A method and system for monitoring moisture ingress in a capacitor transformer bushing

Abstract

The application relates to the technical field of electrical equipment state monitoring, and discloses a capacitor type transformer bushing dampness monitoring method and system. The method comprises the following steps: obtaining an initial monitoring data stream from a monitoring node arranged on a transformer bushing; performing feature extraction based on time evolution analysis on the data stream to obtain a preliminary humidity response sequence, a preliminary temperature response sequence and a potential disturbance identification sequence; performing relevance verification on the humidity and temperature sequences, and calculating the coupling influence degree of the disturbance identification sequence on the humidity and temperature sequences; constructing a dynamic probability response matrix according to the verification result and the coupling influence degree, and deducing a dampness evaluation index; and implementing multi-level signal correction on the initial monitoring data stream according to the index, and analyzing core humidity components and core temperature components reflecting a dampness progress. The application can separate dampness signals from temperature and disturbance effects, and realizes dynamic and accurate evaluation of the dampness state of the transformer bushing.

Description

Technical Field

[0001] This invention relates to the field of electrical equipment condition monitoring technology, specifically to a method and system for monitoring moisture in the bushings of a capacitor transformer. Background Technology

[0002] The bushing of a capacitor transformer is a crucial insulation component of a power transformer. Moisture buildup in its internal insulation can directly lead to a decline in electrical performance and even cause serious accidents. Current monitoring technologies primarily rely on collecting single or a few parameters such as the bushing's dielectric loss factor, equivalent capacitance, or end-screen current, and then determining anomalies by setting fixed thresholds or performing simple trend comparisons. These methods treat the monitoring data as static or quasi-static signals, and their analysis presupposes that parameter changes are mainly caused by alterations in the insulation state.

[0003] Field monitoring data is a dynamic result of the combined effects of multiple factors. Changes in ambient temperature affect the dielectric parameters of insulating materials, and this effect superimposes on the characteristic changes caused by moisture in the time domain. Random disturbances such as electromagnetic interference and mechanical stress continuously mix into the monitoring signal, forming noise and abrupt changes. Current technology cannot effectively separate the three different types of components—moisture characteristics, temperature effects, and transient disturbances—from a continuous data stream, resulting in insensitivity to early, weak moisture signals and susceptibility to false alarms due to interference.

[0004] Existing assessment models typically treat humidity-related parameters in isolation or use fixed temperature and humidity compensation coefficients, lacking a quantitative description of the dynamic correlation between humidity and temperature signals, and failing to consider the impact of random disturbances on this correlation. This makes the assessment results difficult to adapt to complex changes in the operating environment, unable to accurately reflect the true risk level corresponding to the same monitoring value under different operating conditions, and limiting the accuracy and reliability of early warnings. A method is needed that can dynamically decouple the monitoring data stream and establish an adaptive probabilistic assessment model to achieve accurate identification of the casing's moisture status. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for monitoring moisture in the bushings of a capacitor transformer, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides a method for monitoring moisture absorption in the bushings of a capacitor transformer, the method comprising:

[0007] The initial monitoring data stream is obtained from the monitoring nodes deployed on the transformer bushings;

[0008] A feature extraction strategy based on time evolution analysis is performed on the initial monitoring data stream to obtain a preliminary humidity response sequence, a preliminary temperature response sequence, and a potential disturbance identification sequence;

[0009] Perform correlation verification between the preliminary humidity response sequence and the preliminary temperature response sequence, and calculate the coupling influence of the potential perturbation identifier sequence on the preliminary humidity response sequence and the preliminary temperature response sequence;

[0010] Based on the results of the correlation verification and the coupling influence degree, a dynamic probability response matrix is ​​constructed; and based on the dynamic probability response matrix, a moisture assessment index is derived.

[0011] Based on the aforementioned moisture assessment index, multi-level signal correction is performed on the initial monitoring data stream to extract the core humidity component and core temperature component that reflect the moisture absorption process.

[0012] Preferably, the step of performing a feature extraction strategy based on time evolution analysis on the initial monitoring data stream includes the following steps:

[0013] Construct a sliding evaluation window for the initial monitoring data stream;

[0014] Within the sliding evaluation window, identify and segment data segments with continuous stationary properties and data segments with transient non-stationary properties in the initial monitoring data stream;

[0015] For the data segment with continuous and stable properties, a trend stability detection method is used to separate the baseline humidity component and the baseline temperature component. The baseline humidity component and the baseline temperature component constitute part of the basic data of the preliminary humidity response sequence and the preliminary temperature response sequence.

[0016] For the data segment with transient non-stationary properties, an anomaly pattern capture method is initiated to generate initial perturbation labels with time stamps. The set of multiple initial perturbation labels constitutes a potential perturbation identification sequence.

[0017] Preferably, the method for initiating anomaly pattern capture generates an initial perturbation label with a time stamp, comprising the following steps:

[0018] Calculate the local energy entropy and neighborhood variance of each sampling point in the data segment with transient and non-stationary properties;

[0019] The local energy entropy and the neighborhood variance are input into a preset perturbation classification decision tree;

[0020] The initial perturbation type code and initial confidence level are output through the perturbation classification decision tree, and the initial perturbation type code, the initial confidence level and the corresponding sampling point timestamp are bound together to form the initial perturbation label.

[0021] Preferably, the step of performing the correlation verification between the preliminary humidity response sequence and the preliminary temperature response sequence includes the following steps:

[0022] Within the same time domain framework, the rate of change pattern of the initial humidity response sequence is compared with the rate of change pattern of the initial temperature response sequence;

[0023] When the rate of change pattern of the preliminary humidity response sequence is covariant with the rate of change pattern of the preliminary temperature response sequence, the covariance intensity coefficient and the phase delay are calculated, and the covariance intensity coefficient and the phase delay are used as positive correlation features.

