Transformer acoustic small sample fault diagnosis method and system based on mahalanobis distance
By using a convolutional neural network method based on Mahalanobis distance to perform fault diagnosis using transformer acoustic signature signals, the problems of sample scarcity and overfitting are solved, and high-precision and robust transformer fault detection is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- POWER RES INST OF STATE GRID SHAANXI ELECTRIC POWER CO LTD
- Filing Date
- 2026-01-23
- Publication Date
- 2026-05-26
AI Technical Summary
Existing technologies for transformer fault diagnosis suffer from problems such as scarce samples and difficulty in balancing accuracy and generalization ability in classification methods. Traditional methods are complex to install and susceptible to interference, and deep learning models overfit under small sample conditions, leading to unstable decision boundaries.
A convolutional neural network method based on Mahalanobis distance is adopted to collect acoustic fingerprint signals by simulating transformer fault conditions, generate Mel time-frequency maps, extract feature vectors using convolutional neural networks, calculate the average feature vector and covariance matrix, and perform fault diagnosis by combining Mahalanobis distance.
High-precision fault diagnosis was achieved under small sample conditions, which improved the robustness and deployability of the system, reduced the requirements for professional skills, and made it adaptable to new equipment and complex noise environments.
Smart Images
Figure CN122087638A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of power equipment condition monitoring and fault diagnosis, and particularly relates to a method and system for small-sample acoustic fault diagnosis of transformers based on Mahalanobis distance. Background Technology
[0002] As the core equipment for power transmission and distribution in the power grid system, the operational reliability of power transformers directly affects the safety and stability of the entire power system. Among various transformer faults, abnormalities in the internal mechanical structures such as windings and cores (e.g., winding deformation, core loosening, clamp displacement) account for a high proportion, are highly concealed, and are irreversible. These mechanical defects are often caused by long-term electromagnetic vibration, short-circuit impact, or overload operation. If they are not detected in time, they will significantly weaken the transformer's short-circuit withstand capability and may even lead to sudden damage, causing significant economic losses and social impact.
[0003] To effectively monitor the mechanical condition of transformers, traditional methods often rely on vibration sensors installed on the surface of the tank to collect vibration signals, and then use techniques such as spectrum analysis and time-frequency feature extraction for fault identification. However, this approach has significant limitations: on the one hand, vibration sensors need to be in rigid contact with the equipment, making installation complex and susceptible to electromagnetic interference and mechanical coupling noise; on the other hand, single-point vibration signals cannot fully reflect the overall acoustic and vibration characteristics of the transformer, and are insufficiently sensitive to weak or distributed mechanical anomalies.
[0004] In recent years, non-contact monitoring technology based on acoustic signals has gradually attracted attention. Acoustic signals are formed by the radiation of mechanical vibrations inside the transformer through the oil tank into the air. They have rich spectral components and can comprehensively reflect the dynamic response of multiple components. Meanwhile, using microphone arrays for acoustic signature acquisition eliminates the need for physical connections, avoiding electromagnetic coupling risks, and offers advantages such as flexible deployment, high security, and suitability for online monitoring. Therefore, using acoustic anomalies to identify internal transformer faults has become an important research direction for intelligent operation and maintenance.
[0005] However, in practical applications, acoustic diagnostics still faces two major technical bottlenecks:
[0006] First, fault samples are extremely scarce. Due to the long design life and low failure rate of transformers, and the fact that most serious faults are sporadic and destructive, it is difficult to obtain a large number of well-labeled anomalous acoustic samples in real-world operating environments. Existing public databases typically contain fewer than a few dozen valid samples for each type of mechanical fault, representing a typical small-sample learning scenario.
[0007] Secondly, existing classification methods struggle to balance accuracy and generalization ability. Traditional machine learning methods rely on manually designed acoustic features, which are poorly robust in complex industrial noise environments and fail to capture deep discriminative information in high-dimensional time-frequency patterns. While end-to-end deep learning models can automatically extract features, they are prone to overfitting under small sample conditions. The models tend to "memorize" limited samples rather than learn commonalities among categories, leading to unstable decision boundaries and a sharp decline in generalization performance.
[0008] Therefore, there is an urgent need for a new method with high precision and robustness in diagnosis to support condition-based maintenance of transformers. Summary of the Invention
[0009] To overcome the above problems, this invention proposes a method and system for small-sample acoustic fault diagnosis of transformers based on Mahalanobis distance.
[0010] Specifically, the object of the present invention is to provide the following aspects:
[0011] Firstly, a method for diagnosing acoustic faults in a transformer is provided, the method comprising:
[0012] Step 1: Simulate various fault states of the transformer and collect the acoustic fingerprint signals under the corresponding states as a dataset. Divide the dataset into a training set and a validation set.
[0013] Step 2: Preprocess the voiceprint signals of the training set and the validation set respectively to generate the corresponding Mel time-frequency diagrams;
[0014] Step 3: Use a convolutional neural network to extract feature vectors from the Mel time-frequency plots of the training set and the validation set, respectively;
[0015] Step 4: Based on the feature vectors extracted from the training set, calculate the average feature vector and covariance matrix for each fault type to characterize the feature distribution of the fault and realize the training of the convolutional neural network.
[0016] Step 5: For the trained convolutional neural network, based on each sample in the validation set, obtain the Mahalanobis distance between its feature vector and the distribution of each fault type, and determine the fault type with the smallest Mahalanobis distance as the fault diagnosis result to evaluate the accuracy of the diagnosis and realize the validation of the convolutional neural network.
[0017] Step 6: Deploy the validated convolutional neural network, using the preprocessed Mel-time frequency map of the acoustic signature signal of the transformer to be diagnosed as input, and output the fault type according to the principle of minimum Mahalanobis distance to realize the acoustic fault diagnosis of the transformer.
[0018] Optionally, in step 1, the fault conditions include: partial discharge, loose core, DC bias, heavy overload, and winding deformation.
[0019] Optionally, the simulation method for the core loosening fault includes: adjusting the distance between the transformer core clamps to a preset range to simulate different degrees of loosening.
[0020] Optionally, the preset range includes multiple nodes with different spacings from 1mm to 6mm.
