Insulator micro-damage early warning system and method based on acoustoelectric time sequence characteristics

Abstract

The invention provides an insulator micro-damage early warning system and method based on acoustoelectric time sequence characteristics, and relates to the technical field of power transmission line on-line monitoring. Comprising the steps that acoustic signal data and electrical signal data generated in the operation process of an insulator string are collected in real time, and adaptive noise filtering processing is carried out on the acoustic signal data and the electrical signal data to obtain noise-reduced acoustic signal data and noise-reduced electrical signal data; performing feature extraction on the noise-reduced acoustic signal data and the noise-reduced electrical signal data to obtain a time domain synchronism feature, a frequency domain relevance feature and a time-frequency domain feature, and combining the features to form an acoustic-electric joint feature vector; inputting the time sequence convolution attention network model to obtain the operation state of the insulator string and the corresponding probability; and generating a micro-damage early warning signal when the operation state is an incubation period defect and the probability exceeds a preset threshold value in a plurality of continuous time windows. The method has the beneficial effects that real insulator defect signals and random environment noise can be accurately distinguished, and the reliability of signal analysis under complex working conditions is guaranteed.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of transmission line online monitoring, and in particular to an insulator micro-damage early warning system and method based on acoustic-electric time sequence characteristics. BACKGROUND

[0002] Insulators are the key components of overhead transmission lines, and their insulation performance is directly related to the safe and stable operation of the entire power system. Insulators are exposed to complex natural environments for a long time, and are subjected to the continuous effects of mechanical stress, high-voltage electric field, sudden temperature changes, environmental pollution, and rain, snow, and hail, etc. Therefore, performance degradation is inevitable. Common degradation forms include micro-cracks caused by internal stress accumulation of tempered glass insulators, and early fiber breakage of composite insulator core rods, etc. These early and micro physical damages are referred to as "latent period defects", and their external characteristics are extremely weak before they develop into serious faults (such as "self-explosion" of glass insulators or brittle fracture of composite insulator core rods), which are difficult to be effectively discovered by traditional detection methods.

[0003] At present, the online monitoring technology for insulator degradation mainly relies on the detection of a single physical quantity, which has obvious technical bottlenecks: 1) Electrical signal detection method: such as detecting leakage current, ultra-high frequency (UHF) or radio frequency (RF) signals generated by partial discharge, ultraviolet photon counting, etc. These methods are relatively effective for discharge defects that have developed to a certain extent. However, in the latent period, for example, the expansion of glass insulator micro-cracks in the closed state, or the breakage of a small number of fibers of the composite insulator core rod under the protection of a well-hydrophobic sheath, the generated electric pulse signal is extremely weak, and the signal-to-noise ratio is extremely low, completely submerged in the strong background noise generated by line corona, radio interference, and bad weather (rain, fog, wind), etc. In this case, the traditional threshold discrimination method based on signal amplitude or pulse count is prone to false negatives and false positives, and cannot achieve reliable latent period early warning.

[0004] 2) Acoustic detection method: such as using acoustic emission (Acoustic Emission, AE) or ultrasonic technology to detect the acoustic energy released when the internal crack of the material expands. Acoustic emission technology is very sensitive to crack generation and expansion. However, in the strong noise environment of the transmission line, wind noise, rain noise, bird calls, and the sound produced by corona discharge, etc. will cause serious interference to the weak defect acoustic signal. It is also difficult to accurately distinguish between defect signals and environmental noise based on acoustic signals alone.

[0005] 3) Optical and other detection methods: such as infrared temperature measurement, ultraviolet imaging, X-ray or terahertz imaging, etc. These methods are either too expensive to be popularized on a large scale, or can only detect surface defects or specific types of faults (such as heating), and have limited detection capability for internal micro-cracks and other latent mechanical damage.

[0006] In summary, the core dilemma of the prior art is the limitation of a single modal information source. Whether it is an electrical signal or an acoustic signal, it is a weak signal with extremely low signal-to-noise ratio in the latent period, which is easily disturbed by environmental noise. The method of relying on a single signal source for threshold judgment cannot fundamentally solve the problem of distinguishing true from false, resulting in the reliability of the early warning cannot meet the actual demand.

[0007] Therefore, there is an urgent need for a latent period early warning technology that can effectively extract weak defect features from a strong noise background and has high reliability. SUMMARY

[0008] In view of the problems in the prior art, the present application provides an insulator micro-damage early warning system based on acoustic-electric timing characteristics, which is deployed on an insulator string of a power transmission tower, and comprises: a data acquisition module for acquiring acoustic signal data generated by the insulator string during operation and electrical signal data flowing through the insulator string in real time; a data preprocessing module connected to the data acquisition module for performing adaptive noise filtering on the acoustic signal data and the electrical signal data respectively to obtain denoised acoustic signal data and denoised electrical signal data; a feature extraction module connected to the data preprocessing module for extracting time domain synchronization features, frequency domain correlation features and time-frequency domain features from the denoised acoustic signal data and the denoised electrical signal data, and combining the time domain synchronization features, the frequency domain correlation features and the time-frequency domain features to form an acoustic-electric joint feature vector; a defect recognition module connected to the feature extraction module for inputting the acoustic-electric joint feature vector into a pre-trained timing convolution attention network model to obtain the running state of the insulator string and its corresponding probability; and an early warning generation module connected to the defect recognition module for generating a micro-damage early warning signal and uploading it to a monitoring center when the running state is a latent period defect and the corresponding probability exceeds a preset threshold in a plurality of consecutive time windows.

