal air gaps, thus improving defect detectability. Simultaneous acquisition of response signals from different locations on the insulator allows for the quantification of vibration transmission characteristics, phase relationships, and energy distribution differences between measurement points, providing a foundation for subsequent spatial anomaly analysis. Through spectral analysis, complex time-domain vibration signals are mapped to the frequency domain, extracting highly sensitive features such as resonant frequencies, spectral peak positions, and energy distribution. When insulators develop cracks, age, or internal structural damage, their equivalent stiffness and damping change, causing a shift in spectral peak values; spectral shift characteristics directly reflect these changes. Spectral shift characteristics at different acquisition locations can characterize the changes in vibration propagation along the insulator structure, facilitating the location of defects. Anomaly analysis of spectral offset features can identify abnormal characteristics that deviate from the normal structural response distribution, avoiding reliance on human experience for judgment. Anomaly analysis, based on feature deviation rather than absolute values, effectively distinguishes between normal manufacturing differences and abnormal changes caused by actual defects. Anomaly spectral features are uniformly labeled as anomalous acoustic-vibration vectors, allowing defect-induced vibration anomalies to be centrally expressed in the feature space, enhancing the weight of the defect signal. Multi-scale feature decomposition can decompose anomalous acoustic-vibration signals into components at different time or frequency scales, facilitating the simultaneous capture of macroscopic structural damage and microscopic defect features. Early defects such as microcracks and local debonding are typically manifested in high-frequency or local-scale features; multi-scale decomposition can effectively extract these weak features. By constructing a defect feature vector matrix, multi-scale, multi-location, and multi-feature information is uniformly organized to form a defect description with physical meaning and statistical regularity. Using unsupervised learning, normal and anomalous feature distributions can be automatically learned even without known defect labels, making it suitable for practical engineering scenarios. Deep learning architecture can model the nonlinear relationships between defect features, identifying latent defect features that are difficult to detect using traditional threshold or rule-based methods. Attached Figure Description
[0009] Fig. 1 This is a schematic diagram of the steps of a defect detection method for insulators according to the present invention; Fig. 2 This is a detailed flowchart illustrating the implementation steps of step S1. Fig. 3 This is a flowchart illustrating the detailed implementation steps of step S2. Detailed Implementation
[0010] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.
[0011] This application provides a method, apparatus, device, and storage medium for detecting defects in insulators. The execution entities of the method, apparatus, device, and storage medium for detecting defects in insulators include, but are not limited to, mechanical equipment, data processing platforms, cloud server nodes, and network upload devices mounted on the system, which can be considered as general computing nodes in this application. The data processing platform includes, but is not limited to, at least one of an audio-visual management system, an information management system, and a cloud-based data management system.
[0012] Please see Figs. 1 to 3 This invention provides a method for detecting defects in insulators, comprising the following steps: Step S1: Apply a controlled random excitation signal to the insulator under test and simultaneously acquire response signals at different positions of the insulator; Step S2: Perform vibration feature analysis on the response signal to obtain spectral shift feature sets at different locations; Step S3: Perform anomaly deviation analysis on the spectral offset feature set and mark it as anomaly acoustic vibration vector; Step S4: Perform multi-scale feature decomposition on the abnormal acoustic vibration vector to construct a defect feature vector matrix; Step S5: Input the defect feature vector matrix into a preset deep learning architecture for unsupervised learning and output an insulator inspection report.
[0013] In the embodiments of the present invention, see Fig. 1 The diagram below illustrates the steps of a defect detection method for insulators according to the present invention. In this example, the steps of the method include: Step S1: Apply a controlled random excitation signal to the insulator under test and simultaneously acquire response signals at different positions of the insulator; In this embodiment, the insulator under test is mounted on a dedicated test bracket. The test bracket adopts a high-rigidity structure and is vibration-damping and isolated from the ground to avoid interference from external environmental vibrations on the test results. The installation method of the insulator is consistent with its on-site operating conditions. Subsequently, a controlled random excitation signal is applied to the insulator using an electromagnetic or piezoelectric excitation device. The excitation signal is generated by a signal generator, and the frequency range is typically set to 100 Hz to 5 kHz to cover the main structural mode frequency bands of the insulator. The excitation amplitude is controlled according to the mechanical rated strength of the insulator, and the corresponding equivalent excitation force is generally set between 50 N and 300 N. The random excitation signal has randomness in amplitude and phase, but the overall energy distribution remains stable. The excitation duration is set to about 30 seconds to ensure that the structural response enters a steady state. During the excitation process, multiple sensors are deployed at different key structural locations on the insulator for synchronous data acquisition. These include accelerometers for acquiring the vibration response and acoustic sensors for acquiring the accompanying acoustic response. The accelerometers typically have a range of ±50 g, while the acoustic sensors have a frequency response range covering 20 Hz to 20 kHz. All sensors are simultaneously sampled through a multi-channel data acquisition system with a sampling frequency set to no less than 25.6 kHz, thereby obtaining raw response signal data from multiple locations and of various types.
