A transformer fault diagnosis method and device based on a preservation order Gaussian mixture model

Abstract

The application provides a transformer fault diagnosis method and device based on a permutation invariant Gaussian mixture model, and belongs to the field of transformer fault diagnosis in power systems. The method is realized based on a pre-trained fault diagnosis classification model, and the model is trained by a soft label data set generated by a permutation invariant Gaussian mixture model, containing a sequence relationship of transformer fault state categories and label uncertainty. The steps are as follows: collecting transformer multi-modal online monitoring sensor data, obtaining an application feature set through preprocessing and feature extraction, inputting the feature set into the model, and the model outputs the state probability distribution of the transformer at each time based on the learned sequence relationship constraint and uncertainty representation ability. The application effectively integrates the category sequence relationship and the hard label probabilistic process, solves the problem that the hard label cannot reflect the state progression and uncertainty and reduces the discrimination ability of the model boundary samples, and improves the accuracy and reliability of the diagnosis.

Description

Technical Field

[0001] This invention belongs to the field of transformer fault diagnosis technology in power systems, specifically relating to a transformer fault diagnosis method and device based on an order-preserving Gaussian mixture model. Background Technology

[0002] In power systems, transformers are critical equipment, and their operating status directly impacts grid security and power supply reliability. To achieve early fault warning, online monitoring systems are often deployed to collect multimodal data such as temperature, oil level, DGA data, current, voltage, and vibration signals. Traditional transformer fault diagnosis methods are diverse, including empirical rules, threshold alarms, expert systems, DGA-based ratio methods, as well as statistical learning, machine learning (such as support vector machines), and deep learning methods. These methods mostly train classification models using "deterministic labels."

[0003] In industrial settings, different fault states often exhibit continuous or gradual changes. For example, a transformer transitions from normal to partial discharge, or oil temperature changes from normal to overheating. Classifying these states using a single "hard label" fails to capture the uncertainty and transitional nature of these states and can easily lead to overfitting by the classifier, affecting the identification of boundary samples. To address these issues, related fields have proposed methods such as label smoothing, soft labeling / knowledge distillation, Bayesian methods, and credibility calibration, attempting to transform "deterministic labels" into "probability distributions" or perform posterior calibration.

[0004] While the aforementioned methods have improved label processing to some extent, key shortcomings remain in the specific scenario of transformer fault diagnosis. Existing methods fail to effectively combine the "order relationship between categories" with the "hard label probabilistic process," thus failing to generate physically meaningful "state probability distribution labels" suitable for supervised learning. This results in hard labels failing to accurately reflect the asymptotic nature and uncertainty of states in practical applications, leading to insufficient discriminative ability of transformer fault diagnosis models trained on these labels when handling boundary samples, ultimately affecting the accuracy and reliability of the diagnosis. Summary of the Invention

[0005] In view of this, the present invention provides a transformer fault diagnosis method and apparatus based on an order-preserving Gaussian mixture model, aiming to solve the problem in the prior art that the failure to effectively integrate the order relationship between categories and the hard label probabilistic process results in the hard labels failing to reflect the gradual change characteristics of the state, thereby reducing the discrimination ability and diagnostic accuracy of the transformer fault diagnosis model on boundary samples.

[0006] To achieve the above objectives, the technical solution provided by the present invention is as follows:

[0007] In a first aspect, the present invention provides a transformer fault diagnosis method based on an order-preserving Gaussian mixture model, which is implemented based on a pre-trained transformer fault diagnosis classification model. The transformer fault diagnosis classification model is trained on a soft-label dataset generated by the order-preserving Gaussian mixture model. The soft-label dataset is generated by the order-preserving Gaussian mixture model and contains a probability distribution dataset of the order relationship between transformer fault state categories and the uncertainty of labels.

[0008] The method includes:

[0009] Collect multimodal online monitoring sensor data of the transformer;

[0010] Preprocessing is performed on the multimodal online monitoring sensor data to obtain preprocessed data;

[0011] Feature vectors are extracted from the preprocessed data to obtain the application feature set;

[0012] The application feature set is input into the transformer fault diagnosis classification model. Based on the model's ability to represent the order relationship constraints between transformer fault state categories and the label uncertainty learned during the training process on the soft-label dataset, the application feature set is analyzed and processed to obtain the state probability distribution of the transformer at each time step.

[0013] Furthermore, the process of generating soft-label datasets using the order-preserving Gaussian mixture model includes:

[0014] Obtain the sample features corresponding to the application feature set and the original deterministic labels corresponding to the sample features;

[0015] Initialize the Gaussian mixture model to obtain the initial Gaussian mixture model parameters;

[0016] Based on the sample features and the initial Gaussian mixture model parameters, the expectation is solved to obtain the response of each Gaussian component to each sample point;

[0017] The initial failure probability distribution is calculated based on the responsiveness and the indicator function corresponding to the original deterministic label;

[0018] Based on the order relationship constraints between categories in the transformer fault state category, the initial fault probability distribution is calibrated to preserve the order, and the calibrated intermediate fault probability distribution is obtained.

[0019] Based on sample features and responsivity, the parameters of the Gaussian mixture model are updated to maximize the likelihood function, resulting in the updated Gaussian mixture model parameters.

[0020] Repeat the expected solution step, the initial fault probability distribution calculation step, the order-preserving calibration step, and the parameter update step until the change in the likelihood function value of adjacent iterations is less than the preset convergence error, and use the final fault probability distribution as a soft-label dataset.

[0021] Furthermore, multimodal online monitoring sensor data includes, but is not limited to:

[0022] Transformer oil temperature data, tank temperature data, DGA gas concentration data, partial discharge signal characteristic data, load data, current waveform data, and voltage waveform data.

[0023] Furthermore, preprocessing operations are performed on the multimodal online monitoring sensor data, including:

[0024] The multimodal online monitoring sensor data is sequentially processed by filtering and denoising, extracting original signal features, and dividing into time windows to obtain structured data to be processed.

[0025] Furthermore, in extracting feature vectors from the preprocessed data, the feature vectors include at least one of the statistical features, frequency domain features, time-frequency features, and deep learning features of the original signal; among which, the statistical features include the mean and variance, the frequency domain features include the frequency peaks after Fourier transform, and the time-frequency features include the time-frequency domain distribution features after wavelet transform.

[0026] Furthermore, the transformer fault diagnosis classification model is as follows:

[0027] Any of the following: deep neural networks, convolutional neural networks, or graph-based neural networks.

[0028] Furthermore, the total loss used when training the transformer fault diagnosis classification model is:

[0029] L=CE+α·KL

[0030] In the formula, L is the total loss, CE is the standard cross-entropy loss, KL is the KL divergence loss, and α is a hyperparameter used to adjust the weight of the KL divergence loss.

[0031] Secondly, the present invention provides a transformer fault diagnosis device based on an order-preserving Gaussian mixture model, which is implemented based on a pre-trained transformer fault diagnosis classification model. The transformer fault diagnosis classification model is trained on a soft-label dataset generated by the order-preserving Gaussian mixture model. The soft-label dataset is generated by the order-preserving Gaussian mixture model and contains a probability distribution dataset of the order relationship between transformer fault state categories and the uncertainty of labels.

