Transformer testing method and tester

Abstract

The present application discloses a transformer testing method and tester, which relate to the field of transformer testing. It first performs a broadband frequency scan on the specified winding of the transformer to obtain its impedance and phase data within a continuous frequency range. Then, these raw data are preprocessed and feature engineered to extract a feature vector that can fully characterize the broadband state of the transformer. The key is to input this feature vector into a pre-trained VAE model. The VAE model can effectively encode and decode the input data by learning a large number of broadband state patterns of normal transformers. Afterwards, by calculating indicators such as reconstruction error and KL divergence, anomaly scores can be generated and comprehensive judgments can be made, thereby accurately identifying subtle anomalies that are difficult to detect with traditional methods in the early stages of production, effectively improving product reliability, and avoiding the risk of later failures and recalls.

Description

Technical Field

[0001] The present application relates to the field of transformer testing, and more specifically, to a transformer testing method and a tester. Background Art

[0002] Transformers are core components in various electronic devices, especially power supply systems. Their performance and reliability are directly related to the overall quality and operational safety of end products. However, in actual production, traditional transformer testing methods are often limited to measuring parameters such as inductance and turns ratio at a few preset frequency points (e.g., 100kHz). While this point-based or snapshot testing method can effectively screen out products with obvious defects, it struggles to detect rare good products that may perform normally under standard testing but exhibit abnormalities in real-world operating environments. These "rare good" products often only reveal potential problems during assembly or burn-in testing, such as abnormal whistling, overheating, or even premature failure. This not only seriously impacts product reliability but can also lead to large-scale recalls, resulting in significant financial losses and reputational damage to companies. Therefore, a more comprehensive and in-depth transformer testing method is urgently needed to accurately identify and eliminate these potentially defective products at an early stage of production.

[0003] The limitations of existing transformer testing solutions primarily lie in their inability to fully capture the complex behavior of transformers over a wide frequency band. In actual operation, transformers are affected by harmonics of varying frequencies, and their performance is far beyond the full range of what can be fully captured by a single or few frequency points. For example, subtle variations in the manufacturing process, such as uneven winding tension, can cause changes in parasitic capacitance, leading to unexpected resonant peaks at untested high frequencies (such as 850kHz). While these resonances may not affect the inductance at low-frequency test points, when excited by high-order harmonics during power supply operation, they can cause significant energy loss, leading to transformer overheating or system instability. Furthermore, material properties such as the core material's BH curve or dielectric constant exhibit nonlinear variations at different frequencies. Even minor process variations (such as sintering temperature) can exacerbate this nonlinear behavior, causing the shape or smoothness of the transformer's impedance-frequency curve across a wide frequency band to differ significantly from that of standard products. These differences are often difficult to detect using traditional discrete-point testing. These hidden resonances and material nonlinearities lead to the occurrence of out-of-band defects that traditional testing methods cannot effectively identify.

[0004] In order to overcome the above-mentioned defects of the prior art, there is an urgent need in the art for a testing solution that can comprehensively analyze the broadband characteristics of the transformer and effectively identify subtle abnormal patterns. Summary of the Invention

[0005] In order to address the problems in the aforementioned technical background, according to one aspect of the present application, a transformer testing method is provided, which includes: performing a preset broadband frequency scan on a designated winding of the transformer to be tested to obtain an original scanning data packet, wherein the original scanning data packet includes the impedance and phase of each frequency point; preprocessing and feature engineering the original scanning data packet to obtain a broadband state feature vector of the transformer to be tested; inputting the broadband state feature vector of the transformer to be tested into the encoder part of a trained VAE model to obtain a broadband state latent space vector of the transformer to be tested; inputting the broadband state latent space vector of the transformer to be tested into the decoder part of the trained VAE model to obtain a broadband state reconstruction feature vector of the transformer to be tested; and performing abnormality scoring and comprehensive judgment based on the broadband state reconstruction feature vector of the transformer to be tested and the broadband state feature vector of the transformer to be tested to obtain an abnormality judgment result and an abnormality score value.

[0006] According to another aspect of the present application, a transformer tester is provided, the tester including a controller and a memory, the controller being configured to execute computer instructions stored in the memory, the computer instructions being configured to implement the above-mentioned transformer testing method.

[0007] Compared to existing technologies, the present application provides a transformer testing method and tester that first performs a broadband frequency sweep on a designated transformer winding to obtain impedance and phase data over a continuous frequency range. This directly addresses the limitation of traditional testing, which cannot capture the complex behavior of the transformer across the entire frequency band, thereby fully revealing hidden resonance points and material nonlinearities. Next, this raw data is preprocessed and feature-engineered to extract a feature vector that comprehensively characterizes the transformer's broadband state. The key lies in inputting this feature vector into a pre-trained VAE model. By learning the broadband state patterns of a large number of normal transformers, the VAE model can effectively encode and decode the input data. For normal transformers, their feature vectors can be accurately reconstructed by the VAE. However, for the rare good products mentioned in the background technology that appear normal under traditional testing but have potential defects, their broadband state feature vectors will deviate from the normal pattern, significantly increasing the VAE reconstruction error. By calculating metrics such as reconstruction error and KL divergence, this method can generate anomaly scores and perform comprehensive judgments, accurately identifying subtle anomalies that are difficult to detect with traditional methods in the early stages of production, effectively improving product reliability and avoiding the risk of later failures and recalls. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] The above and other purposes, features, and advantages of the present application will become more apparent through a more detailed description of the embodiments of the present application in conjunction with the accompanying drawings. The accompanying drawings are intended to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the drawings, the same reference numerals generally represent the same components or steps.

