Distribution transformer monitoring method and device

Abstract

The invention discloses a distribution transformer monitoring method and device, and relates to the technical field of transformer monitoring. The method comprises the following steps: acquiring a core texture signal under steady-state operation of the distribution transformer, and constructing a reference texture signal through preprocessing; collecting a real-time core texture signal, respectively extracting features of the real-time core texture signal and the reference texture signal through variational mode decomposition and a symmetric point mode algorithm to obtain a real-time feature image and a reference feature image, and comparing the real-time feature image and the reference feature image through an improved scale invariant feature transformation algorithm to obtain a real-time feature image; identifying a suspected fault signal feature image and a corresponding real-time core texture signal; inputting the suspected fault signal feature image into the improved deep residual shrinkage network model, and outputting the fault type of the distribution transformer; and performing space-time synchronization analysis on the suspected fault signal, and calculating a fault diagnosis result in combination with a fault type result, thereby realizing distribution transformer monitoring.

Description

Technical Field

[0001] This invention relates to the field of transformer technology, specifically to a method and device for monitoring distribution transformers. Background Technology

[0002] As a core piece of equipment in the power distribution network, the distribution transformer bears the critical responsibility of voltage transformation, power distribution, and transmission. Its operating status directly affects the reliability of power supply and the safety of electricity consumption. During long-term operation, affected by electromagnetic forces, mechanical vibrations, environmental factors, and material aging, distribution transformers are prone to latent faults such as mechanical loosening, electromagnetic imbalance, and localized overheating. These faults have subtle early characteristics and evolve slowly; if not identified in time, they may gradually develop into serious faults.

[0003] In traditional distribution transformer fault monitoring, a single scale-invariant feature transform algorithm is often used for fault feature identification. This method acquires feature images of the transformer under operating conditions, extracts scale-invariant feature points from the images, constructs a feature description vector, and then compares it with a preset normal state feature vector. Based on the matching result, it determines whether the equipment has a fault. This method has been applied to some scenarios where conventional fault features are relatively obvious.

[0004] However, the fault characteristic signals of distribution transformers are mostly broadband multi-mode signals, containing complex information from multiple frequency bands such as acoustic patterns, vibration patterns, and thermal acoustic patterns. Traditional scale-invariant feature transformation algorithms do not perform targeted frequency band decomposition and feature optimization on the original signals, and cannot effectively separate useful signals from interference noise in different frequency bands. As a result, they cannot accurately capture the subtle feature differences corresponding to early latent faults, which ultimately leads to mismatches and missed matches during the fault identification process, making it difficult to meet the actual needs of accurate early fault monitoring. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method and apparatus for monitoring distribution transformers, thereby resolving the problems existing in the background technology.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method and apparatus for monitoring distribution transformers, comprising the following steps: S1: Collect the core texture signal of the distribution transformer under steady-state operation, and preprocess the core texture signal to obtain the reference texture signal; S2: Acquire real-time core texture signals, extract features from the real-time core texture signals using variational mode decomposition and symmetric point pattern algorithms, extract features from the reference texture signals using variational mode decomposition and symmetric point pattern algorithms, and obtain corresponding real-time feature images and reference feature images; compare the features of the real-time feature images and reference feature images using an improved scale-invariant feature transformation algorithm, identify suspected fault signal feature images, and record the real-time core texture signals corresponding to the suspected fault signal feature images as suspected fault signals; S3: Input the suspected fault signal feature image into the improved deep residual shrinkage network model, and output the distribution transformer fault type result; S4: Perform spatiotemporal synchronization analysis on suspected fault signals to obtain spatiotemporal synchronization analysis results. Based on the fault type results of the distribution transformer and the spatiotemporal synchronization analysis results, calculate the fault diagnosis results to realize the monitoring of the distribution transformer.

[0007] Preferably, the acquisition of the core texture signal of the distribution transformer under steady-state operating conditions includes the following specific steps: Core texture signals of the distribution transformer under steady-state operation are collected. These core texture signals include acoustic fingerprint signals, vibration fingerprint signals, and thermal acoustic fingerprint signals. Acoustic fingerprint signals are collected using electret condenser-type directional microphones. The microphones are symmetrically distributed around the transformer tank, specifically positioned near the core clamping parts, at the winding ends, in the heat sink area, and at four key structural locations: the load-bearing parts of the tank. One microphone is placed at each location, pointing perpendicular to the transformer tank surface, horizontally 30cm to 150cm from the tank, and 80cm above the ground. A temperature sensor is attached to each microphone. Vibration fingerprint signals are collected using piezoelectric vibration acceleration sensors, which are attached to the outer wall of the tank corresponding to the microphone placement location. Thermal acoustic fingerprint signals are collected synchronously using the aforementioned microphones and temperature sensors. Before starting data acquisition, it is necessary to determine whether the transformer is in steady-state operation. The steady-state determination conditions are set as follows: within 24 consecutive hours, the transformer load fluctuation amplitude is ≤ ±5%, the three-phase current imbalance is ≤ 2%, the oil temperature change rate is ≤ 0.5℃ / h, and there is no obvious external interference. When all the above conditions are met, it is determined to be in steady-state operation, data acquisition is started, and the core texture signal is finally obtained.

[0008] Preferably, the specific steps for preprocessing the core texture signal to obtain the reference texture signal are as follows: For the collected voiceprint signals (t) Preprocessing is performed, and based on the preprocessed voiceprint signal, three types of features are extracted to construct a voiceprint benchmark: The formula for calculating spectral complexity is:

[0009] in, This represents the proportion of the power spectral density of the acoustic signature signal at frequency f. The upper limit of the effective frequency, For spectral complexity; The formula for calculating the main frequency is:

[0010] in, The power percentage of the 50Hz harmonic component. For the frequency multiplication factor, Main frequency; The formula for calculating the high-low frequency ratio is:

[0011] in, The harmonic number corresponding to the high-low frequency boundary. High-low frequency ratio; The above three types of features are arranged into a voiceprint feature matrix according to the time series. Statistical analysis was performed on the voiceprint feature matrix to calculate the mean and standard deviation of each feature, and a voiceprint baseline interval was constructed. ,in, This represents the mean of the voiceprint features. The standard deviation of voiceprint features; For the acquired vibration ripple signal (t) Preprocessing: A 50Hz notch filter is used to remove power frequency interference, wavelet threshold denoising is used to eliminate environmental vibration noise, and three types of features are extracted to construct a ripple reference: The formula for calculating the vibration amplitude distribution parameter is:

[0012] in, This represents the peak-to-peak value of the ripple signal. The effective value of the ripple signal. These are the vibration amplitude distribution parameters; The formula for calculating the periodic alternation characteristic parameter is:

[0013] in, Let be the autocorrelation function of the ripple signal. To delay time, These are periodic alternating characteristic parameters; The formula for calculating the energy transfer path parameters is:

[0014] in, For the first The first collection site and the first Cross-correlation coefficients of vibration signals at each acquisition location The number of combinations of 4 sampling sites. For energy transfer path parameters, and Index for the collection site; The three types of features are combined to form a ripple feature matrix. Statistical analysis was performed on the ripple characteristic matrix to construct the ripple reference interval. ,in, The mean value of the vibration ripple characteristics. The standard deviation of the vibration ripple characteristics; For the collected voiceprint signals Preprocessing is performed, and the preprocessing process is consistent with that of the voiceprint signal. Two types of features are extracted to construct a thermal voiceprint benchmark: The formula for calculating the stripe density parameter is:

[0015] in, The total number of fringes in the time spectrum. The effective frequency range of thermal acoustic waveforms. This represents the total number of signal sampling points. For stripe density parameters, This is the correction factor for the rate of temperature change; The formula for calculating the orientation parameter is:

[0016] in, Let be the angle between the k-th stripe and the horizontal direction. These are the arrangement direction parameters; Combining the two types of features into a thermal soundprint feature matrix Statistical analysis was performed on the thermal acoustic signature feature matrix to construct a thermal acoustic signature baseline interval. ,in, The mean value of the vibration ripple characteristics. The standard deviation of the vibration ripple characteristics; By integrating the feature matrices of voiceprints, vibration patterns, and thermal voiceprints with the benchmark interval, a multimodal texture benchmark model is constructed. ,in These are the reference ranges for soundprint, vibration, and thermal soundprint, respectively. Select the signal whose feature value is within ±5% of the center of the reference interval from the core texture signal and use it as the reference texture signal.

[0017] Preferably, the process of acquiring real-time core texture signals, extracting features from the real-time core texture signals using variational mode decomposition and symmetric point pattern algorithms, and extracting features from the reference texture signals using variational mode decomposition and symmetric point pattern algorithms to obtain corresponding real-time feature images and reference feature images includes the following steps: Real-time core texture signals are acquired, including real-time acoustic fingerprint signals, real-time vibration fingerprint signals, and real-time thermal acoustic fingerprint signals, wherein the real-time acoustic fingerprint signal is denoted as... Real-time vibration ripple signal is denoted as Real-time thermal acoustic signal is recorded as The acquisition parameters for the three types of signals remain consistent with step S1. The real-time core texture signals undergo unified preprocessing, and the preprocessed real-time core texture signals are then subjected to variational mode decomposition. The complete mathematical model includes the objective function, constraints, and augmented Lagrangian function.

