Transformer full life cycle state deduction and maintenance optimization decision-making system, method and device and medium

Abstract

The invention relates to the technical field of power equipment state monitoring and intelligent operation and maintenance, and discloses a transformer full life cycle state deduction and maintenance optimization decision system, method, device and medium, and the system comprises a data access and preprocessing module which is used for providing data input, causal structure prior, future driving data and maintenance measure prior; the dynamic causal model building module is used for building an uncertainty structure causal model and outputting parameter posterior probability distribution; the probabilistic state analysis module is used for quantifying future risk indexes based on the parameter posterior probability distribution and future driving data; the anti-fact decision optimization module is used for selecting an optimal maintenance measure by using parameter posterior probability distribution and maintenance measure priori when a future risk index exceeds a threshold value; and a system operation environment and interface module. According to the method, uncertainty can be quantified, probabilistic deduction can be carried out on the future state of the transformer, and an optimal maintenance decision can be obtained.

Description

Technical Field

[0001] This invention relates to the field of power equipment condition monitoring and intelligent operation and maintenance technology, and in particular to a transformer full life cycle condition prediction and maintenance optimization decision-making system, method, equipment and medium. Background Technology

[0002] At present, transformers can be considered the most critical core equipment in the power system, so the operational reliability of transformers affects the stability of the power grid.

[0003] To ensure the safe operation of transformers in the power grid and to optimize the full life cycle cost of transformers in the power grid, current technologies generally use condition monitoring systems to acquire operational data. Then, these acquired data are used to construct health status assessment models or remaining life prediction models for later evaluation.

[0004] However, some existing evaluation methods have limitations in practical applications. Many existing evaluation models are static, and the parameters of these static models are fixed after offline training, making it difficult for them to adapt to the complex operating conditions of transformers and changing environmental factors. This leads to a decrease in model prediction accuracy over time. Moreover, most of these existing evaluation models are used to solve deterministic single-point prediction results and cannot quantify the uncertainty of the prediction itself, so the reliability of the evaluation results of these models cannot be guaranteed.

[0005] In addition, there are some data-driven models. These data-driven models mainly focus on the statistical correlation between variables and do not consider the physical mechanisms and causal relationships of the internal state evolution of the transformer, so they cannot achieve the effect of real prediction.

[0006] Considering the shortcomings of the evaluation models in the existing technologies, the formulation of existing maintenance strategies largely relies on fixed time periods, simple state threshold triggers, or the experience and knowledge of maintenance personnel. This approach cannot flexibly select the optimal maintenance decision among multiple alternative maintenance measures. Summary of the Invention

[0007] In view of the aforementioned existing problems, the present invention is proposed.

[0008] Therefore, this invention provides a transformer full life cycle state extrapolation and maintenance optimization decision-making system, method, equipment and medium, which can solve the problems that existing evaluation models cannot quantify uncertainty, do not consider physical mechanisms and causal relationships, and that existing maintenance strategies are not flexible enough.

[0009] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a transformer full life cycle state prediction and maintenance optimization decision-making system, comprising: The data access and preprocessing module is used to collect and process multi-source data, and output data input, prior knowledge, and maintenance measure priors; A dynamic causal model construction module, connected to the data access and preprocessing module, is used to construct an uncertain structural causal model and, using the data input and prior knowledge provided by the data access and preprocessing module, learn online the strength and uncertainty of causal relationships in the uncertain structural causal model, and output a parametric posterior probability distribution. The probabilistic state analysis module, connected to the dynamic causal model construction module, performs a probabilistic extrapolation of the transformer's health state based on the parameter posterior probability distribution and quantifies a future risk indicator. The counterfactual decision optimization module is used to obtain a set of candidate maintenance measures when the future risk indicator exceeds a preset risk trigger threshold, and to perform an uncertainty counterfactual query on each of the candidate maintenance measures using the parameter posterior probability distribution of the maintenance measures prior, and select an optimal maintenance measure. The system operating environment and interface module provides a hardware operating environment for the data access and preprocessing module, the dynamic causal model construction module, the probabilistic state analysis module, and the counterfactual decision optimization module, and provides external data and decision interaction interfaces.

[0010] As a preferred embodiment of the transformer full life cycle state extrapolation and maintenance optimization decision-making system of the present invention, the data access and preprocessing module includes a priori knowledge base unit, which includes: The physical mechanism knowledge subunit is used to store causal structure priors, which include causal relationship rules, for the dynamic causal model construction module to define the basic causal graph of the U-DSCM. The maintenance measure knowledge subunit is used to store the prior knowledge of the maintenance measures; the prior knowledge of the maintenance measures includes the intervention model definition and direct execution cost corresponding to the candidate maintenance measures.

[0011] This preferred solution can provide accurate prior knowledge for the system of the present invention, thereby enabling the system to perform causal model construction, state inference and decision optimization in a more scientific and reasonable manner during operation.

[0012] As a preferred embodiment of the transformer full life cycle state extrapolation and maintenance optimization decision-making system described in this invention, the dynamic causal model construction module includes: A causal probabilistic unit is used to assign structural equations to the nodes of the U-DSCM and define the initial prior probability distribution of the parameter vectors in the structural equations. The online learning and uncertainty quantification unit is used to receive new observation data at the current time and, according to the sequential form of Bayes' theorem, combine the posterior distribution of the previous time with the likelihood function of the new observation data to update the parameter posterior probability distribution at the current time.

[0013] As a preferred embodiment of the transformer full life cycle state extrapolation and maintenance optimization decision system described in this invention, the online learning and uncertainty quantification unit includes a posterior distribution sequential update subunit, which uses the sequential Monte Carlo method to approximate the posterior probability distribution of the parameters as a set of weighted parameter particles.

[0014] As a preferred embodiment of the transformer full life cycle state extrapolation and maintenance optimization decision system described in this invention, the probabilistic state analysis module includes a probabilistic state extrapolation unit, which is used to extrapolate the predicted probability distribution of key internal states at a future time step based on the parameter posterior probability distribution and the future driving data, and to calculate the future risk index based on a predefined failure threshold. The future driving data is a future external driving prediction sequence; The future risk indicator is quantified as the probability that the critical internal state will be worse than the failure threshold at the future time step.