[0009] Figure 1 Flowchart of a transformer testing method according to an embodiment of the present application.

[0010] Figure 2 4 is a data flow diagram of a transformer testing method according to an embodiment of the present application.

[0011] Figure 3 Flowchart of step S2 in the transformer testing method according to an embodiment of the present application.

[0012] Figure 4 Flowchart of step S5 in the transformer testing method according to an embodiment of the present application.

[0013] Figure 5 Flowchart of step S54 in the transformer testing method according to an embodiment of the present application. DETAILED DESCRIPTION

[0014] The following describes embodiments of the present disclosure in more detail with reference to the accompanying drawings. While the drawings illustrate certain embodiments of the present disclosure, it should be understood that the present disclosure can be implemented in various forms and should not be construed as limited to the embodiments described herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the present disclosure. It should be understood that the drawings and embodiments of the present disclosure are for illustrative purposes only and are not intended to limit the scope of protection of the present disclosure.

[0015] In order to solve the defects in the above technical background, the present application proposes a transformer testing method. Figure 1 Flowchart of a transformer testing method according to an embodiment of the present application. Figure 2 Figure 1 is a data flow diagram of a transformer testing method according to an embodiment of the present application. Figure 1 and Figure 2As shown, the transformer testing method according to the embodiment of the present application includes: S1, performing a preset broadband frequency scan on the designated winding of the transformer to be tested to obtain an original scanning data packet, wherein the original scanning data packet includes the impedance and phase of each frequency point; S2, preprocessing and feature engineering the original scanning data packet to obtain a broadband state feature vector of the transformer to be tested; S3, inputting the broadband state feature vector of the transformer to be tested into the encoder part of the trained VAE model to obtain a broadband state latent space vector of the transformer to be tested; S4, inputting the broadband state latent space vector of the transformer to be tested into the decoder part of the trained VAE model to obtain a broadband state reconstruction feature vector of the transformer to be tested; S5, performing abnormality scoring and comprehensive judgment based on the broadband state reconstruction feature vector of the transformer to be tested and the broadband state feature vector of the transformer to be tested to obtain an abnormality judgment result and an abnormality score value.

[0016] In step S1, a preset broadband frequency scan is performed on the designated winding of the transformer under test to obtain a raw scan data packet. The raw scan data packet includes the impedance and phase at each frequency point. It is understandable that during transformer manufacturing and quality inspection, traditional testing methods often focus on measuring key parameters at a few discrete frequency points, such as inductance or turns ratio at specific frequencies. In practical applications, transformers are subject to complex harmonic environments. Minor internal manufacturing defects, such as parasitic capacitance changes caused by uneven winding tension or nonlinear behavior of the core material at different frequencies, may only exhibit abnormal resonance or impedance characteristics at specific, non-test frequencies. These excellent products may pass traditional testing, but may expose problems such as howling, overheating, or even premature failure during subsequent assembly or aging testing, seriously affecting product reliability and posing a risk of recall. Therefore, to overcome the blind spots of traditional testing methods and achieve early and comprehensive identification of potential transformer defects, this application introduces broadband frequency scanning. This enables the acquisition of the complete impedance and phase response curves of the transformer over a continuous frequency range, revealing subtle frequency-related abnormal patterns that are overlooked in single or a few frequency point tests. This provides comprehensive and high-dimensional basic data for subsequent machine learning-based anomaly detection, ensuring stable and reliable operation of the transformer throughout its life cycle.

[0017] In one embodiment, step S1 is implemented as follows: First, the transformer under test is connected to a precision test instrument with broadband frequency sweep capability, such as a high-precision impedance analyzer or frequency response analyzer. During connection, a specific winding, such as the high-voltage side winding or the low-voltage side winding, is selected as the test object based on the transformer's design and test objectives. Ensure the connection is secure and reliable to avoid errors introduced by contact resistance.

[0018] Before performing a sweep, the broadband frequency scanner needs to be preset. This includes determining the start and end frequencies of the sweep, as well as the frequency stepping method. The setting of these parameters should be based on the designed operating frequency range of the transformer to be tested, potential resonant frequencies, and the accuracy requirements for detecting abnormal modes. For example, for a switching power supply transformer operating at around 100kHz, in order to capture resonances or material nonlinearities that may appear at higher frequencies, the sweep frequency range can be preset to 10kHz to 1MHz, and a logarithmic stepping method can be used, such as setting 100 sweep points per decade, to ensure that a sufficient density of frequency response data is obtained across the entire broadband range.

[0019] Once the parameters are preset, the broadband frequency scanner begins its sweep. Starting from a preset starting frequency, the device gradually increases the frequency while applying a sinusoidal excitation signal of known amplitude and frequency to the specified winding of the transformer under test. At each sweep frequency, the scanner simultaneously measures the current response through the winding and the voltage response across the winding. By precisely measuring the amplitude and phase of these voltage and current signals, the scanner calculates the complex impedance at that frequency in real time. Complex impedance is expressed as the impedance magnitude (ohms) and phase angle (degrees or radians). For example, at a specific frequency f, the scanner measures a voltage V and a current I with a phase difference of θ. Therefore, the impedance at that frequency is Z = V / I, with a phase of θ.

