A transformer intelligent monitoring method and system based on voiceprint vibration

Abstract

This invention discloses a transformer intelligent monitoring method and system based on acoustic vibration, comprising: acquiring the transformer's vibration signal and high-frequency acoustic signal to construct two-dimensional raw data; using wavelet transform to extract vibration harmonic distortion and acoustic spectrum abrupt change features to determine weak anomaly indicators; using a support vector machine algorithm to assign indicator weights and determine potential anomaly types; using a deep neural network to fuse the two-dimensional fundamental frequency and high-frequency details to obtain a comprehensive diagnostic model output; using a multi-channel extension unit to synchronously acquire data from multiple measurement points to ensure signal capture integrity; using a fusion algorithm to process multi-channel data and correlate it with weak anomaly indicators to obtain precise fault location coordinates; adjusting the sampling rate to adapt the weak signal amplitude to determine the priority of the warning level; obtaining an overall assessment of the operating status based on the priority, determining potential fault trends, and obtaining transformer fault monitoring results. This invention significantly improves the early detection capability and maintenance accuracy of latent transformer faults.

Description

Technical Field

[0001] This invention belongs to the field of transformer fault monitoring, and in particular relates to a method and system for intelligent transformer monitoring based on acoustic vibration. Background Technology

[0002] As the core equipment in the power system responsible for the conversion and transmission of electrical energy, the operating status of transformers is directly related to the safety, stability and reliability of the entire power grid. Once a fault occurs, it often leads to large-scale power outages and huge economic losses. Therefore, real-time health monitoring of key components inside transformers, such as the core and windings, has extremely high engineering value.

[0003] Most current transformer monitoring methods rely on vibration signals from the tank surface or acoustic signals in the air for condition assessment. However, these methods face a fundamental contradiction in signal capture: vibration signals can effectively reflect low-frequency macroscopic mechanical changes caused by magnetostriction of the core and overall winding stress, but they are sluggish in responding to minute local distortions that appear in the early stages of a fault; while acoustic signals are sensitive to subtle disturbances in the high-frequency range, they are easily affected by ambient noise and cannot be directly correlated to the stress state of specific internal components. This natural frequency band separation between low-frequency macroscopic information and high-frequency microscopic details means that no single signal path can simultaneously represent the overall trend of the fault and detect early, subtle signs.

[0004] A deeper problem lies in the highly phased nature of the evolution of internal transformer faults. Early stages often manifest as extremely slight magnetostrictive anomalies in the core or minimal initial deformation of the windings. These changes generate extremely low-amplitude mechanical signals, typically below 0.01g, with energy concentrated primarily in specific high-frequency bands or harmonic distortion components. Existing acquisition equipment, limited by resolution and sampling rate, struggles to record these weak and transient signals without distortion. Once these initial disturbances are missed, even the most complex subsequent analyses fail to trace the true source of the fault, leaving the monitoring system largely stuck at a passive response level to late-stage, overt faults.

[0005] Therefore, the key issue in building a truly intelligent early warning capability for transformers is how to accurately capture both the low-frequency vibration fundamental frequency characteristics that reflect the macroscopic operating state and the high-frequency acoustic texture details that characterize the early microscopic anomalies in a complex electromagnetic and mechanical coupling environment, and how to make these two types of information effectively correspond and complement each other at the feature level. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides a method and system for intelligent transformer monitoring based on acoustic vibration. Specifically, the method for intelligent transformer monitoring based on acoustic vibration includes: The vibration signal of the transformer tank is obtained from the surface of the transformer tank by an accelerometer, and the high-frequency acoustic signal is collected non-contactly by an acoustic sensor. The fundamental frequency and high-frequency full-band feature matrix is ​​constructed based on the vibration signal and the high-frequency acoustic signal to obtain two-dimensional raw data. Based on the dual-dimensional raw data, wavelet transform is used to extract the harmonic distortion features of the body vibration signal and the spectral change features of the high-frequency acoustic signature signal. Based on the harmonic distortion features and the spectral change features, weak abnormal indicators associated with early faults are determined. Based on the weak anomaly indicators, the support vector machine algorithm is used to perform weight allocation and correlation analysis on the weak anomaly indicators, invert the operating status of the internal components of the transformer, and determine the potential anomaly type. Based on the potential anomaly type, a multi-feature matrix is ​​obtained, and a deep neural network is used to fuse the fundamental frequency details and high frequency details of the two-dimensional original data to obtain the output result of the comprehensive diagnostic model. Based on the output of the comprehensive diagnostic model, data synchronously collected from multiple measuring points of the transformer by the multi-channel extension unit is obtained, and the integrity of signal capture is determined based on the synchronously collected data. Based on the completeness of the signal capture, the correlation between multi-channel data and the weak anomaly indicators is processed by a fusion algorithm to obtain the precise coordinates for fault location. Based on the precise coordinates of the fault location, a high-resolution module is used to adjust the sampling rate to match the weak signal amplitude, and the priority of the warning level is determined based on the clear signal sequence obtained after the adjustment of the sampling rate. Based on the priority of the warning level, obtain overall assessment data of the transformer's operating status, determine the potential fault trends of internal components of the transformer based on the overall assessment data, and obtain transformer fault monitoring results.

[0007] Preferably, the process of obtaining the two-dimensional raw data includes: Vibration signals of the transformer body are collected from the surface of the transformer tank using an accelerometer, and high-frequency acoustic fingerprint signals are collected from the surface of the transformer tank using an acoustic fingerprint sensor. Initial two-dimensional data is constructed based on the vibration signals of the transformer body and the high-frequency acoustic fingerprint signals. The vibration signal of the body is denoised using a signal preprocessing method to obtain smoothed vibration signal data. Based on the smoothed vibration signal data, the fundamental frequency feature matrix is ​​extracted to construct a feature representation of the vibration dimension; The high-frequency feature matrix is ​​obtained by performing frequency domain transformation on the high-frequency voiceprint signal, and a feature representation of the voiceprint dimension is constructed. If the fundamental frequency feature matrix and the high frequency feature matrix have frequency band overlap, the overlapping part is weighted to obtain the fused full-band feature matrix. Based on the full-band feature matrix, two-dimensional raw data is obtained.

[0008] Preferably, the process of determining the weak anomaly indicators associated with early faults based on the harmonic distortion characteristics and the spectral abrupt change characteristics includes: Based on the two-dimensional raw data, the body vibration signal and the high-frequency acoustic signature signal are classified and stored respectively to obtain a structured signal dataset. Wavelet transform is used to decompose the body vibration signal in the signal dataset and extract harmonic distortion features. Wavelet transform was used to perform frequency domain analysis on the high-frequency acoustic signature signal in the signal dataset to extract spectral abrupt change features. A comprehensive feature matrix is ​​constructed based on the harmonic distortion characteristics and the spectral mutation characteristics; If the eigenvalues ​​of the comprehensive feature matrix exceed the preset threshold range, a weak anomaly is determined to exist, and the corresponding abnormal signal segment is obtained. Based on the abnormal signal segments, the abnormal location and type associated with the fault indicators are determined, and weak abnormal indicators associated with early faults are obtained.

[0009] Preferably, the process of determining the type of potential anomaly includes: Obtain the harmonic distortion related feature value in the weak anomaly index. If the harmonic distortion related feature value exceeds the preset threshold, then use the support vector machine algorithm to assign weights to the weak anomaly index and obtain the priority ranking of each index. Based on the priority ranking, a deep correlation analysis is performed on the high-priority indicators to determine the deviation between the core state and the winding state. If the correlation analysis results show that there is a deviation in the core state or winding state, the corresponding operating status data is extracted to determine the specific manifestation of the potential anomaly. Based on the specific manifestations of the potential anomalies, obtain the distribution of anomaly types and determine whether the anomaly involves multiple internal components. If an anomaly involves multiple internal components, cross-validation is performed on the anomaly types to obtain the final anomaly classification result, and the potential anomaly types are determined based on the final anomaly classification result.

