Substation equipment degradation state evaluation method and system based on multi-modal data fusion

Abstract

The present application relates to the technical field of electric power data processing, in particular to a substation equipment degradation state evaluation method and system based on multi-modal data fusion. A multi-modal prior database is constructed for target substation equipment; based on the multi-modal prior database, heterogeneous fusion features are analyzed and determined, and feature dynamic attention weights under differentiated maintenance cycles are generated to form multi-modal data fusion rules; a degradation state evaluation model based on ensemble learning is constructed, and the ensemble output weights of the degradation state evaluation model are initialized according to the multi-modal data fusion rules for iterative training; multi-modal maintenance data is collected in real time and input into the degradation state evaluation model to obtain a quantitative degradation trend, and the degradation grade is determined by integral calculation, and the output is the degradation state evaluation result. The present application can provide fine support for operation and maintenance decision-making, reduce operation and maintenance cost, and improve power grid operation stability.

Description

Technical Field

[0001] This invention relates to the field of power data processing technology, and in particular to a method and system for assessing the deterioration status of substation equipment using multimodal data fusion. Background Technology

[0002] Traditional substation equipment degradation assessments often rely on single data, fixed-cycle models, or single machine learning models, with features manually selected. They fail to effectively integrate multimodal data, do not adapt to the needs of differentiated maintenance cycles, and suffer from problems such as incomplete data, poor feature adaptation, insufficient assessment accuracy and robustness, and weak practicality of results, making it difficult to support refined operation and maintenance. Summary of the Invention

[0003] This invention addresses the problems of incomplete data, poor feature adaptation, insufficient evaluation accuracy and robustness, and weak practicality of results in existing technologies by providing a method and system for assessing the deterioration status of substation equipment through multimodal data fusion.

[0004] The technical solution of the present invention to solve the above-mentioned technical problems is as follows: In a first aspect, the present invention provides a method for assessing the deterioration status of substation equipment using multimodal data fusion, comprising: constructing a multimodal prior database for target substation equipment; analyzing and determining heterogeneous fusion features based on the multimodal prior database, generating dynamic attention weights for features under differentiated maintenance cycles, and forming multimodal data fusion rules; constructing a deterioration status assessment model based on ensemble learning, initializing the ensemble output weights of the deterioration status assessment model according to the multimodal data fusion rules, and performing iterative training; collecting multimodal maintenance data in real time and inputting it into the deterioration status assessment model to obtain quantitative deterioration trends, calculating the deterioration level by integration, and outputting the deterioration status assessment result.

[0005] Optionally, the multimodal prior database includes multimodal sensing data and operation and maintenance management data.

[0006] Optionally, based on the multimodal prior database, heterogeneous fusion features are analyzed and determined, including: traversing the multimodal prior database to extract prior features, the prior features including basic sensing features and prior fusion features; combining the basic sensing features and the prior fusion features with preset dimensional constraints, randomly combining the basic sensing features and the prior fusion features and performing dimensional deduplication to obtain multiple candidate fusion features, wherein the candidate fusion features contain differentiated configurations of feature dimensions; and based on principal component analysis, combined with the multimodal prior database, analyzing the degradation assessment contribution of the multiple candidate fusion features, and correspondingly selecting and determining multiple heterogeneous fusion features.

[0007] The process of selecting and determining multiple heterogeneous fusion features includes: sorting multiple candidate fusion features in descending order according to the degradation assessment contribution to obtain a sequence feature list; starting from the top of the sequence feature list, accumulating the number of basic sensing feature dimensions involved in multiple candidate fusion features to obtain an accumulated dimension; iteratively calculating the accumulated dimension, and if the accumulated dimension in the Nth iteration is greater than or equal to a preset accumulated dimension constraint, then determining the multiple candidate fusion features included in the N-1th iteration as multiple heterogeneous fusion features.

