A substation equipment state prediction method and system based on multi-source data fusion

Abstract

The present application relates to the technical field of power equipment state data processing, and particularly relates to a substation equipment state prediction method and system based on multi-source data fusion.A substation equipment state prediction method based on multi-source data fusion comprises the following steps: S1: collecting fast variable data and slow variable data, and constructing an initial causal diagram based on the fast variable data and the slow variable data; S2: based on the multi-physical field coupling mechanism of substation equipment, establishing a verification rule of false causal mode; S3: matching the physical stage causal diagram constructed by the initial causal diagram based on the verification rule, and obtaining a fusion causal diagram.The present application fuses time delay scanning and maximum information coefficient to construct a causal diagram, corrects multi-time scale pseudo correlation, extracts a real degradation trend by stripping working condition interference, combines multi-stress acceleration and a Bayesian network, jointly predicts residual life and failure probability, and improves the prediction accuracy of substation equipment.

Description

Technical Field

[0001] This invention relates to the field of power equipment status data processing technology, and in particular to a method and system for predicting the status of substation equipment based on multi-source data fusion. Background Technology

[0002] Multi-source data fusion is a data processing technology that correlates and integrates heterogeneous sensing information from electrical, thermal, and chemical sources. Substation equipment includes core primary equipment such as transformers, circuit breakers, instrument transformers, and gas-insulated switches. Existing condition prediction methods mostly rely on single threshold alarms or simple trend extrapolation. Some progress has been made in digital monitoring and early fault warning, which has promoted the transformation of operation and maintenance models to data-driven approaches.

[0003] However, when processing monitoring data at multiple time scales, existing methods generally mechanically synchronize instantaneous electrical quantities with delayed cumulative chemical quantities, ignoring the causal misalignment caused by the physical response time delay, thus introducing a large number of spurious correlations. At the same time, existing methods are difficult to separate the superimposed interference of seasonal load and ambient temperature fluctuations on monitoring indicators, often misjudging normal operating condition fluctuations as equipment degradation signals. This causal confusion and operating condition confusion together introduce biases at the data feature extraction stage, causing the prediction model to frequently generate false alarms in complex operating scenarios.

[0004] Therefore, there is an urgent need for a data processing method that can verify and correct spurious causal relationships in multi-source asynchronous data, so as to provide a reliable basis for substation equipment status prediction. Summary of the Invention

[0005] To overcome the drawbacks of spurious correlation and operating condition coupling interference, this invention provides a method and system for predicting the state of substation equipment based on multi-source data fusion.

[0006] The technical implementation scheme of the present invention is: a substation equipment status prediction method based on multi-source data fusion, comprising the following steps: S1: Collect fast variable data and slow variable data, and construct an initial cause-effect graph based on the fast variable data and slow variable data; S2: Based on the multi-physics coupling mechanism of substation equipment, establish verification rules for erroneous causal patterns; S3: Match the physical stage causal graph constructed from the initial causal graph based on the verification rules to obtain the fused causal graph; S4: Determine the working condition-state association pair based on the fused causal graph, and identify false degradation signals based on the working condition-state association pair to obtain the true defect deterioration trend; S5: Predict the condition of substation equipment based on the actual deterioration trend of defects.

[0007] Preferably, the collection of fast variable data and slow variable data includes: The fast variable data refers to state representation quantities that respond to system changes instantaneously or on a minute-by-minute scale; The slow variable data refers to physical quantities whose values ​​change monotonically or with a trend as the equipment degrades or accumulates faults over a long period of time.

[0008] Preferably, the construction of the initial causal graph based on fast variable data and slow variable data includes: Construct fast variable sequences based on fast variable data, and construct slow variable sequences based on slow variable data; Define a set of window lengths consisting of multiple candidate window lengths, and define a set of time delay lengths consisting of multiple candidate time delay lengths; Iterate through each candidate window length in the window length set and each candidate time delay length in the time delay length set, and perform the following calculations: Take the current time as the end time of the fast variable sequence, take the subsequence whose fast variable sequence length is equal to the candidate window length before the end time as the first subsequence, calculate the statistical feature value of the first subsequence, and take the statistical feature value as the driving feature value; Take the time after adding the current time and the candidate time delay length as the start time of the slow variable sequence, take the subsequence whose slow variable sequence length is equal to the candidate window length after the start time as the second subsequence, calculate the difference between the end time value and the start time value of the second subsequence as the response increment; Calculate the maximum information coefficient between the driving feature value and the response increment. The candidate correlation is defined as the relationship between the candidate window length and the candidate time delay length that results in the maximum information coefficient being maximized and exceeding the preset strong correlation threshold. An initial causal graph is constructed based on all the candidate correlations.

[0009] Preferably, the verification rules for establishing erroneous causal patterns based on the multi-physics coupling mechanism of substation equipment include: The verification rules for the erroneous causal patterns include the verification rules for positive causal erroneous patterns, the verification rules for reverse causal erroneous patterns, and the verification rules for causal tomography erroneous patterns. The verification rules for the three erroneous causal patterns together constitute the causal pattern verification rule base; the causal pattern verification rule base provides the basis for overall correction judgment in the next step; The positive causal error mode verification rule is defined as the logic for determining whether the time delay between the triggering time of the fast variable and the starting time of the response of the slow variable exceeds the allowable tolerance. The reverse causal error mode verification rule is defined as the logic that determines the cumulative degradation effect time window represented by the slow variable is earlier than the occurrence time of the driving event represented by the fast variable. The causal fault error mode verification rule is defined as the logic for determining whether a matching strongly correlated fast variable driving event is missing during a period when a slow variable experiences a significant interval jump, or whether a matching slow variable response change is missing during a period when a fast variable experiences a significant interval jump.

[0010] Preferably, the step of matching the physical stage causality graph constructed from the initial causality graph based on verification rules to obtain the fused causality graph includes: The fast variable sequence and the slow variable sequence are input into the Bayesian change point detection algorithm. The Bayesian change point detection algorithm outputs a posterior probability sequence of each time point on the complete time axis as a change point. The time points in the posterior probability sequence whose posterior probability exceeds a preset probability threshold are taken as change point positions. All change point positions constitute a change point set. The change point set divides the complete time axis into multiple physical stages. The causal consistency within each physical stage is maximized, and the causal consistency difference between different physical stages is maximized. For each physical stage, perform the following operations separately: extract all candidate correlations from the initial causal graph within the physical stage; construct the physical stage causal graph from all candidate correlations; and match each candidate correlation with the positive causal error pattern verification rules, negative causal error pattern verification rules, and causal tomography error pattern verification rules in the causal pattern verification rule base.