[0010] Preferably, the process of obtaining the output results of the comprehensive diagnostic model includes: Initial data is obtained based on the potential anomaly types, and the initial data is cleaned and standardized to obtain a structured multi-feature matrix. Based on the multi-feature matrix, a hierarchical processing method is performed for the feature extraction process to determine the key feature set. If the distribution of fundamental frequency details and high frequency details in the key feature set is uneven, they are separated by a signal dimension decomposition method to obtain the separated fundamental frequency signal data and high frequency signal data. A deep neural network is used to integrate the separated baseband signal data and high-frequency signal data, and the fused signal features are output. If the fused signal features reach a preset threshold, then the comprehensive diagnostic logic is used for classification processing to determine the specific category of the abnormality, generate diagnostic result data, and obtain the output result of the comprehensive diagnostic model.

[0011] Preferably, the process of determining the integrity of signal acquisition based on synchronously acquired data includes: Based on the output of the comprehensive diagnostic model, obtain the corresponding timestamp interval, and extract synchronously collected data from the multi-channel extension unit based on the timestamp interval; The number of data points in each channel of the synchronously acquired data is counted. If the number of data points reaches the preset number of acquisition points, the corresponding channel is marked as complete; otherwise, it is marked as missing. For channels marked as missing, find the most recent complete acquisition time point and calculate the time difference between the most recent complete acquisition time point and the current timestamp interval; If the time difference is less than the preset allowable interval, the current channel is supplemented with the previous complete data; All the supplemented channel data are time-aligned and arranged to form a unified sequence; Calculate the sampling time deviation between each channel based on the unified sequence; If the sampling time deviations are all less than the preset synchronization threshold, then the signal acquisition integrity is determined to meet the requirements.

[0012] Preferably, the process of obtaining the precise coordinates of the fault location includes: The system acquires a multi-channel raw signal sequence and uses a signal integrity verification module to determine whether the multi-channel raw signal sequence meets the capture requirements. If it does, the system retains the multi-channel raw signal sequence; otherwise, it acquires the signal again. Based on the preserved multi-channel original signal sequences, extract the early abnormal indicator sequences for each channel; A fusion algorithm is used to correlate the multi-channel original signal sequence with the early anomaly index sequence to obtain the spatiotemporal correspondence of each anomaly point in the multi-channel; The channel combination in which the anomaly occurred is determined based on the spatiotemporal correspondence described above; The precise spatial coordinates of the anomaly are calculated based on the spatial geometric distribution of the channel combination, thus obtaining the precise coordinates for fault location.

[0013] Preferably, the process of determining the priority of the warning level based on the clear signal sequence obtained from the adjusted sampling rate includes: The spatial location of the current monitoring point is determined based on the precise coordinates of the fault location, and the original weak signal of the monitoring point is acquired using a high-resolution acquisition module. The current signal strength is calculated based on the amplitude of the original weak signal; The sampling rate parameter is dynamically adjusted based on the current signal strength. The weak signal was reacquired using the adjusted sampling rate to obtain a clear signal sequence. The warning level value is determined based on the amplitude distribution of the clear signal sequence; The warning priority is marked based on the comparison result between the warning level value and the preset threshold, and the priority of the warning level is obtained.

[0014] Preferably, the process of determining the potential failure trends of internal components of the transformer based on the overall assessment data includes: Priority ranking information is obtained from the priority data of the warning level, and the priority ranking information is initially screened. If the priority order is higher than the preset threshold, the corresponding data is marked as a high-risk category. For data marked as high-risk, real-time records of transformer operating status are obtained, and these real-time records are compared and analyzed with historical data. If abnormal fluctuations are found, they are identified as anomalies, and the distribution range of the anomalies is determined. Based on the distribution range of the abnormal points, the operation logs of the internal components of the transformer are obtained. If the operating parameters of the components deviate from the normal range, it is judged as a potential signal of component failure, and the initial fault location is obtained. Based on the component data of the preliminary fault location, the change characteristics of the potential trend are obtained, and the support vector machine algorithm is used to classify the change characteristics to determine whether the potential trend points to continuous deterioration and to determine the direction of the trend prediction. Based on the trend prediction, information for fault judgment is obtained. The information is compared in multiple dimensions. If the comparison results show an abnormal consistency, it is judged as a high-probability fault risk, and the final conclusion of risk assessment is obtained. Based on the final conclusion of the risk assessment, updated data for status analysis is obtained. If the updated data points to a specific component, it is identified as a key monitoring target, and priority targets for internal inspection are determined to obtain the potential failure trends of internal components of the transformer.

[0015] This invention also provides a transformer intelligent monitoring system based on acoustic vibration, comprising: The dual-dimensional data acquisition module is used to acquire the body vibration signal from the surface of the transformer tank through an accelerometer, and at the same time, to acquire the high-frequency acoustic fingerprint signal non-contactly through an acoustic fingerprint sensor. Based on the body vibration signal and the high-frequency acoustic fingerprint signal, a fundamental frequency and high-frequency full-band feature matrix is ​​constructed to obtain dual-dimensional raw data. The weak anomaly index extraction module is used to extract the harmonic distortion features of the body vibration signal and the spectral change features of the high-frequency acoustic text signal using wavelet transform based on the dual-dimensional raw data, and to determine the weak anomaly index associated with the early fault based on the harmonic distortion features and the spectral change features. The early fault analysis module is used to perform weight allocation and correlation analysis on the weak anomaly indicators based on the weak anomaly indicators using the support vector machine algorithm, to invert the operating status of the internal components of the transformer and determine the potential anomaly type. The comprehensive diagnostic module is used to obtain a multi-feature matrix based on the potential abnormality type, and to fuse the fundamental frequency details and high frequency details of the two-dimensional original data using a deep neural network to obtain the output result of the comprehensive diagnostic model. The signal integrity determination module is used to obtain data synchronously collected from multiple measuring points of the transformer by the multi-channel extension unit based on the output results of the comprehensive diagnostic model, and to determine the integrity of the signal capture based on the synchronously collected data; The fault location module is used to process the correlation between multi-channel data and the weak anomaly indicators through a fusion algorithm based on the integrity of the signal capture to obtain the precise coordinates of the fault location. The warning level determination module is used to adjust the sampling rate using a high-resolution module to adapt to the weak signal amplitude based on the precise coordinates of the fault location, and to determine the priority of the warning level based on the clear signal sequence obtained after the adjustment of the sampling rate. The fault trend assessment module is used to obtain overall assessment data of the transformer's operating status according to the priority of the warning level, determine the potential fault trends of the transformer's internal components based on the overall assessment data, and obtain transformer fault monitoring results.

[0016] Compared with the prior art, the present invention has the following advantages and technical effects: This invention uses an accelerometer and an acoustic sensor to simultaneously collect the body vibration signal and high-frequency acoustic signal from the surface of the transformer tank, and constructs a fundamental frequency and high-frequency full-band feature matrix. This achieves the fusion of vibration and acoustic signals in two dimensions, effectively solving the problem of incomplete information coverage in a single signal dimension. It can comprehensively reflect the operating status of internal components such as the core and windings, and significantly improve the completeness and accuracy of fault feature extraction.

[0017] This invention uses wavelet transform to extract the harmonic distortion features of vibration signals and the spectral mutation features of acoustic signals. It also uses the support vector machine algorithm to perform weight allocation and correlation analysis on weak anomaly indicators, which can effectively identify weak anomaly features associated with early faults and significantly advance the fault warning cycle. This invention solves the limitation of traditional algorithms that can only identify mid-to-late stage obvious faults and achieves accurate early warning of early potential faults such as slight loosening of the iron core and initial deformation of the winding.

[0018] This invention fuses fundamental and high-frequency details of dual-dimensional raw data through a deep neural network to obtain the output of a comprehensive diagnostic model, thereby enhancing the multi-dimensional signal fusion processing capability, effectively improving the accuracy of fault identification, and reducing the false positive and false negative rates.

[0019] This invention utilizes a multi-channel extension unit to synchronously collect data from multiple measuring points on the transformer and determine the integrity of signal acquisition. It combines a fusion algorithm to process the correlation between multi-channel data and early abnormal indicators to obtain the precise coordinates of fault location, thereby achieving precise spatial positioning of the fault source and providing accurate fault location information for maintenance personnel.

[0020] This invention employs a high-resolution module to dynamically adjust the sampling rate based on the precise coordinates of fault location to adapt to the amplitude of weak signals, and determines the priority of the warning level. This ensures that weak early fault signals with amplitude ≤0.01g can be captured without distortion, enhancing the ability to capture weak signals and providing high-quality data support for early diagnosis.