[0032] The device includes:

[0033] The data acquisition module is used to collect multi-modal online monitoring sensor data of the transformer;

[0034] The preprocessing module is used to preprocess the multimodal online monitoring sensor data to obtain preprocessed data.

[0035] The feature extraction module is used to extract feature vectors from the preprocessed data to obtain the application feature set;

[0036] The fault diagnosis module is used to input the application feature set into the transformer fault diagnosis classification model. Based on the model's ability to represent the order relationship constraints between transformer fault state categories and the label uncertainty learned during the training process of the soft-label dataset, the application feature set is analyzed and processed to obtain the state probability distribution of the transformer at each time.

[0037] Thirdly, the present invention provides a computer device, the device including a processor and a memory:

[0038] The memory is used to store computer programs and send the instructions of the computer programs to the processor;

[0039] The processor executes, according to the instructions of the computer program, a transformer fault diagnosis method based on a sequence-preserving Gaussian mixture model, as described in the first aspect.

[0040] Fourthly, the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements a transformer fault diagnosis method based on a sequence-preserving Gaussian mixture model as described in the first aspect.

[0041] In summary, this invention provides a transformer fault diagnosis method and apparatus based on an order-preserving Gaussian mixture model. It is implemented using a pre-trained transformer fault diagnosis classification model, which is trained on a soft-label dataset generated by the order-preserving Gaussian mixture model. The soft-label dataset, generated by the order-preserving Gaussian mixture model, contains the order relationships between transformer fault state categories and the probability distribution of label uncertainty. The method includes the following steps: collecting multimodal online monitoring sensor data of the transformer; preprocessing the multimodal online monitoring sensor data to obtain preprocessed data; extracting feature vectors from the preprocessed data to obtain an application feature set; inputting the application feature set into the transformer fault diagnosis classification model; and analyzing and processing the application feature set based on the model's ability to represent the order relationships between transformer fault state categories and label uncertainty learned during the training process on the soft-label dataset, to obtain the state probability distribution of the transformer at each time step. This invention generates a soft-label dataset containing the category order relationship of transformer fault states and label uncertainty through an order-preserving Gaussian mixture model, and trains a transformer fault diagnosis classification model based on this dataset. It can effectively integrate the category order relationship and the hard label probabilistic process, solve the problem that hard labels cannot reflect the state asymptoticity and uncertainty, thus reducing the model's ability to distinguish boundary samples, and improve the model's ability to distinguish boundary samples as well as the accuracy and reliability of diagnosis. Attached Figure Description

[0042] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0043] Figure 1 A flowchart of a transformer fault diagnosis method based on an order-preserving Gaussian mixture model provided in this embodiment of the invention;

[0044] Figure 2 A block diagram of a transformer fault diagnosis device based on an order-preserving Gaussian mixture model provided in this embodiment of the invention;

[0045] Figure 3 This is a block diagram of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0046] To make the objectives, features, and advantages of this invention more apparent and understandable, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described below are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0047] The background technology of this invention will be further introduced below.

[0048] Transformers are crucial equipment in power systems, and their operating status directly impacts grid security and power supply reliability. To achieve early warning, online monitoring systems are typically deployed to collect multimodal data such as temperature, oil level, Dissolved Gas Analysis (DGA) data, current, voltage, and vibration signals. Traditional transformer fault diagnosis methods include empirical rules, threshold alarms, expert systems, DGA-based ratio methods (such as Duval's trigonometric method), statistical learning and machine learning (support vector machines, decision trees, random forests), and deep learning methods (convolutional neural networks, recurrent neural networks, etc.). These methods use deterministic labels (such as normal, overheating, abnormal discharge levels, etc.) as supervised targets to train classification models.

[0049] In the context of modern smart grid and smart substation construction, transformer online monitoring has evolved from single-point monitoring to multi-dimensional, multi-physical quantity, and full life-cycle monitoring. Massive amounts of multimodal monitoring data provide a rich foundation for data-driven fault diagnosis methods. Meanwhile, internal transformer faults exhibit typical gradual evolution characteristics. Whether thermal, electrical, or insulation aging faults, they all follow a continuous gradual process from normal to minor anomalies, and then to severe faults. During this process, state boundaries are blurred and transitional characteristics are obvious, making it difficult for traditional hard-label-based diagnostic models to capture such gradual information.

[0050] As power systems increasingly demand higher reliability, transformer fault diagnosis requires not only accurate identification of known faults but also quantitative assessment and probabilistic early warning of potential risks, early anomalies, and ambiguous boundary states. However, current mainstream industrial field diagnostic systems still predominantly employ discrete category labels for model training and decision-making. This results in insufficient model accuracy in identifying boundary samples, weak anomaly samples, and gradual process samples, leading to false alarms, missed alarms, or unreliable confidence levels. Consequently, these systems struggle to meet the demands for highly reliable, secure, and interpretable intelligent operation and maintenance.

[0051] In many industrial scenarios, there are no clear boundaries between different fault states; the states change continuously or gradually. For example, there is usually a gradual range from "normal" to "partial discharge"; oil temperature also has intermediate states (mild heating, moderate heating) from normal to high-temperature overheating. Classifying samples into a single "hard label" cannot reflect this uncertainty or transition; moreover, the training process may lead to the classifier overfitting to "manually determined boundaries" or poor discrimination of boundary samples. Therefore, several related strategies exist in the machine learning community and domain engineering: label smoothing, probabilistic labels in soft labeling / knowledge distillation, Bayesian methods providing posterior probabilities, and credibility calibration (such as temperature scaling, platt scaling, and isotonic regression). These methods can transform "deterministic labels" into "probability distributions" or perform posterior calibration for model outputs. However, in scenarios with clear order relationships (such as normal → low temperature overheating → medium temperature overheating → high temperature overheating; or discharges of different energy levels) and multimodal sensing data, existing methods typically do not combine "order relationship constraints between state categories" to enhance the labels.

[0052] In transformer fault mechanisms, states not only differ in category but also exhibit strict physical order relationships. For example, overheating faults show a significant increase with rising temperature, partial discharges show an orderly change from low to high energy as energy increases, and insulation aging shows a continuous deterioration trend over time. This order relationship is an inherent physical law of power equipment fault evolution. If this law is ignored in label construction and model training, even with soft labeling or probabilistic calibration, problems such as contradictions between probability distribution and physical mechanisms, uninterpretable model outputs, and difficulties in engineering implementation may arise. Currently, academia and industry have not yet formed a unified, standardized soft label generation framework that fits the physical mechanisms of transformer faults. Most probabilistic methods only improve generalization performance from a pure machine learning perspective without deeply integrating with transformer state evolution mechanisms, fault severity order relationships, and multimodal data characteristics. This results in models performing well on laboratory datasets but lacking robustness and reliability under complex operating conditions in actual substations.

[0053] The following lists several existing solutions that are closest to this invention, and provides a comparative explanation:

[0054] (A) Soft labeling / label smoothing and knowledge distillation methods

[0055] Explanation: Label smoothing replaces one-hot labels with distributions containing small probabilities, reducing model overconfidence and improving generalization; knowledge distillation uses the soft output probabilities of the teacher model as supervision for the student model.

