A power distribution cabinet health state predictive maintenance system based on digital twinning and intelligent sensing

Abstract

The application discloses a power distribution cabinet health state predictive maintenance system based on digital twinning and intelligent sensing, comprising a data acquisition and preprocessing module, which is used for acquiring multi-source operation data and generating standardized input data sets; a causal degradation feature extraction module, which is used for constructing a causal degradation graph and generating a degradation feature vector; a digital twinning modeling module, which is used for constructing a power distribution cabinet digital twinning model and establishing a power distribution cabinet graph structure topology; a constraint fusion prediction module, which is used for outputting component health indexes, degradation parameter estimated values and future state prediction results; a counterfactual maintenance simulation module, which is used for introducing a counterfactual maintenance simulator and generating risk change values and maintenance cost values; and a scheduling and strategy generation module, which is used for generating optimal scheduling and execution strategies. The application realizes accurate prediction of the health state of the power distribution cabinet and optimal generation of maintenance strategies by combining multi-modal intelligent sensing and digital twinning modeling.

Description

Technical Field

[0001] This invention relates to the field of power equipment operation and maintenance and intelligent monitoring technology, and in particular to a predictive maintenance system for the health status of distribution cabinets based on digital twins and intelligent sensing. Background Technology

[0002] Currently, distribution cabinets, as key equipment in power systems, directly affect the safety and stability of the distribution network. With the continuous expansion of the power grid and the improvement of its intelligence level, traditional methods relying on manual inspection or single-sensor monitoring are no longer sufficient to meet the needs of modern power operation and maintenance. In existing technologies, researchers generally collect key parameters during operation by deploying temperature, current, voltage, and partial discharge sensors inside the distribution cabinet, and then combine these with threshold judgment or trend analysis methods for fault early warning. However, these methods often only perform static monitoring of single parameters, lack deep fusion of multi-source data, and are easily affected by load fluctuations and environmental conditions, leading to unstable degradation feature extraction and insufficient reliability of health assessment results.

[0003] With the development of digital twin technology, some existing studies have attempted to introduce digital twin models into the condition assessment of power distribution equipment to achieve a combination of physical models and data-driven approaches. However, existing digital twin models are mostly based on single-domain or simplified physical constraints, failing to fully consider the dual-domain coupling relationship between electrical and thermal aspects, resulting in discrepancies between model predictions and actual operating conditions. Existing models generally lack the embedding of causal mechanisms, relying solely on data correlation, making it difficult to accurately characterize the true evolution of degradation factors and observed signals, leading to poor interpretability and robustness of predictions.

[0004] Regarding maintenance strategies, existing methods typically make simple decisions based on equipment health index thresholds, lacking quantitative assessments of the differences before and after maintenance measures, and failing to provide maintenance personnel with a basis for balancing risks and costs. This leads to maintenance decisions that are often conservative or delayed, increasing unnecessary maintenance costs and making it difficult to avoid potential failure risks in a timely manner. Existing technologies still have significant shortcomings in areas such as robust extraction of degradation features, physical constraints and causal embedding of twin models, and quantitative optimization of maintenance strategies.

[0005] Therefore, how to provide a predictive maintenance system for the health status of power distribution cabinets based on digital twins and intelligent sensing is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a predictive maintenance system for the health status of power distribution cabinets based on digital twins and intelligent sensing. This invention fully utilizes multimodal intelligent sensors, causal degradation maps and invariant constraint methods, and digital twin modeling technology involving electrical-thermal dual-domain coupling. It details the processing flow through robust degradation feature extraction, constraint fusion prediction, and counterfactual maintenance simulation to achieve accurate assessment and predictive maintenance of the power distribution cabinet's operating status. This invention possesses advantages such as high robustness in degradation feature extraction, strong predictive reliability of the twin model, and quantifiable optimization of maintenance decisions.

[0007] According to an embodiment of the present invention, a predictive maintenance system for the health status of a power distribution cabinet based on digital twins and intelligent sensing includes the following modules:

[0008] The data acquisition and preprocessing module is used to collect and preprocess multi-source operational data, generate a standardized input dataset, and divide data samples into multiple operating conditions.

[0009] The causal degradation feature extraction module is used to construct a causal degradation map, extract stable degradation features, and generate degradation feature vectors.

[0010] The digital twin modeling module is used to build a digital twin model of the power distribution cabinet and establish the topology of the power distribution cabinet diagram.

[0011] The constraint fusion prediction module is used to embed a current conservation constraint layer and a heat diffusion constraint layer into the digital twin model of the power distribution cabinet, and output the component health index, degradation parameter estimate and future state prediction results.

[0012] The counterfact maintenance simulation module is used to introduce a counterfact maintenance simulator into the digital twin model, perform counterfactual operations such as parameter reset, replacement or cleaning, and generate risk change values ​​and maintenance cost values.

[0013] The scheduling and strategy generation module is used to generate optimal scheduling and execution strategies by combining maintenance time windows, spare parts supply and power grid reliability constraints.

[0014] Optionally, modules can be integrated using the following methods:

[0015] Multi-source operating data from multimodal smart sensors installed inside the power distribution cabinet is collected, preprocessed, and a standardized input dataset is generated. Data samples are then divided into multiple operating conditions according to load level and environmental conditions.

[0016] Based on the normalized input dataset, a causal degradation map is constructed. The invariant constraint method of multi-condition data samples is used to extract degradation features that remain stable under different loads and environmental conditions, forming a degradation feature vector.

[0017] Based on the normalized input dataset and degenerate feature vectors, a digital twin model of the distribution cabinet is constructed. The busbars, terminals, circuit breakers, and insulation components are used as nodes of the graph structure, and the electrical coupling relationship and heat conduction relationship are used as edges of the graph structure to establish the graph structure topology of the distribution cabinet.