[0008] Optionally, generating feature dynamic attention weights under differentiated maintenance cycles to form multimodal data fusion rules includes: obtaining differentiated maintenance cycles for the target scenario, and extracting data from the multimodal prior database based on the differentiated maintenance cycles to generate multiple cycle prior databases; traversing the multiple cycle prior databases, using multiple heterogeneous fusion features as analysis targets, and evaluating and obtaining the degradation cycle contribution based on principal component analysis, wherein each degradation cycle contribution has a cycle label; normalizing the multiple degradation cycle contributions using the cycle label as the normalization range to obtain the feature dynamic attention weights for each heterogeneous fusion feature; and outputting the multiple feature dynamic attention weights in a structured manner based on the cycle label to form the multimodal data fusion rules.

[0009] Optionally, a degradation state assessment model based on ensemble learning is constructed, including: differentially configuring multiple meta-evaluation models and randomly copying the multiple meta-evaluation models to obtain a meta-evaluation model set; defining the ensemble selection probability of each heterogeneous fusion feature for the meta-evaluation model according to the feature dynamic attention weights of the multiple heterogeneous fusion features; driving the multiple meta-evaluation models to iteratively select from the multiple heterogeneous fusion features according to the ensemble selection probability until the meta-evaluation model has a corresponding heterogeneous fusion feature; configuring the ensemble output layer of the multiple meta-evaluation models as a weighted ensemble, and defaulting the ensemble output weights of the weighted ensemble to equal weights.

[0010] Optionally, the integrated output weights of the degradation state assessment model are initialized according to the multimodal data fusion rules, and iterative training is performed, including: determining the real-time maintenance cycle and the neighborhood maintenance cycle of the target substation equipment, wherein the neighborhood maintenance cycle includes at least one of a first neighborhood maintenance cycle with a cycle greater than the real-time maintenance cycle and a second neighborhood maintenance cycle with a cycle less than the real-time maintenance cycle; constructing a training sample dataset based on the multimodal prior database and the data paradigms of the real-time maintenance cycle and the neighborhood maintenance cycle; initializing the integrated output weights of the degradation state assessment model according to the real-time maintenance cycle and the neighborhood maintenance cycle according to the multimodal data fusion rules, respectively, to obtain a first integrated output layer and several second integrated output layers; performing iterative supervised learning on the degradation state assessment model based on the training sample dataset, and outputting through the first integrated output layer and several second integrated output layers after each iteration to obtain a heterogeneous output set; traversing the heterogeneous output set to verify the assessment accuracy, and iteratively supervising learning until each output in the heterogeneous output set meets the accuracy requirements of the target substation equipment.

[0011] Optionally, real-time acquisition of multimodal maintenance data is input into the degradation state assessment model to obtain quantitative degradation trends, and integral calculation is performed to determine the degradation level, outputting a degradation state assessment result. This includes: acquiring the multimodal maintenance data under the real-time maintenance cycle; standardizing the multimodal maintenance data according to a preset data paradigm for neighboring maintenance cycles; inputting the standardization result into the degradation state assessment model using the maintenance cycle as an index, wherein the degradation state assessment model includes multiple state assessment sub-models for differentiated maintenance cycles; acquiring the degradation state data under the real-time maintenance cycle... The first degradation trend and several second degradation trends under the neighboring maintenance cycle are considered; the residual distributions of several second degradation trends and the first degradation trend are calculated and integrated to obtain a verification parameter. If the verification parameter meets a preset confidence level, the first degradation trend is output as the quantitative degradation trend; the historical quantitative degradation trend and the quantitative degradation trend are combined to perform an integral operation to obtain the cumulative degradation amount and determine the degradation level by matching with a preset degradation grading rule; the degradation level and the quantitative degradation trend are combined and output as the degradation status assessment result.

[0012] The dimensional constraints include: a dimension quantity constraint, used to limit the maximum number of feature dimensions contained in the candidate fusion features; a dimension repetition constraint, used to limit the candidate fusion features from containing intuitively repetitive feature dimensions; and a dimension combination constraint, used to exclude specific feature dimension combinations that are determined to be invalid or inefficient in advance, wherein the dimension combination constraint is determined based on a preset prior knowledge base.