[0009] Preferably, the data acquisition module comprises a piezoelectric ceramic sensor arranged on a metal fitting at the bottom of the insulator string and a current transformer arranged on a lead wire of the insulator string, and a dual-channel synchronous data acquisition card connected to the piezoelectric ceramic sensor and the current transformer respectively; the dual-channel synchronous data acquisition card is used to control the piezoelectric ceramic sensor and the current transformer to synchronously acquire the acoustic signal data and the pulse current signal flowing through the insulator string as the electrical signal data respectively.

[0010] Preferably, the dual-channel synchronous data acquisition card is integrated with a GPS time-providing module, which is used to configure a synchronous time stamp for each acquisition of the acoustic signal data and the electrical signal data.

[0011] Preferably, the data preprocessing module comprises: a first processing unit, which is used to perform multi-layer wavelet decomposition on the acoustic signal data and the electrical signal data respectively to obtain multi-layer approximation coefficients and detail coefficients; a second processing unit connected to the first processing unit, which is used to perform denoising on each layer of the detail coefficients respectively to obtain denoised detail coefficients; a third processing unit connected to the first processing unit and the second processing unit respectively, which is used to perform wavelet reconstruction on each layer of the approximation coefficients and the denoised detail coefficients to obtain preliminary denoised acoustic signal data and preliminary denoised electrical signal data; and a fourth processing unit connected to the third processing unit, which is used to input the preliminary denoised acoustic signal data and the preliminary denoised electrical signal data into a Wiener filter respectively to obtain the denoised acoustic signal data and the denoised electrical signal data.

[0012] Preferably, the time-domain synchrony feature comprises a cross-correlation peak, a time delay and a synchronization entropy; and the feature extraction module comprises: a first extraction unit, which is used to calculate a normalized cross-correlation function of the denoised acoustic signal data and the denoised electrical signal data in a preset short-time window, and extract the cross-correlation peak and the time delay of the normalized cross-correlation function; and a second extraction unit, which is used to calculate the synchronization entropy of a short-time energy sequence of the denoised acoustic signal data and the denoised electrical signal data.

[0013] Preferably, the frequency domain correlation feature comprises a spectrum shape mutual information and a key frequency band energy ratio; the feature extraction module comprises: a third extraction unit, configured to perform fast Fourier transform on the denoised acoustic signal data and the denoised electrical signal data respectively to obtain an acoustic signal spectrum and an electrical signal spectrum, and calculate mutual information of the acoustic signal spectrum and the electrical signal spectrum as the spectrum shape mutual information; and a fourth extraction unit, connected to the third extraction unit, configured to extract a first frequency band energy and a second frequency band energy based on a first frequency band range of the defect signal in the acoustic signal spectrum and a second frequency band range of the defect signal in the electrical signal spectrum respectively, and calculate a first proportion of the first frequency band energy in total energy of the acoustic signal spectrum and a second proportion of the second frequency band energy in total energy of the electrical signal spectrum as the key frequency band energy ratio.

[0014] Preferably, the time-frequency domain feature comprises a wavelet packet energy spectrum synchronism feature and a time-frequency coincidence degree; the feature extraction module comprises: a fifth extraction unit, configured to perform wavelet packet decomposition on the denoised acoustic signal data and the denoised electrical signal data respectively to obtain corresponding acoustic wavelet packet energy spectrum and electrical wavelet packet energy spectrum, and calculate energy distribution similarity of the acoustic wavelet packet energy spectrum and the electrical wavelet packet energy spectrum under the same time and different frequency bands as the wavelet packet energy spectrum synchronism feature; and a sixth extraction unit, configured to obtain instantaneous frequency and amplitude of the denoised acoustic signal data and the denoised electrical signal data by using Hilbert-Huang transform, to construct an acoustic signal Hilbert spectrum and an electrical signal Hilbert spectrum respectively, and further calculate a time-frequency coincidence degree of the acoustic signal Hilbert spectrum and the electrical signal Hilbert spectrum as the time-frequency coincidence degree.

[0015] Preferably, the network structure of the time sequence convolution attention network model comprises an input layer, a one-dimensional time sequence convolution layer, a self-attention mechanism layer and an output layer connected in sequence.

[0016] Preferably, the early warning generation module comprises a pre-warning counter, configured to count once when the running state is a latent defect and the corresponding probability exceeds a preset threshold, and generate the micro-damage early warning signal when the counting result exceeds a preset number of times within a preset time length.

