Insulation Status Assessment Method for Tubular Busbars Based on Multi-Feature Fusion

Abstract

This invention discloses a method for assessing the insulation condition of tubular busbars based on multi-feature fusion, relating to the field of power equipment insulation condition detection technology. The invention includes deploying sensor arrays at various locations on the tubular busbar to synchronously collect multi-source heterogeneous feature quantities. The collected multi-source heterogeneous feature quantities are then time-synchronized and spatially mapped to obtain aligned multi-source heterogeneous feature data. A feature association graph is used as a priori structural input graph attention network to perform deep fusion of the aligned multi-source heterogeneous feature data, generating a fused feature vector and outputting the attention weights between each feature quantity. Based on the fused feature vector, a Bayesian network-based structural causal model is constructed to learn the causal structure between variables. Through a counterfactual reasoning framework, the multi-source heterogeneous feature data is classified as operating condition-driven components and insulation degradation contribution components. This invention solves the problems of fragmented multi-source data, insulation degradation, and difficulty in identifying the root cause of faults in existing technologies.

Description

Technical Field

[0001] This invention relates to the field of electrical equipment insulation condition detection technology, and in particular to a method for evaluating the insulation condition of tubular busbars based on multi-feature fusion. Background Technology

[0002] Insulated tubular busbars have been widely used in power transmission in power plants and other fields due to their advantages such as large current carrying capacity, compact structure and high reliability. However, due to the lack of unified industry technical standards, the production and installation processes of equipment vary, insulation failures occur frequently, and conventional acceptance tests are difficult to effectively detect internal defects. Once a failure occurs, it often causes large-scale power outages and significant economic losses.

[0003] Currently, insulation condition detection of operating tubular busbars mainly relies on single methods such as infrared thermography, partial discharge detection, and ultraviolet imaging. Infrared thermography can only detect faults that have already overheated, making it difficult to predict early deterioration, and it is easily affected by dust and oil. In partial discharge detection, the ultra-high frequency method has complex propagation in metal-shielded environments and is difficult to arrange sensors, while the ultrasonic method is easily drowned out by strong background noise in industrial plants. Some detection methods integrate multiple detection methods, but the data acquisition time of various types is not synchronized and the space cannot be aligned, which makes the judgment rely on human experience. Furthermore, existing methods that integrate multiple single test methods cannot distinguish between changes caused by fluctuations in normal operating conditions and changes caused by insulation deterioration itself, and it is also difficult to identify the root cause of insulation deterioration under multiple stress coupling effects. Summary of the Invention

[0004] The purpose of this invention is to provide a method for evaluating the insulation status of tubular busbars based on multi-feature fusion, which solves the problems of fragmented multi-source detection data, insufficient feature fusion depth, inability to distinguish between changes caused by fluctuations in normal operating conditions and insulation degradation, and difficulty in identifying the root cause of insulation degradation in the prior art.

[0005] To solve the above-mentioned technical problems, the present invention is achieved through the following technical solution: This invention relates to a method for evaluating the insulation status of tubular busbars based on multi-feature fusion, comprising the following steps: S1. Deploy sensor arrays at various locations on the tubular busbar to synchronously collect multi-source heterogeneous characteristic quantities, including electrical characteristic quantities, thermal characteristic quantities, vibration characteristic quantities and environmental characteristic quantities; S2. Perform time synchronization calibration and spatial coordinate mapping on the collected multi-source heterogeneous feature quantities to obtain aligned multi-source heterogeneous feature data. Then, based on the multi-physics coupling mechanism between electrical, thermal, vibration and environmental features, construct a feature association graph with feature quantities as nodes and physical causal action directions as directed edges, and set constraint rules including time, threshold and direction for each directed edge. S3. The feature association graph is used as the prior structure input to the graph attention network to perform deep fusion on the aligned multi-source heterogeneous feature data, generate a fused feature vector, and output the attention weights between each feature quantity. The attention coefficients of the graph attention network are adaptively adjusted according to the current load current and ambient temperature. S4. Construct a structural causal model in the form of a Bayesian network based on the fused feature vectors, learn the causal structure between variables using historical operating data, and use the counterfactual reasoning framework to convert the amount of multi-source heterogeneous feature data into operating condition driving components and insulation degradation contribution components. S5. Based on the insulation degradation contribution component and the attention weight, assess the current insulation status level of the tubular busbar and identify the dominant degradation mode.