[0020] As the frequency gradually increases, the scanner repeats the above measurement process until it reaches the preset end frequency. All impedance amplitude and phase angle data measured at each frequency point is collected and integrated in frequency order to form a complete raw scan data packet. This data packet is essentially a series of frequency-impedance-phase triplets that comprehensively depict the electrical characteristic response curve of the transformer under test across the entire preset broadband frequency range, providing detailed raw information for subsequent feature extraction and anomaly detection.

[0021] In step S2, the raw scan data packets are preprocessed and feature-engineered to obtain a broadband state feature vector for the transformer under test. Accordingly, after obtaining the raw impedance and phase data packets of the transformer within a broadband frequency range, while these data contain rich frequency response information, their raw form may not be directly suitable for efficient pattern recognition by subsequent machine learning models. The raw impedance and phase data may have different dimensions and numerical ranges. Directly inputting the model may cause some features to have an excessive impact on model training, while the influence of other features is weakened, thereby reducing the model's learning efficiency and generalization ability. Furthermore, impedance and phase are fundamental electrical quantities in circuits, while parameters such as inductance and quality factor more intuitively reflect the transformer's energy storage and loss characteristics and are more sensitive to changes in the transformer's physical structure and material properties. Therefore, in order to convert the raw broadband scan data into a more physically meaningful representation that is more conducive to machine learning model learning and identification of abnormal patterns, and to eliminate dimensional differences between different features, the raw scan data packets need to be preprocessed and feature-engineered.

[0022] In one embodiment, Figure 3 FIG. 1 is a flow chart of step S2 in the transformer testing method according to an embodiment of the present application. Figure 3 As shown, step S2, preprocessing and feature engineering the original scan data packet to obtain the broadband state feature vector of the transformer to be tested, including: S21, calculating the inductance and quality factor based on the impedance and phase of each frequency point; S22, splicing the inductance and quality factor of each frequency point into the original broadband state feature vector of the transformer to be tested; S23, normalizing the original broadband state feature vector of the transformer to be tested to obtain the broadband state feature vector of the transformer to be tested.

[0023] The implementation of step S2 in the above embodiment is as follows: First, perform S21. The original scanning data packet contains a series of frequency points , the corresponding impedance and phase For each frequency point in the data packet According to AC circuit theory, the inductance and quality factor at the frequency point can be calculated using this formula. In one embodiment, step S21, based on the impedance and phase of each frequency point, calculates the inductance and quality factor, including: based on the impedance and phase of each frequency point, calculates the inductance and quality factor using the following formula, the formula being: , ;in, is the frequency value, Phase, expressed in radians. If the original data is in angles, it needs to be converted to radians. is pi, is the impedance, is the sine function, To take the absolute value, is the inductor, is the tangent function, is the quality factor. For example, if at a certain frequency The impedance is measured at , phase ,Right now radians, the calculated inductance , quality factor By traversing all the frequency points in the original scan data packet, the inductance value and quality factor value corresponding to each frequency point can be obtained.

[0024] Then, proceed to S22. For each frequency point in the original scan data packet, a corresponding inductance value and a quality factor value will be obtained. If the original scan data packet contains N frequency points, then N inductance values ​​will be obtained. and N quality factor values These calculated inductances and quality factors are concatenated in frequency order to form a one-dimensional vector. This concatenation can be done by concatenating all inductance values ​​first, then all quality factors, or by concatenating the (L, Q) pairs for each frequency point. This concatenated vector is the original broadband eigenvector of the transformer under test, containing the transformer's inductance and quality factor response information across the entire broadband frequency range.

[0025] Then, proceed to S23. Because the numerical ranges of inductance and quality factor can vary significantly, directly inputting these values ​​into a machine learning model may cause certain features to have an excessive impact on model training. Therefore, it is necessary to normalize the original feature vector of the broadband state of the transformer under test. A common normalization method is Z-score normalization (also known as zero-mean normalization), whose formula is: Where X is the original eigenvalue, μ is the mean of the feature in the training dataset, and σ is the standard deviation of the feature in the training dataset. In practical applications, μ and σ need to be determined in advance through statistical analysis of a large number of original broadband state eigenvectors of normal transformers. These mean and standard deviation are fixed as normalization parameters and used to standardize the eigenvectors of all transformers under test. This normalization process scales the values ​​of all features to a similar range, and the resulting normalized vector becomes the final broadband state eigenvector of the transformer under test.

[0026] In step S3, the broadband state feature vector of the transformer under test is input into the encoder portion of the trained VAE model to obtain the broadband state latent space vector of the transformer under test. It should be understood that although these standardized feature vectors already contain rich transformer state information, their dimensionality may still be high, and the data may contain redundancy or noise. Directly performing anomaly detection in high-dimensional space is not only computationally expensive but also susceptible to the curse of dimensionality, making it difficult for the model to capture the true underlying patterns in the data. Furthermore, the normal state of a transformer may exhibit a nonlinear manifold structure in a complex high-dimensional space, making it difficult for traditional linear methods to effectively distinguish between normal and abnormal states. Therefore, to compress the high-dimensional broadband state feature vector into a low-dimensional, continuous, and well-structured latent space while capturing the inherent probability distribution of the data, this method introduces a variational autoencoder (VAE). By inputting the broadband state feature vector of the transformer under test into the encoder portion of the trained VAE model, the goal is to learn and extract the essential characteristics of the transformer's broadband state and map them into the broadband state latent space vector of the transformer under test.