[0011] Preferably, the step of matching each candidate correlation with the positive causal error pattern verification rules, negative causal error pattern verification rules, and causal tomography error pattern verification rules in the causal pattern verification rule base includes: Candidate correlations that match the positive causal error pattern verification rules are marked as time-delay deviation pseudo-correlations. The time delay length of the time-delay deviation pseudo-correlations is adjusted according to the positive causal constraint direction. The adjusted candidate correlations are then re-incorporated into the physical stage causal graph. Candidate correlations that match the reverse causal error pattern validation rule are marked as causal inversion pseudo-correlations, and causal inversion pseudo-correlations are directly removed from the physical stage causal graph; Candidate correlations matching the causal fault error pattern verification rules are marked as causal link breaks. Latent degenerate state nodes are introduced to supplement the causal link break locations. These latent degenerate state nodes are used as missing variables in the physical stage causal graph to fill causal faults. Candidate correlations that do not match any erroneous causal pattern validation rules during the physical phase are retained as valid causal correlations. All valid causal relationships, after adjustment, deletion, and supplementation, constitute the reconstructed physical stage causal graph; all physical stage causal graphs are connected in chronological order to form a fused causal graph that spans the entire timeline.

[0012] Preferably, the step of determining the condition-state correlation pair based on the fused causal graph and identifying false degradation signals based on the condition-state correlation pair to obtain the true defect deterioration trend includes: In the fusion causal graph, variables representing external operating conditions in the fast variables are defined as operating condition variables, and variables representing the internal state of the equipment in the slow variables are defined as state observation variables. In a fused causal graph, the relationship where the fast variable belongs to the operating condition class and the slow variable belongs to the state observation class is defined as an operating condition-state association pair.

[0013] Preferably, the step of defining the association relationship in the fused causal graph where the fast variable belongs to the operating condition class variable and the slow variable belongs to the state observation class variable as an operating condition-state association pair includes: For each condition-state association pair, the following operations are performed: A sliding window combined with standard deviation threshold detection is used to divide the fast variable sequence into a stable condition segment and a condition abrupt change segment. For the stable condition segment, the slow variable sequence segment within the corresponding time interval is defined as a stable condition-state observation segment. Distribution statistics are performed on the stable condition-state observation segment to calculate the normal condition response band of the slow variable under stable conditions. For the condition abrupt change segment, the slow variable sequence segment within the corresponding time interval is defined as a condition abrupt change-state observation segment. The amplitude of the slow variable change in the condition abrupt change-state observation segment is compared with the normal condition response band. If the amplitude of the slow variable change exceeds the normal condition response band and the direction of the slow variable change matches the causal direction recorded in the fused causal graph, then the corresponding change is defined as a condition-driven response. The components are analyzed as follows: If the amplitude of the slow variable change exceeds the normal operating condition response band and the direction of the slow variable change contradicts the causal direction recorded in the fused causal diagram, the corresponding change is marked as a suspected degradation driving component. For each time point on the complete time axis, the difference between the measured value of the slow variable at each time point and the median value of the normal operating condition response band corresponding to the stable operating condition segment at that point is calculated. The operating condition driving response component is stripped from this difference, while the suspected degradation driving component is retained in this difference, resulting in a residual slow variable sequence, which serves as the degradation feature sequence after stripping the operating condition coupling. Monotonicity detection and trend extraction are performed on the degradation feature sequence. The unidirectional cumulative trend that continuously deviates from the zero baseline in the degradation feature sequence is the true defect deterioration trend, and the periodic zeroing fluctuation component in the degradation feature sequence is the residual operating condition coupling residue and is filtered out.

[0014] Preferably, the substation equipment condition prediction based on the actual defect deterioration trend includes: Each actual defect deterioration trend is defined as an independent degradation mechanism, and the rate of change of the actual defect deterioration trend is defined as the current degradation rate of the independent degradation mechanism. The dominant stress variables corresponding to each independent degradation mechanism are obtained. A multi-stress coupling acceleration function is constructed within the equipment. This function uses a generalized Eyring model, with each dominant stress variable and the interaction term between them as the acceleration stress term. The model parameters are used as constant coefficients to fit the nonlinear coupling relationship between the comprehensive acceleration factor and the product of the current degradation rates of multiple independent degradation mechanisms. The dominant stress variables and current degradation rates corresponding to each independent degradation mechanism are input into the multi-stress coupling acceleration function to calculate the comprehensive degradation rate. Based on the comprehensive degradation rate, the linear extrapolated remaining lifetime of each independent degradation mechanism is corrected to obtain the individual remaining lifetime after coupling correction. The causal relationships across different devices are extracted from the fusion causal graph, and a cross-device fault propagation directed graph is constructed. Each node in the cross-device fault propagation directed graph represents a device, and each directed edge represents the propagation direction and propagation condition. The individual remaining lifetimes of each device after coupling correction are transformed into the prior failure probabilities of the corresponding nodes through a preset probability mapping function. The results are then input into a Bayesian network inference algorithm to perform forward probability propagation calculations on the cross-device fault propagation directed graph, calculating the conditional failure probability of each node, which serves as the cascade failure risk prediction result. The individual remaining lifetime after coupling correction and the cascaded failure risk prediction result together constitute the substation equipment status prediction result.

[0015] Preferably, a substation equipment condition prediction system based on multi-source data fusion includes: The causal graph construction module is used to construct an initial causal graph containing true and false correlations based on fast variable data and slow variable data, through time-delay window scanning and maximum information coefficient calculation. The causality verification module is used to establish causality pattern verification rules for forward, reverse, and causality fault errors, correct the causality graph of the physical stage, and obtain the fused causality graph. The false degradation identification module is used to determine the working condition-state association pairs based on the fused causal graph, and to decouple the working conditions to identify the real defect deterioration trend; The condition prediction module is used to perform accelerated correction of multi-stress coupling and prediction of cross-equipment cascade failure probability, and outputs the condition prediction results of substation equipment.