[0014] Step S2: Perform vibration feature analysis on the response signal to obtain spectral shift feature sets at different locations; In this embodiment, the original response signal is preprocessed, including removing the DC component, eliminating low-frequency drift, and suppressing environmental noise interference, to ensure that the analyzed signal mainly contains the effective structural response components caused by random excitation. The processed signal is then segmented for analysis according to a fixed time window, with each segment typically consisting of 2048 or 4096 sampling points, and a certain overlap ratio is set to improve the stability of the spectral analysis results. During the frequency domain analysis, the spectrum of each segment is estimated to obtain the spectral distribution of different measurement points within the entire excitation frequency band. By comparing the current test spectrum with the spectrum of the healthy state or historical baseline, feature information such as changes in spectral peak positions and changes in the energy distribution of the main frequency band is extracted, thereby forming a spectral shift feature set characterizing the structural response changes at different measurement locations.
[0015] Step S3: Perform anomaly deviation analysis on the spectral offset feature set and mark it as anomaly acoustic vibration vector; In this embodiment, the spectral shift characteristics of different measurement points are compared with the corresponding baseline thresholds to analyze whether there are significant deviations in terms of spectral peak shift amplitude, energy distribution pattern, and frequency band expansion. By comprehensively analyzing the characteristics of multiple frequency bands and multiple measurement points, the consistency and stability of the shift characteristics are determined to avoid misjudgments caused by a single frequency point or random noise. When the spectral shift characteristics of a certain measurement point or a certain frequency band significantly exceed the normal fluctuation range in amplitude or trend, and show correlation in adjacent frequency bands or adjacent measurement points, this set of characteristics is marked as abnormal acoustic vibration characteristics. Subsequently, these abnormal characteristics are organized according to the measurement point location and spectral characteristic type to form an abnormal acoustic vibration vector.
[0016] Step S4: Perform multi-scale feature decomposition on the abnormal acoustic vibration vector to construct a defect feature vector matrix; In this embodiment, multi-scale eigenvalue decomposition analysis is performed on the labeled anomalous acoustic vibration vectors to systematically extract key feature parameters related to structural defects. The multi-scale analysis is conducted at both the time and frequency scales. For example, short-time and long-time windows are used to analyze transient anomalies and steady-state characteristics at the time scale, and the excitation frequency band is divided into low-frequency, mid-frequency, and high-frequency sub-bands at the frequency scale. By analyzing the anomalous acoustic vibration vectors at different scales, anomalous parameters such as spectral peak position variation characteristics, main energy bandwidth expansion characteristics, energy attenuation trends, and the proportion of random noise are extracted. Subsequently, the anomalous features extracted at each scale and measurement point are uniformly normalized to ensure comparability of features with different dimensions and numerical ranges. Based on this, the anomalous features are combined and arranged according to the measurement point location and feature type to finally construct a defect feature vector matrix.
[0017] Step S5: Input the defect feature vector matrix into a preset deep learning architecture for unsupervised learning and output an insulator inspection report.
[0018] In this embodiment, a pre-defined deep learning architecture is introduced for unsupervised learning processing to achieve automatic discrimination and detection result output of insulator state. The deep learning architecture employs an unsupervised model with self-organizing and representational capabilities. By learning the intrinsic structure of a large number of defect feature samples, the model can automatically distinguish between normal and abnormal feature patterns. During training, manual labeling of defect types is not required; instead, the acoustic and vibration feature responses related to defects are enhanced through feature reconstruction capabilities and differences in feature distribution. Training parameters can be set according to the experimental scale, for example, the number of training iterations can be set to 100-200 times to ensure model convergence stability. After unsupervised learning, the model outputs a comprehensive discrimination result of the current defect feature vector matrix and, combined with pre-defined diagnostic rules, generates an insulator detection report containing information about the detected object, detection parameters, abnormal feature descriptions, and overall state assessment conclusions.