[0013] Secondly, this invention provides a substation equipment degradation status assessment system based on multimodal data fusion, comprising: The multimodal prior database acquisition module is used to construct a multimodal prior database for target substation equipment; The multimodal data fusion rule acquisition module is used to analyze and determine heterogeneous fusion features based on the multimodal prior database, and generate feature dynamic attention weights under differentiated maintenance cycles to form multimodal data fusion rules. The degradation state assessment model acquisition module is used to construct a degradation state assessment model based on ensemble learning, initialize the ensemble output weights of the degradation state assessment model according to the multimodal data fusion rules, and perform iterative training. The degradation status assessment result acquisition module is used to collect multimodal maintenance data in real time and input it into the degradation status assessment model to obtain quantitative degradation trends, calculate the degradation level by integration, and output the degradation status assessment result.

[0014] By implementing this invention, it is possible to construct a multimodal prior database for target substation equipment; integrate multimodal sensing and operation and maintenance data, break through the limitations of single data, provide a complete and standardized data foundation, and reserve sufficient prior samples for subsequent feature analysis and model training.

[0015] By implementing this invention, it is possible to analyze and determine heterogeneous fusion features based on the multimodal prior database, generate dynamic attention weights for features under differentiated maintenance cycles, and form multimodal data fusion rules; screen core heterogeneous fusion features to reduce redundancy, allocate dynamic weights according to differentiated maintenance cycles, establish unified fusion rules, and improve the targeting of feature applications and the consistency of evaluation.

[0016] By implementing this invention, it is possible to construct a degradation state assessment model based on ensemble learning, initialize the ensemble output weights of the degradation state assessment model according to the multimodal data fusion rules, and perform iterative training; the ensemble multivariate assessment model reduces bias, initializes weights according to fusion rules to improve efficiency, adapts to multiple maintenance cycles, and ensures the robustness and assessment accuracy of the model through iterative verification.

[0017] By implementing this invention, it is possible to collect multimodal maintenance data in real time, input it into the degradation status assessment model, obtain quantitative degradation trends, determine the degradation level through integral calculation, and output the degradation status assessment result. The real-time data is standardized, and the quantitative trends and degradation levels are output. The accuracy is improved through neighborhood periodic verification, providing a precise basis for refined operation and maintenance decisions.

[0018] In summary, by implementing this invention, highly accurate, adaptable, and reliable assessment of the deterioration status of substation equipment can be achieved, providing refined support for operation and maintenance decisions, reducing operation and maintenance costs, and improving the stability of power grid operation. Attached Figure Description

[0019] Figure 1 A flowchart illustrating the substation equipment degradation status assessment method based on multimodal data fusion provided by this invention; Figure 2 This is a schematic diagram of the substation equipment degradation status assessment system based on multimodal data fusion provided by the present invention.

[0020] In the attached diagram, the components represented by each number are as follows: Module 11 for acquiring multimodal prior database, module 12 for acquiring multimodal data fusion rules, module 13 for acquiring degradation state assessment model, and module 14 for acquiring degradation state assessment results. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] In the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0023] In the description of this invention, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this invention is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed herein.

[0024] Example 1, as Figure 1 As shown, this embodiment of the invention provides a method for assessing the deterioration status of substation equipment using multimodal data fusion, including: S100: Construct a multimodal prior database for target substation equipment; S200: Based on the multimodal prior database, analyze and determine heterogeneous fusion features, generate feature dynamic attention weights under differentiated maintenance cycles, and form multimodal data fusion rules; S300: Construct a degradation state assessment model based on ensemble learning, initialize the ensemble output weights of the degradation state assessment model according to the multimodal data fusion rules, and perform iterative training; S400: Real-time acquisition of multimodal maintenance data is input into the degradation status assessment model to obtain quantitative degradation trends, and the degradation level is determined by integral calculation. The output is the degradation status assessment result.

[0025] In step S100 of this application embodiment, it is first necessary to construct a multimodal prior database for the target substation equipment; wherein, the multimodal prior database includes multimodal sensing data and operation and maintenance management data.