[0021] Optionally, the partial discharge fault is simulated by setting up an artificial surface discharge defect.
[0022] Optionally, the heavy overload fault simulation includes: gradually increasing the equipment voltage to 4-6 times the rated voltage using a harmonic source, and collecting multiple sets of overload acoustic signals.
[0023] Optionally, in step 2, the preprocessing includes framing, windowing, Fourier transform, and frequency scaling.
[0024] Optionally, in step 3, the neural network is a deep network with a residual connection structure.
[0025] Secondly, a transformer acoustic fault diagnosis system is provided, the system comprising:
[0026] The signal acquisition module is used to simulate various fault states of the transformer and acquire the acoustic fingerprint signals under the corresponding states as a dataset, which is then divided into a training set and a validation set.
[0027] The preprocessing module is used to preprocess the voiceprint signals of the training set and the validation set respectively to generate the corresponding Mel time-frequency diagrams;
[0028] A feature extraction module is used to extract feature vectors from the Mel time-frequency plots of the training set and the validation set, respectively, using a convolutional neural network.
[0029] The training module is used to calculate the average feature vector and covariance matrix of each fault type based on the feature vector extracted from the training set, so as to characterize the feature distribution of the fault and realize the training of the convolutional neural network.
[0030] The verification module is used to obtain the Mahalanobis distance between the feature vector of the trained convolutional neural network and the distribution of each fault type based on each sample in the verification set, and to determine the fault type with the smallest Mahalanobis distance as the final fault diagnosis result in order to evaluate the accuracy of the diagnosis and realize the verification of the convolutional neural network.
[0031] The diagnostic module deploys a validated convolutional neural network, which takes the preprocessed Mel-time frequency map of the acoustic signature signal of the transformer to be diagnosed as input, and outputs the fault type according to the principle of minimum Mahalanobis distance, thereby realizing the acoustic fault diagnosis of the transformer.
[0032] Thirdly, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the method described in the first aspect.
[0033] The beneficial effects of this invention include:
[0034] (1) This invention combines the deep features automatically extracted by the convolutional neural network (CNN) with Mahalanobis distance, effectively overcoming the problem of overfitting in traditional methods when samples are scarce. Since Mahalanobis distance models the overall distribution of fault features based on the mean vector and covariance matrix within the class, it can stably characterize the statistical characteristics even if there are only 5-20 samples for each type of fault. Thus, in small sample scenarios, it can achieve nearly 100% accuracy for the four states of normal, discharge, deformation, and biased magnetization. It is suitable for actual transformer operation and maintenance scenarios because there are very few fault samples in reality. This method can be quickly deployed and adapted to new equipment, which is significantly better than the end-to-end deep learning model based on the Softmax classifier.
[0035] (2) The entire process from the original voiceprint to the final diagnosis result is automated. No manual design of features or intervention in decision-making is required, which reduces the requirements for the professional skills of operation and maintenance personnel and improves the system's deployability and repeatability. It also enhances the robustness to noise interference and intra-class variation and reduces the dependence on large-scale labeled training data. Attached Figure Description
[0036] Various other advantages and benefits of the present invention will become apparent to those skilled in the art upon reading the detailed description of the preferred embodiments below. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. It is obvious that the drawings described below are merely some embodiments of the invention, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.
[0037] In the attached diagram:
[0038] Figure 1 A flowchart of a transformer acoustic fault diagnosis method based on Mahalanobis distance according to a preferred embodiment of the present invention is shown.
[0039] Figure 2 The steps for drawing the time-frequency spectrum are shown in the diagram.
[0040] Figure 3 This diagram illustrates the steps involved in the Fourier transform of the voiceprint signal in Example 1.
[0041] Figure 4 The Mel-frequency diagrams for three transformer fault states in Example 1 are shown.
[0042] Figure 5 The classification results based on four states—normal operation, discharge, winding deformation, and DC bias—are shown in Example 1.
[0043] Figure 6 The results of the Mahalanobis distance metric classifier in Example 1 are shown.
[0044] Figure 7 The results of classification using SVM are shown in Comparative Example 1.
[0045] Figure 8 Comparative Example 2 shows the classification results using Softmax.
[0046] Figure 9 A simulation flowchart of a DC bias fault according to a preferred embodiment of the present invention is shown. Detailed Implementation
[0047] The following will refer to the appendix. Figures 1 to 9 Specific embodiments of the invention will be described in more detail below. While specific embodiments of the invention are shown in the accompanying drawings, it should be understood that the invention can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the invention and to fully convey the scope of the invention to those skilled in the art.
[0048] It should be noted that certain terms are used in the specification and claims to refer to specific components. Those skilled in the art will understand that different terms may be used to refer to the same component. This specification and claims do not distinguish components based on differences in terminology, but rather on differences in function. The terms "comprising" or "including" used throughout the specification and claims are open-ended and should be interpreted as "comprising but not limited to." The following descriptions are preferred embodiments for carrying out the invention; however, these descriptions are for the purpose of understanding the general principles of the specification and are not intended to limit the scope of the invention. The scope of protection of this invention is determined by the appended claims.
[0049] In the description of this invention, it should be noted that the terms "upper," "lower," "inner," "outer," "front," and "rear," etc., indicate the orientation or positional relationship based on the orientation or positional relationship in the working state of this invention, and are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this invention. Furthermore, the terms "first," "second," "third," and "fourth" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.
[0050] To facilitate understanding of the embodiments of the present invention, further explanations and descriptions will be provided below with reference to the accompanying drawings and specific embodiments. The accompanying drawings do not constitute a limitation on the embodiments of the present invention.
[0051] On the one hand, according to the present invention, a transformer acoustic fault diagnosis method based on Mahalanobis distance is provided, such as... Figure 1 As shown, the method includes:
[0052] Step 1: Simulate various fault states of the transformer and collect the acoustic fingerprint signals under the corresponding states as a dataset. Divide the dataset into a training set and a validation set.
[0053] Step 2: Preprocess the voiceprint signals of the training set and the validation set respectively to generate the corresponding Mel time-frequency diagrams;
[0054] Step 3: Use a convolutional neural network to extract feature vectors from the Mel time-frequency plots of the training set and the validation set, respectively;
[0055] Step 4: Based on the feature vectors extracted from the training set, calculate the average feature vector and covariance matrix for each fault type to characterize the feature distribution of the fault and realize the training of the convolutional neural network.