[0006] Furthermore, in step S1, the electrical characteristics include partial discharge pulse signals and grounding current; the thermal characteristics include bus surface temperature distribution and temperature rise rate; the vibration characteristics include triaxial vibration spectrum and impact pulse amplitude; and the environmental characteristics include ambient temperature and humidity, dust concentration, and corrosive gas concentration. The sensor array includes an ultra-high frequency partial discharge sensor, a high frequency current transformer, a distributed fiber optic temperature sensor or an infrared thermal imager, a triaxial accelerometer, and a multi-parameter environmental sensor.

[0007] Furthermore, spatial coordinate mapping transforms the installation positions of each sensor to a unified coordinate system by establishing a three-dimensional spatial model of the tubular busbar, and interpolates and resamples the feature quantities with different sampling rates to make them have the same time interval on the time axis.

[0008] The specific process of constructing the feature association graph in step S2 is as follows: S2.1. Each of the collected electrical, thermal, vibration, and environmental characteristics is considered a node, forming a node set. S2.2. Based on the multi-physics coupling characteristics between electric field, thermal field, magnetic field and stress field during the operation of insulated tubular busbar, determine the causal relationship between each characteristic quantity, and construct a set with the direction of physical causal interaction as the directed edge. Electrical and thermal coupling edge: A bidirectional causal relationship is established between the partial discharge signal and temperature; Thermal-mechanical coupling: A one-way causal relationship is established between the temperature gradient and the vibration and shock pulse; Mechanical and electrical coupling: One-way causal relationships are established between the vibration spectrum and the grounding current, and between the vibration impact pulse and the partial discharge signal, respectively; Environmental and electrical coupling edges: One-way causal relationships are established between environmental humidity and grounding current, and between dust concentration and partial discharge signal, respectively; Environmental and thermal coupling: Establish unidirectional causal relationships between ambient temperature and temperature rise rate, and between corrosive gas concentration and temperature distribution anomaly, respectively; S2.3 For each directed edge, set constraint rules according to the physical mechanism. The constraint rules include time constraints, threshold constraints and direction constraints. Time constraints are used to specify the response time range of causal action. Threshold constraints are used to specify the numerical conditions of characteristic quantities for the occurrence of causal action. Direction constraints are used to specify the unidirectional and bidirectional nature of causal action.

[0009] Furthermore, the attention coefficients of the graph attention network in step S3 are calculated as follows: For node i and its neighbor node j, the attention coefficient ,in, , Let W be the node feature vector, and W be the weight matrix. This is a learnable attention parameter vector. The activation function is used; then the attention weights are obtained by normalization using the Softmax function. ; Adaptive adjustment refers to: Let exp be the set of neighboring nodes of node i, where exp is the exponential function and the attention coefficient is the sum of the neighboring nodes of node i. Introducing load current I and ambient temperature As a regulatory factor, i.e. ,in This is the operating condition encoding vector obtained by mapping load current and ambient temperature.

[0010] Furthermore, in step S4, the counterfactual reasoning framework decomposes the measured feature quantity X into operating condition-driven components. and insulation degradation contribution : in, Operating condition drive component, where I is the load current. For ambient temperature, The intensity of the external vibration source, The operating condition driving function is obtained through structural causal model learning; Insulation degradation contribution , The estimated operating conditions predicted under counterfactual intervention using a structural causal model; Counterfactual intervention specifically involves: setting intervention variables. Calculate the causal effect size , This represents the conditional expectation after do intervention in the structural causal model. Aging represents the latent variable of insulation aging that cannot be directly observed. Normal represents the normal state, and abnormal represents the abnormal state.