[0016] Beneficial effects: This invention constructs an initial causal graph containing true and false correlations between fast and slow variable data by combining time-delay window scanning and maximum information coefficient calculation. It then corrects the physical stage causal graph segment by segment using three types of erroneous causal patterns: forward, reverse, and causal discontinuity, forming a fused causal graph. This overcomes the causal misalignment and spurious correlation problems caused by synchronous pairing of data across multiple time scales. Furthermore, based on the operating condition-state correlation pair, it performs stationary-abrupt operating condition coupling stripping to extract the true defect deterioration trend from the residual slow variable sequence, eliminating the interference of operating condition fluctuations on degradation identification. Finally, through multi-stress coupling accelerated correction and cross-equipment Bayesian network cascaded inference, it achieves joint prediction of individual unit remaining life and system failure probability, significantly improving the accuracy and reliability of substation equipment state prediction in complex operating environments. Attached Figure Description

[0017] Figure 1 This is a flowchart of the substation equipment status prediction method based on multi-source data fusion according to the present invention; Figure 2 This is a structural diagram of the substation equipment status prediction system based on multi-source data fusion according to the present invention. Detailed Implementation

[0018] The embodiments of the present invention will be described below with reference to the accompanying drawings.

[0019] Example 1: A method for predicting the condition of substation equipment based on multi-source data fusion, such as... Figure 1 As shown, it includes the following steps: S1-1: Collect fast variable data and slow variable data, including: The fast variable data refers to state representation quantities that respond to system changes instantaneously or on a minute-by-minute scale; The slow variable data refers to physical quantities whose values ​​change monotonically or with a trend as the equipment degrades or accumulates faults over a long period of time.

[0020] It should be noted that fast variable data is acquired directly through a high-sampling-rate sensing loop, while slow variable data is extracted periodically using an oil chromatography device. Existing technologies, when processing multi-timescale monitoring data, mechanically synchronize instantaneous electrical quantities with delayed, cumulative chemical quantities, neglecting the causal misalignment caused by physical response time delays. For example, if the power frequency short-circuit current is a fast variable and the dissolved gas concentration in the oil is a slow variable, aligning them at the same time will produce spurious causality, where future consequences are used to infer past causes. Discriminative acquisition preserves the natural time delay between driving events and cumulative consequences from the data source, reducing spurious correlations introduced by synchronization pairing and providing a structured input to the initial causal graph that is not distorted by time.

[0021] S1-2: Constructing an initial causal graph based on fast and slow variable data, including: Construct fast variable sequences based on fast variable data, and construct slow variable sequences based on slow variable data; Define a set of window lengths consisting of multiple candidate window lengths, and define a set of time delay lengths consisting of multiple candidate time delay lengths; Iterate through each candidate window length in the window length set and each candidate time delay length in the time delay length set, and perform the following calculations: Take the current time as the end time of the fast variable sequence, take the subsequence whose fast variable sequence length is equal to the candidate window length before the end time as the first subsequence, calculate the statistical feature value of the first subsequence, and take the statistical feature value as the driving feature value; Take the time after adding the current time and the candidate time delay length as the start time of the slow variable sequence, take the subsequence whose slow variable sequence length is equal to the candidate window length after the start time as the second subsequence, calculate the difference between the end time value and the start time value of the second subsequence as the response increment; Calculate the maximum information coefficient between the driving feature value and the response increment. The candidate correlation is defined as the relationship between the candidate window length and the candidate time delay length that results in the maximum information coefficient being maximized and exceeding the preset strong correlation threshold. An initial causal graph is constructed based on all the candidate correlations.

[0022] It should be noted that the initial causal graph contains true correlations and false correlations. False correlations arise from causal misalignment, causal inversion, or causal absence between fast and slow variables. The initial causal graph explicitly records the direction, strength, and time lag length of each candidate correlation.

[0023] By constructing fast and slow variable sequences, the two types of data acquired during the acquisition phase are organized into time-series computational carriers, preserving the time-delay relationship between driving events and cumulative consequences in a structured manner. The candidate window length set and candidate delay length set provide an explicit spatiotemporal grid for causal search. If only instantaneous value pairing is relied upon, data from different time scales will be forcibly aligned, masking physical response delays. The candidate window length set includes durations of 1 hour, 4 hours, and 8 hours, while the candidate delay length set includes delay values ​​of 0 hours, 24 hours, and 72 hours. The candidate window length is determined by the following criteria: the window should cover the typical duration of the observable cumulative effect of the fast variable on the equipment, and should not exceed the maximum duration of a single operating condition mode. The candidate delay length is determined by the following criteria: the lower bound of the delay range is zero, and the upper bound is determined by the physical upper bound of the response delay of the variable to the involved physical process; for example, the upper bound of the delay between partial discharge and dissolved acetylene concentration in oil should not exceed 72 hours. The density of window and delay values ​​is selected based on a trade-off between accuracy requirements and computational cost.

[0024] The driving characteristic value is obtained by calculating the statistical characteristic value of the first subsequence. It represents the combined effect of fast variables on the device within an observation window. Therefore, the mean or cumulative statistical characteristic is used to summarize this segment, rather than using a single instantaneous sampling point. The response increment is obtained by calculating the difference between the end time value and the start time value of the second subsequence. It represents the net cumulative change generated by the slow variables after the candidate time lag. Therefore, the difference between the end and start of the segment is calculated to measure the response magnitude, avoiding the interference of historical cumulative components in the absolute value of slow variables on causal measurement.

[0025] The maximum information coefficient (MAC) between the driving eigenvalue and the response increment is calculated because the internal physical processes of substation equipment are mostly nonlinearly coupled. The MAC can detect dependencies of any functional form, reducing the limitation of Pearson correlation coefficient, which only captures linear relationships. The preset strong correlation threshold is set by collecting fast and slow variable data under historical normal operating conditions, calculating the MAC value under all window length and time delay length combinations, forming a maximum information coefficient sample set, and taking the upper quartile of this sample set as the preset strong correlation threshold. When the MAC corresponding to a certain combination reaches the maximum value among all candidate combinations, and this value exceeds the preset strong correlation threshold, the correlation represented by this combination is defined as a candidate correlation, and all candidate correlations constitute the initial causal graph.