[0056] Step 5: For the trained convolutional neural network, based on each sample in the validation set, obtain the Mahalanobis distance between its feature vector and the distribution of each fault type, and determine the fault type with the smallest Mahalanobis distance as the fault diagnosis result to evaluate the accuracy of the diagnosis and realize the validation of the convolutional neural network.
[0057] Step 6: Deploy the validated convolutional neural network, using the preprocessed Mel-time frequency map of the acoustic signature signal of the transformer to be diagnosed as input, and output the fault type according to the principle of minimum Mahalanobis distance to realize the acoustic fault diagnosis of the transformer.
[0058] The above methods will be described in detail below.
[0059] Step 1: Simulate various fault states on the transformer and collect the acoustic fingerprint signals under the corresponding states as a dataset. Divide the dataset into a training set and a validation set.
[0060] In step 1, there are C types of fault states, including: partial discharge, core loosening, DC bias, heavy overload, and winding deformation. K samples are collected for each type of fault (K is usually 5-20) to form a small sample dataset. For each fault state in the dataset, the dataset is divided into a training set and a validation set in a ratio of 8:2 or 7:3.
[0061] In one embodiment, the partial discharge fault is simulated by setting up an artificial surface discharge defect, specifically including: setting up an artificial surface discharge defect inside the transformer and applying a power frequency test voltage, and using a stepped voltage increase to make the discharge process go through three stages in sequence: initial discharge, pre-breakdown and breakdown, and collecting acoustic fingerprint signals under different discharge levels, the acoustic fingerprint signals being used to construct a dataset of partial discharge faults.
[0062] In each discharge stage, 10-20 sets of voiceprint samples are collected, with each set of voiceprint samples having a sampling frequency of 20-30kHz, and each collection lasting 5-15 seconds.
[0063] Furthermore, the artificial surface discharge defect is preferably constructed using a tungsten needle as a high-voltage electrode, an oil-impregnated insulating paperboard, and a grounded copper foil. For example, an insulating paperboard with a thickness of 1-1.2 mm is selected, with copper foil attached to its back side and connected to the tank wall and grounded, while the front side is exposed below the needle tip. The surface distance between the needle tip and the surface copper foil is set to 8-15 mm.
[0064] During the test, a power frequency voltage is applied from the high-voltage bushing of the transformer. For example, the voltage is first pre-pressurized at 75-85% partial discharge initiation voltage (PDIV) for 4-5 minutes, and then stepped up at a rate of 0.4-0.6 kV / step, with each step maintained for 2-3 minutes, until breakdown occurs. Acoustic fingerprint signals are collected simultaneously during this process.
[0065] In one embodiment, the simulation of the winding deformation fault includes: applying a momentary high-current short-circuit pulse to the transformer winding outlet end, causing the winding to undergo axial or radial displacement under the action of strong electromagnetic force, thus forming a mechanical deformation fault.
[0066] Furthermore, this simulation was conducted using a short-circuit impact test platform. The test employed a phased, stepped loading strategy, alternating between low-to-medium short-circuit and high-to-medium-to-medium short-circuit connection modes. Frequency response sweeping and reactance measurement were the core monitoring methods throughout the process to ensure the stability of the winding's mechanical state. Specific steps included:
[0067] First, system debugging is carried out, namely: confirming that the current waveform meets the standard requirements by conducting a pre-test impact of 65-75% of the specified short-circuit current;
[0068] Then, the formal testing phase begins, which involves applying short-circuit current surges in ascending order from low to high. After each surge, the reactance value is recorded, and a frequency response sweep test is performed to diagnose whether the winding is deformed. For example, 50%, 70%, 90% to 100% of the rated short-circuit current are applied sequentially, with each surge consisting of one attempt.
[0069] Afterwards, switch to no-load connection and let it stand still. Use the temperature sensor to judge the cooling status of the winding. Then perform no-load and load current verification (first apply 15-20% voltage level and stabilize for 1-2 minutes, then apply 100-200% load current and stabilize for 1-2 minutes) as a functional check of the internal electrical circuit.
[0070] Finally, it enters the next stage of higher current surge.
[0071] The entire process involves a cycle of "impact-monitoring-cooling verification," gradually increasing the destructive force while closely monitoring the mechanical and electrical condition of the transformer.
[0072] During the data acquisition process, the sampling frequency was set to 20-30kHz, and the duration of each acquisition was 10-15 seconds. After the impact test, mechanical deformation faults occurred in the transformer windings; acoustic fingerprint samples were collected as the dataset for these winding deformation faults.
[0073] In one embodiment, the method for simulating the core loosening fault includes: adjusting the distance between the transformer core clamps to a preset range, the preset range including multiple nodes with different spacings from 1mm to 6mm, to simulate different degrees of loosening.
[0074] Furthermore, the simulation operates the transformer under shutdown conditions. 10-20 acoustic signatures are acquired at each preset range node, such as 1 mm, 2 mm, 3 mm, 4 mm, 5 mm, and 6 mm, to construct a dataset of core loosening faults.
[0075] The sampling frequency was set to 20-30kHz, and the duration of a single sampling session was 10-15 seconds.
[0076] In one embodiment, the method for simulating DC bias faults includes: injecting a controllable DC current into the neutral point of the transformer to cause half-cycle saturation of the core magnetic circuit, thereby enhancing the magnetostrictive effect and generating low-frequency abnormal noises dominated by twice the power frequency (e.g., 100Hz in a 50Hz power system and 120Hz in a 60Hz system), forming an acoustic signature signal that can be identified by acoustic acquisition.
[0077] Furthermore, this simulation is achieved by connecting a DC power supply in series between the transformer neutral point and ground, forming a simulated geomagnetic induced current loop. Preferably, in transformers with voltage levels such as 35 kV, 110 kV, 220 kV, and 500 kV, the DC power supply is connected to the neutral point of the high-voltage winding, and a closed loop is formed using the substation grounding system or lines to apply a steady-state DC magnetic flux to the core.