[0026] In this embodiment, the objective of step S100 is to construct a multimodal prior database for the target substation equipment. The core purpose is to provide fundamental data support for subsequent heterogeneous fusion feature analysis, dynamic attention weight generation of features, and training of the degradation state assessment model. The multimodal sensing data and operation and maintenance management data together constitute the database, comprehensively covering equipment operating status-related information. This ensures sufficient and comprehensive data for subsequent analysis and model training, avoiding inaccurate evaluation results due to limited data.

[0027] Specifically, multimodal sensing data can include temperature data, humidity data, vibration frequency data, voltage data, current data, etc., during equipment operation. These data directly reflect the physical state of the equipment in real time.

[0028] Operation and maintenance management data can include equipment maintenance records, such as the scheduled maintenance time and problems found in March 2024; maintenance logs, such as the record of replacing a certain component in May 2024; and fault handling reports, such as the troubleshooting process and handling results of equipment short circuit faults in January 2024. These data record the operation and maintenance status of the equipment throughout its entire life cycle.

[0029] The collection of these two types of data involves, specifically, targeting the equipment in the substation by acquiring multimodal sensor data in real time through various deployed sensors. Simultaneously, it involves compiling past maintenance records, logs, and fault handling reports to obtain operation and maintenance management data. The collected data is then organized and structured for storage, establishing a unified data format and storage standards to create a multimodal prior database for the target substation equipment, facilitating subsequent data extraction and analysis.

[0030] In step S200 of this application embodiment, the heterogeneous fusion characteristics are analyzed and determined based on the multimodal prior database, including: Prior features are extracted by traversing the multimodal prior database, and the prior features include basic sensing features and prior fusion features; Based on preset dimensional constraints, the basic sensing features and the prior fusion features are randomly combined and dimensional deduplication is performed to obtain multiple candidate fusion features, wherein the candidate fusion features contain feature dimension differentiation configurations. Based on principal component analysis and combined with the multimodal prior database, the degradation assessment contribution of multiple candidate fusion features is analyzed, and multiple heterogeneous fusion features are selected accordingly.

[0031] In step S200 of this application embodiment, the dimensional constraint includes: Dimension number constraint, used to limit the maximum number of feature dimensions contained in the candidate fusion features; Dimension repetition constraint is used to limit the candidate fusion features from containing intuitively repetitive feature dimensions. Dimension combination constraints are used to exclude specific feature dimension combinations that are determined to be invalid or inefficient based on prior knowledge. The dimension combination constraints are determined based on a preset prior knowledge base.

[0032] In this embodiment, the purpose of step S200 is to select feature combinations that contribute significantly to the assessment of equipment degradation status from massive amounts of basic data. These heterogeneous fused features are dimensionally differentiated, non-redundant, and effective, accurately reflecting equipment degradation-related information. They provide core analytical targets for subsequent generation of feature dynamic attention weights and construction of degradation status assessment models, improving the relevance and accuracy of the subsequent assessment process.

[0033] To achieve the above steps, firstly, it is necessary to traverse the multimodal prior database to extract prior features, which include basic sensing features and prior fusion features.

[0034] This involves a comprehensive scan of the multimodal prior database, extracting the two types of prior features from multimodal sensing data and operation and maintenance management data to ensure that the features cover relevant information about the equipment's operating status and the entire operation and maintenance process.

[0035] Among them, the basic sensing features are the features corresponding to the raw data directly collected from the sensor. For example, temperature features correspond to the equipment winding temperature data collected by the sensor; voltage features correspond to the equipment operating voltage data collected by the sensor; and vibration features correspond to the equipment housing vibration amplitude data collected by the sensor.

[0036] Prior fusion features are fusion features that have already been applied. Fusion methods include computational fusion, which is specifically divided into two categories. One category is features corresponding to indirect sensing data obtained by calculation based on basic sensing features, such as temperature rise rate features calculated based on temperature features and heat dissipation efficiency features, and power features calculated based on voltage features and current features. The other category is features corresponding to existing combinations of basic sensing features, such as combined features of temperature features and vibration features, and combined features of voltage features and humidity features.