[0026] The initial causal graph is constructed with fast and slow variables as nodes and candidate correlations as directed edges, each edge labeled with direction, strength, and time lag length. Its purpose is to comprehensively record all potential causal hypotheses in the data. Correlations that conform to the physical causal time sequence and direction are considered true correlations; those arising from incorrect time sequences are considered false correlations. Causal misalignment refers to the deviation between the driving moment of the fast variable and the starting moment of the slow variable's response exceeding the allowable range of the physical process; causal inversion refers to the time window of the cumulative degradation effect of the slow variable being placed before the driving event of the fast variable; causal absence refers to the absence of a matching strongly correlated fast variable driving event within the period of a significant interval jump in the slow variable.

[0027] S2: Based on the multi-physics coupling mechanism of substation equipment, establish verification rules for erroneous causal modes, including: The verification rules for the erroneous causal patterns include the verification rules for positive causal erroneous patterns, the verification rules for reverse causal erroneous patterns, and the verification rules for causal tomography erroneous patterns. The verification rules for the three erroneous causal patterns together constitute the causal pattern verification rule base; the causal pattern verification rule base provides the basis for overall correction judgment in the next step; The positive causal error mode verification rule is defined as the logic for determining whether the time delay between the triggering time of the fast variable and the starting time of the response of the slow variable exceeds the allowable tolerance. The reverse causal error mode verification rule is defined as the logic that determines the cumulative degradation effect time window represented by the slow variable is earlier than the occurrence time of the driving event represented by the fast variable. The causal fault error mode verification rule is defined as the logic for determining whether a matching strongly correlated fast variable driving event is missing during a period when a slow variable experiences a significant interval jump, or whether a matching slow variable response change is missing during a period when a fast variable experiences a significant interval jump.

[0028] It should be noted that the positive causal error mode verification rule is established as follows: based on the equipment's thermal-electric multiphysics coupling model, it is determined that there is only a positive thermal accumulation causal relationship between the partial discharge quantity (fast variable) and the dissolved acetylene concentration in the oil (slow variable), and the upper limit of the time delay does not exceed 72 hours. This physical constraint is then converted into a verification rule to determine the situation where the time delay deviation exceeds the allowable tolerance.

[0029] The verification rules extract temporal constraints, directional constraints, and completeness constraints from the multi-physics coupling mechanism of the device, and establish three types of causal judgment logics for positive, negative, and discontinuous errors, which together form a causal pattern verification rule base. The candidate correlations in the initial causal graph are driven only by statistical associations and cannot distinguish between physical causality and spurious correlations. Therefore, a rule base is introduced to implement physical logic verification, providing a definite truth-based judgment basis for each candidate correlation. The positive rule judges cases where the time lag between the fast variable driver and the slow variable response exceeds the allowable tolerance of the physical process as errors; the negative rule judges cases where the time window of the cumulative degradation effect of the slow variable appears before the fast variable driver event; the discontinuity rule judges cases where there is a missing matching strong correlation fast variable driver event or a missing matching slow variable response change logic within a period of significant interval jump in the slow variable. These three types of rules explicitly mark the numerous causal misalignments, causal inversions, and causal omissions that arise in existing technologies due to multi-timescale data synchronization pairing, providing a strict physical judgment basis for the pruning and completion process of the causal graph in the physical stage.

[0030] S3: Based on the verification rules, match the physical stage causality graph constructed from the initial causality graph to obtain the fused causality graph, including: The fast variable sequence and the slow variable sequence are input into the Bayesian change point detection algorithm. The Bayesian change point detection algorithm outputs a posterior probability sequence of each time point on the complete time axis as a change point. The time points in the posterior probability sequence whose posterior probability exceeds a preset probability threshold are taken as change point positions. All change point positions constitute a change point set. The change point set divides the complete time axis into multiple physical stages. The causal consistency within each physical stage is maximized, and the causal consistency difference between different physical stages is maximized. For each physical stage, perform the following operations separately: extract all candidate correlations from the initial causal graph within the physical stage; construct the physical stage causal graph from all candidate correlations; and match each candidate correlation with the positive causal error pattern verification rules, negative causal error pattern verification rules, and causal tomography error pattern verification rules in the causal pattern verification rule base. Candidate correlations that match the positive causal error pattern verification rules are marked as time-delay deviation pseudo-correlations. The time delay length of the time-delay deviation pseudo-correlations is adjusted according to the positive causal constraint direction. The adjusted candidate correlations are then re-incorporated into the physical stage causal graph. Candidate correlations that match the reverse causal error pattern validation rule are marked as causal inversion pseudo-correlations, and causal inversion pseudo-correlations are directly removed from the physical stage causal graph; Candidate correlations matching the causal fault error pattern verification rules are marked as causal link breaks. Latent degenerate state nodes are introduced to supplement the causal link break locations. These latent degenerate state nodes are used as missing variables in the physical stage causal graph to fill causal faults. Candidate correlations that do not match any erroneous causal pattern validation rules during the physical phase are retained as valid causal correlations. All valid causal relationships, after adjustment, deletion, and supplementation, constitute the reconstructed physical stage causal graph; all physical stage causal graphs are connected in chronological order to form a fused causal graph that spans the entire timeline.

[0031] It should be noted that the Bayesian change point detection algorithm is used instead of ordinary mutation detection because this method does not rely on the mean or variance changes of a single variable, but rather infers the moment of a jump in the association pattern between fast and slow variables. This directly corresponds to the goal of dividing the physical stages into stages with the greatest causal consistency and the greatest difference in causal consistency between stages.

[0032] The change point location is used to mark the precise moment when a structural change occurs in the causal pattern. The preset probability threshold is set as the upper quartile of the posterior probability sequence, determined by the posterior probability distribution of the change point in historical operational data. Dividing the data into physical stages serves to segment continuous causal data into processing units with homogeneous internal causal structures and heterogeneous external causal structures. Maximum internal causal consistency means that the driving response relationship between fast and slow variables within the same physical stage maintains the same pattern. Maximum difference in causal consistency between stages means that the causal patterns of adjacent physical stages have undergone identifiable changes. Both factors together ensure that subsequent verification rules are executed within the correct stage boundaries, avoiding misjudgments caused by crossing heterogeneous intervals.

