Defect repeated alarm screening method and system based on twin network and topology analysis

Abstract

The application discloses a defect repeated alarm screening method and system based on a twin network and topological analysis, and relates to the technical field of intelligent substations. The method comprises the following steps: image pair features are extracted through a twin network and matching potential energy is generated; a dynamic reference surface resistant to environmental interference is constructed by using historical data; a weighted coupling graph is constructed based on potential energy residuals and space-time context, potential repeated alarm clusters are screened through belief propagation; time sequence alignment and multi-dimensional consistency evaluation are performed on alarms in the clusters, low-contribution noise nodes are removed through Shapley value game analysis; finally, stable defect chains are identified through topological persistent homology analysis, main alarms are selected, and repeated alarms are screened. The application effectively solves the misjudgment problem caused by light changes, equipment aging and complex interference, greatly improves the recognition accuracy and automation level, and significantly reduces the operation and maintenance review workload.

Description

Technical Field

[0001] This invention relates to the field of intelligent substation technology, and in particular to a method and system for filtering out duplicate defect alarms based on twin networks and topology analysis. Background Technology

[0002] In intelligent inspection systems for power and energy sectors, devices such as drones and fixed cameras frequently collect a large number of equipment images, leading to an explosive increase in alarm information. Furthermore, the same defect is repeatedly captured on camera at different times and under different environmental conditions, generating numerous duplicate alarms. This redundant information severely disrupts the work efficiency of maintenance personnel and increases the time cost of defect troubleshooting.

[0003] Traditional methods for filtering duplicate alarms mainly rely on manual verification or simple threshold matching. Manual verification is extremely inefficient and cannot meet the needs of large-scale inspection scenarios. Simple threshold matching methods, based on a single feature or fixed threshold, struggle to cope with complex environmental interferences such as changes in lighting, differences in shooting angles, and equipment aging, resulting in high false positive rates and poor generalization ability. Furthermore, traditional methods mostly lack physical interpretability and cannot relate to actual scenario factors such as equipment operating status and environmental changes. In inspection scenarios with multiple devices, multiple environments, and long cycles, it is difficult to accurately identify true duplicate alarms, retain core primary alarms, and eliminate redundant information, becoming a key bottleneck restricting the automated closed-loop management of inspections.

[0004] Based on the shortcomings of the existing technology, there is an urgent need for a scheme to screen out repeated alarms at inspection points that can improve the accuracy of repeated alarm identification, anti-interference and generalization ability, and has physical interpretability. Summary of the Invention

[0005] The purpose of this invention is to overcome at least one technical problem existing in the prior art and to provide a method and system for filtering out duplicate alarms based on twin networks and topology analysis.

[0006] On one hand, embodiments of the present invention provide a defect recurrence alarm screening method based on Siamese networks and topology analysis. The method includes: an offline modeling stage for generating a baseline model for auxiliary decision-making based on historical alarm data. The offline modeling stage includes: step S1, generating a historical initial matching potential value sequence using a pre-trained Siamese network based on historical alarm image pairs; step S2, removing pseudo-stable fluctuations caused by periodic interference from the historical initial matching potential value sequence based on a hierarchical Bayesian model and empirical mode decomposition, and constructing a robust time-evolving model through kernel density estimation and interquartile range. Potential energy reference surface; In the online processing stage, it is used to process the current alarm and filter out duplicate alarms. The online processing stage includes the following steps: Step S3: Perform multimodal feature extraction on the current alarm image to obtain three types of features corresponding to each alarm image: spatial distribution map, image texture descriptor, and segmentation region statistics. Integrate the three types of features into high-dimensional heterogeneous feature tensors of the alarm image; Step S4: Input the two high-dimensional heterogeneous feature tensors into the pre-trained Siamese network. Process them independently through each branch of the dual-branch symmetric feature extraction architecture of the Siamese network. A high-dimensional heterogeneous feature tensor is converted into two fixed-length one-dimensional feature vectors; Step S5: Project the feature vectors onto the tangent space of the Riemannian manifold and fuse the geodesic distance and matching potential energy to generate a current initial matching potential energy value that is resistant to environmental interference; Step S6: Map the current initial matching potential energy value to the robust potential energy reference surface to obtain the residual, and construct a dynamic weighted coupling graph by combining the spatiotemporal context information of the current alarm pair. Iterative reasoning is performed on the dynamic weighted coupling graph using the cyclic belief propagation algorithm to filter out potential recurring alarm clusters with high structural and semantic coordination; Step S7: For the... The alarm instances within the potential duplicate alarm cluster undergo time-series alignment and multi-dimensional consistency evaluation to generate a consistency score. Nodes within the potential duplicate alarm cluster are considered as game participants. By calculating their Shapley value for the consistency score and setting a dynamic threshold, abnormal interference nodes with a contribution lower than the dynamic threshold are eliminated to obtain a high-contribution alarm cluster. Step S8: The high-contribution alarm cluster undergoes topological data analysis to verify its structural stability, identifies a stable defect evolution chain, and selects a main alarm from the defect evolution chain based on a preset main alarm selection index. The remaining alarms are marked as duplicate alarms and filtered out.

[0007] Further, step S2 includes: step S21, based on the historical initial matching potential energy value sequence, learning the distribution law of matching potential energy values ​​of different equipment types under different environmental conditions through a hierarchical Bayesian model to obtain a conditional probability density function; step S22, for the historical initial matching potential energy value sequence of each equipment, identifying and removing pseudo-stable fluctuation components caused by periodic environmental interference through empirical mode decomposition, and reconstructing the corrected potential energy sequence of each equipment; step S23, aggregating the corrected potential energy sequences of all equipment under that type, and based on the distribution law represented by the conditional probability density function, performing kernel density estimation and interquartile range analysis on the aggregated corrected potential energy sequence to construct a robust potential energy reference surface that evolves slowly with the inspection time; wherein, the reference surface is used to provide a dynamic benchmark for judging whether a new alarm conforms to the repetitive pattern.

[0008] Furthermore, the current alarm image in step S3 is an image collected at different times from the same patrol point. Step S3 includes: Step S31, after grayscale processing, applying gamma correction to uniform image brightness, and using the Sobel operator to calculate the gradient values ​​in the horizontal and vertical directions for each pixel in the image; Step S32, dividing the image into 8×8 pixel cells, generating a 9-dimensional directional histogram vector in each cell based on the gradient direction and size of the pixel, and stitching them together to form a spatial distribution map; Step S33, extracting the contrast, energy, and correlation of the grayscale co-occurrence matrix, averaging them in multiple directions to obtain the image texture descriptor; Step S34, calling the pre-trained semantic segmentation model, assigning a category label to each pixel in the image to generate a pixel-level category mask, and calculating the mean, variance, and area ratio of pixel values ​​in each region according to the device region as segmentation region statistics.

[0009] Furthermore, step S4 includes: step S41, converting the high-dimensional heterogeneous feature tensor... and The symmetric dual-branch structure of the pre-trained Siamese network is input, and two one-dimensional feature vectors are obtained by mapping through convolution and pooling operations with shared weights. and Step S5 includes: Step S51, calculating the weighted Euclidean distance and cosine similarity between the two one-dimensional feature vectors to obtain the matching potential energy; Step S52, projecting the two one-dimensional feature vectors onto the Riemannian manifold space, calculating the geodesic distance and fusing it with the matching potential energy to generate the current initial matching potential energy value that is resistant to environmental interference.

[0010] Furthermore, the matching potential energy is calculated as follows:

[0011] ;

[0012] in, The weights of the Euclidean distance term are... For vectors The model, For vectors The model, Represents the geometric distance between two eigenvectors, with the denominator introducing... Dynamic normalization is implemented to prevent misjudgment due to excessive differences in feature amplitudes. For vectors and Cosine similarity between them The dot product of vectors reflects the degree of coordination between two vectors in a direction. The transposed vector , To match the potential energy.

[0013] Further, step S6 includes: Step S61, calculating the difference between the current initial matching potential energy value and the expected value of the robust potential energy reference surface of its corresponding equipment type at the same time, as the core residual; Step S62, standardizing and splicing the core residual with the geographical location, inspection order, time interval, quantified meteorological factors, and equipment maintenance cycle information of the current alarm pair to form a graph node attribute vector; Step S63, establishing connection edges between nodes based on the rule that the feature similarity between nodes is higher than a first threshold and the spatiotemporal distance is lower than a second threshold, and the weight of the edge... A dynamic weighted coupling graph is constructed by dynamically calculating feature similarity and temporal proximity. Step S64: Define a global energy function containing unary and binary terms, and perform iterative reasoning on the dynamic weighted coupling graph using a ring belief propagation algorithm to calculate the posterior probability of each node being a repeat alarm. The unary term is based on the node core residual, and the binary term is based on the penalty for label differences based on attribute similarity between connected nodes. Step S65: Identify the largest connected subgraph with a posterior probability higher than the confidence threshold and an average core residual lower than the residual threshold as the potential repeat alarm cluster.