[0027] In one implementation, the specific process of step S3 is as follows: First, a trained VAE model is required. This VAE model is obtained by unsupervised training on a large number of broadband state feature vector data sets of normal transformers. The training process is intended to allow VAE to learn the intrinsic probability distribution of normal transformer broadband state data. The VAE model consists of two parts: an encoder and a decoder. The encoder's responsibility is to cleverly map high-dimensional input data, that is, the broadband state feature vector of the transformer to be tested, into a low-dimensional latent space. However, unlike traditional autoencoders, the encoder does not directly output a certain latent vector, but outputs the mean of each dimension in the latent space. and the logarithm of the variance , which is the natural logarithm with the natural constant e as the base. This probabilistic representation is a key feature of VAE, making the latent space continuous and sampleable. The decoder is responsible for randomly sampling a vector from this latent space and attempting to reconstruct it back into the original data space. Throughout the training process, the parameters of the VAE model, including all weights and biases within the encoder and decoder, are continuously adjusted and optimized through the backpropagation algorithm and optimizer until the model performance reaches a convergence state, that is, the reconstruction error is minimized and the latent space distribution approaches a standard normal distribution.

[0028] The encoder part of the trained VAE model is composed of multiple fully connected layers or convolutional layers. The specific number of layers and the number of neurons in each layer depend on the complexity of the input data and the desired latent space dimension. For example, the encoder architecture may include: first, an input layer for receiving the broadband state feature vector of the transformer to be tested with a dimension of D. This is followed by the first hidden layer, which is a fully connected layer, for example, containing 256 neurons, and using a nonlinear ReLU activation function to enhance the expressive power of the model. This is followed by the second hidden layer, which is also a fully connected layer, for example, containing 128 neurons, and also using a ReLU activation function. Finally, the output layer of the encoder is divided into two parallel fully connected layers, which are responsible for outputting the latent space mean vector, respectively, denoted as And the latent space variance log vector is recorded as , the dimensions of these two vectors are much smaller than D, for example, set to 32 dimensions to represent a compact and informative low-dimensional representation of the data in a high-dimensional space.

[0029] Specifically, the encoding process is as follows: When the standardized broadband state feature vector of the transformer to be tested is input into the encoder, it will pass through the above layers in sequence according to the preset hierarchical structure. In each fully connected layer, the input vector will be linearly multiplied with the weight matrix of the layer, and then the corresponding bias vector will be added to complete a linear transformation. Subsequently, the result of the linear transformation will be nonlinearly transformed through activation functions such as ReLU, so that the model can learn and capture complex nonlinear relationships in the data. After processing all hidden layers, the output layer of the encoder will generate two key vectors: the latent space mean vector and the latent space variance log vector In order to obtain the final broadband state latent space vector of the transformer to be tested from these two vectors, a reparameterization technique is required. Specifically, a vector with the same value as is randomly sampled from a standard normal distribution N(0,1). and Noise vector of the same dimension Then, the latent space vector is calculated by the following formula ,in, The vector obtained is the final latent space vector of the broadband state of the transformer under test. It is a low-dimensional, continuous vector that represents a sampling point of the broadband state of the transformer under test in the latent space learned by the VAE.

[0030] In step S4, the broadband state latent space vector of the transformer under test is input into the decoder portion of the trained VAE model to obtain a broadband state reconstructed feature vector of the transformer under test. Accordingly, after the broadband state feature vector of the transformer under test is mapped to a low-dimensional latent space by the encoder portion of the VAE model to obtain the broadband state latent space vector of the transformer under test, the key next step is to use this latent space representation to determine the health of the transformer. The core concept of VAE lies in its ability to learn the inherent generation mechanism of normal data. If a transformer under test is normal, then its broadband state feature vector, after being compressed into the latent space by the encoder and then reconstructed from the latent space back to the original data space by the decoder, should be highly similar to the original input feature vector. However, if the transformer under test has an abnormality, even a subtle defect, the representation of its broadband state feature vector in the latent space may deviate from the distribution of normal data, making it difficult for the decoder to accurately reconstruct it back to its original form. Therefore, in order to evaluate the normality or abnormality of the broadband state feature vector of the transformer under test, it is necessary to input the broadband state latent space vector of the transformer under test into the decoder portion of the trained VAE model to obtain the broadband state reconstructed feature vector of the transformer under test.