[0019] In this embodiment, see Fig. 2 The diagram below illustrates the detailed implementation steps of step S1. In this embodiment, the detailed implementation steps of step S1 include: A controlled random excitation signal is applied to the insulator under test, and response signals at different positions of the insulator are collected simultaneously; the response signals include vibration acceleration signals and accompanying acoustic response signals. The response signal is subjected to time synchronization and amplitude normalization processing to generate an optimized signal; Power spectral density estimation is performed on the optimized signal to extract energy distribution curves at different locations; The energy distribution curves are analyzed for morphological consistency and energy concentration trend to construct an excitation energy distribution baseline.
[0020] In this embodiment, the insulator under test is installed on a dedicated mechanical testing support. The testing support adopts a high-rigidity steel structure and is isolated from the ground by vibration damping pads to reduce environmental vibration interference. The installation method of the insulator is consistent with its on-site operating condition. An electromagnetic random exciter is selected as the excitation device. The exciter is connected to the designated excitation point of the insulator through a rigid connector. The excitation point is usually selected at the metal end or a position with high structural rigidity to ensure that the excitation energy can be effectively transferred to the overall structure of the insulator. The excitation signal is generated by a signal generator. The signal type is a band-limited random excitation signal with a frequency range of 100 Hz to 5 kHz. The excitation amplitude is controlled according to the rated mechanical strength of the insulator, usually corresponding to an equivalent excitation force of 50 N to 300 N. The excitation duration is set to more than 30 s to ensure that the structural response reaches a steady state. During the excitation process, multiple vibration acceleration sensors are deployed at different key structural locations on the insulator, such as at the base of the skirt, the middle of the skirt, and the core rod area. The acceleration sensors have a range of ±50 g and a frequency response range of no less than 10 kHz. Simultaneously, airborne acoustic sensors are placed at reasonable distances around the insulator to collect the accompanying acoustic response signals caused by the excitation. The frequency response range of these acoustic sensors covers 20 Hz to 20 kHz. All sensor signals are synchronously acquired through a multi-channel data acquisition system. The sampling frequency is uniformly set to no less than 25.6 kHz, and a unified trigger mechanism is used to start the acquisition to ensure the consistency of the excitation signal and each response signal on the time axis, thereby obtaining raw response data from multiple locations and of multiple types. Time synchronization correction is performed on the signals of each channel. Although the acquisition system uses synchronous sampling, slight time offsets may still exist due to factors such as sensor response delays and differences in installation conditions. Therefore, the excitation input signal is used as a time reference. Correlation analysis is used to determine the time delay of each response signal relative to the excitation signal, and the signals are uniformly aligned to ensure that signals from different measurement points can be directly compared under the same time reference. After time synchronization, the signal amplitude is normalized to eliminate inconsistencies caused by differences in sensor sensitivity, installation stiffness, and local structural characteristics at different measuring points. During normalization, DC and trend term removal are performed on each channel signal to eliminate low-frequency drift. Then, the effective value or statistical energy index of each channel signal over the entire sampling period is calculated, and this index is used as the normalization benchmark to map the original signal amplitude to a uniform scale range. For vibration acceleration signals and acoustic response signals, normalization is performed separately within their respective signal types to ensure comparability between signals of the same type.