[0011] Furthermore, in step S5, the insulation condition level includes five levels: good, attention, abnormal, serious, and critical, which are determined based on the comprehensive score of the amplitude and trend of the insulation degradation contribution component and the attention weight. The method for identifying the dominant degradation mode is as follows: calculate the score for each degradation mode, and the mode with the highest score is the current dominant degradation mode; the score of the k-th mode... The calculation formula is: in, This represents the set of features corresponding to the k-th degradation mode. For set The number of features in the middle The attention weights for the i-th feature output in step S3 are... The insulation degradation contribution component is obtained for the i-th feature quantity decomposed in step S4. The degradation modes include thermal aging-dominated mode, vibration fatigue-dominated mode, electrical stress-dominated mode, and environmental corrosion-dominated mode. Among them, the characteristic quantity set corresponding to the thermal aging-dominated mode includes temperature and temperature rise rate; the characteristic quantity set corresponding to the vibration fatigue-dominated mode includes vibration spectrum and impact pulse amplitude; the characteristic quantity set corresponding to the electrical stress-dominated mode includes partial discharge pulse signal and grounding current; and the characteristic quantity set corresponding to the environmental corrosion-dominated mode includes ambient temperature and humidity, dust concentration, and corrosive gas concentration.

[0012] Furthermore, the construction and updating of the structural causal model in step S4 adopts an adaptive time-varying mechanism, specifically including the following steps: S4.1. Using historical data from the initial commissioning period of the tubular busbar when its insulation condition is good, construct an initial structural causal model. Including cause-effect graph structure and the conditional probability parameters between variables ; S4.2 Set the time window length L. Every L time interval, using the running data within the current time window, re-estimate the causal effect strength corresponding to each directed edge in the feature association graph; for any directed edge... The strength of its causal effect Defined as: Where A represents the causal characteristic and B represents the outcome characteristic. As the baseline value of A, For the preset intervention increment, This indicates that an intervention is being performed in the outcome-causal model. Expressing expectations; S4.3 Calculate the rate of change of the intensity of the causal effect: in, The strength of the causal effect in the previous time window. If is a non-zero small constant, Exceeding the preset threshold If so, then the directed edge is determined to have significant causal drift; S4.4 For directed edges exhibiting significant causal drift, the following update operations are performed sequentially: First, using the current window data, the conditional probability parameter of the edge is updated through Bayesian online learning; then, if the drift of the edge persists for more than two time windows, a conditional independence test is further performed, and a decision is made based on the test result to delete the directed edge; the updated structural causal model It replaces the current model and is used for subsequent counterfactual decomposition.

[0013] Furthermore, it also includes insulation remaining life prediction step S4.5: The calculated rate of change of causal effect intensity of each causal edge in different time windows A drift trend sequence for each causal edge is constructed. For the key causal edge corresponding to the dominant degradation mode, a time series prediction model is used to extrapolate its future trend, predicting that the causal effect strength of the causal edge first exceeds a preset danger threshold. The difference between the current time and the specified time point is used as the estimated remaining insulation lifetime (RUL). Time series forecasting models employ exponential smoothing or linear regression, and their forecasting formulas are as follows: in, For the future Predicted value of the causal effect strength at time t. The strength of the causal effect at the current moment. The average rate of change over the past k-th time window. is the preset decay weight coefficient, and m is the number of backtracking windows; The formula for calculating the remaining useful life (RUL) is: Where c is the set of key causal edges, Let be the predicted value of the i-th causal edge. The preset danger threshold corresponding to this edge. Prediction time required to reach the threshold; Finally, the remaining lifetime estimate, RUL machine confidence interval, is output.

[0014] The present invention has the following beneficial effects: This invention solves the problems of data fragmentation and difficulty in correlation analysis in existing detection methods by synchronously collecting four-dimensional features of electrical, thermal, vibration and environmental data and aligning them in time and space, thus providing a complete information foundation for comprehensive evaluation.

[0015] This invention constructs a feature association graph based on physical causal relationships and uses it as the graph structure of a graph attention network. This makes the fusion process constrained by the physical laws of insulation aging, which is different from the pure data-driven black box model. This improves the interpretability of the evaluation results. At the same time, the attention coefficient is adaptively adjusted according to the load current and ambient temperature, which enhances the adaptability under different operating conditions.

[0016] This invention introduces a structural causal model and counterfactual reasoning to decompose measured characteristic quantities into operating condition-driven components and insulation degradation contribution components, effectively distinguishing between normal operating condition fluctuations and true insulation degradation, significantly reducing false alarm rates, and solving the technical problem that existing methods cannot quantify the degree of degradation.