[0031] In one implementation, the specific process of step S4 is as follows: As previously described, the trained VAE model is obtained through unsupervised training on a large dataset of broadband state feature vectors of normal transformers, with the goal of learning the intrinsic probability distribution of normal data. This model consists of two parts: an encoder and a decoder. The decoder is responsible for mapping vectors in the latent space back to the original data space. During training, the decoder's weights and biases are optimized along with the encoder to ensure that it can reconstruct the original features of normal samples as accurately as possible. The decoder portion of the trained VAE model also consists of multiple fully connected layers. Its architecture is symmetrical or nearly symmetrical to that of the encoder, but its function is reversed, aiming to gradually expand the low-dimensional latent space vector back to the high-dimensional original feature space. For example, the decoder architecture may include: first, an input layer specifically designed to receive the broadband state latent space vector of the transformer under test from the encoder output. This is followed by the first hidden layer, a fully connected layer that may contain, for example, 128 neurons and use a ReLU activation function to introduce the necessary nonlinear transformation capabilities. Following this is the second hidden layer, also a fully connected layer, which may contain, for example, 256 neurons and also uses the Reluctant Unit (ReLU) activation function, further expanding the feature dimensionality and enhancing the model's expressiveness. Finally, there is the output layer, also a fully connected layer, with the number of neurons precisely set to match the dimension D of the original input feature vector. Notably, this output layer uses no activation function, or instead uses a linear activation function. This ensures that the reconstructed feature vector covers its original numerical range. Because the normalized broadband state latent space vector of the transformer under test may contain positive and negative values, the output layer needs to be able to freely generate arbitrary values. The decoding process is as follows: When the broadband state latent space vector of the transformer under test is input into the trained decoder, it passes through each of the aforementioned fully connected layers in a pre-set hierarchical structure. Within each fully connected layer, a core linear transformation is performed: the input vector is multiplied by the pre-trained weight matrix for that layer, and then the corresponding bias vector is added. The result of this linear transformation is then processed using a nonlinear activation function, such as the Reluctant Unit (ReLU), enabling the model to learn and reconstruct the complex nonlinear patterns in the original data. Through these progressive linear transformations and nonlinear activations, the decoder gradually transforms the low-dimensional latent space representation into a high-dimensional feature representation, achieving data upsampling or reconstruction. Ultimately, when the data flows through all hidden layers of the decoder and reaches the output layer, the output layer generates a vector with exactly the same dimensions as the original input feature vector: the reconstructed feature vector of the broadband state of the transformer under test. This reconstructed feature vector represents the broadband state of the transformer under test, generated from the latent space by the VAE model based on its understanding of the normal broadband state of the transformer.

[0032] In particular, considering that inductance and quality factor are calculated based on impedance and phase at frequency, the original eigenvectors contain rich phase information. When these eigenvectors are mapped to a low-dimensional latent space via the VAE model's encoder, the resulting broadband state latent space vector of the transformer under test is expected to not only compress the data but, more importantly, be able to sensitively capture phase-related statistical probability changes in the original data. If the broadband state latent space vector of the transformer under test fails to fully reflect these subtle probability distribution differences, even minor anomalies caused by material nonlinearities or structural defects in the transformer may be smoothed out in the latent space, making it difficult for the model to accurately identify them. Therefore, to improve the VAE model's understanding of the essence of the transformer's broadband state characteristic distribution, especially the sensitivity of its probability distribution parameters (mean and logarithmic variance), the broadband state latent space vector of the transformer under test needs to be optimized for statistical probability characteristics before decoding.

[0033] In another embodiment, step S4, inputting the broadband state latent space vector of the transformer to be tested into the decoder part of the trained VAE model to obtain the broadband state reconstruction feature vector of the transformer to be tested, includes: first, probabilizing the broadband state latent space vector of the transformer to be tested based on the mean and logarithmic variance to obtain the broadband state latent space mean probabilistic vector of the transformer to be tested and the broadband state latent space variance probabilistic vector of the transformer to be tested, that is: , ;in, are the eigenvalues ​​in the broadband state latent space vector of the transformer to be tested, is the latent space mean vector Each eigenvalue in is the logarithmic vector of the latent space variance Each eigenvalue in are the eigenvalues ​​in the mean probabilistic vector of the broadband state latent space of the transformer to be tested, are the individual eigenvalues ​​in the probabilistic variance vector of the broadband state latent space of the transformer under test. It should be understood that the broadband state latent space vector of the transformer under test itself is a combination of the mean and logarithmic variance of the encoder output, obtained through reparameterization. To ensure that each eigenvalue in the latent space better reflects its relative position and importance in the probability distribution, it needs to be probabilistically processed. This probabilistic operation is equivalent to establishing a phase anchor point for each eigenvalue in the latent space based on its mean and logarithmic variance, thereby ensuring that the baseline distribution of different eigenvalues ​​can be more precisely modeled relative to statistical probability characteristics. In this way, the original broadband state latent space vector of the transformer under test is converted into two probabilistic components and, focusing on their relative relationships with the mean and variance, respectively.

[0034] Next, the mean probabilistic vector of the broadband state latent space of the transformer to be tested and the variance probabilistic vector of the broadband state latent space of the transformer to be tested are respectively phased relative to the phase feature to obtain the mean phased vector of the broadband state latent space of the transformer to be tested and the variance phased vector of the broadband state latent space of the transformer to be tested, that is: , ;in, are the eigenvalues ​​in the phased vector of the mean value of the broadband state latent space of the transformer to be tested, are the individual eigenvalues ​​in the phased variance vector of the broadband state latent space of the transformer under test. Accordingly, after the first step of probabilization, a probabilistic vector based on the mean and logarithmic variance is obtained. However, since these probabilistic components may still be misaligned in their distribution, relative phaseization is required to further enhance the sensitivity of the eigenvalues ​​to statistical probability characteristics, especially considering the phase information contained in the original input features. This essentially distributes the confidence level of the latent space distribution by introducing symmetry, allowing the eigenvalues ​​to better adapt to changes in the statistical characteristics of the probability distribution parameters. Through this normalization process, a relative relationship is established between the two, giving them a more phase-based meaning, that is, their relative contribution to the sum. These vectors make the eigenvalues ​​in the latent space more sensitive to statistical probability characteristics, capable of more finely capturing subtle changes in the feature distribution.