[0014] Furthermore, the mathematical expression for the global energy function is:

[0015] ;

[0016] In the formula, Let be the matching potential residual of node i. To adjust the coefficients, the weight of environmental disturbances in the node energy is controlled. This is a meteorological interference factor used to enhance the penalty for mismatches under unfavorable imaging conditions. The overall smoothness strength coefficient controls the influence of prior structural information on the control chart. Let be the feature similarity between node i and node j. >1 is used for non-linear enhancement of low similarity penalty. Measure the contextual similarity between nodes i and j in a multidimensional feature space. Let be the Euclidean distance between the multidimensional attribute vectors of nodes i and j. For scale parameters, Let E be the activation penalty term when two nodes are assigned different labels, and let E be the set of edges. This represents the total energy.

[0017] Further, step S7 includes: Step S71, for the potential duplicate alarm cluster, sorting the alarm instances it contains according to the collection time, using the high-dimensional embedding vector output by the Siamese network in the feature extraction stage to construct a corresponding feature vector sequence for each alarm instance, and constructing a feature trajectory based on the feature vector sequence; Step S72, using the earliest alarm feature trajectory as a benchmark, using the dynamic time warping algorithm to align the remaining feature trajectories with it to find the optimal alignment path, and calculating the minimum cumulative distance and the standard deviation of the slope change of the alignment path; Step S73, fusing the minimum cumulative distance and the slope change... The standard deviation, stability index of semantic segmentation region between alarms, and phase offset of alarm time in device load cycle are used to calculate the geometric, semantic, and temporal triple consistency score; Step S74: Each alarm instance in the cluster is regarded as a participant in cooperative game, and the total payoff of its alliance is the triple consistency score. The marginal contribution value of each alarm instance is calculated by Shapley value formula; Step S75: A dynamic threshold is set based on the mean and standard deviation of Shapley values ​​of all alarm instances. Alarm instances with contribution values ​​lower than the dynamic threshold are identified as noise nodes and removed. The remaining alarm instances constitute the high contribution alarm cluster.

[0018] Further, step S8 includes: Step S81, treating the feature vector of each alarm instance in the high-contribution alarm cluster as a point in a high-dimensional space, constructing a multi-scale Vietoris-Rips complex by gradually increasing the neighborhood radius, recording the birth and death process of connected components as the radius increases, and generating a persistent coherent barcode; Step S82, analyzing the lifetime of the longest bar and the topological entropy of the barcode in the persistent coherent barcode. If the lifetime of the longest bar exceeds the historical statistical threshold and the topological entropy is lower than the set entropy threshold, then the cluster is determined to be a defect evolution chain with a stable topological structure; Step S83, for the cluster determined to be a defect evolution chain, calculating the topological centrality of each node in the cluster in the maximum connected component of the chain, and calculating the main alarm selection index according to the preset main alarm selection index formula based on the alarm timestamp of each node and its consistency score ranking in the cluster; Step S84, selecting the alarm instance with the highest main alarm selection index as the main alarm.

[0019] Secondly, embodiments of the present invention provide a defect repetition alarm screening system based on Siamese networks and topology analysis. The system is implemented using the aforementioned defect repetition alarm screening method based on Siamese networks and topology analysis. The system includes: a robust potential energy reference surface generation module, suitable for generating a historical initial matching potential energy value sequence based on historical alarm image pairs using a pre-trained Siamese network; the module removes pseudo-stable fluctuations caused by periodic interference from the historical initial matching potential energy value sequence using a hierarchical Bayesian model and empirical mode decomposition, and constructs a robust potential energy that evolves over time through kernel density estimation and interquartile range. The system includes a reference surface; a high-dimensional heterogeneous feature tensor extraction module, suitable for performing multimodal feature extraction on the current alarm pair image to obtain three types of features for each alarm image: spatial distribution map, image texture descriptor, and segmentation region statistics. These three types of features are then integrated into a high-dimensional heterogeneous feature tensor for the alarm image. An initial matching potential value generation module is used to input two high-dimensional heterogeneous feature tensors into the pre-trained Siamese network. Each branch of the Siamese network's bi-branch symmetric feature extraction architecture processes these tensors independently, converting them into two fixed-length binary vectors. A one-dimensional feature vector is generated; the feature vector is projected onto the tangent space of the Riemannian manifold and fused with the geodesic distance and matching potential energy to generate a current initial matching potential energy value that is resistant to environmental interference; a potential repeated alarm cluster generation module is used to map the current initial matching potential energy value to the robust potential energy reference surface to obtain residuals, construct a dynamic weighted coupling graph by combining the spatiotemporal context information of the current alarm pair, and perform iterative reasoning on the dynamic weighted coupling graph through a cyclic belief propagation algorithm to filter out potential repeated alarm clusters with high structural and semantic coordination; a high contribution alarm cluster generation module is used to generate the potential repeated alarm clusters. Alarm instances within an alarm cluster undergo time-series alignment and multi-dimensional consistency assessment to generate a consistency score. Nodes within a potentially duplicate alarm cluster are considered game participants. By calculating their Shapley value for the consistency score and setting a dynamic threshold, abnormal interference nodes with contributions below the dynamic threshold are eliminated to obtain high-contribution alarm clusters. The duplicate alarm filtering module is suitable for performing topological data analysis on the high-contribution alarm clusters to verify their structural stability, identifying stable defect evolution chains, and selecting a main alarm from the defect evolution chain based on a preset main alarm selection index. The remaining alarms are marked as duplicate alarms and filtered out.

[0020] Thirdly, embodiments of the present invention also provide an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the above-described defect duplicate alarm screening method based on twin networks and topology analysis.

[0021] Fourthly, embodiments of the present invention also provide a readable storage medium, which, when the instructions in the storage medium are executed by the processor of an electronic device, enables the electronic device to perform the above-described defect duplicate alarm screening method based on twin networks and topology analysis.

[0022] The advantages of this invention compared to the prior art are as follows:

[0023] (1) High accuracy: Robust deep features are extracted through Siamese networks and combined with traditional visual and semantic features, providing a fine basis for similarity measurement.

[0024] (2) Strong environmental adaptability: The time-varying reference surface constructed in the offline stage can filter out periodic environmental interference, and the Riemann manifold projection in the online stage further enhances the robustness of the model to nonlinear changes.

[0025] (3) Strong global correlation analysis capability: By introducing graph model and belief propagation algorithm, the correlation of alarm groups can be discovered from the perspective of global spatiotemporal context, avoiding the limitations of isolated judgment.

[0026] (4) Outstanding noise resistance and purification capabilities: Through time alignment and Shapley value analysis, it can effectively identify and eliminate noise alarms that have mistakenly entered the cluster, ensuring the purity of the core evolutionary chain.

[0027] (5) Good decision interpretability: The topological stability is finally verified based on persistent cohomology, and the main alarm is scientifically elected by combining multiple indicators. The whole decision-making process has a clear mathematical and geometric interpretation and conforms to the operation and maintenance logic.

[0028] (6) High degree of systematization: It has formed a complete closed loop of "offline learning benchmark - online multi-level reasoning", realizing full-process automation from alarm access to final screening. Attached Figure Description

[0029] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0030] Figure 1 This is a flowchart of a defect duplicate alarm screening method based on twin network and topology analysis provided in Embodiment 1 of this application.

[0031] Figure 2 This is a schematic diagram of a defect repetition alarm screening system based on twin network and topology analysis provided in Embodiment 2 of this application.

[0032] Figure 3 This is a partial block diagram of the electronic device provided in Embodiment 3 of the present invention. Detailed Implementation

[0033] 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, not all, of the embodiments of the present invention. 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.

[0034] Example 1

[0035] like Figure 1 The flowchart shown is a method for filtering out duplicate defect alarms based on twin networks and topology analysis provided in this application embodiment.