[0024] When the rate of change pattern of the preliminary humidity response sequence is inversely related to the rate of change pattern of the preliminary temperature response sequence, the frequency and magnitude of the inverse change event are recorded, and the frequency and magnitude of the inverse change event are used as abnormal correlation features.

[0025] Preferably, calculating the coupling influence of the potential perturbation identifier sequence on the preliminary humidity response sequence and the preliminary temperature response sequence includes the following steps:

[0026] Based on the timestamp in the initial disturbance label, the corresponding disturbed data interval is located in the preliminary humidity response sequence and the preliminary temperature response sequence;

[0027] Calculate the data distribution distortion degree of the disturbed data interval in the preliminary humidity response sequence and the data distribution distortion degree of the disturbed data interval in the preliminary temperature response sequence, respectively.

[0028] By integrating the data distribution distortion of the preliminary humidity response sequence, the data distribution distortion of the preliminary temperature response sequence, and the initial confidence level in the initial perturbation label, the local coupling influence factor of a single initial perturbation label is calculated using a weighted evaluation model.

[0029] By summarizing the local coupling influence factors of all the initial perturbation labels and combining them with timestamp density analysis based on spatial neighborhood relationships, a global coupling influence degree is generated.

[0030] Preferably, constructing the dynamic probability response matrix includes the following steps:

[0031] The positive correlation features and the abnormal correlation features are used as the basic elements of the row vectors of the matrix;

[0032] The global coupling influence degree is used as the column vector adjustment coefficient of the matrix;

[0033] By using matrix transformation rules, the basic elements of the row vector and the adjustment coefficients of the column vector are fused and calculated to construct a dynamic probability response matrix whose elements represent the probability of the occurrence of a damp state.

[0034] Preferably, before constructing the dynamic probability response matrix based on the results of the correlation verification and the coupling influence degree, the following steps are further included:

[0035] Check whether there is a group of initial perturbation labels with the same initial perturbation type encoding and consecutive timestamps in the potential perturbation label sequence;

[0036] If the initial perturbation label group exists, then the perturbation event merging and confidence enhancement operations are performed on the initial perturbation label group to form a composite perturbation identifier, and the potential perturbation identifier sequence is updated using the composite perturbation identifier.

[0037] Preferably, the step of performing perturbation event merging and confidence enhancement operations on the initial perturbation label group to form a composite perturbation label includes the following steps:

[0038] Extract the timestamps of all initial perturbation labels in the initial perturbation label group, and determine the composite time interval by taking the earliest timestamp as the starting point and the latest timestamp as the ending point.

[0039] Calculate the weighted average of the initial confidence scores of all initial perturbation labels in the initial perturbation label group, and multiply the weighted average by a preset enhancement coefficient to obtain the enhanced confidence score;

[0040] The initial perturbation type encoding, the composite time interval, and the enhanced confidence level are encapsulated to generate a composite perturbation identifier.

[0041] Preferably, the step of performing multi-level signal correction on the initial monitoring data stream based on the moisture assessment index includes the following steps:

[0042] First-level correction: Based on the moisture assessment index, adaptively set multiple signal amplitude filtering thresholds in the initial monitoring data stream;

[0043] Second-level correction: The data stream after the first-level correction is compared with the composite time interval of the composite disturbance identifier, data points falling within the composite time interval are removed, and the data gaps after removal are smoothed by context-based interpolation.

[0044] Third-level correction: The data stream after the second-level correction is input to an optimized filter with the dynamic probability response matrix as a parameter, and finally the core humidity component and the core temperature component are separated.

[0045] Preferably, the present invention also includes a capacitive transformer bushing moisture monitoring system, the system including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein when the processor executes the computer program, it implements the steps of the capacitive transformer bushing moisture monitoring method described above.

[0046] Compared with the prior art, the beneficial effects of the present invention are:

[0047] A feature extraction strategy based on time evolution analysis was employed to synchronously process the initial monitoring data stream, separating the preliminary humidity response sequence, the preliminary temperature response sequence, and potential disturbance marker sequences. This technique can identify and remove interference components weakly correlated with or unrelated to the damping process from the original mixed signal, and preliminarily decouple the temperature change trend from the damping signal in the time domain. This means that the source data relied upon for subsequent analysis is no longer a direct reference to the original observations, but rather a feature sequence with clear physical orientation after preliminary purification, laying a clear signal foundation for accurately perceiving the essential characteristics of damping.

[0048] The correlation between humidity and temperature response sequences is verified, and the coupling influence of disturbance identification sequences on the two is calculated, thereby constructing a dynamic probabilistic response matrix. This technique abandons fixed thresholds and static models, and constructs a probabilistic assessment framework that adapts to environmental and operating conditions by quantifying the dynamic statistical correlation between humidity and temperature signals and the real-time impact of disturbances on the strength of this correlation. The moisture assessment index derived from this matrix is ​​essentially a risk metric that integrates multi-dimensional information and undergoes probability normalization. It can dynamically distinguish between parameter changes caused by normal temperature fluctuations and parameter degradation caused by insulation moisture, improving the specificity of state judgment. Multi-level signal correction based on this index can reversely analyze the core humidity and temperature components stripped of temperature effects and random disturbances, thus intuitively and reliably reflecting the true state of the moisture absorption process. Attached Figure Description

[0049] Figure 1 This is a schematic diagram illustrating the working principle of the capacitive transformer bushing moisture monitoring method described in this invention.

