CN-121980414-A - AmRMR-based transformer self-adaptive fault diagnosis method and system

CN121980414ACN 121980414 ACN121980414 ACN 121980414ACN-121980414-A

Abstract

The invention relates to the technical field of intelligent operation and maintenance of power equipment, and discloses a AmRMR-based transformer self-adaptive fault diagnosis method and a AmRMR-based transformer self-adaptive fault diagnosis system, wherein the method comprises the steps of obtaining detection data of dissolved gas in transformer oil and constructing a candidate ratio feature set; the method comprises the steps of respectively carrying out standardization treatment on key gas concentration characteristics and candidate ratio characteristic sets, carrying out discretization operation on the standardized candidate ratio characteristic sets, carrying out redundancy compression and information contribution evaluation on the discretized candidate ratio characteristics based on AmRMR algorithm, outputting the key ratio characteristic sets, inputting the key gas concentration characteristics and the key ratio characteristic sets into a pre-constructed DSD-DQN fault diagnosis model, and outputting a transformer fault diagnosis result. The invention can effectively avoid the highest engineering risk of misjudging the fault as normal, obviously improves the identification capability of few samples such as serious faults and the like, and realizes the diagnosis of the transformer faults with engineering safety and diagnosis accuracy.

Inventors

GU WENBIN
Li Nuandong
TAO JUNHAO
ZHANG WEIWEI
YUAN MINGHAI
PEI FENGQUE

Assignees

河海大学

Dates

Publication Date: 20260505
Application Date: 20260108

Claims (10)