[0078] After the above wiring is completed, as follows Figure 9As shown, the following operations were performed to simulate a DC bias fault: First, initial state data of the equipment were collected, including oil chromatography, oil temperature, voltage, and current; then, starting from 1 A, the injected DC current amplitude was gradually increased in steps of 2-3 A, maintaining static injection for 4-5 minutes at each current level until the iron core magnetization state entered a steady state; then, the excitation current, branch current, node voltage, and acoustic signature were monitored synchronously; finally, the test was terminated when the injected current reached a low-frequency abnormal noise dominated by twice the power frequency.
[0079] In the transformer used in this invention, the low-frequency abnormal noise, which is mainly at twice the power frequency, usually occurs when the injected current reaches 10A to 30A.
[0080] Furthermore, for each injected DC current, the sampling rate is 20-30kHz, and the single acquisition time is 2-3 minutes, which is used to construct a dataset for simulating DC bias faults.
[0081] In one embodiment, the heavy overload fault simulation includes: gradually increasing the equipment voltage to 4-6 times the rated voltage using a harmonic source, and collecting multiple sets of overload acoustic signature signals.
[0082] Furthermore, the voltage boosting method adopts a step-by-step increasing approach, maintaining steady-state operation for 5-6 minutes at each voltage level. At each overload level (such as 4×, 5×, 6× rated voltage), the sampling rate is 20-30kHz, and the single acquisition duration is 2-3 minutes, which is used to construct a dataset for simulating heavy overload faults.
[0083] Step 2: Preprocess the voiceprint signals of the training set and validation set respectively to generate the corresponding Mel time-frequency diagrams.
[0084] In step 2, as Figure 2 As shown, the preprocessing includes framing, windowing, Fourier transform, and frequency scaling.
[0085] In step 2, when processing the voiceprint signal, it is first necessary to perform frame segmentation. This is because acoustic signals have temporal variability; by dividing them into shorter time periods (frames), the changing characteristics of the voiceprint signal over time can be better captured. This can be understood as dividing the original one-dimensional voiceprint signal into several short time segments in chronological order, i.e., "frame segmentation".
[0086] Typically, the length of each frame is set to 40-50 milliseconds. This range is suitable for the dynamic characteristics of acoustic signals from most power equipment: long enough to cover a complete acoustic cycle, yet not so long as to lose transient information. Combined with a sampling frequency of 20-30 kHz, each frame typically contains N = 400-1200 sampling points. Adjacent frames usually overlap to a certain extent, typically 30%-70%, to ensure smooth transitions between frames and retain sufficient information.
[0087] Taking a sampling frequency of 25.6kHz as an example, there are N=1024 sampling points; the overlap between adjacent frames is 50%, that is, 512 points.
[0088] In step 2, windowing is applied to each frame of the voiceprint signal to reduce spectral leakage caused by signal truncation. Preferably, a Hamming window function is used for windowing, and the windowing processing is expressed as follows:
[0089] Equation (1)
[0090] In formula (1):
[0091] h(n) is the weighting coefficient of the Hamming window function at the nth sampling point, which is used to smooth and weight the original frame signal to suppress spectral leakage;
[0092] n is the index of the current sampling point, n=0,1,…N-1;
[0093] N is the total number of sampling points;
[0094] 'a' represents the weight, which is typically between 0.4 and 0.5, for example, 0.46.
[0095] The framed signal is represented as s(n), n=0,1,…N-1. Multiplying it by the window function h(n) yields the windowed signal:
[0096] Equation (2)
[0097] Subsequently, a Fourier transform is performed on the windowed signal S(n) to obtain its frequency domain representation X(k), which is expressed as follows:
[0098] Equation (3)
[0099] In formula (3):
[0100] X(k) is the spectrum signal after Fourier transform;
[0101] k represents the frequency index, indicating the k-th frequency component, with a value ranging from 0 to N-1, corresponding to the discrete frequency f. k= k•f s / N, where f s The sampling frequency;
[0102] n represents the sequence number of each frame sampling point;
[0103] N represents the total number of sampling points per frame;
[0104] j is the imaginary unit, satisfying j 2 = -1.
[0105] In step 2, since the human ear's perception of frequency is non-linear, especially sensitive in the low-frequency range, the linear frequency f is converted into a non-linear Mel frequency f. mel (f) In order to better simulate the working method of the human auditory system, after the conversion, the low-frequency details will be amplified, while the remaining high-frequency components will be reduced in weight, the resolution will decrease, and the information focus will be concentrated in the range that people can hear, which is exactly the effective working condition information needed.
[0106] Furthermore, the conversion formula is as follows:
[0107] Equation (4)
[0108] In equation (4): f is the linear frequency (Hz); f mel (f) is the nonlinear Mel frequency (dimensionless).
[0109] Based on the above mapping, a set of triangular bandpass filters is constructed in the frequency domain, forming a Mel filter bank. The frequency response of each filter m is denoted as H. m (k), where m = 0, 1, ..., M-1. The number of filter banks M is set according to the signal bandwidth and characteristic resolution requirements, and is usually between 64 and 258, for example, 128.
[0110] The center frequency of each triangular filter is denoted as f(m), and its transfer function is H. m The relationship between (k) and the center frequency of the adjacent filter is as follows (expressed as equation (5)):
[0111]
[0112] In equation (5):
[0113] k is the frequency index, representing the kth frequency component;
[0114] f(m) is the center frequency of the m-th filter;
[0115] f(m-1) is the center frequency of the (m-1)th filter;
[0116] f(m+1) is the center frequency of the (m+1)th filter;
[0117] H m (k) represents the weight value of the m-th filter at the k-th frequency point.
[0118] The energy output of each filter is accumulated, and the natural logarithm is taken to obtain the logarithmic energy spectrum:
[0119] Equation (6)
[0120] In formula (6):
[0121] X(k) is the spectrum signal after Fourier transform;
[0122] N represents the number of points in the Fourier transform;
[0123] H m (k) represents the weight value of the m-th filter at the k-th frequency point.
[0124] Next, a Discrete Cosine Transform (DCT) is performed on s(m) to extract its cepstral domain features, yielding the Mel-frequency cepstral coefficients (MFCC) vector:
[0125] Equation (7)
[0126] In equation (7):
[0127] C(n) represents the nth order MFCC coefficient;
[0128] L is the order of the FCC coefficients;
[0129] m is the current filter number.