[0021] The optimized time-domain signal is divided into several data segments of fixed length, typically with 2048 or 4096 sampling points per segment, corresponding to a frequency resolution of approximately 6 Hz to 12 Hz. Segments are overlapped by about 50% to improve the stability of the spectral estimation. A window function is applied to each data segment to suppress spectral leakage; commonly used window functions include the Hanning window or the Blackman window. Subsequently, frequency domain transformation is performed on each data segment, and the power spectral density is calculated. The power spectral density results of all data segments are then averaged to obtain a stable power spectral density estimate for each measurement point across the entire excitation frequency band. Based on the power spectral density results, energy statistics are further performed on the frequency axis. By integrating or summing the power spectral density within a specified frequency band, the energy values corresponding to each frequency band are extracted, forming energy distribution curves characterizing the energy distribution features of different measurement points. Comparative analysis of energy distribution curves at similar measuring points was conducted, focusing on the frequency location of the main energy peaks, the trend of peak amplitude variation, and the overall morphological characteristics of energy changes with frequency. Under structurally sound conditions, the energy distribution curves at each measuring point typically exhibit high consistency within the main structural modal frequency bands, with concentrated and gently changing peak frequencies. Further analysis of the energy concentration in the low, mid, and high frequency regions was performed to determine whether the excitation energy was primarily concentrated in the structural's inherent modal frequency bands or whether there was any abnormal frequency band energy enhancement. For acoustic response signals, the energy concentration trend should be relatively consistent with the structural vibration energy distribution. Abnormal concentration of acoustic energy in local frequency bands may reflect internal structural discontinuities or damage characteristics. Through comprehensive morphological consistency analysis of energy distribution curves at multiple locations and with multiple signal types, a set of energy distribution patterns capable of stably characterizing the dynamic response of healthy insulators was extracted, and this pattern was used as the baseline for excitation energy distribution.
[0022] In this embodiment, see Fig. 3 The diagram below illustrates the detailed implementation steps of step S2. In this embodiment, the detailed implementation steps of step S2 include: Vibration characteristic analysis was performed on the associated acoustic response signal to extract the acoustic power spectrum; Based on the acoustic power spectrum, the baseline of the excitation energy distribution is compared and analyzed to extract the energy shift ratio, spectral peak migration trend and energy diffusion degree. The distribution differences at multiple locations are calculated based on the energy shift ratio, spectral peak migration trend, and energy diffusion degree to obtain spectral shift feature sets at different locations.
[0023] In this embodiment, the original acoustic signal is preprocessed, including removing environmental background noise and low-frequency wind noise interference. Typically, bandpass filtering is used to limit the signal frequency range to 100 Hz–10 kHz to highlight the effective acoustic components generated by the vibration radiation of the insulator structure. Then, based on the acoustic-vibration coupling theory, the acoustic signal is considered as the radiation result of structural vibration in the air medium. By analyzing the spectral characteristics of sound pressure changing over time, the distribution of structural vibration energy in the frequency domain is indirectly characterized. In the specific implementation, the processed acoustic signal is segmented for analysis according to a fixed time window. The length of a single segment is generally selected as 2048 or 4096 sampling points, and a certain overlap ratio is set between segments to ensure the stability of the spectrum estimation. A window function is applied to each segment to reduce the impact of spectral leakage. Subsequently, frequency domain analysis is performed on the acoustic signal of each data segment, and its power spectral density is calculated. Finally, a stable acoustic-vibration power spectrum is obtained by averaging the results of multiple segments. Under the conditions of uniform frequency resolution and frequency band division, the acoustic-vibration power spectrum is aligned with the baseline energy distribution curve to ensure a one-to-one correspondence between the two on the frequency axis. Subsequently, the energy variation of the acoustic power spectrum relative to the baseline was calculated within multiple preset frequency bands. By comparing the ratio of actual acoustic energy to baseline energy within the same frequency band, the energy shift ratio was extracted to characterize the relative enhancement or attenuation of excitation energy in the acoustic response. In peak migration trend analysis, the focus was on the frequency positions corresponding to the main energy peaks in the acoustic power spectrum, comparing them with the corresponding peak frequencies in the baseline to analyze whether there were phenomena such as overall peak frequency shift, disappearance of peaks in local frequency bands, or the appearance of new peaks. This reflects the impact of structural stiffness changes or internal damage on vibration characteristics. By analyzing the distribution width and smoothness of the acoustic power spectrum energy along the frequency axis, the degree of energy diffusion was assessed, i.e., whether energy diffused from a concentrated distribution to a wider frequency band. This is usually closely related to internal discontinuities or changes in damping characteristics within the structure.