[0018] A current conservation constraint layer and a heat diffusion constraint layer are embedded in the digital twin model of the power distribution cabinet, and a multi-head attention mechanism is introduced. The degradation feature vector is used as input and combined with the normalized input dataset to output the health index, degradation parameter estimate and state prediction results of each component of the power distribution cabinet within the future time window.

[0019] A counterfactual maintenance simulator is introduced into the digital twin model of the power distribution cabinet. Counterfactual operations such as parameter reset, replacement or cleaning are performed on the target components. The temperature rise change, partial discharge evolution and failure probability before and after maintenance are calculated respectively, and risk change value and maintenance cost value are generated.

[0020] Based on the risk change value and maintenance cost value, combined with the operation and maintenance time window, spare parts supply and power grid reliability constraints, the optimal scheduling and execution strategy for predictive maintenance of the distribution cabinet is generated.

[0021] Optionally, the multimodal smart sensor includes a temperature sensor, a current sensor, a voltage sensor, a partial discharge sensor, and a humidity sensor.

[0022] Optionally, the multi-source operating data includes temperature data, current data, voltage data, partial discharge data, and humidity data, and the preprocessing of the multi-source operating data includes time synchronization processing, noise filtering processing, and normalization processing.

[0023] Optionally, forming the degenerate feature vector includes:

[0024] It receives a normalized input dataset and multi-condition data samples. Each condition domain contains temperature data, current data, voltage data, partial discharge data, and humidity data, as well as corresponding load levels, ambient temperature, and ambient humidity labels.

[0025] Constructing a causal degradation map:

[0026] The degradation factor nodes are defined as the increase in contact resistance, the increase in thermal interface resistance, and the increase in insulation loss factor.

[0027] Set the load level, ambient temperature, and ambient humidity as external operating condition factor nodes;

[0028] Busbar temperature rise, terminal temperature rise, partial discharge amplitude, current fluctuation characteristics, and voltage deviation are set as observation signal nodes;

[0029] Establish directed causal relationships between degradation factor nodes and observation signal nodes, and between external condition factor nodes and observation signal nodes, and determine the direction and strength parameters of each causal relationship.

[0030] Based on multi-condition data samples and causal degradation maps, cross-condition invariant constraint feature extraction is performed. Consistency constraints are applied to the degradation-related feature statistics of each condition domain and the cross-domain overall statistics to ensure that the deviation of the mean and covariance of degradation-related features of each condition domain from the cross-domain overall statistics does not exceed the preset range, thus obtaining a degradation candidate feature set.

[0031] The degraded candidate feature set is subjected to distribution alignment processing. Alignment is performed using a distribution difference metric to ensure that the feature distribution difference between any two working domains does not exceed a preset threshold, thus forming a degraded feature set.

[0032] The set of degradation features consistent across operating conditions is aggregated according to the mapping relationship between the sensing channel and the distribution cabinet components to form a degradation feature vector.

[0033] Optionally, the construction of the digital twin model of the power distribution cabinet includes:

[0034] Based on the normalized input dataset and degenerate feature vector, a digital twin model of the distribution cabinet is established. A hierarchical heterogeneous multi-graph structure is adopted to define the graph structure topology. The equipment layer is used to represent circuits and segments, the component layer is used to represent busbars, terminals, circuit breakers and insulation components, and the sensing layer is used to represent sensing channels. The connection relationship includes electrical coupling relationship and thermal conduction relationship.

[0035] Establish a mapping relationship between sensing channels and component nodes, register the data in the normalized input dataset as node observation vectors in chronological order, establish a mapping relationship between degradation feature vectors and corresponding component nodes, and create a component parameter registry to record geometric parameters, material parameters and rated parameters;

[0036] In the digital twin model of the power distribution cabinet, the electric domain kernel and the thermal domain kernel are instantiated, a shared degradation state interface is set to realize dual-domain coupling, and physical constraint interface, boundary condition interface and cross-domain coupling mapping table are configured to identify the correspondence between electric domain nodes and thermal domain nodes.

[0037] Register the basic initial state for the digital twin model of the power distribution cabinet. The basic initial state includes the initial electrical state and initial thermal state of the component nodes, the load level and environmental condition occupancy data under the current operating conditions, register the topology version number and connection relationship verification information, and set the calculation step size and update order for state updates.

[0038] A degraded mirror node generation module is loaded into the digital twin model of the power distribution cabinet to generate a healthy baseline mirror node for each component node in the component layer. The healthy baseline state is set according to the component parameter registry and the registered operating condition data. A one-to-one mapping relationship between the mirror node and the component node is established. The difference between the component node and the healthy baseline state is recorded through differential links. The digital twin model of the power distribution cabinet carrying the mirror node and differential state channels is output.

[0039] Optionally, the health index, degradation parameter estimates, and state prediction results for each component of the output distribution cabinet within a future time window include:

[0040] In the digital twin model of the power distribution cabinet, a current conservation constraint layer and a heat diffusion constraint layer are loaded, and the normalized input dataset and degenerate feature vector are read.

[0041] In the current conservation constraint layer, for each component node, the algebraic sum of the inflow current and outflow current is constrained to be zero. For each connection relationship, the voltage drop is determined based on the component's equivalent resistance and the current passing through it. The contact resistance, thermal interface-related electrical parameters, and insulation loss-related electrical parameters obtained by mapping the degradation feature vector are recorded, and the electrical state quantities are updated.

[0042] In the thermal diffusion constraint layer, a dual-path dynamic conduction method is used to update the node temperature state:

[0043] On the one hand, conventional conduction updates are performed based on the temperature difference between adjacent nodes, node heat capacity, and thermal conductivity.

[0044] On the other hand, a degradation-sensitive path is introduced to add additional thermal resistance compensation and heat source correction to the components characterized by degradation feature vectors;

[0045] A temperature update logic is formed that includes normal conduction and degradation compensation, and the updated thermal state variables are output.