[0015] As a preferred embodiment of the transformer full life cycle state extrapolation and maintenance optimization decision system described in this invention, the probabilistic state extrapolation unit includes a Monte Carlo forward simulation subunit, which generates the predicted probability distribution by performing multiple simulation runs, wherein each simulation run includes: Draw a parameter sample from the posterior probability distribution of the parameter; By combining the parameter sample and the future external driving prediction sequence, a future state sample is obtained through time-series iterative deduction on the U-DSCM.

[0016] As a preferred embodiment of the transformer full life cycle state extrapolation and maintenance optimization decision-making system of the present invention, the counterfactual decision optimization module includes a maintenance intervention modeling unit, which is used for: Map each of the proposed maintenance measures in the group to one or a group of do-operators; The do-operator is then applied to the baseline U-DSCM to generate a set of corresponding intervention models.

[0017] Secondly, the present invention provides a method for transformer full life cycle state prediction and maintenance optimization decision-making, including: Collect and process multi-source data, and output data input, prior knowledge, and maintenance measure priors; Construct an uncertain structural causal model, and use the data input and the prior knowledge to learn the causal strength and uncertainty in the uncertain structural causal model online, and output a parametric posterior probability distribution; Based on the posterior probability distribution of the parameters, the health status of the transformer is probabilistically extrapolated, and a future risk indicator is quantified. When the future risk indicator exceeds a preset risk trigger threshold, a set of candidate maintenance measures is obtained, and an uncertainty counterfactual query is performed on each of the candidate maintenance measures using the parameter posterior probability distribution prior to select an optimal maintenance measure.

[0018] Thirdly, the present invention provides an electronic device including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the method described above.

[0019] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described above.

[0020] Compared with existing technologies, the beneficial effects of this invention are that it proposes a transformer full life cycle state prediction and maintenance optimization decision-making system. The system includes: a data access and preprocessing module for providing data input, causal structure priors, future driving data, and maintenance measure priors; a dynamic causal model construction module for constructing an uncertain structural causal model and outputting the parameter posterior probability distribution; a probabilistic state analysis module for quantifying future risk indicators based on the parameter posterior probability distribution and the future driving data; a counterfactual decision optimization module for selecting the optimal maintenance measure when the future risk indicator exceeds a threshold, using the parameter posterior probability distribution and the maintenance measure priors; and a system operating environment and interface module. This invention can quantify uncertainty, realize probabilistic prediction of the future state of transformers, and provide optimal maintenance decisions that balance cost and risk. Attached Figure Description

[0021] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a system module framework diagram of a transformer full life cycle state extrapolation and maintenance optimization decision-making system provided in one embodiment of the present invention.

[0023] Figure 2 This is a schematic diagram of the data access and preprocessing module of a transformer full life cycle state extrapolation and maintenance optimization decision-making system provided in one embodiment of the present invention.

[0024] Figure 3 This is an internal structure diagram of an electronic device for a transformer full life cycle state prediction and maintenance optimization decision-making system, provided as an embodiment of the present invention. Detailed Implementation

[0025] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0026] Example 1, referring to Figures 1-2 This is the first embodiment of the present invention, which provides a transformer full life cycle state prediction and maintenance optimization decision-making system, including: Figure 1 This diagram illustrates a system module framework for a transformer full life-cycle state prediction and maintenance optimization decision-making system, including: The data access and preprocessing module 100 is used to collect and process multi-source data and output data input, prior knowledge, and maintenance measure priors. The dynamic causal model construction module 200 is connected to the data access and preprocessing module. It is used to construct an uncertain structural causal model and, using the data input and prior knowledge provided by the data access and preprocessing module, learns online the strength of causal relationships and uncertainty in the uncertain structural causal model, and outputs a parametric posterior probability distribution. The probabilistic state analysis module 300 is connected to the dynamic causal model construction module. Based on the posterior probability distribution of parameters, it performs a probabilistic inference on the health status of the transformer and quantifies a future risk indicator. The counterfactual decision optimization module 400 is used to obtain a set of candidate maintenance measures when the future risk indicator exceeds a preset risk trigger threshold, and to perform an uncertainty counterfactual query on each candidate maintenance measure using the parameter posterior probability distribution maintenance measure prior to select an optimal maintenance measure. The system operating environment and interface module 500 provides the hardware operating environment for the data access and preprocessing module, the dynamic causal model construction module, the probabilistic state analysis module, and the counterfactual decision optimization module, and provides external data and decision interaction interfaces.

[0027] In an embodiment of the present invention, the data access and preprocessing module includes a priori knowledge base unit, which includes: The physical mechanism knowledge subunit is used to store the causal structure prior, which includes causal relationship rules, providing the dynamic causal model building module with the basic causal graph for defining the U-DSCM. The maintenance measure knowledge subunit is used to store the prior knowledge of maintenance measures. The prior knowledge of maintenance measures includes the intervention model definition and direct execution cost corresponding to the candidate maintenance measures.

[0028] In this embodiment of the invention, the dynamic causal model construction module includes: Causal probabilistic unit: The causal probabilistic unit is used to assign structural equations to the nodes of U-DSCM and define the initial prior probability distribution of the parameter vectors in the structural equations. The online learning and uncertainty quantization unit is used to receive new observation data at the current time and, according to the sequential form of Bayes' theorem, combine the posterior distribution of the previous time step with the likelihood function of the new observation data to update the parameter posterior probability distribution at the current time step.

[0029] In an embodiment of the present invention, the online learning and uncertainty quantification unit includes a posterior distribution sequential update subunit. The posterior distribution sequential update subunit adopts the sequential Monte Carlo method to approximate the posterior probability distribution of the parameters as a set of weighted parameter particles.

[0030] In an embodiment of the present invention, the probabilistic state analysis module includes a probabilistic state deduction unit, which is used to deduce the predicted probability distribution of the key internal state at a future time step based on the posterior probability distribution of the parameters and future driving data, and to calculate the future risk index based on a predefined failure threshold. Future-driven data is a future external-driven predictive sequence; The future risk indicator is quantified as the probability that a critical internal state will deteriorate below the failure threshold in future time steps.

[0031] In an embodiment of the present invention, the probabilistic state deduction unit includes a Monte Carlo forward simulation subunit, which generates a predicted probability distribution by performing multiple simulation runs, wherein each simulation run includes: Draw a parameter sample from the posterior probability distribution of the parameters; By combining parameter samples and future external driving prediction sequences, a future state sample is obtained through time-series iterative extrapolation on U-DSCM.