[0021] In summary, this invention achieves a closed-loop diagnosis across the entire chain, from synchronous acquisition of dual-dimensional signals, early identification of weak anomalies, accurate fault location to trend prediction. This significantly improves the early detection capability and maintenance accuracy of latent transformer faults, effectively reduces unplanned transformer downtime, lowers the risk of power outages and economic losses caused by fault escalation, extends transformer service life, and provides reliable technical support for the safe and stable operation of the power grid.

[0022] This invention achieves synchronous acquisition and deep fusion of vibration and acoustic signature signals. Through multi-feature weight allocation and correlation analysis, it effectively identifies early weak anomalies in the core and windings, significantly advancing the fault warning cycle. It employs a 24-bit high-resolution acquisition module with a maximum sampling rate of 256kHz to ensure distortion-free capture of weak early fault signals. A flexible and scalable 8-16 channel synchronous monitoring architecture is constructed to achieve synchronous data acquisition from multiple measurement points, significantly improving fault identification accuracy and reducing false positive and false negative rates, providing reliable technical support for the safe and stable operation of the power grid. Attached Figure Description

[0023] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the system structure according to an embodiment of the present invention. Detailed Implementation

[0024] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0025] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0026] Example 1 like Figure 1 As shown, this embodiment provides a transformer intelligent monitoring method based on acoustic vibration, including: The vibration signal of the transformer tank is obtained from the surface of the transformer tank by an accelerometer, and the high-frequency acoustic signal is collected non-contactly by an acoustic sensor. The fundamental frequency and high-frequency full-band feature matrix is ​​constructed based on the vibration signal and the high-frequency acoustic signal to obtain two-dimensional raw data. Based on the two-dimensional raw data, wavelet transform was used to extract the harmonic distortion features of the body vibration signal and the spectral change features of the high-frequency acoustic signature signal. Based on the harmonic distortion features and the spectral change features, weak anomaly indicators associated with early faults were determined. Based on the weak anomaly indicators, the support vector machine algorithm is used to perform weight allocation and correlation analysis on the weak anomaly indicators, invert the operating status of the internal components of the transformer, and determine the potential anomaly type. Based on the potential anomaly type, a multi-feature matrix is ​​obtained, and a deep neural network is used to fuse the fundamental frequency details and high frequency details of the two-dimensional original data to obtain the output of the comprehensive diagnostic model. Based on the output of the comprehensive diagnostic model, data synchronously collected from multiple measuring points of the transformer by the multi-channel extension unit is obtained, and the integrity of signal acquisition is determined based on the synchronously collected data. Based on the integrity of the signal capture, the correlation between multi-channel data and weak anomaly indicators is processed by a fusion algorithm to obtain the precise coordinates for fault location. Based on the precise coordinates of the fault location, a high-resolution module is used to adjust the sampling rate to match the weak signal amplitude, and the priority of the warning level is determined based on the clear signal sequence obtained after the adjustment of the sampling rate. Based on the priority of the warning level, obtain overall assessment data of the transformer's operating status, determine the potential fault trends of internal components of the transformer based on the overall assessment data, and obtain transformer fault monitoring results.

[0027] Furthermore, the process of obtaining the two-dimensional raw data includes: Vibration signals of the transformer body are collected from the surface of the transformer tank using an accelerometer, and high-frequency acoustic fingerprint signals are collected from the surface of the transformer tank using an acoustic fingerprint sensor. Initial two-dimensional data is constructed based on the vibration signals of the transformer body and the high-frequency acoustic fingerprint signals. The vibration signal of the body is denoised using a signal preprocessing method to obtain smoothed vibration signal data. The fundamental frequency feature matrix is ​​extracted from the smoothed vibration signal data to construct a feature representation of the vibration dimension. By performing frequency domain transformation on the high-frequency voiceprint signal, a high-frequency feature matrix is ​​obtained, and a feature representation of the voiceprint dimension is constructed. If the fundamental frequency feature matrix and the high frequency feature matrix have overlapping frequency bands, the overlapping part is weighted to obtain the fused full-band feature matrix. Based on the full-band feature matrix, two-dimensional raw data is obtained.

[0028] Furthermore, in this embodiment, in the scenario of transformer tank condition monitoring, signals are collected using an accelerometer and an acoustic fingerprint sensor to construct two-dimensional data. The accelerometer is mainly used to capture mechanical vibration signals on the surface of the tank, measured in millimeters per second squared. Assuming a peak value of 2.5 for a particular vibration signal, the frequency range is between 10 and 100 Hz. The acoustic fingerprint sensor, on the other hand, focuses on high-frequency acoustic fingerprint signals, with a frequency range potentially between 1 kHz and 10 kHz, used to capture subtle changes in sound waves. These two signals respectively represent the mechanical vibration of the transformer and the possible internal discharge or bubble sounds, forming complementary data dimensions.

[0029] In one possible implementation, this embodiment employs wavelet transform for denoising the vibration signal. The original signal is decomposed into different frequency components, and high-frequency noise is removed to obtain smoothed vibration signal data. Assuming the original signal contains a 50 Hz fundamental frequency vibration and some stray noise, after processing, the fundamental frequency signal is preserved, and the noise is effectively filtered out. This processing significantly improves the accuracy of subsequent feature extraction, ensuring that the fundamental frequency feature matrix reflects the true vibration characteristics.

[0030] Specifically, in this embodiment, when extracting the fundamental frequency feature matrix, the main frequency components in the range of 10 to 100 Hz are separated from the smoothed vibration signal to form a matrix containing frequency, amplitude and phase information.

[0031] For example, the fundamental frequency may be concentrated at 50 Hz, with an amplitude of approximately 1.8, reflecting the vibration pattern under normal transformer operation. This matrix, as a characteristic representation of the vibration dimension, lays the foundation for subsequent analysis.

[0032] For example, in processing high-frequency acoustic signature signals, this embodiment uses Fast Fourier Transform to convert them from the time domain to the frequency domain, obtaining a high-frequency feature matrix in the range of 1 kHz to 10 kHz. Suppose that in a certain measurement, a significant amplitude peak is found at 5 kHz, which may be related to internal partial discharge. This feature matrix, as a representation of the acoustic signature dimension, can reveal potential anomalies inside the transformer.

[0033] In one possible implementation, if the fundamental frequency feature matrix and the high-frequency feature matrix overlap in the frequency band, for example, there is a small amount of overlapping data near 100 Hz, then the overlapping part is weighted. In this embodiment, the vibration signal is given a higher weight, such as 0.7, while the acoustic signature signal is weighted at 0.3, to highlight the reliability of the mechanical vibration. The fused full-band feature matrix covers the entire frequency band from 10 Hz to 10 kHz, improving the comprehensiveness of the features.

[0034] Specifically, in this embodiment, the support vector machine algorithm is applied to classify the fused feature matrix, and the signal features are compared with preset anomaly patterns.

[0035] For example, assuming an abnormal pattern is defined as a high-frequency amplitude exceeding a certain threshold or a fundamental frequency offset exceeding the normal range, a classifier can be used to determine whether the current signal is abnormal. This method can effectively distinguish between normal and abnormal states, improving diagnostic accuracy.

[0036] For example, in the final operational status determination, if the classification result shows that the signal conforms to an abnormal pattern, the system will generate an identification data, such as "Status Abnormal - High Frequency Abnormality," indicating a possible internal discharge problem. This status identification data provides maintenance personnel with an intuitive reference, significantly reducing the risk of misjudgment and improving the safety and reliability of transformer operation.

[0037] Furthermore, the process of determining weak anomaly indicators associated with early faults based on harmonic distortion characteristics and spectral abrupt change characteristics includes: Based on the two-dimensional raw data, the body vibration signal and high-frequency acoustic signature signal are classified and stored separately to obtain a structured signal dataset; Wavelet transform is used to decompose the body vibration signal in the signal dataset and extract harmonic distortion features; Wavelet transform was used to perform frequency domain analysis on the high-frequency acoustic signature signal in the signal dataset to extract spectral abrupt change features. A comprehensive feature matrix is ​​constructed based on harmonic distortion characteristics and spectral abrupt change characteristics; If the eigenvalues ​​of the comprehensive feature matrix exceed the preset threshold range, a weak anomaly is determined to exist, and the corresponding abnormal signal segment is obtained. Based on the abnormal signal segments, the location and type of the abnormality associated with the fault indicators are determined, and weak abnormal indicators associated with early faults are obtained.