[0036] As an example, the method includes: an offline modeling stage for generating a baseline model for decision support based on historical alarm data, the offline modeling stage including: step S1, generating a sequence of historical initial matching potential energy values ​​using a pre-trained Siamese network based on historical alarm image pairs; step S2, removing pseudo-stable fluctuations caused by periodic interference from the historical initial matching potential energy value sequence based on a hierarchical Bayesian model and empirical mode decomposition, and constructing a robust potential energy reference surface that evolves over time through kernel density estimation and interquartile range; and an online processing stage for processing the current alarm. The online processing stage, which removes duplicate alarms, includes the following steps: Step S3: Multimodal feature extraction is performed on the current alarm image to obtain three types of features: spatial distribution map, image texture descriptor, and segmentation region statistics. These three types of features are then integrated into high-dimensional heterogeneous feature tensors for the alarm image. Step S4: The two high-dimensional heterogeneous feature tensors are input into the pre-trained Siamese network. Each branch of the Siamese network's two-branch symmetric feature extraction architecture processes the two high-dimensional heterogeneous feature tensors independently, converting them into fixed-length tensors. Two one-dimensional feature vectors; Step S5: Project the feature vectors onto the tangent space of the Riemannian manifold and fuse the geodesic distance and matching potential energy to generate a current initial matching potential energy value resistant to environmental interference; Step S6: Map the current initial matching potential energy value to the robust potential energy reference surface to obtain the residual, and construct a dynamic weighted coupling graph by combining the spatiotemporal context information of the current alarm pair. Perform iterative reasoning on the dynamic weighted coupling graph using the cyclic belief propagation algorithm to filter out potential duplicate alarm clusters with high structural and semantic coordination; Step S7: Within the potential duplicate alarm clusters... The alarm instances are time-series aligned and multi-dimensionally consistent to generate a consistency score. Nodes within potential duplicate alarm clusters are considered as game participants. By calculating their Shapley values ​​for the consistency score and setting a dynamic threshold, abnormal interference nodes with contributions lower than the dynamic threshold are eliminated to obtain high-contribution alarm clusters. Step S8: The high-contribution alarm clusters are subjected to topological data analysis to verify their structural stability, a stable defect evolution chain is identified, and a main alarm is selected from the defect evolution chain based on a preset main alarm selection index. The remaining alarms are marked as duplicate alarms and filtered out.

[0037] In some feasible implementations, the steps of step S1, which involves generating a sequence of historical initial matching potential values ​​using a pre-trained Siamese network based on historical alarm image pairs, are the same as those of step S5. For details, please refer to the explanation of steps S3-S5 below.

[0038] In some feasible implementations, step S2 includes: step S21, based on the historical initial matching potential energy value sequence, learning the distribution law of matching potential energy values ​​of different equipment types under different environmental conditions through a hierarchical Bayesian model to obtain a conditional probability density function; step S22, for the historical initial matching potential energy value sequence of each equipment, identifying and removing pseudo-stable fluctuation components caused by periodic environmental interference through empirical mode decomposition, and reconstructing the corrected potential energy sequence of each equipment; step S23, aggregating the corrected potential energy sequences of all equipment under that type, and based on the distribution law represented by the conditional probability density function, performing kernel density estimation and interquartile range analysis on the aggregated corrected potential energy sequence to construct a robust potential energy reference surface that evolves slowly with the inspection time; wherein, the reference surface is used to provide a dynamic benchmark for judging whether a new alarm conforms to the repetitive pattern.

[0039] Preferably, the objective of step S2 is to address the issue that the initial matching potential energy value is affected by periodic environmental disturbances (such as daily variations in sunlight and peak-valley cycles in equipment load), which can lead to "pseudo-stable fluctuations" (such as the difference in potential energy values ​​for the same defect under strong / weak light). This step constructs a robust potential energy reference surface that evolves over time through "historical data modeling - disturbance removal - dynamic surface generation," providing an objective benchmark for new alarm judgment.

[0040] Preferably, in step S21, based on the historical initial matching potential energy value sequence, a hierarchical Bayesian model is used to learn the distribution law of matching potential energy values ​​of different equipment types under different environmental conditions, and the conditional probability density function is obtained. This includes: to improve the environmental adaptability and equipment generalization ability of the recurring alarm identification system, a labeled image pair sample library is constructed covering various equipment types T (transformers, circuit breakers, insulators, etc.), lighting conditions L (strong light, weak light, cloudy, nighttime, etc.), and resolution levels R (high definition, standard definition, blurry, etc.) to ensure the representativeness of the data in real inspection scenarios. Subsequently, a trained twin network dual-branch architecture is used to perform batch inference on all collected samples to extract the initial matching potential energy value. This generates a large-scale potential energy dataset containing metadata. Based on this, a three-layer Bayesian hierarchical model is established: the first layer groups devices by type, assuming the initial matching potential energy values ​​follow a Beta distribution and fitting its shape parameters. The second layer introduces illumination and resolution as covariates, and models the model using a super-prior model. and A dynamic adjustment mechanism adapts to environmental changes; the third layer sets prior distributions for the global mean and variance to enable knowledge transfer between small sample devices. Finally, the posterior distribution is solved using the Markov chain Monte Carlo method to obtain the conditional probability density, which is used to characterize the statistical regularity of normal repetitive patterns under different operating conditions.

[0041] Specifically, the first layer (data layer) is grouped by device type, assuming an initial matching potential value. Following a Beta distribution (a probability distribution fitting the [0,1] interval), the shape-fit shape parameters for each device type are estimated using variational inference or Markov chain Monte Carlo (MCMC) methods. The second layer (operating condition layer) introduces illumination (L) and resolution (R) as covariates, and models the model using a priori model. and Dynamic adjustment mechanisms in response to environmental changes (such as under strong light) Increase the weight of Euclidean distance); third layer (global layer): set The global prior distribution (such as a normal distribution) is used to achieve knowledge transfer from small sample devices (e.g., using a large amount of transformer data to assist in insulator model training). Solution method: The posterior distribution is solved using the Markov Chain Monte Carlo (MCMC) algorithm, outputting the conditional probability density function. To characterize the potential energy distribution pattern of normal repetitive modes under different working conditions.

[0042] Preferably, step S22, for each device's historical initial matching potential energy value sequence, identifies and removes pseudo-stable fluctuation components caused by periodic environmental disturbances through empirical mode decomposition, and reconstructs the corrected potential energy sequence for each device, including: the initial matching potential energy time sequence accumulated by a certain device during periodic inspections. To separate the true defect evolution trend from pseudo-stationary fluctuations caused by environmental disturbances, empirical mode decomposition (EMD) is employed for adaptive time-frequency analysis of the non-stationary signal. This decomposes the signal into multiple intrinsic mode functions (IMFs) and a trend term, with each IMF representing an oscillation component at a different time scale. Hilbert transform analysis is used to analyze the dominant frequency characteristics of each IMF. If the period is highly correlated with the daily variation in sunlight or the peak and trough of equipment load, it is identified as a pseudo-similar signal caused by external disturbances. During reconstruction, such components are removed, while high-frequency noise, low-frequency degradation-related components, and the trend term are retained. These components are then superimposed to generate a corrected potential energy sequence. This effectively filters out periodic environmental interference and enhances the sequence's ability to represent changes in the actual device state.

[0043] Specifically, during periodic inspections of power equipment, the similarity (matching potential value) of its defect characteristics forms a fluctuating curve over time. This curve is actually a mixture of two types of changes: Real changes: the genuine trend of the equipment defect itself slowly deteriorating, repairing, or stabilizing. False changes: periodic fluctuations caused by external environmental factors such as solar cycles (day-night cycle) and equipment load cycles (peak and off-peak electricity consumption), which are not inherently defective. Without processing, this periodic interference can "contaminate" the data, causing the system to mistakenly interpret short-term environmental fluctuations as genuine changes in the defect state, leading to misjudgments. Traditional Fourier transform excels at handling stationary periodic signals, but equipment state signals are typically non-stationary (frequency and amplitude vary over time). EMD was developed for this purpose; it adaptively peels away layers of complex signals, decomposing them into a series of intrinsic mode functions (IMFs) arranged from high to low frequency, and a final trend term. IMF: Each IMF represents a specific time-scale oscillation mode in the signal. The first IMF has the highest frequency (e.g., random noise), and subsequent IMFs gradually decrease in frequency (e.g., daily or weekly cycles). The trend term represents the slowest change in the signal, typically corresponding to the long-term evolution of equipment conditions (e.g., the slow deterioration of defects). In other words, through EMD, the original signal is broken down into "parts," allowing us to analyze each component independently. For each IMF component, its instantaneous frequency can be calculated using Hilbert transform, thus analyzing its main periodic characteristics.

[0044] For example, known environmental cycles are: Diurnal cycle of sunlight: approximately 24 hours; Diurnal cycle of equipment load: approximately 24 hours (peak electricity consumption usually occurs during the day); Weekly cycle of load: approximately 168 hours (load patterns differ between weekdays and weekends). We compare the principal cycle of each IMF component with these known environmental cycles. If the cycle of a certain IMF (e.g., 23.5 hours) closely matches the diurnal cycle of sunlight, it is highly likely to be a "pseudo-signal" caused by diurnal light variations, rather than a true reflection of defects. Once these "pseudo-signals" (IMF components highly correlated with environmental cycles) are identified, they are removed during signal reconstruction. The remaining components include: High-frequency noise IMFs: representing random disturbances, usually retained; Low-frequency degradation-related IMFs: potentially corresponding to the slow process of equipment aging and defect evolution, which are key information; Trend term: the overall direction of long-term changes in equipment condition. By re-superimposing these retained components, the corrected potential energy sequence is obtained. This new sequence appears to have been "de-noiseed" and "de-periodized," stripping away environmental interference to the greatest extent possible, thus highlighting the true evolutionary information of the device's state.