[0017] This invention combines attention weights and degradation contribution components to identify dominant degradation modes such as thermal aging, vibration fatigue, electrical stress, and environmental corrosion, and outputs root cause diagnosis results and targeted operation and maintenance suggestions. It changes the limitation of existing technologies that only output health scores and realizes a leap from condition assessment to root cause diagnosis. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0020] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0021] See Figure 1 This invention provides a method for evaluating the insulation status of tubular busbars based on multi-feature fusion, comprising the following steps: S1. Deploy sensor arrays at various locations on the tubular busbar to synchronously collect multi-source heterogeneous characteristic quantities, including electrical characteristic quantities, thermal characteristic quantities, vibration characteristic quantities and environmental characteristic quantities; In step S1, the electrical characteristics include partial discharge pulse signals and grounding current; the thermal characteristics include bus surface temperature distribution and temperature rise rate; the vibration characteristics include triaxial vibration spectrum and impact pulse amplitude; and the environmental characteristics include ambient temperature and humidity, dust concentration, and corrosive gas concentration. The sensor array includes an ultra-high frequency partial discharge sensor, a high frequency current transformer, a distributed fiber optic temperature sensor or an infrared thermal imager, a triaxial accelerometer, and a multi-parameter environmental sensor.

[0022] S2. Perform time synchronization calibration and spatial coordinate mapping on the collected multi-source heterogeneous feature quantities to obtain aligned multi-source heterogeneous feature data. Then, based on the multi-physics coupling mechanism between electrical, thermal, vibration and environmental features, construct a feature association graph with feature quantities as nodes and physical causal action directions as directed edges, and set constraint rules including time, threshold and direction for each directed edge. Spatial coordinate mapping transforms the installation positions of each sensor to a unified coordinate system by establishing a three-dimensional spatial model of the tubular busbar, and interpolates and resamples feature quantities with different sampling rates to make them have the same time interval on the time axis.

[0023] The specific process of constructing the feature association graph in step S2 is as follows: S2.1. Each of the collected electrical, thermal, vibration, and environmental characteristics is considered a node, forming a node set. S2.2. Based on the multi-physics coupling characteristics between electric field, thermal field, magnetic field and stress field during the operation of insulated tubular busbar, determine the causal relationship between each characteristic quantity, and construct a set with the direction of physical causal interaction as the directed edge. Electrical and thermal coupling edge: A bidirectional causal relationship is established between the partial discharge signal and temperature; Thermal-mechanical coupling: A one-way causal relationship is established between the temperature gradient and the vibration and shock pulse; Mechanical and electrical coupling: One-way causal relationships are established between the vibration spectrum and the grounding current, and between the vibration impact pulse and the partial discharge signal, respectively; Environmental and electrical coupling edges: One-way causal relationships are established between environmental humidity and grounding current, and between dust concentration and partial discharge signal, respectively; Environmental and thermal coupling: Establish unidirectional causal relationships between ambient temperature and temperature rise rate, and between corrosive gas concentration and temperature distribution anomaly, respectively; S2.3 For each directed edge, set constraint rules according to the physical mechanism. The constraint rules include time constraints, threshold constraints and direction constraints. Time constraints are used to specify the response time range of causal action. Threshold constraints are used to specify the numerical conditions of characteristic quantities for the occurrence of causal action. Direction constraints are used to specify the unidirectional and bidirectional nature of causal action.

[0024] S3. The feature association graph is used as the prior structure input to the graph attention network to perform deep fusion on the aligned multi-source heterogeneous feature data, generate a fused feature vector, and output the attention weights between each feature quantity. The attention coefficients of the graph attention network are adaptively adjusted according to the current load current and ambient temperature. The attention coefficient of the graph attention network in step S3 is calculated as follows: For node i and its neighbor node j, the attention coefficient ,in, , Let W be the node feature vector, and W be the weight matrix. This is a learnable attention parameter vector. The activation function is used; then the attention weights are obtained by normalization using the Softmax function. ; Adaptive adjustment refers to: Let exp be the set of neighboring nodes of node i, where exp is the exponential function and the attention coefficient is the sum of the neighboring nodes of node i. Introducing load current I and ambient temperature As a regulatory factor, i.e. ,in This is the operating condition encoding vector obtained by mapping load current and ambient temperature.

