Electrical equipment fault early warning method and system driven by mathematical fusion

Abstract

The invention discloses a mathematical fusion driven power equipment fault early warning method and system, and belongs to the technical field of power equipment operation and maintaining.The method comprises the steps that by extracting fault evolution laws of key state variables such as partial discharge and gas dissolved in oil from historical operation data, an index change mathematical model, a linear change mathematical model and a statistical distribution mathematical model are established respectively; difference processing is carried out on various laws, and the deviation between an actual value and a fitting curve is calculated; carrying out weighted association on the deviation and a model prediction probability, and constructing a comprehensive loss function containing cross entropy loss and a multi-source law loss item; a neural network based on an attention mechanism is adopted, and the comprehensive loss function is introduced in the training process for optimization, so that the model follows a physical evolution path while learning data features; meanwhile, a lightweight edge calculation framework is designed, and efficient reasoning and real-time early warning of the mathematical fusion model on the site of the converter station are achieved.

Description

Technical Field

[0001] This application relates to the field of power equipment operation and maintenance technology, and in particular to a power equipment fault early warning method and system driven by mathematical fusion. Background Technology

[0002] In the field of power system operation and maintenance, converter transformers, as core equipment in ultra-high voltage direct current (UHVDC) transmission projects, directly affect the safety and stability of the power grid. To achieve early identification and warning of potential faults in converter transformers, current mainstream technologies fall into two main categories: one is the mechanism-driven approach based on multi-physics coupled simulation and fault mechanism analysis. This method establishes multi-field coupled models involving electromagnetics, thermodynamics, and fluid dynamics to simulate the development process of internal defects in the equipment, uncover fault evolution patterns, and thus achieve condition diagnosis. The other is the data-driven approach based on big data and artificial intelligence. This method uses historical monitoring data to train deep learning models and constructs a fault classifier through the fusion analysis of multi-modal information such as electrical, non-electrical, and chemical parameters to achieve automatic identification. These two approaches represent technical paths for modeling equipment condition from the two dimensions of "physical mechanism" and "data characteristics," respectively, and have already been initially applied in some online monitoring systems.

[0003] However, the aforementioned existing technologies all suffer from unavoidable technical bottlenecks in practical applications. Mechanism-driven methods rely on complex multiphysics modeling and high-precision numerical solutions, resulting in high computational costs and long processing times. Furthermore, they are sensitive to boundary conditions and material parameters, leading to weak model generalization capabilities and difficulty in adapting to complex and variable operating conditions in the field. Simultaneously, fault equivalent modeling is difficult, failing to comprehensively cover various defect morphologies and limiting its engineering applicability. While data-driven methods offer advantages such as convenient modeling and fast inference speed, the lack of physical understanding of the equipment degradation process makes model training susceptible to factors like sample imbalance and noise interference. This results in low accuracy in identifying unseen fault types, and the diagnostic results exhibit "black box" characteristics, making it difficult to interpret the decision-making basis and impacting the trust and adoption of these methods by maintenance personnel. Moreover, due to the slow development and weak initial signals of converter transformer faults, traditional methods often struggle to capture early signs, leading to delayed warnings. Therefore, how to effectively integrate the physical evolution law of equipment fault development with multi-source sensing data to construct a fault early warning model that combines high accuracy, strong interpretability and engineering feasibility has become a key issue that needs to be addressed to improve the intelligent diagnosis level of converter transformers. Summary of the Invention

[0004] This application proposes a mathematical and physical fusion-driven method and system for early warning of power equipment faults. This application solves the technical problems of existing power equipment fault early warning methods, such as the lack of interpretability of data-driven models and the computational complexity and difficulty in engineering application of mechanism-driven models.

[0005] To solve the above-mentioned technical problems, the technical solution adopted in this application is as follows: In a first aspect, this application provides a method for early warning of power equipment faults driven by mathematical and physical fusion, including: Acquire multimodal monitoring data of the converter transformer under different operating conditions; The fault evolution law of key state quantities is extracted from the multimodal monitoring data. For partial discharge parameters that exhibit exponential changes, dissolved gas parameters that exhibit linear changes, and total hydrocarbon concentration parameters that conform to statistical distribution characteristics, corresponding mathematical fitting models are established respectively. Differential processing is performed on each type of fault evolution law to calculate the deviation between the actual sampled value and the corresponding fitted curve. The deviation is then weighted and correlated with the predicted probability output by the corresponding mathematical fitting model to form a law loss term that characterizes the degree of deviation from the mechanism. The law loss term is then weighted and fused with a cross-entropy loss term based on the real label to construct a comprehensive loss function. A neural network with an attention mechanism is used as the basic model. During the training process, the comprehensive loss function is used for backpropagation optimization, so that the basic model can adaptively learn and follow the physical evolution path of fault development while fitting data features. Finally, a fault early warning model for power transmission and transformation equipment is trained. The operating status of power transmission and transformation equipment is obtained, and fault warnings are given to the power transmission and transformation equipment based on the fault warning model of the power transmission and transformation equipment.

[0006] As a further improvement to this application, the monitoring data includes electrical signals, non-electrical parameters, and chemical parameters; The key state parameters include partial discharge pulse repetition rate, maximum discharge value, high-frequency signal amplitude, ultra-high-frequency signal amplitude, hydrogen concentration, total hydrocarbon concentration, oil temperature, and trace water content.

[0007] As a further improvement of this application, the partial discharge parameters are fitted using an exponential function, and the equipment is divided into three stages according to its operating status: normal, general defect, and severe defect. Corresponding exponential evolution models are established for each stage.

[0008] As a further improvement to this application, the concentration of dissolved gas in the oil is fitted with a linear function, and linear evolution models are constructed for three operating states: normal, general defects, and severe defects.

[0009] As a further improvement of this application, the total hydrocarbon concentration is statistically modeled using the Weibull distribution function, and Weibull distribution evolution models are constructed for three operating states: normal, general defect, and severe defect.

[0010] As a further improvement to this application, the step of performing differential processing on the evolution law of each type of fault and calculating the deviation between the actual sampled value and the corresponding fitted curve includes: The mean square error between the actual sampled value and the corresponding fitted value of the evolution law is calculated as an indicator to measure the degree to which the monitoring data deviates from the ideal evolution path.

[0011] As a further improvement to this application, the step of weightedly associating the deviation with the predicted probability output by the corresponding mathematical fitting model to form a regularity loss term characterizing the degree of deviation from the mechanism includes: The mean squared error of each state variable is compared with the predicted probability output by the model at the current time. Multiplying them together yields the weighted regularity loss term.