[0046] A multi-head attention mechanism is introduced to construct the input features of each component node. The input features consist of node observation data, the degradation feature vector corresponding to the component node, and the load level and environmental conditions. The information of adjacent nodes is weighted and aggregated according to the attention weight to generate the node representation.

[0047] Under the combined constraints of the current conservation constraint layer and the thermal diffusion constraint layer, the system takes node representation as input and outputs component-level health index, component degradation parameter estimates, and state prediction results of temperature, partial discharge, and failure probability within the future time window.

[0048] Optionally, the generation of risk change values ​​and maintenance cost values ​​includes:

[0049] Obtain a snapshot of the digital twin model of the power distribution cabinet at the current moment, and synchronously read the registered differential status channels;

[0050] The target component set is selected based on the health index and differential status channel. The maintenance action library is called to define the maintenance action set to be evaluated. The maintenance action set includes four types of actions: parameter reset, replacement, cleaning and tightening. In the parameter update rule library, the corresponding degradation parameter update rules and maintenance cost descriptions are specified for each target component and each maintenance action.

[0051] On the digital twin model snapshot of the power distribution cabinet, a counterfactual simulation branch is created for each target component and each maintenance action. The set time step and boundary conditions are kept unchanged. The degradation parameters of the target components are updated according to the parameter update rules. The digital twin model of the power distribution cabinet is run to generate the predicted trajectory of temperature, partial discharge and failure probability after maintenance.

[0052] Based on the preset risk assessment rules, the predicted trajectories of temperature, partial discharge and failure probability before and after maintenance are converted into comparable risk scores. The risk reduction of each counterfactual simulation branch is calculated. Combined with the maintenance cost description and the set cost weights, a benefit score is generated.

[0053] For each target component, the maintenance action with the highest benefit score is selected from the candidate maintenance actions to form a maintenance plan package that includes component identification, selected maintenance action, expected risk reduction, estimated maintenance cost, post-maintenance predicted trajectory summary, required resources and expected time window.

[0054] Optionally, the optimal scheduling and execution strategy for generating predictive maintenance of the distribution cabinet includes:

[0055] Based on the maintenance plan package, the risk change value and maintenance cost value of each target component are sorted by benefit, forming a candidate job list and initial execution order arranged by priority;

[0056] The candidate job list is matched and verified with the operation and maintenance time window, spare parts supply and power grid reliability constraints. Specific time periods and resources are allocated to each target component according to the constraint feasibility. If the constraints are not met, the execution order, maintenance actions or job batches are adjusted, and the optimal schedule is output.

[0057] Execution strategies are generated based on the optimal schedule, and specific maintenance actions, schedules and resource requirements are determined according to component priority and resource configuration results.

[0058] The beneficial effects of this invention are:

[0059] This invention deploys multimodal intelligent sensors inside the distribution cabinet to collect multi-source operational data, including temperature, current, voltage, partial discharge, humidity, operating load, and environmental conditions. Combined with causal degradation mapping and invariant constraint methods, it achieves causal modeling and stable feature extraction of degradation factors and external operating conditions. Unlike existing technologies that rely on only a single parameter or static feature, this invention effectively reduces the interference of load fluctuations and environmental changes on degradation features, thereby improving the robustness and consistency of health status identification and providing a more reliable data foundation for prediction.

[0060] In terms of model construction, this invention proposes a digital twin model of a distribution cabinet that couples electrical and thermal domains, embedding a current conservation constraint layer and a heat diffusion constraint layer within it. Simultaneously, a multi-head attention mechanism is introduced, which not only ensures physical consistency during the model's prediction process but also dynamically highlights the weights of key features under different operating conditions, improving the accuracy and interpretability of the predictions. Compared to traditional schemes that rely on simplified single-domain models, the twin model of this invention can more realistically reflect the evolution of the distribution cabinet's operating state, enhancing the credibility of the prediction results.

[0061] In terms of maintenance strategy formulation, this invention introduces a counterfactual maintenance simulator to perform virtual maintenance operations such as parameter reset, replacement, or cleaning on target components. It calculates the risk changes and maintenance costs before and after maintenance, and generates optimal scheduling and execution strategies by combining maintenance time windows, spare parts supply, and power grid reliability constraints. This not only overcomes the limitations of existing technologies that rely on experience or single thresholds for maintenance decisions, but also enables quantifiable comparison of the effects of maintenance measures. It can reduce unnecessary maintenance costs while ensuring equipment safety, and improve the scientific and economical operation and maintenance of distribution cabinets. Attached Figure Description

[0062] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0063] Figure 1 This is a schematic diagram of the structure of a predictive maintenance system for the health status of a power distribution cabinet based on digital twins and intelligent sensing, as proposed in this invention.

[0064] Figure 2 This is a flowchart illustrating a predictive maintenance method for the health status of a power distribution cabinet based on digital twins and intelligent sensing, as proposed in this invention. Detailed Implementation

[0065] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0066] refer to Figure 1 A predictive maintenance system for the health status of power distribution cabinets based on digital twins and intelligent sensing includes the following modules:

[0067] The data acquisition and preprocessing module is used to collect and preprocess multi-source operational data, generate a standardized input dataset, and divide data samples into multiple operating conditions.

[0068] The causal degradation feature extraction module is used to construct a causal degradation map, extract stable degradation features, and generate degradation feature vectors.

[0069] The digital twin modeling module is used to build a digital twin model of the power distribution cabinet and establish the topology of the power distribution cabinet diagram.

[0070] The constraint fusion prediction module is used to embed a current conservation constraint layer and a heat diffusion constraint layer into the digital twin model of the power distribution cabinet, and output the component health index, degradation parameter estimate and future state prediction results.