[0017] The application also provides an insulator micro-damage early warning method based on acoustic-electric timing characteristics, which is applied to the insulator micro-damage early warning system, and comprises the following steps: S1, the insulator micro-damage early warning system collects acoustic signal data and electrical signal data generated by the insulator string during operation; S2, the insulator micro-damage early warning system respectively performs adaptive noise filtering on the acoustic signal data and the electrical signal data to obtain denoised acoustic signal data and denoised electrical signal data; S3, the insulator micro-damage early warning system extracts time domain synchronism features, frequency domain correlation features and time-frequency domain features from the denoised acoustic signal data and the denoised electrical signal data, and combines the time domain synchronism features, the frequency domain correlation features and the time-frequency domain features to form an acoustic-electric joint feature vector; S4, the insulator micro-damage early warning system inputs the acoustic-electric joint feature vector into a pre-trained timing convolution attention network model to obtain the operating state of the insulator string and the corresponding probability; and S5, the insulator micro-damage early warning system generates a micro-damage early warning signal and uploads it to a monitoring center when the operating state is a latent defect and the corresponding probability exceeds a preset threshold in continuous multiple time windows.

[0018] The above technical solution has the following advantages or beneficial effects: 1) By synchronously acquiring the acoustic signal and the electrical signal of the insulator string, and deeply mining the time domain synchronism features, the frequency domain correlation features and the time-frequency domain fusion features of the two signals, the limitations of traditional single modal signal monitoring are broken through, the cross-modal fixed correlation attributes of acoustic and electrical signals are utilized, the real insulator defect signal and random environmental noise (such as wind and rain interference, external electromagnetic interference, surrounding mechanical noise, etc.) can be accurately distinguished, the signal-to-noise ratio is improved from the bottom logic of signal recognition, the influence of environmental interference on the monitoring result is effectively avoided, and the reliability of signal analysis under complex working conditions is ensured; 2) The adaptive noise filtering can effectively retain the weak acoustic signal (such as the ultrasonic signal of early fiber fracture of a composite insulator core rod) and the weak electrical signal (such as the partial discharge signal before self-explosion of a tempered glass insulator) generated during the insulator micro-damage process; combined with the deep mining capability of the timing convolution attention network model for the acoustic-electric joint feature vector, the operating state of the insulator fault germination stage (such as the latent period before self-explosion of a tempered glass insulator and the early fiber fracture stage of a composite insulator core rod) can be accurately identified, the early warning window is advanced from the traditional fault occurrence to the defect latent period, sufficient emergency disposal time is provided for power transmission line operation and maintenance, and the risk of fault expansion is significantly reduced; 3) This system is deployed on the insulator strings of transmission towers to achieve localized and efficient computation of data acquisition, signal preprocessing, feature extraction and defect identification. It features low power consumption, low cost, and ease of deployment and promotion, meeting the practical requirements of large-scale online monitoring. 4) Adaptive filtering technology is used to dynamically filter out environmental noise, and a specially designed deep learning network model is used to achieve in-depth mining and intelligent decision-making of the combined sound and electricity features. This effectively avoids judgment bias caused by single feature analysis, thereby significantly reducing the false alarm rate and the false alarm rate, ensuring the credibility of early warning information, and providing reliable support for the precise operation and maintenance of transmission lines. Attached Figure Description

[0019] Figure 1 A schematic diagram of the structure of an insulator micro-loss early warning system based on acoustic-electric timing characteristics is shown in a preferred embodiment of the present invention. Figure 2 This is a flowchart illustrating a preferred embodiment of the present invention for an insulator micro-loss early warning method based on acoustic-electric timing characteristics. Detailed Implementation

[0020] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. The present invention is not limited to this embodiment; other embodiments that conform to the spirit of the present invention may also fall within the scope of the present invention.

[0021] In a preferred embodiment of the present invention, based on the aforementioned problems existing in the prior art, an insulator micro-damage early warning system based on acoustic-electric timing characteristics is provided. This system is deployed on the insulator strings of transmission towers to achieve accurate identification and early warning of micro-damage defects during the insulator's latency period. Figure 1 As shown, the system includes: a data acquisition module 1, used to acquire acoustic signal data and electrical signal data flowing through the insulator string in real time during operation; a data preprocessing module 2, connected to the data acquisition module 1, used to perform adaptive noise filtering on the acoustic signal data and electrical signal data respectively to obtain denoised acoustic signal data and denoised electrical signal data; a feature extraction module 3, connected to the data preprocessing module 2, used to extract features from the denoised acoustic signal data and denoised electrical signal data to obtain time-domain synchronization features, frequency-domain correlation features, and time-frequency domain features, and combine the time-domain synchronization features, frequency-domain correlation features, and time-frequency domain features to form a joint acoustic-electric feature vector; a defect identification module 4, connected to the feature extraction module 3, used to input the joint acoustic-electric feature vector into a pre-trained temporal convolutional attention network model to obtain the operating state of the insulator string and its corresponding probability; and an early warning generation module 5, connected to the defect identification module 4, used to generate a minor damage early warning signal and upload it to the monitoring center when the operating state is a latent defect and the corresponding probability exceeds a preset threshold for multiple consecutive time windows.

[0022] In a preferred embodiment of the present invention, the data acquisition module 1 includes a piezoelectric ceramic sensor 11 deployed on a metal fitting at the bottom of the insulator string and a current transformer 12 deployed on the leads of the insulator string, as well as a dual-channel synchronous data acquisition card 13 connected to the piezoelectric ceramic sensor 11 and the current transformer 12 respectively; the dual-channel synchronous data acquisition card 13 is used to control the piezoelectric ceramic sensor 11 and the current transformer 12 to synchronously acquire acoustic signal data and pulse current signal flowing through the insulator string as electrical signal data respectively.