[0033] During forward rule matching, candidate correlations with time lag bias are corrected for time lag according to the physical constraint direction and then re-included. During reverse rule matching, candidate correlations with inverted causality are directly deleted. During fault rule matching, the positions of missing variables are filled with latent degradation state nodes to fill the causal link. Candidate correlations that do not match any rule are directly retained as valid causal correlations. The latent degradation state node represents a potential degradation state variable in the current set of monitored variables that is not directly sensed but is confirmed as a necessary link in the causal link through physical causal inference. Its value is not used as independent monitoring data, but as a placeholder node in the causal graph, which is indirectly characterized by the residual abnormal patterns of observable variables in the subsequent S4 trend extraction stage. The physical meaning of this node is indicated by the causal fault error pattern verification rule during matching (for example, if the rule identifies partial discharge leading to insulation decomposition but there is no corresponding dielectric loss monitoring, the latent node is marked as the degree of insulation dielectric degradation), which is used to guide the addition of sensing points or manual inspection items when the system is expanded. After the above adjustments, deletions, and additions, the causal diagrams of each stage are connected end to end in chronological order to form a fused causal diagram that runs through the entire timeline.

[0034] The fused causal graph is a complete causal structure after verification and correction. It carries all valid causal relationships between fast and slow variables that have been verified by physical logic. It solves the problems of causal misalignment and causal inversion introduced by multi-timescale data synchronization pairing in existing technologies, and provides structured constraints on causal direction and time delay length for subsequent working condition coupling and decoupling.

[0035] S4: Based on the fused causal graph, determine the working condition-state correlation pairs, and identify spurious degradation signals based on the working condition-state correlation pairs to obtain the true defect deterioration trend, including: In the fusion causal graph, variables representing external operating conditions in the fast variables are defined as operating condition variables, and variables representing the internal state of the equipment in the slow variables are defined as state observation variables. In the fusion causal graph, the relationship in which the fast variable belongs to the operating condition class and the slow variable belongs to the state observation class is defined as the operating condition-state association pair; For each condition-state association pair, the following operations are performed: A sliding window combined with standard deviation threshold detection is used to divide the fast variable sequence into a stable condition segment and a condition abrupt change segment. For the stable condition segment, the slow variable sequence segment within the corresponding time interval is defined as a stable condition-state observation segment. Distribution statistics are performed on the stable condition-state observation segment to calculate the normal condition response band of the slow variable under stable conditions. For the condition abrupt change segment, the slow variable sequence segment within the corresponding time interval is defined as a condition abrupt change-state observation segment. The amplitude of the slow variable change in the condition abrupt change-state observation segment is compared with the normal condition response band. If the amplitude of the slow variable change exceeds the normal condition response band and the direction of the slow variable change matches the causal direction recorded in the fused causal graph, then the corresponding change is defined as a condition-driven response. The components are analyzed as follows: If the amplitude of the slow variable change exceeds the normal operating condition response band and the direction of the slow variable change contradicts the causal direction recorded in the fused causal diagram, the corresponding change is marked as a suspected degradation driving component. For each time point on the complete time axis, the difference between the measured value of the slow variable at each time point and the median value of the normal operating condition response band corresponding to the stable operating condition segment at that point is calculated. The operating condition driving response component is stripped from this difference, while the suspected degradation driving component is retained in this difference, resulting in a residual slow variable sequence, which serves as the degradation feature sequence after stripping the operating condition coupling. Monotonicity detection and trend extraction are performed on the degradation feature sequence. The unidirectional cumulative trend that continuously deviates from the zero baseline in the degradation feature sequence is the true defect deterioration trend, and the periodic zeroing fluctuation component in the degradation feature sequence is the residual operating condition coupling residue and is filtered out.

[0036] It's important to note that the core of S4 is to distinguish between fluctuations caused by external operating conditions and the actual gradual changes within the equipment, based on the corrected causal structure of the fused causal graph. In the fused causal graph, fast variables whose changes are primarily driven by external conditions such as load and ambient temperature are defined as operating condition variables, with load current being a typical example; slow variables reflecting the equipment's own condition, such as insulation, thermal, or chemical accumulation, are defined as state observation variables. Only when the cause of a causal relationship belongs to the operating condition category and the effect belongs to the state observation category is it identified as an operating condition-state relationship pair. This accurately extracts the relationships that need to be decoupled from the causal graph, avoiding blind decomposition of all variables. The distinction logic is as follows: if a fast variable changes significantly during normal equipment operation and only when external power grid conditions fluctuate, it is classified as an operating condition variable; if a slow variable only shows a trend of cumulative change after defects occur within the equipment, it is classified as a state observation variable.

[0037] A sliding window combined with standard deviation threshold detection divides the fast variable sequence into stable and abruptly changing operating conditions. Segments within the window with a standard deviation below the threshold are considered stable segments, while those above are considered abruptly changing segments. The standard deviation threshold is determined based on the upper limit of the standard deviation distribution within the sliding window of historical stable operating condition data segments. Distribution statistics are performed on the slow variable segments corresponding to the stable segments, and the median and main dispersion range are used to construct the normal operating condition response band, serving as the benchmark for normal fluctuations in equipment status under unchanged operating conditions. The normal operating condition response band employs a sliding update mechanism: whenever a new stable-state observation segment, verified to have no defect deterioration trend, is added, this segment's data is included in the rolling statistical sample pool of the normal operating condition response band. The distribution statistics parameters are updated with stable-state observation segments within the most recent preset time window (e.g., the past 365 days), enabling the normal response band to adaptively track the slow aging of the equipment itself. Upon entering the transition phase, the magnitude and direction of the slow variable changes are compared: if the magnitude exceeds the normal response band and the direction is consistent with the causal direction recorded in the fused causal graph, this change is identified as a condition-driven response component, belonging to the normal equipment response caused by changes in external conditions; if the magnitude also exceeds the limit but the direction contradicts the causal direction, it is marked as a suspected degradation-driven component, indicating that abnormal changes that violate normal physical laws have occurred inside the equipment. Subsequently, the difference between the measured value of the slow variable and the median value of the corresponding normal response band is calculated point by point on the complete time axis. The condition-driven response component is stripped from this difference, while the suspected degradation-driven component is retained in this difference, resulting in the residual slow variable sequence.