1. The self-adaptive fault diagnosis method for the transformer based on AmRMR is characterized by comprising the following steps of: Obtaining detection data of dissolved gas in transformer oil to obtain key gas concentration characteristics, and constructing a candidate ratio characteristic set based on the gas detection data; The method comprises the steps of respectively carrying out standardization treatment on key gas concentration characteristics and candidate ratio characteristic sets to obtain standardized key gas concentration characteristics and standardized candidate ratio characteristic sets; Based on the obtained discretized candidate ratio feature set, redundant compression and information contribution evaluation are carried out on the discretized candidate ratio feature based on AmRMR algorithm, key ratio features which are most representative to fault classification are screened out, and then the key ratio feature set is output; and inputting the standardized key gas concentration characteristics and the obtained key ratio characteristic set into a pre-constructed DSD-DQN fault diagnosis model, and outputting a transformer fault diagnosis result.
2. The method for adaptively diagnosing a fault in a transformer based on AmRMR as set forth in claim 1, wherein said constructing a candidate ratio feature set based on gas detection data includes: the structural form is shown as the following formula: ; Wherein, the Representing the logarithmic ratio feature value of the construction; represent the first Concentration value of seed gas; represent the first A concentration value of a gas or a concentration value of total hydrocarbons.
3. The method for adaptively diagnosing a fault of a transformer based on AmRMR as set forth in claim 1, wherein the discretizing the normalized candidate ratio feature set to obtain a discretized candidate ratio feature set includes: The mathematical definition of the discretization operation is as follows: ; Wherein, the Characteristic label representing discretized candidate ratio and having a value range of ; Representing the normalized candidate ratio characteristic value; Representing the number of discretization steps; Respectively represent 1 st to 1 st A plurality of quantiles.
4. The method for adaptively diagnosing the transformer self-adaptive fault based on AmRMR of claim 1, wherein the step of performing redundancy compression and information contribution evaluation on the discretized candidate ratio features based on the AmRMR algorithm based on the obtained discretized candidate ratio feature set, and screening out key ratio features most representative of fault classification to output the key ratio feature set comprises the following steps: the degree of dependence between a single candidate ratio feature and a fault tag is measured by mutual information, and is defined as follows: ; Wherein, the Is a joint probability distribution; 、 are distributed for the respective edges; Representing characteristic variables And fault labels Mutual information value between the two; Representing characteristic variables Specific values of (1), namely discretized candidate ratio feature labels, the value set is ; Representing fault tag variables The specific value of (1) is the real fault label of the transformer; Feature screening is carried out in a round iteration mode, net information gain of candidate ratio features is calculated through design condition mutual information, and definition of the condition mutual information is shown as follows: ; Wherein, the Is a joint probability distribution of the probability distribution, Is a conditional probability distribution given Z, based on which candidate ratio features are defined relative to a selected feature set Net information gain of (2) The following formula is shown: ; Wherein, the Represent the first Candidate ratio features; is a fault label; representing the selected feature set; In each iteration, the gain is first calculated based on the net information obtained Calculate the first Single feature information contribution rate of individual candidate ratio features The following formula is shown: ; defining cumulative information contribution rate Is in front of The ratio of the sum of the net information gains of the candidate ratio features to the aggregate net information gain of the currently selected feature is used to evaluate the saturation of the feature subset as shown in the following equation: ; Wherein: Representing a currently selected feature set The total number of features in (a); represent the first Candidate ratio features At a given set The net information gain; represent the first Candidate ratio features; representing the remaining candidate ratio feature set; Truncated index variable representing cumulative summation, where Representing the front of the order of contribution rate Candidate ratio features; Before representation Accumulated information contribution rates of the candidate ratio features; setting double self-adaptive constraint in the screening process: Constraint 1. First Single feature net information contribution rate for individual candidate ratio features Not lower than the maximum information contribution rate of the first round Is 0.3 times that of (2); constraint 2, that the cumulative information contribution rate reaches a preset target threshold; If the candidate ratio feature meets the two constraint conditions, the candidate ratio feature is added into the selected feature set and then the next round of screening is carried out, and if the candidate ratio feature does not meet any constraint condition, iteration is stopped and the final key ratio feature set is output.
5. The method for adaptively diagnosing a fault of a transformer based on AmRMR as set forth in claim 1, wherein the inputting the normalized key gas concentration characteristic and the obtained key ratio characteristic set into the pre-constructed DSD-DQN fault diagnosis model to output the fault diagnosis result of the transformer comprises: The DSD-DQN fault diagnosis model comprises a convolutional neural network, and the convolutional neural network performs deep representation learning on the selected characteristics; Taking the standardized key gas concentration characteristics and the obtained key ratio characteristics as the input of a DSD-DQN fault diagnosis model, wherein the specific characteristic extraction process sequentially comprises a first convolution layer convolution, a first maximum pooling, a second convolution layer convolution, a second maximum pooling, flattening operation and full-connection layer mapping treatment, and mapping the flattened characteristic vector to a low-dimensional embedding space which is used as the input of a reinforcement learning part and used for the subsequent state cost function Dominance function with each action Respectively, to merge state cost functions Dominance function with each action Obtaining the final product A function of the value, The definition of the value function is: ; Wherein, the The parameters representing the shared feature extraction layer, 、 Parameters of the dominant function branch and the state cost function branch are respectively; Representing a motion variable for calculating an average of the merit function; representing the average value of all action dominant values in the current state; Target object The value calculation formula is shown as follows: ; Wherein, the Parameters representing the target network; identifying a variable for the termination state; representing a next time state to which the motion is transferred after the motion is performed; Indicating the current time Is of (1) A value; indicating the time of day of the agent Instant rewards obtained after the action is executed; Representing overall parameters of the current online network; target at current moment based on calculation Value of Constructing a loss function for updating network parameters, wherein the loss function is used for minimizing the output of the current online network Value of With the object Value of The prediction error between the two is calculated, Indicating time of day State of (2); indicating the time of day of the agent And (3) performing actions.
6. The method for adaptively diagnosing a fault of a transformer based on AmRMR as set forth in claim 1, wherein the training optimization method of the DSD-DQN fault diagnosis model is as follows: Introducing a priority experience playback mechanism, and allocating priority to each state transition sample by calculating time difference error (TD error) of each state transition sample, wherein the larger the error is, the higher the learning value of the state transition sample to the current strategy is, the state transition sample is Is determined according to the following formula, wherein Is the first TD error of the individual state transition samples, Non-uniform sampling is carried out according to the probability in the experience playback process, so that the model gives higher attention to the key state and the samples difficult to classify, thereby improving the convergence rate and the classification accuracy; ; Wherein, the Represent the first Probability of each state transition sample being sampled; representing the first of the experience playback pools The absolute value of the time difference error of each state transition sample; Index variables representing samples in the experience playback pool; the normalized denominator is represented by the term, Traversing the sum index variable of all stored samples in the experience playback pool; constructing a cost sensitive rewarding function fused with engineering risk priors, wherein the expression of the cost sensitive rewarding function is shown in the following formula As a true class of fault, In order to predict the type of fault, Is the first The number of samples for each failure category, The number of fault class samples with the largest number of samples; ; ; ; Wherein, the Representing an instant rewarding value after the DSD-DQN fault diagnosis model makes diagnosis action in the current state; representing a cost sensitivity penalty coefficient reflecting engineering risks; Representing class imbalance weight factors reflecting the distribution of the data; represent the first Class imbalance ratio of class fault samples; Representing the average unbalance ratio of all fault categories for normalizing the weights; a total number of samples representing the class with the greatest number of samples in the training set; represent training set The total number of samples for class failures; In defining penalty coefficients Before, according to the actual fault type of the transformer, the faults are divided into three categories of discharge faults, thermal faults and normal faults, and the fault is classified into three categories by punishment coefficients The engineering hazard degree is integrated into the DSD-DQN fault diagnosis model learning process, and the hazard degrees of different misjudgment types are classified into missed report > cross-class misdistribution > similar misdistribution > misdistribution, wherein the missed report is truly a fault and predicted to be normal, the cross-class misdistribution is truly a fault and is predicted to be a fault, and the similar misdistribution is mistakenly a fault type within the same large class.
7. A AmRMR-based transformer adaptive fault diagnosis system, characterized in that a AmRMR-based transformer adaptive fault diagnosis method according to any one of claims 1 to 6 is implemented, comprising: The data preprocessing module is configured to acquire detection data of dissolved gas in the transformer oil, obtain key gas concentration characteristics and construct a candidate ratio characteristic set based on the gas detection data; the method comprises the steps of respectively carrying out standardization treatment on key gas concentration characteristics and candidate ratio characteristic sets to obtain standardized key gas concentration characteristics and standardized candidate ratio characteristic sets; the self-adaptive feature screening module is configured to carry out redundancy compression and information contribution evaluation on the discretized candidate ratio features based on the obtained discretized candidate ratio feature set and a AmRMR algorithm, screen out key ratio features which are most representative of fault classification, and further output the key ratio feature set; The fault identification module is configured to input the standardized key gas concentration characteristics and the obtained key ratio characteristic set into a pre-constructed DSD-DQN fault diagnosis model and output a transformer fault diagnosis result.
8. A computer readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the AmRMR-based adaptive fault diagnosis method for a transformer of any one of claims 1 to 6.
9. A computer device, comprising: A memory for storing a computer program; A processor for executing the computer program to implement the steps of the AmRMR-based transformer adaptive fault diagnosis method according to any one of claims 1-6.
10. A computer program product comprising a computer program, characterized in that the computer program, when being executed by a processor, implements the steps of the AmRMR-based adaptive fault diagnosis method for a transformer according to any one of claims 1 to 6.