[0037] Next, based on preset dimensional constraints, the basic sensing features and the prior fusion features are randomly combined and deduplicated in terms of dimensions to obtain multiple candidate fusion features, wherein the candidate fusion features contain feature dimension differentiation configurations.

[0038] The dimensional constraints may include dimensional number constraints, dimensional repetition constraints, and dimensional combination constraints.

[0039] Specifically, the dimension number constraint is: limiting the maximum number of feature dimensions contained in the candidate fusion feature. For example, if the dimension number constraint is set to 5, then any candidate fusion feature cannot contain more than 5 feature dimensions.

[0040] The dimensional repetition constraint stipulates that candidate fusion features must not contain intuitively repetitive feature dimensions. For example, the temperature feature in the basic sensing features and the temperature-derived features calculated based on the temperature feature in the prior fusion features are not intuitively repetitive and can coexist; however, two completely identical temperature features cannot appear in the same candidate fusion feature at the same time.

[0041] The dimension combination constraint is to exclude specific feature dimension combinations that are determined to be invalid or inefficient in advance. This constraint is determined based on a preset prior knowledge base. For example, if the combination of voltage feature and equipment manufacturing date feature recorded in the prior knowledge base is meaningless for degradation assessment, it is an invalid combination and will be excluded during the combination process.

[0042] Then, the basic sensing features and prior fusion features are randomly combined. Due to the differentiated configuration of feature dimensions, the number and dimensions of features in different combinations can be different.

[0043] For example, combination A includes three dimensions: temperature characteristics, vibration characteristics, and power characteristics; combination B includes four dimensions: voltage characteristics, humidity characteristics, temperature rise rate characteristics, and vibration characteristics.

[0044] The combined features are checked for dimensional deduplication to eliminate combinations with intuitively repetitive dimensions, resulting in multiple candidate fusion features that meet the dimensional constraints.

[0045] Then, based on the principal component analysis method and combined with the multimodal prior database, it is necessary to analyze the degradation assessment contribution of multiple candidate fusion features, and correspondingly screen and determine multiple heterogeneous fusion features.

[0046] First, the contribution of the degradation assessment is calculated. This is done by using principal component analysis to correlate each candidate fusion feature with data in the multimodal prior database and to evaluate the degree of influence of each candidate fusion feature on the equipment degradation status assessment result. This degree of influence is the contribution of the degradation assessment.

[0047] For example, analysis revealed that the alternative fusion feature of temperature characteristics + vibration characteristics + power characteristics accounted for 35% of the variance in explaining the changes in equipment degradation status, and its contribution to degradation assessment was 35%; the alternative fusion feature of voltage characteristics + humidity characteristics accounted for 12% of the variance in explaining the changes in equipment degradation status, and its contribution to degradation assessment was 12%.

[0048] Next, it is necessary to screen and determine several of the heterogeneous fusion features, including: Based on the degradation assessment contribution, multiple candidate fusion features are sorted in descending order to obtain a list of sequence features; Starting from the top of the sequence feature list, the number of basic sensing feature dimensions involved in multiple candidate fusion features is accumulated and calculated to obtain the accumulated dimension. The cumulative dimension number is calculated iteratively. If the cumulative dimension number in the Nth iteration is greater than or equal to the preset cumulative dimension number constraint, then the multiple candidate fusion features included in the N-1th iteration are determined to be multiple heterogeneous fusion features.

[0049] In this embodiment, the purpose of the above steps is to select feature combinations from the analyzed candidate fusion features that have both a high contribution to degradation assessment and control over the total number of basic sensor feature dimensions. This ensures that the selected features can provide accurate support for subsequent degradation state assessment, while avoiding data redundancy and increased model computational complexity due to too many basic sensor feature dimensions, thus achieving a balance between assessment accuracy and computational efficiency.

[0050] First, based on the degradation assessment contribution, multiple candidate fusion features need to be sorted in descending order to obtain a list of sequence features.