[0038] The residual slow variable sequence uses zero as a baseline, representing complete consistency between the equipment state and the normal operating condition response; that is, the measured value of the slow variable at the current moment is exactly equal to the median value of the normal response under that operating condition. The sign and magnitude of the residuals reflect the direction and degree of deviation of the equipment state from the normal baseline. Monotonicity detection and trend extraction are performed on this sequence, with the following criteria: If, after the Mann-Kendall trend test, the absolute value of the test statistic is greater than the critical value and the p-value is less than the preset significance level (e.g., 0.05), it indicates that the series has a significant monotonic trend, exhibiting a unidirectional cumulative characteristic of continuously deviating from the zero baseline without regression. This trend is then judged as a true defect worsening trend. Unidirectional means that the direction of change remains consistent, with the residuals consistently positive or negative without directional reversal; cumulative means that the deviation gradually increases over time, with the effect continuously accumulating, rather than fluctuating within a fixed range. Physically, this characteristic corresponds to an irreversible degradation process within the equipment (such as insulation aging and a continuous decrease in heat dissipation capacity), and the degree of defect is gradually deepening.

[0039] If the p-value is greater than the preset significance level, it indicates that the sequence has no significant monotonic trend; or if the sequence fails the trend consistency test within any sliding sub-window, the sequence is judged to be a residual coupled residual and is filtered out. Residual coupled residuals manifest as residuals fluctuating back and forth around the zero baseline and periodically returning to the vicinity of the zero line. The components of this fluctuation are seasonal temperature changes, residual unwashed components after stripping external operating condition fluctuations from the diurnal load cycle, and random noise from the measurement system. These fluctuations do not reflect irreversible internal equipment degradation and are therefore not output as a defect worsening signal.

[0040] Thus, the two types of interferences, causal confusion and operating condition coupling, have been eliminated one after another, and the degradation trend output directly supports subsequent lifetime prediction.

[0041] S5: Predicting the condition of substation equipment based on the actual deterioration trend of defects, including: Each actual defect deterioration trend is defined as an independent degradation mechanism, and the rate of change of the actual defect deterioration trend is defined as the current degradation rate of the independent degradation mechanism. The dominant stress variables corresponding to each independent degradation mechanism are obtained. A multi-stress coupling acceleration function is constructed within the equipment. This function uses a generalized Eyring model, with each dominant stress variable and the interaction term between them as the acceleration stress term. The model parameters are used as constant coefficients to fit the nonlinear coupling relationship between the comprehensive acceleration factor and the product of the current degradation rates of multiple independent degradation mechanisms. The dominant stress variables and current degradation rates corresponding to each independent degradation mechanism are input into the multi-stress coupling acceleration function to calculate the comprehensive degradation rate. Based on the comprehensive degradation rate, the linear extrapolated remaining lifetime of each independent degradation mechanism is corrected to obtain the individual remaining lifetime after coupling correction.

[0042] The causal relationships across different devices are extracted from the fusion causal graph, and a cross-device fault propagation directed graph is constructed. Each node in the cross-device fault propagation directed graph represents a device, and each directed edge represents the propagation direction and propagation condition. The individual remaining lifetimes of each device after coupling correction are transformed into the prior failure probabilities of the corresponding nodes through a preset probability mapping function. The results are then input into a Bayesian network inference algorithm to perform forward probability propagation calculations on the cross-device fault propagation directed graph, calculating the conditional failure probability of each node, which serves as the cascade failure risk prediction result.

[0043] The individual remaining lifetime after coupling correction and the cascaded failure risk prediction result together constitute the substation equipment status prediction result.

[0044] It should be noted that S5 transforms the actual defect deterioration trend obtained after removing the interference of the operating conditions in S4 into clear prediction results on two levels.

[0045] At the first level, each real defect deterioration trend corresponds to a degradation process driven by specific physical or chemical stresses, which is identified as an independent degradation mechanism. The rate of change of the trend reflects the current degradation speed of this mechanism and is directly used as the current degradation rate. In actual operation, the insulation thermal aging and partial discharge degradation mechanisms do not develop independently, but rather exhibit a nonlinear coupling acceleration effect. Thermal stress accelerates discharge corrosion, and discharge products catalyze insulation decomposition. If each rate is directly extrapolated linearly, the prediction results will be systematically overly optimistic. To quantify this coupling relationship, a multi-stress coupling acceleration function based on the generalized Eyring model is introduced. An exemplary form of the multi-stress coupling acceleration function is a comprehensive acceleration factor AF that satisfies... ,in Temperature (in K). Electric stress (unit: kV / mm). , , , For undetermined model parameters, This is a scale parameter (or proportionality constant), an empirical constant related to the specific material and failure mode. The activation energy is a physical parameter characterizing the sensitivity of the degradation process to temperature stress, typically measured in eV. , These are model constants, representing the intensity of non-thermal stress (such as electric field V) and the effect of the interaction between this stress and temperature, respectively. , The dimensions of the exponent are determined through data fitting, and during the fitting process, the dimensions are automatically determined to be the reciprocals of the dimensions of the stress term being multiplied, thus keeping the exponent terms dimensionless. Boltzmann constant The model itself is suitable for describing the combined effects of temperature, electric field, and other stresses and their interactions on the degradation rate. Here, the dominant stress variables and stress interaction terms corresponding to each independent degradation mechanism are used as accelerating stress terms. The nonlinear relationship between the comprehensive acceleration factor and the product of each degradation rate is fitted to calculate the comprehensive degradation rate. The comprehensive degradation rate reflects the true degradation rhythm under the combined action of multiple mechanisms. It is used to replace each independent rate to correct the linear extrapolated remaining lifetime. The minimum value of the corrected remaining lifetime among all mechanisms is taken as the coupled corrected individual remaining lifetime of the device, reflecting the estimated healthy operating limit of a single device after considering the interaction of internal multiple mechanisms. The preset probability mapping function is used to convert the individual remaining lifetime (time dimension) into a priori failure probability (dimensionless, value range 0-1). A specific mapping method is: a failure probability function based on the Weibull distribution is established according to the device type. The shape and scale parameters of the Weibull distribution are obtained by fitting historical failure data of similar devices. For any device, its coupled-corrected individual remaining lifetime is substituted into the reliability function of the Weibull distribution to obtain its conditional failure probability within that future time period, which serves as the prior failure probability of the corresponding node. In principle, the longer the individual remaining lifetime, the smaller the mapped prior failure probability.