[0056] The difference from this invention is that these methods typically do not utilize order constraints between categories, nor do they design probabilistic generation mechanisms for the physical meaning of industrial transformer states.

[0057] Label smoothing achieves softening only through uniform distribution or fixed noise, and does not have the physical meaning of fault state evolution; knowledge distillation relies on pre-trained teacher models, which cannot generate soft labels that fit the transformer mechanism from the data layer, and cannot explicitly guarantee that the state probability distribution satisfies the order relation.

[0058] (B) Posterior probability calibration methods (temperature scaling, Platt, order-preserving regression)

[0059] Note: Temperature Scaling adjusts the output confidence by multiplying the softmax logits by a temperature parameter; Platt Scaling uses logistic regression for calibration; Isotonic Regression is used for monotonicity calibration.

[0060] The difference from this invention is that these methods are mainly used for post-processing of model output to calibrate confidence, and are not directly used to generate label distributions of band order relationships based on sample features (target supervision).

[0061] Such post-processing methods do not change the model training objective and supervision signal, and cannot fundamentally improve the model's learning ability for boundary samples and weakly abnormal samples. They also cannot introduce fault sequence relationship constraints during the training phase, so their improvement effect on progressive transformer faults is limited.

[0062] (C) Probabilistic modeling based on Gaussian mixture models (GMM) or mixture density networks (MDN)

[0063] Note: GMM is used to model multimodal phenomena of certain continuous features or latent variables; Hybrid density networks combine neural networks with GMM to predict conditional probability density.

[0064] The difference from this invention is that ordinary GMM / MDN does not introduce "inter-category order relation" constraints to generate category probability distributions; nor is it specifically designed to extend discrete state labels into continuous state distributions.

[0065] Traditional GMM / MDN only fits the feature space distribution without considering the ordered structure of the label space. The generated probability distribution may have high probability jumps and violate the fault gradual change law, so it cannot be directly used for transformer status label enhancement.

[0066] (D) Ordinal Regression Based on Order Relations

[0067] Note: Ordinal regression is used to handle ordered categories (e.g., ratings 1-5). Common methods include thresholding models (proportional odds), continuous latent variable mapping, or converting ordinal numbers into multiple binary classifiers (e.g., cumulative link models).

[0068] The difference from this invention is that ordinal regression mainly focuses on the classification problem itself and rarely generates "state probability distribution" label enhancement methods for supervising model training.

[0069] Ordinal regression focuses primarily on classification task optimization, failing to generate probabilistic labels for supervised training from multimodal monitoring data. It also lacks integration with Gaussian mixture models to achieve unified modeling of distribution fitting and order constraints, making it difficult to directly transfer to transformer fault diagnosis scenarios.

[0070] (E) Papers or patents specifically for transformers / power equipment (regarding fault classification)

[0071] Common practice: Train a classifier directly based on the collected features and output a defined label.

[0072] Patents: Multiple patents for transformer fault diagnosis based on DGA rules, expert systems, or machine learning (specific implementations can be found in patent databases of various countries).

[0073] The difference from this invention is that these schemes typically treat tags as discrete hard tags, ignore the continuity or transition between states, and do not provide a probabilistic tag generation method based on order relations.

[0074] Existing patents related to power equipment are mostly focused on feature extraction, model structure improvement, and rule fusion, with very few involving label space optimization and order-constrained probability label generation. In particular, there is a lack of systematic solutions for joint training of order-preserving mechanisms, probability density fitting, and soft label supervision.

[0075] In summary, the closest existing technologies to this invention are: confidence calibration methods such as label smoothing / knowledge distillation, temperature scaling / order-preserving regression, general GMM / MDN modeling, ordinal regression methods, and conventional transformer fault classification methods. While each addresses some problems, none combine the "order relationship between categories (order preservation)" with the "hard label probabilistic process" to generate physically meaningful "state probability distribution labels" suitable for supervised learning. This invention fills this technological gap by proposing a "transformer state label enhancement method based on an order-preserving Gaussian mixture model," and jointly training it with a classifier (cross-entropy + KL regularization) to improve diagnostic performance.

[0076] The main drawbacks of the aforementioned prior art are as follows:

[0077] (1) Hard labels (one-hot) cannot reflect the asymptotic nature and uncertainty of the state: In actual engineering, there are continuous transition intervals between states. A single fixed label loses valuable supervision information, resulting in poor discrimination of boundary samples, false alarms or false alarms.

[0078] (2) Confidence calibration methods are usually post-processing and cannot enhance the supervision signal during the training stage: for example, temperature scaling is to recalibrate the model output after training, without changing the training process or supervision target, and cannot improve the model's learning of boundary samples.

[0079] (3) Existing probabilistic modeling (GMM / MDN) and ordinal regression usually do not simultaneously satisfy (a) order preservation constraints (i.e., probability distribution forms with clear order relationships for different categories) and (b) hard label probabilistic modeling. Therefore, the generated soft labels may be inconsistent with the engineering semantics.

[0080] (4) For equipment such as transformers, diagnostic methods also need to take into account the interpretability and the consistency of engineering experience in output (for example, temperature-related states should meet certain monotonic / sequential rules), which is difficult to guarantee with existing black-box probabilistic methods.

[0081] (5) Insufficient utilization of multimodal data: Existing methods do not design label enhancement mechanisms for the heterogeneous characteristics of multimodal data such as oil temperature, DGA, partial discharge, vibration, and electrical quantities, making it difficult to fully integrate multi-source information to improve the accuracy of state probability estimation.

[0082] (6) Insufficient engineering feasibility: Most methods are only verified in laboratory environments without considering practical issues such as on-site noise, missing data, and fluctuations in operating conditions. The generated probability labels are unstable and unreproducible, which is not conducive to the deployment and long-term stable operation of online monitoring systems.

[0083] (7) The model has weak ability to express uncertainty: Traditional classification models only output category results and do not provide the confidence level and risk level distribution of each state, which cannot support maintenance personnel to make quantitative decisions, graded early warnings and priority handling.

[0084] This invention proposes a transformer fault diagnosis method and device based on an order-preserving Gaussian mixture model. The primary objective is to enhance the original deterministic state labels into state probability distributions (soft labels), enabling the labels to represent the probability of the transformer being in each state, thus more comprehensively expressing state uncertainty and asymptotic transition zones, and improving the supervision of boundary samples. A secondary objective is to design a label enhancement algorithm combining class-preserving constraints and Gaussian mixture modeling (i.e., a "state label enhancement algorithm based on an order-preserving Gaussian mixture model"), ensuring that the generated state probabilities reflect both the multimodal characteristics of the observed data and the order relationships between states in engineering (e.g., normal → low temperature overheating → medium temperature → high temperature). Furthermore, the generated soft labels are used as supervision targets, optimized with classifier training through joint loss (cross-entropy + KL divergence), ensuring that the model's predicted distribution can both correctly classify and approximate the label probability distribution, improving diagnostic accuracy and uncertainty estimation; enhancing the system's robustness and interpretability (probabilistic output is more conducive to subsequent alarm threshold setting and maintenance decisions), while simultaneously considering implementation complexity and computational efficiency.