[0038] Furthermore, in this embodiment, when processing the two-dimensional raw data collected from the device, the vibration signal and the acoustic signature signal are first initially stored and classified. Assuming a transformer monitoring scenario, the vibration signal is stored as time-series data, with 1000 data points collected per second, while the acoustic signature signal is stored in the frequency domain, generating a spectrum every minute. This classification method facilitates the subsequent separate processing of the two types of signals, ensuring structured data storage while reducing the risk of data confusion.

[0039] For example, this embodiment focuses on wavelet transform decomposition of vibration signals, extracting harmonic distortion features through multi-level decomposition. Assuming periodic fluctuations exist in the vibration signal, wavelet transform can decompose it into multiple frequency bands, identifying anomalous harmonic components appearing around 50Hz. This feature extraction helps to discover potential mechanical loosening or imbalance problems during equipment operation, providing crucial information for subsequent analysis.

[0040] For example, in the frequency domain analysis of the voiceprint signal in this embodiment, wavelet transform is also used to extract spectral abrupt change features. Suppose that in the high-frequency band, the voiceprint signal experiences a significant spectral jump at a certain moment, abruptly rising from 2000Hz to 2500Hz. This may indicate phenomena such as partial discharge or bubble rupture within the device. By capturing these irregular change points, important clues can be provided for anomaly early warning.

[0041] For example, in this embodiment, the harmonic distortion features of the vibration signal and the spectral abrupt change features of the acoustic signature signal are integrated in the feature fusion processing to construct a comprehensive feature matrix. Assuming the vibration features show distortion at 50Hz, while the acoustic signature features show a high-frequency abrupt change at the corresponding time point, it can be inferred that the device may have fatigue damage in a specific component. This fusion method can reveal potential problems from multiple dimensions, improving the comprehensiveness of the analysis.

[0042] For example, if the eigenvalues ​​in the comprehensive feature matrix exceed a preset threshold, such as a vibration eigenvalue exceeding the normal range of 0.5, it is identified as a weak anomaly, and a signal segment corresponding to that time period is extracted. Assuming the anomaly segment lasts for 10 seconds, further analysis of its waveform changes can pinpoint the source of the anomaly to a connecting component on one side of the fuel tank. This method helps to quickly locate the problem area.

[0043] For example, in this embodiment, the location and type of the anomaly are determined by combining the typical patterns of early faults during in-depth analysis of abnormal signal segments. Assuming that historical data comparison reveals a high similarity between the current signal segment and the insulation aging pattern inside the tank, the anomaly type can be inferred to be insulation degradation. This method provides precise guidance for equipment maintenance, preventing further escalation of the fault. Through the aforementioned multi-dimensional signal processing and feature analysis methods, the monitoring capability of transformer operating status can be effectively improved, providing strong support for the safe and stable operation of equipment, while reducing the risk of misjudgment and extending equipment lifespan.

[0044] Furthermore, the process of determining the type of potential anomaly includes: Obtain the harmonic distortion related feature values ​​in the weak anomaly indicators. If the harmonic distortion related feature values ​​exceed the preset threshold, the weak anomaly indicators are weighted using the support vector machine algorithm to obtain the priority ranking of each indicator. Based on priority ranking, in-depth correlation analysis is performed on high-priority indicators to determine the deviation between the core condition and the winding condition. If the correlation analysis results show that there is a deviation in the core state or winding state, the corresponding operating status data is extracted to determine the specific manifestation of the potential anomaly. Based on the specific manifestations of potential anomalies, obtain the distribution of anomaly types and determine whether the anomaly involves multiple internal components. If an anomaly involves multiple internal components, cross-validation of the anomaly types is performed to obtain the final anomaly classification result, and the potential anomaly types are determined based on the final anomaly classification result.

[0045] Furthermore, this embodiment acquires key characteristic values ​​such as harmonic distortion rate in real time by collecting electrical signal data during transformer operation.

[0046] In one embodiment, when the third harmonic distortion rate is detected to reach 7.2%, while the normal operating threshold is set to 4.5%, a preliminary abnormality is determined, triggering the subsequent diagnostic process.

[0047] In one possible implementation, this embodiment uses a support vector machine algorithm to assign weights to the extracted multiple harmonic feature values.

[0048] Specifically, the third, fifth, and seventh harmonic distortion rates, as well as the total harmonic distortion rate, are input into the model. After training, weight coefficients for each feature are obtained. For example, the weight for the third harmonic is 0.42, the weight for the fifth harmonic is 0.31, the weight for the seventh harmonic is 0.18, and the weight for the total distortion rate is 0.09. Based on the weight ranking, the third harmonic distortion is preferentially identified as the core anomaly indicator.

[0049] Preferably, this embodiment further conducts in-depth correlation analysis on the third harmonic distortion, which has a higher weight.

[0050] Understandably, by comparing historical normal data with the current deviation, it was found that the magnetic saturation of the iron core increased by about 18%, while the probability of inter-turn short circuits in the windings rose to the medium-risk range. This correlation analysis helps to clarify whether the anomaly mainly originates from the iron core or the windings.

[0051] For example, when the core condition deviation is manifested as an abnormal increase in magnetic flux density and the winding condition shows signs of local overheating, multi-dimensional operating status data such as temperature, vibration and noise for the corresponding time period can be extracted.

[0052] In one embodiment, the core temperature rises by 12°C from the reference temperature, and the winding hot spot temperature exceeds the warning value by 9°C. These specific manifestations all point to a potential anomaly where core loosening and winding insulation aging coexist.

[0053] It should be noted that after classifying the aforementioned potential anomalies in this embodiment, a detailed distribution of anomaly types can be obtained. For example, core-related anomalies account for 61%, winding-related anomalies account for 33%, and other comprehensive factors account for 6%. If the cross-validation results show that the anomaly involves both the core and winding, two key components, it is determined to be a composite internal fault. Based on the final anomaly classification results, a targeted status monitoring strategy is generated.

[0054] For example, this embodiment focuses subsequent data acquisition on the trend of third harmonic variation, the increase in core vibration amplitude, and the rate of temperature rise in winding hot spots, thereby achieving more accurate tracking and early warning of early complex faults. This hierarchical and focused monitoring method effectively improves the speed of anomaly location and the pertinence of preventive maintenance, significantly reducing the risk of sudden downtime.

[0055] Furthermore, the process of obtaining the output results of the comprehensive diagnostic model includes: Initial data is obtained based on potential anomaly types. The initial data is then cleaned and standardized to obtain a structured multi-feature matrix. Based on the multi-feature matrix, the feature extraction process is hierarchically processed to determine the set of key features. If the distribution of fundamental frequency details and high frequency details in the key feature set is uneven, then the signal dimension decomposition method is used to separate them to obtain the separated fundamental frequency signal data and high frequency signal data. A deep neural network is used to integrate the separated fundamental frequency signal data and high frequency signal data, and output the fused signal features. If the fused signal features reach a preset threshold, then the comprehensive diagnostic logic is used for classification processing to determine the specific category of the abnormality, generate diagnostic result data, and obtain the output result of the comprehensive diagnostic model.

[0056] Furthermore, in this embodiment, data cleaning and standardization are crucial steps in processing initial data of anomaly types. Assuming a set of raw signal data containing noise and redundant information is collected from the operation of power equipment, the cleaning process removes irrelevant interference through filtering methods, such as eliminating low-frequency noise below 0.5 Hz. Simultaneously, the data is normalized to ensure that feature values ​​of different dimensions are within the same range, facilitating subsequent analysis. This improves data quality and lays the foundation for subsequent feature extraction.

[0057] For example, this embodiment addresses the hierarchical processing of structured multi-feature matrices by using principal component analysis to reduce the dimensionality of the features and extract the most representative set of key features. Suppose that the operating data of a transformer contains multi-dimensional information such as voltage, current, and temperature. After hierarchical processing, it is found that voltage fluctuations and temperature anomalies are the main influencing factors. These sets of key features can help focus on core issues, reduce the burden of redundant calculations, and improve analysis efficiency.