[0024] Based on the sensor placement, the characteristic parameters corresponding to the acoustic response signals at each location are categorized and organized to form a multi-location characteristic parameter matrix. On this basis, the distribution of the same feature at different locations is compared to analyze its spatial variation patterns. For example, it is determined whether a certain location exhibits a significantly higher energy shift ratio than other locations, or whether there is an abnormally large increase in the amplitude of local spectral peak migration. Subsequently, by calculating the relative differences of characteristic parameters at each location, distribution difference indicators reflecting local structural anomalies are extracted, thus elevating the variation characteristics of a single measurement point to a multi-location collaborative characterization result. For the analysis of energy diffusion, the focus is on its consistency or dispersion at different locations. If significant broadband energy diffusion occurs at some locations while other locations maintain a concentrated distribution, it may indicate the presence of local defects or damaged areas inside the insulator. By comprehensively organizing multiple features such as energy shift ratio, spectral peak migration trend, and energy diffusion degree in the spatial dimension, a set of spectral shift features is finally formed to comprehensively characterize the anomalous acoustic-vibration response characteristics of the insulator under controlled excitation conditions.
[0025] In this embodiment, step S3 includes the following steps: Set multiple random excitation periods; Based on the multiple random excitation cycles, the spectral offset feature set is subjected to time-series stability analysis. When the spectral offset feature set shows consistent offset differences within multiple random excitation cycles, the insulator under test is determined to be a potential structural anomaly. The response signal is subjected to acoustic vibration feature analysis and labeled as an abnormal acoustic vibration vector.
[0026] In this embodiment, to improve the reliability of insulator testing results and the ability to resist accidental interference, multiple independent random excitation cycles are set and executed on the same insulator under test. Each random excitation cycle uses the same excitation frequency range and amplitude control conditions, for example, the excitation frequency range is maintained between 100 Hz and 5 kHz, and the equivalent excitation amplitude is controlled between 100 N and 300 N. However, different random phase and amplitude distribution characteristics are introduced into the specific random excitation sequence to ensure that each excitation cycle is statistically independent. The duration of each random excitation cycle is usually set to 20 s to 40 s, and sufficient structural recovery time, such as 5 s to 10 s, is reserved between adjacent excitation cycles to avoid the residual vibration of the previous cycle affecting subsequent tests. During the experimental implementation, all excitation cycles are carried out under the same installation conditions, sensor placement positions, and acquisition parameters, and the sampling frequency is maintained at no less than 25.6 kHz to ensure the comparability of data between different excitation cycles. The spectral shift feature sets obtained under different excitation cycles are sorted according to the excitation cycle number to form a multi-cycle feature sequence, ensuring that each cycle feature is based on the same frequency band division and parameter calculation method. Based on this, the changes of key features such as energy shift ratio, spectral peak migration trend, and energy diffusion degree within multiple excitation cycles are compared and analyzed, focusing on whether these features repeatedly exhibit similar shift patterns at the same measurement location or within the same frequency band. When a certain type of spectral shift feature shows consistent direction and similar amplitude shift differences in multiple random excitation cycles, and this shift feature significantly deviates from the baseline range under healthy conditions, it can be determined that the feature has high temporal stability. Since random excitation signals are statistically unpredictable, if abnormal spectral shifts in the structural response can be stably reproduced under multiple independent excitation conditions, it usually reflects that they originate from the structural characteristics rather than accidental disturbances.
[0027] For excitation cycles identified as exhibiting abnormal offset characteristics, feature information related to the abnormal frequency band is extracted from the original vibration acceleration signal and the associated acoustic response signal. This includes energy distribution characteristics, spectral peak position characteristics, and changes in acoustic-vibration coupling strength within the corresponding frequency band. During feature extraction, the vibration signal and acoustic signal are analyzed according to a unified time window and frequency division, and normalization is used to ensure the comparability of different physical quantity characteristics within the same feature space. Subsequently, multiple acoustic-vibration features within the same measurement location and the same excitation cycle are combined to form a feature vector describing the abnormal state of the acoustic-vibration response at that location. When the spectral offset characteristics corresponding to this feature vector are consistent with the aforementioned time-series stability analysis results, it is marked as an abnormal acoustic-vibration vector. This abnormal acoustic-vibration vector not only reflects the abnormal response characteristics of the insulator under specific excitation conditions but also serves as an important input for subsequent defect type analysis, condition assessment, or historical database comparison.
[0028] In this embodiment, step S4 includes the following steps: The anomalous acoustic vibration vector is subjected to multi-scale feature decomposition to obtain multiple anomalous feature sets; the multiple anomalous feature sets specifically include spectral peak position changes, bandwidth expansion characteristics, energy attenuation slope, and the proportion of random noise. The multiple abnormal feature sets are scaled and defect-related filtered to construct a defect feature vector matrix.