[0012] As a further improvement to this application, the comprehensive loss function consists of the sum of a cross-entropy loss term based on the true label and multiple weighted regularity loss terms, and its expression is:

[0013] In the formula, N is the total number of samples. , , These are the adjustment weights for the loss functions related to the evolution of the maximum discharge amplitude of partial discharge, the evolution of H2, and the distribution of total hydrocarbons; i=1, 2, 3, corresponding to the normal state, general defect state, and severe defect state of the converter transformer; j Here, K is the sample sequence number, and K is the total number of running status categories; a i For the first i The initial characteristic coefficients under each state; b j For the first j Baseline offset of each sample; c j Indicates the first j The evolutionary decay constant of each sample; y ij For the first i The true fault label of the j-th sample under each state; β i Let be the shape parameter of the Weibull distribution of total hydrocarbon concentration in the i-th state; p ij Q is the predicted probability output by the Softmax function of the early warning model; ij It is the first i The first state j The actual value of the maximum discharge amplitude of each sample x ij It's a timestamp; N i They represent the first i The number of state samples, H2ij It is the first i The first state j The actual H2 concentration of each sample; It is the first i The first state j The probability distribution of the actual total hydrocarbon values ​​for each sample; For the first i The scale parameter of the Weibull distribution of total hydrocarbon concentration under various conditions; x i Indicates the first i The timestamp of the state; During training, the weights between different regular loss patterns are dynamically adjusted based on the gradient changes of various regular loss patterns.

[0014] As a further improvement of this application, the neural network with an attention mechanism is used as the basic model, and the self-attention mechanism is used to capture the long-range dependency relationship and cross-modal correlation features between channels in multimodal time series data. The multimodal monitoring data, after normalization and alignment, is constructed into a multi-channel time series tensor, which serves as the input to the basic model. Each channel corresponds to a historical observation sequence of a state variable. The output of the basic model includes fault type classification labels for the converter transformer and deviation scores of the evolution patterns of various state variables, which are used to indicate whether the current operating state deviates from the normal development path and the corresponding severity.

[0015] As a further improvement of this application, the step of performing differential processing on each type of fault evolution law, calculating the deviation between the actual sampled value and the corresponding fitted curve, and weightedly associating the deviation with the predicted probability output by the corresponding mathematical fitting model to form a law loss term characterizing the degree of deviation from the mechanism; and weightedly fusing the law loss term with the cross-entropy loss term based on the true label to construct a comprehensive loss function; including: Obtain various evolutionary laws ； Evolutionary patterns include: constructing exponential difference, linear difference, statistical difference, and pattern loss function; Extract case data of various evolutionary patterns and features from the evolutionary patterns of various state variables, including feature transformation, feature regularization, and feature filtering; The feature transformation includes obtaining the query information, key information, and value information of the evolutionary feature, and performing linear transformation; multiplying the linear transformation result of the evolutionary feature query information with the linear transformation result of the evolutionary feature key information by matrix, and performing probability classification according to the actual case labels; multiplying the result of the probability classification with the linear transformation result of the evolutionary feature value information again; performing linear transformation on the multiplication result again, and superimposing it with the evolutionary pattern again. Feature regularization normalizes feature values ​​by calculating the mean and variance of the input features. Feature filtering performs a linear transformation on the input features and then performs numerical filtering. The result is then linearly transformed and further superimposed with the input feature information. The obtained features are fused, and the extracted features are further linearly transformed by calculating the mean to construct a comprehensive loss function.

[0016] Secondly, this application provides a power equipment fault early warning system driven by mathematical fusion, and the power equipment fault early warning method driven by mathematical fusion includes: The acquisition module is used to acquire multi-mode monitoring data of the converter transformer under different operating conditions; The model building module is used to extract the fault evolution law of key state quantities from the multimodal monitoring data, and to establish corresponding mathematical fitting models for partial discharge parameters that exhibit exponential changes, dissolved gas parameters that exhibit linear changes, and total hydrocarbon concentration parameters that conform to statistical distribution characteristics in the fault evolution law. The function construction module is used to perform differential processing on the evolution law of each type of fault, calculate the deviation between the actual sampled value and the corresponding fitted curve, and weight the deviation with the predicted probability output by the corresponding mathematical fitting model to form a law loss term that characterizes the degree of deviation of the mechanism; the law loss term is weighted and fused with the cross-entropy loss term based on the real label to construct a comprehensive loss function. The model training module is used to use a neural network with an attention mechanism as the base model. During the training process, the comprehensive loss function is used for backpropagation optimization, so that the base model can adaptively learn and follow the physical evolution path of fault development while fitting data features, and finally train a fault early warning model for power transmission and transformation equipment. The fault early warning module is used to obtain the operating status of power transmission and transformation equipment and to provide fault early warning for the power transmission and transformation equipment based on the fault early warning model of the power transmission and transformation equipment.

[0017] Thirdly, this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the mathematical fusion-driven power equipment fault early warning method.

[0018] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the mathematical fusion-driven power equipment fault early warning method.

[0019] Fifthly, this application provides a computer program product, the computer program product including computer instructions, the computer instructions instructing a computer to execute the mathematical fusion-driven power equipment fault early warning method.

[0020] Sixthly, this application provides a power equipment fault early warning device driven by mathematical fusion, including a data acquisition unit, an edge computing unit, and an alarm output unit; wherein, the data acquisition unit is used to acquire multi-source monitoring signals, the edge computing unit runs the method described, and the alarm output unit is used to display the fault type and risk level and trigger an audible and visual alarm.

[0021] The advantages of this application over the prior art are as follows: The mathematical fusion-driven power equipment fault early warning method provided in this application extracts the fault evolution laws of key state variables of converter transformers from historical operating data and fault cases. These state variables include partial discharge pulse repetition rate, maximum discharge value, high-frequency signal amplitude, ultra-high-frequency signal amplitude, dissolved gas concentration in oil, oil temperature, and trace water content. For the changing characteristics of different state variables, mathematical models of evolution laws are established for exponential, linear, and statistical distribution types, respectively. Differential processing is performed on each type of law to calculate the mean square error between the actual monitored value and the ideal evolution curve. The error term of the evolution law of each state variable is then compared with the error term of the classification task. A comprehensive loss function is constructed by combining the cross-entropy loss function with data fitting terms and mechanism constraint terms. A fault diagnosis model is built based on the Transformer network architecture. During training, the comprehensive loss function is used for backpropagation optimization, so that the model can learn sample features while following the physical laws of fault development. A lightweight computing framework for edge deployment is designed, which includes an evolution law extraction module, a feature transformation module, a regularization module, a feature filtering module, and a fusion post-processing module. The modular structure reduces redundant computation, improves inference efficiency, and supports fast, accurate, and interpretable fault early warning at the converter station site.