[0130] In step 2, the MFCC features of multiple consecutive frames are stacked in chronological order to form a two-dimensional image structure, namely the Mel time-frequency map.
[0131] Step 3: Use a convolutional neural network to extract feature vectors from the Mel time-frequency plots of the training set and the validation set, respectively.
[0132] In step 3, instead of traditional manual feature design, a more robust and discriminative feature representation is automatically obtained. That is, a convolutional neural network is used to automatically learn and extract a deep feature vector from the Mel time-frequency map. This vector comprehensively reflects the difference patterns of the voiceprint signal in the time domain, frequency domain, and auditory perception scale under different fault states. This feature vector serves as the basis for Mahalanobis distance calculation, enabling accurate identification of fault types.
[0133] In step 3, the neural network is preferably a deep network with a residual connection structure. This network introduces cross-layer skip connections, directly passing the outputs of the first few layers to subsequent layers, thereby effectively alleviating the vanishing gradient problem and improving network training stability and feature extraction capabilities. The network includes convolutional layers, batch normalization layers, and activation function layers, ultimately outputting a fixed-dimensional deep feature vector.
[0134] The Mel-frequency plot has a size of P×M, where the first dimension represents the time direction (P frames) and the second dimension represents the frequency direction (M Mel filters). To adapt to the input requirements of convolutional neural networks, this two-dimensional matrix is expanded into a three-channel image, that is, it is copied three times in the third dimension to form an RGB format image of size P×M×3, which is used as the network input. In other words, the input image is a three-dimensional tensor of P×M×3, where the first two dimensions correspond to the time direction (P frames) and the frequency direction (M Mel filters), respectively, and the third dimension is an artificially constructed three-channel (RGB) format used to adapt to the input requirements of convolutional neural networks.
[0135] Step 4: Based on the feature vectors extracted from the training set, calculate the average feature vector and covariance matrix for each fault type to characterize the feature distribution of the fault and realize the training of the convolutional neural network.
[0136] In step 4, the mean vector and covariance matrix of each fault type are calculated based on the deep feature vectors extracted from the training set.
[0137] For the c-th type of fault, all samples belonging to this type are selected from the training set, and after processing in step 3, a set of deep feature vectors is obtained: {z1 (c) z2 (c) ,…,z nc (c)}, where z i (c) ∈R d , where represents the d-dimensional deep feature vector of the i-th sample, which is automatically learned by the convolutional neural network from the corresponding Mel time-frequency plot.
[0138] Furthermore, the average feature vector of this fault type is represented as:
[0139] (Equation 8)
[0140] In equation (8):
[0141] z i (c) Let represent the d-dimensional depth feature vector of the i-th sample;
[0142] μ c Let be a d-dimensional vector, representing the average value of the fault characteristics in each dimension;
[0143] n is the number of training samples for each fault type.
[0144] In step 4, the covariance matrix is used to characterize the correlation between each feature dimension and the shape of the data distribution. The specific calculation is as follows:
[0145] Equation (9)
[0146] In equation (9):
[0147] X i and X j Let i and j represent the eigenvectors of the i-th and j-th dimensions, respectively.
[0148] μ i and μ j Its expected value is usually replaced by the sample mean;
[0149] E[] represents the expected value of a mathematical expression, which is approximated by the sample mean in actual calculations.
[0150] Step 5: For the trained convolutional neural network, based on each sample in the validation set, obtain the Mahalanobis distance between its feature vector and the distribution of each fault type. The fault type with the smallest Mahalanobis distance is determined as the fault diagnosis result to evaluate the accuracy of the diagnosis and realize the validation of the convolutional neural network.
[0151] In step 5, Mahalanobis distance is used to measure the similarity between the feature vector of a new sample and the overall distribution of a certain fault category. Essentially, it is the distance from a point to a multivariate distribution. By considering the covariance structure of the data, this distance can effectively suppress the influence of irrelevant features and improve classification robustness. When a new sample is input, Mahalanobis distance determines whether it follows the overall distribution pattern of the target fault category.
[0152] Furthermore, the Mahalanobis distance is obtained by the following formula:
[0153] Equation (10)
[0154] In equation (10), S is the covariance matrix; μ i Let be the mean vector of the distribution, and i represent the number of features. Therefore, it describes the distance from the i-th feature point to the i-th feature distribution.
[0155] In step 5, the Mahalanobis distance reflects the similarity between the validation set and the distribution of various fault characteristics; the smaller the distance, the higher the similarity. Therefore, the category with the smallest distance is taken as the diagnostic result.
[0156] In step 5, if the diagnostic accuracy evaluated in step 5 is lower than a preset threshold (e.g., 80%), it indicates that the current convolutional neural network's generalization ability on this dataset is insufficient, which may be due to factors such as insufficient sample size, low feature discrimination, or noise interference. In this case, the dataset can be expanded by simulating more fault conditions or introducing transfer learning; or the structure of the convolutional neural network can be adjusted (e.g., increasing the number of convolutional layers or introducing an attention mechanism) to improve feature representation ability.
[0157] It should be noted that when 5 to 20 sets of voiceprint samples are collected for each type of fault and the above steps are followed, the present invention has generally achieved a diagnostic accuracy of over 90% in multiple experimental batches, and can reach 100% under some conditions. This performance is significantly better than traditional classifiers (such as Softmax and SVM), verifying the effectiveness and practicality of the method under small sample conditions.
[0158] Step 6: Deploy the validated convolutional neural network, using the preprocessed Mel-time frequency map of the acoustic signature signal of the transformer to be diagnosed as input, and output the fault type according to the principle of minimum Mahalanobis distance to realize the acoustic fault diagnosis of the transformer.
[0159] In step 6, a preset confidence threshold of approximately 3.0 is set (determined based on the 95th percentile of the Mahalanobis distance for each category in the training set). When the minimum Mahalanobis distance of the sample under test exceeds this threshold, it indicates that it does not belong to a known fault mode and may be a novel fault, a compound fault, or strong noise interference. The system automatically marks it as "requiring expert intervention" and generates an abnormal alarm log to avoid operational and maintenance errors due to misjudgment. Through the above methods, highly reliable fault diagnosis is achieved while ensuring safety.