[0046] At the second level, the fused causal graph already contains valid causal correlations across different device boundaries. These cross-device associations are extracted, and a directed graph of cross-device fault propagation is constructed based on their propagation direction and triggering conditions. For example, the causal path of a transformer bushing fault leading to a fire is represented in the graph as a directed edge from the bushing node to the transformer node. The remaining lifetime of each device after coupling correction is converted into the prior failure probability of the corresponding node in the Bayesian network. This prior probability represents the likelihood of failure based solely on its own degradation state, without considering the influence of adjacent devices. Subsequently, through forward probability propagation calculation, given the failure of an upstream node, the conditional failure probability of downstream nodes is derived step by step. This conditional probability is the cascade failure risk prediction result, reflecting the comprehensive risk level of a single device failure inducing sequential failures of other devices along the physical causal link.

[0047] Ultimately, the coupled-corrected individual remaining lifetime answers the question of the available time of a single device, while the cascaded failure risk prediction results answer the systemic question of fault propagation within the substation. Together, they constitute the substation equipment condition prediction results.

[0048] Example 2: Based on Example 1, a substation equipment status prediction system based on multi-source data fusion, such as... Figure 2 As shown, it includes: The causal graph construction module is used to construct an initial causal graph containing true and false correlations based on fast variable data and slow variable data, through time-delay window scanning and maximum information coefficient calculation. The causality verification module is used to establish causality pattern verification rules for forward, reverse, and causality fault errors, correct the causality graph of the physical stage, and obtain the fused causality graph. The false degradation identification module is used to determine the working condition-state association pairs based on the fused causal graph, and to decouple the working conditions to identify the real defect deterioration trend; The condition prediction module is used to perform accelerated correction of multi-stress coupling and prediction of cross-equipment cascade failure probability, and outputs the condition prediction results of substation equipment.

[0049] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting the condition of substation equipment based on multi-source data fusion, characterized in that, Includes the following steps: S1: Collect fast variable data and slow variable data, and construct an initial cause-effect graph based on the fast variable data and slow variable data; S2: Based on the multi-physics coupling mechanism of substation equipment, establish verification rules for erroneous causal patterns; S3: Match the physical stage causal graph constructed from the initial causal graph based on the verification rules to obtain the fused causal graph; S4: Determine the working condition-state association pair based on the fused causal graph, and identify false degradation signals based on the working condition-state association pair to obtain the true defect deterioration trend; S5: Predict the condition of substation equipment based on the actual deterioration trend of defects.

2. The substation equipment status prediction method based on multi-source data fusion according to claim 1, characterized in that, The collection of fast variable data and slow variable data includes: The fast variable data refers to state representation quantities that respond to system changes instantaneously or on a minute-by-minute scale; The slow variable data refers to physical quantities whose values ​​change monotonically or with a trend as the equipment degrades or accumulates faults over a long period of time.

3. The substation equipment status prediction method based on multi-source data fusion according to claim 1, characterized in that, The construction of the initial causal graph based on fast and slow variable data includes: Construct fast variable sequences based on fast variable data, and construct slow variable sequences based on slow variable data; Define a set of window lengths consisting of multiple candidate window lengths, and define a set of time delay lengths consisting of multiple candidate time delay lengths; Iterate through each candidate window length in the window length set and each candidate time delay length in the time delay length set, and perform the following calculations: Take the current time as the end time of the fast variable sequence, take the subsequence whose fast variable sequence length is equal to the candidate window length before the end time as the first subsequence, calculate the statistical feature value of the first subsequence, and take the statistical feature value as the driving feature value; Take the time after adding the current time and the candidate time delay length as the start time of the slow variable sequence, take the subsequence whose slow variable sequence length is equal to the candidate window length after the start time as the second subsequence, calculate the difference between the end time value and the start time value of the second subsequence as the response increment; Calculate the maximum information coefficient between the driving feature value and the response increment. The candidate correlation is defined as the relationship between the candidate window length and the candidate time delay length that results in the maximum information coefficient reaching its maximum value and exceeding the preset strong correlation threshold. An initial causal graph is constructed based on all the candidate correlations.

4. The substation equipment status prediction method based on multi-source data fusion according to claim 1, characterized in that, The verification rules for establishing erroneous causal patterns based on the multi-physics coupling mechanism of substation equipment include: The verification rules for the erroneous causal patterns include the verification rules for positive causal erroneous patterns, the verification rules for reverse causal erroneous patterns, and the verification rules for causal tomography erroneous patterns. The verification rules for the three erroneous causal patterns together constitute the causal pattern verification rule base; the causal pattern verification rule base provides the basis for overall correction judgment in the next step; The positive causal error mode verification rule is defined as the logic for determining whether the time delay between the triggering time of the fast variable and the starting time of the response of the slow variable exceeds the allowable tolerance. The reverse causal error mode verification rule is defined as the logic that determines the cumulative degradation effect time window represented by the slow variable is earlier than the occurrence time of the driving event represented by the fast variable. The causal fault mode check rule is defined as the logic for determining whether a matching strongly correlated fast variable driving event is missing during a period when a slow variable experiences a significant interval jump, or whether a matching slow variable response change is missing during a period when a fast variable experiences a significant interval jump.

5. The substation equipment condition prediction method based on multi-source data fusion according to claim 1, characterized in that, The process of matching the physical stage causality graph constructed from the initial causality graph based on verification rules to obtain the fused causality graph includes: The fast variable sequence and the slow variable sequence are input into the Bayesian change point detection algorithm. The Bayesian change point detection algorithm outputs a posterior probability sequence of each time point on the complete time axis as a change point. The time points in the posterior probability sequence whose posterior probability exceeds a preset probability threshold are taken as change point positions. All change point positions constitute a change point set. The change point set divides the complete time axis into multiple physical stages. The causal consistency within each physical stage is maximized, and the causal consistency difference between different physical stages is maximized. For each physical stage, perform the following operations separately: extract all candidate correlations from the initial causal graph within the physical stage; construct the physical stage causal graph from all candidate correlations; and match each candidate correlation with the positive causal error pattern verification rules, negative causal error pattern verification rules, and causal tomography error pattern verification rules in the causal pattern verification rule base.