[0023] Specifically, in this embodiment, a low-cost, high-sensitivity sensor combination is used to synchronously acquire acoustic and electrical signals: 1) For acoustic signal acquisition: A broadband piezoelectric ceramic (PZT) ultrasonic sensor is used, with a preferred center frequency range of 40 kHz to 200 kHz. This frequency band can effectively avoid low-frequency environmental noise (usually below 20 kHz) generated by line corona discharge and wind and rain, while also exhibiting high sensitivity to acoustic emission signals generated by microscopic damage such as microcrack propagation and core rod fiber breakage inside the insulator. During deployment, it is preferable to use magnetic attraction or a special clamp to tightly attach the piezoelectric ceramic sensor 11 to the metal fitting surface at the bottom of the insulator, ensuring effective acoustic coupling between the sensor and the fitting and reducing energy attenuation of sound waves during propagation.

[0024] 2) For electrical signal acquisition: An open-type high-frequency current transformer (HFCT) is selected as the current transformer 12. Its operating bandwidth range is 1MHz to 100MHz, which can accurately detect weak high-frequency pulse currents flowing through the insulator string leads (such pulse currents are generated by partial discharge accompanying insulator micro-damage). During deployment, the HFCT 12 is directly snap-fitted onto the insulator string leads without disassembling the wiring, achieving power-free deployment and improving the safety and convenience of on-site construction. If the electromagnetic interference at the site is strong, a specially designed UHF antenna can be used to further enhance the anti-interference acquisition capability of electrical signals.

[0025] Furthermore, the signal output terminals of the piezoelectric ceramic sensor 11 and the current transformer 12 are respectively connected to the two signal input channels of the dual-channel synchronous data acquisition card 13 via shielded signal lines. This acquisition card enables synchronous control and data acquisition of the two signals. The core component of the dual-channel synchronous data acquisition card 13 is a synchronous controller with nanosecond-level synchronization accuracy, ensuring that the sampling operations of acoustic and electrical signals are strictly aligned in time, with timing deviation controlled within 10ns, providing a foundation for subsequent synchronization analysis of acoustic and electrical signals. Simultaneously, the dual-channel synchronous data acquisition card 13 integrates a GPS timing module 131. This module receives high-precision UTC time signals broadcast by satellites in real time, calibrates the internal clock of the acquisition card, and, after each sampling, adds a unified high-precision GPS timestamp to the corresponding acoustic and electrical signal data, denoted as [time stamp]. This enables absolute time correlation between the two signals.

[0026] Among them, sampling frequency The setting follows the Nyquist sampling theorem, and the specific value is more than twice the highest frequency of the electrical signal. In this embodiment, it is preferably set to 200MSa / s, which can completely preserve the high-frequency pulse characteristics in the electrical signal and avoid feature loss caused by signal aliasing.

[0027] The raw acoustic and electrical signal data collected above typically contain a large amount of noise and need to be preprocessed by the data preprocessing module 2. In a preferred embodiment of the present invention, the data preprocessing module 2 includes: a first processing unit 21, used to perform multi-level wavelet decomposition on the acoustic and electrical signal data respectively to obtain multi-level approximation coefficients and detail coefficients; a second processing unit 22, connected to the first processing unit 21, used to denoise each level of detail coefficients to obtain denoised detail coefficients; a third processing unit 23, connected to the first processing unit 21 and the second processing unit 22 respectively, used to perform wavelet reconstruction on each level of approximation coefficients and denoised detail coefficients to obtain pre-denoised acoustic signal data and pre-denoised electrical signal data; and a fourth processing unit 24, connected to the third processing unit 23, used to input the pre-denoised acoustic signal data and pre-denoised electrical signal data into a Wiener filter for filtering to obtain denoised acoustic signal data and denoised electrical signal data.

[0028] Specifically, for changing environmental noise (especially in rainy or foggy weather), this embodiment employs a wavelet-based adaptive Wiener filter algorithm to adaptively filter acoustic and electrical signal data, including: The first processing unit 21 receives the acoustic signal data segment A(t) and the electrical signal data segment E(t) output by the data acquisition module 1, and performs multi-level wavelet decomposition on the two signals to obtain approximation coefficients and detail coefficients at different scales. Among them, the approximation coefficients characterize the low-frequency backbone features of the signal (mainly the effective signal components), and the detail coefficients characterize the high-frequency details of the signal (mainly the environmental noise components).

[0029] The second processing unit 22 preferably first calculates the signal-to-noise ratio (SNR) of the detail coefficients of each layer, and then dynamically selects a denoising strategy based on the SNR estimation results: When the signal-to-noise ratio (SNR) is greater than or equal to the preset SNR value (e.g., 30dB, where the noise ratio is small), a soft thresholding function is used to denoise the detail coefficients, which can filter out noise while retaining some useful high-frequency details; when the SNR is less than the preset SNR value (e.g., 30dB, where the noise ratio is large), a hard thresholding function is used to denoise the detail coefficients, achieving complete noise removal.