[0160] Secondly, according to the present invention, a transformer acoustic fault diagnosis system based on Mahalanobis distance is provided, the system comprising:
[0161] The signal acquisition module is used to simulate various fault states of the transformer and acquire the acoustic fingerprint signals under the corresponding states as a dataset, which is then divided into a training set and a validation set.
[0162] The preprocessing module is used to preprocess the voiceprint signals of the training set and the validation set respectively to generate the corresponding Mel time-frequency diagrams;
[0163] A feature extraction module is used to extract feature vectors from the Mel time-frequency plots of the training set and the validation set, respectively, using a convolutional neural network.
[0164] The feature distribution characterization module is used to calculate the average feature vector and covariance matrix of each fault type based on the feature vectors extracted from the training set, so as to characterize the feature distribution of the fault.
[0165] The verification module is used to obtain the Mahalanobis distance between the feature vector of each sample in the verification set and the distribution of each fault type, and to determine the fault type with the smallest Mahalanobis distance as the final fault diagnosis result.
[0166] The diagnostic module is used to collect acoustic fingerprint signals of the transformer under its operating state, perform feature extraction and feature distribution characterization processes, and output the fault type according to the principle of minimum Mahalanobis distance to complete the fault diagnosis.
[0167] Thirdly, according to the present invention, a computer-readable storage medium is provided thereon having a computer program stored thereon, which, when executed by a processor, implements the method described in the first aspect. Example
[0168] The present invention is further described below through specific examples; however, these examples are merely exemplary and do not constitute any limitation on the scope of protection of the present invention. Example 1
[0169] This embodiment uses a set of experimental scaled-down transformer prototypes, including three devices with capacities of 2000 kVA, 2000 kVA, and 1000 kVA, corresponding to voltage levels of 35 kV on the high-voltage side, 15.3 kV on the medium-voltage side, and 3.08 kV on the low-voltage side, respectively. All prototypes are based on the same core structure (1 / 2 scale) and coil arrangement (low-voltage / common / series) to ensure comparability of fault characteristics. The parameters of the scaled-down transformer prototypes are shown in Table 1.
[0170] Table 1
[0171]
[0172] (1) Fault simulation and voiceprint signal acquisition
[0173] Five typical fault states were simulated for each transformer in sequence: partial discharge, core loosening, DC bias, heavy overload and winding deformation, and acoustic fingerprint signals were collected simultaneously to construct a dataset; each fault state in the dataset was divided into a training set and a validation set in an 8:2 ratio.
[0174] (1-1) Partial Discharge Fault Simulation
[0175] Artificial surface discharge defects are set inside the transformer, specifically using a plate-needle electrode structure consisting of a tungsten needle electrode and the tank wall. The insulating medium is a 1 mm thick insulating cardboard. Copper foil is attached to one side of the cardboard and is in close contact with the tank wall (plate electrode) for reliable grounding. Copper foil is attached to the other side as the high-voltage end, and the surface distance between the cardboard and the tip of the tungsten needle is set to 10 mm.
[0176] During the test, high voltage was applied from the transformer bushing, and the core, clamps, and shell were all reliably grounded. First, the voltage was pre-pressurized for 5 minutes at 80% PDIV (partial discharge initiation voltage) corresponding to various defects; then, the voltage was stepped up from the PDIV initiation voltage, with each step being 0.5 kV and each step lasting for 3 minutes (maintained at a rate of 1 kV / 6 min).
[0177] If the power system trips automatically after a breakdown occurs, the test should be stopped immediately; if it does not trip, the spark or arc discharge state should be maintained for 1 minute to obtain the acoustic characteristics of the high-energy discharge stage.
[0178] In the partial discharge fault simulation phase, the sampling frequency was set to 25.6kHz, each acquisition lasted 10s, and 10 sets of acoustic fingerprint signals were collected for each type of discharge state to construct a dataset of partial discharge faults.
[0179] (1-2) Simulation of winding deformation fault
[0180] This fault was simulated using a short-circuit impact test platform, employing a phased, stepped loading strategy, alternating between low-to-medium short-circuit and high-to-medium-to-medium short-circuit connection modes. The entire process utilized frequency response sweeping and reactance measurement as core monitoring methods to ensure the stability of the winding's mechanical state. Specifically:
[0181] First, perform system debugging by applying a pre-test impact of 70% of the specified short-circuit current to confirm that the waveform meets the requirements;
[0182] Then, the formal test phase begins, in which short-circuit current surges are applied sequentially from low to high, in increments of 50%, 70%, 90% to 100% of the specified current, with each surge being performed once. After each surge, the reactance value is recorded, and a frequency response sweep test is completed in 10 minutes to diagnose whether the winding has undergone mechanical deformation due to axial or radial displacement.
[0183] Afterwards, switch to no-load connection and let it stand still. Use the temperature sensor to judge the cooling status of the winding. Perform a short no-load and load current check. First, apply 20% voltage level and stabilize for 1 minute, then apply 150% load current and stabilize for 1 minute as a functional check of the internal electrical circuit.
[0184] After completing the above process, the system proceeds to the next level of higher current impact, forming a closed-loop cycle of "impact-monitoring-cooling verification".
[0185] During the data acquisition process, the sampling frequency was set to 25.6 kHz, and the duration of each acquisition was 10 seconds. Due to the brief short-circuit time, the data needed to be truncated, and a 1-second short-circuit segment was extracted as the effective acoustic signature signal for subsequent analysis.
[0186] (1-3) Simulation of core loosening fault
[0187] Open the transformer top cover and loosen the top cover screws. Use an overhead crane to lift the iron core and loosen the fastening devices on the upper left and right sides of the iron core, while keeping the lower fastening devices in place. Gradually increase the distance between the upper iron core clamps to three nodes: 1mm, 3mm, and 6mm. In actual adjustment, it was found that the upper bolts were completely loose when loosened to 6mm. At each node, the sampling frequency was set to 25.6 kHz, the single acquisition time was 10 seconds, and 10 sets of iron core loosening samples were collected at each node as acoustic fingerprint signals.