[0029] In this embodiment, the anomalous acoustic vibration vector is divided into scales. The time scale is typically distinguished according to short-time response and steady-state response, for example, selecting various analysis windows with lengths of 10 ms, 50 ms, and 200 ms to respectively characterize local transient features and overall spectral change trends. The frequency scale is divided hierarchically according to the excitation frequency band, such as further subdividing the 100 Hz to 5 kHz band into low-frequency, mid-frequency, and high-frequency sub-bands. By decomposing and analyzing the anomalous acoustic vibration vector at different scales, the omission of local anomalies or broadband changes by single-scale analysis can be effectively avoided. In the multi-scale analysis, key anomaly characteristic parameters are extracted at each scale. Spectral peak position changes characterize the shift of the main resonant frequency relative to the baseline, typically described in Hertz. Bandwidth expansion reflects the trend of energy distribution shifting from concentrated to dispersed under anomaly conditions, which can be characterized by comparing the main energy bandwidth under anomaly and baseline conditions. Energy decay slope describes the rate of energy decay with frequency or time, and its variation is often related to internal structural damping or crack development. The proportion of random noise characterizes the degree of disordered response enhancement caused by the anomaly by analyzing the proportion of non-structural high-frequency components in the total energy. Dimensional unification is performed on various anomaly characteristics. Since spectral peak position changes are represented by frequency shift, bandwidth expansion characteristics by bandwidth, and energy decay slope and random noise proportion have different physical meanings and numerical ranges, each characteristic needs to be normalized or standardized to map it to a unified characteristic scale range. This prevents any one type of characteristic from dominating the comprehensive analysis due to excessively large numerical amplitudes. After achieving scale uniformity, based on the insulator structure mechanism and experimental experience, defect-related screening was performed on anomalous features. Priority was given to retaining feature parameters that are sensitive to structural anomalies and remain stable across multiple random excitation cycles, such as spectral peak migration characteristics closely related to structural stiffness changes, and bandwidth expansion and energy attenuation slope characteristics reflecting internal damage propagation. Features significantly affected by environmental noise or with poor repeatability were weakened or eliminated. After screening, the retained anomalous features were arranged and combined in an ordered manner according to measurement location, excitation cycle, and feature category to construct a defect feature vector matrix. Each row of the matrix corresponds to an anomalous acoustic-vibration sample, and each column corresponds to a selected defect-related feature.
[0030] In this embodiment, step S5 includes the following steps: The defect feature vector matrix is input into a preset deep learning architecture for unsupervised learning, and defect discrimination response feature enhancement is performed on the defect feature vector matrix to extract defect enhancement samples with different acoustic and vibration features. Based on the defect-enhanced samples, defect acoustic-vibration fingerprint matching is performed, and defect identification results are output; the defect identification results include defect type and defect severity. Based on the defect identification results, a graded diagnosis is performed, and an insulator inspection report is output.
[0031] In this embodiment, the defect feature vector matrix constructed in the previous stage is used as input data, and a preset deep learning architecture is introduced to perform unsupervised learning processing to achieve automatic enhancement of defect discrimination response features. The deep learning architecture is preferably an unsupervised model structure with feature self-organization capabilities, such as a deep network based on a multi-layer encoder-decoder mechanism, whose input layer dimension is consistent with the feature dimension of the defect feature vector matrix. During model training, manually labeled defect category information is not introduced; instead, the model learns from the distribution differences of a large number of normal and abnormal acoustic and vibration feature samples in the feature space to achieve self-learning of potential defect patterns. In specific implementation, the defect feature vector matrix is input into the model according to the sample dimension, and reasonable training parameters are set, such as setting the number of training iterations to 100-300 rounds and controlling the learning rate within the range of 0.001-0.01 to ensure that the model can converge sufficiently without overfitting. During the unsupervised learning process, the model adaptively optimizes the reconstruction error of the input features, feature compression, and recombination capabilities, so that acoustic and vibration features highly correlated with structural anomalies are significantly amplified in the hidden layers, while random noise or weakly correlated features are gradually suppressed. Through this process, the original defect feature vector is mapped to a set of defect enhancement samples that more prominently highlight the defect response characteristics, further widening the response differences of different acoustic and vibration features under defect conditions. Based on historical experimental data or standard test samples, an acoustic and vibration fingerprint database containing various typical defect states is pre-constructed. Each defect type in this fingerprint database corresponds to a set of defect enhancement feature patterns stably extracted under multiple random excitation conditions. Subsequently, the defect enhancement sample of the current insulator under test is input into the fingerprint matching module and similarity analysis is performed with various defect acoustic and vibration fingerprints in the fingerprint database, focusing on comparing the matching degree of key feature dimensions such as spectral peak position change pattern, energy distribution pattern, bandwidth expansion degree, and noise ratio. During the matching process, by comprehensively evaluating the similarity trends in different feature dimensions, the fingerprint category that best matches the characteristics of the current defect enhancement sample is determined, thereby completing the defect type determination. Based on the deviation level of the defect enhancement sample from the corresponding fingerprint benchmark in terms of feature amplitude and energy shift, the defect severity is quantitatively assessed, for example, distinguishing it as mild, moderate, or severe defects.