[0050] Figure 2 A flowchart for generating initial perturbation labels for an anomaly pattern capture method;

[0051] Figure 3 A flowchart for calculating the coupling effect of the perturbation sequence on the temperature and humidity sequence;

[0052] Figure 4 A heatmap showing the probability of transformer bushings becoming damp.

[0053] Figure 5This is a comparison chart showing the effect of the second-level signal correction for transformer bushing monitoring data. Detailed Implementation

[0054] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0055] Please see Figure 1 This invention provides a method for monitoring moisture absorption in capacitive transformer bushings. The method includes: continuously acquiring an initial monitoring data stream from monitoring nodes deployed on the transformer bushings, the data stream containing raw signals related to the bushing state. A feature extraction strategy based on time evolution analysis is performed on the acquired initial monitoring data stream. This process identifies the temporal characteristics of the data stream, thereby obtaining a preliminary humidity response sequence characterizing humidity changes, a preliminary temperature response sequence characterizing temperature changes, and a potential disturbance identifier sequence identifying abnormal events. Subsequently, a correlation verification is performed on the preliminary humidity response sequence and the preliminary temperature response sequence to analyze the correlation between their change patterns. Simultaneously, the coupling influence degree of the potential disturbance identifier sequence on the extracted preliminary humidity response sequence and preliminary temperature response sequence is calculated to quantify the impact of abnormal events on the monitoring data. Based on the results of the correlation verification and the calculated coupling influence degree, a dynamic probability response matrix is ​​constructed. This matrix is ​​used to map the probability of moisture absorption under different conditions, and a comprehensive moisture absorption assessment index is derived based on this matrix. Finally, based on the obtained moisture assessment indicators, multi-level signal correction was performed on the initial monitoring data stream to gradually filter out interference and correct the data, thereby resolving the core humidity component and core temperature component that can accurately reflect the moisture absorption process inside the casing.

[0056] Example 1: See Figure 2Within the sliding evaluation window, data segments with continuous stationary properties and those with transient non-stationary properties in the initial monitoring data stream are identified and segmented. For data segments with continuous stationary properties, a trend stability detection method is used to separate the baseline humidity component and the baseline temperature component. These components constitute part of the basic data for the preliminary humidity response sequence and the preliminary temperature response sequence. For data segments with transient non-stationary properties, an anomaly pattern capture method is activated to calculate the local energy entropy and neighborhood variance of each sampling point in the data segment. The calculated local energy entropy and neighborhood variance are input into a pre-defined perturbation classification decision tree. The perturbation classification decision tree outputs the initial perturbation type code and initial confidence level, and binds the initial perturbation type code, initial confidence level, and corresponding sampling point timestamp to form an initial perturbation label. A set of multiple initial perturbation labels constitutes a potential perturbation identification sequence.

[0057] In practice, monitoring nodes deployed on transformer bushings collect raw capacitance and environmental sensor data once per minute; this data stream is referred to as the initial monitoring data stream. A sliding evaluation window is constructed for the initial monitoring data stream, with a time length set to 120 consecutive sampling points, i.e., a two-hour data span. The sliding evaluation window slides forward by 60 sampling points each time. Within the sliding evaluation window, data segments with continuous stationary properties and data segments with transient non-stationary properties are identified and segmented. Specifically, the standard deviation and mean skew ratio of the data within the window are calculated. When the standard deviation of 50 consecutive sampling points is below a threshold and the mean skew ratio is less than a preset threshold, the data segment is identified as having continuous stationary properties. Conversely, when a data point experiences a sudden change with an amplitude exceeding three times the historical average fluctuation, the area is identified as having transient non-stationary properties.

[0058] In practical implementation, for identified data segments with continuous and stable properties, a trend stability detection method is employed. This method, based on first-order difference and moving average filtering, separates the baseline humidity component and the baseline temperature component. The baseline humidity and temperature components are low-frequency trend signals after removing short-term high-frequency fluctuations. They constitute part of the basic data for the preliminary humidity response sequence and the preliminary temperature response sequence. For example, a continuous and stable data segment from 10:00 AM to 2:00 PM, after processing, yields a smoothly rising baseline temperature component and a relatively flat baseline humidity component. These two components are then filled into the corresponding time positions of the preliminary temperature response sequence and the preliminary humidity response sequence, respectively.

[0059] In some embodiments, for data segments identified as having transient non-stationary properties, an anomaly pattern capture method is initiated. The local energy entropy and neighborhood variance of each sampling point in the data segment with transient non-stationary properties are calculated. The local energy entropy is calculated based on the spectral distribution of the signal within a time window of 20 sampling points centered on the current point. The calculated local energy entropy and neighborhood variance are input into a predefined perturbation classification decision tree. The perturbation classification decision tree matches predefined classification rules according to the input numerical features. The perturbation classification decision tree outputs an initial perturbation type code and an initial confidence level. For example, the initial perturbation type code "E001" represents transient electrical interference, and "T001" represents a sudden temperature change event. The initial perturbation type code, initial confidence level, and corresponding sampling point timestamps are bound together to form an initial perturbation label. In specific implementations, a set of multiple initial perturbation labels constitutes a potential perturbation identification sequence. For example, a group of dense perturbation labels around 3 PM collectively describes the same external interference event.

[0060] The formula used to calculate local energy entropy is:

[0061] Where: characters A numerical value representing local energy entropy; character Index representing frequency components; character Represents the total number of frequency components; character This represents the normalized energy percentage of the k-th frequency component.