[0085] This invention also aims to address problems in practical engineering such as the difficulty of multimodal heterogeneous data fusion, low accuracy of boundary sample identification, uninterpretable model output, and probability distributions that do not conform to fault evolution mechanisms. It upgrades from "hard-label single-point supervision" to "soft-label distribution supervision" by using an order-preserving Gaussian mixture model, improving model reliability and generalization ability from the source of the supervision signal. Furthermore, the method of this invention is lightweight, modular, and scalable, and can be seamlessly embedded into existing substation online monitoring systems. Upgrades can be achieved without significant hardware modifications, facilitating widespread deployment on transformers of different voltage levels and models, and promoting the upgrade of transformer fault diagnosis from "classification and identification" to "quantitative assessment and probabilistic early warning."

[0086] The various embodiments of the present invention will be described in detail below.

[0087] Please see Figure 1 This invention provides a transformer fault diagnosis method based on an order-preserving Gaussian mixture model, which is implemented based on a pre-trained transformer fault diagnosis classification model. The transformer fault diagnosis classification model is trained on a soft-label dataset generated by the order-preserving Gaussian mixture model. The soft-label dataset is generated by the order-preserving Gaussian mixture model and contains a probability distribution dataset of the order relationship between transformer fault state categories and the uncertainty of labels.

[0088] It should be noted that the Gaussian Mixture Model (GMM) is a probability-based unsupervised learning model that fits the probability density distribution of data through a linear combination of multiple Gaussian distributions. Order preservation indicates that there is a natural ordinal relationship between categories, such as temperature level or discharge energy level. In this embodiment, order preservation refers to the inherent order relationship between transformer fault state categories, that is, the transformer fault state has a progressive evolution law from "normal → mild abnormality → moderate abnormality → severe fault". When fitting the data distribution, the order-preserving Gaussian Mixture Model will force the Gaussian components corresponding to different fault states to satisfy the order relationship constraint to avoid the situation of probability distribution crossing and confusion.

[0089] In multimodal data scenarios, different monitoring signals exhibit varying sensitivities to different fault types. For example, DGA gas is sensitive to overheating and discharge, temperature signals are sensitive to thermal fault gradients, and partial discharge signals are sensitive to discharge intensity and development stage. The ordinal-preserving Gaussian mixture model can fit the multimodal distribution at the feature level and force the probability to follow the order of fault evolution and increasing severity at the label level, enabling soft labels to possess both data-driven objectivity and the rationality of physical mechanisms.

[0090] This embodiment introduces an order-preserving constraint to improve the traditional GMM, forming an order-preserving Gaussian mixture model. When fitting the data distribution, the order-preserving Gaussian mixture model forces the Gaussian components corresponding to different fault states to satisfy the order relationship constraint, thus avoiding the situation of probability distribution crossover and disorder.

[0091] The soft-label dataset is generated by an order-preserving Gaussian mixture model, unlike traditional deterministic hard-label datasets that only label "normal" or "fault". Its features include: an order relationship between fault state categories, meaning that each sample label in the dataset is associated with the sequential logic of the gradual evolution of transformer faults, enabling the model to learn the patterns of state transitions; and a probability distribution of label uncertainty, where each sample's label is a probability vector, with each element corresponding to the probability that the sample belongs to a certain fault state (e.g., a sample has a probability of 0.2 for "normal", 0.6 for "mildly abnormal", and 0.2 for "moderately abnormal"). This dataset can directly quantify the uncertainty and gradual transition characteristics of states.

[0092] Soft-label datasets also have the following advantages: they can preserve the fuzzy characteristics of weakly anomalous samples, avoiding subjective bias caused by manual labeling; they can provide smoother supervised gradients for the model, reducing training oscillations and overfitting risks; and they can align with on-site operation and maintenance experience, making the probability output interpretable and facilitating the formation of hierarchical alarm strategies.

[0093] The transformer fault diagnosis classification model uses a soft-label dataset as a supervision signal. During training, it not only learns the mapping relationship between multimodal data and fault states, but also learns the ability to represent state uncertainty under the constraints of order relations. That is, the model understands that "mild anomaly" is a transitional state from "normal" to "moderate anomaly", so it outputs a more reasonable probability distribution for samples at the boundary (such as samples between normal and mild anomaly), rather than forcibly classifying them into a certain category.

[0094] The method includes:

[0095] S1: Collect multi-modal online monitoring sensor data of the transformer.

[0096] It should be noted that multimodal refers to the data collected covering different state characterization dimensions of the transformer. Common modes include electrical quantities (such as winding temperature, partial discharge, three-phase current / voltage), oiling characteristics (such as dissolved gas concentration and moisture content in oil), and mechanical quantities (such as vibration frequency and noise intensity).

[0097] Online monitoring sensor data refers to the continuous collection of operating status data through sensors deployed on the transformer body or auxiliary equipment.

[0098] In practical engineering applications, multimodal data often have different sampling frequencies, different dimensions, and different noise levels. For example, oil temperature is a slowly changing signal, partial discharge and electrical waveforms are fast-changing signals, and DGA is an intermittently sampled signal. This step supports the alignment and fusion of heterogeneous sampling frequency data.

[0099] S2: Perform preprocessing operations on the multimodal online monitoring sensor data to obtain preprocessed data.

[0100] It should be noted that preprocessing refers to a series of data cleaning and standardization operations performed on the raw sensor data to improve data quality and meet the requirements of subsequent feature extraction.

[0101] S3: Extract feature vectors from the preprocessed data to obtain the application feature set.

[0102] It should be noted that the feature vector refers to the key numerical indicators extracted from the preprocessed data that can characterize the operating status of the transformer.

[0103] The application feature set pointer refers to the set of feature vectors filtered for actual diagnostic scenarios.

[0104] S4: Input the application feature set into the transformer fault diagnosis classification model. Based on the model's ability to represent the order relationship constraints between transformer fault state categories and the label uncertainty learned during the training process on the soft-label dataset, analyze and process the application feature set to obtain the state probability distribution of the transformer at each time step.

[0105] It should be noted that the state probability distribution refers to the probability vector output by the model that corresponds to the transformer belonging to each fault state at a certain moment. The vector dimension is equal to the number of fault state categories, and the sum of the vector elements is 1.

[0106] Order relation constraints refer to the gradual evolution law of transformer fault states learned during model training, which constrains the rationality of probability distribution during the inference stage.

[0107] This step involves inputting the feature set into the classification model trained on the soft-label dataset. The model first uses the mapping relationship between multimodal features and fault states learned during training to initially determine the state tendency corresponding to the feature vector. Then, combining the order relation constraints and label uncertainty representation capabilities learned during training, the model adjusts the initial judgment result and outputs a probability distribution that conforms to the state evolution law. For example, for boundary samples between "mild anomaly" and "moderate anomaly," the model does not force the output of a specific label, but instead outputs high probability values ​​for both states (e.g., 0.55 for mild anomaly and 0.4 for moderate anomaly), thereby characterizing the transitional characteristics of the sample.