[0025] S4. Construct a structural causal model in the form of a Bayesian network based on the fused feature vectors, learn the causal structure between variables using historical operating data, and use the counterfactual reasoning framework to convert the amount of multi-source heterogeneous feature data into operating condition driving components and insulation degradation contribution components. In step S4, the counterfactual reasoning framework decomposes the measured feature quantity X into operating condition-driven components. and insulation degradation contribution : in, Operating condition drive component, where I is the load current. For ambient temperature, The intensity of the external vibration source, The operating condition driving function is obtained through structural causal model learning; Insulation degradation contribution , The estimated operating conditions predicted under counterfactual intervention using a structural causal model; Counterfactual intervention specifically involves: setting intervention variables. Calculate the causal effect size , This represents the conditional expectation after do intervention in the structural causal model. Aging represents the latent variable of insulation aging that cannot be directly observed. Normal represents the normal state, and abnormal represents the abnormal state.

[0026] S5. Based on the insulation degradation contribution component and the attention weight, assess the current insulation status level of the tubular busbar and identify the dominant degradation mode.

[0027] In step S5, the insulation condition level includes five levels: good, attention, abnormal, serious, and critical. It is determined based on the comprehensive score of the amplitude and trend of the insulation degradation contribution component and the attention weight. The method for identifying the dominant degradation mode is as follows: calculate the score for each degradation mode, and the mode with the highest score is the current dominant degradation mode; the score of the k-th mode... The calculation formula is: in, This represents the set of features corresponding to the k-th degradation mode. For set The number of features in the middle The attention weights for the i-th feature output in step S3 are... The insulation degradation contribution component is obtained for the i-th feature quantity decomposed in step S4. The degradation modes include thermal aging-dominated mode, vibration fatigue-dominated mode, electrical stress-dominated mode, and environmental corrosion-dominated mode. Among them, the characteristic quantity set corresponding to the thermal aging-dominated mode includes temperature and temperature rise rate; the characteristic quantity set corresponding to the vibration fatigue-dominated mode includes vibration spectrum and impact pulse amplitude; the characteristic quantity set corresponding to the electrical stress-dominated mode includes partial discharge pulse signal and grounding current; and the characteristic quantity set corresponding to the environmental corrosion-dominated mode includes ambient temperature and humidity, dust concentration, and corrosive gas concentration.

[0028] The working principle of this invention is as follows: First, multiple types of sensors are deployed at key locations on the tubular busbar to simultaneously collect four types of characteristic quantities: electrical, thermal, vibration, and environmental. Electrical characteristics include partial discharge pulse signals and grounding current; thermal characteristics include surface temperature distribution and temperature rise rate; vibration characteristics include triaxial vibration spectrum and impact pulse amplitude; and environmental characteristics include temperature, humidity, dust concentration, and corrosive gas concentration. All sensors are triggered by the same clock source to ensure strict time alignment of the data. Subsequently, the collected data is spatially mapped to transform the sensor data from different installation locations into a unified three-dimensional spatial model. Signals with different sampling rates are interpolated and resampled to obtain aligned multi-source characteristic data. Based on the physical causal relationship between electrical, thermal, vibration, and environmental factors, a feature correlation graph is constructed: each characteristic quantity is a node, and the direction of physical causal action is a directed edge. For example, vibration causes loosening of connections, which in turn causes temperature rise; the temperature rise accelerates insulation aging, which in turn strengthens the partial discharge pulse signal. The feature association graph is used as the graph structure of the graph attention network. The network adaptively adjusts the attention coefficient according to the current load current and ambient temperature, performs deep fusion on the aligned multi-source features, generates a fused feature vector, and outputs the attention weights between each feature quantity. After fusion, a structural causal model in the form of a Bayesian network is constructed based on the fused feature vector. The causal structure between variables is learned using historical operating data, and each measured feature quantity is decomposed into a working condition driving component and an insulation degradation contribution component through counterfactual reasoning: the working condition driving component is caused by load current, ambient temperature and external vibration sources, and the insulation degradation contribution component is caused by material aging or defects. Finally, the insulation status level of the tubular busbar is evaluated based on the insulation degradation contribution component and the attention weights. By comparing the weighted scores of the degradation contribution of each feature group with the attention weights, the dominant degradation mode is identified as thermal aging, vibration fatigue, electrical stress or environmental corrosion, and the evaluation results and targeted operation and maintenance suggestions are output.