[0018]

[0019] in, Let K be the k-th intrinsic modal component, and K be the number of modes. Let the center frequency of the k-th IMF component be . For the Dirac impulse function, The imaginary unit, Refers to the preprocessed real-time signal. As a penalty factor, For Lagrange multipliers, For inner product operations, To augment the Lagrange function; The five IMF components obtained from the decomposition of the real-time core texture signal and the original signal are subjected to symmetric point mode transformation to convert the one-dimensional time-domain signal into a two-dimensional polar coordinate feature image. The core transformation formula is as follows:

[0020]

[0021]

[0022] in, For the first The radius of each sampling point in polar coordinates Let be the amplitude of the i-th sampling point of the IMF component. and These are the maximum and minimum amplitudes of the IMF component, respectively. For the first The counterclockwise rotation angle of each mirror-symmetrical part The number of symmetrical parts is 6. The angle is clockwise rotation. The lag factor has a value of 20. This is the angle gain factor, with a value of 30 degrees. The maximum radius of the polar coordinate feature image. For sampling point index, For mirror cell index; For the real-time core texture signal, extract a set of spiral arms from each of its K IMF components and the SDP transformation result of the original signal, and fuse them to obtain a real-time feature image. The fusion rule is: according to the center frequency of the IMF components from low to high, the spiral arms of each component are superimposed on the corresponding mirror unit of the polar coordinate image. Similarly, feature extraction of the baseline texture signal is performed using variational mode decomposition and symmetric point pattern algorithms to obtain the baseline feature image.

[0023] Preferably, the step of comparing the real-time feature image and the reference feature image using an improved scale-invariant feature transform algorithm to identify the suspected fault signal feature image, and recording the real-time core texture signal corresponding to the suspected fault signal feature image as the suspected fault signal, includes the following steps: The feature images are subjected to Gaussian blurring and downsampling at different scales to generate a 6-layer Gaussian pyramid, each layer containing 5 Gaussian difference images, with a scale factor of [missing information]. =1.6. In the difference of Gaussian image, each pixel is compared with its 8 neighboring pixels and corresponding pixels in the upper and lower layers. Extreme points are selected as candidate feature points. Candidate points with low contrast and edge response are eliminated, and stable feature points are retained. A 16×16 neighborhood window is constructed with the feature points as the center, which is divided into 4×4 sub-regions. Gradient histograms in 8 directions are calculated for each sub-region to generate a 128-dimensional feature description vector. Calculate the Euclidean distance between the feature descriptor vector of the real-time feature image and the feature descriptor vector of the reference feature image. Perform initial matching using the nearest neighbor to second nearest neighbor ratio criterion, retaining those that satisfy the criteria. The matching pairs, where Nearest neighbor distance The next nearest neighbor distance; Three feature points A, B, and C, and their corresponding matching points A', B', and C', are randomly selected from the initial matching pairs to construct a triangle. ABC and The match A'B'C' is validated using the following formula:

[0024] in, and Triangles ABC The lengths of the sides corresponding to A'B'C' and These are the included angles of the corresponding vertices. The threshold for consistency in side length ratios is 0.1. and As the angular consistency threshold, The value is 1. The value is 20; Iterate through all initial matching pairs, repeat the similar triangle verification, and count the number of voiceprint matching points. Number of ripple matching points thermal voiceprint matching points ; Calculate the average number of matching points between the baseline feature image and the real-time feature image. Set a single-mode matching threshold ; When the number of matching points in a feature image is lower than the corresponding threshold, it is marked as a suspected fault signal feature image, and the real-time core texture signal corresponding to the suspected fault signal feature image is recorded as a suspected fault signal.

[0025] Preferably, the base inputs the suspected fault signal feature image into the improved deep residual shrinkage network model and outputs the distribution transformer fault type result, including the following steps: The feature images of suspected fault signals are normalized, and channels are stitched together in the order of acoustic pattern-vibration pattern-thermal acoustic pattern to obtain a fused feature matrix. For the fused feature matrix Principal component analysis was performed to reduce the dimensionality, resulting in the core feature matrix. The feature map is then reshaped into a 32×32×1 two-dimensional feature map, and an improved deep residual shrinkage network model is constructed. The loss function of the improved deep residual shrinkage network adopts the multi-class cross-entropy loss function, and the calculation formula is as follows:

[0026] in, For the batch sample size, For the first The true label of each sample For the network prediction of the first Class failure probability, For multi-class cross-entropy loss, For sample index, Index for fault categories; core feature matrix Inputting the trained improved deep residual shrinkage network model, the output yields the probability distribution of various fault types. These correspond to normal operation, mechanical loosening, electromagnetic imbalance, and localized overheating, respectively. The category with the highest probability is selected as the result of the distribution transformer fault type.

[0027] Preferably, the construction of the improved deep residual shrinkage network model includes the following specific steps: The overall structure of the improved deep residual shrinking network consists of: an input layer, a convolutional layer, four improved residual shrinking units, a BN layer, an APReLu layer, a global average pooling layer, a fully connected layer, and an output layer. The adaptive parameter correction linear unit APReLu is used, and the core formula is:

[0028]

[0029]

[0030] in, This is the output feature vector of the APReLu layer. For adaptive weights, Learn intermediate variables for weights. This is the weight matrix. For bias terms, For the input feature vector, This is the output result after performing global average pooling on the input feature vector x; Each Improved Residual Shrinkage Unit (IRSBU) contains the following main path: convolutional layer, Batch Normalization (BN) layer, APReLU layer, convolutional layer again; and the threshold learning path: absolute value layer, GAP layer, fully connected layer, sigmoid layer. The specific formula is as follows:

[0031] in, Let c be the feature matrix of the c-th channel. This is the feature matrix of the c-th channel after soft thresholding. The adaptive threshold for the c-th channel. This is the channel-independent scaling factor.

[0032] Preferably, the step of performing spatiotemporal synchronization analysis on the suspected fault signal to obtain the spatiotemporal synchronization analysis results includes the following specific steps: The spatiotemporal synchronization analysis results include a time synchronization coefficient and a spatial synchronization coefficient. The formula for calculating the time synchronization coefficient is as follows:

[0033] in, For the first Class modality and the first The time synchronization coefficient for modalities; the closer the absolute value is to 1, the stronger the time synchronization. For the first Modal fault characteristic time series, For time sampling point index, For the first The mean of the time series of characteristics of modal faults. For the first Modal fault characteristic time series, For the first The mean of the time series of characteristics of modal faults. and For modal fault feature index; Settings | | 0.85 indicates strong synchronization. It is a weak synchronization. For asynchronous conditions, if two or more modes meet the strong synchronization condition, it is determined to be a composite fault; if only one mode exhibits fault characteristics and is asynchronous with other modes, it is determined to be a single fault. The formula for calculating the spatial synchronization coefficient is:

[0034] in, For the first Class modality and the first The spatial synchronization coefficient of a modality, the closer it is to 1, the higher the spatial overlap. For the first The spatial region corresponding to the modal fault characteristics for and The number of sensor points in the intersection area. for and The number of sensor points in the union region. For the first Spatial region corresponding to modal fault characteristics; when =1 indicates that the fault characteristics are concentrated in the same spatial region; when =0 indicates that the fault characteristics are distributed in different regions. Combined with the time synchronization coefficient, composite faults can be further divided into intra-regional composite faults and cross-regional composite faults.

[0035] Preferably, the fault diagnosis result is calculated based on the distribution transformer fault type result and the spatiotemporal synchronization analysis result, specifically as follows: The formula for locating mechanical loosening faults is:

[0036] in, The results of locating the mechanical loosening fault. For sensor indexing, The average amplitude of the low-frequency IMF component of the ripple pattern acquired by the l-th sensor. The average amplitude of the low-frequency IMF component of the ripple corresponding to the l-th sensor in the baseline model. Let be the absolute value of the deviation between the amplitude of the vibration ripple in the l-th region and the reference. The deviation rate; The formula for locating electromagnetic off-center load faults is:

[0037] in, The results of electromagnetic off-center load fault location. The spectral distortion of the intermediate frequency (IMF) component of the acoustic signature acquired by the l-th sensor is given. is the local time synchronization coefficient between the acoustic signature and the vibration signature in the l-th region; The formula for locating localized overheating faults is:

[0038]

[0039]

[0040] in, The results show the location of the localized overheating fault. The density of the thermoacoustic stripes collected by the l-th sensor. The mean stripe density in the baseline model. For density increment, The total energy of the high-frequency IMF component of the thermal acoustic signature collected by the l-th sensor. The energy mean in the baseline model. For energy increment; Composite fault localization employs a regional overlay and weight allocation strategy: for composite faults within the same region, the intersection of the localization results of each individual fault is used; for composite faults across regions, the network output probability of each individual fault is used. Assign weights, the weight formula is as follows Where n is the number of composite fault types, the region with the largest weight is the primary fault region, and the rest are secondary fault regions. Finally, the fault diagnosis results are obtained, which include the fault type results and fault location results of the distribution transformer.

[0041] A distribution transformer monitoring transducer includes 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 above method.

[0042] This invention provides a method for monitoring distribution transformers, which involves machine learning and deep learning technologies, and has the following beneficial effects: (1) VMD-SDP feature extraction adapts to the wideband multimodal signal characteristics of distribution transformers, can decompose complex acoustic, vibration and thermal signals into narrowband components and convert them into texture images, clearly present the frequency band feature differences corresponding to different faults, avoid feature distortion caused by signal aliasing, provide intuitive and accurate feature support for subsequent fault identification, and fit the signal monitoring scenario of equipment operation.