[0188] (1-4) Simulation of DC bias fault
[0189] like Figure 9 As shown, a DC power supply is connected in series with the neutral point of the 110 kV winding, forming a simulated geomagnetic induced current loop within the substation through the 110 kV line, applying a steady-state DC flux to the transformer core. A simulated injection power supply outputs a steady-state DC current of 1A, gradually increasing the injection current amplitude in 2A increments. Static injection is maintained for 4 minutes at each current level until the core magnetization reaches a steady state, and then acoustic signature signals are collected. For each injected DC current, the sampling frequency is 25.6 kHz, the sampling duration is 2 minutes, and 10 sets of acoustic signature signals are collected for each injected DC current.
[0190] (1-5) Simulation of heavy overload faults
[0191] The device voltage was gradually increased to 4×, 5×, and 6× times the rated voltage using a harmonic source. Steady-state operation was maintained for 6 minutes at each voltage level, and 10 sets of acoustic signature signals were collected at each voltage. During the acquisition process, the sampling frequency was set to 25.6kHz, and the duration of each acquisition session was 2 minutes.
[0192] (2) Voiceprint signal preprocessing
[0193] First, the collected voiceprint signals are processed by framing: the frame length is set to 40 milliseconds; since the sampling frequency is 25.6kHz, the number of sampling points is N=1024; there is 50% overlap between adjacent frames, that is, 512 points.
[0194] Next, the Hamming window is applied, where a = 0.46 in equation (1), i.e., equation (1) is expressed as: h(n) = 0.54 - 0.46cos(2πn / 1023); the framed signal is multiplied by the window function to obtain the windowed signal; then, the framed and windowed signal is subjected to Fourier transform to obtain the spectrum signal X(k);
[0195] Then, the linear frequency f is converted into the nonlinear Mel frequency f according to equation (4). mel(f) Based on this mapping, a set of triangular bandpass filters is constructed in the frequency domain to form a Mel filter bank, totaling 128 filters;
[0196] Finally, the logarithmic energy of each filter output is calculated and subjected to discrete cosine transform to obtain the Mel frequency cepstral coefficients (MFCC) vector. The MFCC features from multiple consecutive frames are stacked in chronological order to form a two-dimensional image structure, namely the Mel time-frequency map. The Mel time-frequency map has a size of P×M = (128×128). To adapt to the input requirements of the convolutional neural network, this two-dimensional matrix is expanded into a three-channel image, that is, copied three times in the third dimension, forming an RGB format image with a size of 128×128×3, which serves as the network input. In other words, the input image is a 128×128×3 three-dimensional tensor used to adapt to the input requirements of the convolutional neural network.
[0197] Figure 3 The diagram illustrates the specific steps of Fourier transforming the voiceprint signal. The upper part shows the time-domain waveform of the original voiceprint signal, the left side of the middle part shows the single-frame signal after framing and windowing, the right side of the middle part shows the spectrum of the frame, and the lower part shows the voiceprint time-frequency diagram formed by splicing multiple frames.
[0198] Figure 4 The Mel time-frequency diagrams are shown for three transformer fault states (partial discharge, winding deformation, and DC bias). It can be seen that winding deformation of the transformer will cause the energy distribution to be more dispersed and the vibration to be more unstable. As time changes, in the mid-to-high frequency band, partial discharge faults will show non-periodic impact characteristics. The frequency band energy of DC bias has shifted significantly, and the energy in the mid-frequency band has increased significantly.
[0199] (3) A convolutional neural network (with residual connection structure, consisting of convolutional layers, batch normalization layers, and activation function layers) is constructed using a residual network model. The input of the network is a 128×128×3 three-dimensional tensor, and the output is a fixed-dimensional depth feature vector with a dimension of 64. This vector comprehensively reflects the discriminative differences of the voiceprint signal on the time-frequency-auditory perception scale, serving as the basis for subsequent Mahalanobis distance calculation. The parameters of the last layer feature vector output by the network model are set to 64.
[0200] (4) Based on the feature vectors extracted from the training set, calculate the average feature vector and covariance matrix of each fault type to characterize the feature distribution of the fault and realize the training of the convolutional neural network.
[0201] (5) For the trained convolutional neural network, based on each sample in the validation set, obtain the Mahalanobis distance between its feature vector and the distribution of each fault type, and determine the fault type with the smallest Mahalanobis distance as the fault diagnosis result, so as to evaluate the accuracy of the diagnosis and realize the validation of the convolutional neural network.
[0202] (6) Deploy the verified convolutional neural network, take the preprocessed Mel time-frequency map of the acoustic fingerprint signal of the transformer to be diagnosed as input, and output the fault type according to the principle of minimum Mahalanobis distance to realize the acoustic fault diagnosis of the transformer.
[0203] Figure 5 The classification results based on four states—normal operation, discharge, winding deformation, and DC bias—are shown, demonstrating that this technology can achieve approximately 100% accuracy.
[0204] The results of the Mahalanobis distance classifier are as follows: Figure 6 As shown, its stability can be observed.
[0205] Comparative Example 1
[0206] Using traditional methods, Mel-frequency cepstral coefficient features are manually extracted from the voiceprint signal, and then machine learning methods are used for classification. Specifically:
[0207] (1) The collected voiceprint signals are preprocessed, including framing, windowing, Fourier transform, Mel filtering and discrete cosine transform, to obtain the MFCC feature vector of each frame;
[0208] (2) Stack the MFCC features of multiple consecutive frames in chronological order to form a two-dimensional feature matrix as sample input;
[0209] (3) Support Vector Machine (SVM) was selected as the classification model. Radial Basis Function (RBF) was used, and the penalty coefficient C and kernel function parameter γ were optimized by grid search. The search range of the penalty coefficient C was set to [0.1, 1, 10, 100], and the optimal value was selected as 10. The search range of the kernel function parameter γ was set to [0.001, 0.01, 0.1, 1], and the optimal value was selected as 0.1.
[0210] (4) Train the SVM model using the handcrafted feature vectors from the training set;
[0211] (5) Input the hand-crafted feature vectors of the validation set into the trained model to obtain the fault category prediction results for each sample.