[0071] The counterfact maintenance simulation module is used to introduce a counterfact maintenance simulator into the digital twin model, perform counterfactual operations such as parameter reset, replacement or cleaning, and generate risk change values ​​and maintenance cost values.

[0072] The scheduling and strategy generation module is used to generate optimal scheduling and execution strategies by combining maintenance time windows, spare parts supply and power grid reliability constraints.

[0073] refer to Figure 2 A predictive maintenance method for the health status of power distribution cabinets based on digital twins and intelligent sensing includes:

[0074] Multi-source operating data from multimodal smart sensors installed inside the power distribution cabinet is collected, preprocessed, and a standardized input dataset is generated. Data samples are then divided into multiple operating conditions according to load level and environmental conditions.

[0075] Based on the normalized input dataset, a causal degradation map is constructed. The invariant constraint method of multi-condition data samples is used to extract degradation features that remain stable under different loads and environmental conditions, forming a degradation feature vector.

[0076] Based on the normalized input dataset and degenerate feature vectors, a digital twin model of the distribution cabinet is constructed. The busbars, terminals, circuit breakers, and insulation components are used as nodes of the graph structure, and the electrical coupling relationship and heat conduction relationship are used as edges of the graph structure to establish the graph structure topology of the distribution cabinet.

[0077] A current conservation constraint layer and a heat diffusion constraint layer are embedded in the digital twin model of the power distribution cabinet, and a multi-head attention mechanism is introduced. The degradation feature vector is used as input and combined with the normalized input dataset to output the health index, degradation parameter estimate and state prediction results of each component of the power distribution cabinet within the future time window.

[0078] A counterfactual maintenance simulator is introduced into the digital twin model of the power distribution cabinet. Counterfactual operations such as parameter reset, replacement or cleaning are performed on the target components. The temperature rise change, partial discharge evolution and failure probability before and after maintenance are calculated respectively, and risk change value and maintenance cost value are generated.

[0079] Based on the risk change value and maintenance cost value, combined with the operation and maintenance time window, spare parts supply and power grid reliability constraints, the optimal scheduling and execution strategy for predictive maintenance of the distribution cabinet is generated.

[0080] In this embodiment, the multimodal smart sensor includes a temperature sensor, a current sensor, a voltage sensor, a partial discharge sensor, and a humidity sensor.

[0081] In this embodiment, the multi-source operating data includes temperature data, current data, voltage data, partial discharge data, and humidity data. The preprocessing of the multi-source operating data includes time synchronization processing, noise filtering processing, and normalization processing.

[0082] In this embodiment, forming the degenerate feature vector includes:

[0083] It receives a normalized input dataset and multi-condition data samples. Each condition domain contains temperature data, current data, voltage data, partial discharge data, and humidity data, as well as corresponding load levels, ambient temperature, and ambient humidity labels.

[0084] Constructing a causal degradation map:

[0085] The degradation factor nodes are defined as the increase in contact resistance, the increase in thermal interface resistance, and the increase in insulation loss factor.

[0086] Set the load level, ambient temperature, and ambient humidity as external operating condition factor nodes;

[0087] Busbar temperature rise, terminal temperature rise, partial discharge amplitude, current fluctuation characteristics, and voltage deviation are set as observation signal nodes;

[0088] Directed causal relationships are established between degradation factor nodes and observed signal nodes, and between external condition factor nodes and observed signal nodes. The direction and strength parameters of each causal relationship are determined. Specifically, determining the direction and strength parameters of each causal relationship involves:

[0089] Based on the temporal sequence and delay characteristics of multi-source operational data, the causal direction between degradation factors, external operating condition factors and observed signals is determined.

[0090] By combining the electrical and thermal physical constraints of the power distribution cabinet, the initially determined causal direction is corrected, and causal edges that do not conform to the operating law are eliminated.

[0091] The sensitivity of factor changes to the observed signal is calculated using disturbance response data under multiple operating conditions, and quantified as the strength parameter of the causal correlation edge.

[0092] Based on multi-condition data samples and causal degradation maps, cross-condition invariant constraint feature extraction is performed. Consistency constraints are applied to the degradation-related feature statistics of each condition domain and the cross-domain overall statistics to ensure that the deviation of the mean and covariance of degradation-related features of each condition domain from the cross-domain overall statistics does not exceed the preset range, thus obtaining a degradation candidate feature set.

[0093] The degraded candidate feature set is subjected to distribution alignment processing, which uses a distribution difference metric for alignment to ensure that the feature distribution difference between any two operating domains does not exceed a preset threshold, thus forming a degraded feature set. Specifically, the alignment using the distribution difference metric involves:

[0094] Based on the degradation candidate feature sets of different working conditions, the probability distribution of each feature under different working conditions is calculated, and the distribution difference measure between features is obtained.

[0095] The distribution difference metric is compared with a preset threshold, and the distribution shift or scaling operation is performed on features with excessive differences to reduce the differences between operating conditions.

[0096] After distribution correction, the aligned features are re-aggregated to form a set of degenerate features that are stable across operating conditions;

[0097] The set of degradation features consistent across operating conditions is aggregated according to the mapping relationship between the sensing channel and the distribution cabinet components to form a degradation feature vector, where the mapping relationship is as follows:

[0098] Temperature sensing channels correspond to busbars, terminals, and insulating components, and are used to characterize thermal aging and contact degradation features;

[0099] Current and voltage sensing channels correspond to circuit breakers and busbars, and are used to characterize electrical load and conductivity degradation characteristics.

[0100] The partial discharge and humidity sensing channels correspond to insulating components, used to characterize insulation degradation and environmental sensitivity degradation features.