[0108] In engineering applications, state probability distributions can be directly used to determine risk levels: for example, setting probability thresholds to trigger early warnings, determining fault types based on the highest probability category, judging state stability based on probability entropy values, and judging the speed of fault development based on probability change trends, providing maintenance personnel with multi-dimensional decision-making support. Compared with traditional hard-label outputs, the probability distribution output by this invention can detect weak anomalies and potential risks earlier, reducing the false negative rate; at the same time, it reduces misjudgments of boundary samples, lowering the false alarm rate, and is particularly suitable for the high-reliability monitoring needs of critical equipment such as 220kV and above main transformers.

[0109] This embodiment provides a transformer fault diagnosis method based on an order-preserving Gaussian mixture model. This method innovatively introduces an order-preserving Gaussian mixture model to generate a soft-label dataset that includes both the natural order relationship between transformer fault state categories and label uncertainty, solving the problem that hard labels cannot express state uncertainty and asymptotic transition regions. Simultaneously, a transformer fault diagnosis classification model is trained based on this soft-label dataset, enabling the model to not only learn the mapping relationship between multimodal features and fault states, but also to acquire the ability to constrain the order relationship of fault states and represent label uncertainty, effectively improving the supervision effect of boundary samples. Furthermore, by collecting multimodal online monitoring sensor data and performing preprocessing and multi-dimensional feature extraction, the method comprehensively captures the subtle features of the transformer's operating state, especially the hidden features of boundary samples between different states, providing high-quality data support for the model to output accurate probability distributions. Finally, the method outputs the state probability distribution of the transformer at each moment, rather than a single deterministic state, achieving refined diagnosis of transformer fault states and providing a more comprehensive and scientific basis for operation and maintenance decisions.

[0110] In one embodiment of the present invention, the multimodal online monitoring sensor data includes, but is not limited to:

[0111] Transformer oil temperature data, tank temperature data, DGA gas concentration data, partial discharge signal characteristic data, load data, current waveform data, and voltage waveform data.

[0112] In this embodiment, oil temperature and tank temperature data can intuitively reflect the thermal operating status of the transformer, DGA gas concentration data is the main basis for judging transformer insulation faults, partial discharge signal characteristic data can reflect the intensity and type of partial discharge inside the transformer, and load, current waveform and voltage waveform data can characterize the electrical operating conditions of the transformer.

[0113] In one embodiment of the present invention, preprocessing of multimodal online monitoring sensor data includes:

[0114] The multimodal online monitoring sensor data is sequentially processed by filtering and denoising, extracting original signal features, and dividing into time windows to obtain structured data to be processed.

[0115] In this embodiment, the filtering and denoising operation can effectively eliminate environmental interference noise mixed in during the sensor acquisition process and avoid noise masking the effective signal. The original signal feature extraction is to perform preliminary feature extraction on the denoised modal data and filter out the basic information related to the transformer state. The time window segmentation process divides the continuously acquired time series data into time window data of fixed duration to form structured data to be processed. This preprocessing process can effectively improve data quality, unify data format, and provide a high-quality and standardized data source for subsequent feature vector extraction and model diagnosis.

[0116] In one embodiment of the present invention, in extracting feature vectors from preprocessed data, the feature vectors include at least one of statistical features, frequency domain features, time-frequency features, and deep learning features of the original signal; wherein, the statistical features include mean and variance, the frequency domain features include frequency peaks after Fourier transform, and the time-frequency features include time-frequency domain distribution features after wavelet transform.

[0117] In this embodiment, statistical features include mean and variance, which can reflect the overall distribution characteristics of the data. Frequency domain features are the frequency peaks after Fourier transform, which can capture the key features of the data in the frequency domain. Time-frequency features are the time-frequency domain distribution features after wavelet transform, which can simultaneously reflect the time and frequency characteristics of the data. This multi-type feature extraction method not only covers basic statistical and transform features, but also combines deep learning features to mine the deep correlation information of the data. It can comprehensively capture the feature differences of transformers under different states, especially the subtle features of boundary samples between different states, which provides core support for the classification model to output accurate state probability distribution and effectively improves the supervision effect of boundary samples.

[0118] In one embodiment of the present invention, the process of generating a soft-labeled dataset using an order-preserving Gaussian mixture model includes:

[0119] Step 1: Obtain the sample features corresponding to the application feature set and the original deterministic labels corresponding to the sample features.

[0120] For example, the sample feature corresponding to each sample i is x. i It belongs to the application feature set X={x n}, n=1,2,...,N, where N is the number of sample features; the original deterministic label is y. i (This can be obtained through manual annotation), belonging to the category set C={1,2,...,M}, where M is the number of fault state categories. The training set {(x1,y1),(x2,y2),...,(x...}... N ,y N Convergence error and hyperparameters (>0) as input, failure probability The output represents the probability that the nth sample belongs to the mth type of fault state.

[0121] Step 2: Initialize the Gaussian mixture model to obtain the initial Gaussian mixture model parameters.

[0122] For example, the initialization process includes:

[0123] (1) Initialization of mixing coefficients: The mixing coefficients of the K Gaussian components are initialized using a Dirichlet distribution. .

[0124] (2) Mean initialization: The D-dimensional mean vector of the k-th (k=1,2,3,...,K) Gaussian component Initialize with a D-dimensional standard normal distribution with zero mean and unit variance;

[0125] (3) Initialization of the covariance matrix: The D-dimensional covariance matrix of the k-th Gaussian component Initialize it as an identity matrix (to ensure the stability of the initial distribution);

[0126] (4) Calculation of the initial likelihood function:

[0127]

[0128] These are all the parameters of the GMM, and f(.) is the Gaussian probability density function. This formula calculates the log-likelihood of all samples under the initial parameters, which can measure the degree of model fit to the samples.

[0129] (5) Iteration index initialization: t←1, mark the iteration round.

[0130] Step 3: Based on the sample features and the initial Gaussian mixture model parameters, calculate the expectation to obtain the response of each Gaussian component to each sample point.

[0131] For example, the expected solution is:

[0132]

[0133] in, It is the response (posterior probability) of the k-th Gaussian component to the n-th sample, representing the sample x. n The probability of belonging to the k-th Gaussian component; the numerator is the probability of the k-th Gaussian component on x. n The goodness of fit (mixture coefficient × Gaussian density); the denominator is the sum of all Gaussian components for x. n The sum of the goodness of fits; n=1,2,3,…,N, k=1,2,3,…,K.

[0134] Step 4: Calculate the initial failure probability distribution based on the responsiveness and the indicator function corresponding to the original deterministic label.

[0135] For example, the initial failure probability distribution is calculated as follows:

[0136]

[0137] in, It is an indicator function (yi (Take 1 when =m, otherwise take 0). Let be the initial probability that the nth sample belongs to the mth type of fault state.

[0138] Step 5: Based on the order relationship constraints between categories in the transformer fault state categories, perform order-preserving calibration on the initial fault probability distribution to obtain the calibrated intermediate fault probability distribution.