[0032] In an embodiment of the present invention, the counterfactual decision optimization module includes a maintenance intervention modeling unit, which is used for: Map each of the candidate maintenance measures in the group to one or a group of do-operators; The do-operator is then applied to the baseline U-DSCM to generate a set of corresponding intervention models.

[0033] In an optional implementation, the counterfactual decision optimization module includes an uncertainty counterfactual query unit, which is used to calculate the counterfactual probability distribution at future time steps for each intervened model by calling the Monte Carlo forward simulation logic. The simulated parameter samples are drawn from the posterior probability distribution of the original parameters output by the dynamic causal model building module.

[0034] In an optional implementation, the counterfactual decision optimization module includes a risk perception decision unit, which is used for: Transform the counterfactual probability distribution of each measure into a counterfactual loss distribution; A scalar risk value is calculated from the counterfactual loss distribution; the scalar risk value is the conditional risk value. The optimal maintenance measure is selected by solving an optimization function that minimizes the total cost, which is obtained by summing the direct execution cost of the measure and the conditional risk value multiplied by a risk preference coefficient.

[0035] In one optional implementation, the system operating environment and interface module includes a hardware resource configuration unit, which includes: One or more central processing units are used to perform data I / O and task scheduling for the data access and preprocessing module; One or more graphics processors and one or more tensor processors are used to accelerate parallel computing tasks in the dynamic causal model building module, the probabilistic state analysis module, and the counterfactual decision optimization module.

[0036] Example 2: Based on the above examples, a specific implementation of a transformer full life cycle state prediction and maintenance optimization decision-making system can be designed as follows: See attached document Figure 1 The data access and preprocessing module 100 in this embodiment of the invention functions to provide standardized data input and prior knowledge input for the subsequent dynamic causal model construction module 200. In a specific embodiment, this module 100 includes a multi-source data acquisition unit, a data preprocessing unit, and a prior knowledge base unit.

[0037] In one optional implementation, the multi-source data acquisition unit is used to aggregate data from multiple physical and information systems related to the transformer. In one embodiment, the unit enables data access by configuring various data interface adapters. For example, the unit can access the substation's backend real-time or historical database through an OPC (Open Platform Communication) interface, Modbus TCP / IP protocol, or ODBC (Open Database Connectivity) to obtain the required data.

[0038] In one optional implementation, the multi-source data acquisition unit may specifically include the following sub-units: Sensor data ( The data acquisition subunit is responsible for acquiring sensor measurements that reflect the internal state of the transformer. These measurements are time-series data.

[0039] Specifically, the types of data collected include data from online dissolved gas analysis (DGA) monitoring devices, such as volume fractions (ppm) of H2 (hydrogen), CH4 (methane), C2H4 (ethylene), C2H2 (acetylene), CO (carbon monoxide), and CO2 (carbon dioxide); data from online partial discharge (PD) monitoring devices, such as discharge amplitude (mV), discharge pulse count, and phase spectral decomposition (PRPD) statistics from ultra-high frequency (UHF) sensors; and data from trace water in the oil. ) and oil temperature ( Data from sensors. Data from fiber optic temperature measurement systems, such as winding hotspot temperatures ( Direct measurements of (if equipped). Data from the casing monitoring system, such as the casing's medium loss factor ( ) and capacitance value ( ).

[0040] In an optional implementation, operating conditions may also be included ( The data acquisition subunit is responsible for acquiring the variables that define the transformer's operating boundary conditions and external stresses. These variables are also time-series data.

[0041] Specifically, the data it collects includes data from SCADA systems or power quality monitoring systems, such as the three-phase currents on the high and low voltage sides of the transformer. ), three-phase voltage ( ), active power ( ) and reactive power ( Data from the cooler control system, such as the operating status of the cooler (fan, oil pump). Data from local weather stations or environmental sensors, such as ambient air temperature. ).

[0042] In an alternative implementation, a static data acquisition subunit may also be included, which is responsible for retrieving the inherent attribute information of the transformer from the asset management system (EAM) or equipment ledger database on a one-time or low-frequency basis. This information remains constant or changes slowly throughout the lifecycle of a single transformer.

[0043] Specifically, this information includes the device ID, manufacturer, year of manufacture, rated voltage level, rated capacity, cooling method (e.g., ONAN / ONAF), connection group, and impedance parameters. It also includes historical operation and maintenance records, such as historical oil filtration, oil change, and overhaul dates and details, which are used for model initialization or verification.

[0044] In an optional implementation, a data preprocessing unit may also be included. This unit receives raw data from the multi-source data acquisition unit and performs standardization processing to generate a dataset that is uniformly formatted, of reliable quality, and usable by the dynamic causal model building module 200. This unit may specifically include the following sub-units: In an optional implementation, a data cleaning subunit may also be included to address data quality issues in the raw data. For missing values ​​(e.g., data point loss due to communication interruption): depending on the type and duration of the missing data, forward imputation, backward imputation, or regression interpolation based on the variable and other relevant variables (e.g., the correlation between load and oil temperature) may be used.

[0045] For outliers (e.g., outliers that deviate significantly from physical laws due to instantaneous sensor jumps or acquisition errors): they can be identified using statistical distribution-based methods (e.g., the 3-sigma criterion) or smoothed using time-series filtering methods (e.g., median filters).

[0046] In an optional implementation, a data alignment and resampling subunit may also be included, due to the sensor data ( ) and operating conditions ( The sampling frequencies (sampling periods) of the data differ; for example, load data is collected in minutes, while DGA data is collected in hours or days. This subunit is responsible for unifying all time-series data to a preset time resolution. (For example, 1 hour).

[0047] For sampling frequency higher than For data (such as minute-level load): time window aggregation methods can be used, such as calculating the mean, maximum, or integral value within the window, as the data. The representative value of the period is used to achieve downsampling.

[0048] For sampling frequency lower than Data (such as daily DGA): can be generated using zero-order hold or linear interpolation methods. The data at a higher resolution is upsampled to align with other data on the time axis.