6. The substation equipment status prediction method based on multi-source data fusion according to claim 5, characterized in that, The step of matching each candidate correlation with the positive causal error pattern verification rules, negative causal error pattern verification rules, and causal tomography error pattern verification rules in the causal pattern verification rule base includes: Candidate correlations that match the positive causal error pattern verification rules are marked as time-delay deviation pseudo-correlations. The time delay length of the time-delay deviation pseudo-correlations is adjusted according to the positive causal constraint direction. The adjusted candidate correlations are then re-incorporated into the physical stage causal graph. Candidate correlations that match the reverse causal error pattern validation rule are marked as causal inversion pseudo-correlations, and causal inversion pseudo-correlations are directly removed from the physical stage causal graph; Candidate correlations matching the causal fault error pattern verification rules are marked as causal link breaks. Latent degenerate state nodes are introduced to supplement the causal link break locations. These latent degenerate state nodes are used as missing variables in the physical stage causal graph to fill causal faults. Candidate correlations that do not match any erroneous causal pattern validation rules during the physical phase are retained as valid causal correlations. All valid causal relationships, after adjustment, deletion, and supplementation, constitute the reconstructed physical stage causal graph; all physical stage causal graphs are connected in chronological order to form a fused causal graph that spans the entire timeline.

7. The substation equipment status prediction method based on multi-source data fusion according to claim 1, characterized in that, The process of determining the condition-state correlation pairs based on the fused causal graph, and identifying false degradation signals based on the condition-state correlation pairs to obtain the true defect deterioration trend, includes: In the fusion causal graph, variables representing external operating conditions in the fast variables are defined as operating condition variables, and variables representing the internal state of the equipment in the slow variables are defined as state observation variables. In a fused causal graph, the relationship where the fast variable belongs to the operating condition class and the slow variable belongs to the state observation class is defined as an operating condition-state association pair.

8. The method for predicting the condition of substation equipment based on multi-source data fusion according to claim 7, characterized in that, The relationship in the fused causal graph where the fast variable belongs to the operating condition class and the slow variable belongs to the state observation class is defined as an operating condition-state association pair, including: For each condition-state association pair, the following operations are performed: A sliding window combined with standard deviation threshold detection is used to divide the fast variable sequence into a stable condition segment and a condition abrupt change segment. For the stable condition segment, the slow variable sequence segment within the corresponding time interval is defined as a stable condition-state observation segment. Distribution statistics are performed on the stable condition-state observation segment to calculate the normal condition response band of the slow variable under stable conditions. For the condition abrupt change segment, the slow variable sequence segment within the corresponding time interval is defined as a condition abrupt change-state observation segment. The amplitude of the slow variable change in the condition abrupt change-state observation segment is compared with the normal condition response band. If the amplitude of the slow variable change exceeds the normal condition response band and the direction of the slow variable change matches the causal direction recorded in the fused causal graph, then the corresponding change is defined as a condition-driven response. The components are analyzed as follows: If the amplitude of the slow variable change exceeds the normal operating condition response band and the direction of the slow variable change contradicts the causal direction recorded in the fused causal diagram, the corresponding change is marked as a suspected degradation driving component. For each time point on the complete time axis, the difference between the measured value of the slow variable at each time point and the median value of the normal operating condition response band corresponding to the stable operating condition segment at that point is calculated. The operating condition driving response component is stripped from this difference, while the suspected degradation driving component is retained in this difference, resulting in a residual slow variable sequence, which serves as the degradation feature sequence after stripping the operating condition coupling. Monotonicity detection and trend extraction are performed on the degradation feature sequence. The unidirectional cumulative trend that continuously deviates from the zero baseline in the degradation feature sequence is the true defect deterioration trend, and the periodic zeroing fluctuation component in the degradation feature sequence is the residual operating condition coupling residue and is filtered out.

9. The method for predicting the condition of substation equipment based on multi-source data fusion according to claim 1, characterized in that, The substation equipment condition prediction based on the actual defect deterioration trend includes: Each actual defect deterioration trend is defined as an independent degradation mechanism, and the rate of change of the actual defect deterioration trend is defined as the current degradation rate of the independent degradation mechanism. The dominant stress variables corresponding to each independent degradation mechanism are obtained. A multi-stress coupling acceleration function is constructed within the equipment. This function uses a generalized Eyring model, with each dominant stress variable and the interaction term between them as the acceleration stress term. The model parameters are used as constant coefficients to fit the nonlinear coupling relationship between the comprehensive acceleration factor and the product of the current degradation rates of multiple independent degradation mechanisms. The dominant stress variables and current degradation rates corresponding to each independent degradation mechanism are input into the multi-stress coupling acceleration function to calculate the comprehensive degradation rate. Based on the comprehensive degradation rate, the linear extrapolated remaining lifetime of each independent degradation mechanism is corrected to obtain the individual remaining lifetime after coupling correction. The causal relationships across different devices are extracted from the fusion causal graph, and a cross-device fault propagation directed graph is constructed. Each node in the cross-device fault propagation directed graph represents a device, and each directed edge represents the propagation direction and propagation condition. The individual remaining lifetimes of each device after coupling correction are transformed into the prior failure probabilities of the corresponding nodes through a preset probability mapping function. The results are then input into a Bayesian network inference algorithm to perform forward probability propagation calculations on the cross-device fault propagation directed graph, calculating the conditional failure probability of each node, which serves as the cascade failure risk prediction result. The individual remaining lifetime after coupling correction and the cascaded failure risk prediction result together constitute the substation equipment status prediction result.

10. A substation equipment condition prediction system based on multi-source data fusion, used to implement the substation equipment condition prediction method based on multi-source data fusion as described in any one of claims 1-9, characterized in that, include: The causal graph construction module is used to construct an initial causal graph containing true and false correlations based on fast variable data and slow variable data, through time-delay window scanning and maximum information coefficient calculation. The causality verification module is used to establish causality pattern verification rules for forward, reverse, and causality fault errors, correct the causality graph of the physical stage, and obtain the fused causality graph. The false degradation identification module is used to determine the working condition-state association pairs based on the fused causal graph, and to decouple the working conditions to identify the real defect deterioration trend; The condition prediction module is used to perform accelerated correction of multi-stress coupling and prediction of cross-equipment cascade failure probability, and outputs the condition prediction results of substation equipment.

Images