[0043] (2) Improved SIFT feature matching relies on the feature images extracted by VMD-SDP. It eliminates mismatch points through unique similarity verification criteria, closely matches the comparison requirements of multimodal texture benchmark models, accurately captures the texture shift of early hidden faults, makes up for the lack of accuracy of traditional matching methods in complex equipment operating environments, and makes suspected fault screening more in line with actual monitoring scenarios.

[0044] (3) Based on the improved deep residual shrinkage network, it can effectively integrate multimodal features, suppress redundant information of environmental interference through adaptive activation and threshold mechanism, accurately distinguish various fault types such as mechanical loosening and electromagnetic off-center load, solve the problem of fuzzy identification of complex faults in distribution transformers, provide clear and reliable type judgment for on-site maintenance, and adapt to the actual needs of equipment fault handling. Attached Figure Description

[0045] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 This is a flowchart of the steps of a power distribution transformer monitoring method proposed in this invention; Figure 2 This is a step hierarchy diagram of obtaining a suspected fault signal feature image in a power distribution transformer monitoring method proposed in this invention; Figure 3 This is a hierarchical diagram of the steps involved in obtaining fault diagnosis results in a power distribution transformer monitoring method proposed in this invention. Detailed Implementation

[0047] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0048] Please see Figures 1-3 The present invention provides a technical solution: a method for monitoring distribution transformers.

[0049] S1: Collect the core texture signal of the distribution transformer under steady-state operation, and preprocess the core texture signal to obtain the reference texture signal; The core texture signal of the distribution transformer under steady-state operating conditions is collected. The core texture signal includes acoustic fingerprint signal, vibration fingerprint signal, and thermal acoustic fingerprint signal. The steps are as follows: The first step is to deploy the monitoring system and determine the steady-state operating conditions. A non-contact monitoring system was constructed to collect data: acoustic signature signals were acquired using electret condenser gun-type directional microphones with a frequency response range of 20Hz~20kHz, sensitivity ≤-33dB±3dB, and dynamic range ≥107dB. These microphones were evenly distributed around the transformer tank according to a symmetrical arrangement, specifically near the core clamping parts, winding ends, heat sink area, and load-bearing parts of the tank—four key structural locations. One microphone was placed at each location, pointing perpendicular to the transformer tank surface, horizontally 30cm~150cm from the tank, and 80cm above the ground to avoid external interference sources such as the water cooling system. At the microphone locations in the heat sink area and load-bearing parts of the tank, patch-type temperature sensors (measuring range -20℃~150℃, accuracy ±0.5℃) were simultaneously attached, ensuring perfect alignment with the microphone timestamps. Vibration ripple signals were acquired using piezoelectric vibration accelerometers with a measurement range of ±250g, a maximum frequency response of 30kHz, and a turns ratio of 9.80665 / 0.025 (m / s²). 2 The data value is pasted on the outer wall of the enclosure at the corresponding microphone placement position to ensure spatial alignment of the acquisition positions of the vibration signal and the acoustic signature signal; the thermal acoustic signature signal is synchronously acquired through the aforementioned microphones, and indirectly captured by utilizing the modulation effect of the hot air flow caused by the temperature change in the heat sink area on the acoustic signature features.

[0050] Before initiating data acquisition, it is necessary to determine whether the transformer is in steady-state operation. The steady-state determination criteria are: within 24 consecutive hours, the transformer load fluctuation amplitude is ≤ ±5%, the three-phase current imbalance is ≤ 2%, the oil temperature change rate is ≤ 0.5℃ / h, and there is no significant external interference (such as the start-up and shutdown of nearby equipment, sudden load impacts, etc.). If all the above conditions are met, it is determined to be in steady-state operation, and data acquisition is initiated; if not, monitoring continues until the operating conditions stabilize to ensure that the acquired reference signal is not affected by transient operating conditions.

[0051] The data acquisition parameters were set as follows: microphone sampling frequency was set to 48kHz, sampling accuracy to 24-bit, single-segment acquisition duration to 10s, acquisition interval to 5min, and continuous acquisition for 30 days; vibration acceleration sensor sampling frequency was set to 50kHz, sampling accuracy to 16-bit. The acquisition parameters were synchronized with the acoustic signature signal to ensure complete alignment of the timestamps of the two types of signals, laying the foundation for subsequent correlation analysis. During the acquisition process, transformer operating parameters (including load current, voltage, oil temperature, ambient temperature, etc.) were recorded in real time as auxiliary calibration information for the benchmark model.

[0052] Next, texture benchmarks for acoustic texture, vibration texture, and thermal acoustic texture are constructed separately. The construction process of each modal benchmark is independent but interrelated: The acoustic signature benchmark focuses on reflecting the inherent acoustic characteristics of the coupling between the electromagnetic force and mechanical structure of a transformer, and is modeled by extracting multidimensional features from the acoustic signature signal. First, the original acoustic signature signal is collected... (t) Preprocessing: A Hamming window (2048 points, 50% overlap) is used for frame segmentation to suppress spectral leakage; a first-order Butterworth filter is used to filter low-frequency noise below 20Hz and high-frequency interference above 18kHz, retaining the effective signal frequency band. It should be noted that the core characteristics of the acoustic signature signal of the distribution transformer depend on the authenticity of the spectral structure, and the phase distortion of the first-order filter is much lower than that of the higher-order filter. The smooth roll-off characteristic of the first-order filter can avoid excessive attenuation of edge signals of the effective frequency band; the acoustic signature signal of the distribution transformer is dominated by electromagnetic force and mechanical vibration, and the highest effective frequency does not exceed 18kHz.

[0053] Based on the preprocessed signal, three types of core features are extracted to construct a voiceprint benchmark: The formula for calculating spectral complexity is:

[0054] in, This represents the proportion of the power spectral density of the acoustic signature signal at frequency f. The upper limit of the effective frequency is set at 18kHz. The complexity is spectral.

[0055] This formula quantifies the complexity of the spectrum distribution by calculating the information entropy of the power spectrum of the acoustic signal. Under normal operation, the electromagnetic force and mechanical vibration of the transformer are stable, and the spectrum complexity shows a stable distribution. The value range is determined to be 7.15~7.6 by statistically analyzing 30 days of steady-state data.

[0056] The formula for calculating the main frequency is:

[0057] in, The power percentage of the 50Hz harmonic component. For the frequency multiplication factor, The main frequency.

[0058] The core significance of this formula is to locate the main frequency position where the energy is most concentrated in the acoustic signal. The main frequency of the distribution transformer is mainly dominated by the magnetostrictive effect of the iron core and the electromagnetic force of the winding. Under normal operating conditions, the main frequency is concentrated within 1kHz. This parameter is used to capture the inherent acoustic characteristics of the equipment.

[0059] The formula for calculating the high-low frequency ratio is:

[0060] in, The harmonic number corresponding to the high-low frequency boundary. =24, This refers to the high-to-low frequency ratio.

[0061] This formula reflects the frequency distribution characteristics of the voiceprint signal by calculating the ratio of power proportions in the high and low frequency bands. During normal operation, the high-low frequency ratio is stable between 1.2 and 1.8.

[0062] The above three types of features are arranged into a voiceprint feature matrix according to the time series. Statistical analysis was performed on the feature matrix collected over 30 days, and the mean of each feature was calculated. and standard deviation Constructing the voiceprint reference range (k=1,2,3), while preserving the time evolution trend of the feature matrix to form a dynamic voiceprint benchmark.

[0063] Ripple references are used to characterize the inherent vibration characteristics of transformer mechanical structures. Modeling is achieved by analyzing the amplitude distribution, periodic alternation characteristics, and energy transfer paths of the ripple signals. The raw ripple signals acquired by the vibration acceleration sensor are used for this purpose. (t) Preprocessing: A 50Hz notch filter is used to remove power frequency interference, and wavelet threshold denoising is used to eliminate environmental vibration noise (Sym8 wavelet basis is selected, the number of decomposition layers is 4, and the threshold is determined by the improved Birge-Massart criterion).

[0064] The following steps are taken to extract three types of core features to construct the vibration criterion benchmark: The formula for calculating the vibration amplitude distribution parameter is:

[0065] in, This represents the peak-to-peak value of the ripple signal. The effective value of the ripple signal. These are the vibration amplitude distribution parameters.

[0066] This parameter reflects the uniformity of vibration amplitude distribution. Under normal operation, the mechanical structure is tight, and the amplitude distribution is stable. The value range is 3.2 to 4.5.

[0067] The formula for calculating the periodic alternation characteristic parameter is:

[0068] in, Let be the autocorrelation function of the ripple signal. To delay time, These are periodic alternating characteristic parameters.

[0069] It should be noted that the formula for calculating the autocorrelation function of the ripple signal is: ,in, The signal acquisition duration is set to 10 seconds. This formula represents the capture of the natural period of the vibration signal. The vibration period of a distribution transformer is related to the power frequency and its harmonics. Under normal operating conditions... It stabilized at around 0.02 seconds.

[0070] The formula for calculating the energy transfer path parameters is:

[0071] in, For the first The first collection site and the first Cross-correlation coefficients of vibration signals at each acquisition location The number of combinations of 4 sampling sites. These are the parameters for the energy transfer path.

[0072] It should be noted that, , =1, 2, 3, 4 correspond to the vicinity of the core clamping parts, the winding ends, the heat sink area, and the load-bearing parts of the housing, respectively. This parameter reflects the consistency of vibration energy transmission on the equipment surface during normal operation. 0.85.