[0062] Example 2: See Figure 3Within the same time domain framework, the rate of change patterns of the initial humidity response sequence and the initial temperature response sequence are compared. When the rate of change patterns of the initial humidity response sequence and the initial temperature response sequence are covariant in the same direction, the covariance intensity coefficient and phase delay are calculated, and these are used as positive correlation features. When the rate of change patterns of the initial humidity response sequence and the initial temperature response sequence are inversely related, the frequency and magnitude of the inverse events are recorded, and these are used as abnormal correlation features. Based on the timestamps in the initial perturbation labels, the corresponding perturbed data intervals are located in the initial humidity and initial temperature response sequences. The data distribution distortion degree of the perturbed data intervals in the initial humidity response sequence and the initial temperature response sequence are calculated separately. By fusing the data distribution distortion degree of the initial humidity response sequence, the data distribution distortion degree of the initial temperature response sequence, and the initial confidence level in the initial perturbation labels, the local coupling influence factor of a single initial perturbation label is calculated using a weighted evaluation model. By summarizing the local coupling influence factors of all initial perturbation labels and combining them with timestamp density analysis based on spatial neighborhood relationships, a global coupling influence degree is generated.

[0063] In practice, the correlation between the initial humidity response sequence and the initial temperature response sequence is verified. The rate of change patterns of the initial humidity response sequence and the initial temperature response sequence are compared within the same time domain framework. The rate of change pattern is obtained by calculating the first-order difference of the sequences over fixed time intervals, such as comparing the direction and magnitude of humidity and temperature changes at ten-minute intervals. When the rate of change patterns of the initial humidity response sequence and the initial temperature response sequence exhibit a covariant relationship (i.e., both rise or fall simultaneously), the covariance intensity coefficient and phase delay are calculated. The covariance intensity coefficient quantifies the closeness of the covariance, while the phase delay describes the time difference between humidity and temperature changes. The calculated covariance intensity coefficient and phase delay are used as positive correlation features. For example, in a gradual increase in ambient temperature, if the rate of increase of the initial temperature response sequence is 0.1 degrees Celsius per minute and the rate of increase of the initial humidity response sequence is 0.05% relative humidity per minute, the calculated covariance intensity coefficient is 0.78, and the phase delay is 2 minutes. These two values ​​are recorded as the positive correlation features for the current time period.

[0064] In practice, when the rate of change pattern of the primary humidity response sequence and the rate of change pattern of the primary temperature response sequence show an inverse relationship (i.e., one increases while the other decreases), the frequency and amplitude of the inverse event are recorded. The frequency of the inverse event is recorded as the number of times the inverse event occurs per unit time, and the amplitude is recorded as the average absolute value of the product of the humidity and temperature change rates in each event. The frequency and amplitude of the inverse event are used as anomaly correlation features. For example, if three brief events of humidity increase and temperature decrease are detected within five minutes, with an average amplitude of 0.15, the anomaly correlation feature is recorded as "Frequency: 3 times / 5 minutes, Average amplitude: 0.15".

[0065] In some embodiments, the coupling influence of potential disturbance label sequences on the preliminary humidity response sequence and the preliminary temperature response sequence is calculated. Based on the timestamp in the initial disturbance label, the corresponding disturbed data interval is located in the preliminary humidity response sequence and the preliminary temperature response sequence. The disturbed data interval is extended forward and backward by a predetermined number of sampling points centered on the timestamp. For example, an initial disturbance label labeled "E001" with a timestamp of 14:30 will locate the disturbed data interval from 14:25 to 14:35 in the preliminary humidity response sequence. The data distribution distortion degree of the disturbed data interval in the preliminary humidity response sequence and the data distribution distortion degree of the disturbed data interval in the preliminary temperature response sequence are calculated respectively. The data distribution distortion degree measures the degree of deviation of the statistical characteristics of the data in the disturbed interval from the normal background intervals before and after it.

[0066] In some embodiments, the data distribution distortion of the preliminary humidity response sequence, the data distribution distortion of the preliminary temperature response sequence, and the initial confidence level in the initial perturbation labels are integrated to calculate the local coupling influence factor of a single initial perturbation label using a weighted evaluation model. The weighted evaluation model is a linear weighted model, assigning a high weight to the data distribution distortion. The local coupling influence factors of all initial perturbation labels are summarized, and combined with timestamp density analysis based on spatial neighborhood relationships, the timestamp density analysis examines the clustering effect of multiple initial perturbation labels appearing within a short period, generating a global coupling influence degree. The global coupling influence degree is a comprehensive scalar value reflecting the overall impact strength of all perturbation events on the humidity and temperature sequences during the monitoring period.

[0067] It is understandable that correlation verification generates positive correlation features and anomalous correlation features, which describe the intrinsic relationship pattern between humidity and temperature signals from different perspectives. The formula used to calculate the covariance intensity coefficient θ_c is:

[0068]

[0069] Where: characters Represents the numerical value of the covariance intensity coefficient; character Index representing a point in time; character Indicates the total number of data points within the time window used for calculation; character This represents the first-order difference value of the initial humidity response sequence at index i; character This represents the average of the first-order differences of the initial humidity response sequence within this time window; character This represents the first-order difference value of the initial temperature response sequence at index i; character This represents the average of the first-order differences of the initial temperature response sequence within that time window. The formula describes the process of calculating the degree of linear correlation between the changes in two sequences using the Pearson correlation coefficient.