[0022] Furthermore, this application achieves a deep fusion of data-driven and mechanism-driven approaches by explicitly embedding the evolution law of multidimensional state variables during the fault development process of converter transformers into the training mechanism of the deep learning model. Compared with traditional pure data-driven methods, this scheme not only relies on labeled data for classification learning but also introduces physical laws reflecting the equipment degradation trend as constraints, significantly improving the model's sensitivity to early minor defects and diagnostic accuracy, while enhancing the interpretability of the decision-making process—maintenance personnel can trace the cause of the fault by analyzing which type of state variable deviates from its normal evolution path. Compared with traditional mechanism-driven methods, this application avoids complex multi-physics coupled simulations and high-dimensional numerical solutions, only requiring the fitting of empirical laws based on historical data and the integration of loss functions in a lightweight manner, greatly reducing computational overhead and possessing good engineering practicality. The designed edge early warning computing framework further optimizes the model structure and computing process, removes unnecessary complex operations, adapts to resource-constrained field equipment, and realizes efficient inference and real-time deployment of the mathematical-physical fusion model at the edge of the converter station. This technical solution can be widely applied to the condition assessment and intelligent operation and maintenance system of power grid main equipment, supporting the construction of an intelligent diagnostic system that integrates "end-edge-cloud" collaboration, and promoting the transformation and upgrading of the power system towards digitalization and intelligence. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the following description of the relevant technical solutions in the embodiments of this application or the prior art is provided with accompanying drawings. It should be understood that the accompanying drawings described below are only for the convenience of clearly illustrating some embodiments of the technical solutions in this application. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0024] Figure 1 Flowchart of a fault early warning model for power transmission and transformation equipment driven by data and mechanism fusion provided in the embodiments of this application; Figure 2 The evolution diagram of the partial discharge pulse repetition rate, maximum discharge quantity, high-frequency signal amplitude, and ultra-high-frequency signal amplitude of the converter transformer provided in the embodiments of this application; Figure 3 The following diagram illustrates the evolution of dissolved gas, oil temperature, and trace moisture in converter transformer oil, as provided in this embodiment of the application; wherein, (a) dissolved gas in oil; (b) oil temperature; and (c) trace moisture. Figure 4 The statistical distribution curve of total hydrocarbon concentration provided for the embodiments of this application; Figure 5 A flowchart of the mathematical fusion edge early warning device provided in the embodiments of this application; Figure 6A flowchart illustrating the evolution law extraction of the mathematical fusion edge early warning device provided in this application embodiment; Figure 7 A flowchart illustrating the feature extraction process of the mathematical fusion edge early warning device provided in this application embodiment; Figure 8 A flowchart illustrating the feature transformation of the mathematical fusion edge early warning device provided in this application embodiment; Figure 9 A flowchart illustrating the feature regularization process of the mathematical fusion edge early warning device provided in this application embodiment; Figure 10 A flowchart illustrating the feature filtering process of the mathematical fusion edge early warning device provided in this application embodiment. Detailed Implementation

[0025] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain this application, and should not be construed as limiting this application. The step numbers in the following embodiments are set only for ease of explanation, and there is no limitation on the order between the steps. The execution order of each step in the embodiments can be adaptively adjusted according to the understanding of those skilled in the art.

[0026] In the description of this application, unless otherwise expressly defined, terms such as "setup," "installation," and "connection" should be interpreted broadly, and those skilled in the art can reasonably determine the specific meaning of the above terms in this application in conjunction with the specific content of the technical solution.

[0027] Terminology Explanation: MSE stands for Mean-Square Error.

[0028] TOT stands for Total Hydrocarbons.

[0029] like Figure 1 As shown, this application provides a power equipment fault early warning method driven by mathematical fusion. It constructs a deep learning model based on multi-source sensing data to classify and identify the operating status and fault type of power transmission and transformation equipment. It includes the following steps S1 to S5, and each step is described in detail below.

[0030] S1, acquire multi-mode monitoring data of the converter transformer under different operating conditions, including electrical signals, non-electrical parameters and chemical parameters; During training, a converter transformer is used as the target, but the power transmission and transformation equipment targeted in this embodiment includes, but is not limited to, ultra-high voltage transformers, converter transformers, converter valves, and other equipment.

[0031] In the above scheme, key state quantities include partial discharge pulse repetition rate, maximum discharge quantity, high-frequency signal amplitude, ultra-high-frequency signal amplitude, hydrogen (H2) concentration, total hydrocarbon concentration, oil temperature, and trace water content. The specific physical parameters involved in modeling the evolutionary patterns encompass multi-dimensional information from electrical, chemical, and thermodynamic perspectives, forming the data foundation for mathematical-physical fusion diagnostics and ensuring the comprehensiveness and representativeness of the model input.

[0032] S2, extract the fault evolution law of key state variables, and establish corresponding mathematical fitting models for partial discharge parameters that exhibit exponential changes, dissolved gas parameters in oil that exhibit linear changes, and total hydrocarbon concentration parameters that conform to statistical distribution characteristics. In the above scheme, exponential functions are used to fit partial discharge parameters, and the equipment operating status is divided into three stages: normal, general defect, and severe defect, with corresponding exponential evolution models established for each stage. To address the typical characteristic of partial discharge accelerating over time, exponential functions are chosen to accurately describe its nonlinear growth trend. Stage-based modeling can capture the dynamic changes under different degrees of degradation, improving the precision of mechanistic constraints.

[0033] Linear function fitting was used to fit the concentration of dissolved gases such as hydrogen (H2) in oil, and linear evolution curves were constructed for three operating states: normal, general defects, and severe defects. For gas components exhibiting approximately linear growth in the early stages of a fault, the linear model is simple and effective. State-specific modeling reflects the differences in development trends under different health levels, enhancing the ability to identify regular loss terms.

[0034] The total hydrocarbon concentration was statistically modeled using the Weibull distribution function, and shape and scale parameters were extracted to characterize its probabilistic evolution at different stages of fault development. The total hydrocarbon concentration exhibits significant statistical distribution characteristics, and the Weibull distribution can well fit its cumulative probability change process. Parametric modeling enables a probabilistic description of the overall degradation trend, suitable for early warning analysis in imbalanced sample scenarios.

[0035] Specifically, this embodiment analyzes cases of defects or faults in converter transformers; the process of the fault early warning model for power transmission and transformation equipment driven by data and mechanism fusion is as follows: Figure 1 As shown; specifically including: (1) Obtaining the Fault Evolution Law of Converter Transformer Analyzing the fault development process of converter transformers from defect or fault cases, we can see the progression from "minor defect to serious defect to minor fault to serious fault to near breakdown." Figure 2 and Figure 3As shown, the evolution law of key state variables can be grasped through data fitting methods. Figure 2 The evolution of partial discharge pulse repetition rate, maximum discharge quantity, high-frequency signal amplitude, and ultra-high-frequency signal amplitude of converter transformers; Figure 3 The evolution of dissolved gas, oil temperature, and trace water in converter transformer oil is shown, where (a) dissolved gas in oil, (b) oil temperature, and (c) trace water.