[0051] First, we organize the degradation assessment contribution of all candidate fusion features, and then sort the candidate fusion features in descending order of degradation assessment contribution to form an ordered list of sequence features.

[0052] For example, suppose there are 5 alternative fusion features, each with a degradation assessment contribution of 35%, 30%, 20%, 10%, and 5%, respectively. The resulting sequence feature list after sorting in descending order is as follows: feature with a degradation assessment contribution of 35%, feature with a degradation assessment contribution of 30%, feature with a degradation assessment contribution of 20%, feature with a degradation assessment contribution of 10%, and feature with a degradation assessment contribution of 5%.

[0053] Next, starting from the top of the sequence feature list, the number of basic sensing feature dimensions involved in multiple candidate fusion features is accumulated to obtain the accumulated dimension.

[0054] Starting with the candidate fusion feature at the top of the sequence feature list, the number of basic sensing feature dimensions contained in each candidate fusion feature is counted sequentially and accumulated. Each time a feature is accumulated, the accumulated dimension is obtained.

[0055] The basic sensing feature dimension refers to the number of original features directly acquired from sensors in the candidate fusion features, excluding the non-direct sensing feature dimensions calculated from the prior fusion features.

[0056] For example, the preset cumulative dimension constraint is 8. In the sequence feature list, the first candidate fusion feature has a degradation evaluation contribution of 35%, including a basic sensing feature dimension of 3, so the cumulative dimension is 3; the second candidate fusion feature has a degradation evaluation contribution of 30%, including a basic sensing feature dimension of 2, so the cumulative dimension is updated to 3+2=5; the third candidate fusion feature has a degradation evaluation contribution of 20%, including a basic sensing feature dimension of 3, so the cumulative dimension is updated to 5+3=8; the fourth candidate fusion feature has a degradation evaluation contribution of 10%, including a basic sensing feature dimension of 2, so the cumulative dimension is updated to 8+2=10.

[0057] Then, the cumulative dimension number is calculated iteratively. If the cumulative dimension number in the Nth iteration is greater than or equal to the preset cumulative dimension number constraint, then the multiple candidate fusion features included in the N-1th iteration are determined to be multiple heterogeneous fusion features.

[0058] That is, the above accumulation process is repeated continuously. After each accumulation, it is determined whether the current accumulated dimension number is greater than or equal to the preset accumulated dimension number constraint. The iteration stops when the condition is met, and finally all candidate fusion features included in the N-1th iteration are determined as heterogeneous fusion features.

[0059] For example, continuing the previous example, the preset cumulative dimension constraint is 8. In the third iteration, the cumulative dimension reaches 8, equal to the preset constraint, and the iteration stops. The Nth iteration is the third iteration, and the N-1th iteration is the second iteration. The candidate fusion features included in the second iteration are features with a contribution of 35% and features with a contribution of 30%. These two features are the selected heterogeneous fusion features. If the cumulative dimension in the third iteration is 9, which is greater than the preset constraint of 8, the candidate fusion features included in the first two iterations are again taken as the heterogeneous fusion features.

[0060] In step S200 of this application embodiment, feature dynamic attention weights under differentiated maintenance cycles are generated to form multimodal data fusion rules, including: Obtain the differentiated maintenance cycle of the target scenario, and extract data from the multimodal prior database according to the differentiated maintenance cycle to generate multiple cycle prior databases; By traversing multiple prior databases of the aforementioned cycles and using multiple heterogeneous fusion features as the analysis targets, the contribution of the degradation cycle is evaluated and obtained based on the principal component analysis method, wherein each contribution of the degradation cycle has a cycle label. Using the period marker as the normalization range, the contribution of multiple degradation periods is normalized to obtain the feature dynamic attention weight of each heterogeneous fusion feature. Based on the periodic markers, multiple dynamic attention weights of the features are output in a structured manner to form the multimodal data fusion rules.