[0139] For example, the sequence-preserving calibration process includes:

[0140] (1) Define the set of order relations :

[0141]

[0142] According to the original label y n The order constraint pairs for the defined states are 0 to 6, representing normal, low-temperature overheating, medium-temperature overheating, high-temperature overheating, partial discharge, low-energy discharge, and high-energy discharge, respectively. For example, when y=0 (normal state), (1,2) indicates that the probability of low-temperature overheating should be greater than that of medium-temperature overheating; (1,3) indicates that the probability of low-temperature overheating should be greater than that of high-temperature overheating.

[0143] (2) The expression for order-preserving calibration is:

[0144]

[0145]

[0146] in, yes (Initial probability) and KL divergence of (calibrated probabilities) (ensuring that the calibrated probabilities are close to the initial values); regularization term This is to enforce the satisfaction of order relations. This represents the intermediate failure probability distribution under the current step.

[0147] Step 6: Based on the sample features and responsivity, update the parameters of the Gaussian mixture model to maximize the likelihood function, and obtain the updated Gaussian mixture model parameters.

[0148] For example, the steps for maximizing the likelihood function include:

[0149] (1) Mixing coefficient update:

[0150]

[0151] This formula indicates that the mixing coefficient of the k-th Gaussian component is equal to the mean of the responses of all samples to that component.

[0152] (2) Mean update:

[0153]

[0154] This formula represents the mean of the k-th Gaussian component, which is the value of the sample x. n The average of weighted response rates.

[0155] (3) Covariance update:

[0156]

[0157] This formula represents the covariance of the k-th Gaussian component, which is the weighted covariance of the sample deviation from the mean.

[0158] (4) Likelihood function update:

[0159]

[0160] This formula is used to calculate the log-likelihood value after updating the parameters.

[0161] Step 7: Repeat the expected solution step, the initial fault probability distribution calculation step, the order-preserving calibration step, and the parameter update step until the change in the likelihood function value of adjacent iterations is less than the preset convergence error, and use the final fault probability distribution as the soft-label dataset.

[0162] This step iteratively calculates the expected value of the Gaussian component responsivity, calculates the initial fault probability based on the responsivity and indicator function, performs order-preserving calibration with embedded transformer fault state order constraints, and updates model parameters by maximizing the likelihood function. By repeatedly executing these steps, the model's fit to the sample features is continuously optimized, and the order of the labels is corrected until the change in the log-likelihood function between two consecutive iterations is less than the preset convergence error. At this point, the final output fault probability distribution is the soft-label dataset, i.e. .

[0163] Through the above iterative process, the order-preserving Gaussian mixture model can simultaneously optimize the data distribution fit and the rationality of the label order in each round, so that the generated soft labels not only conform to the statistical laws of the data, but also strictly follow the transformer fault evolution mechanism, thus solving the problems of unclear physical meaning and chaotic order relationship of traditional soft labels.

[0164] This embodiment addresses the problem in traditional transformer fault diagnosis where hard labels fail to reflect the gradual nature and uncertainty of states by employing a state label enhancement method based on an Order-preserving Gaussian Mixture Model (OP-GMM). This method builds upon the observed characteristics of the transformer, combining the category order relationship of fault states with a Gaussian Mixture Model (GMM) for modeling. Monotonicity constraints are added during the EM (Expectation-Maximization) iteration or update of the GMM parameters to maintain order preservation. Then, order-preserving projection and normalization are used to process the category scores, ultimately generating a probability distribution reflecting the different fault states of each transformer sample as an enhanced soft label. This label not only captures the complex distribution of sample features using GMM, reflecting the uncertainty and gradual transition characteristics of states, but also relies on order relationship constraints to ensure that the label conforms to the natural logic of fault severity. This provides more reasonable supervisory data for subsequent transformer fault diagnosis models, effectively improving the model's ability to discriminate boundary samples, and ultimately optimizing the accuracy and reliability of fault diagnosis.

[0165] In one embodiment of the present invention, the transformer fault diagnosis classification model is as follows:

[0166] Any of the following: deep neural networks, convolutional neural networks, or graph-based neural networks.

[0167] In this embodiment, three types of models—Deep Neural Network (DNN), Convolutional Neural Network (CNN), or graph-based neural network—can all adapt to the training requirements of soft-label datasets generated based on order-preserving Gaussian mixture models. Deep neural networks can perform high-dimensional mapping and nonlinear fitting of transformer multimodal features through multilayer perceptron structures, fully exploring the complex correlation between features and fault states. Convolutional neural networks can extract local correlation information from temporal or spatial monitoring features (such as current / voltage waveforms and partial discharge signal features), enhancing the ability to identify fault features. Graph-based neural networks can treat monitoring features of different dimensions of the transformer as graph nodes, capturing the topological relationships and mutual influences between features. After being trained on the soft-label dataset, all three types of models can learn the order relation constraints of fault states and the ability to represent label uncertainty, ultimately realizing the analysis of the input feature set and outputting the state probability distribution of the transformer at each time step.

[0168] In a further embodiment of the present invention, the total loss used when training the transformer fault diagnosis classification model is:

[0169] L=CE+α·KL

[0170] In the formula, L is the total loss, CE is the standard cross-entropy loss (using the original one-hot labels as the target), KL is the KL divergence loss, used to constrain the model's predicted distribution to be close to the label distribution of the order-preserving GMM enhancement; α is a hyperparameter used to adjust the weights of the KL divergence loss.

[0171] The hyperparameter α can be adjusted according to the characteristics of the field data and engineering requirements: when α=0, it degenerates into traditional hard label training; when α increases, the model pays more attention to learning order relations and probability distribution, which is suitable for scenarios with many boundary samples and many weak anomalies; by reasonably selecting α, the optimal balance between classification accuracy and probability distribution rationality can be achieved, and the robustness of the model under actual working conditions can be improved.

[0172] After training the transformer fault diagnosis classification model based on the above loss function, the model performance is comprehensively evaluated on the validation and test sets using metrics such as classification accuracy, recall, precision, and confusion matrix. Once the model performance meets the preset requirements, it is deployed to the industrial system to achieve online diagnosis of transformer faults, and finally outputs the probability distribution corresponding to the transformer operating status at each monitoring time.

[0173] Based on the above embodiments, it can be seen that the present invention has the following advantages compared with the prior art:

[0174] (1) Richer supervision information: Compared with hard labels / one-hot, the soft labels generated by the present invention through order-preserving GMM can represent the probability distribution of samples in multiple states, making up for the defect of binary label of boundary samples, thereby improving the classifier's ability to discriminate boundary samples and its generalization.

[0175] (2) Engineering semantic consistency and interpretability: The introduction of order-preserving constraints makes the generated probability distribution consistent with engineering knowledge (such as the monotonicity / sequentiality of temperature level and discharge energy level), improving the credibility and interpretability of the output and facilitating operation and maintenance decisions.

[0176] (3) Flexible expansion and deployment: This invention can be implemented under different features, different classifier structures and different GMM configurations; and the generated probability labels can be directly used for online alarms, risk classification and visualization, which is conducive to engineering implementation.

[0177] (4) Deep fusion of multimodal data: The method of the present invention can effectively fuse heterogeneous data such as oil temperature, DGA, partial discharge, and electrical quantity, and make full use of multi-dimensional information to improve the accuracy of state probability estimation and adapt to the characteristics of modern transformer online monitoring data.