[0045] In some feasible implementations, step S23, aggregating the corrected potential energy sequences of all devices under a given device type, and performing kernel density estimation and interquartile range analysis on the aggregated corrected potential energy sequences based on the distribution law represented by the conditional probability density function, constructs a robust potential energy reference surface that evolves slowly with inspection time; wherein, the reference surface is used to provide a dynamic benchmark for determining whether a new alarm conforms to a recurring pattern, including: taking device type T as the unit, collecting corrected matching potential energy sequences from multiple devices of the same type at different time periods. Aggregate analysis was performed. To characterize the typical behavior pattern of recurring alarms in this type of device under normal operating conditions, all devices belonging to type T were analyzed. Kernel density estimation is performed on numerical datasets by stacking kernel functions around each data point to smoothly estimate the overall probability density function. From the estimated probability density function, identify its global peak location. This serves as a typical potential energy level for recurring alarms in this type of equipment. To further define the normal fluctuation range, the interquartile range (IQR) is introduced to define the robustness interval for this equipment type. Where k is an empirical coefficient. Considering the slow degradation of equipment performance over time, its normal matching behavior will drift; therefore, the above static interval is extended to a time-dependent robust potential energy reference surface. Specifically, this can be achieved by calculating local values ​​through a sliding window along the time dimension. and A low-order polynomial is fitted to construct upper and lower boundary envelopes that change slowly over time. This reference surface serves as the core benchmark for determining whether a new alarm is a repeat: when the matching potential value of a newly acquired alarm falls within... If the alarm is within the specified range, it is determined to conform to the historical repetitive pattern and is considered a repetitive alarm; if it deviates significantly, it indicates an abnormal status or a new defect, requiring in-depth investigation.

[0046] Specifically, the core idea of ​​step S23 is to learn the "normal pattern" of recurring alarms for each type of equipment and its reasonable fluctuation range from historical data, and to consider the slow changes caused by equipment aging, ultimately forming a "benchmark" that evolves over time. This includes: Step 1: Data Preparation – Aggregating historical data by equipment type: Object: Multiple devices of the same type (denoted as T) (e.g., all transformers of the same model). Data: "Corrected matching potential energy sequences" generated during inspections of these devices at different time periods in the past. This sequence has undergone pre-processing (e.g., empirical mode decomposition) to filter out periodic environmental interference such as sunlight and load, better reflecting the true state of equipment defects. Operation: Gathering the historical corrected potential energy values ​​of all devices belonging to type T to form a large dataset.

[0047] Step 2: From the aggregated historical data, find out the most frequently occurring matching potential value (i.e., typical value) for this type of device during "normal recurring alarms".

[0048] Method: Kernel density estimation is employed. This is a nonparametric statistical method used to estimate the probability density function of a random variable. Unlike parametric estimation, it does not assume that the data follows a specific distribution (such as a normal distribution). Instead, it places a smooth "mound" (i.e., the kernel function, often a Gaussian bell curve) at each data point, and then superimposes all the "mounds" to form a smooth, continuous probability density curve. It can more faithfully and flexibly reflect the true distribution shape of the data, especially when the distribution is multimodal or asymmetrical. Based on this, a curve describing the "probability of the occurrence of matching potential energy values" is obtained. The potential energy value corresponding to the highest point of the curve (global peak) is the most concentrated and frequently occurring value in historical data, which is identified as the typical potential energy level of normal recurring alarms for this type of equipment. It can be understood as the "standard similarity" of the recurring defects of this type of equipment in terms of image features.

[0049] Step 3: Knowing only the typical value is not enough; we also need to define a reasonable fluctuation range around the typical value to accommodate normal random fluctuations.

[0050] Method: We introduce the robust statistic of interquartile range (IQR). First, we sort all the data in ascending order and find the first quartile (Q1, 25th quartile) and the third quartile (Q3, 75th quartile). The IQR = Q3 - Q1 describes the dispersion of the middle 50% of the data and is insensitive to extreme values.

[0051] Interval construction: A robust interval is defined as... Where k is an empirical coefficient (usually taken as 1.5, but can be adjusted according to the actual data distribution). Values ​​outside this interval can be considered potential outliers. This interval provides a static, time-independent "corridor of normal behavior," initially defining the range of potential values ​​that conform to historical recurring patterns.

[0052] Step 4: As equipment performance slowly degrades over time (e.g., insulation aging, component wear), its "normal" matching potential energy level may drift overall (e.g., the similarity of image features of the same defect may slightly decrease with aging). Static ranges cannot capture this slow trend.

[0053] Therefore, the static analysis method described above is applied "sliding" along the time axis to construct a dynamic reference surface that changes slowly over time. The sliding window is defined as follows: a fixed-length window (e.g., every 6 months) is set along the time axis and slids within the window. Local statistics: Within each time window, only historical corrected potential energy data falling within that window are used to recalculate the local typical potential energy level and local robustness interval. Curve fitting: The local typical values, local upper boundaries, and local lower boundaries corresponding to the center points of all time windows are extracted, forming three scatter plot sequences that change over time. Surface generation: These three sequences (typical value sequence, upper boundary sequence, and lower boundary sequence) are fitted with low-order polynomials (such as quadratic or cubic curves) to obtain three smooth, continuous function curves with respect to time. These three curves together constitute a band-shaped robust potential energy reference surface. The middle curve represents the evolution trajectory of the typical potential energy level, while the upper and lower curves constitute the boundaries of a dynamic "normal behavior corridor."

[0054] For a specific example, when a new alarm occurs, its device type T and alarm occurrence time t are first determined. On the reference surface corresponding to type T, the upper and lower boundary values ​​corresponding to time t are queried. The matching potential value of the new alarm is compared with this dynamic range. Decision logic: If the matching potential value of the new alarm is within the dynamic range, it indicates that the matching potential of the new alarm is highly consistent with the historical recurring pattern of the device during the same period, supporting the determination of it as a recurring alarm. If the matching potential value of the new alarm is significantly lower than the lower boundary value or significantly higher than the upper boundary value, it indicates that its behavior deviates from the historical normal recurring pattern. This suggests two possibilities: either the device state has changed abnormally (such as a defect rapidly worsening), or a completely new defect (non-recurring) has appeared. Either way, a focused investigation is required.

[0055] In some feasible implementations, the current alarm image in step S3 is an image collected at different times from the same patrol point. Step S3 includes: Step S31, after grayscale processing, applying gamma correction to uniform image brightness, and using the Sobel operator to calculate the gradient values ​​in the horizontal and vertical directions for each pixel in the image; Step S32, dividing the image into 8×8 pixel cells, generating a 9-dimensional directional histogram vector in each cell based on the gradient direction and size of the pixel, and stitching them together to form a spatial distribution map; Step S33, extracting the contrast, energy, and correlation of the grayscale co-occurrence matrix, averaging them in multiple directions to obtain the image texture descriptor; Step S34, calling a pre-trained semantic segmentation model, assigning a category label to each pixel in the image to generate a pixel-level category mask, and calculating the mean, variance, and area ratio of pixel values ​​in each region according to the device region as segmentation region statistics.

[0056] Preferably, images of the same inspection point acquired at different times are paired. and Multimodal feature extraction is performed as follows: First, the local gradient orientation histogram is calculated. After grayscale processing, gamma correction is applied to uniform image brightness. For each pixel in the image, the Sobel operator is used to calculate its horizontal and vertical gradient values. The image is divided into 8×8 pixel cells. Within each cell, a 9-dimensional orientation histogram vector is generated based on the gradient direction and magnitude of the pixel, and these vectors are concatenated to form a spatial distribution map. Second, the distance between pixels is set, and three statistics—contrast, energy, and correlation—of the gray-level co-occurrence matrix are extracted and averaged across multiple directions to serve as the image texture descriptor. Finally, a pre-trained semantic segmentation model is called to assign a category label to each pixel in the image, generating a pixel-level category mask. The mean, variance, and area ratio of pixel values ​​in each region are statistically analyzed according to the device region. The spatial distribution map, image texture descriptor, and segmentation region statistics are integrated into a unified high-dimensional heterogeneous feature tensor. and The dimension is H×W×C, which serves as the input to the two branches of the Siamese network. H represents the height of the tensor, i.e. the number of pixels in the vertical direction, W represents the width of the tensor, i.e. the number of pixels in the horizontal direction, representing the spatial resolution, and C is the number of feature channels, i.e. the number of different feature dimensions contained at each spatial location.

[0057] In some feasible implementations, step S4 includes: step S41, converting the high-dimensional heterogeneous feature tensor... and The symmetric dual-branch structure of the pre-trained Siamese network is input, and two one-dimensional feature vectors are obtained by mapping through convolution and pooling operations with shared weights. and Step S5 includes: Step S51, calculating the weighted Euclidean distance and cosine similarity between the two one-dimensional feature vectors to obtain the matching potential energy; Step S52, projecting the two one-dimensional feature vectors onto the Riemannian manifold space, calculating the geodesic distance and fusing it with the matching potential energy to generate the current initial matching potential energy value that is resistant to environmental interference.