[0061] In step S200 of this application embodiment, the purpose of the above steps is to assign weights to heterogeneous fusion features under different maintenance cycles that match their actual assessment value. The equipment degradation patterns and influencing factors differ across maintenance cycles. By dynamically allocating weights, the more critical features for degradation assessment in each cycle can be highlighted, avoiding assessment bias caused by fixed weights. Simultaneously, structured fusion rules provide a clear basis for the weight initialization of the subsequent degradation state assessment model, ensuring that the model adapts to different maintenance scenarios.

[0062] To achieve the above objectives, firstly, it is necessary to obtain the differentiated maintenance cycle of the target scenario, and then extract data from the multimodal prior database based on the differentiated maintenance cycle to generate multiple periodic prior databases.

[0063] First, the differentiated maintenance cycles of the target substation equipment are defined, i.e., different maintenance time intervals. Then, according to the time range of each maintenance cycle, all data within the corresponding time period are extracted from the multimodal prior database, and an independent cycle prior database is generated for each cycle.

[0064] For example, the differentiated maintenance cycle for the target scenario is set as 30 days per month, 90 days per quarter, and 365 days per year. Multimodal sensing data and operation and maintenance management data for each month over the past two years are extracted from the multimodal prior database to generate a 24-monthly prior database; relevant data for each quarter are extracted to generate an 8-quarterly prior database; and relevant data for each year are extracted to generate an 2-yearly prior database.

[0065] Next, the multiple cycle prior databases are traversed, and the multiple heterogeneous fusion features are used as analysis targets. The degradation cycle contribution is evaluated and obtained based on the principal component analysis method, wherein each degradation cycle contribution has a cycle label.

[0066] That is, the prior database of each cycle is traversed one by one, and multiple heterogeneous fusion features that have been selected and determined in the previous period are used as the analysis objects. The principal component analysis method is used to evaluate the contribution of each heterogeneous fusion feature to the equipment degradation status assessment in the cycle, and the degradation cycle contribution is obtained. At the same time, a corresponding cycle mark is added to each degradation cycle contribution to clarify the maintenance cycle to which it belongs.

[0067] For example, two heterogeneous fused features, temperature + vibration + power and voltage + humidity, are selected. Traversing the monthly prior database, principal component analysis reveals that the temperature + vibration + power feature contributes 40% to the degradation cycle within that monthly cycle, thus adding a monthly cycle label; the voltage + humidity feature contributes 25%, also adding a monthly cycle label. Traversing the quarterly prior database, the temperature + vibration + power feature contributes 35% to the degradation cycle, thus adding a quarterly cycle label; the voltage + humidity feature contributes 30%, also adding a quarterly cycle label.

[0068] Then, using the period marker as the normalization range, the contribution of multiple degradation periods is normalized to obtain the feature dynamic attention weight of each heterogeneous fusion feature.

[0069] That is, the features are grouped according to the period label. All the degradation period contributions corresponding to the same period label are grouped together. Each group is used as the normalization range. The degradation period contributions within the group are normalized so that the sum of the weights of all heterogeneous fusion features in each group is 1. The processed value is the feature dynamic attention weight of the corresponding heterogeneous fusion feature under that period.

[0070] For example, the degradation cycle contribution group corresponding to the monthly cycle marker is 40% 35% 25%. During normalization, the total is first calculated as 40% + 35% + 25% = 100%. Then, each contribution is divided by the total, resulting in a dynamic attention weight of 40% ÷ 100% = 0.4 for the temperature + vibration + power feature, a weight of 35% ÷ 100% = 0.35 for the other feature, and a weight of 25% ÷ 100% = 0.25 for the voltage + humidity feature. The degradation cycle contribution group corresponding to the quarterly cycle marker is 35% 30% 20%, with a total of 85%. After normalization, the weights of each feature are approximately 35% ÷ 85% ≈ 0.41, 30% ÷ 85% ≈ 0.35, and 20% ÷ 85% ≈ 0.24.

[0071] Based on the periodic markers, multiple dynamic attention weights of the features are output in a structured manner to form the multimodal data fusion rules.