[0030] The specific values ​​of the soft / hard thresholds are preferably obtained by adaptive calculation based on the variance of the detail coefficients of each layer.

[0031] The third processing unit 23 performs wavelet reconstruction on the unprocessed approximation coefficients (low-frequency effective components) of each layer and the denoised detail coefficients output by the second processing unit 22 to obtain the acoustic signal data and electrical signal data after preliminary denoising.

[0032] The fourth processing unit 24 inputs the acoustic signal data and electrical signal data after preliminary noise reduction into the Wiener filter for secondary filtering. The parameters (such as the order) of the Wiener filter are dynamically adjusted according to the statistical characteristics (such as the autocorrelation function) of the signal to achieve optimal filtering of residual stationary noise. Finally, the noise-reduced acoustic signal data and electrical signal data are output.

[0033] Furthermore, feature extraction module 3 extracts characteristic parameters that characterize the micro-loss state of the insulator from the denoised acoustic and electrical signals. In a preferred embodiment of the present invention, the time-domain synchronization features include cross-correlation peak value, time delay, and synchronization entropy. Feature extraction module 3 includes: a first extraction unit 31, used to calculate the normalized cross-correlation function of the denoised acoustic signal data and the denoised electrical signal data within a preset short time window, and extract the cross-correlation peak value and time delay of the normalized cross-correlation function; and a second extraction unit 32, used to calculate the synchronization entropy of the short-time energy sequence of the denoised acoustic signal data and the denoised electrical signal data.

[0034] Specifically, in this embodiment, the aforementioned preset short window can be set according to requirements, such as being set to 10ms, corresponding to the normalized cross-correlation function. The expression is as follows:

[0035] in, The acoustic signal data after noise reduction. For the noise-reduced electrical signal data, For time delay.

[0036] Extracting the maximum peak value of the normalized cross-correlation function and corresponding delay Real defect signals originate from the same physical event; therefore, their acoustic and electrical signals should have a very high correlation. Approaching 1) and a fixed physical propagation delay Environmental noise originates from different sources, and their correlation is usually very low.

[0037] Furthermore, the second extraction unit 32 treats the short-time energy sequences of the acoustic and electrical signals as two random variables for comparison, and preferably calculates their synchronization entropy using the Shannon entropy formula to quantify the degree of synchronization between the two in energy bursts.

[0038] In a preferred embodiment of the present invention, the frequency domain correlation features include spectral morphology mutual information and key frequency band energy ratio; the feature extraction module 3 includes: a third extraction unit 33, used to perform fast Fourier transform on the denoised acoustic signal data and the denoised electrical signal data respectively to obtain the acoustic signal spectrum and the electrical signal spectrum, treating the two spectra as two probability distributions, and calculating the mutual information of the acoustic signal spectrum and the electrical signal spectrum as spectral morphology mutual information through the mutual information formula, which measures the similarity of the spectrum shape. The larger the spectral morphology mutual information value, the stronger the frequency domain feature correlation of the two signals; a fourth extraction unit 34, connected to the third extraction unit 33, used to extract the energy in the first frequency band and the energy in the second frequency band based on the preset first frequency band range of the defect signal in the acoustic signal spectrum and the second frequency band range in the electrical signal spectrum respectively, and calculate the first proportion of the energy in the first frequency band in the total energy in the acoustic signal spectrum and the second proportion of the energy in the second frequency band in the total energy in the electrical signal spectrum as the key frequency band energy ratio.

[0039] Specifically, in this embodiment, it is preferable to preset the characteristic frequency bands of the defect signal in the acoustic and electrical spectrum based on experimental data, such as the first frequency band range of 60-120kHz and the second frequency band range of 5-20MHz.

[0040] In a preferred embodiment of the present invention, the time-frequency domain features include wavelet packet energy spectrum synchronization features and time-frequency overlap; the feature extraction module 3 includes: a fifth extraction unit 35, used to perform wavelet packet decomposition on the denoised acoustic signal data and the denoised electrical signal data respectively to obtain the corresponding acoustic wavelet packet energy spectrum and electrical wavelet packet energy spectrum, and calculate the energy distribution similarity of the acoustic wavelet packet energy spectrum and electrical wavelet packet energy spectrum at the same time and in different frequency bands as the wavelet packet energy spectrum synchronization feature; a sixth extraction unit 36, used to obtain the instantaneous frequency and amplitude of the denoised acoustic signal data and the denoised electrical signal data using Hilbert-Huang transform, so as to construct the acoustic signal Hilbert spectrum and the electrical signal Hilbert spectrum respectively, and then calculate the overlap of the acoustic signal Hilbert spectrum and the electrical signal Hilbert spectrum in time and frequency as the time-frequency overlap.

[0041] Finally, the feature extraction module 3 normalizes the seven types of feature parameters extracted above—cross-correlation peak, time delay, synchronization energy entropy, spectral morphology mutual information, key frequency band energy ratio, wavelet packet energy spectrum synchronization features, and time-frequency overlap—to the [0,1] interval, and combines them to form a high-dimensional acoustic-electric joint feature vector AEJFV (AEJFV), which serves as the input to the subsequent deep learning model.