[0029] In the above embodiments, the static structural causal model assumes that the intensity of the causal effect between characteristic quantities does not change with time. However, during the long-term operation of the tubular busbar, insulation degradation will cause the causal relationship between various physical quantities to drift. For example, the contribution of vibration to temperature increases with aging. The static model cannot adapt to this time-varying characteristic, resulting in a gradual decrease in decomposition accuracy. Therefore, in order to solve this problem, in this embodiment, the construction and updating of the structural causal model in step S4 adopts an adaptive time-varying mechanism, which specifically includes the following steps: S4.1. Using historical data from the initial commissioning period of the tubular busbar when its insulation condition is good, construct an initial structural causal model. Including cause-effect graph structure and the conditional probability parameters between variables ; S4.2 Set the time window length L. Every L time interval, using the running data within the current time window, re-estimate the causal effect strength corresponding to each directed edge in the feature association graph; for any directed edge... The strength of its causal effect Defined as: Where A represents the causal characteristic and B represents the outcome characteristic. As the baseline value of A, For the preset intervention increment, This indicates that an intervention is being performed in the outcome-causal model. Expressing expectations; S4.3 Calculate the rate of change of the intensity of the causal effect: in, The strength of the causal effect in the previous time window. If is a non-zero small constant, Exceeding the preset threshold If so, then the directed edge is determined to have significant causal drift; S4.4 For directed edges exhibiting significant causal drift, the following update operations are performed sequentially: First, using the current window data, the conditional probability parameter of the edge is updated through Bayesian online learning; then, if the drift of the edge persists for more than two time windows, a conditional independence test is further performed, and a decision is made based on the test result to delete the directed edge; the updated structural causal model It replaces the current model and is used for subsequent counterfactual decomposition.

[0030] In this embodiment, although causal drift can be monitored and the model updated, the core operational issue of how long the insulation can continue to operate safely cannot be predicted. Operations personnel not only need to know the current status and causes of degradation, but also need to estimate the remaining lifespan to develop a maintenance plan. Therefore, to solve this problem, this embodiment also includes an insulation remaining lifespan prediction step S4.5: The calculated rate of change of causal effect intensity of each causal edge in different time windows A drift trend sequence for each causal edge is constructed. For the key causal edge corresponding to the dominant degradation mode, a time series prediction model is used to extrapolate its future trend, predicting that the causal effect strength of the causal edge first exceeds a preset danger threshold. The difference between the current time and the specified time point is used as the estimated remaining insulation lifetime (RUL). Time series forecasting models employ exponential smoothing or linear regression, and their forecasting formulas are as follows: in, For the future Predicted value of the causal effect strength at time t. The strength of the causal effect at the current moment. The average rate of change over the past k-th time window. is the preset decay weight coefficient, and m is the number of backtracking windows; The formula for calculating the remaining useful life (RUL) is: Where c is the set of key causal edges, Let be the predicted value of the i-th causal edge. The preset danger threshold corresponding to this edge. Prediction time required to reach the threshold; Finally, the remaining lifetime estimate, RUL machine confidence interval, is output.