[0049] In an optional implementation, a priori knowledge base unit may also be included. This unit is a structured knowledge repository for storing domain knowledge related to transformers. The output of this knowledge base is used to guide the initialization of the basic causal graph in the dynamic causal model building module 200, and the definition of decision parameters in the counterfactual decision optimization module 400. This unit may specifically include the following sub-units: In an alternative implementation, a physical mechanism knowledge subunit may also be included, which stores the recognized physicochemical laws governing the transformer's internal structure. In one embodiment, this knowledge is stored as a set of structured causal rules used to define the edge set EE of the underlying causal graph.

[0050] For example, storage rules (thermal effects): In other words, the load current is the direct cause of the temperature rise in the winding hot spot.

[0051] For example, storage rules (aging effect): In other words, the hot spot temperature and the moisture content of the insulating paper are the common direct causes of the decrease in the degree of polymerization (DP) of the insulating paper.

[0052] For example, storage rules (gas generation effect): In other words, high-temperature overheating or partial discharge events are the direct cause of acetylene (C2H2) production.

[0053] In an optional implementation, a maintenance action knowledge subunit may also be included, which stores attributes of standard maintenance operations. In one embodiment, this knowledge is stored as a lookup table or key-value database for the counterfactual decision optimization module 400 to define intervention operations. and cost function .

[0054] For example, storage (measure ID: Name: "Online Oil Filtration", Execution Cost: Intervention model: This intervention forces the trace water content in the oil to a low value. And reduce the moisture in the paper. .

[0055] For example, storage (measure ID: M2, name: "Replace Cooler", execution cost: Intervention model: This means that the intervention modified the causal strength parameter from ambient temperature to oil temperature. .

[0056] In an optional implementation, the dynamic causal model building module 200 is connected to the data access and preprocessing module 100, and to the subsequent probabilistic state analysis module 300 and counterfactual decision optimization module 400.

[0057] The function of module 200 is to receive time-series data, static data, and prior knowledge provided by module 100, and construct a dynamic, probabilistic structural causal model (U-DSCM) based on this. This model is used to characterize the intrinsic physicochemical causal relationships of individual transformers, and is responsible for updating the posterior probability distribution of the model parameters online, thereby quantifying the cognitive uncertainty of the system regarding the strength of various causal relationships.

[0058] In one specific embodiment, the dynamic causal model construction module 200 includes: a basic causal graph definition unit, a causal relationship probabilization unit, and an online learning and uncertainty quantification unit.

[0059] Among them, the basic causal graph definition unit is responsible for defining the static topological structure of the U-DSCM, namely a directed acyclic graph (DAG). The structure is determined based on the physical mechanism knowledge provided by the prior knowledge base unit in the data access and preprocessing module 100.

[0060] In one alternative implementation, the node set ( Defining sub-units is used to define diagrams. All nodes .node It is the collection of all variables collected and processed by the data access and preprocessing module 100. In one embodiment, this collection... It is divided into three mutually exclusive subsets: Exogenous driving variables : Represents input variables that are not affected by other variables in the model, such as load current. Ambient temperature These nodes are in the graph. There is no parent node.

[0061] Endogenous state variables These represent key physicochemical state variables inside the transformer that are typically impossible to measure directly or are too costly to measure. For example, winding hot spot temperature. Degree of polymerization of insulating paper Moisture content of insulating paper .

[0062] Observable characterization variables : Represents the actual measured value of the sensor, its value is or The result (effect) of the interaction of variables. For example, DGA gas components (H2, C2H2, etc.), partial discharge signals. .

[0063] Total Node Set .

[0064] In an optional implementation, the prior edge set ( Defining sub-units is used to define diagrams. The set of directed edges between nodes This edge set Obtained from the physical mechanism knowledge subunit of the prior knowledge base unit.

[0065] A directed edge express yes One direct physical reason ( yes (parent node). For example, the child unit defines the following edges based on the physical mechanism: This indicates that the load current is the direct cause of the hot spot temperature.

[0066] and This indicates that hot spot temperature and moisture content in the paper are the common direct causes of the decrease in DP.

[0067] This indicates that the hot spot temperature is the direct cause of the production of ethylene (a DGA gas).

[0068] In an optional implementation, a causal probabilistic unit may also be included, responsible for generating a static graph output by the basic causal graph definition unit. This is transformed into a computable dynamic probability model.

[0069] In an optional implementation, structural equation (SCM) modeling of sub-elements may also be included, which are shown in the figure. Each endogenous node in Assign a structural equation This equation defines how the value of a node is inherited from its parent node. Decision: In the formula: It is a node exist The numerical value of the moment. yes The parent node (by the edge set) (Confirmed) A set of numerical values ​​at any given time. It is a functional form characterizing this causal mechanism. This functional form can be based on physical formulas (e.g., Aging models such as Dakins' or Montsinger's can be used, or a general function approximator (e.g., a neural network or a Gaussian process) can be employed. It is a function The parameter vector, whose physical meaning is right The causal strength. For example, in the Arrhenius equation for DP aging, It includes activation energy and pre-exponential factor. It is an exogenous noise term, representing all those not present in Modeling in China, for The random influence of.

[0070] In an optional implementation, a prior distribution of parameters may also be included ( Define a sub-element that contains the parameter vectors of all structure equations. It is considered as a random variable to be estimated.

[0071] This subunit is responsible for Set an initial prior probability distribution The prior distribution The settings can be constructed based on domain knowledge, for example, for those based on physical formulas. Its parameters The prior can be set as a distribution within its physically reasonable range (e.g., the mean of the activation energy parameter is set to a literature value, and the variance to a large value). For parameters lacking domain knowledge, a broad, uninformative, or weakly informative prior distribution can be constructed (e.g., a Gaussian distribution with a mean of 0 and a large variance). This prior distribution... As the initial input for online learning and uncertainty quantification units.

[0072] In an optional implementation, an online learning and uncertainty quantification unit may also be included, which is the core of the U-DSCM's dynamic updates. This unit receives new observation data from the data access and preprocessing module 100. And used to sequentially update parameters. The probability distribution.

[0073] In an alternative implementation, a likelihood function may also be included. Construct sub-units that are responsible for defining "given a specific set of parameters" In the case of observing new data The probability.

[0074] The likelihood function (can be simplified to) ) from structural equations and noise The probability distributions are jointly determined. Specifically, given... and Structural equations It will output the correct answer. A predicted distribution; this sub-unit calculates the actual observed distribution. The probability density under this predicted distribution is the likelihood value.