[0070] It is understandable that the calculation of data distribution distortion depends on the statistical comparison between the disturbed data interval and the normal background interval. The data distribution distortion of the preliminary humidity response sequence may differ from that of the preliminary temperature response sequence. The local coupling influence factor combines the degree of distortion with the initial confidence level. The higher the initial confidence level of the initial disturbance label, the more likely the corresponding data distortion is caused by the actual disturbance, and therefore it has a higher weight in the calculation of the local coupling influence factor. The global coupling influence level integrates the local effects and spatiotemporal clustering effects of all disturbances, providing a quantitative adjustment basis for the subsequent construction of the dynamic probability response matrix.

[0071] Example 3: Using positive correlation features and abnormal correlation features as the basic elements of the matrix's row vectors. The global coupling influence degree is used as the adjustment coefficient of the matrix's column vectors. Through matrix transformation rules, the basic elements of the row vectors and the adjustment coefficients of the column vectors are fused to construct a dynamic probability response matrix whose elements represent the probability of dampness occurrence. In specific implementation, the process of constructing the dynamic probability response matrix uses positive correlation features and abnormal correlation features as the basic elements of the matrix's row vectors. Positive correlation features include the values ​​of the covariance intensity coefficient and phase delay obtained from correlation verification, while abnormal correlation features include the values ​​of the frequency and amplitude of reverse change events. Taking a specific monitoring period as an example, positive correlation features might be represented by a covariance intensity coefficient of 0.82 and a phase delay of 90 seconds, while abnormal correlation features might be represented by a frequency of 2 times per hour and an amplitude of 0.12. These values ​​together constitute the original data of the basic elements of the row vectors for the current period.

[0072] In practice, the global coupling influence is used as the column vector adjustment coefficient of the matrix. The global coupling influence is a comprehensive scalar value, such as 0.75, which reflects the overall influence intensity of all identified disturbances on the humidity and temperature sequences within the current monitoring period. Through matrix transformation rules, the basic elements of the row vectors and the column vector adjustment coefficients are fused. These rules define how to scale and combine the basic elements of the row vectors using the column vector adjustment coefficients. The fusion operation of the basic elements of the row vectors and the column vector adjustment coefficients generates an intermediate matrix. Each element of this intermediate matrix is ​​derived from different components of the basic elements of the row vectors and the column vector adjustment coefficients through specific operations. For example, one element might be the product of the covariance intensity coefficient and the global coupling influence, while another element might be the phase delay divided by the global coupling influence.

[0073] In some embodiments, the basic elements of the row vector need to be normalized before participating in matrix transformation, so that values ​​of different dimensions and ranges, such as covariance intensity coefficient, phase delay, occurrence frequency, and variation amplitude, are on the same scale. Through fusion operations performed according to matrix transformation rules, a dynamic probability response matrix is ​​finally constructed, with elements representing the probability of a damp state occurring. The dynamic probability response matrix is ​​a structured representation of a multidimensional probability distribution. Its rows typically correspond to different combinations of associated features, and its columns may correspond to different time segments or operating conditions. Each element value in the matrix represents an estimated probability that the transformer bushing is in a damp state under the conditions defined by a specific row and column; this probability value is a value between 0 and 1.

[0074] Optionally, the fusion operation in the matrix transformation rule may include a weighted summation and a nonlinear mapping step. The weighted summation step assigns a weight to each component of the row vector's fundamental elements, which may be associated with the global coupling influence. The nonlinear mapping step converts the weighted summation result into probability values ​​using an activation function. These probabilities are then used to calculate the element probability values ​​in the dynamic probability response matrix. The formula can be expressed as:

[0075]

[0076] Where: characters This represents the probability of the damp state occurring, represented by the element in the m-th row and n-th column of the dynamic probability response matrix; character Represents a non-linear activation function, character Indicates the index of the basic element component of a row vector; character Represents the total number of components of the basic elements of a row vector; character Represents the weight coefficient associated with the x-th component; character This represents the normalized value of the fundamental element component of the x-th row vector in the m-th state; character This represents the column vector adjustment coefficient under the nth condition or time segment, i.e., the global coupling influence. The formula describes how to combine the normalized feature components, weights, and global coupling influence, and map them to probability values ​​through a nonlinear function.

[0077] Optionally, the dimensions of the dynamic probability response matrix can be adjusted according to the granularity of monitoring needs. The row vector basic elements can be further broken down into finer-grained sub-features, thereby increasing the number of rows in the dynamic probability response matrix to characterize more complex correlation patterns. The global coupling influence of the column vector adjustment coefficients can also be calculated into multiple values ​​based on different time windows, serving as adjustment coefficients for different columns, thus constructing a probability matrix with multiple rows and columns. It can be understood that the core of constructing the dynamic probability response matrix lies in mathematically combining the qualitative features generated by correlation verification with the quantitative measurement of disturbance effects. The row vector basic elements provide the feature basis for state discrimination, while the global coupling influence of the column vector adjustment coefficients modulates the reliability or importance of these features based on external disturbances. It can be understood that the probability of dampness occurrence generated through matrix transformation rules is a comprehensive evaluation value, simultaneously considering the inherent correlation pattern between humidity and temperature and the degree to which this pattern is affected by external disturbances. The dynamic probability response matrix provides a probabilistic basis for subsequently deriving a single dampness assessment index.

[0078] See Figure 4 This is a heatmap showing the probability of moisture in transformer bushings, with the core feature being the visualization of the dynamic probability response matrix. The probability of moisture ingress peaks between 12-16 hours, indicating the highest risk during this period; while the probabilities are relatively lower between 0-4 hours and 20-24 hours. The probability of moisture in "abnormally high frequency" states is significantly higher than that in "positively correlated" states, indicating that abnormal correlation characteristics have a stronger impact on the risk of moisture ingress. This heatmap-like approach is used for risk assessment of transformer bushing moisture monitoring, quickly identifying high-risk periods and key correlated states, and assisting in the formulation of equipment maintenance strategies.