[0036] (2) Differential processing of fault evolution law of converter transformer 1) Differential processing of exponential change-type evolutionary patterns For state variables such as partial discharge pulse repetition rate, maximum discharge quantity, high-frequency signal amplitude, and ultra-high-frequency signal amplitude, which exhibit exponential changes, the differential processing process is as follows: The variation curves are obtained by fitting the key state variables. Taking the maximum discharge quantity as an example, the fitting curve is shown in equation (1): (1) Where a, b, and c represent the parameter values ​​of the fitting curve of the maximum partial discharge amplitude of the transformer, x represents time, and Q represents the maximum partial discharge amplitude at time x. As can be seen from the formula, the exponential function of the maximum discharge amplitude exhibits an exponential variation law.

[0037] Based on the development process, partial discharge is divided into three states: normal, general defect, and severe defect. Referring to the exponential change form of equation (1), the maximum amplitude of partial discharge corresponding to the three states is fitted, and the fitting curves corresponding to the three states are obtained as follows: (2) Where i=1 / 2 / 3: corresponds to the normal state / general defect state / severe defect state of the converter transformer; a i b i and c i Let x represent the parameter values ​​of the fitted curve of the maximum discharge amplitude of partial discharge in the i-th state of the transformer. i Q represents the timestamp of the i-th state. i It is the xth i Fitted value of the maximum discharge amplitude of partial discharge over time.

[0038] The mean-square error (MSE) between the actual and fitted values ​​of the maximum partial discharge amplitude for the three state samples (normal, general defect, and severe defect) is calculated as follows: (3) in, The mean squared error of the i-th class of state samples, N i(i=1,2,3) represent the number of samples in the three states: normal, general defect, and severe defect, respectively, and Q ij It is the actual value of the maximum discharge amplitude of the j-th sample in the i-th state, x ij It is Q ij The corresponding timestamp.

[0039] By correlating the mean square error between the actual value and the fitted value calculated in equation (3) for each state with the predicted probability output by the Softmax function of the early warning model, a loss function is obtained after differential processing of the evolution law of the maximum discharge amplitude of partial discharge. for: (4) in, p ij It is the predicted probability output by the Softmax function of the early warning model.

[0040] 2) Differential processing of linear change-type evolution laws For state variables such as dissolved gas in oil and oil temperature that exhibit linear changes, the differential processing procedure is as follows: The change curves are obtained by fitting the key state variables. Taking H2 concentration as an example, the fitting curve is shown in equation (5): (5) Where a: linearly changing state variable, b: linearly changing state variable, x represents time, and H2 represents the H2 concentration at time x. As can be seen from the formula, the H2 concentration exhibits a linear change pattern. Based on the development process, dissolved gases in oil are divided into three states: normal, general defect, and severe defect. Referring to the linear change form of equation (5), the H2 (hydrogen) corresponding to the three states of normal, general defect, and severe defect are fitted respectively, and the fitting curves for the three states are as follows: (6) Among them, a i b i and c i Let x represent the parameter values ​​of the H2 concentration fitting curve for the i-th state of the transformer. i H represents the timestamp of the i-th state. 2i It is the xth i Fitted H2 concentration values ​​over time.

[0041] Calculate the root mean square error between the actual and fitted values ​​of H2 concentration for samples in three states: normal, moderately defective, and severely defective. for: (7) Where, N i (i =1,2,3) represent the number of samples in the three states of normal, general defect, and severe defect, respectively, and H is the number of samples in the three states of normal, general defect, and severe defect. 2ij It is the i-th state. j The actual H2 concentration of each sample. x ij It is H 2ij The corresponding timestamp.

[0042] By correlating the mean square error between the actual value and the fitted value for each state calculated in equation (7) with the predicted probability output by the Softmax function of the early warning model, the loss function after differential processing of the H2 concentration change pattern is obtained as follows: (8) 3) Differential processing of statistical distribution evolution patterns: For state variables such as dissolved gases in oil that exhibit statistical distribution patterns, the difference processing procedure is as follows: Statistical analysis was performed on the key state variables to obtain statistical distribution curves. Taking total hydrocarbon concentration as an example, the fitted curve is as follows: Figure 4 As shown. According to Figure 4 As can be seen from the fitted curve, the total hydrocarbon concentration generally exhibits a Weibull distribution.

[0043] The Weibull distribution curve was obtained by fitting the total hydrocarbons, and the fitted curve is shown in equation (9): (9) Where x represents time, This represents the probability of the total hydrocarbon concentration at time x. 、 These represent the current status parameter and scale parameter of the total hydrocarbons in the i-th state of the transformer, respectively.

[0044] Based on the development process, dissolved gases in oil are divided into three states: normal, moderately defective, and severely defective. The total hydrocarbons in each of these three states are fitted with their respective fitting curves, resulting in the following: (10) in, i 、 Let x represent the current state parameter and scale parameter of the total hydrocarbons in the i-th state of the transformer, respectively. i The timestamp representing the i-th state. It is the first x i The total probability of hydrocarbons over time.

[0045] Furthermore, the root mean square error between the actual distribution and the fitted distribution of total hydrocarbon values ​​for the three state samples (normal, general defect, and severe defect) is calculated as follows: (11) Where, N i (i=1,2,3) represent the number of samples in the three states: normal, general defect, and severe defect, respectively. It represents the probability distribution of the total hydrocarbon actual value for the j-th sample in the i-th state. x i yes The corresponding timestamp.

[0046] Furthermore, the mean square error between the actual value and the fitted value calculated in equation (11) for each state is correlated with the predicted probability output by the Softmax function of the early warning model to obtain the loss function after differential processing of the total hydrocarbon distribution probability. for: (12) S3. For each type of evolutionary pattern, differential processing is performed to calculate the deviation between the actual sampled value and the corresponding fitted curve. This deviation is then weighted and correlated with the predicted probability output by the model to form a pattern loss term that characterizes the degree of deviation from the mechanism. A comprehensive loss function is constructed to weight and fuse the pattern loss term with the cross-entropy loss term based on the true label. In the above scheme, the mean square error (MSE) between the actual sampled value and the corresponding fitted value of the evolutionary law is calculated as an indicator to measure the degree to which the monitoring data deviates from the ideal evolutionary path. MSE provides a mathematical tool for quantifying the "degree of deviation" and is a key bridge connecting the sensing data and physical laws, providing a calculable and optimizable objective function for the subsequent construction of the law loss term.