[0058] For example, signal dimensionality decomposition is particularly important when the fundamental frequency signal and high-frequency details are unevenly distributed. Imagine a set of signal data where the fundamental frequency signal reflects the basic operating state of the equipment, while the high-frequency details may contain early signs of localized faults. This embodiment uses wavelet decomposition to separate the signal into different frequency bands, extracting feature values ​​from the fundamental frequency and high-frequency components separately. For instance, the fundamental frequency signal amplitude is 10 volts, while the high-frequency details fluctuate within 0.2 volts. The advantage of this separation method is that it can more accurately capture the source of abnormal signals, providing a clear basis for subsequent diagnosis.

[0059] For example, in this embodiment, for the separated signal data, signal integration through a deep neural network can further uncover potential relationships between features. Assuming the separated signal data contains features from multiple frequency bands, the hierarchical structure of the neural network can fuse these features into a comprehensive feature vector. For instance, it can combine the stability of the fundamental frequency signal with abnormal fluctuations in high-frequency details to form an indicator reflecting the overall status of the device. This fusion method can enhance the expressive power of features and provide more comprehensive data support for anomaly classification.

[0060] For example, once the fused signal characteristics reach a preset threshold, the classification processing of the comprehensive diagnostic logic becomes particularly crucial. Suppose the anomaly index of the fused characteristics reaches 0.8, exceeding the preset threshold of 0.6. At this point, historical data and a rule base can be used to classify the anomaly type into categories such as core overheating or winding short circuits. The advantage of this classification processing is that it can quickly pinpoint the problem type, providing a clear direction for subsequent maintenance.

[0061] For example, when generating diagnostic result data in this embodiment, the classification results are combined with the equipment operation log to form a detailed report. Assuming the diagnostic results indicate an anomaly type of winding short circuit, the report can indicate the time period of the anomaly and relevant parameter values, such as the peak current reaching 1.5 times the rated value. This detailed output helps technicians quickly develop countermeasures, reduce equipment downtime, and improve maintenance efficiency.

[0062] Furthermore, the process of determining the integrity of signal acquisition based on synchronously acquired data includes: Based on the output of the comprehensive diagnostic model, obtain the corresponding timestamp interval, and extract synchronously collected data from the multi-channel extended unit based on the timestamp interval; The number of data points in each channel of the synchronously collected data is counted. If the number of data points reaches the preset number of collection points, the corresponding channel is marked as complete; otherwise, it is marked as missing. For channels marked as missing, find the most recent complete acquisition time point and calculate the time difference between the most recent complete acquisition time point and the current timestamp interval; If the time difference is less than the preset allowable interval, the current channel will be supplemented with the previous complete data; All the supplemented channel data are time-aligned and arranged to form a unified sequence; The sampling time deviation between each channel is calculated based on the unified sequence; If the sampling time deviation is less than the preset synchronization threshold, then the signal acquisition integrity is determined to meet the requirements.

[0063] Furthermore, in this embodiment, when processing the output of the comprehensive diagnostic model, the corresponding timestamp interval is first obtained. For example, if a device's diagnostic result shows an abnormal signal between 10:00:00 and 10:05:00, then synchronously acquired data is extracted from the multi-channel expansion unit. Assuming the device has three channels monitoring vibration, temperature, and pressure respectively, the extracted data shows that the number of sampling points for each channel within this time period is 500, 480, and 510 respectively, while the preset number of sampling points is 500. For the temperature channel with insufficient data, it is marked as missing, and the most recent complete acquisition time point is found, for example, 09:55:00. The time difference is 5 minutes, less than the preset allowable interval of 10 minutes, so the data for the current channel is supplemented with the previous complete data.

[0064] Specifically, this embodiment aligns the supplemented data by time to form a unified sequence. Assuming the vibration channel samples once per second, the temperature channel at 0.9 times per second, and the pressure channel at 1.1 times per second, calculating the sampling time deviation between each channel reveals a deviation of 0.1 seconds between the temperature and vibration channels, and 0.2 seconds between the pressure and vibration channels, while the preset synchronization threshold is 0.15 seconds. Clearly, the pressure channel's deviation exceeds the threshold, therefore it is marked as a synchronization anomaly, and the anomaly channel number is output as 3, with a deviation value of 0.2 seconds. This method helps to quickly locate the problematic channel, ensuring the accuracy of subsequent analysis.

[0065] For example, in practical applications, determining the timestamp interval can be based on equipment operation logs, combined with abnormal time points output by the diagnostic model, to accurately pinpoint the data range requiring analysis. The extraction of multi-channel data relies on the real-time acquisition capabilities of the sensor network to ensure data comprehensiveness. The statistical analysis of data points and the marking of missing data points can be understood as a data integrity check, preventing misjudgments due to missing data. The data completion operation in this embodiment demonstrates the rational use of historical data, reducing the impact of temporary data loss. The calculation of time alignment and sampling time deviation ensures the consistency of multi-channel signals in the time dimension, which is particularly important in equipment status monitoring. The marking and output of synchronization anomalies provide clear direction for subsequent maintenance; for example, priority can be given to checking the clock synchronization settings or hardware connection issues of the pressure sensors.

[0066] Specifically, the implementation of each of the above steps helps to improve the reliability of data processing.

[0067] For example, time-aligned data arrangement can avoid signal misalignment caused by time deviations, ensuring that the data from each channel reflects the equipment status at the same moment during analysis. The identification and output of synchronization anomalies provide technicians with intuitive anomaly information, saving troubleshooting time. This meticulous data processing method not only improves diagnostic accuracy but also provides more targeted basis for equipment maintenance, reducing unnecessary resource waste.

[0068] Furthermore, the process of obtaining the precise coordinates of the fault location includes: The system acquires the original signal sequence of multiple channels and uses a signal integrity verification module to determine whether the original signal sequence of multiple channels meets the capture requirements. If it does, the original signal sequence of multiple channels is retained; otherwise, it is acquired again. Based on the preserved multi-channel original signal sequences, extract the early abnormal indicator sequences for each channel; A fusion algorithm is used to correlate the original multi-channel signal sequence with the early anomaly index sequence to obtain the spatiotemporal correspondence of each anomaly point in the multi-channel; Determine the channel combination for the occurrence of anomalies based on spatiotemporal correspondence; The precise spatial coordinates of the anomaly are calculated based on the spatial geometric distribution of the channel combination, thus obtaining the precise coordinates for fault location.

[0069] Furthermore, in this embodiment, in the scenario of monitoring the vibration of the core and windings of a power transformer, multi-channel raw signal sequences are first acquired. These sequences come from accelerometers and acoustic sensors deployed at different parts of the transformer. A signal integrity verification module quickly assesses the number of sampling points and temporal continuity for each channel. If the preset requirement is that each channel acquires at least 1024 valid points without any interruptions exceeding 50 milliseconds, and all channels meet the conditions, the entire raw signal sequence is retained for subsequent analysis. If any channel has insufficient points or is missing for an extended period, the acquired sequence is discarded, triggering a re-acquisition command to avoid misjudgments caused by incomplete data.

[0070] In one embodiment, the early anomaly indicator sequence of each channel is extracted from the preserved multi-channel original signal sequence. Specifically, this includes indicators such as short-term energy spikes, abnormal increases in peak factor, and rapid increases in the amplitude of specific harmonic components in the spectrum. For example, if a channel detects a sudden energy spike to 3.2 times the normal value at 0.8 seconds, while the peak factor rises from 1.4 to 4.1, this can be marked as an early anomaly.

[0071] Preferably, a fusion algorithm is used to perform spatiotemporal correlation calculations on the original signal sequence and the early anomaly indicator sequence. During the calculation, the time difference and amplitude change trend of the anomaly points in each channel are compared to determine the propagation order and intensity correspondence of the anomalies across multiple channels.