[0075] In an optional implementation, a posterior distribution sequential update subunit may also be included, which is the core of the Bayesian inference execution. It receives... posterior parameters at time (As a priori at the current moment) and the calculated likelihood function, then calculate posterior distribution of parameters at time 1 .

[0076] Update the sequential form based on Bayes' theorem: in, It is up to At any given moment, all data was observed. After that, parameters The posterior probability distribution. yes The likelihood function of the new data at each time step. yes The posterior distribution at time (i.e., the current time) (Prior distribution at time).

[0077] In one embodiment, due to If no analytical solution exists, an approximate inference algorithm is used for this sub-unit. For example, the Sequential Monte Carlo (SMC, or particle filter) method can be employed. This method will use the posterior... Approximately a set A weighted parameter particle .exist At that moment, the sub-unit executes: Resampling: based on Time weight For particles Perform resampling.

[0078] Propagation: The process of state transition for parameters (if the parameters are modeled as dynamic, e.g.) ).

[0079] Weighting: using the likelihood function Calculate new weights .

[0080] In an alternative implementation, uncertainty may also be included. ) Quantize the output subunit, which receives the complete posterior distribution of the parameters from the aforementioned output. (For example, Group of particles .

[0081] Its function is to extract information about each specific causal relationship (e.g., ...) from this high-dimensional distribution. The statistical information is obtained and output to the probabilistic state analysis module 300 and the counterfactual decision optimization module 400.

[0082] for Causal strength parameter The sub-unit calculates its marginal posterior distribution. (For example, through a histogram of the particle set). Based on this marginal distribution, the sub-unit calculates and outputs: Posterior mean : As exist The most likely estimate of the time.

[0083] Posterior variance : This variance value is the system's tolerance. Quantification of the uncertainty of the estimated value. The larger the value, the weaker the data support for the strength of the causal relationship, and the more ambiguous the system's understanding of the relationship.

[0084] It should be noted that the output of this unit is... (and its statistical summary) , The data is passed to the probabilistic state analysis module 300 and the counterfactual decision optimization module 400 as the model basis for their inference and decision-making.

[0085] In an optional implementation, the probabilistic state analysis module 300 has its input connected to the dynamic causal model construction module 200 and its output connected to the counterfactual decision optimization module 400 and the system operating environment and interface module 500.

[0086] The function of this module 300 is to receive and utilize the posterior probability distribution characterizing the individual causal relationships of the transformer, output by the dynamic causal model construction module 200. It performs two core analytical tasks: first, when abnormal characteristics are detected, it performs probabilistic fault root cause attribution; second, it analyzes the future critical health status of the transformer (e.g., We will conduct probabilistic simulations and quantify future risks.

[0087] In one specific embodiment, the probabilistic state analysis module 300 includes a probabilistic root cause attribution unit and a probabilistic state deduction unit.

[0088] It should be noted that the probabilistic root cause attribution unit is a monitoring representation This function is activated when an anomaly (e.g., C2H2 concentration) is identified as abnormal. Its function is to trace the anomaly's characterization backwards based on the U-DSCM. In endogenous state variables The most likely root cause (e.g., "high-temperature overheating" or "high-energy discharge"). This unit receives the posterior distribution of parameters from module 200. (In one embodiment, for) One parameter particle ), and perform the following operations: In an optional implementation, a total causal effect (TCE) distribution calculation subunit may also be included, which is used to calculate the distribution from a potential root cause. To anomaly representation Total causal effect The probability distribution.

[0089] The specific calculation steps are as follows: Obtain the posterior distribution of parameters from module 200. Particles .

[0090] for arrive Every particle : Using this particle Instantiate a defined structural causal model .

[0091] In Above, calculation right Point estimates of total causal effect .

[0092] In one embodiment, if the model Structural equations in It is nonlinear, Through calculation right The sensitivity to change (i.e., the gradient) is obtained by using the chain rule along the graph. middle arrive Calculate all directed paths.

[0093] In another embodiment, if the model is linear, i.e. ,but for arrive Causality strength on all paths ( The sum of the products.

[0094] This subunit outputs A set of point estimates of the total causal effect This set is The numerical representation of the posterior probability distribution.

[0095] In an alternative implementation, an attribution list generation subunit may also be included, which receives one or more potential root causes. corresponding The distribution is calculated, and a sorted list of attributions is generated.

[0096] The specific calculation steps are as follows: For each potential root cause , obtain its Distribution set .

[0097] Define an attribution ranking metric. In one embodiment, this metric is the "probability of a significant causal effect," i.e. ,in It is a preset threshold that characterizes the significance of the effect (e.g.) ).

[0098] This probability is obtained through... The number of particles is counted to estimate: the calculation satisfies number of particles Then the probability is .

[0099] This subunit uses this probability index to analyze all potential root causes. Sort in descending order.

[0100] Output the sorted attribution list (e.g., [("High-energy discharge", 0.85, ("High-temperature overheating", 0.15)]) to the system operating environment and interface module 500 for fault diagnosis display.

[0101] In an optional implementation, a probabilistic state deduction unit may also be included, which is used to predict the future. After one time step (i.e.) Key internal states at any given time (e.g., degree of polymerization of insulating paper) The probability distribution of ) and quantify future risks.

[0102] This unit receives the posterior distribution of parameters from module 200. ), and future external driving prediction sequences from module 100. (For example, load curve forecast for the next 24 hours).

[0103] In an alternative implementation, a Monte Carlo forward simulation subunit may also be included, which performs... To conduct independent simulations to deduce A possible future state.

[0104] The specific calculation steps are as follows: for arrive Each simulation run: Parameter sampling: from the posterior distribution provided by module 200 Randomly select a parameter sample (In the particle filtering embodiment, it can be seen from...) One particle is randomly selected from the particles. ).

[0105] State initialization: Get the current state Internal state at time 1 (For example, currently) (estimated mean).

[0106] Iterative deduction of time series: For arrive Every future time step: Obtain the external driver for this time step ; For each endogenous node in U-DSCM From its noise distribution (As defined in module 221) Extract a noise sample ; According to the structural equation Sampling parameters and noise Calculate the node in Value at time: .

[0107] Storage number The simulation ran in Key state results at time 1 .

[0108] This subunit outputs A sample set of future states .

[0109] In an optional implementation, a future state prediction distribution generation subunit may also be included, which receives data from... a sample set and convert it to Predicted probability distribution .