[0073] Similarly, the three types of features are combined into a ripple feature matrix. Calculate the average of 30 days of data. and standard deviation Constructing the vibration ripple reference range (k=1,2,3) forms a dynamic vibration ripple reference.

[0074] The thermal acoustic signature benchmark focuses on reflecting the thermal state of the transformer's heat dissipation structure. It achieves modeling by capturing the modulation effect of hot airflow on acoustic signature features in the heat sink area, based on acoustic signature signals collected by microphones in the heat sink area. Based on the temperature signal, the temperature change rate is calculated, and time periods with a temperature change rate > 0.1℃ / s are selected (the temperature change during this period has a significant modulation effect on the acoustic signal, making it a key time period for thermal acoustic fingerprints). For the acoustic fingerprint signal during this period, wavelet packet decomposition is used to extract frequency bands positively correlated with the temperature change rate (by calculating the Pearson correlation coefficient between the energy of each frequency band and the temperature change rate, frequency bands with a correlation coefficient > 0.6 are selected). These frequency bands are the effective frequency bands of the thermal acoustic fingerprint signal. The selected effective frequency bands are reconstructed to obtain a pure thermal acoustic fingerprint signal. The remaining frequency bands are reconstructed into ordinary acoustic fingerprint signals, achieving accurate separation of the two types of signals and avoiding aliasing. Two core features are extracted from the pure thermal acoustic fingerprint signal: The formula for calculating the stripe density parameter is:

[0075] in, The total number of stripes in the time spectrum is obtained through an edge detection algorithm. The effective frequency range of thermal acoustic signature is 50Hz~8kHz. This represents the total number of signal sampling points. For stripe density parameters, This is a temperature change rate correction coefficient, which enhances the effect of temperature modulation on thermal acoustic characteristics.

[0076] It should be noted that the formula for calculating the temperature change rate correction factor is as follows: The stripe density parameter reflects the intensity of hot air flow; heat dissipation is stable during normal operation. The value range is 15~25 lines / (kHz・s).

[0077] The formula for calculating the orientation parameter is:

[0078] in, Let be the angle between the k-th stripe and the horizontal direction. This refers to the arrangement direction parameter.

[0079] This parameter reflects the dominant direction of hot air flow; under normal operating conditions, hot air flows upwards in the heat sink area. It stabilizes between 75° and 85°.

[0080] Combining the two types of features into a thermal soundprint feature matrix Calculate the average of 30 days of data. and standard deviation Constructing a thermal acoustic signature reference range This forms a dynamic thermal acoustic signature.

[0081] Finally, a multimodal benchmark model fusion and optimization are performed. The feature matrices of acoustic texture, vibration texture, and thermal acoustic texture are integrated with the benchmark interval to construct a multimodal texture benchmark model. ,in The benchmark intervals are defined for acoustic signatures, vibration signatures, and thermal signatures. To improve the adaptability of the benchmark model, a sliding window update mechanism is adopted: the window length is set to 7 days, and the benchmark model is updated every 24 hours. The mean and standard deviation of the features within the new window are calculated, and the corresponding parameters in the original benchmark are replaced. This allows the benchmark model to adapt to the natural structural evolution and load fluctuations of the transformer during long-term operation. Simultaneously, a benchmark validity verification mechanism is established: if the mean change of a certain feature exceeds 5% of the initial mean after three consecutive updates, manual verification is triggered to eliminate the influence of sensor malfunctions or latent equipment defects, ensuring the reliability of the benchmark model.

[0082] The 10 sets of data with the most stable features (satisfying that each feature value is within ±5% of the center of the baseline interval) are selected from the 30-day core texture signal and used as the baseline texture signal.

[0083] The multimodal texture benchmark model constructed through the above process can comprehensively capture the acoustic-vibration-thermal coupling characteristics of distribution transformers during steady-state operation, providing a unified and dynamic reference framework for texture offset identification and fault diagnosis in subsequent steps.

[0084] S2: Acquire real-time core texture signals, extract features from the real-time core texture signals using variational mode decomposition and symmetric point pattern algorithms, extract features from the reference texture signals using variational mode decomposition and symmetric point pattern algorithms, and obtain corresponding real-time feature images and reference feature images; compare the features of the real-time feature images and reference feature images using an improved scale-invariant feature transformation algorithm, identify suspected fault signal feature images, and record the real-time core texture signals corresponding to the suspected fault signal feature images as suspected fault signals; Real-time core texture signals are acquired, including real-time acoustic fingerprint signals, real-time vibration fingerprint signals, and real-time thermal acoustic fingerprint signals, wherein the real-time acoustic fingerprint signal is denoted as... Real-time vibration ripple signal is denoted as Real-time thermal acoustic signal is recorded as The acquisition parameters for the three types of signals are kept consistent with those in step 1 (soundprint sampling frequency 48kHz, vibration ripple sampling frequency 50kHz, single-segment acquisition duration 10s) to ensure complete matching with the feature dimensions and time scale of the benchmark model and avoid feature shifts caused by parameter differences.

[0085] First, the real-time core texture signal undergoes unified preprocessing. The core objective is to suppress environmental interference and enhance signal sparsity, providing high-quality input for subsequent VMD decomposition and SDP transformation. right , , Hamming windows were used for framing, with a window length of 2048 points (corresponding to a window length of approximately 42.67ms for acoustic signature signals and approximately 40.96ms for vibration signature signals), and an overlap rate of 50%. This parameter setting is based on the fact that the acoustic and vibration signals of distribution transformers are mainly concentrated in the low-to-mid-frequency band. An excessively long window length would result in insufficient time resolution, making it impossible to capture instantaneous fault characteristics; an excessively short window length would lead to spectral leakage. A 50% overlap rate can compensate for signal discontinuities caused by framing, ensuring spectral integrity.

[0086] The framed signal is filtered using a first-order Butterworth bandpass filter. The filtering frequency band for acoustic and thermal acoustic signals is 20Hz~18kHz, and the filtering frequency band for vibration signal is 0~25kHz, which is consistent with the effective frequency band of the reference signal in step S1. This filters out low-frequency vibration interference (such as ground vibration) and high-frequency electromagnetic noise (such as inverter interference) in the environment.

[0087] The filtered signal is subjected to single-source time-frequency point detection based on phase angle. By calculating the ratio deviation between the real and imaginary parts of the signal, signals that meet the requirements are selected. The time and frequency points, among which , These represent the time and frequency points of the i-th type of signal, respectively. The real and imaginary parts of the vector at the given point. =0.25. The significance of this operation is to highlight the signal component corresponding to a single fault source, suppress signal aliasing caused by multi-source interference, and enable subsequent VMD decomposition to extract fault characteristic frequency bands more accurately. The value of 0.25 is based on verification through 100 sets of fault simulation experiments. This threshold can eliminate more than 90% of the interference time and frequency points while retaining the effective signal.

[0088] Feature extraction is performed on the real-time core texture signal and the reference features of the distribution transformer based on variational mode decomposition and symmetric point pattern algorithms, respectively, to obtain the corresponding real-time feature image and reference feature image. The steps are as follows: The preprocessed real-time core texture signals are subjected to VMD decomposition. The core of this process is to construct an augmented Lagrangian function with a penalty factor to decompose the broadband acoustic, vibration, and thermal acoustic signals into multiple narrow-band intrinsic mode components (IMFs), highlighting the characteristic frequency band differences corresponding to different fault types. The complete mathematical model includes the objective function, constraints, and augmented Lagrangian function.

[0089]

[0090] in, Let K be the k-th intrinsic mode component, corresponding to a narrow frequency band of the signal, where K is the number of modes. The center frequency of the k-th IMF component is adaptively updated by the algorithm. The Dirac impulse function is used to construct the analytic signal form of a signal. The imaginary unit corresponds to the imaginary component of an analytic signal. Refers to the preprocessed real-time signal. As a penalty factor, its physical meaning is to balance the frequency domain bandwidth constraint of the modal components with the reconstruction error of the original signal. These are Lagrange multipliers used to constrain the condition that the sum of all IMF components equals the original signal. This is an inner product operation used to calculate the coupling term between the Lagrange multipliers and the reconstruction error. To augment the Lagrange function.

[0091] It should be noted that the first term of the augmented Lagrange function controls the frequency domain bandwidth of each IMF component through a penalty factor to avoid decomposition failure caused by multi-band aliasing of the distribution transformer signal. The second term minimizes the reconstruction error between the original signal and the sum of each IMF component to ensure that no effective information is lost after decomposition. The third term forces the constraint conditions to be met through Lagrange multipliers to avoid excessive deviation between the reconstructed signal and the original signal due to noise interference.

[0092] VMD decomposition solves for the saddle points of the augmented Lagrangian function using the alternating direction multiplier algorithm. The specific process is as follows: Initialization: Set the number of modes K=5 and the penalty factor. =1500, Lagrange multipliers =0, randomly initialize each IMF component. and center frequency (0).

[0093] It should be noted that the value of K was determined based on: designing multiple sets of comparative experiments with K values ​​(K=3, 4, 5, 6, 7), calculating the cross-correlation coefficients of adjacent IMF components and the signal reconstruction error, and comprehensively judging that K=5 is the optimal value with no aliasing, low error, and no redundancy; the value of the penalty factor was determined based on: traversal The value range is [500, 2500]. The optimal value is selected through three indicators: frequency domain bandwidth stability. When α≥1500, the center frequency fluctuation of each IMF component is ≤±2%, the bandwidth fluctuation is ≤±5%, and the stability meets the requirements. When the frequency is <1500, the center frequency fluctuation is ≥±8%, and the bandwidth is unstable; spectral kurtosis is used to measure the prominence of fault characteristics in the IMF components. When the value is 1500, the kurtosis of the IMF component spectrum corresponding to the fault is ≥5.8, which is significantly higher than other values, indicating clearer fault characteristics; computational efficiency, When the value is 1500, the time taken for a single VMD decomposition is approximately 0.8 seconds, which meets the requirements for real-time monitoring. When the value is greater than 1500, the decomposition time increases linearly, affecting real-time performance. In summary... =1500 is the optimal value with high stability, excellent feature separation, and high computational efficiency.