[0042] After forming the aforementioned acoustic-electric joint feature vector, deep analysis is performed through the defect identification module 4 to output the operating state of the insulator string and its corresponding probability. Specifically, this invention designs a deep learning model called Temporal Convolutional Attention Network (TCAN) specifically for processing and identifying the constructed acoustic-electric joint feature vector sequence. The network structure of the Temporal Convolutional Attention Network model includes: 1) Input layer: The input is continuous A sequence of acoustic and electronic joint feature vectors at each time step .

[0043] 2) One-dimensional temporal convolutional layer (1D-TCN): This layer uses multiple layers of causal convolution and dilated convolution to capture the long-term dependencies of feature vectors over time. Causal convolution ensures that the model uses only past information when making predictions, while dilated convolution achieves a large receptive field with fewer parameters, effectively learning the temporal evolution patterns of signals.

[0044] 3) Self-Attention Layer: Following the TCN layer, a self-attention mechanism is introduced. This layer calculates the mutual importance between feature vectors at different time steps in the sequence and performs a weighted sum. This allows the model to focus on the key time points in the sequence that best indicate the occurrence of defects (e.g., a highly correlated acoustic-electric pulse pair) and assign them higher weights, thereby effectively suppressing the interference of random noise time points. The calculation method is as follows:

[0045] in, These represent the query vector, key vector, and value vector, respectively, which are obtained from the output of TCN through a linear transformation.

[0046] 4) Output layer: After passing through a fully connected layer and a Softmax activation function, the output layer finally outputs the probability that the signal segment belongs to "normal state", "latency defect" or "interference noise".

[0047] Preferably, known defect signal samples and noise samples are used to generate a large amount of training data through data augmentation methods such as mixing, superposition, and noise addition to construct a training dataset and improve the model's generalization ability. The TCAN model is then trained using the labeled dataset under supervised learning conditions. The trained model is deployed in an edge computing module on the tower. This module processes the collected data in real time and outputs the current operating status of the insulator string (normal state, latent defect, interference noise) and its corresponding probability.

[0048] After identifying the current operating status of the insulator string, the system also includes a defect warning through the warning generation module 5. In a preferred embodiment of the present invention, the warning generation module 5 includes a warning counter 51, which is used to count once when the operating status is a latent defect and the corresponding probability exceeds a preset threshold, and generate a minor loss warning signal when the counting result exceeds a preset number of times within a preset time.

[0049] Specifically, in this embodiment, the minor damage early warning signal preferably includes information such as the tower number where the insulator is located, the time of the defect occurrence (based on GPS timestamp), and the probability of the defect; the early warning signal is uploaded to the power grid monitoring center through a LoRa wireless communication module (in remote areas) or a 4G / 5G module (in areas with good signal coverage), and at the same time triggers a local audible and visual alarm (optional) to remind maintenance personnel to handle the situation in a timely manner.

[0050] This invention also provides a method for early warning of minor insulator damage based on acoustic-electric timing characteristics, applicable to the aforementioned early warning system for minor insulator damage, such as... Figure 2As shown, the insulator minor damage early warning method includes: Step S1, the insulator minor damage early warning system collects acoustic signal data and electrical signal data flowing through the insulator string in real time during operation; Step S2, the insulator minor damage early warning system performs adaptive noise filtering on the acoustic signal data and electrical signal data respectively to obtain denoised acoustic signal data and denoised electrical signal data; Step S3, the insulator minor damage early warning system extracts features from the denoised acoustic signal data and denoised electrical signal data to obtain time-domain synchronization features, frequency-domain correlation features, and time-frequency domain features, and combines the time-domain synchronization features, frequency-domain correlation features, and time-frequency domain features to form a joint acoustic-electric feature vector; Step S4, the insulator minor damage early warning system inputs the joint acoustic-electric feature vector into a pre-trained temporal convolutional attention network model to obtain the operating state of the insulator string and its corresponding probability; Step S5, the insulator minor damage early warning system generates a minor damage early warning signal and uploads it to the monitoring center when the operating state is a latent defect and the corresponding probability exceeds a preset threshold for multiple consecutive time windows.

[0051] Example: Early warning of latent microcracks in 110kV tempered glass insulator strings Step 1: System Hardware Deployment The monitoring terminal of this invention is deployed on the insulator string of a 110kV transmission tower: 1) A PZT piezoelectric ceramic ultrasonic sensor with a center frequency of 150kHz is attached to the metal cap of the bottommost piece of the insulator string using a strong magnet base.

[0052] 2) Insert an open-type HFCT (bandwidth 5-50MHz) onto the conductor leading from the crossarm to the insulator string.

[0053] 3) Connect the signal lines of the PZT sensor and HFCT to CH1 and CH2 of the dual-channel synchronous data acquisition module, respectively. This module has a built-in GPS timing module and wireless communication via a 4G module. The entire terminal is powered by a small solar panel and a battery.

[0054] Step Two: Data Acquisition and Transmission Set the sampling frequency of the acquisition card to 100 MSa / s, and collect data of 10 milliseconds (ms) in length as a data frame each time.

[0055] The edge computing unit within the terminal (such as an NVIDIA Jetson Nano or Raspberry Pi 4B) triggers a data acquisition once per second.