[0047] Furthermore, the mean square error of each state variable is compared with the predicted probability output by the model at the current time. Multiplying these results yields a weighted regularity loss term, which imposes a stronger penalty on predictions with high confidence but deviating from the expected pattern. Introducing prediction probability as a weighting factor enables a joint evaluation of "classification confidence" and "regularity conformity," preventing the model from outputting results that violate physical common sense with high confidence, and significantly improving the reliability of inference.

[0048] Furthermore, for each type of fault evolution pattern, differential processing is performed to calculate the deviation between the actual sampled value and the corresponding fitted curve. This deviation is then weighted and correlated with the predicted probability output by the corresponding mathematical fitting model to form a pattern loss term characterizing the degree of deviation from the mechanism. This pattern loss term is then weighted and fused with a cross-entropy loss term based on the true labels to construct a comprehensive loss function, including: Obtain various evolutionary laws ；Evolutionary patterns include: constructing exponential difference, linear difference, statistical difference, and pattern loss function; Extract case data of various evolutionary patterns and features from the evolutionary patterns of various state variables, including feature transformation, feature regularization, and feature filtering; The feature transformation includes obtaining the query information, key information, and value information of the evolutionary feature, and performing linear transformation; multiplying the linear transformation result of the evolutionary feature query information with the linear transformation result of the evolutionary feature key information by matrix, and performing probability classification according to the actual case labels; multiplying the result of the probability classification with the linear transformation result of the evolutionary feature value information again; performing linear transformation on the multiplication result again, and superimposing it with the evolutionary pattern again. Feature regularization normalizes feature values ​​by calculating the mean and variance of the input features. Feature filtering performs a linear transformation on the input features and then performs numerical filtering. The result is then linearly transformed and further superimposed with the input feature information. The obtained features are fused, and the extracted features are further linearly transformed by calculating the mean to construct a comprehensive loss function.

[0049] In the process of constructing the integrated mathematical loss function, a cross-entropy loss function for converter transformer fault samples is constructed. : (13) in, y ij These are genuine fault labels. p ij This is the predicted probability output by the Softmax function of the diagnostic model, where N is the number of samples and K is the number of categories.

[0050] Considering the loss functions of various evolutionary laws in equations (4), (8) and (12) and combining them with the data sample loss in equation (13), a comprehensive loss function is constructed. for: (14) In the formula, N is the total number of samples. , , These are the adjustment weights for the loss functions related to the evolution of the maximum discharge amplitude of partial discharge, the evolution of H2, and the distribution of total hydrocarbons; i=1, 2, 3, corresponding to the normal state, general defect state, and severe defect state of the converter transformer; j Here, K is the sample sequence number, and K is the total number of running status categories; a i For the first i The initial characteristic coefficients under each state; bj For the first j Baseline offset of each sample; c j Indicates the first j The evolutionary decay constant of each sample; y ij For the first i The true fault label of the j-th sample under each state; β i Let be the shape parameter of the Weibull distribution of total hydrocarbon concentration in the i-th state; p ij Q is the predicted probability output by the Softmax function of the early warning model; ij It is the first i The first state j The actual value of the maximum discharge amplitude of each sample x ij It's a timestamp; N i They represent the first i The number of state samples, H 2ij It is the first i The first state j The actual H2 concentration of each sample; It is the first i The first state j The probability distribution of the actual total hydrocarbon values ​​for each sample; For the first i The scale parameter of the Weibull distribution of total hydrocarbon concentration under various conditions; x i Indicates the first i The timestamp of the state; During training, the weights of different loss patterns are dynamically adjusted to optimize the state variables and their evolution patterns.

[0051] This loss structure achieves an explicit fusion of data fitting and mechanistic constraints, which is the core mechanism for realizing "mathematical-mechanistic fusion training" in this application and directly determines the learning orientation and generalization performance of the model.

[0052] S4 uses a neural network with an attention mechanism as the basic model. During training, a comprehensive loss function is used for backpropagation optimization, which enables the model to adaptively learn and follow the physical evolution path of fault development while fitting data features. In the above scheme, a mathematical fusion fault early warning model classification model for converter transformers is constructed based on the Transformer basic model. In the back propagation process of model training, the comprehensive loss function shown in Equation (14) is used to replace the traditional cross-entropy loss function to realize model training with evolution law constraints.

[0053] In this model training process, the parameters are dynamically adjusted based on the gradient changes of various rule losses to balance the relationship between data fitting accuracy and the degree of adherence to physical laws. Dynamic weight adjustment avoids the optimization imbalance problem caused by fixed ratios, allowing the model to focus on data learning in the early stages of training and strengthen rule constraints in the later stages, thereby improving convergence stability and final performance.

[0054] A neural network incorporating an attention mechanism forms the basic model architecture, utilizing self-attention to capture long-range dependencies and cross-modal correlation features among channels in multimodal time-series data. The Transformer possesses powerful sequence modeling capabilities, effectively handling the temporal evolution patterns of various heterogeneous signals such as partial discharge, oil temperature, and gas, providing a high-performance feature extraction platform for fusion mechanism constraints. Multimodal monitoring data, after normalization and alignment, is constructed into a multi-channel time-series tensor, serving as model input. Each channel corresponds to a historical observation sequence of a state variable. Standardizing the input data format to ensure that signals with different dimensions and sampling frequencies can be processed within a unified framework is a prerequisite for achieving deep fusion of multi-source information. The output results include fault type classification labels for the converter transformer and deviation scores for the evolution patterns of various state variables, indicating whether the current operating state deviates from the normal development path and its severity. Therefore, the training process in this application not only provides "black-box" classification results but also outputs interpretable deviation evaluations, helping maintenance personnel understand the judgment basis and supporting early fault intervention decisions.

[0055] S5 acquires the operating status of power transmission and transformation equipment and provides fault warnings based on the equipment fault warning model. This enables high-accuracy and interpretable early warning diagnosis of converter transformer faults.

[0056] This application proposes a fault early warning model and device for power transmission and transformation equipment driven by evolution law and perception data fusion. (1) The evolution law of converter transformer fault development is embedded into the data-driven model to guide the model to adaptively and directionally train, and the advantages of mechanism-driven and data-driven paradigms are combined to form a mathematical-physical fusion diagnostic paradigm for converter transformers, thereby improving the diagnostic accuracy and interpretability of fault early warning for power transmission and transformation equipment. (2) A converter transformer edge early warning computing architecture oriented towards the mathematical-physical fusion paradigm is designed to reduce the necessary computation of the data and mechanism fusion model, accelerate inference computation, and support the rapid deployment and application of the mathematical-physical fusion diagnostic model for converter transformers at the edge.

[0057] The method in this application is a fault early warning model for power transmission and transformation equipment driven by the fusion of data and mechanism. It solves the problems of large simulation calculation, difficulty in fault equivalence and low fitting accuracy in the mechanism-driven early warning paradigm, as well as the problems of low accuracy, lack of mechanism guidance and lack of interpretability in the data-driven early warning paradigm. It enables interpretable and accurate identification of faults in the main equipment of the power grid.