[0079] Example 4: Check if there exists a group of initial perturbation labels with the same initial perturbation type code and consecutive timestamps in the potential perturbation label sequence. If an initial perturbation label group exists, perform perturbation event merging and confidence enhancement operations on the initial perturbation label group. Extract the timestamps of all initial perturbation labels in the initial perturbation label group, using the earliest timestamp as the starting point and the latest timestamp as the ending point to determine the composite time interval. Calculate the weighted average of the initial confidence of all initial perturbation labels in the initial perturbation label group, and multiply this weighted average by a preset enhancement coefficient to obtain the enhanced confidence. Encapsulate the initial perturbation type code, composite time interval, and enhanced confidence to generate a composite perturbation label. Update the potential perturbation label sequence using the generated composite perturbation label.

[0080] In practical implementation, before constructing the dynamic probability response matrix, it is checked whether there are groups of initial perturbation tags with the same initial perturbation type code and consecutive timestamps in the potential perturbation tag sequence. Consecutive timestamps are defined as the time interval between any two adjacent initial perturbation tags within the group being less than a preset merging time threshold, such as three minutes. There may be multiple groups of initial perturbation tags that meet this condition in the potential perturbation tag sequence. The checking process traverses all initial perturbation tags in the sequence, classifies them according to their initial perturbation type code, and then scans the time interval within each category after sorting by timestamp.

[0081] In some embodiments, if there exists a group of initial perturbation labels with the same initial perturbation type code and consecutive timestamps, then a perturbation event merging and confidence enhancement operation is performed on this initial perturbation label group. The timestamps of all initial perturbation labels in the initial perturbation label group are extracted, with the earliest timestamp as the starting point and the latest timestamp as the ending point, to determine the composite time interval. The weighted average of the initial confidence levels of all initial perturbation labels in the initial perturbation label group is calculated. The weighted average can be calculated based on time density or a simple arithmetic mean. The calculated weighted average is multiplied by a preset enhancement coefficient to obtain the enhanced confidence level. The enhancement coefficient is a value greater than 1, such as 1.2, used to improve the overall confidence level of the merged event. The initial perturbation type code, composite time interval, and enhanced confidence level are encapsulated to generate a composite perturbation identifier.

[0082] In practice, refer to Table 1, which shows the process of merging an initial disturbance label group. It demonstrates how three initial disturbance labels with the same initial disturbance type code and consecutive timestamps are merged into a composite disturbance label.

[0083] Table 1: Initial Disturbance Label Group Merging Table

[0084] Optionally, the weighted average calculation can assign different weights to the initial confidence levels at different time points, with initial confidence levels closer to the group center being given higher weights. Calculate the reinforced confidence level. The formula used is:

[0085] Where: characters Represents the numerical value of reinforced confidence; character This represents the preset enhancement coefficient, a constant with a value greater than 1; character Indicates the index of the initial perturbation label within the initial perturbation label group; character Indicates the total number of initial perturbation labels within the group; character The weight coefficient representing the j-th initial perturbation label can be determined by its timestamp position in the group; character This represents the initial confidence level of the j-th initial perturbation label. The formula describes the calculation process of weighted averaging of the initial confidence levels within the group followed by coefficient reinforcement.

[0086] Optionally, when determining the composite time interval, the start and end points can be slightly extended forward and backward based on the timestamps of the initial perturbation labels to cover the potential event impact boundaries. For example, the earliest timestamp minus a fixed number of seconds can be used as the actual start point, and the latest timestamp plus a fixed number of seconds can be used as the actual end point. The generated composite perturbation label is used to update the potential perturbation label sequence. The update operation involves deleting all the original initial perturbation labels in the initial perturbation label group and inserting the newly generated composite perturbation label at the corresponding time position. It can be understood that the checking and merging operations can effectively reduce the number of labels in the potential perturbation label sequence, integrating multiple fragmented initial perturbation labels describing the same continuous perturbation event into a more representative composite perturbation label. It can also be understood that the composite perturbation label generated through the confidence enhancement operation usually has a higher confidence level than the original single initial perturbation label. This reflects that the certainty of the event judgment after merging is improved due to the corroboration of multiple labels, and the composite perturbation label provides a simpler and more reliable input for subsequent calculation of the coupling impact degree.

[0087] Example 5: The first-level correction adaptively sets multiple signal amplitude filtering thresholds in the initial monitoring data stream based on the moisture assessment index. The second-level correction compares the data stream processed by the first-level correction with the composite time interval of the composite disturbance identifier, removes data points falling within the composite time interval, and performs context-based smoothing interpolation on the data gaps after removal. The third-level correction inputs the data stream processed by the second-level correction into an optimized filter with a dynamic probability response matrix as a parameter, ultimately separating the core humidity component and the core temperature component.

[0088] In practice, multi-level signal correction is applied to the initial monitoring data stream based on the moisture assessment index. The first-level correction adaptively sets multiple signal amplitude filtering thresholds in the initial monitoring data stream according to the moisture assessment index. The moisture assessment index is a numerical value that comprehensively reflects the probability of bushing moisture absorption, for example, 0.65. The setting of the signal amplitude filtering threshold is positively correlated with the value of the moisture assessment index. When the moisture assessment index is high, the set filtering threshold is more stringent to filter out more possible interference fluctuations. The initial monitoring data stream is compared with these set signal amplitude filtering thresholds in sequence. Data points with amplitudes exceeding the corresponding threshold are regarded as abnormal pulses and are suppressed or smoothed. For example, for an initial monitoring data stream containing raw capacitance measurements, the first-level correction may set an upper threshold and a lower threshold, replacing spike data exceeding the upper threshold or falling below the lower threshold with the average of adjacent data.