[0056] Step 3: Real-time analysis and early warning at the edge The core algorithm program of this invention runs on the edge computing unit.

[0057] 1. Receiving a data frame: The unit receives a 10ms acoustic signal data frame. and electrical signal data frames .

[0058] 2. Perform preprocessing: Call the pre-configured adaptive Wiener filter program based on wavelet thresholding to perform preprocessing. and Noise reduction is performed to obtain and For example, a 5-level decomposition is performed using the db4 wavelet, and the hard threshold is dynamically adjusted based on the calculated signal-to-noise ratio.

[0059] 3. Extract joint feature vector: 1) Calculation and The normalized cross-correlation function yields the peak value. and delay Assuming we get , .

[0060] 2) Calculate the synchronization entropy of the short-time energy sequences of the two signals, and obtain... .

[0061] Perform FFT transformation on the signal to calculate frequency domain characteristics such as the energy ratio in the acoustic characteristic band (80-160kHz) and the electrical characteristic band (10-30MHz).

[0062] 3) Combine these features into a 20-dimensional feature vector. .

[0063] 4. Input the TCAN model for recognition: The edge computing unit maintains a first-in-first-out (FIFO) queue of length 10 to store the 10 most recent feature vectors. The current feature vector is then... Push into the queue to form a The feature matrix is ​​used as the input to the TCAN model.

[0064] The TCAN model computes this sequence. The model's temporal convolutional layers capture the changing trends of features within these 10 seconds, while the self-attention layer identifies one of the vectors... and It was significantly higher than at other times, thus giving it a higher weight.

[0065] Finally, the result given by the Softmax output layer is: {"Normal state": 0.1, "Late latency defect": 0.85, "Interference noise": 0.05}.

[0066] 5. Generate early warning decisions: The program sets up an alert counter, alert_count. Since the probability of "latency defect" (0.85) exceeds the preset threshold (e.g., 0.7), alert_count is incremented by 1.

[0067] The system continuously monitors the system. If the alert_count accumulates more than 5 times within 1 minute, the system determines it as a valid latent defect event.

[0068] The terminal immediately encapsulates the warning information, including the device ID, GPS location, timestamp, and feature vectors of the data frames that triggered the warning, and sends it to the backend monitoring center via the 4G network.

[0069] 6. Background analysis and confirmation: After receiving an alert, engineers at the monitoring center can retrieve historical data and detailed feature vectors from the terminal for verification. For example, if engineers find that the terminal experienced three similar weak alert events in the past week, and the timing of these events coincided with peak line load periods, the reliability of the alerts can be further confirmed.

[0070] The maintenance department dispatched a drone carrying a high-definition camera to the tower for visual inspection, or used portable ultrasonic imaging equipment to conduct a focused inspection of the insulator string during the next power outage maintenance. Ultimately, they discovered a tiny crack inside one of the glass insulators and successfully replaced it before it "self-exploded," thus avoiding a line fault.

[0071] The above description is merely a preferred embodiment of the present invention and does not limit the implementation and protection scope of the present invention. Those skilled in the art should realize that any equivalent substitutions and obvious changes made using the content of this specification and illustrations should be included within the protection scope of the present invention.

Claims

1. An insulator micro-loss early warning system based on acoustic-electrical timing characteristics, characterized in that, Deployed on an insulator string on a transmission tower, the system includes: a data acquisition module for real-time acquisition of acoustic signal data generated by the insulator string during operation and electrical signal data flowing through the insulator string; a data preprocessing module connected to the data acquisition module for adaptive noise filtering of the acoustic signal data and the electrical signal data to obtain denoised acoustic signal data and denoised electrical signal data; and a feature extraction module connected to the data preprocessing module for feature extraction of the denoised acoustic signal data and the denoised electrical signal data to obtain time-domain synchronization features and frequency characteristics. The system includes a time-domain correlation feature and a time-frequency domain feature, and combines the time-domain synchronization feature, the frequency-domain correlation feature, and the time-frequency domain feature to form a joint acoustic-electric feature vector; a defect identification module, connected to the feature extraction module, is used to input the joint acoustic-electric feature vector into a pre-trained temporal convolutional attention network model to obtain the operating state of the insulator string and its corresponding probability; an early warning generation module, connected to the defect identification module, is used to generate a minor damage early warning signal and upload it to the monitoring center when the operating state is a latent defect and the corresponding probability exceeds a preset threshold in multiple consecutive time windows.

2. The insulator micro-damage early warning system according to claim 1, characterized in that, The data acquisition module includes a piezoelectric ceramic sensor deployed on a metal fitting at the bottom of the insulator string and a current transformer deployed on the leads of the insulator string, as well as a dual-channel synchronous data acquisition card connected to the piezoelectric ceramic sensor and the current transformer respectively; the dual-channel synchronous data acquisition card is used to control the piezoelectric ceramic sensor and the current transformer to synchronously acquire the acoustic signal data and the pulse current signal flowing through the insulator string respectively as the electrical signal data.

3. The insulator micro-damage early warning system according to claim 2, characterized in that, The dual-channel synchronous data acquisition card integrates a GPS timing module, which is used to configure a synchronization timestamp for each acquired acoustic signal data and electrical signal data.