[0031] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A method for evaluating the insulation status of tubular busbars based on multi-feature fusion, characterized in that, Includes the following steps: S1. Deploy sensor arrays at various locations on the tubular busbar to synchronously collect multi-source heterogeneous characteristic quantities, including electrical characteristic quantities, thermal characteristic quantities, vibration characteristic quantities and environmental characteristic quantities; S2. Perform time synchronization calibration and spatial coordinate mapping on the collected multi-source heterogeneous feature quantities to obtain aligned multi-source heterogeneous feature data. Then, based on the multi-physics coupling mechanism between electrical, thermal, vibration and environmental features, construct a feature association graph with feature quantities as nodes and physical causal action directions as directed edges, and set constraint rules including time, threshold and direction for each directed edge. S3. The feature association graph is used as the prior structure input to the graph attention network to perform deep fusion on the aligned multi-source heterogeneous feature data, generate a fused feature vector, and output the attention weights between each feature quantity. The attention coefficients of the graph attention network are adaptively adjusted according to the current load current and ambient temperature. S4. Construct a structural causal model in the form of a Bayesian network based on the fused feature vectors, learn the causal structure between variables using historical operating data, and use a counterfactual reasoning framework to convert the amount of multi-source heterogeneous feature data into operating condition driving components and insulation degradation contribution components. S5. Based on the insulation degradation contribution component and the attention weight, assess the current insulation status level of the tubular busbar and identify the dominant degradation mode.

2. The method for evaluating the insulation status of tubular busbars based on multi-feature fusion according to claim 1, characterized in that, In step S1, the electrical characteristics include partial discharge pulse signals and grounding current; the thermal characteristics include bus surface temperature distribution and temperature rise rate; the vibration characteristics include triaxial vibration spectrum and impact pulse amplitude; and the environmental characteristics include ambient temperature and humidity, dust concentration, and corrosive gas concentration. The sensor array includes an ultra-high frequency partial discharge sensor, a high frequency current transformer, a distributed fiber optic temperature sensor or an infrared thermal imager, a triaxial accelerometer, and a multi-parameter environmental sensor.

3. The method for evaluating the insulation status of tubular busbars based on multi-feature fusion according to claim 1, characterized in that, The specific process of constructing the feature association graph in step S2 is as follows: S2.

1. Each of the collected electrical, thermal, vibration, and environmental characteristics is considered a node, forming a node set. S2.

2. Based on the multi-physics coupling characteristics between electric field, thermal field, magnetic field and stress field during the operation of insulated tubular busbar, determine the causal relationship between each characteristic quantity, and construct a set with the direction of physical causal interaction as the directed edge. Electrical and thermal coupling edge: A bidirectional causal relationship is established between the partial discharge signal and temperature; Thermal-mechanical coupling: A one-way causal relationship is established between the temperature gradient and the vibration and shock pulse; Mechanical and electrical coupling: One-way causal relationships are established between the vibration spectrum and the grounding current, and between the vibration impact pulse and the partial discharge signal, respectively; Environmental and electrical coupling edges: One-way causal relationships are established between environmental humidity and grounding current, and between dust concentration and partial discharge signal, respectively; Environmental and thermal coupling: Establish unidirectional causal relationships between ambient temperature and temperature rise rate, and between corrosive gas concentration and temperature distribution anomaly, respectively; S2.3 For each directed edge, set constraint rules according to the physical mechanism. The constraint rules include time constraints, threshold constraints and directional constraints. The time constraints are used to specify the response time range of the causal action. The threshold constraints are used to specify the characteristic quantity numerical conditions for the occurrence of the causal action. The directional constraints are used to specify the unidirectionality and bidirectionality of the causal action.

4. The method for evaluating the insulation status of tubular busbars based on multi-feature fusion according to claim 1, characterized in that, The attention coefficients of the graph attention network described in step S3 are calculated as follows: For node i and its neighbor node j, the attention coefficient ,in, , Let W be the node feature vector, and W be the weight matrix. This is a learnable attention parameter vector. The activation function is used; then the attention weights are obtained by normalization using the Softmax function. ; The adaptive adjustment refers to: Let exp be the set of neighboring nodes of node i, where exp is the exponential function and the attention coefficient is the sum of the neighboring nodes of node i. Introducing load current I and ambient temperature As a regulatory factor, i.e. ,in This is the operating condition encoding vector obtained by mapping load current and ambient temperature.

5. The method for evaluating the insulation status of tubular busbars based on multi-feature fusion according to claim 1, characterized in that, The counterfactual reasoning framework described in step S4 decomposes the measured feature quantity X into operating condition-driven components. and insulation degradation contribution : in, Operating condition drive component, where I is the load current. For ambient temperature, The intensity of the external vibration source, The operating condition driving function is obtained through structural causal model learning; Insulation degradation contribution , The estimated operating conditions predicted under counterfactual intervention using a structural causal model; The counterfactual intervention specifically involves: setting intervention variables. Calculate the causal effect size , This represents the conditional expectation after do intervention in the structural causal model. Aging represents the latent variable of insulation aging that cannot be directly observed. Normal represents the normal state, and abnormal represents the abnormal state.