[0089] In some embodiments, the second-level correction compares the data stream processed by the first-level correction with the composite time interval of the composite perturbation identifier, and removes data points falling within the composite time interval. The composite perturbation identifier comes from an updated sequence of potential perturbation identifiers, and each composite perturbation identifier contains a composite time interval; for example, the composite time interval for a composite perturbation identifier labeled "E001" is "08:12:05 to 08:13:30". The correction process scans the timestamps of the entire data stream, marking all data points with timestamps within this interval as invalid and removing them. Context-based smoothing interpolation is performed on the removed data gaps. Context-based smoothing interpolation utilizes the trends and values ​​of valid data points before and after the data gaps, employing linear interpolation or spline interpolation algorithms to generate new data to fill the gaps, ensuring the continuity of the data stream.

[0090] In practice, the third-stage correction inputs the data stream processed by the second-stage correction to an optimized filter that uses the dynamic probability response matrix as a parameter, ultimately separating the core humidity component and the core temperature component. The optimized filter using the dynamic probability response matrix as a parameter is an adaptive digital filter whose filter coefficients or bandwidth are dynamically adjusted according to the probability of moisture occurrence indicated by the elements in the dynamic probability response matrix. When the dynamic probability response matrix indicates a high probability of moisture in the current time period, the optimized filter uses a preset filtering mode to more finely separate low-frequency or characteristic frequency band signals related to moisture; when the probability of moisture is low, the optimized filter uses another filtering mode to preserve the original characteristics of a wider frequency band. The output signal processed by the optimized filter is analyzed into two independent components, which are the core humidity component and the core temperature component, respectively. The output of the optimized filter is described in terms of a single-step output. The formula can be expressed as:

[0091]

[0092] Where: characters This represents the output signal value of the optimized filter at time t; character Indicates the filter tap index; character Indicates the order of the filter; character This represents the k-th filter coefficient, which is the corresponding probability in the dynamic probability response matrix. Functions; characters This represents the probability element of the occurrence of the damp state in the dynamic probability response matrix at time t; character This represents the input value of the data stream at time tk after the second-stage correction process. The formula describes the process where the filter output is a weighted sum of multiple past input values, and the weighting coefficients are dynamically determined by the probability values ​​in the dynamic probability response matrix.

[0093] Optionally, the adaptively set multiple signal amplitude filtering thresholds in the first-level correction can include independent thresholds for different frequency components. For example, one threshold can be set for high-frequency noise components, and another threshold can be set for components near the power frequency. The context-based smoothing interpolation in the second-level correction has a context window length that can be dynamically adjusted according to the size of the data gap. For longer data gaps, a longer historical context is used for trend fitting. It can be understood that multi-level signal correction is a progressive process. The first-level correction focuses on removing obvious amplitude anomalies, the second-level correction focuses on clearing time-period data affected by known disturbance events, and the third-level correction uses the moisture assessment index and dynamic probability response matrix obtained from the previous steps to perform the finest signal separation related to the probability of moisture status. It can be understood that the core humidity component and core temperature component are the final results after multi-level correction and filtering. They minimize the impact of random interference and known disturbance events, and more directly reflect the true humidity and temperature changes related to the moisture absorption process of the transformer bushing and insulation material.

[0094] See Figure 5This is a comparison chart of the second-level signal correction effect for transformer bushing monitoring data, focusing on the processing and interpolation effect of the composite disturbance interval. The pink area (approximately 12-13.5 minutes) corresponds to the composite disturbance period. After the first-level correction, the data (blue) exhibits high-frequency fluctuations (disturbance characteristics) in this interval. The second-level correction, by removing data from this interval and using smooth interpolation, corrects the green line into a continuous trend curve, eliminating the influence of the disturbance. The curve after interpolation (green line) is consistent with the data trend before and after the disturbance interval, indicating that smooth interpolation effectively fills the data gap and ensures the continuity of the data flow. This step is a key part of multi-level correction, making the data closer to the true state of the transformer bushing by removing the influence of known disturbances, laying the foundation for subsequent core component extraction. This type of chart is used in the signal preprocessing stage of transformer bushing moisture monitoring, visually demonstrating the effect of disturbance removal, assisting in verifying the rationality of the correction algorithm, and providing a clean data source for subsequent moisture status assessment.

[0095] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0096] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for monitoring moisture in the bushings of a capacitor transformer, characterized in that, Includes the following steps: The initial monitoring data stream is obtained from the monitoring nodes deployed on the transformer bushings; A feature extraction strategy based on time evolution analysis is performed on the initial monitoring data stream to obtain a preliminary humidity response sequence, a preliminary temperature response sequence, and a potential disturbance identification sequence; Perform correlation verification between the preliminary humidity response sequence and the preliminary temperature response sequence, and calculate the coupling influence of the potential perturbation identifier sequence on the preliminary humidity response sequence and the preliminary temperature response sequence; Based on the results of the correlation verification and the coupling influence degree, a dynamic probability response matrix is ​​constructed; and based on the dynamic probability response matrix, a moisture assessment index is derived. Based on the moisture assessment index, multi-level signal correction is performed on the initial monitoring data stream to extract the core humidity component and core temperature component that reflect the moisture process. The feature extraction strategy based on time evolution analysis applied to the initial monitoring data stream includes the following steps: Construct a sliding evaluation window for the initial monitoring data stream; Within the sliding evaluation window, identify and segment data segments with continuous stationary properties and data segments with transient non-stationary properties in the initial monitoring data stream; For the data segment with continuous and stable properties, a trend stability detection method is used to separate the baseline humidity component and the baseline temperature component. The baseline humidity component and the baseline temperature component constitute part of the basic data of the preliminary humidity response sequence and the preliminary temperature response sequence. For the data segment with transient non-stationary properties, an anomaly pattern capture method is initiated to generate initial perturbation labels with time stamps. The set of multiple initial perturbation labels constitutes a potential perturbation identification sequence.