[0101] In this embodiment, constructing the digital twin model of the power distribution cabinet includes:

[0102] Based on the normalized input dataset and degenerate feature vector, a digital twin model of the distribution cabinet is established. A hierarchical heterogeneous multi-graph structure is adopted to define the graph structure topology. The equipment layer is used to represent circuits and segments, the component layer is used to represent busbars, terminals, circuit breakers and insulation components, and the sensing layer is used to represent sensing channels. The connection relationship includes electrical coupling relationship and thermal conduction relationship.

[0103] Establish a mapping relationship between sensing channels and component nodes. Register the data in the normalized input dataset as node observation vectors in chronological order. Establish a mapping relationship between degenerate feature vectors and corresponding component nodes. Create a component parameter registry to record geometric parameters, material parameters, and rated parameters, where:

[0104] The mapping relationship between sensing channels and component nodes is as follows:

[0105] Temperature sensing channels correspond to busbar nodes, terminal nodes, and insulating component nodes;

[0106] The current and voltage sensing channels correspond to the busbar nodes and circuit breaker nodes;

[0107] Partial discharge and humidity sensing channels correspond to insulating component nodes;

[0108] The mapping relationship between the degenerate feature vector and the corresponding component node is as follows:

[0109] The conductivity degradation feature vector corresponds to the busbar node and the terminal node;

[0110] The thermal aging degradation feature vector corresponds to the busbar node and the insulating component node;

[0111] The insulation degradation feature vector corresponds to the node of the insulating component;

[0112] The mechanical wear degradation feature vector corresponds to the circuit breaker node;

[0113] In the digital twin model of the power distribution cabinet, the electric domain kernel and the thermal domain kernel are instantiated, a shared degradation state interface is set to realize dual-domain coupling, and physical constraint interface, boundary condition interface and cross-domain coupling mapping table are configured to identify the correspondence between electric domain nodes and thermal domain nodes.

[0114] Register the basic initial state for the digital twin model of the power distribution cabinet. The basic initial state includes the initial electrical state and initial thermal state of the component nodes, the load level and environmental condition occupancy data under the current operating conditions, register the topology version number and connection relationship verification information, and set the calculation step size and update order for state updates.

[0115] A degraded mirror node generation module is loaded into the digital twin model of the power distribution cabinet to generate a healthy baseline mirror node for each component node in the component layer. The healthy baseline state is set according to the component parameter registry and the registered operating condition data. A one-to-one mapping relationship between the mirror node and the component node is established. The difference between the component node and the healthy baseline state is recorded through differential links. The digital twin model of the power distribution cabinet carrying the mirror node and differential state channels is output.

[0116] In this embodiment, the health index, degradation parameter estimates, and state prediction results within the future time window for each component of the output distribution cabinet include:

[0117] In the digital twin model of the power distribution cabinet, a current conservation constraint layer and a heat diffusion constraint layer are loaded, and the normalized input dataset and degenerate feature vector are read.

[0118] In the current conservation constraint layer, for each component node, the algebraic sum of the inflow current and outflow current is constrained to be zero. For each connection relationship, the voltage drop is determined based on the component's equivalent resistance and the current passing through it. The contact resistance, thermal interface-related electrical parameters, and insulation loss-related electrical parameters obtained by mapping the degradation feature vector are recorded, and the electrical state quantities are updated.

[0119] In the thermal diffusion constraint layer, a dual-path dynamic conduction method is used to update the node temperature state:

[0120] On the one hand, conventional conduction updates are performed based on the temperature difference between adjacent nodes, node heat capacity, and thermal conductivity.

[0121] On the other hand, a degradation-sensitive path is introduced to add additional thermal resistance compensation and heat source correction to the components characterized by degradation feature vectors;

[0122] A temperature update logic is formed that includes normal conduction and degradation compensation, and the updated thermal state variables are output.

[0123] A multi-head attention mechanism is introduced to construct the input features of each component node. The input features consist of node observation data, the degradation feature vector corresponding to the component node, and the load level and environmental conditions. The information of adjacent nodes is weighted and aggregated according to the attention weight to generate the node representation.

[0124] Under the combined constraints of the current conservation constraint layer and the thermal diffusion constraint layer, the system takes node representation as input and outputs component-level health index, component degradation parameter estimates, and state prediction results of temperature, partial discharge, and failure probability within the future time window.

[0125] In this embodiment, generating risk change values ​​and maintenance cost values ​​includes:

[0126] Obtain a snapshot of the digital twin model of the power distribution cabinet at the current moment, and synchronously read the registered differential status channels;

[0127] The target component set is selected based on the health index and differential status channel. The maintenance action library is called to define the maintenance action set to be evaluated. The maintenance action set includes four types of actions: parameter reset, replacement, cleaning and tightening. In the parameter update rule library, the corresponding degradation parameter update rules and maintenance cost descriptions are specified for each target component and each maintenance action.

[0128] On the digital twin model snapshot of the power distribution cabinet, a counterfactual simulation branch is created for each target component and each maintenance action. The set time step and boundary conditions are kept unchanged. The degradation parameters of the target components are updated according to the parameter update rules. The digital twin model of the power distribution cabinet is run to generate the predicted trajectory of temperature, partial discharge and failure probability after maintenance.

[0129] Based on preset risk assessment rules, the predicted trajectories of temperature, partial discharge, and failure probability before and after maintenance are converted into comparable risk scores. The risk reduction amount for each counterfactual simulation branch is calculated. Combined with the maintenance cost description and the set cost weights, a benefit score is generated. Specifically, the calculation of the risk reduction amount for each counterfactual simulation branch is as follows:

[0130] The temperature peak, partial discharge amplitude, and failure probability sequences were extracted from the pre-maintenance and post-maintenance predicted trajectories, and converted into corresponding risk scores.

[0131] The risk change value is obtained by calculating the difference between the risk score before and after maintenance for the same component;

[0132] The change in risk is recorded as the risk reduction amount of this counterfactual simulation branch;

[0133] For each target component, the maintenance action with the highest benefit score is selected from the candidate maintenance actions to form a maintenance plan package that includes component identification, selected maintenance action, expected risk reduction, estimated maintenance cost, post-maintenance predicted trajectory summary, required resources and expected time window.