[0110] In one embodiment, the sub-unit achieves this through... Kernel density estimation (KDE) or histogram construction is performed on each sample to generate a numerical representation of the predicted distribution.

[0111] This subunit can also calculate a statistical summary of the distribution, such as predicting the mean. Prediction variance and the 5th, 50th, and 95th percentiles The predicted distribution or its statistical summary is output to module 500 for visualization of health status (e.g., confidence intervals for life expectancy prediction).

[0112] In an optional implementation, a future risk indicator quantification subunit may also be included, which receives data from... A set of samples, based on a predefined failure threshold. To calculate a scalar risk indicator .

[0113] In one embodiment, the failure threshold is Lifetime end threshold (e.g.) ).

[0114] The risk indicators calculated by this subunit for The probability of failure at any given moment: This probability is obtained through... To estimate, count the samples: statistics Among the samples, the following conditions are met: Number of samples .

[0115] Calculate risk indicators: .

[0116] This subunit will calculate the scalar risk index. (For example The output is sent to the counterfactual decision optimization module 400. This module 400 will... With a preset risk trigger threshold If a comparison is made, If this occurs, module 400 will be triggered to perform maintenance decision optimization.

[0117] The counterfactual decision optimization module 400 has its input connected to the dynamic causal model construction module 200 and the probabilistic state analysis module 300, and its output connected to the system operating environment and interface module 500.

[0118] The function of this counterfactual decision optimization module 400 is to, when it receives future risk indicators from the future risk indicator quantification subunit of the probabilistic state analysis module 300... Exceeding a preset risk trigger threshold Time (i.e.) ), module 400 is activated.

[0119] Upon activation, the counterfactual decision optimization module 400 obtains a set of candidate maintenance measures from the maintenance measures knowledge subunit of the data access and preprocessing module 100. Subsequently, module 400 utilizes the dynamic causal model to construct the posterior distribution of the parameters output by module 200. For each of the candidate measures An uncertainty counterfactual query is performed to extrapolate the probability distribution of future states after taking the measure. Finally, this module 400 calculates and selects an optimal maintenance measure based on a preset risk-adjusted cost function. .

[0120] In one specific embodiment, the counterfactual decision optimization module 400 includes: a maintenance intervention modeling unit, an uncertainty counterfactual query unit, and a risk perception decision unit.

[0121] In an alternative implementation, a maintenance intervention modeling unit may also be included, which functions to convert real-world maintenance operations obtained from the maintenance action knowledge subunit into formal "intervention" operations on the U-DSCM.

[0122] The unit receives the set of alternative maintenance measures. At least include measures of "inaction". and one or more active interventions, such as ="Online oil filtration" ="Replace the cooler".

[0123] This unit specifically includes the following sub-units: In an optional implementation, it may also include - Operator definition subunit, which is used to define each maintenance measure Mapped to one or a group - Operator. - An operator is a mathematical operation used to represent a forced modification to the structure or parameters of a U-DSCM.

[0124] For example, for =“Online oil filtration”, its operating effect is to forcibly reduce the trace water content in the oil. This subunit defines it as ),in It is a micro-water node in oil. It is a target threshold (e.g., 5 ppm).

[0125] For example, for ="Replacing the cooler" improves heat dissipation efficiency. This sub-unit defines it as... ,in It is the causal strength parameter in U-DSCM that characterizes the effect of ambient temperature on oil temperature. These are the efficiency parameters of the new cooler.

[0126] For example, for = "inaction", its - The operator is an empty operation, that is .

[0127] In an alternative implementation, an intervention model may also be included. Generate a subunit, which receives data from... - Operator definition, and its application to the baseline U-DSCM (Figure 200) constructed by the Dynamic Causal Model Building Module 200. and structural equations This generates a set of corresponding "intervention models". .

[0128] for This subunit is generated. In the diagram In, all pointing to Directed edges of a node (e.g., from) The edge was removed. And, Structural equations Replaced with a new equation .

[0129] for This subunit is generated. .exist In the diagram structure It remains unchanged, but Structural equations of nodes Modified to ,in Parameters in Forced to be replaced with .

[0130] for The subunit generated This is the original U-DSCM ( ).

[0131] This unit outputs the group of intervention models. To the Uncertainty Counterfactual Query Unit.

[0132] In an optional implementation, an uncertainty counterfactual query unit may also be included, the function of which is to query each intervened model output by the maintenance intervention modeling unit. Calculate its future The probabilistic consequences of a given moment.

[0133] This unit receives the posterior distribution of parameters from module 200. and future external driving forecasts from module 100 .

[0134] In an alternative implementation, a counterfactual distribution may also be included ( ) Calculate the sub-unit, which is for Each of the candidate measures (include ), calculate its corresponding critical state (e.g. )exist Counterfactual probability distribution at time 1 .

[0135] This calculation is implemented by calling the logic of the Monte Carlo forward simulation sub-unit in the probabilistic state analysis module 300, but its simulation process is performed within the intervened model. Execute above: for arrive Each simulation run: Parameter sampling: from the posterior distribution of the original parameters provided by module 200 Extract a parameter sample from .

[0136] State initialization: Get the current state Internal state at time 1 .

[0137] Iterative deduction of time series: For arrive At each future time step, the intervention model is applied. Structural equations (For example, in) In the simulation, Always forced to be set ), and combined Iterative calculation .

[0138] Storage number The simulation ran in Counterfactual results of moments .

[0139] This sub-unit is for each measure Output a string containing A set of samples This set is the numerical representation of the counterfactual distribution.

[0140] This unit outputs the counterfactual distribution set of all measures to the risk perception decision-making unit.

[0141] In an optional implementation, a risk perception decision-making unit may also be included, which functions to quantitatively evaluate multiple counterfactual distributions of the output and select an optimal maintenance measure. .

[0142] This unit receives from One counterfactual sample, and each measure from the maintenance measures knowledge subunit of module 100. direct execution costs .

[0143] In an optional implementation, a loss distribution and risk metric (CVaR) calculation subunit may also be included, which first converts the output state distribution into a loss distribution and then calculates its risk metric.

[0144] for Each measure and A counterfactual sample : Define a loss function In one embodiment, the function is ,in This is the DP failure threshold.

[0145] calculate Loss samples This set constitutes The counterfactual loss distribution.