2. The method for monitoring moisture absorption in the bushing of a capacitor transformer according to claim 1, characterized in that, The method for capturing abnormal startup modes generates an initial disturbance label with a time stamp, including the following steps: Calculate the local energy entropy and neighborhood variance of each sampling point in the data segment with transient and non-stationary properties; The local energy entropy and the neighborhood variance are input into a preset perturbation classification decision tree; The initial perturbation type code and initial confidence level are output through the perturbation classification decision tree, and the initial perturbation type code, the initial confidence level and the corresponding sampling point timestamp are bound together to form the initial perturbation label.

3. The method for monitoring moisture absorption in the bushing of a capacitor transformer according to claim 2, characterized in that, The process of verifying the correlation between the preliminary humidity response sequence and the preliminary temperature response sequence includes the following steps: Within the same time domain framework, the rate of change pattern of the initial humidity response sequence is compared with the rate of change pattern of the initial temperature response sequence; When the rate of change pattern of the preliminary humidity response sequence is covariant with the rate of change pattern of the preliminary temperature response sequence, the covariance intensity coefficient and the phase delay are calculated, and the covariance intensity coefficient and the phase delay are used as positive correlation features. When the rate of change pattern of the preliminary humidity response sequence is inversely related to the rate of change pattern of the preliminary temperature response sequence, the frequency and magnitude of the inverse change event are recorded, and the frequency and magnitude of the inverse change event are used as abnormal correlation features.

4. The method for monitoring moisture in the bushing of a capacitor transformer according to claim 3, characterized in that, The calculation of the coupling influence of the potential perturbation identifier sequence on the preliminary humidity response sequence and the preliminary temperature response sequence includes the following steps: Based on the timestamp in the initial disturbance label, the corresponding disturbed data interval is located in the preliminary humidity response sequence and the preliminary temperature response sequence; Calculate the data distribution distortion degree of the disturbed data interval in the preliminary humidity response sequence and the data distribution distortion degree of the disturbed data interval in the preliminary temperature response sequence, respectively. By integrating the data distribution distortion of the preliminary humidity response sequence, the data distribution distortion of the preliminary temperature response sequence, and the initial confidence level in the initial perturbation label, the local coupling influence factor of a single initial perturbation label is calculated using a weighted evaluation model. By summarizing the local coupling influence factors of all the initial perturbation labels and combining them with timestamp density analysis based on spatial neighborhood relationships, a global coupling influence degree is generated.

5. The method for monitoring moisture in the bushing of a capacitor transformer according to claim 4, characterized in that, The construction of the dynamic probability response matrix includes the following steps: The positive correlation features and the abnormal correlation features are used as the basic elements of the row vectors of the matrix; The global coupling influence degree is used as the column vector adjustment coefficient of the matrix; By using matrix transformation rules, the basic elements of the row vector and the adjustment coefficients of the column vector are fused and calculated to construct a dynamic probability response matrix whose elements represent the probability of the occurrence of a damp state.

6. The method for monitoring moisture in the bushing of a capacitor transformer according to claim 5, characterized in that, Before constructing the dynamic probability response matrix based on the results of the correlation verification and the coupling influence degree, the following steps are also included: Check whether there is a group of initial perturbation labels with the same initial perturbation type encoding and consecutive timestamps in the potential perturbation label sequence; If the initial perturbation label group exists, then the perturbation event merging and confidence enhancement operations are performed on the initial perturbation label group to form a composite perturbation identifier, and the potential perturbation identifier sequence is updated using the composite perturbation identifier.

7. The method for monitoring moisture in a capacitor transformer bushing according to claim 6, characterized in that, The step of performing perturbation event merging and confidence enhancement operations on the initial perturbation label group to form a composite perturbation label includes the following steps: Extract the timestamps of all initial perturbation labels in the initial perturbation label group, and determine the composite time interval by taking the earliest timestamp as the starting point and the latest timestamp as the ending point. Calculate the weighted average of the initial confidence scores of all initial perturbation labels in the initial perturbation label group, and multiply the weighted average by a preset enhancement coefficient to obtain the enhanced confidence score; The initial perturbation type encoding, the composite time interval, and the enhanced confidence level are encapsulated to generate a composite perturbation identifier.

8. The method for monitoring moisture in the bushing of a capacitor transformer according to claim 7, characterized in that, The step of performing multi-level signal correction on the initial monitoring data stream based on the moisture assessment index includes the following steps: First-level correction: Based on the moisture assessment index, adaptively set multiple signal amplitude filtering thresholds in the initial monitoring data stream; Second-level correction: The data stream after the first-level correction is compared with the composite time interval of the composite disturbance identifier, data points falling within the composite time interval are removed, and the data gaps after removal are smoothed by context-based interpolation. Third-level correction: The data stream after the second-level correction is input to an optimized filter with the dynamic probability response matrix as a parameter, and finally the core humidity component and the core temperature component are separated.

9. A moisture monitoring system for capacitive transformer bushings, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method for monitoring moisture in the bushing of a capacitor transformer as described in any one of claims 1 to 8.

Images