[0058] The second objective of this application is to provide a power equipment fault early warning system driven by mathematical fusion. Based on the aforementioned power equipment fault early warning method driven by mathematical fusion, the system includes: The acquisition module is used to acquire multi-mode monitoring data of the converter transformer under different operating conditions; The model building module is used to extract the fault evolution law of key state variables from multimodal monitoring data. For partial discharge parameters that exhibit exponential changes, dissolved gas parameters that exhibit linear changes, and total hydrocarbon concentration parameters that conform to statistical distribution characteristics, corresponding mathematical fitting models are established respectively. The function construction module is used to perform differential processing on the evolution law of each type of fault, calculate the deviation between the actual sampled value and the corresponding fitted curve, and weight the deviation with the predicted probability output by the corresponding mathematical fitting model to form a law loss term that characterizes the degree of deviation of the mechanism; the law loss term is weighted and fused with the cross-entropy loss term based on the real label to construct a comprehensive loss function. The model training module uses a neural network with an attention mechanism as the base model. During the training process, a comprehensive loss function is used for backpropagation optimization, which enables the base model to adaptively learn and follow the physical evolution path of fault development while fitting data features, and finally trains a fault early warning model for power transmission and transformation equipment. The fault early warning module is used to obtain the operating status of power transmission and transformation equipment and to provide fault early warning for the power transmission and transformation equipment based on the fault early warning model of the power transmission and transformation equipment.

[0059] This application provides a mathematical-physical fusion-driven computational framework for converter transformer edge early warning of power equipment faults, mainly described below: (1) Mathematical-physical fusion edge early warning computing framework: The overall computing process of the mathematical-physical fusion edge early warning computing framework is as follows: Figure 5 As shown, the process includes three steps: evolutionary pattern extraction, feature extraction, and feature post-processing. The evolutionary pattern extraction step is used to obtain various evolutionary patterns; feature extraction is used to extract features from the case data and the evolutionary patterns of various state variables; and feature post-processing is used to fuse the obtained features.

[0060] (2) Evolutionary Law Extraction Process of Mathematical-Physical Fusion Edge Early Warning Computation Framework: The process of extracting the evolutionary law of the mathematical-physical fusion edge early warning device is as follows: Figure 6 As shown, it includes the construction processes of exponential difference, linear difference, statistical difference, and the regularity loss function.

[0061] (3) Feature extraction process of the mathematical-physical fusion edge early warning computing framework, and the feature extraction process of the mathematical-physical fusion edge early warning device are as follows: Figure 7As shown, it includes three parts: feature transformation, feature regularization, and feature filtering. Each part is explained below: 1) The process of feature transformation is as follows: Figure 8 As shown, for a certain evolutionary law, firstly, the query information, key information, and value information of the evolutionary law feature are obtained and linearly transformed; then, the linear transformation result of the evolutionary law feature query information is matrix-multiplied with the linear transformation result of the evolutionary law feature key information, and probabilistic classification is performed according to the actual case labels. The result after probability classification is multiplied again with the linear transformation result of the evolutionary law feature value information; furthermore, the multiplication result is linearly transformed again and superimposed with the evolutionary law.

[0062] 2) Feature regularization process as follows Figure 9 As shown, by calculating the mean and variance of the input features, the feature values ​​are normalized to avoid overfitting and improve generalization ability.

[0063] 3) Feature filtering process as follows Figure 10 As shown, firstly, the input features are linearly transformed. Then, numerical filtering is performed, with the principle that feature values ​​greater than 0 remain unchanged and those less than 0 are assigned a value of 0. Finally, the result is linearly transformed and then superimposed with the input feature information to complete the feature filtering.

[0064] 4) Feature post-processing process of the mathematical-physical fusion edge early warning computing framework: Feature post-processing performs a further linear transformation on the extracted features by calculating the mean.

[0065] The edge early warning computing architecture for converter transformers, which is oriented towards the mathematical-physical fusion paradigm, solves the problem that traditional edge computing devices cannot adapt to the mathematical-physical fusion computing paradigm and cannot perform high-speed inference at the edge. It supports the rapid deployment and efficient inference of the fault early warning model of power transmission and transformation equipment driven by data and mechanism fusion at the edge of the station.

[0066] A third objective of this application is to provide an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the aforementioned mathematical fusion-driven power equipment fault early warning method. It also includes a communication interface and a bus. The mathematical fusion-driven power equipment fault early warning method specifically includes: Acquire multimodal monitoring data of the converter transformer under different operating conditions; The fault evolution law of key state quantities is extracted from the multimodal monitoring data. For partial discharge parameters that exhibit exponential changes, dissolved gas parameters that exhibit linear changes, and total hydrocarbon concentration parameters that conform to statistical distribution characteristics, corresponding mathematical fitting models are established respectively. Differential processing is performed on each type of fault evolution law to calculate the deviation between the actual sampled value and the corresponding fitted curve. The deviation is then weighted and correlated with the predicted probability output by the corresponding mathematical fitting model to form a law loss term that characterizes the degree of deviation from the mechanism. The law loss term is then weighted and fused with a cross-entropy loss term based on the real label to construct a comprehensive loss function. A neural network with an attention mechanism is used as the basic model. During the training process, the comprehensive loss function is used for backpropagation optimization, so that the basic model can adaptively learn and follow the physical evolution path of fault development while fitting data features. Finally, a fault early warning model for power transmission and transformation equipment is trained. The operating status of power transmission and transformation equipment is obtained, and fault warnings are given to the power transmission and transformation equipment based on the fault warning model of the power transmission and transformation equipment.

[0067] A fourth objective of this application is to provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned mathematical fusion-driven power equipment fault early warning method. The mathematical fusion-driven power equipment fault early warning method specifically includes: Acquire multimodal monitoring data of the converter transformer under different operating conditions; The fault evolution law of key state quantities is extracted from the multimodal monitoring data. For partial discharge parameters that exhibit exponential changes, dissolved gas parameters that exhibit linear changes, and total hydrocarbon concentration parameters that conform to statistical distribution characteristics, corresponding mathematical fitting models are established respectively. Differential processing is performed on each type of fault evolution law to calculate the deviation between the actual sampled value and the corresponding fitted curve. The deviation is then weighted and correlated with the predicted probability output by the corresponding mathematical fitting model to form a law loss term that characterizes the degree of deviation from the mechanism. The law loss term is then weighted and fused with a cross-entropy loss term based on the real label to construct a comprehensive loss function. A neural network with an attention mechanism is used as the basic model. During the training process, the comprehensive loss function is used for backpropagation optimization, so that the basic model can adaptively learn and follow the physical evolution path of fault development while fitting data features. Finally, a fault early warning model for power transmission and transformation equipment is trained. The operating status of power transmission and transformation equipment is obtained, and fault warnings are given to the power transmission and transformation equipment based on the fault warning model of the power transmission and transformation equipment.