[0058] Preferably, in the dual-branch structure of a Siamese network, the two branches share the exact same network architecture and weight parameters, ensuring that the input image is processed as expected. and The feature extraction process is strictly symmetric. When high-dimensional heterogeneous feature tensors... and After inputting into the two branches, lightweight convolutional layers are used for dimensionality reduction to enhance the semantic response capability at each location. Subsequently, batch normalization and ReLU activation are applied to improve the non-linear expressiveness of the features and training stability. Finally, global average pooling is performed to convert the feature map into a fixed-length one-dimensional feature vector. and Based on this, a matching potential is defined to quantify the similarity between two images in the feature space. The specific formula is:

[0059] ;

[0060] in, The weights of the Euclidean distance term are... For vectors The model, For vectors The model, Represents the geometric distance between two eigenvectors, with the denominator introducing... Dynamic normalization is implemented to prevent misjudgment due to excessive differences in feature amplitude. For vectors and Cosine similarity between them The dot product of vectors reflects the degree of coordination between two vectors in a direction. The transposed vector When two images have highly similar features, the matching potential... If the value is close to 1, the difference is significant, and it approaches 0.

[0061] To suppress environmental interference such as lighting and imaging conditions, the feature vector is... and Treating points as points on a Riemannian manifold, we project them onto a local tangent space using local linear embedding, calculate the geodesic distance reflecting the real nonlinear structure, and then adaptively weight and fuse this distance with the matching potential using the Sigmoid function to generate an enhanced matching score that is robust to environmental changes. Finally, we use neighborhood-preserving embedding to reduce the dimensionality of the fused high-dimensional features, remove redundant information and retain the local geometric structure, obtain a low-dimensional compact representation, and normalize it to the initial matching potential value.

[0062] In some feasible implementations, step S6 includes: step S61, calculating the difference between the current initial matching potential energy value and the expected value of the robust potential energy reference surface of its corresponding equipment type at the same time, as the core residual; step S62, standardizing and splicing the core residual with the geographical location, inspection order, time interval, quantified meteorological factors, and equipment maintenance cycle information of the current alarm pair to form a graph node attribute vector; step S63, establishing connection edges between nodes based on the rule that the feature similarity between nodes is higher than a first threshold and the spatiotemporal distance is lower than a second threshold. The weights are dynamically calculated based on feature similarity and time proximity to construct a dynamic weighted coupling graph; Step S64: Define a global energy function containing unary and binary terms, and perform iterative reasoning on the dynamic weighted coupling graph using a ring belief propagation algorithm to calculate the posterior probability of each node being a repeat alarm; wherein, the unary term is based on the node core residual, and the binary term is based on the penalty for label differences based on attribute similarity between connected nodes; Step S65: Identify the largest connected subgraph with a posterior probability higher than the confidence threshold and an average core residual lower than the residual threshold as the potential repeat alarm cluster.

[0063] Preferably, step S61 includes: after completing the construction of the robust potential energy reference surface for the device type, for each newly generated alarm pair, firstly mapping its initial matching potential energy value to the reference surface of its device type at the corresponding time point, and calculating the residual. The residual is the core indicator for determining whether it is abnormal or a new defect.

[0064] Preferably, step S62 includes: to further enhance the context awareness capability of the judgment, the system simultaneously extracts multiple auxiliary information: including the precise geographical coordinates pos of the alarm point, the path sequence number seq in the inspection task, and the time interval between the last alarm of the adjacent point. Meteorological influencing factors This is to compensate for the negative impact of severe weather on image quality and matching stability, as well as the historical maintenance closed-loop cycle C of the equipment. All these values ​​are standardized to eliminate dimensional differences and then concatenated into a structured multidimensional attribute vector. , as a node in the graph model.

[0065] Preferably, step S63 includes: constructing a dynamic weighted coupling graph G=(V,E) to achieve global consistency modeling of large-scale inspection alarm data, wherein each node in the node set V This represents a specific alarm instance, whose characteristics are defined by the aforementioned multidimensional attribute vector. Full description. The edge set E connects alarm pairs belonging to the same defect evolution process by comprehensively judging spatiotemporal proximity and feature similarity. If two alarms are geographically close (e.g., the geographical distance between two alarms is less than 5m), have similar inspection times (e.g., the difference in alarm times between two alarms is less than 30 days), and the cosine similarity between their feature vectors is higher than a set threshold (e.g., 0.7), then an undirected edge is established between them. And assign dynamic weights to the edge. This indicates the overall confidence level of the match between the alarms.

[0066] Specifically, suppose the inspection system detects 5 alarms (A, B, C, D, E) at a substation. The graph construction process is as follows: Calculate attributes: Calculate the geographical location, timestamp, and feature vector for each alarm. Perform pairwise comparisons:

[0067] A and B: distance 3 meters, time difference 5 days, cosine similarity 0.85 → construct edges;

[0068] A and C: Distance 20 meters (exceeds threshold) → No edge built;

[0069] B and C: distance 4 meters, time difference 8 days, cosine similarity 0.78 → construct edges;

[0070] C and D: distance 2 meters, time difference 25 days, cosine similarity 0.9 → construct edges;

[0071] D and E: distance 1 meter, time difference 40 days (exceeds threshold) → No edge built

[0072] ...The same applies to other combinations.

[0073] Graph structure formation: This may ultimately result in two connected subgraphs (clusters): Cluster 1: A—B—C—D (connected by edges); Cluster 2: E (isolated node, potentially a completely new defect). The system automatically categorizes A, B, C, and D as a group of "multiple alarms suspected of being for the same defect," while E is listed separately. Subsequent steps will perform more refined analysis on Cluster 1 (such as time alignment and noise removal) to ultimately confirm whether they are truly duplicate alarms and select the primary alarm.

[0074] Preferably, step S64 includes: based on this graph structure, defining a global energy function to quantify the configuration rationality of the entire graph, and inferring the label of each node, specifically using the following formula:

[0075] ;

[0076] The goal of the formula is to find a set of label configurations. This makes the total energy Minimum. Let be the matching potential residual of node i. The adjustment coefficient controls the weight of environmental disturbances in node energy. This is a meteorological interference factor used to enhance the penalty for mismatches under unfavorable imaging conditions. The overall smoothness strength coefficient controls the influence of the prior structure of the control chart. Let be the feature similarity between node i and node j. >1 is used to penalize low similarity in non-linear enhancement. We measure the contextual similarity between nodes i and j in the multidimensional feature space and ensure that enhanced smoothing is applied only between nodes whose features and contexts are similar. That is, if the two nodes have different labels, one is a repeat and the other is a non-repeat, they will suffer a large energy penalty, forcing them to be assigned the same label. Let be the Euclidean distance between the multidimensional attribute vectors of nodes i and j. This is the scale parameter. This is the activation penalty when two nodes are assigned different labels.

[0077] To further achieve globally consistent reasoning, a ring-shaped belief propagation algorithm is used for approximate probabilistic inference on the graph. The prior probability of each node belonging to the repeating alarm category is initialized based on the matching potential residual of each node, such that nodes with smaller residuals have higher initial confidence. Subsequently, iterative message passing is performed along each edge of the graph. Each node collects messages from its neighbors, updates the edge probability of each node being a repeating or non-repeating alarm, and updates its posterior probability estimate for its own label by combining its own prior and edge weights. The message update rule follows the Bayesian principle, propagating the current optimal belief to neighbors while minimizing local energy. A fixed number of iterations is set. After each iteration, for each connected subgraph, the proportion of high-confidence nodes whose posterior probability is greater than a threshold is counted. If this proportion exceeds 70%, and the average potential residual of all nodes in the subgraph is less than the set threshold, then the subgraph is determined to be a potential repeating alarm cluster, i.e., a set of alarms highly likely to originate from the same defect captured at different times.

[0078] In some feasible implementations, step S7 includes: Step S71, sorting the alarm instances contained in the potential duplicate alarm cluster by collection time, constructing a corresponding feature vector sequence for each alarm instance using the high-dimensional embedding vector output by the Siamese network during the feature extraction stage, and constructing a feature trajectory based on the feature vector sequence; Step S72, using the earliest alarm feature trajectory as a benchmark, using a dynamic time warping algorithm to align the remaining feature trajectories with it to find the optimal alignment path, and calculating the minimum cumulative distance and the standard deviation of the slope change of the alignment path; Step S73, fusing the minimum cumulative distance and the standard deviation of the slope change of the alignment path. The standard deviation of the slope change, the stability index of the semantic segmentation region between alarms, and the phase offset of the alarm time in the device load cycle are used to calculate the geometric, semantic, and temporal triple consistency score; Step S74: Each alarm instance in the cluster is regarded as a participant in a cooperative game, and the total payoff of its alliance is the triple consistency score. The marginal contribution value of each alarm instance is calculated using the Shapley value formula; Step S75: A dynamic threshold is set based on the mean and standard deviation of the Shapley values ​​of all alarm instances. Alarm instances with contribution values ​​lower than the dynamic threshold are identified as noise nodes and removed. The remaining alarm instances constitute the high contribution alarm cluster.