[0072] For example, after an anomaly occurs in channel 1 at time t1, a similar but slightly weaker anomaly occurs in channel 3 at t1 plus 12 milliseconds, while channel 5 shows a stronger response 18 milliseconds later. Based on this, it can be determined that the source of the anomaly is closer to the area covered by channels 1 and 3. According to the spatiotemporal correspondence, the channel combination in which the anomaly occurred can be further determined. For example, this anomaly mainly involves the triangular region formed by channels 1, 3, and 5. Next, the spatial geometric distribution of the channel combination is used for location calculation. Assuming the spatial coordinates of the three sensors are known, channel 1 is located at (0.4 m, 0.2 m, 0.1 m), channel 3 at (0.7 m, 0.5 m, 0.1 m), and channel 5 at (0.5 m, 0.8 m, 0.3 m), combined with the anomaly propagation time difference and the speed of sound or vibration wave velocity, the precise spatial coordinates of the anomaly can be calculated to be approximately (0.52 m, 0.47 m, 0.18 m).

[0073] Understandably, this complete link, from signal integrity control to spatiotemporal fusion to precise geometric positioning, can significantly improve the accuracy and reliability of fault location, avoid mislocation caused by missing or asynchronous data from individual channels, and in practical applications, the positioning error can be controlled at the centimeter level, providing more accurate spatial guidance for transformer live-line maintenance and defect early warning.

[0074] Furthermore, the process of determining the priority of the warning level based on the clear signal sequence obtained from the adjusted sampling rate includes: The spatial location of the current monitoring point is determined based on the precise coordinates of the fault location, and the original weak signal of the monitoring point is acquired using a high-resolution acquisition module. Calculate the current signal strength based on the amplitude of the original weak signal; The sampling rate parameter is dynamically adjusted based on the current signal strength. The weak signal was reacquired using the adjusted sampling rate to obtain a clear signal sequence. The warning level value is determined based on the amplitude distribution of the clear signal sequence; The warning priority is marked based on the comparison result between the warning level value and the preset threshold, and the priority of the warning level is obtained.

[0075] Furthermore, in one possible implementation, the precise coordinates output by the fault location module in this embodiment can be directly mapped to the spatial location of the current monitoring point. For example, in a distributed monitoring scenario of a large transformer array, assuming the fault location module calculates the precise coordinates of the anomaly as x=12.5 meters, y=8.7 meters, and z=3.2 meters, the system immediately confirms that these coordinates correspond to the specific monitoring point location near the B-phase winding of transformer group 3, thereby avoiding the inefficient process of traditional point-by-point investigation.

[0076] Preferably, for the determined monitoring point location, a high-resolution acquisition module is used to capture the original weak signal of the target point.

[0077] Specifically, when the monitoring point is in an environment with strong electromagnetic interference, the high-resolution acquisition module operates with 24-bit precision and an initial sampling rate of 10kHz, capturing partial discharge pulse signals with an amplitude of only a few microvolts. This weak characteristic often indicates early insulation degradation.

[0078] Understandably, calculating the current signal strength based on the amplitude of the original weak signal is crucial for subsequent optimization. For example, when the acquired pulse peak values ​​are 3.2μV, 5.8μV, and 12.1μV, the system estimates the signal strength indices as 0.15, 0.28, and 0.62 by taking the maximum amplitude and combining it with background noise. Higher signal strength indicates more significant anomalous energy.

[0079] In one embodiment, the system dynamically adjusts the sampling rate parameter based on the signal strength. When the strength index is below 0.3, the sampling rate is maintained at a low 5kHz to conserve resources; when the strength index is between 0.3 and 0.6, the sampling rate is increased to 20kHz; and when the strength exceeds 0.6, it is further increased to 50kHz to ensure complete waveform details are captured. After adjustment and re-acquisition, the resulting clear signal sequence exhibits steeper pulse rising edges and clearly discernible oscillation period characteristics.

[0080] For example, the clear signal sequences acquired after adjusting the sampling rate show a distinct three-segment amplitude distribution: 80% of the sample amplitudes are concentrated in the normal noise range of -2μV to 2μV, 15% of the samples are in the moderately abnormal range of ±5μV to ±15μV, and 5% of the samples exceed the high-risk range of ±20μV. Based on this amplitude distribution pattern, the system calculates the warning level value; for example, using a segmented weighting method, the comprehensive warning level is 0.78.

[0081] It should be noted that if the warning level value is higher than the preset threshold of 0.75, it is marked as high priority, which means that there may be a rapidly developing partial discharge defect; if the value is in the middle range of 0.45 to 0.75, it is marked as medium priority, which means that the anomaly has developed to a certain extent but has not yet gotten out of control; if it is lower than 0.45, it is marked as low priority, which only requires routine attention.

[0082] Specifically, the system sorts all monitoring points in the current batch according to their priority values ​​from highest to lowest. For example, if high-priority point A has a warning level of 0.82, medium-priority point C has a warning level of 0.61, and low-priority point B has a warning level of 0.31, the output sequence will be A→C→B. This sorted warning processing sequence allows maintenance personnel to prioritize responding to the highest-risk points, significantly improving the timeliness of fault prevention and resource allocation efficiency, thereby effectively reducing the probability of sudden equipment downtime and extending the overall service life.

[0083] Furthermore, the process of determining potential failure trends of internal transformer components based on overall assessment data includes: Priority ranking information is obtained from the priority data of the warning level. The priority ranking information is initially screened. If the priority order is higher than the preset threshold, the corresponding data is marked as a high-risk category. For data marked as high-risk, real-time records of transformer operating status are obtained, and these records are compared and analyzed with historical data. If abnormal fluctuations are found, they are identified as anomalies, and the distribution range of the anomalies is determined. Based on the distribution range of abnormal points, the operation logs of the internal components of the transformer are obtained. If the operating parameters of the components deviate from the normal range, it is judged as a potential signal of component failure, and the initial fault location is obtained. Based on the component data of the initial fault location, the change characteristics of the potential trend are obtained. The support vector machine algorithm is used to classify the change characteristics, determine whether the potential trend points to continuous deterioration, and determine the direction of the trend prediction. Based on the trend prediction, information is obtained to determine the fault. The information is compared in multiple dimensions. If the comparison results show an abnormal consistency, it is judged as a high-probability fault risk, and the final conclusion of the risk assessment is obtained. Based on the final conclusions of the risk assessment, updated data for condition analysis is obtained. If the updated data points to a specific component, it is identified as a key monitoring target, and priority targets for internal inspection are determined to obtain potential failure trends of internal components of the transformer.

[0084] For example, in this embodiment, priority information of each monitoring point is extracted from the warning level value. For example, the warning level of high priority point is 0.81, medium priority point is 0.59, and low priority point is 0.27. After sorting the values ​​in descending order, preliminary screening is performed.

[0085] Preferably, the screening threshold is set to 0.70. When the priority is higher than this threshold, it is immediately marked as a high-risk category, thereby obtaining a preliminary risk assessment result and retaining only the locations that are truly worth further attention.

[0086] In one possible implementation, for data marked as high-risk, the system acquires real-time operational status records of the corresponding monitoring points, including historical curves of multiple parameters such as voltage, current, temperature, and vibration. By comparing the data from the most recent week with the normal baseline, it was found that the voltage fluctuation at a certain point suddenly increased from the normal ±1.2% to ±4.8%, accompanied by a local temperature increase of 2.3℃, which was determined to be an anomaly in the status analysis.

[0087] Understandably, such anomalies tend to occur in specific phase areas of the same transformer group, with their distribution range limited to a range of approximately 4 meters from the winding end to the bushing.

[0088] Specifically, this embodiment further retrieves the operating logs of various components inside the transformer based on the distribution range of the anomalies, such as the core grounding current, oil gas concentration, and winding hot spot temperature. Internal inspection of the log data reveals that the hot spot temperature of a certain high-voltage winding has consistently deviated from the upper limit of the normal range by 3.1°C over the past 48 hours. Simultaneously, the acetylene content in the oil shows a slight cumulative upward trend, which is determined to be a potential signal of component failure. The initial assessment suggests the fault may originate from localized aging of the high-voltage winding insulation.

[0089] In one embodiment, this embodiment extracts potential trend change features, such as amplitude increment rate, pulse repetition frequency, and waveform distortion degree, from continuous monitoring data of the component where the fault is initially located. A support vector machine algorithm is used to classify these feature vectors, categorizing the trend into three types: stable, slowly deteriorating, and rapidly deteriorating. When the classification result points to rapid deterioration, the predicted trend is determined to be continuous degradation, indicating that the defect may evolve into severe discharge in the short term.