[0068] A fifth objective of this application is to provide a computer program product, which includes computer instructions that instruct a computer to execute the aforementioned mathematical fusion-driven power equipment fault early warning method. The mathematical fusion-driven power equipment fault early warning method specifically includes: Acquire multimodal monitoring data of the converter transformer under different operating conditions; The fault evolution law of key state quantities is extracted from the multimodal monitoring data. For partial discharge parameters that exhibit exponential changes, dissolved gas parameters that exhibit linear changes, and total hydrocarbon concentration parameters that conform to statistical distribution characteristics, corresponding mathematical fitting models are established respectively. Differential processing is performed on each type of fault evolution law to calculate the deviation between the actual sampled value and the corresponding fitted curve. The deviation is then weighted and correlated with the predicted probability output by the corresponding mathematical fitting model to form a law loss term that characterizes the degree of deviation from the mechanism. The law loss term is then weighted and fused with a cross-entropy loss term based on the real label to construct a comprehensive loss function. A neural network with an attention mechanism is used as the basic model. During the training process, the comprehensive loss function is used for backpropagation optimization, so that the basic model can adaptively learn and follow the physical evolution path of fault development while fitting data features. Finally, a fault early warning model for power transmission and transformation equipment is trained. The operating status of power transmission and transformation equipment is obtained, and fault warnings are given to the power transmission and transformation equipment based on the fault warning model of the power transmission and transformation equipment.

[0069] A sixth objective of this application is to provide a power equipment fault early warning device driven by mathematical fusion, comprising a data acquisition unit, an edge computing unit, and an alarm output unit. The data acquisition unit acquires multi-source monitoring signals, the edge computing unit executes the aforementioned power equipment fault early warning method driven by mathematical fusion, and the alarm output unit displays the fault type and risk level and triggers an audible and visual alarm. The power equipment fault early warning method driven by mathematical fusion specifically includes: Acquire multimodal monitoring data of the converter transformer under different operating conditions; The fault evolution law of key state quantities is extracted from the multimodal monitoring data. For partial discharge parameters that exhibit exponential changes, dissolved gas parameters that exhibit linear changes, and total hydrocarbon concentration parameters that conform to statistical distribution characteristics, corresponding mathematical fitting models are established respectively. Differential processing is performed on each type of fault evolution law to calculate the deviation between the actual sampled value and the corresponding fitted curve. The deviation is then weighted and correlated with the predicted probability output by the corresponding mathematical fitting model to form a law loss term that characterizes the degree of deviation from the mechanism. The law loss term is then weighted and fused with a cross-entropy loss term based on the real label to construct a comprehensive loss function. A neural network with an attention mechanism is used as the basic model. During the training process, the comprehensive loss function is used for backpropagation optimization, so that the basic model can adaptively learn and follow the physical evolution path of fault development while fitting data features. Finally, a fault early warning model for power transmission and transformation equipment is trained. The operating status of power transmission and transformation equipment is obtained, and fault warnings are given to the power transmission and transformation equipment based on the fault warning model of the power transmission and transformation equipment.

[0070] This application develops a data and mechanism-driven fault early warning model, computing architecture, and device for power transmission and transformation equipment. This enhances the real-time analysis capability of substation / converter station equipment operation status, timely predicts equipment fault development trends, and issues alarms to remind maintenance personnel to troubleshoot, further improving equipment reliability, extending equipment lifespan, and reducing equipment depreciation and downtime replacement time. This application can be integrated into substation / converter station digital platforms, undertaking the important function of real-time intelligent analysis of complex multimodal data at the edge, improving equipment operational reliability. This application supports the construction of an edge-cloud collaborative intelligent inspection system integrating "end-side modal fusion perception, edge mathematical fusion computing, and cloud-side human-machine fusion decision-making," improving the level of unmanned and intelligent equipment operation and maintenance management, and achieving reliable prediction of equipment operation status. Ultimately, by strengthening the application research and demonstration of the data and mechanism fusion computing paradigm in equipment operation and maintenance, it explores new applications of online updated heterogeneous fusion edge intelligent computing, supporting the digital transformation and upgrading of the power grid.

[0071] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0072] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0073] This application may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application may take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, readable storage media, optical storage, etc.) containing computer-usable program code.

[0074] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0075] Obviously, the described embodiments are only some, not all, of the embodiments in this application. All other embodiments obtained by those skilled in the art based on the embodiments in this application without inventive effort should fall within the scope of protection of this application.

[0076] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application and not to limit them. Although this application has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of this application. Any modifications or equivalent substitutions that do not depart from the spirit and scope of this application should be covered within the protection scope of the claims of this application.

Claims

1. A mathematical fusion driven power equipment fault early warning method, characterized in that, The method comprises the following steps: acquiring multi-modal monitoring data of a converter transformer under different operating states; extracting fault evolution rules of key state quantities from the multi-modal monitoring data, and establishing corresponding mathematical fitting models for partial discharge parameters showing exponential changes, oil dissolved gas parameters showing linear changes, and total hydrocarbon concentration parameters conforming to statistical distribution characteristics in the fault evolution rules; performing differential processing on each type of fault evolution rule, calculating deviations between actual sampling values and corresponding fitting curves, and weighting and associating the deviations with prediction probabilities output by the corresponding mathematical fitting models to form a rule loss term representing the degree of mechanism deviation; weighting and fusing the rule loss term and a cross-entropy loss term based on a true label to construct a comprehensive loss function; using a neural network containing an attention mechanism as a basic model, using the comprehensive loss function for back propagation optimization during the training process, so that the basic model fits the data characteristics while adaptively learning and following the physical evolution path of fault development, and finally training a power transmission and transformation equipment fault early warning model; acquiring an operating state of the power transmission and transformation equipment, and performing fault early warning on the power transmission and transformation equipment based on the power transmission and transformation equipment fault early warning model.

2. The mathematically fused drive-based power equipment failure warning method of claim 1, wherein, The monitoring data includes electrical quantity signals, non-electrical quantity parameters, and chemical parameters. The key state quantities include partial discharge pulse repetition rate, discharge quantity maximum value, high-frequency signal amplitude, ultra-high frequency signal amplitude, hydrogen concentration, total hydrocarbon concentration, oil temperature, and micro-water content.

3. The mathematically fused drive-based power equipment failure warning method of claim 1, wherein, Exponential functions are used to fit the partial discharge parameters, and the device operating states are divided into three stages of normal, general defect, and severe defect, and corresponding exponential evolution models are established.