[0094] Iterative updates are performed using the alternating direction multiplier algorithm: updating modal components. The augmented Lagrangian function is minimized by solving the problem in the frequency domain using Fourier transform. The formula is as follows: ,in This is the Fourier transform of the original signal.

[0095] Update center frequency The frequency domain center is calculated using the power spectral density of each IMF component as a weight, as shown in the formula: ,in, For the updated modal components.

[0096] Update Lagrange multipliers The formula is: ,in The iteration step size is set to 0.001 to ensure convergence stability.

[0097] When satisfied < When the iteration terminates, K=5 IMF components are output.

[0098] The core significance of this process is to break down the originally complex multi-band signal into simple components of a single frequency band, using a penalty factor. Precise control of the frequency domain width of each component highlights fault characteristics such as low-frequency component shift due to mechanical loosening, mid-frequency component distortion due to electromagnetic off-center loading, and high-frequency component enhancement due to local overheating, providing clear frequency band characteristic input for subsequent SDP conversion.

[0099] The five IMF components obtained from the decomposition of the real-time core texture signal and the original signal are subjected to SDP transformation to convert the one-dimensional time-domain signal into a two-dimensional polar coordinate feature image. The texture morphology differences of the image are used to reflect the fault characteristics. The core transformation formula is as follows:

[0100]

[0101]

[0102] in, For the first The radius of each sampling point in polar coordinates reflects the magnitude of the IMF component, and is normalized and mapped to [0, ... ] interval, The maximum radius of the polar coordinate feature image. The value is 100. Let be the amplitude of the i-th sampling point of the IMF component. and These are the maximum and minimum amplitudes of the IMF component, respectively. For the first The counterclockwise rotation angle of each mirror-symmetrical part The value is 6, representing the number of symmetrical parts, meaning each 60° segment is a mirror unit, resulting in a six-fold symmetrical structure in the image. The clockwise rotation angle is modulated by the amplitude difference between adjacent sampling points. The lag factor has a value of 20. This is the angle gain factor, with a value of 30 degrees. For sampling point index, This is the index for the mirror unit.

[0103] It should be noted that the lag factor value is calculated by traversing the [0,45] interval and performing different... The sum of mutual information between the SDP images of various fault signals and normal signals, when When the sum of mutual information is 20, the sum of mutual information is minimized. The value of the angle gain factor is verified by experiments. This angle can avoid the overlap of adjacent spiral arms and make the texture features of different frequency bands clearly distinguishable.

[0104] For the real-time core texture signal, extract a set of spiral arms from each of its K IMF components and the SDP transform result of the original signal, and fuse them to generate a six-fold symmetrical feature image (voiceprint feature image). Vibration ripple feature image Thermal voiceprint feature image As a real-time feature image, the fusion rule is as follows: according to the center frequency of the IMF components from low to high, the spiral arms of each component are sequentially superimposed onto the corresponding mirror unit of the polar coordinate image to ensure that low, medium, and high frequency features do not interfere with each other in the image, and to fully preserve the multi-band texture information of the signal. For example, the IMF1 component (low frequency, corresponding to the fundamental frequency of mechanical vibration) of the vibration signal corresponds to the inner spiral arm of the image, and the IMF5 component (high frequency, corresponding to thermoacoustic flow) corresponds to the outer spiral arm of the image. After fusion, the differences in the thickness, curvature, and edge smoothness of the spiral arms can be used to intuitively reflect the loosening state of the mechanical structure or the abnormal changes of the thermodynamic system.

[0105] Similarly, feature extraction of the baseline texture signal is performed using variational mode decomposition and symmetric point pattern algorithms to obtain the baseline feature image.

[0106] By comparing the features of the real-time feature image and the reference feature image using an improved Scale Invariant Feature Transform (SIFT) algorithm, suspected fault signal feature images are identified. The real-time core texture signal corresponding to the suspected fault signal feature image is then recorded as the suspected fault signal. The steps are as follows: The feature images are subjected to Gaussian blurring and downsampling at different scales to generate a 6-layer Gaussian pyramid, each layer containing 5 Gaussian difference images, with a scale factor of [missing information]. =1.6, ensuring the scale invariance of feature points and adapting to scale changes in image texture under different fault conditions.

[0107] In the difference of Gaussian image, each pixel is compared with its 8 neighboring pixels and corresponding pixels in the upper and lower layers. Extreme points are selected as candidate feature points. Candidate points with low contrast (contrast threshold set to 0.04) and edge response (edge ​​threshold set to 10) are removed, and stable feature points are retained to avoid false features caused by noise interference.

[0108] Centered on the feature point, a 16×16 neighborhood window is constructed and divided into 4×4 sub-regions. The gradient histogram of each sub-region is calculated in 8 directions to generate a 128-dimensional feature description vector, ensuring the rotation invariance and brightness invariance of the feature point, and adapting to slight changes in ambient light during the operation of the distribution transformer.

[0109] To address the issue of "one-to-many" mismatches that easily occur in traditional SIFT algorithms in textured images, a dual matching criterion based on similar triangles is introduced. The specific steps are as follows: Calculate the Euclidean distance between the feature description vector of the real-time feature image and the feature description vector of the reference feature image. Initial matching is performed using the nearest neighbor to second nearest neighbor ratio criterion (ratio H=0.8), retaining those that satisfy the criteria. The matching pairs, where Nearest neighbor distance The value of H=0.8 is the second nearest neighbor distance, which is chosen to balance the matching accuracy and the number of matching points, avoiding missed matches due to an overly strict threshold or an increase in false matches due to an overly lenient threshold.

[0110] Similar triangle verification: Randomly select 3 feature points A, B, C (from the real-time image) and their corresponding matching points A', B', C' (from the reference image) from the initial matching pairs to construct triangles. ABC and The match A'B'C' is validated using the following formula:

[0111] in, and Triangles ABC The lengths of the sides corresponding to A'B'C' are calculated using the pixel coordinates of the feature points. and These are the included angles between the corresponding vertices, calculated using the vector dot product formula. The threshold for consistency in side length ratios is 0.1. and As the angular consistency threshold, The value is 1. The value is 20.

[0112] It should be noted that the threshold value was determined through a matching experiment of 50 sets of normal signals and 50 sets of fault signals, which can eliminate more than 95% of mismatched points while retaining more than 90% of the correct matching points.

[0113] Iterate through all initial matching pairs, repeat the similar triangle verification, retain matching pairs that satisfy all threshold conditions, and count the final number of matching points and voiceprint matching points. Number of ripple matching points thermal voiceprint matching points .

[0114] The matching results are combined for a comprehensive judgment, and the specific judgment rules are as follows: Calculate the average number of matching points between the baseline feature image and the real-time feature image. Set a single-mode matching threshold The threshold is set at 70% of the average because experiments have shown that the number of matching points decreases by about 30% when there is a minor fault. This threshold can effectively distinguish between normal and minor fault states and avoid missing early hidden faults.

[0115] like and and If the status is determined to be normal, return to the baseline model update process in step S1, continuously monitor real-time signals, and ensure the dynamic adaptability of the baseline model.

[0116] If the number of matching points in any one type of feature image is lower than the corresponding threshold, it is marked as a suspected fault signal feature image; if the number of matching points in two or more types of feature images is lower than the threshold, it is marked as a composite suspected fault feature image.

[0117] The real-time core texture signal corresponding to the suspected fault signal feature image is recorded as the suspected fault signal. The number of matching points of the suspected fault signal and the operating parameters (load, oil temperature, etc.) at the time of signal acquisition are recorded. At the same time, the real-time signal is reacquired (the acquisition interval is shortened to 1 minute) to verify the persistence of the fault characteristics and avoid misjudgment caused by occasional interference (such as instantaneous electromagnetic shock).

[0118] S3: Input the suspected fault signal feature image into the improved deep residual shrinkage network model, and output the distribution transformer fault type result; The core purpose of this step is to take the suspected fault signal output from step S2, enhance the complementarity of fault features through multimodal feature fusion, and use an improved deep residual shrinkage network (IDRSN) with embedded adaptive parameter correction linear unit (APReLu) to achieve accurate classification of mechanical loosening, electromagnetic off-center loading, and local overheating of distribution transformers. This solves the problems of insufficient single-modal feature information and weak ability of traditional networks to suppress redundant information, and provides a clear basis for fault type determination.

[0119] The input to this step is the suspected fault signal feature image generated in step S2: acoustic signature image, vibration signature image, and thermal acoustic signature image. All three types of images are 128×128 pixel two-dimensional matrices, containing core information such as frequency band texture and spatial morphology corresponding to the fault. Multimodal fusion integrates the complementary information of the three types of features, improving the network's ability to identify complex faults.