[0032] Based on a pre-defined diagnostic rule system, the defect identification results are compared and analyzed with operational safety thresholds. For example, minor defects are classified as allowing continued operation, moderate defects as requiring close monitoring, and severe defects as requiring repair or replacement. The diagnostic rules are typically formulated by combining insulator structural design parameters, material performance indicators, and long-term operational experience to ensure the engineering feasibility of the diagnostic conclusions. Furthermore, the diagnostic results from multiple measurement points and multiple excitation cycles are comprehensively evaluated to avoid misjudgments based on a single measurement point or a single excitation anomaly. The final insulator inspection report systematically includes basic information about the inspected object, test excitation parameters, acoustic-vibration response analysis results, defect type determination conclusions, defect severity classification, and corresponding operational recommendations.
[0033] In this embodiment, the specific steps for performing graded diagnosis based on defect identification results and outputting an insulator inspection report are as follows: Based on the defect identification results, structural integrity and operational reliability are assessed to obtain the defect safety level. Based on the defect safety level, a graded diagnosis is performed, and an insulator inspection report is output.
[0034] In this embodiment, the identified defect type and severity are used as core input parameters. Combined with the insulator's structural design parameters, material performance indicators, and actual operating conditions, the impact of defects on structural load-bearing capacity and insulation performance is comprehensively analyzed. For example, for different types of defects such as core rod cracks, shed detachment, or interface aging, their impact on mechanical strength, electrical insulation level, and long-term operational stability is assessed. During the assessment process, a reliability evaluation method based on acoustic and vibration response characteristics is introduced. Characteristic indicators reflecting changes in structural stiffness, increased energy attenuation, and enhanced random noise in the defect-enhanced samples are compared with baseline thresholds under healthy conditions to determine whether the current defect has substantially weakened structural integrity. The defect impact can be modified by considering the insulator's rated mechanical load, design life, and on-site operating environment conditions; for example, the sensitivity of the safety assessment can be appropriately increased in high-humidity and high-pollution environments. By weighted and fused evaluation results from multiple measurement points and multiple excitation cycles, the evaluation bias caused by a single abnormal feature is avoided. Finally, the evaluation results are mapped to a preset safety level system, such as dividing them into multiple defect safety levels, including safe, concern, risk and serious risk, to quantitatively or semi-quantitatively characterize the current structural safety and operational reliability level of the insulator.
[0035] In this embodiment, an insulator defect detection device is provided for performing the insulator defect detection method described above, including: The application unit is used to apply a controlled random excitation signal to the insulator under test and simultaneously acquire response signals at different positions of the insulator. The vibration analysis unit is used to perform vibration characteristic analysis on the response signal to obtain a spectral shift feature set at different locations; The deviation analysis unit is used to perform abnormal deviation analysis on the spectral offset feature set and mark it as an abnormal acoustic vibration vector. The decomposition unit is used to perform multi-scale feature decomposition on the abnormal acoustic vibration vector to construct a defect feature vector matrix. The learning unit is used to input the defect feature vector matrix into a preset deep learning architecture for unsupervised learning and output an insulator inspection report.
[0036] The present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the defect detection method for insulators described in any of the above claims.
[0037] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the defect detection method for insulators described in any of the preceding claims.
[0038] Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of the equivalents of the application are intended to be included within the invention.
[0039] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement it. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein are implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features of the invention herein.