[0146] The sub-unit then calculates a risk metric from the loss distribution. In one embodiment, Conditional Value at Risk (CVaR) is used.

[0147] The steps for calculating CVaR are as follows: Set a confidence level (For example =0.95).

[0148] Calculate the loss distribution -Value-at-Risk That is, the 100th loss sample percentile.

[0149] calculate That is, all that are greater than or equal to The arithmetic mean of the loss samples.

[0150] This subunit is Each measure Output a scalar risk value .

[0151] In an alternative implementation, optimal decision-making may also be included. Select a sub-unit, which is responsible for performing the final decision optimization to select the measure with the lowest total cost.

[0152] This subunit is Each measure Define a total cost function : in, It is execution Direct costs (such as labor, materials, and downtime costs). It is calculated, in Expected future loss (risk cost) at the confidence level. This is a preset risk preference coefficient used to balance current direct costs and future risk costs. This sub-unit calculates the total cost of all candidate measures and selects the one that... Minimization measures : This subunit will select the optimal decision (For example, the ID or name of the measure) is output to the decision application interface unit of the system operating environment and interface module 500.

[0153] The system operating environment and interface module 500 provides physical hardware resources for all computing and storage tasks in the data access and preprocessing module 100, the dynamic causal model construction module 200, the probabilistic state analysis module 300, and the counterfactual decision optimization module 400, and is responsible for outputting the analysis results and decision instructions of this system to external information systems or operating terminals.

[0154] In one specific embodiment, the system operating environment and interface module 500 can be deployed on a local server, an edge computing device, or a cloud platform. The module 500 includes a hardware resource configuration unit and a decision application interface unit.

[0155] In an optional implementation, a hardware resource configuration unit may also be included, which provides the computing and storage resources required for the operation of the system of the present invention. In one embodiment, this unit is one or a group of physical servers or virtual machine instances.

[0156] In an optional implementation, a computing processor subunit may also be included, which is responsible for performing all data processing and model calculation steps in the system of the present invention. This subunit may include: One or more central processing units (CPUs): used to perform data I / O, cleaning, and alignment tasks in the data access and preprocessing module 100, as well as to perform task scheduling, control flow logic, and data formatting and communication tasks of the decision application interface unit for the entire system.

[0157] One or more graphics processing units (GPUs) or tensor processors (TPUs): used to accelerate the parallel computational tasks of the posterior distribution sequential update subunit 232 in the dynamic causal model building module 200 (e.g., large-scale particle operations in sequential Monte Carlo / particle filtering), and to accelerate the execution of the Monte Carlo forward simulation subunit in the probabilistic state analysis module 300 and the uncertainty counterfactual query unit in the counterfactual decision optimization module 400. Sub-Monte Carlo forward simulation.

[0158] High-speed internal buses (such as PCIe) and network interfaces (such as Ethernet) are used for data exchange between the CPU, GPU / TPU, and data storage subunits.

[0159] In an optional implementation, a data storage subunit may also be included, which is responsible for providing persistent storage and high-speed memory caching of data for the system of the present invention. This subunit may include: Volatile memory (RAM): Used to provide high-speed working memory for computing processor subunits, for example, for storing the posterior distribution of parameters output by the dynamic causal model building module 200 in real time during computation. of Particles .

[0160] Non-volatile memory (e.g., SSD or HDD): used for persistent data storage. In one embodiment, the memory is configured as follows: A time series database (TSDB) is used to store sensor data acquired and processed by the data access and preprocessing module 100. and operating conditions .

[0161] A relational database or key-value store is used to store the equipment ledger information acquired by the static data acquisition subunit, as well as the physical mechanism knowledge and maintenance measures knowledge in the prior knowledge base unit.

[0162] A file system or object store for storing model checkpoints and historical computation results (e.g., historical data). and (and system logs).

[0163] In an optional implementation, a decision application interface unit may also be included, which is responsible for converting the analysis results and decision recommendations generated by modules 300 and 400 into a standardized data format and transmitting them to an external system.

[0164] In an optional implementation, a risk warning output subunit may also be included, which receives output from the probabilistic state analysis module 300. Specifically: This subunit receives scalar risk indicators calculated by the future risk indicator quantification subunit. .

[0165] This subunit receives a statistical summary of the predicted distribution (e.g., the predicted mean) generated by the future state prediction distribution generation subunit. and confidence interval ).

[0166] This subunit encapsulates the above data into data packets in a predefined format (such as JSON, XML, or messages conforming to specific industry standards such as IEC61850).

[0167] The subunit sends the data packet to one or more external systems, such as the operator station HMI (human-machine interface) of a substation DCS (distributed control system) or the dashboard of an asset health management (AHM) system, via a network interface (e.g., via HTTP / REST API calls, MQTT message push, or writing to a shared database).

[0168] In an optional implementation, a maintenance decision work order generation subunit may also be included, which receives output from the counterfactual decision optimization module 400. Specifically: This subunit receives the optimal maintenance measures determined by the optimal decision-making subunit. Identifier (ID).

[0169] This subunit uses The identifier is used to query the database of the maintenance measure knowledge subunit in the data access and preprocessing module 100 to obtain a complete description of the measure (e.g., “Perform online oil filtration operation on [device ID]”), the relevant operating procedure number, and the estimated cost. .

[0170] This subunit assembles this information into a structured data record that conforms to the work order data pattern required by external enterprise asset management (EAM) systems or computerized maintenance management systems (CMMS).

[0171] This subunit pushes the work order data record to EAM or CMMS through a standard Enterprise Application Integration (EAI) interface (e.g., WebService (SOAP), REST API, or Message Queuing (MQ)) to trigger the creation or approval process of a new maintenance work order.

[0172] Example 3, referring to Figure 3This embodiment also provides a method for transformer full life cycle state extrapolation and maintenance optimization decision-making, including: Collect and process multi-source data, and output data input, prior knowledge, and maintenance measure priors; Construct an uncertain structural causal model, and use data input and prior knowledge to learn the strength of causal relationships and uncertainty in the uncertain structural causal model online, and output a parametric posterior probability distribution; Based on the posterior probability distribution of parameters, the health status of the transformer is probabilistically extrapolated, and a future risk indicator is quantified. When a future risk indicator exceeds a preset risk trigger threshold, a set of candidate maintenance measures is obtained. Then, using the parameter posterior probability distribution maintenance measure prior, an uncertainty counterfactual query is performed on each candidate maintenance measure to select the optimal maintenance measure.