[0035] Afterwards, based on the phased mean vector of the broadband state latent space of the transformer to be tested and the phased variance vector of the broadband state latent space of the transformer to be tested, the KL divergence regularization constraint of the overall distribution modulation is performed on the broadband state latent space vector of the transformer to be tested to obtain the broadband state latent space constraint vector of the transformer to be tested, that is: ;in, is the phased vector of the latent space mean value of the broadband state of the transformer to be tested, is the phased vector of the latent space variance of the broadband state of the transformer to be tested, is to calculate the KL divergence, are the eigenvalues ​​in the latent space constraint vector of the broadband state of the transformer under test. It should be understood that the probabilistic and phasing steps of the previous two steps have enhanced the sensitivity of the latent space features to statistical probability characteristics. To further optimize the consistency of the latent space features' sensitivity to statistical changes with the global distribution and establish global co-optimization, a KL divergence regularization constraint is introduced. This constraint ensures that the eigenvalues ​​in the latent space are not only sensitive to local statistical characteristics, but also that their overall distribution is consistent with the model's expected prior distribution (such as the standard normal distribution). This makes the latent space vector of the broadband state of the transformer under test more sensitive to the statistical probability characteristics of the encoder latent space. This regularization term incorporates the phased mean and variance information into the optimization of the original latent space vector. Specifically, a modulation factor can be calculated based on the KL divergence of the two phased vectors with the standard normal distribution and applied to each dimension of the original latent space vector. This form of regularization aligns the latent space features' sensitivity to statistical changes with the global distribution, thereby establishing global co-optimization. Therefore, the eigenvalues ​​within the obtained latent space constraint vector of the broadband state of the transformer to be tested are extremely sensitive to changes in the statistical probability characteristics of the broadband state of the transformer, and can more effectively capture subtle abnormal patterns, providing a more discriminative latent space representation for subsequent reconstruction and abnormality judgment.

[0036] Finally, the broadband state latent space constraint vector of the transformer to be tested is input into the decoder part of the trained VAE model to obtain the broadband state reconstructed feature vector of the transformer to be tested. That is to say, after the above series of optimizations and constraints, a highly sensitive and structured broadband state latent space constraint vector of the transformer to be tested is obtained. In order to use this optimized latent space representation to evaluate the normality of the transformer, it is necessary to remap it back to the original feature space. The decoder, as the other half of the VAE model, can restore the vector in the latent space to the representation of the original data space. Since the latent space constraint vector that has been sensitively optimized for statistical probability characteristics is input into the decoder, the reconstructed feature vector will more accurately reflect the true state of the transformer to be tested, and for abnormal samples, its reconstruction error will be more significant, thereby providing a more reliable basis for subsequent abnormal scoring and judgment. In particular, this decoding process is the same as the above decoding process.

[0037] In step S5, anomaly scoring and comprehensive judgment are performed based on the broadband state reconstruction feature vector of the transformer under test and the broadband state feature vector of the transformer under test to obtain an anomaly judgment result and an anomaly score. It is understandable that although the high-dimensional data has been compressed and reconstructed, quantifying the quality of this reconstruction and using it as a basis for determining whether the transformer is abnormal is a key step in anomaly detection. When training the VAE model, its goal is to minimize the reconstruction error of normal data and make the latent space distribution close to the standard normal distribution. This means that for normal transformers, their original feature vectors can be well reconstructed by the VAE model, resulting in a small reconstruction error. However, for abnormal transformers, since their features deviate from the distribution of normal data, the VAE model has difficulty accurately reconstructing them, resulting in a large reconstruction error. In addition, the degree of deviation of the latent space distribution from the standard normal distribution can also reflect the abnormality of the sample. Therefore, in order to convert the output of the reconstruction process into a quantifiable anomaly indicator and ultimately obtain a clear anomaly judgment result, it is necessary to perform anomaly scoring and comprehensive judgment based on the broadband state reconstruction feature vector of the transformer under test and the broadband state feature vector of the transformer under test. That is, by calculating indicators such as reconstruction error and KL divergence, the degree of deviation of the transformer under test from the normal mode is comprehensively evaluated, and a comprehensive abnormality score value is generated. Based on the comparison of the score value with the preset threshold, the final judgment result of whether the transformer is abnormal is given, thereby achieving accurate identification of potentially defective transformers.

[0038] In one embodiment, step S5, Figure 4 FIG. 5 is a flow chart of step S5 in the transformer testing method according to an embodiment of the present application. Figure 4 As shown, step S5, based on the broadband state reconstruction feature vector of the transformer to be tested and the broadband state feature vector of the transformer to be tested, performs abnormal scoring and comprehensive judgment to obtain an abnormal judgment result and an abnormal score value, including: S51, calculating the reconstruction error between the broadband state reconstruction feature vector of the transformer to be tested and the broadband state feature vector of the transformer to be tested; S52, calculating the KL divergence between the broadband state reconstruction feature vector of the transformer to be tested and the broadband state feature vector of the transformer to be tested; S53, calculating the abnormal score value based on the reconstruction error and the KL divergence; S54, generating the abnormal judgment result based on the comparison between the abnormal score value and the abnormal threshold.