[0120] Multimodal feature fusion is achieved by using a channel-dimensional concatenation method. The core logic is to retain the complete feature information of each modality while constructing cross-modal correlation features. The specific process is as follows: The three types of feature images are normalized to map pixel values ​​to the [0,1] interval, as shown in the formula. ,in For the image at the pixel point The original pixel values, , These are the maximum and minimum pixel values ​​of the image, respectively. Normalization can eliminate the imbalance of feature weights caused by the difference in amplitude between different modal images.

[0121] Channels are stitched together in the order of voiceprint-vibration pattern-thermal voiceprint to generate a fused feature matrix. The three channels correspond to three types of modal features, ensuring that the texture information of each modality is independent and can be learned by the network simultaneously.

[0122] To reduce network computational complexity and suppress redundant information interference, the fusion feature matrix is... Principal Component Analysis (PCA) dimensionality reduction is performed, and the specific steps and parameters are explained below: Will Flattened into a one-dimensional feature vector Preserve feature information across all pixel dimensions and calculate mean and standard deviation The standardized formula is Construct the covariance matrix of the standardized eigenvectors Where N is the feature dimension, the covariance matrix is ​​solved. The eigenvalues ​​and corresponding eigenvectors were used to select principal components based on their cumulative contribution rate being greater than 90%, and the final number of principal components was determined to be k=1024.

[0123] Standardize the feature vector Multiplying the core feature matrix by the projection matrix P formed by the first k eigenvectors yields the dimensionality-reduced core feature matrix. The feature map is then reshaped into a 32×32×1 two-dimensional feature map (adapting to the network input dimension) and used as the input for the improved deep residual shrinkage network.

[0124] The overall network structure consists of an input layer → convolutional layer → four improved residual shrinking units (IRSBU) → BN layer → APReLu layer → global average pooling layer (GAP) → fully connected layer (FC) → output layer. The parameters and improved logic of each layer are designed specifically for the multimodal characteristics of distribution transformers, as detailed below: The input dimension of the input layer is 32×32×1; the convolutional layer kernel is 5×5, stride=1, number of channels=8, padding=SAME; the kernel of the improved residual shrinkage unit 1 is 3×3, stride=2, number of channels=8; the kernel of the improved residual shrinkage unit 2 (IRSBU2) is 3×3, stride=1, number of channels=8; the kernel of the improved residual shrinkage unit 3 (IRSBU3) is 3×3, stride=2, number of channels=16; the kernel of the improved residual shrinkage unit 4 (IRSBU4) is 3×3, stride=1, number of channels=16; the momentum of the BN layer is 0.9, epsilon= To avoid the denominator being zero; the APReLu layer uses adaptive weight learning; the pooling kernel of the GAP layer is 8×8; the number of neurons in the FC layer is 5, corresponding to 5 types of faults (normal, mechanical loosening, electromagnetic bias, local overheating, and compound faults), the L2 regularization coefficient is 0.0001, and the activation function of the output layer is Softmax.

[0125] An adaptive parameter-corrected linear unit (APReLU) is used to replace the traditional ReLU activation function, solving the problem of ReLU directly discarding negative features. The core formula is:

[0126]

[0127]

[0128] in, This is the output feature vector of the APReLu layer. The weights are adaptive and dynamically adjusted through network training. Intermediate variables are used to learn the weights, and are calculated from the feature vector after global average pooling through a fully connected layer. This is the weight matrix. For bias terms, For the input feature vector, This is the output result after performing global average pooling on the input feature vector x.

[0129] Each IRSBU contains a main path (convolutional layer + BN layer + APReLU layer + convolutional layer) and a threshold learning path (absolute value layer + GAP layer + fully connected layer + Sigmoid layer). The core function is to eliminate redundant information (such as invalid textures left by environmental interference) in multimodal features using an adaptive soft thresholding function. The specific formula is as follows:

[0130] in, Let c be the feature matrix of the c-th channel. This is the feature matrix of the c-th channel after soft thresholding. The adaptive threshold for the c-th channel. This is the channel-independent scaling factor.

[0131] The threshold-independent design allows the threshold of each channel to be adapted to its own feature distribution. Compared with the shared threshold, it can more accurately eliminate the cross-redundant information after multimodal fusion. Experimental results show that it can improve the network's accuracy in identifying complex faults by more than 4%.

[0132] The training dataset contains laboratory simulated fault data and field measured data, with a total of 5 types of samples (normal, mechanical loosening, electromagnetic off-center load, local overheating, and combined faults), with 1000 samples in each type. The training set accounts for 70% (3500 samples) and the test set accounts for 30% (1500 samples). The samples cover different loads (30%~100% of rated load) and different ambient temperatures (-10℃~45℃) to ensure the network's generalization ability.

[0133] The loss function used is the multi-class cross-entropy loss function, and the calculation formula is as follows:

[0134] in, For the batch sample size, For the first The true label of each sample For the network prediction of the first Class failure probability, For multi-class cross-entropy loss, For sample index, Index for fault categories.

[0135] Select the Adam optimizer and set the parameters to learning=0.001 and momentum. =0.9、 =0.999, weight decay =0.0001, batch size =32, number of iterations =100.

[0136] The core feature matrix of the test set The trained IDRSN is input, processed through various layers, and the fully connected (FC) layer outputs a 5-dimensional vector, which is then activated by Softmax to obtain the probability distribution of various fault types. (Corresponding to normal, mechanical loosening, electromagnetic off-center load, local overheating).

[0137] The category with the highest probability is selected as the result of the distribution transformer fault type. If the maximum probability is less than 0.8, it is marked as a suspected unidentified fault and manual review is triggered; if the maximum probability is greater than 0.8, the specific fault type is output.

[0138] If the fault is suspected to be unidentified, return to step S2 to re-extract features (adjust VMD / SDP parameters).

[0139] S4: Perform spatiotemporal synchronization analysis on suspected fault signals to obtain spatiotemporal synchronization analysis results. Based on the fault type results of the distribution transformer and the spatiotemporal synchronization analysis results, calculate the fault diagnosis results to realize the monitoring of the distribution transformer.

[0140] Spatiotemporal synchronization analysis is performed on suspected fault signals to obtain the results. The core of spatiotemporal synchronization analysis is to determine whether the fault is a single type or a compound type by quantifying the temporal correlation and spatial consistency of the fault characteristics of acoustic patterns, vibration patterns, and thermal acoustic patterns. The specific steps are as follows: The spatiotemporal synchronization analysis results include temporal synchronization coefficients and spatial synchronization coefficients. Temporal synchronization focuses on whether the occurrence times of the three types of modal fault characteristics are consistent, and is measured using cross-correlation coefficients. The formula for calculating the temporal synchronization coefficient is as follows:

[0141] in, For the first Class modality and the first The time synchronization coefficient for modalities; the closer the absolute value is to 1, the stronger the time synchronization. For the first Modal fault characteristic time series, For time sampling point index, For the first The mean of the time series of characteristics of modal faults. For the first Modal fault characteristic time series, For the first The mean of the time series of characteristics of modal faults. and This is an index for modal fault characteristics.

[0142] Verification was conducted through 50 sets of single fault experiments and 30 sets of compound fault experiments, setting | | 0.85 indicates strong synchronization. It is a weak synchronization. To determine if a mode is out of sync, if two or more modes meet the strong synchronization condition, it is determined to be a composite fault; if only one mode exhibits fault characteristics and is out of sync with other modes, it is determined to be a single fault.

[0143] It should be noted that this is not a subjective setting, but rather derived from regression analysis of fault data. The samples are labeled, with composite fault samples marked as "1" and single fault samples marked as "0," using a synchronicity coefficient (…). Using the absolute value of the value as the independent variable, a binary logistic regression model is constructed. The "optimal threshold" is calculated through ROC curve analysis. The horizontal axis of the ROC curve is the false positive rate (the probability of misclassifying a single fault as a compound fault), and the vertical axis is the true positive rate (the correct identification rate of compound faults). The point where the "Yorden index (true positive rate - false positive rate) is the largest" is taken as the dividing point between strong synchronization and weak synchronization, and the point where the "false positive rate ≤ 5%" is taken as the dividing point between weak synchronization and asynchrony.

[0144] Spatial synchronization focuses on whether the spatial distributions of the three types of fault characteristics overlap. It is quantified by the overlap of characteristic locations, and the calculation formula is as follows:

[0145] in, For the first Class modality and the first The spatial synchronization coefficient of a modality, the closer it is to 1, the higher the spatial overlap. For the first The spatial region corresponding to the modal fault characteristics for and The number of sensor points in the intersection area. for and The number of sensor points in the union region. For the first The spatial region corresponding to the modal fault characteristics.

[0146] It should be noted that the spatial regions are divided based on the sensor deployment location and signal acquisition coverage area. Each region corresponds one-to-one with the acquisition range of one sensor, ensuring clear and non-overlapping region boundaries. There are four regions: Region 1 is near the iron core clamping part (covering the iron core and a 50cm surrounding area); Region 2 is the winding end (covering the winding end and a 50cm surrounding area); Region 3 is the heat sink area (covering the entire heat sink area); and Region 4 is the load-bearing part of the enclosure (covering the bottom load-bearing structure of the enclosure and a 30cm surrounding area). The boundaries are defined by the effective signal acquisition radius, centered on the sensor installation location. The effective acquisition radius of the microphone / vibration sensor is 50cm. Signals exceeding this radius are considered to be outside the region.

[0147] like =1 indicates that the fault characteristics are concentrated in the same spatial region; if =0 indicates that the fault characteristics are distributed in different areas. Combined with the time synchronization results, the compound fault can be further divided into "compound fault in the same area" (such as mechanical loosening and local overheating of the iron core clamping parts at the same time) and "compound fault across areas" (such as electromagnetic off-center load at the winding end and local overheating in the heat sink area).