4. The mathematically fused drive-based power equipment failure warning method of claim 1, wherein, Linear functions are used to fit the oil dissolved gas concentrations, and linear evolution models are constructed for the three operating states of normal, general defect, and severe defect.

5. The mathematically fused drive-based power equipment failure warning method of claim 1, wherein, Weibull distribution functions are used to statistically model the total hydrocarbon concentrations, and Weibull distribution evolution models are constructed for the three operating states of normal, general defect, and severe defect.

6. The mathematically fused drive-based power equipment failure warning method of claim 1, wherein, The differential processing of each type of fault evolution rule to calculate the deviations between the actual sampling values and the corresponding fitting curves comprises: calculating the mean square error between the actual sampling values and the fitting values of the corresponding evolution rules as an index for measuring the degree of deviation of the monitoring data from the ideal evolution path.

7. The mathematically fused drive-based power equipment failure warning method of claim 1, wherein, The weighting and association of the deviations with the prediction probabilities output by the corresponding mathematical fitting models to form the rule loss term representing the degree of mechanism deviation comprises: The mean square error of each state quantity is multiplied by the prediction probability of the model output at the current time to obtain the weighted regular loss term.

8. The mathematically fused drive-based power equipment failure warning method of claim 1, wherein, The comprehensive loss function is composed of a cross-entropy loss term based on a true label and a plurality of weighted regular loss terms, and the expression is: In the formula, N is the total number of samples, , , are the adjustment weights of the partial discharge maximum discharge amplitude evolution law loss, H2 evolution law loss and total hydrocarbon distribution law loss function respectively; i=1, 2, 3, corresponding to the normal state, general defect state and serious defect state of the converter transformer; j is the sample serial number, and K is the total number of running state categories; a i is the initial characteristic coefficient under the first i state; i > b j is the baseline offset for the i-th sample; j c j is the evolution decay constant for the i-th sample; j y ij is the true fault label for the i-th sample in the i-th state; i j β i is the shape parameter of the Weibull distribution of the total hydrocarbon concentration in the i-th state; p ij is the prediction probability of the early warning model Softmax function output; Q ij is the actual value of the maximum discharge amplitude of the i-th sample in the i-th state, i j x ij is the time stamp; N​​​​​​ i They represent the first i The number of state samples, H 2ij It is the first i The first state j The actual H2 concentration of each sample; It is the first i The first state j The probability distribution of the actual total hydrocarbon values ​​for each sample; For the first i The scale parameter of the Weibull distribution of total hydrocarbon concentration under various conditions; x i Indicates the first i The timestamp of the state; During the training process, the weights between different rule losses are dynamically adjusted according to the gradient changes of each type of rule loss.

9. The mathematically fused drive-based power equipment failure warning method of claim 1, wherein, The neural network containing an attention mechanism is used as the basic model, and the self-attention mechanism is used to capture the long-range dependency relationships and cross-modal correlation characteristics between channels in the multi-modal time series data; After normalization and alignment processing, the multi-modal monitoring data are constructed into a multi-channel time series tensor as the input of the basic model, and each channel corresponds to a historical observation sequence of a state quantity; the output of the basic model includes fault type classification labels of the converter transformer and deviation scores of the evolution rules of each state quantity, which are used to indicate whether the current operating state deviates from the normal development path and the corresponding severity.

10. The mathematically fused drive-based power equipment failure warning method of claim 1, wherein, The deviation between the actual sampling value and the corresponding fitting curve is calculated by differentiating each type of fault evolution law, and the deviation is weighted and associated with the prediction probability output by the corresponding mathematical fitting model to form a law loss term representing the degree of mechanism deviation; the law loss term and the cross-entropy loss term based on the real label are weighted and fused to construct a comprehensive loss function; comprising: Obtain each type of evolution law; the evolution law includes: constructing an exponential law difference, a linear law difference, a statistical law difference, and a law loss function; Extracting case data of each type of evolution law and features in each type of state quantity evolution law, including feature transformation, feature regularization, and feature filtering; The feature transformation includes obtaining the query information, key information and value information of the evolution law feature, and performing linear transformation; the linear transformation result of the evolution law feature query information is multiplied by the linear transformation result of the evolution law feature key information, and the result after probability classification according to the actual case label is multiplied again with the linear transformation result of the evolution law feature value information; the multiplication result is linearly transformed again and stacked with the evolution law again; The feature regularization is to normalize the feature value by calculating the mean and variance of the input feature; The feature filtering performs linear transformation on the input feature and performs numerical filtering, and further stacks the input feature information after linear transformation of the obtained result; Fuse the obtained features, further linearly transform the extracted features by calculating the mean, and construct a comprehensive loss function.

11. A mathematical fusion driven power equipment fault early warning system, characterized in that, Comprising: The acquisition module is configured to acquire multi-modal monitoring data of the converter transformer under different operating states; The model establishing module is configured to extract a fault evolution law of a key state quantity from the multi-modal monitoring data, and establish a corresponding mathematical fitting model for a partial discharge parameter showing exponential change, an oil dissolved gas parameter showing linear change, and a total hydrocarbon concentration parameter conforming to a statistical distribution characteristic in the fault evolution law; The function constructing module is configured to differentiate each type of fault evolution law, calculate the deviation between the actual sampling value and the corresponding fitting curve, and weight and associate the deviation with the prediction probability output by the corresponding mathematical fitting model to form a law loss term representing the degree of mechanism deviation; the law loss term and the cross-entropy loss term based on the real label are weighted and fused to construct a comprehensive loss function; The model training module is configured to use a neural network containing an attention mechanism as a basic model, use the comprehensive loss function for back propagation optimization during training, make the basic model fit the data features while adaptively learning and follow the physical evolution path of fault development, and finally train a power transmission and transformation equipment fault warning model; The fault warning module is configured to acquire the operating state of the power transmission and transformation equipment, and perform fault warning on the power transmission and transformation equipment based on the power transmission and transformation equipment fault warning model.

12. A mathematical fusion driven power equipment fault early warning device, characterized in that, It includes a data acquisition unit, an edge computing unit, and an alarm output unit; wherein, the data acquisition unit is used to acquire multi-source monitoring signals, the edge computing unit runs the power equipment fault early warning method driven by mathematical fusion as described in any one of claims 1 to 10, and the alarm output unit is used to display the fault type and risk level and trigger an audible and visual alarm.

13. An electronic device, comprising: The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the power equipment fault early warning method driven by mathematical fusion as described in any one of claims 1 to 10.

14. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the power equipment fault early warning method driven by mathematical fusion as described in any one of claims 1 to 10.

15. A computer program product comprising computer instructions, characterized in that, The computer instructions instruct the computer to execute the power equipment fault early warning method driven by mathematical fusion as described in any one of claims 1 to 10.

Images