[0079] Preferably, step S71 includes: for each identified potential duplicate alarm cluster, arranging the alarm instances contained therein in ascending order according to timestamps to form a time-series sequence. Subsequently, using the high-dimensional embedding vector output by the Siamese network in the feature extraction stage, constructing a corresponding feature vector sequence for each alarm, and concatenating these vectors in chronological order to form a feature trajectory.

[0080] Preferably, step S72 includes: selecting the earliest alarm as the reference trajectory, aligning all other trajectories sequentially with it using a dynamic time warping algorithm, and calculating the corresponding minimum cumulative distance, which reflects the overall degree of geometric deformation from the reference state to the current observation. Simultaneously, the standard deviation of the alignment path slope change is calculated to quantify the jitter or variability during the alignment process: if the standard deviation is significantly high, it indicates that some inspection points have imaging anomalies, occlusions, or sudden changes in equipment status, thereby disrupting the normal evolutionary pattern. The minimum cumulative distance and the standard deviation of the alignment path slope change together constitute a quantitative assessment of the stability of the repeating pattern.

[0081] Preferably, step S73 includes: introducing the external device's operating load curve and analyzing the relative position of the alarm occurrence time within the load cycle. Specifically, the main cycle of the load signal is extracted through Fourier analysis, and the phase shift of each alarm time relative to that cycle is calculated. If the alarm is close to the load peak, it indicates that the alarm occurred during the period of greatest thermal or electrical stress on the equipment. The resulting imaging anomalies more accurately reflect defect heating or deformation, thus making the judgment more reliable. Based on this, a triple consistency scoring system is constructed to quantify the stability of repetitive patterns from different dimensions, including:

[0082] Geometric consistency: The minimum cumulative distance is normalized and mapped to the [0,1] interval. Combined with the standard deviation of slope change, the smaller the standard deviation of slope change and the smaller the distance, the smoother the evolution of the feature trajectory and the higher the morphological consistency, resulting in a higher geometric consistency score. Specifically, the geometric consistency score is calculated as follows:

[0083] ;

[0084] in, , It is a preset maximum cost threshold used for linear normalization. For the minimum cumulative distance, The standard deviation of the slope variation This is the preset slope fluctuation tolerance threshold.

[0085] Semantic consistency: The target device region is extracted using a semantic segmentation model. The intersection-union ratio (IUU) sequence of segmentation results between adjacent alarms is calculated, and the average of its moving Pearson correlation coefficient with the time series is taken. A higher value indicates more stable structural evolution. Specifically, the semantic consistency score is calculated as follows:

[0086] IOU sequence calculation: For the semantic segmentation results of adjacent alarms within a cluster, calculate their Intersection over Union (IOU) sequence. Calculate the moving Pilcarson correlation coefficient between the IOU sequence and the time series, and take the mean as the semantic consistency score. .

[0087] Timing consistency: defined as When the phase offset is close to 0, that is, when the alarm occurs near the load peak, the timing consistency score is the highest, reflecting the prior knowledge of the fault-prone window.

[0088] The three factors are weighted together to obtain a triple consistency score, which improves the accuracy of identifying true repeating alarms and reduces misjudgments caused by environmental fluctuations or non-faulty load changes.

[0089] Preferably, step S74 includes: treating alarm instances within each repeating alarm cluster as participants in a game, whose cooperative goal is to jointly improve the triple consistency score of the cluster. To quantify the marginal contribution of each alarm to overall performance, a Shapley value method is used for fair allocation: enumerating all possible subset combinations, calculating the average improvement in the triple consistency score after an alarm is added to different alliances, and calculating its expected value under all permutations to obtain the Shapley value of that node.

[0090] Preferably, step S75 includes: setting a dynamic removal threshold, identifying low-contribution nodes with Shapley values ​​below the threshold as noise outliers and removing them. Nodes with high Shapley values ​​are retained to form a high-contribution alarm cluster, providing high-quality input for subsequent primary alarm selection and closed-loop management. The dynamic removal threshold is dynamically calculated based on the distribution of Shapley values ​​(such as mean and standard deviation) of all nodes in the current cluster. For example, the threshold can be set as: mean - β × standard deviation, which allows the decision to adapt to the characteristics of different clusters.

[0091] In some feasible implementations, step S8 includes: Step S81, treating the feature vector of each alarm instance in the high-contribution alarm cluster as a point in a high-dimensional space, constructing a multi-scale Vietoris-Rips complex by gradually increasing the neighborhood radius, recording the birth and death process of connected components as the radius increases, and generating a persistent homogeneous barcode; Step S82, analyzing the lifetime of the longest bar and the topological entropy of the barcode in the persistent homogeneous barcode. If the lifetime of the longest bar exceeds the historical statistical threshold and the topological entropy is lower than the set entropy threshold, the cluster is determined to be a defect evolution chain with a stable topological structure; Step S83, for the cluster determined to be a defect evolution chain, calculating the topological centrality of each node in the cluster in the largest connected component of the chain, and calculating the main alarm selection index according to the preset main alarm selection index formula based on the alarm timestamp of each node and its consistency score ranking in the cluster; Step S84, selecting the alarm instance with the highest main alarm selection index as the main alarm.

[0092] Preferably, step S81 includes: inputting the high-contribution alarm cluster into a scale-adaptive topological clustering framework, treating each node in the high-dimensional feature space extracted from the Siamese network as a basic unit for topological analysis, and constructing VR complexes under different neighborhood radii: setting an incremental distance threshold. When the Euclidean distance between any two nodes in the feature space is less than At that time, edges are established between them, forming a graph structure. With... Starting from 0 and gradually increasing, the number of connections between nodes increases, and connected components are gradually generated and merged in the graph. The formation and disappearance scales of each connected component are recorded, and these intervals are represented in the form of a bar chart, which is the persistent homology barcode. Long bars represent stable subgroups that remain independent at large scales, reflecting potential sub-patterns or staged evolution; short bars correspond to noise or loose associations. By analyzing this topological feature, it is possible to identify whether there are multi-stage evolutions or anomalous branches within the cluster, enhancing the ability to analyze complex defect evolution paths.

[0093] Preferably, step S82 includes: in persistent homology analysis, analyzing the lifetime of the longest bar in the persistent homology barcode, i.e., the scale difference from its formation to its disappearance. If it exceeds the dynamic threshold of the normal range of historical defect clusters, then the connected component has significant persistence in the topological evolution process, indicating the existence of a long-term stable topological structure. Simultaneously, the topological entropy of all bar length distributions is calculated. If the topological entropy is less than a set threshold, it indicates that most connections are concentrated in a few long-lived components, with low noise interference. When both indicators meet the conditions, the alarm cluster is determined to have a highly consistent stable defect evolution chain and possess physically interpretable repeatability characteristics.

[0094] Preferably, steps S83 and S84 include: defining topological centrality. For nodes The sum of the reciprocals of the shortest paths to other nodes in the most connected component, combined with the first occurrence time. Construct the primary alarm selection index, the specific formula is as follows:

[0095] ;

[0096] in, This means using b to iterate through all nodes in the cluster and find the maximum value of the centrality. The attenuation coefficient controls the suppression strength of the time term. For nodes Alarm time, These represent the minimum and maximum alarm times within the cluster. The weights for the consistency scoring ranking items, Indicates the i-th alarm instance by Ranked in descending order, with the highest score ranked as 1. The total number of nodes in the high-contribution alarm cluster. (Select) The alarm with the highest value is designated as the primary alarm, and all others are marked as duplicate alarms and removed from the system.

[0097] In the above implementation, by fusing multimodal features through a Siamese network and introducing Riemannian manifold projection, feature extraction and matching become more robust to environmental interferences such as changes in illumination, differences in viewing angles, and fluctuations in imaging conditions. Hierarchical Bayesian models and empirical mode decomposition effectively eliminate "pseudo-repetition" signals caused by periodic environmental factors (such as illumination and load), allowing the matching potential energy to more realistically reflect the evolution of the defect itself. Based on this, the accuracy of identification and robustness to complex environments are significantly improved. A dynamically evolving robust potential energy reference surface is constructed based on historical data, rather than using a fixed threshold. This allows the system to automatically adjust judgment criteria according to different equipment types and different operating stages, adapting to long-term factors such as equipment aging and changes in operating conditions, significantly improving generalization ability in different scenarios. By constructing a weighted coupling graph and applying a global energy function and belief propagation algorithm, the system can comprehensively analyze the spatiotemporal and semantic relationships between alarms, filtering out repetitive alarm clusters with high structural and semantic coordination from a global perspective, avoiding fragmentation errors that may result from judging solely based on pairwise similarity. By introducing dynamic time warping and triple consistency scoring, combined with Shapley value analysis, the system can accurately identify and remove noisy nodes and abnormal observations within clusters, ensuring the purity and reliability of the final alarm clusters. Employing scale-adaptive topological clustering and persistent cohomology analysis, it identifies stable defect evolution chains at the topological level. This means that the selection of the primary alarm no longer relies on simple rules or single scores, but is based on the structural stability of the evolutionary pattern, resulting in more interpretable and stable results. The system automates the entire process from feature extraction, duplicate detection, noise removal to primary alarm selection, significantly reducing the workload of manually reviewing massive amounts of duplicate alarms and improving the closed-loop management efficiency and intelligent operation and maintenance level of power intelligent inspection.