[0134] In this embodiment, the generation of the optimal scheduling and execution strategy for predictive maintenance of the power distribution cabinet includes:

[0135] Based on the maintenance plan package, the risk change value and maintenance cost value of each target component are sorted by benefit, forming a candidate job list and initial execution order arranged by priority;

[0136] The candidate job list is matched and verified with the operation and maintenance time window, spare parts supply and power grid reliability constraints. Specific time periods and resources are allocated to each target component according to the constraint feasibility. If the constraints are not met, the execution order, maintenance actions or job batches are adjusted, and the optimal schedule is output.

[0137] Execution strategies are generated based on the optimal schedule, and specific maintenance actions, schedules and resource requirements are determined according to component priority and resource configuration results.

[0138] Example 1:

[0139] To verify the feasibility of this invention in practice, it was applied to the power supply workshop of a large manufacturing enterprise. The distribution cabinets operate under high load and complex environments for extended periods, and maintenance personnel frequently encounter two prominent problems: first, monitoring data is easily affected by load fluctuations and changes in environmental humidity, leading to significant fluctuations in health assessment indicators and making it difficult to reflect the true degradation status; second, existing maintenance strategies mainly rely on experience and single threshold judgments, often resulting in over-maintenance or delayed maintenance, increasing maintenance costs and posing safety hazards. Therefore, this embodiment introduces a predictive maintenance method based on digital twins and intelligent sensing on the five main distribution cabinets in the workshop to comprehensively monitor and analyze the equipment's health status.

[0140] In the application scenario, multimodal intelligent sensors were first deployed within the distribution cabinet to collect multi-source operational data, including busbar and terminal temperature, current, voltage, partial discharge signals, humidity, and real-time load levels. The data collection cycle was set to once per second, and time synchronization, noise filtering, and normalization were performed by edge computing devices to form a standardized input dataset. After three months of continuous operation and monitoring, over 30 million valid data points were collected. Among these data, the busbar temperature peaked at 92℃ during high-load periods in summer, while the highest ambient temperature was 38℃, and humidity varied between 65% and 90%. Traditional methods, relying solely on temperature thresholds, frequently resulted in false alarms, especially when humidity was high, as the partial discharge signal fluctuated drastically, leading to a false alarm rate exceeding 15%.

[0141] To overcome this problem, the method of this invention further constructs a causal degradation map, taking factors such as increased contact resistance, thermal interface aging, and insulation loss as degradation factors, load level and environmental conditions as external operating condition factors, and temperature rise, partial discharge amplitude, and current fluctuation as observed signals. Through causal correlation and invariant constraint methods, degradation characteristics that remain stable under different operating conditions are extracted.

[0142] In the digital twin modeling process, an electrical-thermal dual-domain model of the distribution cabinet was established. Busbars, terminals, circuit breakers, and insulation components were designated as nodes, while electrical coupling and thermal conduction relationships were defined as edges. Current conservation and thermal diffusion constraints were embedded, and a multi-head attention mechanism was introduced, enabling the model to dynamically adjust feature weights under different operating conditions. In a three-month predictive experiment, the twin model's average prediction error for busbar temperature was 1.8℃, significantly lower than the 5.7℃ error of the traditional linear regression model. The prediction of circuit breaker insulation degradation probability showed a 93.5% agreement rate with actual fault inspection results, an improvement of over 12% compared to the model without dual-domain constraints and the attention mechanism.

[0143] In terms of maintenance strategies, counterfactual maintenance simulators are introduced into the twin model to perform virtual "parameter reset, replacement, or cleaning" operations on target components. For example, performing a "parameter reset" on busbar connection points, that is, restoring the contact resistance to a healthy baseline state, the simulator predicts that the peak temperature rise can decrease from 92°C to 78°C within the next month, and the failure rate can be reduced from 12% to 3%. Similarly, performing a "replacement" operation on insulation components predicts that the partial discharge amplitude can decrease from an average of 23dB to 8dB, and the failure rate can be reduced from 9% to less than 2%. These quantitative indicators provide maintenance personnel with a clear cost-risk comparison, making maintenance no longer dependent on a single threshold.

[0144] Ultimately, based on the risk change value and maintenance cost value, combined with the factory's operation and maintenance time window and spare parts supply, the system generated the optimal maintenance schedule.

[0145] Table 1. Statistical Table of Predictive Maintenance Experiment Data

[0146]

[0147] As shown in Table 1, under traditional methods, the false alarm rate for risk assessment of various components in the distribution cabinet is generally high. The false alarm rate for insulation components reaches 20%, while busbars and circuit breakers reach 15% and 18%, respectively. This high false alarm rate indicates that existing methods relying on single thresholds or static monitoring are easily affected by load and environmental fluctuations, leading to inaccurate health status assessments. However, the method of this invention, by extracting stable degradation features through causal degradation maps and invariant constraints, significantly reduces the feature variance of each component, reaching a minimum of 0.02 and a maximum of only 0.05. This improves the stability of degradation features and provides a more reliable basis for health assessment.

[0148] Regarding prediction accuracy, this method embeds current conservation and thermal diffusion constraints into the digital twin model of the distribution cabinet, and combines this with a multi-head attention mechanism to make the prediction results more consistent with actual operating conditions. Experimental results show that the average prediction errors for busbars, terminals, circuit breakers, and insulation components are controlled between 1.8℃ and 2.7℃, while the prediction errors of traditional models typically exceed 5℃. This significant improvement in accuracy not only makes the predictions of health indices and degradation trends more reliable, but also enhances the interpretability and operability of the prediction results, providing a solid basis for maintenance decisions.