[0212] Figure 7 The classification results show that hand-designed features are sensitive to complex noise and have limited representation capabilities; the feature extraction process relies on prior knowledge and is difficult to adapt to diverse failure modes; and the classifier performance is limited by feature quality and has poor generalization ability.
[0213] Comparative Example 2
[0214] Using a deep learning framework, discriminative features of the voiceprint signal are automatically learned through end-to-end training, and a Softmax classifier is used to output probability scores. Specifically:
[0215] (1) Convert the voiceprint signal into a Mel-time frequency map and use it as the input to the convolutional neural network;
[0216] (2) Construct a convolutional neural network consistent with Example 1;
[0217] (3) The Softmax function is used to map the raw scores output by the fully connected layer to the probability distribution of each type of fault (the Softmax classifier is the core classification module of the model, and its core principle is to map the raw scores (Logits) output by the fully connected layer to the probability distribution of each fault category through the Softmax function). The specific formula is as follows:
[0218]
[0219] In the formula:
[0220] z i The original score corresponding to the i-th type of fault output by the fully connected layer;
[0221] j represents the total number of fault categories;
[0222] exp(z i ) represents the predicted probability of the i-th type of fault, and the sum of the predicted probabilities of all types is 1.
[0223] (4) Through this mapping, the model can directly output the probability of each sample belonging to each type of fault, making it easy to intuitively determine the category to which the sample belongs. However, there are obvious defects in small sample scenarios. Since the end-to-end deep learning model has a large number of trainable parameters, and the classification task depends on the weight fitting of the last fully connected layer, when the number of training samples for a certain fault category is small, the model is difficult to fully learn the common features of the category, and tends to "memorize" the specific patterns of a few samples, resulting in serious overfitting.
[0224] At this point, the decision boundary defined by the Softmax classifier becomes extremely unstable. Figure 8 The classification results show that even slight data perturbations can cause drastic changes in the boundary, reflecting poor model generalization ability and an unstable training process. In contrast, the Mahalanobis distance classification method proposed in this invention is based on statistical distribution modeling, does not rely on large-scale parameter fitting, and exhibits stronger robustness and stability under small sample conditions.
[0225] The present invention has been described in detail above with reference to preferred embodiments and exemplary examples. However, it should be noted that these specific embodiments are merely illustrative explanations of the invention and do not constitute any limitation on the scope of protection of the invention. Various improvements, equivalent substitutions, or modifications can be made to the technical content and embodiments of the present invention without departing from the spirit and scope of protection of the invention, and all such modifications fall within the scope of protection of the present invention. The scope of protection of the present invention is defined by the appended claims.
Claims
1. A method for diagnosing acoustic faults in a transformer, characterized in that, The method includes: Step 1: Simulate various fault states of the transformer and collect the acoustic fingerprint signals under the corresponding states as a dataset. Divide the dataset into a training set and a validation set. Step 2: Preprocess the voiceprint signals of the training set and the validation set respectively to generate the corresponding Mel time-frequency diagrams; Step 3: Use a convolutional neural network to extract feature vectors from the Mel time-frequency plots of the training set and the validation set, respectively; Step 4: Based on the feature vectors extracted from the training set, calculate the average feature vector and covariance matrix for each fault type to characterize the feature distribution of the fault and realize the training of the convolutional neural network. Step 5: For the trained convolutional neural network, based on each sample in the validation set, obtain the Mahalanobis distance between its feature vector and the distribution of each fault type, and determine the fault type with the smallest Mahalanobis distance as the fault diagnosis result to evaluate the accuracy of the diagnosis and realize the validation of the convolutional neural network. Step 6: Deploy the validated convolutional neural network, using the preprocessed Mel-time frequency map of the acoustic signature signal of the transformer to be diagnosed as input, and output the fault type according to the principle of minimum Mahalanobis distance to realize the acoustic fault diagnosis of the transformer.
2. The method according to claim 1, characterized in that, Preferably, in step 1, the fault states include: partial discharge, loose iron core, DC bias, heavy overload, and winding deformation.
3. The method according to claim 2, characterized in that, The simulation method for core loosening faults includes: adjusting the distance between transformer core clamps to a preset range to simulate different degrees of loosening.
4. The method according to claim 3, characterized in that, The preset range includes multiple nodes with different spacings from 1mm to 6mm.
5. The method according to claim 2, characterized in that, The partial discharge fault is simulated by setting up an artificial surface discharge defect.
6. The method according to claim 2, characterized in that, The heavy overload fault simulation includes: gradually increasing the equipment voltage to 4-6 times the rated voltage using a harmonic source, and collecting multiple sets of overload acoustic signals.
7. The method according to claim 1, characterized in that, In step 2, the preprocessing includes framing, windowing, Fourier transform, and frequency scaling.
8. The method according to claim 1, characterized in that, In step 3, the neural network is a deep network with a residual connection structure.
9. A transformer acoustic fault diagnosis system, characterized in that, The system includes: The signal acquisition module is used to simulate various fault states of the transformer and acquire the acoustic fingerprint signals under the corresponding states as a dataset, which is then divided into a training set and a validation set. The preprocessing module is used to preprocess the voiceprint signals of the training set and the validation set respectively to generate the corresponding Mel time-frequency diagrams; A feature extraction module is used to extract feature vectors from the Mel time-frequency plots of the training set and the validation set, respectively, using a convolutional neural network. The training module is used to calculate the average feature vector and covariance matrix of each fault type based on the feature vector extracted from the training set, so as to characterize the feature distribution of the fault and realize the training of the convolutional neural network. The verification module is used to obtain the Mahalanobis distance between the feature vector of the trained convolutional neural network and the distribution of each fault type based on each sample in the verification set, and to determine the fault type with the smallest Mahalanobis distance as the final fault diagnosis result in order to evaluate the accuracy of the diagnosis and realize the verification of the convolutional neural network. The diagnostic module is used to deploy a validated convolutional neural network. It takes the preprocessed Mel-time frequency map of the acoustic signature signal of the transformer to be diagnosed as input, and outputs the fault type according to the principle of minimum Mahalanobis distance, thereby realizing the acoustic fault diagnosis of the transformer.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1-8.