Claims
1. A method for detecting defects in insulators, characterized in that, Includes the following steps: Step S1: Apply a controlled random excitation signal to the insulator under test and simultaneously acquire response signals at different positions of the insulator; Step S2: Perform vibration feature analysis on the response signal to obtain spectral shift feature sets at different locations; Step S3: Perform anomaly deviation analysis on the spectral offset feature set and mark it as anomaly acoustic vibration vector; Step S4: Perform multi-scale feature decomposition on the abnormal acoustic vibration vector to construct a defect feature vector matrix; Step S5: Input the defect feature vector matrix into a preset deep learning architecture for unsupervised learning and output an insulator inspection report.
2. The defect detection method for insulators according to claim 1, characterized in that, The specific steps of step S1 are as follows: A controlled random excitation signal is applied to the insulator under test, and response signals at different positions of the insulator are collected simultaneously; the response signals include vibration acceleration signals and accompanying acoustic response signals. The response signal is subjected to time synchronization and amplitude normalization processing to generate an optimized signal; Power spectral density estimation is performed on the optimized signal to extract energy distribution curves at different locations; The energy distribution curves are analyzed for morphological consistency and energy concentration trend to construct an excitation energy distribution baseline.
3. The defect detection method for insulators according to claim 1, characterized in that, The specific steps of step S2 are as follows: Vibration characteristic analysis was performed on the associated acoustic response signal to extract the acoustic power spectrum; Based on the acoustic power spectrum, the baseline of the excitation energy distribution is compared and analyzed to extract the energy shift ratio, spectral peak migration trend and energy diffusion degree. The distribution differences at multiple locations are calculated based on the energy shift ratio, spectral peak migration trend, and energy diffusion degree to obtain spectral shift feature sets at different locations.
4. The defect detection method for insulators according to claim 1, characterized in that, Step S3 is as follows: Set multiple random excitation periods; Based on the multiple random excitation cycles, the spectral offset feature set is subjected to time-series stability analysis. When the spectral offset feature set shows consistent offset differences within multiple random excitation cycles, the insulator under test is determined to be a potential structural anomaly. The response signal is subjected to acoustic vibration feature analysis and labeled as an abnormal acoustic vibration vector.
5. The defect detection method for insulators according to claim 1, characterized in that, The specific steps of step S4 are as follows: The anomalous acoustic vibration vector is subjected to multi-scale feature decomposition to obtain multiple anomalous feature sets; the multiple anomalous feature sets specifically include spectral peak position changes, bandwidth expansion characteristics, energy attenuation slope, and the proportion of random noise. The multiple abnormal feature sets are scaled and defect-related filtered to construct a defect feature vector matrix.
6. The defect detection method for insulators according to claim 1, characterized in that, The specific steps of step S5 are as follows: The defect feature vector matrix is input into a preset deep learning architecture for unsupervised learning, and defect discrimination response feature enhancement is performed on the defect feature vector matrix to extract defect enhancement samples with different acoustic and vibration features. Based on the defect-enhanced samples, defect acoustic-vibration fingerprint matching is performed, and defect identification results are output. The defect identification results include the defect type and the defect severity. Based on the defect identification results, a graded diagnosis is performed, and an insulator inspection report is output.
7. The defect detection method for insulators according to claim 6, characterized in that, The specific steps for performing graded diagnosis based on defect identification results and outputting an insulator inspection report are as follows: Based on the defect identification results, structural integrity and operational reliability are assessed to obtain the defect safety level. Based on the defect safety level, a graded diagnosis is performed, and an insulator inspection report is output.
8. A defect detection device for insulators, characterized in that, A method for performing defect detection of an insulator as described in claim 1, comprising: The application unit is used to apply a controlled random excitation signal to the insulator under test and simultaneously acquire response signals at different positions of the insulator. The vibration analysis unit is used to perform vibration characteristic analysis on the response signal to obtain a spectral shift feature set at different locations; The deviation analysis unit is used to perform abnormal deviation analysis on the spectral offset feature set and mark it as an abnormal acoustic vibration vector. The decomposition unit is used to perform multi-scale feature decomposition on the abnormal acoustic vibration vector to construct a defect feature vector matrix. The learning unit is used to input the defect feature vector matrix into a preset deep learning architecture for unsupervised learning and output an insulator inspection report.
9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the insulator defect detection method according to any one of claims 1 to 7.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the insulator defect detection method according to any one of claims 1 to 7.