[0149] Regarding maintenance effectiveness, this invention uses a counterfactual maintenance simulator to calculate the risk changes and costs before and after maintenance, achieving a quantitative assessment of the maintenance measures' effectiveness. Data shows that by resetting parameters on the busbar, the risk value decreased from 12% to 3%, with a cost of only 300 yuan; by replacing insulation components, the risk value decreased from 9% to 2%, although the cost was higher, the risk reduction was also the greatest. Overall, the average risk of the entire distribution cabinet decreased from 11.5% to 3.5%, a risk reduction of 8%, while the total maintenance cost was only 2100 yuan, saving approximately 25% compared to traditional experience-based decision-making methods. This invention achieves an effective balance between risk reduction and cost control, possessing significant economic and reliability advantages.

[0150] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A predictive maintenance system for the health status of a power distribution cabinet based on digital twins and intelligent sensing, characterized in that, Includes the following modules: The data acquisition and preprocessing module is used to collect and preprocess multi-source operational data, generate a standardized input dataset, and divide data samples into multiple operating conditions. The causal degradation feature extraction module is used to construct a causal degradation map, extract stable degradation features, and generate degradation feature vectors. The digital twin modeling module is used to build a digital twin model of the power distribution cabinet and establish the topology of the power distribution cabinet diagram. The digital twin modeling module constructs a digital twin model of the power distribution cabinet, including: Based on the normalized input dataset and degenerate feature vector, a digital twin model of the distribution cabinet is established. A hierarchical heterogeneous multi-graph structure is adopted to define the graph structure topology. The equipment layer is used to represent circuits and segments, the component layer is used to represent busbars, terminals, circuit breakers and insulation components, and the sensing layer is used to represent sensing channels. The connection relationship includes electrical coupling relationship and thermal conduction relationship. Establish a mapping relationship between sensing channels and component nodes, register the data in the normalized input dataset as node observation vectors in chronological order, establish a mapping relationship between degradation feature vectors and corresponding component nodes, and create a component parameter registry to record geometric parameters, material parameters and rated parameters; In the digital twin model of the power distribution cabinet, the electric domain kernel and the thermal domain kernel are instantiated, a shared degradation state interface is set to realize dual-domain coupling, and physical constraint interface, boundary condition interface and cross-domain coupling mapping table are configured to identify the correspondence between electric domain nodes and thermal domain nodes. Register the basic initial state for the digital twin model of the power distribution cabinet. The basic initial state includes the initial electrical state and initial thermal state of the component nodes, the load level and environmental condition occupancy data under the current operating conditions, register the topology version number and connection relationship verification information, and set the calculation step size and update order for state updates. A degraded mirror node generation module is loaded into the digital twin model of the power distribution cabinet to generate a healthy baseline mirror node for each component node in the component layer. The healthy baseline state is set according to the component parameter registry and the registered operating condition data. A one-to-one mapping relationship between the mirror node and the component node is established. The difference between the component node and the healthy baseline state is recorded through differential links. The digital twin model of the power distribution cabinet carrying the mirror node and differential state channels is output. The constraint fusion prediction module is used to embed a current conservation constraint layer and a heat diffusion constraint layer into the digital twin model of the power distribution cabinet, and output the component health index, degradation parameter estimate and future state prediction results. The counterfact maintenance simulation module is used to introduce a counterfact maintenance simulator into the digital twin model, perform counterfactual operations such as parameter reset, replacement or cleaning, and generate risk change values ​​and maintenance cost values. The scheduling and strategy generation module is used to generate optimal scheduling and execution strategies by combining maintenance time windows, spare parts supply and power grid reliability constraints.

2. A predictive maintenance method for the health status of a distribution cabinet based on digital twins and intelligent sensing, applied to the predictive maintenance system for the health status of a distribution cabinet based on digital twins and intelligent sensing as described in claim 1, characterized in that, include: Multi-source operating data from multimodal smart sensors installed inside the power distribution cabinet is collected, preprocessed, and a standardized input dataset is generated. Data samples are then divided into multiple operating conditions according to load level and environmental conditions. Based on the normalized input dataset, a causal degradation map is constructed. The invariant constraint method of multi-condition data samples is used to extract degradation features that remain stable under different loads and environmental conditions, forming a degradation feature vector. Based on the normalized input dataset and degenerate feature vectors, a digital twin model of the distribution cabinet is constructed. The busbars, terminals, circuit breakers, and insulation components are used as nodes of the graph structure, and the electrical coupling relationship and heat conduction relationship are used as edges of the graph structure to establish the graph structure topology of the distribution cabinet. A current conservation constraint layer and a heat diffusion constraint layer are embedded in the digital twin model of the power distribution cabinet, and a multi-head attention mechanism is introduced. The degradation feature vector is used as input and combined with the normalized input dataset to output the health index, degradation parameter estimate and state prediction results of each component of the power distribution cabinet within the future time window. A counterfactual maintenance simulator is introduced into the digital twin model of the power distribution cabinet. Counterfactual operations such as parameter reset, replacement or cleaning are performed on the target components. The temperature rise change, partial discharge evolution and failure probability before and after maintenance are calculated respectively, and risk change value and maintenance cost value are generated. Based on the risk change value and maintenance cost value, combined with the operation and maintenance time window, spare parts supply and power grid reliability constraints, the optimal scheduling and execution strategy for predictive maintenance of the distribution cabinet is generated.

3. The method for predictive maintenance of distribution cabinet health status based on digital twin and intelligent sensing according to claim 2, characterized in that, The multimodal smart sensor includes a temperature sensor, a current sensor, a voltage sensor, a partial discharge sensor, and a humidity sensor.

4. The method for predictive maintenance of distribution cabinet health status based on digital twin and intelligent sensing according to claim 2, characterized in that, The multi-source operating data includes temperature data, current data, voltage data, partial discharge data, and humidity data. The preprocessing of the multi-source operating data includes time synchronization processing, noise filtering processing, and normalization processing.