[0148] Based on the distribution transformer fault type results and spatiotemporal synchronization analysis results, fault diagnosis results are calculated, including distribution transformer fault type results and fault location results. The steps are as follows: The core characteristics of mechanical loosening faults are reflected in the amplitude shift of the low-frequency IMF component of the vibration signal and the change in the morphology of the rotating arm in the SDP image. The localization formula is:

[0149] in, This indicates the location of the mechanical loosening fault, specifically the area where the sensor with the largest amplitude offset is located. For sensor indexing, The average amplitude of the low-frequency IMF component of the ripple pattern acquired by the l-th sensor. The average amplitude of the low-frequency IMF component of the ripple corresponding to the l-th sensor in the baseline model. Let be the absolute value of the deviation between the amplitude of the vibration ripple in the l-th region and the reference. The deviation rate indicates that the larger the deviation rate, the more severe the mechanical loosening in that area.

[0150] Location results The corresponding sensor area, combined with the rotation arm offset direction of the vibration pattern SDP image (e.g., the rotation arm offset clockwise corresponds to the left bolt being loose, and the offset counterclockwise corresponds to the right bolt being loose), further locks the sub-region.

[0151] The core characteristics of electromagnetic off-center load faults are reflected in the distortion of the intermediate frequency IMF component of the acoustic waveform signal and the synchronous jump of the vibration waveform signal. The localization formula is:

[0152] in, The results of electromagnetic off-center load fault location. The spectral distortion of the intermediate frequency (IMF) component of the acoustic signature acquired by the l-th sensor is given. is the local time synchronization coefficient between the acoustic signature and the vibration signature in the l-th region.

[0153] The core characteristics of localized overheating faults are the increased stripe density and enhanced energy of the high-frequency IMF component in the thermoacoustic signal. The localization formula is as follows:

[0154]

[0155]

[0156] in, The results show the location of the localized overheating fault. The density of the thermoacoustic stripes collected by the l-th sensor. The mean stripe density in the baseline model. This represents the density increment; a larger increment indicates more vigorous hot air movement. The total energy of the high-frequency IMF component of the thermal acoustic signature collected by the l-th sensor. The energy mean in the baseline model. This represents the energy increment; a larger increment indicates a higher local temperature. Based on the arrangement direction of the thermal stripes, if the stripes are arranged vertically, it is determined that the normal heat dissipation area of ​​the heat sink is overheating; if the stripe arrangement direction is disordered, it is determined that the internal part of the cabinet is overheating.

[0157] Composite fault location employs a region overlay and weight allocation strategy: For compound faults in the same area, the intersection area of ​​the location results of each individual fault is taken.

[0158] Cross-regional composite faults are classified according to the network output probability of each individual fault. Assign weights, the weight formula is as follows , where n is the number of composite fault types, the region with the largest weight is the primary fault region, and the rest are secondary fault regions.

[0159] The final fault diagnosis results include the fault type and fault location results of the distribution transformer.

[0160] Furthermore, based on the above method embodiments, the present invention also provides an apparatus, including a memory, a processor, and a computer program stored in the memory, which is adapted to be loaded and executed by the processor to implement the above-described method for monitoring a distribution transformer.

[0161] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, the phrase "comprising an element defined as..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0162] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for monitoring distribution transformers, characterized in that: Includes the following steps: S1: Collect the core texture signal of the distribution transformer under steady-state operation, and preprocess the core texture signal to obtain the reference texture signal; S2: Acquire real-time core texture signals, extract features from the real-time core texture signals using variational mode decomposition and symmetric point pattern algorithms, extract features from the reference texture signals using variational mode decomposition and symmetric point pattern algorithms, and obtain corresponding real-time feature images and reference feature images; compare the features of the real-time feature images and reference feature images using an improved scale-invariant feature transformation algorithm, identify suspected fault signal feature images, and record the real-time core texture signals corresponding to the suspected fault signal feature images as suspected fault signals; S3: Input the suspected fault signal feature image into the improved deep residual shrinkage network model, and output the distribution transformer fault type result; S4: Perform spatiotemporal synchronization analysis on suspected fault signals to obtain spatiotemporal synchronization analysis results. Based on the fault type results of the distribution transformer and the spatiotemporal synchronization analysis results, calculate the fault diagnosis results to realize the monitoring of the distribution transformer.

2. The method for monitoring a distribution transformer according to claim 1, characterized in that: The acquisition of the core texture signal of the distribution transformer under steady-state operating conditions includes the following specific steps: The core texture signals of the distribution transformer under steady-state operation are collected. These core texture signals include acoustic fingerprint signals, vibration fingerprint signals, and thermal acoustic fingerprint signals. The acoustic fingerprint signals are collected using electret capacitor-type gun-shaped directional microphones, which are evenly distributed around the transformer tank according to a symmetrical arrangement principle. Specifically, they are located near the core clamping parts, the winding ends, the heat sink area, and the load-bearing parts of the tank. The vibration fingerprint signals are collected using piezoelectric vibration acceleration sensors, which are attached to the outer wall of the tank corresponding to the microphone placement positions. The thermal acoustic fingerprint signals are collected synchronously using the aforementioned microphones and temperature sensors. Before starting data acquisition, it is necessary to determine whether the transformer is in steady-state operation. The steady-state determination conditions are set as follows: within 24 consecutive hours, the transformer load fluctuation amplitude is ≤ ±5%, the three-phase current imbalance is ≤ 2%, the oil temperature change rate is ≤ 0.5℃ / h, and there is no obvious external interference. When the conditions are met, it is determined to be in steady-state operation, data acquisition is started, and the core texture signal is finally obtained.

3. The method for monitoring a distribution transformer according to claim 2, characterized in that: The specific steps for preprocessing the core texture signal to obtain the reference texture signal are as follows: For the collected voiceprint signals (t) Preprocessing is performed, and based on the preprocessed voiceprint signal, three types of features are extracted to construct a voiceprint benchmark: The formula for calculating spectral complexity is: ； in, This represents the proportion of the power spectral density of the acoustic signature signal at frequency f. The upper limit of the effective frequency, For spectral complexity; The formula for calculating the main frequency is: ； in, The power percentage of the 50Hz harmonic component. For the frequency multiplication factor, Main frequency; The formula for calculating the high-low frequency ratio is: ； in, The harmonic number corresponding to the high-low frequency boundary. High-low frequency ratio; The above three types of features are arranged into a voiceprint feature matrix according to the time series. Statistical analysis was performed on the voiceprint feature matrix to calculate the mean and standard deviation of each feature, and a voiceprint baseline interval was constructed. ,in, This represents the mean of the voiceprint features. The standard deviation of voiceprint features; For the acquired vibration ripple signal (t) Preprocessing: A 50Hz notch filter is used to remove power frequency interference, wavelet threshold denoising is used to eliminate environmental vibration noise, and three types of features are extracted to construct a ripple reference: The formula for calculating the vibration amplitude distribution parameter is: ； in, This represents the peak-to-peak value of the ripple signal. The effective value of the ripple signal. These are the vibration amplitude distribution parameters; The formula for calculating the periodic alternation characteristic parameter is: ； in, Let be the autocorrelation function of the ripple signal. To delay time, These are periodic alternating characteristic parameters; The formula for calculating the energy transfer path parameters is: ； in, For the first The first collection site and the first Cross-correlation coefficients of vibration signals at each acquisition location The number of combinations of 4 sampling sites. For energy transfer path parameters, and Index for the collection site; The three types of features are combined to form a ripple feature matrix. Statistical analysis was performed on the ripple characteristic matrix to construct the ripple reference interval. ,in, The mean value of the vibration ripple characteristics. The standard deviation of the vibration ripple characteristics; For the collected voiceprint signals Preprocessing is performed, and the preprocessing process is consistent with that of the voiceprint signal. Two types of features are extracted to construct a thermal voiceprint benchmark: The formula for calculating the stripe density parameter is: ； in, The total number of fringes in the time spectrum. The effective frequency range of thermal acoustic waveforms. This represents the total number of signal sampling points. For stripe density parameters, This is the correction factor for the rate of temperature change; The formula for calculating the orientation parameter is: ； in, Let be the angle between the k-th stripe and the horizontal direction. These are the arrangement direction parameters; Combining the two types of features into a thermal soundprint feature matrix Statistical analysis was performed on the thermal acoustic signature feature matrix to construct a thermal acoustic signature baseline interval. ,in, The mean value of the vibration ripple characteristics. The standard deviation of the vibration ripple characteristics; By integrating the feature matrices of voiceprints, vibration patterns, and thermal voiceprints with the benchmark interval, a multimodal texture benchmark model is constructed. ,in These are the reference ranges for soundprint, vibration, and thermal soundprint, respectively. Select the signal whose feature value is within ±5% of the center of the reference interval from the core texture signal and use it as the reference texture signal.