[0090] It should be noted that, based on trend prediction, the system obtains multi-dimensional information for fault judgment, including historical discharge patterns, the influence of ambient temperature and humidity, and the corresponding relationship with load changes. These indicators are compared, and if three or more indicators show abnormal consistency, such as the discharge phase being concentrated around 180° of the power frequency and the amplitude significantly increasing with load, it is determined to be a high-probability fault risk, and the final risk assessment conclusion is an urgent defect.

[0091] Preferably, in this embodiment, the status analysis update data is obtained through the final conclusion of the risk assessment, and relevant abnormal characteristics, location coordinates, trend categories, and other information are recorded and integrated. If the integrated results consistently point to the same area of ​​the high-voltage winding, it is identified as a key monitoring target, and the priority targets for internal inspection are determined to be the end of the winding and the main insulation layer. This targeted monitoring method allows maintenance resources to be concentrated on the highest-risk areas, significantly improving the timeliness of defect detection and effectively reducing the probability of sudden insulation breakdown.

[0092] This embodiment proposes a precise diagnosis and location method for early latent faults in transformers. It collects vibration signals from the transformer tank surface and high-frequency acoustic signature signals using an accelerometer and an acoustic signature sensor, respectively, constructing a full-band, two-dimensional feature matrix covering the fundamental frequency to high frequencies. Wavelet transform is used to extract harmonic distortion features of the vibration signals and spectral abrupt changes in the acoustic signature signals to capture early weak anomalies. When the harmonic distortion features exceed a threshold, a support vector machine is used for feature weight allocation and correlation analysis to invert the operating status of internal components such as the core windings. A deep neural network is then used to deeply fuse the two-dimensional multi-feature matrix to achieve a comprehensive diagnostic output. Multi-channel synchronous data acquisition is introduced to verify signal integrity. A fusion algorithm is used to complete the correlation calculation between early anomalies and multi-measurement data, accurately locating the fault's spatial coordinates. Finally, based on high-resolution adaptive sampling and early warning priority judgment, a comprehensive assessment and early warning of potential fault trends in the transformer's internal components are formed. This embodiment achieves a closed-loop diagnosis across the entire chain, from weak anomaly signal capture to precise internal fault location and trend prediction, significantly improving the early detection capability and operational accuracy of latent faults in transformers.

[0093] Example 2 like Figure 2 As shown, based on the same inventive concept, this embodiment also provides a transformer intelligent monitoring system based on acoustic vibration, including: The dual-dimensional data acquisition module is used to acquire the body vibration signal from the surface of the transformer tank through an accelerometer, and at the same time, to acquire the high-frequency acoustic fingerprint signal non-contactly through an acoustic fingerprint sensor. Based on the body vibration signal and the high-frequency acoustic fingerprint signal, a fundamental frequency and high-frequency full-band feature matrix is ​​constructed to obtain dual-dimensional raw data. The weak anomaly index extraction module is used to extract the harmonic distortion features of the body vibration signal and the spectral change features of the high-frequency acoustic signature signal from the two-dimensional raw data using wavelet transform, and to determine the weak anomaly indexes associated with early faults based on the harmonic distortion features and spectral change features. The early fault analysis module is used to perform weight allocation and correlation analysis on weak anomaly indicators based on the support vector machine algorithm, invert the operating status of the internal components of the transformer, and determine the type of potential anomaly. The comprehensive diagnostic module is used to obtain a multi-feature matrix based on potential anomaly types, and uses a deep neural network to fuse the fundamental frequency details and high frequency details of the two-dimensional raw data to obtain the output of the comprehensive diagnostic model. The signal integrity determination module is used to obtain data synchronously collected from multiple measuring points of the transformer by the multi-channel expansion unit based on the output results of the comprehensive diagnostic model, and to determine the integrity of the signal capture based on the synchronously collected data. The fault location module is used to obtain the precise coordinates of the fault location by processing the correlation between multi-channel data and weak anomaly indicators through a fusion algorithm based on the integrity of the signal capture. The warning level judgment module is used to adjust the sampling rate using a high-resolution module to adapt to the weak signal amplitude based on the precise coordinates of the fault location, and to determine the priority of the warning level based on the clear signal sequence obtained after the adjustment of the sampling rate. The fault trend assessment module is used to obtain overall assessment data of the transformer's operating status based on the priority of the warning level, determine the potential fault trends of the transformer's internal components based on the overall assessment data, and obtain transformer fault monitoring results.

[0094] The intelligent transformer monitoring system based on acoustic vibration provided in this embodiment has all the advantages of the intelligent transformer monitoring method based on acoustic vibration provided in Embodiment 1.

[0095] Example 3 This embodiment also discloses a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method described in Embodiment 1.

[0096] Example 4 This embodiment also discloses a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the method described in Embodiment 1.

[0097] Example 5 This embodiment also discloses a computer program product, including a computer program that, when executed by a processor, implements the steps of the method described in Embodiment 1.

[0098] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for intelligent monitoring of transformers based on acoustic vibration, characterized in that, include: The vibration signal of the transformer tank is obtained from the surface of the transformer tank by an accelerometer, and the high-frequency acoustic signal is collected non-contactly by an acoustic sensor. The fundamental frequency and high-frequency full-band feature matrix is ​​constructed based on the vibration signal and the high-frequency acoustic signal to obtain two-dimensional raw data. Based on the dual-dimensional raw data, wavelet transform is used to extract the harmonic distortion features of the body vibration signal and the spectral change features of the high-frequency acoustic signature signal. Based on the harmonic distortion features and the spectral change features, weak abnormal indicators associated with early faults are determined. Based on the weak anomaly indicators, the support vector machine algorithm is used to perform weight allocation and correlation analysis on the weak anomaly indicators, invert the operating status of the internal components of the transformer, and determine the potential anomaly type. Based on the potential anomaly type, a multi-feature matrix is ​​obtained, and a deep neural network is used to fuse the fundamental frequency details and high frequency details of the two-dimensional original data to obtain the output result of the comprehensive diagnostic model. Based on the output of the comprehensive diagnostic model, data synchronously collected from multiple measuring points of the transformer by the multi-channel extension unit is obtained, and the integrity of signal capture is determined based on the synchronously collected data. Based on the completeness of the captured signal, the correlation between the multi-channel data and the weak anomaly indicators is processed by a fusion algorithm to obtain the precise coordinates for fault location. Based on the precise coordinates of the fault location, a high-resolution module is used to adjust the sampling rate to match the weak signal amplitude, and the priority of the warning level is determined based on the clear signal sequence obtained after the adjustment of the sampling rate. Based on the priority of the warning level, obtain overall assessment data of the transformer's operating status, determine the potential fault trends of internal components of the transformer based on the overall assessment data, and obtain transformer fault monitoring results.

2. The method according to claim 1, characterized in that, The process of obtaining the two-dimensional raw data includes: Vibration signals of the transformer body are collected from the surface of the transformer tank using an accelerometer, and high-frequency acoustic fingerprint signals are collected from the surface of the transformer tank using an acoustic fingerprint sensor. Initial two-dimensional data is constructed based on the vibration signals of the transformer body and the high-frequency acoustic fingerprint signals. The vibration signal of the body is denoised using a signal preprocessing method to obtain smoothed vibration signal data. Based on the smoothed vibration signal data, the fundamental frequency feature matrix is ​​extracted to construct a feature representation of the vibration dimension; The high-frequency feature matrix is ​​obtained by performing frequency domain transformation on the high-frequency voiceprint signal, and a feature representation of the voiceprint dimension is constructed. If the fundamental frequency feature matrix and the high frequency feature matrix have frequency band overlap, the overlapping part is weighted to obtain the fused full-band feature matrix. Based on the full-band feature matrix, two-dimensional raw data is obtained.