[0178] (5) Strong ability to detect weak anomalies and early faults: Soft label supervision makes the model more sensitive to weak changes and transition states, and can identify potential faults earlier, reducing the risk of missed fault detection and sudden shutdown.

[0179] (6) Enhancement of model robustness and generalization: The combined loss and order constraint reduce the sensitivity of the model to noise and annotation errors, enabling it to maintain stable performance under different loads, in different environments, and for different transformer models.

[0180] Based on the above technical solutions of the present invention, in order to further expand the applicable scenarios of the method and adapt to different engineering practice conditions and data characteristics, there are also various equivalent alternative technical solutions that do not deviate from the core order-preserving idea of the present invention. These solutions provide different technical implementation paths for key links such as label generation, probability model selection, order-preserving constraint implementation, loss function design, feature engineering, and model structure construction, and specifically include the following categories:

[0181] (1) Alternative methods for label generation (still retaining the order-preserving idea)

[0182] Use an ordinal regression model to output the cumulative probabilities of each category, then obtain the category probability distribution through differentiation, and perform smoothing after output to reflect the characteristics of multimodal observations. That is, first predict the ordinal regression function of P(Y≤k | x), and then obtain the probability by P(Y=k | x)=P(Y≤k)-P(Y≤k-1).

[0183] Model with a cumulative distribution function (CDF): For continuous latent variable z ~ GMM, define category thresholds, and the category probabilities are given by the probabilities that the latent variable falls within the corresponding intervals. Probability labels are obtained through joint learning of GMM parameters and thresholds.

[0184] (2) Other probability models to replace GMM

[0185] Use a mixture exponential family model (such as a mixture Laplace) or kernel density estimation (Kernel Density Estimation, KDE) to estimate the conditional densities of each category to obtain category likelihoods and perform order constraint correction.

[0186] Replace the traditional GMM with a mixture density network (Mixture Density Network, MDN), allowing the neural network to directly output the mixture component parameters and combine order-preserving constraints to generate probability labels.

[0187] (3) Alternative ways to implement order preservation

[0188] Use other ranking losses to make the category scores satisfy the order relationship (for example, for each pair of categories (i<j), add a ranking constraint based on Hinge Loss loss L_rank = max(0, margin - s_j + s_i)).

[0189] (4) Loss function substitution

[0190] In addition to KL divergence, Jensen-Shannon divergence (JS divergence) and other methods can be used to make the model's predictions closer to the augmented labels.

[0191] Temperature scaling can be used to dynamically adjust the softmax temperature during training, making the model output probability smooth and consistent with the augmented label.

[0192] (5) Feature engineering and model structure replacement

[0193] Features can be replaced by traditional statistical features with deep learning automatic features (CNN / RNN / Transformer) or a combination of both; the classifier can use any probabilistic output model (softmax DNN, Bayesian neural network, probabilistic output of random forest, etc.).

[0194] Based on the same inventive concept, this application also provides a transformer fault diagnosis device based on the order-preserving Gaussian mixture model for implementing the above-mentioned transformer fault diagnosis method based on the order-preserving Gaussian mixture model. The solution provided by this device is similar to the solution described in the above method. Therefore, the specific limitations in the embodiments of the transformer fault diagnosis device based on the order-preserving Gaussian mixture model provided below can be found in the limitations of the transformer fault diagnosis method based on the order-preserving Gaussian mixture model described above, and will not be repeated here.

[0195] Please see Figure 2 This invention also provides a transformer fault diagnosis device based on an order-preserving Gaussian mixture model, which is implemented based on a pre-trained transformer fault diagnosis classification model. The transformer fault diagnosis classification model is trained on a soft-label dataset generated by the order-preserving Gaussian mixture model. The soft-label dataset is generated by the order-preserving Gaussian mixture model and contains a dataset of probability distributions of the order relationship between transformer fault state categories and the uncertainty of labels.

[0196] The device includes:

[0197] The data acquisition module is used to collect multi-modal online monitoring sensor data of the transformer;

[0198] The preprocessing module is used to preprocess the multimodal online monitoring sensor data to obtain preprocessed data.

[0199] The feature extraction module is used to extract feature vectors from the preprocessed data to obtain the application feature set;

[0200] The fault diagnosis module is used to input the application feature set into the transformer fault diagnosis classification model. Based on the model's ability to represent the order relationship constraints between transformer fault state categories and the label uncertainty learned during the training process of the soft-label dataset, the application feature set is analyzed and processed to obtain the state probability distribution of the transformer at each time.

[0201] Furthermore, the process of generating soft-label datasets using the order-preserving Gaussian mixture model includes:

[0202] Obtain the sample features corresponding to the application feature set and the original deterministic labels corresponding to the sample features;

[0203] Initialize the Gaussian mixture model to obtain the initial Gaussian mixture model parameters;

[0204] Based on the sample features and the initial Gaussian mixture model parameters, the expectation is solved to obtain the response of each Gaussian component to each sample point;

[0205] The initial failure probability distribution is calculated based on the responsiveness and the indicator function corresponding to the original deterministic label;

[0206] Based on the order relationship constraints between categories in the transformer fault state category, the initial fault probability distribution is calibrated to preserve the order, and the calibrated intermediate fault probability distribution is obtained.

[0207] Based on sample features and responsivity, the parameters of the Gaussian mixture model are updated to maximize the likelihood function, resulting in the updated Gaussian mixture model parameters.

[0208] Repeat the expected solution step, the initial fault probability distribution calculation step, the order-preserving calibration step, and the parameter update step until the change in the likelihood function value of adjacent iterations is less than the preset convergence error, and use the final fault probability distribution as a soft-label dataset.

[0209] Furthermore, multimodal online monitoring sensor data includes, but is not limited to:

[0210] Transformer oil temperature data, tank temperature data, DGA gas concentration data, partial discharge signal characteristic data, load data, current waveform data, and voltage waveform data.

[0211] Furthermore, preprocessing operations are performed on the multimodal online monitoring sensor data, including:

[0212] The multimodal online monitoring sensor data is sequentially processed by filtering and denoising, extracting original signal features, and dividing into time windows to obtain structured data to be processed.

[0213] Furthermore, in extracting feature vectors from the preprocessed data, the feature vectors include at least one of the statistical features, frequency domain features, time-frequency features, and deep learning features of the original signal; among which, the statistical features include the mean and variance, the frequency domain features include the frequency peaks after Fourier transform, and the time-frequency features include the time-frequency domain distribution features after wavelet transform.

[0214] Furthermore, the transformer fault diagnosis classification model is as follows:

[0215] Any of the following: deep neural networks, convolutional neural networks, or graph-based neural networks.

[0216] Furthermore, the total loss used when training the transformer fault diagnosis classification model is:

[0217] L=CE+α·KL

[0218] In the formula, L is the total loss, CE is the standard cross-entropy loss, KL is the KL divergence loss, and α is a hyperparameter used to adjust the weight of the KL divergence loss.