[0039] Step S5 in the above embodiment is implemented as follows: First, proceed to S51. The reconstruction error is used to measure the accuracy of the reconstruction, that is, the difference between the decoder output and the original input. In one embodiment, the reconstruction error is the mean square error or square absolute error between the reconstructed feature vector of the broadband state of the transformer under test and the broadband state feature vector of the transformer under test. If the mean square error is used, the calculation formula is: ;in, are the eigenvalues ​​in the broadband state eigenvector of the transformer to be tested, Each eigenvalue in the broadband state reconstruction eigenvector of the transformer to be tested, is the dimension of the broadband state feature vector of the transformer under test and the broadband state reconstruction feature vector of the transformer under test, is the mean square error.

[0040] If the square absolute error is used, the calculation formula is: ;in, is the squared absolute error. For example, if and =[0.12,0.48,0.75], then the mean square error is ((0.1-0.12)^2+(0.5-0.48)^2+(0.8-0.75)^2) / 3≈0.0011, and the squared absolute error is (|0.1-0.12|+|0.5-0.48|+|0.8-0.75|) / 3=0.03.

[0041] Then, proceed to S52. In particular, the KL divergence here does not directly calculate the divergence between the reconstructed feature vector of the broadband state of the transformer to be tested and the feature vector of the broadband state of the transformer to be tested, but specifically refers to the KL divergence between the latent space distribution in the VAE model, that is, the latent space vector of the broadband state of the transformer to be tested and the standard normal distribution. and the latent space variance log vector The KL divergence is used to measure the difference between the latent space Gaussian distribution defined by the latent space mean and the logarithm of the latent space variance and the standard normal distribution N(0,1). Its calculation formula is: ;in, are the eigenvalues ​​in the latent space variance log vector, are the eigenvalues ​​in the latent space mean vector, is the KL divergence. That is, Logarithm of the variance Convert to actual variance (ensuring positive value), Measure mean shift, Eliminate the bias introduced by the logarithmic transformation. This KL divergence value reflects the degree to which the latent space representation of the test sample deviates from the center of the normal data latent space distribution.

[0042] Then, proceed to S53. In order to comprehensively evaluate the degree of abnormality of the transformer, the reconstruction error and KL divergence need to be fused to obtain the final abnormality score value. In one embodiment, step S53, based on the reconstruction error and the KL divergence, calculates the abnormality score value, including: calculating the weighted sum between the reconstruction error and the KL divergence to obtain the abnormality score value, that is, abnormality score value = α*reconstruction error + β*KL, wherein α and β are preset weight coefficients for balancing the contribution of reconstruction error and KL divergence in the abnormality score. These coefficients need to be determined by experiments and tuning on the validation set. For example, α=0.7 and β=0.3 can be set, indicating that the reconstruction error dominates the abnormality score, and KL divergence serves as an auxiliary indicator.

[0043] Finally, S54. When making a judgment, an abnormality threshold needs to be preset. This threshold is determined by analyzing the abnormality score distribution of a large number of normal samples after the VAE model training is completed. In one embodiment, Figure 5 FIG. 5 is a flow chart of step S54 in the transformer testing method according to an embodiment of the present application. Figure 5 As shown, step S54 generates the abnormality determination result based on a comparison between the abnormality score value and the abnormality threshold, including: S541, in response to the abnormality score value being less than the abnormality threshold, determining the abnormality determination result as normal; S542, in response to the abnormality score value being greater than or equal to the abnormality threshold, determining the abnormality determination result as abnormal. In other words, in response to the abnormality score value being less than the abnormality threshold, the abnormality determination result is determined to be normal. This means that the broadband state of the transformer under test is highly consistent with the pattern of a normal transformer. In response to the abnormality score value being greater than or equal to the abnormality threshold, the abnormality determination result is determined to be abnormal. This means that the broadband state of the transformer under test has significantly deviated from the normal pattern, and a potential defect may exist. For example, if the preset abnormality threshold is 0.05, and the abnormality score value calculated for a transformer under test is 0.08, the determination result is abnormal; if the abnormality score value is 0.03, the determination result is normal.

[0044] In summary, the transformer testing method based on the embodiment of the present application is explained. It first performs a broadband frequency scan on the specified winding of the transformer to obtain its impedance and phase data within a continuous frequency range. This directly addresses the limitation of traditional testing that cannot capture the complex behavior of the transformer across the entire frequency band, thereby fully revealing hidden resonance points and material nonlinear characteristics. Next, these raw data are preprocessed and feature-engineered to extract a feature vector that can fully characterize the broadband state of the transformer. The key is to input this feature vector into a pre-trained VAE model. By learning the broadband state patterns of a large number of normal transformers, the VAE model can effectively encode and decode the input data. For normal transformers, their feature vectors can be accurately reconstructed by the VAE; however, for the rare good products mentioned in the background technology that perform normally under traditional testing but have potential defects, their broadband state feature vectors will deviate from the normal pattern, resulting in a significant increase in the VAE reconstruction error. By calculating indicators such as reconstruction error and KL divergence, this method can generate anomaly scores and perform comprehensive judgments, thereby accurately identifying subtle anomalies that are difficult to detect with traditional methods in the early stages of production, effectively improving product reliability and avoiding the risk of later failures and recalls.

[0045] Specifically, in a specific example of the present application, a transformer tester is further provided, which includes a controller and a memory. The controller is used to execute computer instructions stored in the memory, and the computer instructions are used to implement the above-mentioned transformer testing method.