3. The method according to claim 1, characterized in that, The process of determining weak anomaly indicators associated with early faults based on the harmonic distortion characteristics and the spectral abrupt change characteristics includes: Based on the two-dimensional raw data, the body vibration signal and the high-frequency acoustic signature signal are classified and stored respectively to obtain a structured signal dataset. Wavelet transform is used to decompose the body vibration signal in the signal dataset and extract harmonic distortion features. Wavelet transform was used to perform frequency domain analysis on the high-frequency acoustic signature signal in the signal dataset to extract spectral abrupt change features. A comprehensive feature matrix is ​​constructed based on the harmonic distortion characteristics and the spectral mutation characteristics; If the eigenvalues ​​of the comprehensive feature matrix exceed the preset threshold range, a weak anomaly is determined to exist, and the corresponding abnormal signal segment is obtained. Based on the abnormal signal segments, the abnormal location and type associated with the fault indicators are determined, and weak abnormal indicators associated with early faults are obtained.

4. The method according to claim 1, characterized in that, The process of determining the type of potential anomaly includes: Obtain the harmonic distortion related feature value in the weak anomaly index. If the harmonic distortion related feature value exceeds the preset threshold, then use the support vector machine algorithm to assign weights to the weak anomaly index and obtain the priority ranking of each index. Based on the priority ranking, a deep correlation analysis is performed on the high-priority indicators to determine the deviation between the core state and the winding state. If the correlation analysis results show that there is a deviation in the core state or winding state, the corresponding operating status data is extracted to determine the specific manifestation of the potential anomaly. Based on the specific manifestations of the potential anomalies, obtain the distribution of anomaly types and determine whether the anomaly involves multiple internal components. If an anomaly involves multiple internal components, cross-validation is performed on the anomaly types to obtain the final anomaly classification result, and the potential anomaly types are determined based on the final anomaly classification result.

5. The method according to claim 1, characterized in that, The process of obtaining the output results of the comprehensive diagnostic model includes: Initial data is obtained based on the potential anomaly types, and the initial data is cleaned and standardized to obtain a structured multi-feature matrix. Based on the multi-feature matrix, a hierarchical processing method is performed for the feature extraction process to determine the key feature set. If the distribution of fundamental frequency details and high frequency details in the key feature set is uneven, they are separated by a signal dimension decomposition method to obtain the separated fundamental frequency signal data and high frequency signal data. A deep neural network is used to integrate the separated baseband signal data and high-frequency signal data, and the fused signal features are output. If the fused signal features reach a preset threshold, then the comprehensive diagnostic logic is used for classification processing to determine the specific category of the abnormality, generate diagnostic result data, and obtain the output result of the comprehensive diagnostic model.

6. The method according to claim 1, characterized in that, The process of determining the integrity of signal capture based on synchronously acquired data includes: Based on the output of the comprehensive diagnostic model, obtain the corresponding timestamp interval, and extract synchronously collected data from the multi-channel extension unit based on the timestamp interval; The number of data points in each channel of the synchronously acquired data is counted. If the number of data points reaches the preset number of acquisition points, the corresponding channel is marked as complete; otherwise, it is marked as missing. For channels marked as missing, find the most recent complete acquisition time point and calculate the time difference between the most recent complete acquisition time point and the current timestamp interval; If the time difference is less than the preset allowable interval, the current channel is supplemented with the previous complete data; All the supplemented channel data are time-aligned and arranged to form a unified sequence; Calculate the sampling time deviation between each channel based on the unified sequence; If the sampling time deviations are all less than the preset synchronization threshold, then the signal acquisition integrity is determined to meet the requirements.

7. The method according to claim 1, characterized in that, The process of obtaining the precise coordinates of the fault location includes: The system acquires a multi-channel raw signal sequence and uses a signal integrity verification module to determine whether the multi-channel raw signal sequence meets the capture requirements. If it does, the system retains the multi-channel raw signal sequence; otherwise, it acquires the signal again. Based on the preserved multi-channel original signal sequences, extract the early abnormal indicator sequences for each channel; A fusion algorithm is used to correlate the multi-channel original signal sequence with the early anomaly index sequence to obtain the spatiotemporal correspondence of each anomaly point in the multi-channel; The channel combination in which the anomaly occurred is determined based on the spatiotemporal correspondence described above; The precise spatial coordinates of the anomaly are calculated based on the spatial geometric distribution of the channel combination, thus obtaining the precise coordinates for fault location.

8. The method according to claim 1, characterized in that, The process of determining the priority of the warning level based on the clear signal sequence obtained from the adjusted sampling rate includes: The spatial location of the current monitoring point is determined based on the precise coordinates of the fault location, and the original weak signal of the monitoring point is acquired using a high-resolution acquisition module. The current signal strength is calculated based on the amplitude of the original weak signal; The sampling rate parameter is dynamically adjusted based on the current signal strength. The weak signal was reacquired using the adjusted sampling rate to obtain a clear signal sequence. The warning level value is determined based on the amplitude distribution of the clear signal sequence; The warning priority is marked based on the comparison result between the warning level value and the preset threshold, and the priority of the warning level is obtained.

9. The method according to claim 1, characterized in that, The process of determining potential failure trends of internal components of the transformer based on the overall assessment data includes: Priority ranking information is obtained from the priority data of the warning level, and the priority ranking information is initially screened. If the priority order is higher than the preset threshold, the corresponding data is marked as a high-risk category. For data marked as high-risk, real-time records of transformer operating status are obtained, and these real-time records are compared and analyzed with historical data. If abnormal fluctuations are found, they are identified as anomalies, and the distribution range of the anomalies is determined. Based on the distribution range of the abnormal points, the operation logs of the internal components of the transformer are obtained. If the operating parameters of the components deviate from the normal range, it is judged as a potential signal of component failure, and the initial fault location is obtained. Based on the component data of the preliminary fault location, the change characteristics of the potential trend are obtained, and the support vector machine algorithm is used to classify the change characteristics to determine whether the potential trend points to continuous deterioration and to determine the direction of the trend prediction. Based on the trend prediction, information for fault judgment is obtained. The information is compared in multiple dimensions. If the comparison results show an abnormal consistency, it is judged as a high-probability fault risk, and the final conclusion of risk assessment is obtained. Based on the final conclusion of the risk assessment, updated data for status analysis is obtained. If the updated data points to a specific component, it is identified as a key monitoring target, and priority targets for internal inspection are determined to obtain the potential failure trends of internal components of the transformer.

10. A transformer intelligent monitoring system based on acoustic vibration, characterized in that, include: The dual-dimensional data acquisition module is used to acquire the body vibration signal from the surface of the transformer tank through an accelerometer, and at the same time, to acquire the high-frequency acoustic fingerprint signal non-contactly through an acoustic fingerprint sensor. Based on the body vibration signal and the high-frequency acoustic fingerprint signal, a fundamental frequency and high-frequency full-band feature matrix is ​​constructed to obtain dual-dimensional raw data. The weak anomaly index extraction module is used to extract the harmonic distortion features of the body vibration signal and the spectral change features of the high-frequency acoustic text signal using wavelet transform based on the dual-dimensional raw data, and to determine the weak anomaly index associated with the early fault based on the harmonic distortion features and the spectral change features. The early fault analysis module is used to perform weight allocation and correlation analysis on the weak anomaly indicators based on the weak anomaly indicators using the support vector machine algorithm, to invert the operating status of the internal components of the transformer and determine the potential anomaly type. The comprehensive diagnostic module is used to obtain a multi-feature matrix based on the potential abnormality type, and to fuse the fundamental frequency details and high frequency details of the two-dimensional original data using a deep neural network to obtain the output result of the comprehensive diagnostic model. The signal integrity determination module is used to obtain data synchronously collected from multiple measuring points of the transformer by the multi-channel extension unit based on the output results of the comprehensive diagnostic model, and to determine the integrity of the signal capture based on the synchronously collected data; The fault location module is used to process the correlation between multi-channel data and the weak anomaly indicators through a fusion algorithm based on the integrity of the signal capture to obtain the precise coordinates of the fault location. The warning level judgment module is used to adjust the sampling rate using a high-resolution module to adapt to the weak signal amplitude based on the precise coordinates of the fault location, and to judge the priority of the warning level based on the clear signal sequence obtained after the adjustment of the sampling rate. The fault trend assessment module is used to obtain overall assessment data of the transformer's operating status according to the priority of the warning level, determine the potential fault trends of the transformer's internal components based on the overall assessment data, and obtain transformer fault monitoring results.

Images