4. The method for monitoring a distribution transformer according to claim 3, characterized in that: The process of acquiring real-time core texture signals, extracting features from these signals using variational mode decomposition and symmetric point pattern algorithms, and extracting features from a reference texture signal using the same algorithms, to obtain real-time feature images and reference feature images, includes the following steps: Real-time core texture signals are acquired, including real-time acoustic fingerprint signals, real-time vibration fingerprint signals, and real-time thermal acoustic fingerprint signals, wherein the real-time acoustic fingerprint signal is denoted as... Real-time vibration ripple signal is denoted as Real-time thermal acoustic signal is recorded as The acquisition parameters for the three types of signals remain consistent with step S1. The real-time core texture signals undergo unified preprocessing, and the preprocessed real-time core texture signals are then subjected to variational mode decomposition. The complete mathematical model includes the objective function, constraints, and augmented Lagrangian function. ； ； in, Let K be the k-th intrinsic modal component, and K be the number of modes. Let the center frequency of the k-th IMF component be . For the Dirac impulse function, The imaginary unit, Refers to the preprocessed real-time signal. As a penalty factor, For Lagrange multipliers, For inner product operations, To augment the Lagrange function; The five IMF components obtained from the decomposition of the real-time core texture signal and the original signal are subjected to symmetric point mode transformation to convert the one-dimensional time-domain signal into a two-dimensional polar coordinate feature image. The core transformation formula is as follows: ； ； ； in, For the first The radius of each sampling point in polar coordinates Let be the amplitude of the i-th sampling point of the IMF component. and These are the maximum and minimum amplitudes of the IMF component, respectively. For the first The counterclockwise rotation angle of each mirror-symmetrical part The number of symmetrical parts is 6. The angle is clockwise rotation. The lag factor has a value of 20. This is the angle gain factor, with a value of 30 degrees. The maximum radius of the polar coordinate feature image. For sampling point index, For mirror cell index; For the real-time core texture signal, extract a set of spiral arms from each of its K IMF components and the SDP transformation result of the original signal, and fuse them to obtain a real-time feature image. The fusion rule is: according to the center frequency of the IMF components from low to high, the spiral arms of each component are superimposed on the corresponding mirror unit of the polar coordinate image. The baseline texture signal is used to extract features by variational mode decomposition and symmetric point pattern algorithm to obtain the baseline feature image.

5. The method for monitoring a distribution transformer according to claim 4, characterized in that: The step of comparing real-time feature images and baseline feature images using an improved scale-invariant feature transform algorithm to identify suspected fault signal feature images, and recording the real-time core texture signal corresponding to the suspected fault signal feature image as the suspected fault signal, includes the following steps: The feature images are subjected to Gaussian blurring and downsampling at different scales to generate a 6-layer Gaussian pyramid, each layer containing 5 Gaussian difference images, with a scale factor of [missing information]. =1.

6. In the difference of Gaussian image, each pixel is compared with its 8 neighboring pixels and corresponding pixels in the upper and lower layers. Extreme points are selected as candidate feature points. Candidate points with low contrast and edge response are eliminated, and stable feature points are retained. A 16×16 neighborhood window is constructed with the feature points as the center, which is divided into 4×4 sub-regions. Gradient histograms in 8 directions are calculated for each sub-region to generate a 128-dimensional feature description vector. Calculate the Euclidean distance between the feature descriptor vector of the real-time feature image and the feature descriptor vector of the reference feature image. Perform initial matching using the nearest neighbor to second nearest neighbor ratio criterion, retaining those that satisfy the criteria. The matching pairs, where Nearest neighbor distance The next nearest neighbor distance; Three feature points A, B, and C, and their corresponding matching points A', B', and C', are randomly selected from the initial matching pairs to construct a triangle. ABC and The match A'B'C' is validated using the following formula: ； in, and Triangles ABC The lengths of the sides corresponding to A'B'C' and These are the included angles of the corresponding vertices. The threshold for consistency in side length ratios is 0.

1. and As the angular consistency threshold, The value is 1. The value is 20; Iterate through all initial matching pairs, repeat the similar triangle verification, and count the number of voiceprint matching points. Number of ripple matching points thermal voiceprint matching points ; Calculate the average number of matching points between the baseline feature image and the real-time feature image. Set a single-mode matching threshold ; When the number of matching points in a feature image is lower than the corresponding threshold, it is marked as a suspected fault signal feature image, and the real-time core texture signal corresponding to the suspected fault signal feature image is recorded as a suspected fault signal.

6. The method for monitoring a distribution transformer according to claim 5, characterized in that: The process of inputting the suspected fault signal feature image into the improved deep residual shrinkage network model and outputting the distribution transformer fault type result includes the following steps: The feature images of suspected fault signals are normalized, and channels are stitched together in the order of acoustic pattern-vibration pattern-thermal acoustic pattern to obtain a fused feature matrix. For the fused feature matrix Principal component analysis was performed to reduce the dimensionality, resulting in the core feature matrix. The feature map is then reshaped into a 32×32×1 two-dimensional feature map, and an improved deep residual shrinkage network model is constructed. The loss function of the improved deep residual shrinkage network adopts the multi-class cross-entropy loss function, and the calculation formula is as follows: ； in, For the batch sample size, For the first The true label of each sample For the network prediction of the first Class failure probability, For multi-class cross-entropy loss, For sample index, Index for fault categories; core feature matrix Inputting the trained improved deep residual shrinkage network model, the output yields the probability distribution of various fault types. These correspond to normal operation, mechanical loosening, electromagnetic imbalance, and localized overheating, respectively. The category with the highest probability distribution is selected as the result of the distribution transformer fault type.

7. The method for monitoring a distribution transformer according to claim 6, characterized in that: The construction of the improved deep residual shrinkage network model includes the following specific steps: The overall structure of the improved deep residual shrinking network consists of: an input layer, a convolutional layer, four improved residual shrinking units, a BN layer, an APReLu layer, a global average pooling layer, a fully connected layer, and an output layer. The adaptive parameter correction linear unit APReLu is used, and the core formula is: ； ； ； in, This is the output feature vector of the APReLu layer. For adaptive weights, Learn intermediate variables for weights. This is the weight matrix. For bias terms, For the input feature vector, This is the output result after performing global average pooling on the input feature vector x; Each Improved Residual Shrinkage Unit (IRSBU) contains the following main path: convolutional layer, Batch Normalization (BN) layer, APReLU layer, convolutional layer again; and the threshold learning path: absolute value layer, GAP layer, fully connected layer, sigmoid layer. The specific formula is as follows: ； in, Let c be the feature matrix of the c-th channel. This is the feature matrix of the c-th channel after soft thresholding. The adaptive threshold for the c-th channel. This is the channel-independent scaling factor.

8. The method for monitoring a distribution transformer according to claim 7, characterized in that: The process of performing spatiotemporal synchronization analysis on suspected fault signals to obtain spatiotemporal synchronization analysis results includes the following specific steps: The spatiotemporal synchronization analysis results include a time synchronization coefficient and a spatial synchronization coefficient. The formula for calculating the time synchronization coefficient is as follows: ； in, For the first Class modality and the first The time synchronization coefficient for modalities; the closer the absolute value is to 1, the stronger the time synchronization. For the first Modal fault characteristic time series, For time sampling point index, For the first The mean of the time series of characteristics of modal faults. For the first Modal fault characteristic time series, For the first The mean of the time series of characteristics of modal faults. and For modal fault feature index; Settings | | 0.85 indicates strong synchronization. It is a weak synchronization. For asynchronous conditions, if two or more modes meet the strong synchronization condition, it is determined to be a composite fault; if only one mode exhibits fault characteristics and is asynchronous with other modes, it is determined to be a single fault. The formula for calculating the spatial synchronization coefficient is: ； in, For the first Class modality and the first The spatial synchronization coefficient of a modality, the closer it is to 1, the higher the spatial overlap. For the first The spatial region corresponding to the modal fault characteristics for and The number of sensor points in the intersection area. for and The number of sensor points in the union region. For the first Spatial region corresponding to modal fault characteristics; when =1 indicates that the fault characteristics are concentrated in the same spatial region; when =0 indicates that the fault characteristics are distributed in different regions. Combined with the time synchronization coefficient, composite faults can be further divided into intra-regional composite faults and cross-regional composite faults.

9. A method for monitoring a distribution transformer according to claim 8, characterized in that: The fault diagnosis results, calculated based on the distribution transformer fault type results and spatiotemporal synchronization analysis results, are as follows: The formula for locating mechanical loosening faults is: ； in, The results of locating the mechanical loosening fault. For sensor indexing, The average amplitude of the low-frequency IMF component of the ripple pattern acquired by the l-th sensor. The average amplitude of the low-frequency IMF component of the ripple corresponding to the l-th sensor in the baseline model. Let be the absolute value of the deviation between the amplitude of the vibration ripple in the l-th region and the reference. The deviation rate; The formula for locating electromagnetic off-center load faults is: ； in, The results of electromagnetic off-center load fault location. The spectral distortion of the intermediate frequency (IMF) component of the acoustic signature acquired by the l-th sensor is given. is the local time synchronization coefficient between the acoustic signature and the vibration signature in the l-th region; The formula for locating localized overheating faults is: ； ； ； in, The results show the location of the localized overheating fault. The density of the thermoacoustic stripes collected by the l-th sensor. The mean stripe density in the baseline model. For density increment, The total energy of the high-frequency IMF component of the thermal acoustic signature collected by the l-th sensor. The energy mean in the baseline model. For energy increment; Composite fault localization employs a regional overlay and weight allocation strategy: for composite faults within the same region, the intersection of the localization results of each individual fault is used; for composite faults across regions, the network output probability of each individual fault is used. Assign weights, the weight formula is as follows Where n is the number of composite fault types, the region with the largest weight is the primary fault region, and the rest are secondary fault regions. Finally, the fault diagnosis results are obtained, which include the fault type results and fault location results of the distribution transformer.

10. A monitoring transceiver for distribution transformers, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method according to any one of claims 1-9.

Images