[0098] Example 2

[0099] Please see Figure 2 This embodiment provides a structural diagram of a defect repetition alarm screening system based on twin networks and topology analysis.

[0100] As an example, the system is implemented using the defect duplicate alarm screening method based on twin network and topology analysis described in Example 1. The system includes:

[0101] The robust potential energy reference surface generation module 20 is suitable for generating a sequence of historical initial matching potential energy values ​​based on historical alarm image pairs using a pre-trained Siamese network; the historical initial matching potential energy value sequence is used to remove pseudo-stable fluctuations caused by periodic interference based on a hierarchical Bayesian model and empirical mode decomposition, and a robust potential energy reference surface that evolves over time is constructed by kernel density estimation and interquartile range.

[0102] The high-dimensional heterogeneous feature tensor extraction module 21 is suitable for performing multimodal feature extraction on the current alarm pair image to obtain three types of features corresponding to each alarm image: spatial distribution map, image texture descriptor, and segmentation region statistics. The three types of features are then integrated into the high-dimensional heterogeneous feature tensor of the alarm image.

[0103] The current initial matching potential value generation module 22 is suitable for inputting two high-dimensional heterogeneous feature tensors into the pre-trained Siamese network, and processing them independently through each branch of the dual-branch symmetric feature extraction architecture of the Siamese network to convert the two high-dimensional heterogeneous feature tensors into two one-dimensional feature vectors of fixed length; projecting the feature vectors onto the tangent space of the Riemannian manifold and fusing the geodesic distance and matching potential to generate a current initial matching potential value that is resistant to environmental interference;

[0104] The potential repeated alarm cluster generation module 23 is suitable for mapping the current initial matching potential energy value to the robust potential energy reference surface to obtain the residual, constructing a dynamic weighted coupling graph by combining the spatiotemporal context information of the current alarm pair, and performing iterative reasoning on the dynamic weighted coupling graph through the cyclic belief propagation algorithm to filter out potential repeated alarm clusters with high structural and semantic coordination.

[0105] The high-contribution alarm cluster generation module 24 is suitable for performing time-series alignment and multi-dimensional consistency evaluation on alarm instances within the potential duplicate alarm cluster, generating a consistency score, and treating nodes within the potential duplicate alarm cluster as game participants. By calculating their Shapley value for the consistency score and setting a dynamic threshold, abnormal interference nodes with a contribution lower than the dynamic threshold are eliminated to obtain the high-contribution alarm cluster.

[0106] The duplicate alarm filtering module 25 is suitable for performing topological data analysis on the high-contribution alarm cluster to verify its structural stability, identifying a stable defect evolution chain, selecting a main alarm from the defect evolution chain based on a preset main alarm selection index, and marking the remaining alarms as duplicate alarms and filtering them out.

[0107] It is not difficult to see that this embodiment is a system implementation corresponding to the first embodiment, and this embodiment can be implemented in conjunction with the first embodiment. The relevant technical details mentioned in the first embodiment are still valid in this embodiment, and will not be repeated here to reduce repetition. Accordingly, the relevant technical details mentioned in this embodiment can also be applied to the first embodiment.

[0108] It is worth mentioning that all modules involved in this embodiment are logical units. In practical applications, a logical unit can be a physical unit, a part of a physical unit, or a combination of multiple physical units. Furthermore, to highlight the innovative aspects of this invention, this embodiment does not introduce units that are not closely related to solving the technical problem proposed by this invention; however, this does not mean that other units are absent from this embodiment.

[0109] Example 3

[0110] Please see Figure 3 The present invention also provides an electronic device, including: a memory and a processor; the memory stores at least one program instruction; the processor loads and executes the at least one program instruction to implement the defect duplicate alarm screening method based on twin network and topology analysis provided in Embodiment 1.

[0111] The memory 302 and processor 301 are connected via a bus, which may include any number of interconnecting buses and bridges, connecting various circuits of one or more processors 301 and memory 302 together. The bus may also connect various other circuits, such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and therefore will not be described further herein. A bus interface provides an interface between the bus and the transceiver. The transceiver may be a single element or multiple elements, such as multiple receivers and transmitters, providing a unit for communicating with various other devices over a transmission medium. Data processed by processor 301 is transmitted over a wireless medium via an antenna, which further receives data and transmits it to processor 301.

[0112] Processor 301 is responsible for managing the bus and general processing, and can also provide various functions, including timing, peripheral interfaces, voltage regulation, power management, and other control functions. Memory 302 can be used to store data used by processor 301 during operation.

[0113] Example 4

[0114] This invention also proposes a storage medium storing a method for filtering duplicate defect alarms based on twin networks and topology analysis. When executed by a processor, the program for filtering duplicate defect alarms based on twin networks and topology analysis implements the steps of the method described above. Since this storage medium employs all the technical solutions of the above embodiments, it possesses at least all the beneficial effects brought about by the technical solutions of the above embodiments, which will not be elaborated upon further here.

[0115] The above descriptions are merely embodiments of the present invention. Commonly known structures and characteristics are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are aware of all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, based on the guidance provided in this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A method for filtering out duplicate defect alarms based on twin networks and topology analysis, characterized in that, The method includes: The offline modeling phase, used to generate a baseline model for decision support based on historical alarm data, includes: Step S1: Based on historical alarm image pairs, generate a sequence of historical initial matching potential values ​​using a pre-trained Siamese network; Step S2: Based on the hierarchical Bayesian model and empirical mode decomposition, the pseudo-stable fluctuations caused by periodic disturbances are eliminated from the historical initial matching potential energy value sequence. A robust potential energy reference surface that evolves over time is constructed by kernel density estimation and interquartile range. The online processing phase is used to process the current alarm and filter out duplicate alarms. The online processing phase includes the following steps: Step S3: Perform multimodal feature extraction on the current alarm image to obtain three types of features corresponding to each alarm image: spatial distribution map, image texture descriptor, and segmentation region statistics. Integrate the three types of features into a high-dimensional heterogeneous feature tensor of the alarm image. Step S4: Input the two high-dimensional heterogeneous feature tensors into the pre-trained Siamese network respectively, and process them independently through each branch of the dual-branch symmetric feature extraction architecture of the Siamese network to convert the two high-dimensional heterogeneous feature tensors into two one-dimensional feature vectors of fixed length. Step S5: Project the feature vectors onto the tangent space of the Riemannian manifold and fuse the geodesic distance and matching potential energy to generate the current initial matching potential energy value that is resistant to environmental interference. Step S6: Map the current initial matching potential energy value to the robust potential energy reference surface to obtain the residual. Combine the spatiotemporal context information of the current alarm pair to construct a dynamic weighted coupling graph. Perform iterative reasoning on the dynamic weighted coupling graph through the cyclic belief propagation algorithm to filter out potential duplicate alarm clusters with high structural and semantic coordination. Step S7: Perform time-series alignment and multi-dimensional consistency evaluation on alarm instances within the potential duplicate alarm cluster to generate a consistency score. Treat nodes within the potential duplicate alarm cluster as game participants. Calculate their Shapley value for the consistency score and set a dynamic threshold to remove abnormal interference nodes with a contribution lower than the dynamic threshold, thus obtaining a high-contribution alarm cluster. Step S8: Perform topological data analysis on the high-contribution alarm cluster to verify its structural stability, identify a stable defect evolution chain, and select a main alarm from the defect evolution chain based on the preset main alarm selection index, while marking the remaining alarms as duplicate alarms and filtering them out.

2. The defect duplicate alarm screening method based on twin network and topology analysis according to claim 1, characterized in that, Step S2 includes: Step S21: Based on the historical initial matching potential energy value sequence, the distribution law of matching potential energy values ​​of different equipment types under different environmental conditions is learned through a hierarchical Bayesian model to obtain the conditional probability density function; Step S22: For the historical initial matching potential energy value sequence of each device, identify and remove pseudo-stable fluctuation components caused by periodic environmental disturbances through empirical mode decomposition, and reconstruct the corrected potential energy sequence of each device. Step S23: Aggregate the correction potential energy sequences of all devices under the device type, and perform kernel density estimation and interquartile range analysis on the aggregated correction potential energy sequences based on the distribution law represented by the conditional probability density function to construct a robust potential energy reference surface that evolves slowly with the inspection time; wherein, the reference surface is used to provide a dynamic benchmark for judging whether a new alarm conforms to the repetitive pattern.