4. The insulator micro-damage early warning system according to claim 1, characterized in that, The data preprocessing module includes: a first processing unit, used to perform multi-level wavelet decomposition on the acoustic signal data and the electrical signal data respectively to obtain multi-level approximation coefficients and detail coefficients; a second processing unit, connected to the first processing unit, used to denoise the detail coefficients at each level respectively to obtain denoised detail coefficients; a third processing unit, connected to the first processing unit and the second processing unit respectively, used to perform wavelet reconstruction on the approximation coefficients at each level and the denoised detail coefficients to obtain pre-denoised acoustic signal data and pre-denoised electrical signal data; and a fourth processing unit, connected to the third processing unit, used to input the pre-denoised acoustic signal data and the pre-denoised electrical signal data into a Wiener filter for filtering to obtain the denoised acoustic signal data and the denoised electrical signal data.

5. The insulator minor damage early warning system according to claim 1, characterized in that, The time-domain synchronization characteristics include cross-correlation peak value, time delay, and synchronization entropy; The feature extraction module includes: a first extraction unit, used to calculate the normalized cross-correlation function of the denoised acoustic signal data and the denoised electrical signal data within a preset short time window, and extract the cross-correlation peak value and the time delay of the normalized cross-correlation function; and a second extraction unit, used to calculate the synchronization entropy of the short-time energy sequence of the denoised acoustic signal data and the denoised electrical signal data.

6. The insulator micro-damage early warning system according to claim 1, characterized in that, The frequency domain correlation characteristics include spectral morphology mutual information and key frequency band energy ratio; The feature extraction module includes: a third extraction unit, used to perform fast Fourier transform on the denoised acoustic signal data and the denoised electrical signal data respectively to obtain the acoustic signal spectrum and the electrical signal spectrum, and calculate the mutual information of the acoustic signal spectrum and the electrical signal spectrum as the spectral morphology mutual information; and a fourth extraction unit, connected to the third extraction unit, used to extract the energy in the first frequency band and the energy in the second frequency band based on a preset defect signal in the first frequency band range of the acoustic signal spectrum and the second frequency band range of the electrical signal spectrum respectively, and calculate the first proportion of the energy in the first frequency band in the total energy in the acoustic signal spectrum and the second proportion of the energy in the second frequency band in the total energy in the electrical signal spectrum as the key frequency band energy ratio.

7. The insulator minor damage early warning system according to claim 1, characterized in that, The time-frequency domain features include wavelet packet energy spectrum synchronization features and time-frequency overlap. The feature extraction module includes: a fifth extraction unit, used to perform wavelet packet decomposition on the denoised acoustic signal data and the denoised electrical signal data respectively to obtain the corresponding acoustic wavelet packet energy spectrum and electrical wavelet packet energy spectrum, and calculate the energy distribution similarity of the acoustic wavelet packet energy spectrum and the electrical wavelet packet energy spectrum at the same time and in different frequency bands as the synchronization feature of the wavelet packet energy spectrum; a sixth extraction unit, used to obtain the instantaneous frequency and amplitude of the denoised acoustic signal data and the denoised electrical signal data using Hilbert-Huang transform, so as to construct the acoustic signal Hilbert spectrum and the electrical signal Hilbert spectrum respectively, and then calculate the overlap of the acoustic signal Hilbert spectrum and the electrical signal Hilbert spectrum in time and frequency as the time-frequency overlap.

8. The insulator micro-damage early warning system according to claim 1, characterized in that, The network structure of the temporal convolutional attention network model includes an input layer, a one-dimensional temporal convolutional layer, a self-attention mechanism layer, and an output layer connected in sequence.

9. The insulator micro-damage early warning system according to claim 1, characterized in that, The early warning generation module includes an early warning counter, which is used to count once when the operating state is a latent defect and the corresponding probability exceeds a preset threshold, and generate the minor damage early warning signal when the counting result exceeds a preset number within a preset time.

10. A method for early warning of minor insulator damage based on acoustic-electrical timing characteristics, characterized in that, The insulator minor loss early warning system, applicable to any one of claims 1-9, comprises the following steps: Step S1, the insulator minor loss early warning system real-time acquisition of acoustic signal data generated by the insulator string during operation and electrical signal data flowing through the insulator string; Step S2, the insulator minor loss early warning system performs adaptive noise filtering on the acoustic signal data and the electrical signal data respectively to obtain noise-reduced acoustic signal data and noise-reduced electrical signal data; Step S3, the insulator minor loss early warning system further processes the noise-reduced acoustic signal data and the noise-reduced electrical signal data... Feature extraction yields temporal synchronization features, frequency domain correlation features, and time-frequency domain features. These features are then combined to form a joint acoustic-electric feature vector. In step S4, the insulator micro-damage early warning system inputs the joint acoustic-electric feature vector into a pre-trained temporal convolutional attention network model to obtain the operating state of the insulator string and its corresponding probability. In step S5, when the operating state is a latent defect and the corresponding probability exceeds a preset threshold for multiple consecutive time windows, the insulator micro-damage early warning system generates a micro-damage early warning signal and uploads it to the monitoring center.

Images