[0219] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the system can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0220] Reference Figure 3 The present invention also provides a computer device, including: a memory and a processor, and a computer program stored in the memory. When the computer program is executed on the processor, it implements the transformer fault diagnosis method based on the order-preserving Gaussian mixture model as described in any of the above methods.

[0221] The computer device may be a desktop computer, laptop, handheld computer, or cloud server, etc. This computer device may include, but is not limited to, a processor and memory. Those skilled in the art will understand that... Figure 3The examples of computer devices are merely examples and do not constitute a limitation on computer devices. They may include more or fewer components than shown in the illustration, or combinations of certain components, or different components. For example, they may also include input / output devices, network access devices, etc.

[0222] The processor referred to can be a Central Processing Unit (CPU), but it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.

[0223] In some embodiments, the memory may be an internal storage unit of the computer device, such as a hard drive or RAM. In other embodiments, the memory may be an external storage device of the computer device, such as a plug-in hard drive, Smart Media Card (SMC), Secure Digital (SD) card, or Flash Card. Furthermore, the memory may include both internal and external storage units of the computer device. The memory is used to store the operating system, applications, boot loader, data, and other programs, such as the program code of the computer program. The memory can also be used to temporarily store data that has been output or will be output.

[0224] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the transformer fault diagnosis method based on the order-preserving Gaussian mixture model as described in any of the above methods.

[0225] In this embodiment, if the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include at least: any entity or device capable of carrying computer program code to a photographing device / terminal device, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium. Examples include USB flash drives, portable hard drives, magnetic disks, or optical disks. In some jurisdictions, according to legislation and patent practice, computer-readable media cannot be electrical carrier signals or telecommunication signals.

[0226] This invention provides a computer program product, including a computer program that, when executed by a processor, implements the transformer fault diagnosis method based on the order-preserving Gaussian mixture model as described in any of the above methods.

[0227] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0228] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0229] In the embodiments disclosed in this application, it should be understood that the disclosed devices / terminal equipment and methods can be implemented in other ways. For example, the device / terminal equipment embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling or direct coupling or communication connection may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0230] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A transformer fault diagnosis method based on an order-preserving Gaussian mixture model, characterized in that, The transformer fault diagnosis classification model is implemented based on a pre-trained model. The model is trained on a soft-label dataset generated by an order-preserving Gaussian mixture model. The soft-label dataset is generated by the order-preserving Gaussian mixture model and includes the order relationship between transformer fault state categories and the probability distribution of label uncertainty. The method includes: Collect multimodal online monitoring sensor data of the transformer; The multimodal online monitoring sensor data is preprocessed to obtain preprocessed data; Feature vectors are extracted from the preprocessed data to obtain the application feature set; The application feature set is input into the transformer fault diagnosis classification model. Based on the model's ability to represent the order relationship constraints between transformer fault state categories and the label uncertainty learned during the training process of the soft-label dataset, the application feature set is analyzed and processed to obtain the state probability distribution of the transformer at each time.

2. The transformer fault diagnosis method based on the order-preserving Gaussian mixture model according to claim 1, characterized in that, The process of generating soft-labeled datasets using the order-preserving Gaussian mixture model includes: Obtain the sample features corresponding to the application feature set and the original deterministic labels corresponding to the sample features; Initialize the Gaussian mixture model to obtain the initial Gaussian mixture model parameters; Based on the sample features and the initial Gaussian mixture model parameters, the expectation is calculated to obtain the response of each Gaussian component to each sample point. Based on the responsiveness and the indicator function corresponding to the original deterministic label, calculate the initial fault probability distribution; Based on the order relationship constraints between categories in the transformer fault state category, the initial fault probability distribution is calibrated in order preservation to obtain the calibrated intermediate fault probability distribution. Based on the sample features and the responsivity, the parameters of the Gaussian mixture model are updated to maximize the likelihood function, resulting in the updated Gaussian mixture model parameters. Repeat the expected solution step, the initial fault probability distribution calculation step, the order-preserving calibration step, and the parameter update step until the change in the likelihood function value of adjacent iterations is less than the preset convergence error, and use the final fault probability distribution as the soft-label dataset.

3. The transformer fault diagnosis method based on the order-preserving Gaussian mixture model according to claim 1, characterized in that, The multimodal online monitoring sensor data includes, but is not limited to: Transformer oil temperature data, tank temperature data, DGA gas concentration data, partial discharge signal characteristic data, load data, current waveform data, and voltage waveform data.

4. The transformer fault diagnosis method based on the order-preserving Gaussian mixture model according to claim 1, characterized in that, Preprocessing operations are performed on the multimodal online monitoring sensor data, including: The multimodal online monitoring sensor data is sequentially processed by filtering and denoising, extracting original signal features, and dividing into time windows to obtain structured data to be processed.

5. The transformer fault diagnosis method based on the order-preserving Gaussian mixture model according to claim 1, characterized in that, In extracting feature vectors from the preprocessed data, the feature vectors include at least one of the statistical features, frequency domain features, time-frequency features, and deep learning features of the original signal; wherein, the statistical features include the mean and variance, the frequency domain features include the frequency peaks after Fourier transform, and the time-frequency features include the time-frequency domain distribution features after wavelet transform.

6. The transformer fault diagnosis method based on the order-preserving Gaussian mixture model according to any one of claims 1-5, characterized in that, The transformer fault diagnosis classification model is as follows: Any of the following: deep neural networks, convolutional neural networks, or graph-based neural networks.

7. The transformer fault diagnosis method based on the order-preserving Gaussian mixture model according to claim 5, characterized in that, The total loss used when training the transformer fault diagnosis classification model is: L=CE+α·KL In the formula, L is the total loss, CE is the standard cross-entropy loss, KL is the KL divergence loss, and α is a hyperparameter used to adjust the weight of the KL divergence loss.

8. A transformer fault diagnosis device based on an order-preserving Gaussian mixture model, characterized in that, The transformer fault diagnosis classification model is implemented based on a pre-trained model. The model is trained on a soft-label dataset generated by an order-preserving Gaussian mixture model. The soft-label dataset is generated by the order-preserving Gaussian mixture model and includes the order relationship between transformer fault state categories and the probability distribution of label uncertainty. The device includes: The data acquisition module is used to collect multi-modal online monitoring sensor data of the transformer; The preprocessing module is used to perform preprocessing operations on the multimodal online monitoring sensor data to obtain preprocessed data; The feature extraction module is used to extract feature vectors from the preprocessed data to obtain an application feature set; The fault diagnosis module is used to input the application feature set into the transformer fault diagnosis classification model. Based on the model's ability to represent the order relationship constraints between transformer fault state categories and the uncertainty of labels learned during the training process of the soft-label dataset, the application feature set is analyzed and processed to obtain the state probability distribution of the transformer at each time.

9. A computer device, characterized in that, The device includes a processor and a memory: The memory is used to store computer programs and send the instructions of the computer programs to the processor; The processor executes, according to the instructions of the computer program, a transformer fault diagnosis method based on a sequence-preserving Gaussian mixture model as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, implements a transformer fault diagnosis method based on a sequence-preserving Gaussian mixture model as described in any one of claims 1-7.

Images