6. The method for evaluating the insulation status of tubular busbars based on multi-feature fusion according to claim 1, characterized in that, The insulation condition level mentioned in step S5 includes five levels: good, attention, abnormal, serious, and critical. It is determined based on a comprehensive score of the amplitude and trend of the insulation degradation contribution component and the attention weight. The method for identifying the dominant degradation mode is as follows: calculate the score for each degradation mode, and the mode with the highest score is the current dominant degradation mode; the score of the kth mode... The calculation formula is: in, This represents the set of features corresponding to the k-th degradation mode. For set The number of features in the middle The attention weights for the i-th feature output in step S3 are... The insulation degradation contribution component is obtained for the i-th feature quantity decomposed in step S4. The degradation modes include thermal aging-dominated mode, vibration fatigue-dominated mode, electrical stress-dominated mode, and environmental corrosion-dominated mode. Among them, the characteristic quantity set corresponding to the thermal aging-dominated mode includes temperature and temperature rise rate; the characteristic quantity set corresponding to the vibration fatigue-dominated mode includes vibration spectrum and impact pulse amplitude; the characteristic quantity set corresponding to the electrical stress-dominated mode includes partial discharge pulse signal and grounding current; and the characteristic quantity set corresponding to the environmental corrosion-dominated mode includes ambient temperature and humidity, dust concentration, and corrosive gas concentration.

7. The method for evaluating the insulation status of tubular busbars based on multi-feature fusion according to claim 5, characterized in that, The construction and updating of the structural causal model in step S4 adopts an adaptive time-varying mechanism, which specifically includes the following steps: S4.

1. Using historical data from the initial commissioning period of the tubular busbar when its insulation condition is good, construct an initial structural causal model. Including cause-effect graph structure and the conditional probability parameters between variables ; S4.

2. Set a time window length L. Every L time interval, using the running data within the current time window, re-estimate the causal effect strength corresponding to each directed edge in the feature association graph; for any directed edge... The strength of its causal effect Defined as: Where A represents the causal characteristic and B represents the outcome characteristic. As the baseline value of A, For the preset intervention increment, This indicates that an intervention is being performed in the outcome-causal model. Expressing expectations; S4.3 Calculate the rate of change of the intensity of the causal effect: in, The strength of the causal effect in the previous time window. If is a non-zero small constant, Exceeding the preset threshold If so, then the directed edge is determined to have significant causal drift; S4.4 For directed edges exhibiting significant causal drift, the following update operations are performed sequentially: First, using the current window data, the conditional probability parameter of the edge is updated through Bayesian online learning; then, if the drift of the edge persists for more than two time windows, a conditional independence test is further performed, and a decision is made based on the test result to delete the directed edge; the updated structural causal model It replaces the current model and is used for subsequent counterfactual decomposition.

8. The method for evaluating the insulation status of tubular busbars based on multi-feature fusion according to claim 7, characterized in that, It also includes insulation remaining life prediction step S4.5: The calculated rate of change of causal effect intensity of each causal edge in different time windows A drift trend sequence for each causal edge is constructed. For the key causal edge corresponding to the dominant degradation mode, a time series prediction model is used to extrapolate its future trend, predicting that the causal effect strength of the causal edge first exceeds a preset danger threshold. The difference between the current time and the specified time point is used as the estimated remaining insulation lifetime (RUL). The time series prediction model employs either exponential smoothing or linear regression, and its prediction formula is as follows: in, For the future Predicted value of causal effect strength at time t. The strength of the causal effect at the current moment. The average rate of change over the past k-th time window. is the preset decay weight coefficient, and m is the number of backtracking windows; The formula for calculating the remaining lifetime (RUL) is as follows: Where c is the set of key causal edges, Let be the predicted value of the i-th causal edge. The preset danger threshold corresponding to this edge. Prediction time required to reach the threshold; Finally, the remaining lifetime estimate, RUL machine confidence interval, is output.

Images