5. A predictive maintenance method for the health status of a power distribution cabinet based on digital twins and intelligent sensing, as described in claim 2, is characterized in that... The formation of the degenerate feature vector includes: It receives a normalized input dataset and multi-condition data samples. Each condition domain contains temperature data, current data, voltage data, partial discharge data, and humidity data, as well as corresponding load levels, ambient temperature, and ambient humidity labels. Constructing a causal degradation map: The degradation factor nodes are defined as the increase in contact resistance, the increase in thermal interface resistance, and the increase in insulation loss factor. Set the load level, ambient temperature, and ambient humidity as external operating condition factor nodes; Busbar temperature rise, terminal temperature rise, partial discharge amplitude, current fluctuation characteristics, and voltage deviation are set as observation signal nodes; Establish directed causal relationships between degradation factor nodes and observation signal nodes, and between external condition factor nodes and observation signal nodes, and determine the direction and strength parameters of each causal relationship. Based on multi-condition data samples and causal degradation maps, cross-condition invariant constraint feature extraction is performed. Consistency constraints are applied to the degradation-related feature statistics of each condition domain and the cross-domain overall statistics to ensure that the deviation of the mean and covariance of degradation-related features of each condition domain from the cross-domain overall statistics does not exceed the preset range, thus obtaining a degradation candidate feature set. The degraded candidate feature set is subjected to distribution alignment processing. Alignment is performed using a distribution difference metric to ensure that the feature distribution difference between any two working domains does not exceed a preset threshold, thus forming a degraded feature set. The set of degradation features consistent across operating conditions is aggregated according to the mapping relationship between the sensing channel and the distribution cabinet components to form a degradation feature vector.

6. The method for predictive maintenance of distribution cabinet health status based on digital twin and intelligent sensing according to claim 2, characterized in that, The health index, degradation parameter estimates, and state prediction results for each component of the output distribution cabinet within the future time window include: In the digital twin model of the power distribution cabinet, a current conservation constraint layer and a heat diffusion constraint layer are loaded, and the normalized input dataset and degenerate feature vector are read. In the current conservation constraint layer, for each component node, the algebraic sum of the inflow current and outflow current is constrained to be zero. For each connection relationship, the voltage drop is determined based on the component's equivalent resistance and the current passing through it. The contact resistance, thermal interface-related electrical parameters, and insulation loss-related electrical parameters obtained by mapping the degradation feature vector are recorded, and the electrical state quantities are updated. In the thermal diffusion constraint layer, a dual-path dynamic conduction method is used to update the node temperature state: On the one hand, conventional conduction updates are performed based on the temperature difference between adjacent nodes, node heat capacity, and thermal conductivity. On the other hand, a degradation-sensitive path is introduced to add additional thermal resistance compensation and heat source correction to the components characterized by degradation feature vectors; A temperature update logic is formed that includes normal conduction and degradation compensation, and the updated thermal state variables are output. A multi-head attention mechanism is introduced to construct the input features of each component node. The input features consist of node observation data, the degradation feature vector corresponding to the component node, and the load level and environmental conditions. The information of adjacent nodes is weighted and aggregated according to the attention weight to generate the node representation. Under the combined constraints of the current conservation constraint layer and the thermal diffusion constraint layer, the system takes node representation as input and outputs component-level health index, component degradation parameter estimates, and state prediction results of temperature, partial discharge, and failure probability within the future time window.

7. A predictive maintenance method for the health status of a power distribution cabinet based on digital twins and intelligent sensing, as described in claim 2, is characterized in that... The generation of risk change values ​​and maintenance cost values ​​includes: Obtain a snapshot of the digital twin model of the power distribution cabinet at the current moment, and synchronously read the registered differential status channels; The target component set is selected based on the health index and differential status channel. The maintenance action library is called to define the maintenance action set to be evaluated. The maintenance action set includes four types of actions: parameter reset, replacement, cleaning and tightening. In the parameter update rule library, the corresponding degradation parameter update rules and maintenance cost descriptions are specified for each target component and each maintenance action. On the digital twin model snapshot of the power distribution cabinet, a counterfactual simulation branch is created for each target component and each maintenance action. The set time step and boundary conditions are kept unchanged. The degradation parameters of the target components are updated according to the parameter update rules. The digital twin model of the power distribution cabinet is run to generate the predicted trajectory of temperature, partial discharge and failure probability after maintenance. Based on the preset risk assessment rules, the predicted trajectories of temperature, partial discharge and failure probability before and after maintenance are converted into comparable risk scores. The risk reduction of each counterfactual simulation branch is calculated. Combined with the maintenance cost description and the set cost weights, a benefit score is generated. For each target component, the maintenance action with the highest benefit score is selected from the candidate maintenance actions to form a maintenance plan package that includes component identification, selected maintenance action, expected risk reduction, estimated maintenance cost, post-maintenance predicted trajectory summary, required resources and expected time window.

8. A method for predictive maintenance of the health status of a power distribution cabinet based on digital twins and intelligent sensing, as described in claim 2, is characterized in that... The optimal scheduling and execution strategy for generating predictive maintenance of the distribution cabinet includes: Based on the maintenance plan package, the risk change value and maintenance cost value of each target component are sorted by benefit, forming a candidate job list and initial execution order arranged by priority; The candidate job list is matched and verified with the operation and maintenance time window, spare parts supply and power grid reliability constraints. Specific time periods and resources are allocated to each target component according to the constraint feasibility. If the constraints are not met, the execution order, maintenance actions or job batches are adjusted, and the optimal schedule is output. Execution strategies are generated based on the optimal schedule, and specific maintenance actions, schedules and resource requirements are determined according to component priority and resource configuration results.

Images