Description

AmRMR-based transformer self-adaptive fault diagnosis method and system Technical Field The invention belongs to the technical field of intelligent operation and maintenance of power equipment, relates to a self-adaptive fault diagnosis method and system for a transformer based on AmRMR, and is particularly suitable for intelligent multi-class fault identification scenes of transformers based on gas detection data in oil. Background Transformers are key devices in power systems, and the insulation state and the operation condition of the transformers are directly related to the safety and reliability of a power grid. The analysis (Dissolved GAS ANALYSIS, DGA) of the dissolved gas in the oil has been widely used for monitoring the state of the transformer and diagnosing faults because of the convenience in sampling, high sensitivity and capability of reflecting internal defects such as partial discharge, overheat and the like in early stage. By detecting the content of characteristic gases such as hydrogen, methane, ethane, ethylene, acetylene and the like in the oil and the ratio relation thereof, the possible fault types in the transformer can be deduced. The existing DGA fault diagnosis method mainly comprises rule methods such as an empirical ratio method, a key gas method, a three-ratio method, a criterion based on IEC standards and the like, and the methods depend on expert experience and a preset threshold value, so that the adaptability to complex operation conditions and novel equipment is limited. With the development of machine learning and deep learning, more and more researches are conducted to model the fault diagnosis of the transformer DGA as a multi-classification problem, and a support vector machine, a neural network, a random forest, integrated learning, a deep neural network and other supervised learning models are adopted to improve the automation and the accuracy of diagnosis. In such methods, the input features are typically selected manually based on experience, such as direct use of raw gas concentrations, typical ratios, or a few combined features, lacking systematic assessment of feature-to-fault type correlation and inter-feature redundancy. Meanwhile, the model adopts a static classification frame, takes overall accuracy or average loss as a main optimization target, and is difficult to reflect the differences of class unbalance and different fault misjudgment costs under the same frame. In actual operation, the DGA data of the transformer often show unbalanced classification characteristics that the number of normal and slight fault samples is large, serious faults and malignant defect samples are relatively rare, and the risk and cost of misdiagnosing the serious faults as normal or low-risk faults are obviously higher than those of general misclassification. In the traditional supervised learning model, although the problems can be considered to a certain extent through means such as oversampling, undersampling, category weights or cost sensitive rewarding functions, a plurality of groups of weight parameters are usually required to be manually set through repeated experiments, category imbalance and misjudgment cost are easily mixed in the same weight, and the relationship between the data distribution characteristics and the engineering hazard degree is difficult to clearly describe. Disclosure of Invention Aiming at the problems that in the existing transformer fault diagnosis, the characteristic information quantity of a single DGA gas component is deficient, a complex fault mode is difficult to fully represent, and the existing model ignores misjudgment cost difference and lacks consideration on engineering risks, the invention provides a self-adaptive fault diagnosis method and system for a transformer based on AmRMR, which can effectively avoid the highest engineering risk of misjudging a fault as normal, remarkably improve the identification capability of few samples such as serious faults and the like, and realize the transformer fault diagnosis which takes engineering safety and diagnosis accuracy into consideration. In order to solve the technical problems, the invention is realized by adopting the following technical scheme. In a first aspect, the present invention provides a method for adaptively diagnosing a fault of a transformer based on AmRMR, including the following steps: Obtaining detection data of dissolved gas in transformer oil to obtain key gas concentration characteristics, and constructing a candidate ratio characteristic set based on the gas detection data; The method comprises the steps of respectively carrying out standardization treatment on key gas concentration characteristics and candidate ratio characteristic sets to obtain standardized key gas concentration characteristics and standardized candidate ratio characteristic sets; Based on the obtained discretized candidate ratio feature set, redundant compression and information contribution eva