[0046] As described above, the transformer tester according to the embodiments of the present application can be implemented in various wireless terminals, such as a server equipped with a transformer tester algorithm. In one possible implementation, the transformer tester according to the embodiments of the present application can be integrated into the wireless terminal as a software module and / or hardware module. For example, the transformer tester can be a software module within the operating system of the wireless terminal, or an application developed specifically for the wireless terminal. Of course, the transformer tester can also be one of the many hardware modules of the wireless terminal.

[0047] Alternatively, in another example, the transformer tester and the wireless terminal may be separate devices, and the transformer tester may be connected to the wireless terminal via a wired and / or wireless network and transmit interactive information in accordance with an agreed data format.

[0048] While various embodiments of the present disclosure have been described above, the above descriptions are illustrative, non-exhaustive, and not intended to be limiting of the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is selected to best explain the principles of the embodiments, their practical applications, or improvements to existing technologies, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A transformer testing method, characterized in that: include: Performing a preset broadband frequency scan on a designated winding of the transformer to be tested to obtain an original scan data packet, wherein the original scan data packet includes the impedance and phase at each frequency point; Preprocessing and feature engineering the original scan data packet to obtain a broadband state feature vector of the transformer to be tested; Inputting the broadband state feature vector of the transformer to be tested into the encoder part of the trained VAE model to obtain the broadband state latent space vector of the transformer to be tested; Inputting the broadband state latent space vector of the transformer to be tested into the decoder part of the trained VAE model to obtain the broadband state reconstruction feature vector of the transformer to be tested; Based on the broadband state reconstruction feature vector of the transformer to be tested and the broadband state feature vector of the transformer to be tested, abnormality scoring and comprehensive judgment are performed to obtain an abnormality judgment result and an abnormality score value; The broadband state latent space vector of the transformer to be tested is input into the decoder part of the trained VAE model to obtain the broadband state reconstruction feature vector of the transformer to be tested, including: Probabilistically converting the broadband state latent space vector of the transformer to be tested based on the mean and logarithmic variance to obtain a probabilistic mean vector of the broadband state latent space of the transformer to be tested and a probabilistic variance vector of the broadband state latent space of the transformer to be tested; Respectively performing relative phasing on the phase feature of the broadband state latent space mean probabilistic vector of the transformer to be tested and the broadband state latent space variance probabilistic vector of the transformer to be tested to obtain the broadband state latent space mean phased vector of the transformer to be tested and the broadband state latent space variance phased vector of the transformer to be tested; Based on the phased mean vector of the broadband state latent space of the transformer to be tested and the phased variance vector of the broadband state latent space of the transformer to be tested, the KL divergence regularization constraint of the overall distribution modulation is performed on the broadband state latent space vector of the transformer to be tested to obtain the broadband state latent space constraint vector of the transformer to be tested; The broadband state latent space constraint vector of the transformer to be tested is input into the decoder part of the trained VAE model to obtain the broadband state reconstruction feature vector of the transformer to be tested.

2. The transformer testing method according to claim 1, characterized in that: The raw scan data packet is preprocessed and feature engineered to obtain a broadband state feature vector of the transformer to be tested, including: Calculating the inductance and quality factor based on the impedance and phase at each frequency point; Splicing the inductance and quality factor of each frequency point into the original characteristic vector of the broadband state of the transformer to be tested; The original characteristic vector of the broadband state of the transformer to be measured is normalized to obtain the characteristic vector of the broadband state of the transformer to be measured.

3. The transformer testing method according to claim 2, characterized in that: Calculating the inductance and the quality factor based on the impedance and the phase at each frequency point includes: calculating the inductance and the quality factor based on the impedance and the phase at each frequency point using the following formula, wherein the formula is: ;in, is the frequency value, is the phase, is pi, is the impedance, is the sine function, To take the absolute value, is the inductor, is the tangent function, is the quality factor.

4. The transformer testing method according to claim 1, characterized in that: Based on the broadband state reconstruction feature vector of the transformer to be tested and the broadband state feature vector of the transformer to be tested, anomaly scoring and comprehensive judgment are performed to obtain an anomaly judgment result and an anomaly score value, including: Calculating a reconstruction error between the broadband state reconstruction feature vector of the transformer to be tested and the broadband state feature vector of the transformer to be tested; Calculating the KL divergence between the broadband state reconstruction eigenvector of the transformer to be tested and the broadband state eigenvector of the transformer to be tested; Calculating the anomaly score value based on the reconstruction error and the KL divergence; The abnormality determination result is generated based on a comparison between the abnormality score value and an abnormality threshold.

5. The transformer testing method according to claim 4, characterized in that: The reconstruction error is the mean square error or square absolute error between the broadband state reconstructed feature vector of the transformer to be measured and the broadband state feature vector of the transformer to be measured.

6. The transformer testing method according to claim 5, characterized in that: Calculating the anomaly score value based on the reconstruction error and the KL divergence includes: calculating a weighted sum of the reconstruction error and the KL divergence to obtain the anomaly score value.

7. The transformer testing method according to claim 6, characterized in that: Generating the abnormality determination result based on the comparison between the abnormality score value and the abnormality threshold value includes: In response to the abnormality score value being less than the abnormality threshold, determining that the abnormality determination result is normal; In response to the abnormality score value being greater than or equal to the abnormality threshold, the abnormality determination result is determined to be abnormal.

8. A transformer tester, comprising a controller and a memory, wherein the controller is configured to execute computer instructions stored in the memory, wherein: The computer instructions are used to implement the transformer testing method according to any one of claims 1 to 7.

Images