3. The defect duplicate alarm screening method based on twin network and topology analysis according to claim 1, characterized in that, The current alarm pair image in step S3 is an image collected at the same patrol point at different times. Step S3 includes: Step S31: After grayscale processing of the image, apply gamma correction to uniform image brightness. For each pixel in the image, use the Sobel operator to calculate its gradient values ​​in the horizontal and vertical directions. Step S32: Divide the image into 8×8 pixel cells. In each cell, generate a 9-dimensional directional histogram vector based on the gradient direction and size of the pixels, and stitch them together to form a spatial distribution map. Step S33: Extract the three statistics of contrast, energy and correlation of the gray-level co-occurrence matrix, and average them in multiple directions to use as the image texture descriptor; Step S34: Call the pre-trained semantic segmentation model, assign a category label to each pixel in the image to generate a pixel-level category mask, and calculate the mean, variance and area ratio of pixel values ​​in each region according to the device region as the segmentation region statistics.

4. The defect duplicate alarm screening method based on twin network and topology analysis according to claim 1, characterized in that, Step S4 includes: Step S41: Convert the high-dimensional heterogeneous feature tensor and The symmetric dual-branch structure of the pre-trained Siamese network is input, and two one-dimensional feature vectors are obtained by mapping through convolution and pooling operations with shared weights. and ; Step S5 includes: Step S51: Calculate the weighted Euclidean distance and cosine similarity between the two one-dimensional feature vectors to obtain the matching potential. Step S52: Project the two one-dimensional feature vectors onto the Riemannian manifold space, calculate the geodesic distance and fuse it with the matching potential energy to generate the current initial matching potential energy value that is resistant to environmental interference.

5. The defect duplicate alarm screening method based on twin network and topology analysis according to claim 4, characterized in that, The matching potential energy is calculated as follows: ； in, The weights of the Euclidean distance term are... For vectors The model, For vectors The model, Represents the geometric distance between two eigenvectors, with the denominator introducing... Dynamic normalization is implemented to prevent misjudgment due to excessive differences in feature amplitudes. For vectors and Cosine similarity between them The dot product of vectors reflects the degree of coordination between two vectors in a direction. The transposed vector , To match the potential energy.

6. The defect duplicate alarm screening method based on twin network and topology analysis according to claim 1, characterized in that, Step S6 includes: Step S61: Calculate the difference between the current initial matching potential energy value and the expected value of the robust potential energy reference surface of its corresponding device type at the same time, and use it as the core residual; Step S62: Standardize and splice the core residual with the geographical location, inspection sequence, time interval, quantitative meteorological factors and equipment maintenance cycle information of the current alarm pair to form a graph node attribute vector; Step S63: Establish connection edges between nodes based on the rule that the feature similarity between nodes is higher than the first threshold and the spatiotemporal distance is lower than the second threshold. The weight of the edge is dynamically calculated based on feature similarity and temporal proximity, thereby constructing a dynamic weighted coupling graph. Step S64: Define a global energy function containing unary and binary terms, and perform iterative reasoning on the dynamic weighted coupling graph using the circular belief propagation algorithm to calculate the posterior probability of each node being a repeat alarm; wherein, the unary term is based on the node core residual, and the binary term is based on the penalty for label differences based on the attribute similarity between connected nodes. Step S65: Identify the largest connected subgraph with a posterior probability higher than the confidence threshold and an average core residual lower than the residual threshold as the potential repeat alarm cluster.

7. The defect duplicate alarm screening method based on twin network and topology analysis according to claim 6, characterized in that, The mathematical expression for the global energy function is: ； In the formula, Let be the matching potential residual of node i. To adjust the coefficients, the weight of environmental disturbances in the node energy is controlled. This is a meteorological interference factor used to enhance the penalty for mismatches under unfavorable imaging conditions. The overall smoothness strength coefficient controls the influence of prior structural information on the control chart. Let be the feature similarity between node i and node j. >1 is used for non-linear enhancement of low similarity penalty. Measure the contextual similarity between nodes i and j in a multidimensional feature space. Let be the Euclidean distance between the multidimensional attribute vectors of nodes i and j. For scale parameters, Let E be the activation penalty term when two nodes are assigned different labels, and let E be the set of edges. This represents the total energy.

8. The defect duplicate alarm screening method based on twin network and topology analysis according to claim 1, characterized in that, Step S7 includes: Step S71: For the potential repeating alarm cluster, sort the alarm instances contained therein according to the collection time, and use the high-dimensional embedding vector output by the Siamese network in the feature extraction stage to construct a corresponding feature vector sequence for each alarm instance, and construct a feature trajectory based on the feature vector sequence. Step S72: Using the earliest alarm feature trajectory as a benchmark, use the dynamic time warping algorithm to align the remaining feature trajectories with it to find the optimal alignment path, and calculate the minimum cumulative distance and the standard deviation of the slope change of the alignment path. Step S73: Combine the minimum cumulative distance, the standard deviation of the slope change, the stability index of the semantic segmentation region between alarms, and the phase offset of the alarm time in the device load cycle to calculate the geometric, semantic, and temporal triple consistency score. Step S74: Treat each alarm instance within the cluster as a participant in a cooperative game, and the total payoff of the alliance is the triple consistency score. Calculate the marginal contribution value of each alarm instance using the Shapley value formula. Step S75: Set a dynamic threshold based on the mean and standard deviation of the Shapley values ​​of all alarm instances. Alarm instances with contribution values ​​lower than the dynamic threshold are identified as noise nodes and removed. The remaining alarm instances constitute the high contribution alarm cluster.

9. The defect duplicate alarm screening method based on twin network and topology analysis according to claim 1, characterized in that, Step S8 includes: Step S81: Treat the feature vector of each alarm instance in the high contribution alarm cluster as a point in a high-dimensional space, construct a multi-scale Vietoris-Rips complex by gradually increasing the neighborhood radius, record the birth and death process of connected components as the radius increases, and generate a persistent coherent barcode. Step S82: Analyze the lifetime of the longest bar in the persistent coherent barcode and the topological entropy of the barcode. If the lifetime of the longest barcode exceeds the historical statistical threshold and the topological entropy is lower than the set entropy threshold, then the cluster is determined to be a defect evolution chain with a stable topological structure. Step S83: For a cluster that is determined to be a defect evolution chain, calculate the topological centrality of each node in the cluster in the maximum connected component of the chain, and calculate the main alarm selection index according to the preset main alarm selection index formula by combining the alarm timestamp of each node and the consistency score ranking in the cluster. Step S84: Select the alarm instance with the highest priority index as the main alarm.

10. A defect duplicate alarm filtering system based on twin networks and topology analysis, wherein the system is implemented using the defect duplicate alarm filtering method based on twin networks and topology analysis as described in any one of claims 1-9, characterized in that, The system includes: The robust potential energy reference surface generation module is suitable for generating a sequence of historical initial matching potential energy values ​​based on historical alarm image pairs using a pre-trained Siamese network. The module removes pseudo-stable fluctuations caused by periodic interference from the historical initial matching potential energy value sequence using a hierarchical Bayesian model and empirical mode decomposition, and constructs a robust potential energy reference surface that evolves over time through kernel density estimation and interquartile range. The high-dimensional heterogeneous feature tensor extraction module is suitable for performing multimodal feature extraction on the current alarm pair image to obtain three types of features corresponding to each alarm image: spatial distribution map, image texture descriptor, and segmentation region statistics. The three types of features are then integrated into a high-dimensional heterogeneous feature tensor of the alarm image. The current initial matching potential value generation module is suitable for inputting two high-dimensional heterogeneous feature tensors into the pre-trained Siamese network, and processing them independently through each branch of the Siamese network's two-branch symmetric feature extraction architecture to convert the two high-dimensional heterogeneous feature tensors into two one-dimensional feature vectors of fixed length; projecting the feature vectors onto the tangent space of the Riemannian manifold and fusing the geodesic distance and matching potential to generate a current initial matching potential value that is resistant to environmental interference; The potential repeated alarm cluster generation module is suitable for mapping the current initial matching potential energy value to the robust potential energy reference surface to obtain the residual, constructing a dynamic weighted coupling graph by combining the spatiotemporal context information of the current alarm pair, and performing iterative reasoning on the dynamic weighted coupling graph through the cyclic belief propagation algorithm to filter out potential repeated alarm clusters with high structural and semantic coordination. The high-contribution alarm cluster generation module is suitable for performing time-series alignment and multi-dimensional consistency evaluation on alarm instances within the potential duplicate alarm cluster, generating a consistency score, and treating nodes within the potential duplicate alarm cluster as game participants. By calculating their Shapley value for the consistency score and setting a dynamic threshold, abnormal interference nodes with a contribution lower than the dynamic threshold are eliminated to obtain the high-contribution alarm cluster. The duplicate alarm filtering module is suitable for performing topological data analysis on the high-contribution alarm cluster to verify its structural stability, identifying a stable defect evolution chain, selecting a main alarm from the defect evolution chain based on a preset main alarm selection index, and marking the remaining alarms as duplicate alarms and filtering them out.

Images