[0173] The above-mentioned unit modules can be embedded in the processor of the electronic device in hardware form or independent of it, or they can be stored in the memory of the electronic device in software form, so that the processor can call and execute the corresponding operations of the above modules.

[0174] This embodiment also provides an electronic device, which can be a terminal, and its internal structure diagram can be as follows: Figure 3 As shown, the electronic device includes a processor, memory, communication interface, display screen, and input device connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. When the computer program is executed by the processor, it implements a method for transformer lifecycle state prediction and maintenance optimization decision-making. The display screen can be an LCD screen or an e-ink screen. The input device can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the device's casing, or an external keyboard, touchpad, or mouse.

[0175] This embodiment also provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, it performs the following steps: Collect and process multi-source data, and output data input, prior knowledge, and maintenance measure priors; Construct an uncertain structural causal model, and use data input and prior knowledge to learn the strength of causal relationships and uncertainty in the uncertain structural causal model online, and output a parametric posterior probability distribution; Based on the posterior probability distribution of parameters, the health status of the transformer is probabilistically extrapolated, and a future risk indicator is quantified. When a future risk indicator exceeds a preset risk trigger threshold, a set of candidate maintenance measures is obtained. Then, using the parameter posterior probability distribution maintenance measure prior, an uncertainty counterfactual query is performed on each candidate maintenance measure to select the optimal maintenance measure.

[0176] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

[0177] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the invention.

[0178] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A transformer full life cycle state prediction and maintenance optimization decision-making system, characterized in that, include: The data access and preprocessing module is used to collect and process multi-source data, and output data input, prior knowledge, and maintenance measure priors; A dynamic causal model construction module, connected to the data access and preprocessing module, is used to construct an uncertain structural causal model and, using the data input and prior knowledge provided by the data access and preprocessing module, learn online the strength and uncertainty of causal relationships in the uncertain structural causal model, and output a parametric posterior probability distribution. The probabilistic state analysis module, connected to the dynamic causal model construction module, performs a probabilistic extrapolation of the transformer's health state based on the parameter posterior probability distribution and quantifies a future risk indicator. The counterfactual decision optimization module is used to obtain a set of candidate maintenance measures when the future risk indicator exceeds a preset risk trigger threshold, and to perform an uncertainty counterfactual query on each of the candidate maintenance measures using the parameter posterior probability distribution of the maintenance measures prior, and select an optimal maintenance measure. The system operating environment and interface module provides a hardware operating environment for the data access and preprocessing module, the dynamic causal model construction module, the probabilistic state analysis module, and the counterfactual decision optimization module, and provides external data and decision interaction interfaces.

2. The transformer full life cycle state prediction and maintenance optimization decision-making system as described in claim 1, characterized in that, The data access and preprocessing module includes a prior knowledge base unit, which includes: The physical mechanism knowledge subunit is used to store causal structure priors, which include causal relationship rules, for the dynamic causal model construction module to define the basic causal graph of the U-DSCM. The maintenance measure knowledge subunit is used to store the prior knowledge of the maintenance measures; the prior knowledge of the maintenance measures includes the intervention model definition and direct execution cost corresponding to the candidate maintenance measures.

3. The transformer full life cycle state prediction and maintenance optimization decision-making system as described in claim 2, characterized in that, The dynamic causal model construction module includes: A causal probabilistic unit is used to assign structural equations to the nodes of the U-DSCM and define the initial prior probability distribution of the parameter vectors in the structural equations. The online learning and uncertainty quantification unit is used to receive new observation data at the current time and, according to the sequential form of Bayes' theorem, combine the posterior distribution of the previous time with the likelihood function of the new observation data to update the parameter posterior probability distribution at the current time.

4. The transformer full life cycle state prediction and maintenance optimization decision-making system as described in claim 3, characterized in that, The online learning and uncertainty quantification unit includes a posterior distribution sequential update subunit, which uses the sequential Monte Carlo method to approximate the posterior probability distribution of the parameters as a set of weighted parameter particles.

5. The transformer full life cycle state prediction and maintenance optimization decision-making system as described in claim 4, characterized in that, The probabilistic state analysis module includes a probabilistic state deduction unit, which is used to deduce the predicted probability distribution of key internal states at a future time step based on the posterior probability distribution of the parameters and the future driving data, and to calculate the future risk index based on a predefined failure threshold. The future driving data is a future external driving prediction sequence; The future risk indicator is quantified as the probability that the critical internal state will be worse than the failure threshold at the future time step.

6. The transformer full life cycle state prediction and maintenance optimization decision-making system as described in claim 5, characterized in that, The probabilistic state deduction unit includes a Monte Carlo forward simulation subunit, which generates the predicted probability distribution by performing multiple simulation runs, wherein each simulation run includes: Draw a parameter sample from the posterior probability distribution of the parameter; By combining the parameter sample and the future external driving prediction sequence, a future state sample is obtained through time-series iterative deduction on the U-DSCM.

7. The transformer full life cycle state prediction and maintenance optimization decision-making system as described in claim 6, characterized in that, The counterfactual decision optimization module includes a maintenance intervention modeling unit, which is used for: Map each of the proposed maintenance measures in the group to one or a group of do-operators; The do-operator is then applied to the baseline U-DSCM to generate a set of corresponding intervention models.

8. A method for transformer full life cycle state extrapolation and maintenance optimization decision-making, using the system described in any one of claims 1 to 7, characterized in that, include: Collect and process multi-source data, and output data input, prior knowledge, and maintenance measure priors; Construct an uncertain structural causal model, and use the data input and the prior knowledge to learn the causal strength and uncertainty in the uncertain structural causal model online, and output a parametric posterior probability distribution; Based on the posterior probability distribution of the parameters, the health status of the transformer is probabilistically extrapolated, and a future risk indicator is quantified. When the future risk indicator exceeds a preset risk trigger threshold, a set of candidate maintenance measures is obtained, and an uncertainty counterfactual query is performed on each of the candidate maintenance measures using the parameter posterior probability distribution prior to select an optimal maintenance measure.

9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the transformer full life cycle state deduction and maintenance optimization decision-making method as described in claim 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the transformer full life cycle state deduction and maintenance optimization decision-making method as described in claim 8.

Images