[0072] The heterogeneous fusion features and their dynamic attention weights corresponding to each cycle mark are structured and output in a unified format to form clear multimodal data fusion rules. The rules clearly correspond to the weight allocation of each heterogeneous fusion feature under each maintenance cycle.

[0073] For example, the structured output format is: Periodicity label: Heterogeneous fusion feature - Feature dynamic attention weight. The final multimodal data fusion rules include: Monthly: Temperature + Vibration + Power -0.4, Feature A -0.35, Voltage + Humidity -0.25; Quarterly: Temperature + Vibration + Power -0.41, Feature A -0.35, Voltage + Humidity -0.24; Annually: Feature X -0.5, Feature Y -0.3, Feature Z -0.2, etc.

[0074] In step S300 of this application embodiment, constructing a degradation state assessment model based on ensemble learning includes: Multiple meta-evaluation models are configured differently, and the multiple meta-evaluation models are randomly copied to obtain a meta-evaluation model set; Based on the feature dynamic attention weights of multiple heterogeneous fusion features, the integration selection probability of the meta-evaluation model for each heterogeneous fusion feature is defined; Based on the integration selection probability, multiple meta-evaluation models are driven to iteratively select multiple heterogeneous fusion features until the meta-evaluation model has a corresponding heterogeneous fusion feature. The integrated output layer of multiple meta-evaluation models is configured as a weighted integration, and the integrated output weights of the weighted integration are defaulted to equal weights.

[0075] In this embodiment, the purpose of step S300 is to improve the stability and generalization ability of equipment degradation state assessment through the synergistic effect of multiple meta-assessment models. Differentiated configuration of meta-assessment models and random replication to form a meta-assessment model set enriches model diversity; combining feature dynamic attention weights to define the integration selection probability allows meta-assessment models to prioritize high-value heterogeneous fusion features; the iterative selection mechanism ensures that each meta-assessment model has suitable features; the configuration of the weighted integration output layer integrates the advantages of each meta-assessment model, and the default setting of equal weights provides a basis for subsequent weight adjustments based on fusion rules, ultimately achieving a more accurate and reliable degradation state assessment.

[0076] To achieve the above objectives, firstly, it is necessary to differentiate and configure multiple meta-evaluation models, and then randomly replicate these multiple meta-evaluation models to obtain a meta-evaluation model set.

[0077] This involves selecting multiple different types of models as meta-evaluation models, achieving differentiated configurations by adjusting the core parameters of the models, and then randomly replicating each differentiated meta-evaluation model to generate a meta-evaluation model set containing multiple independent copies of the same model structure.

[0078] For example, support vector machines (SVMs), decision trees, and neural networks were selected as the meta-evaluation models. The differentiated configurations were as follows: the kernel function of the SVM was set to a linear kernel with a penalty coefficient of 1.0; the maximum depth of the decision tree was set to 5, and the minimum number of sample splits was set to 10; the number of hidden layer nodes in the neural network was set to 64, and the activation function was ReLU. Each meta-evaluation model with these configurations was randomly replicated 5 times, resulting in a meta-evaluation model set containing 15 models in total, with 5 copies of each of the three types.

[0079] Next, based on the feature dynamic attention weights of the multiple heterogeneous fusion features, it is necessary to define the integration selection probability of the meta-evaluation model for each of the heterogeneous fusion features.

[0080] That is, extract the feature dynamic attention weight of each heterogeneous fusion feature under the corresponding maintenance cycle, and define the proportion of the feature dynamic attention weight of each heterogeneous fusion feature to the sum of the feature dynamic attention weights of all heterogeneous fusion features as the integration selection probability of the meta-evaluation model for that heterogeneous fusion feature. The higher the feature dynamic attention weight, the greater the probability of the feature being selected by the model.

[0081] For example, suppose there are three heterogeneous fusion features: feature a: temperature + vibration + power, with a dynamic attention weight of 0.4; feature b: voltage + humidity, with a dynamic attention weight of 0.35; and feature c: pressure + insulation resistance, with a dynamic attention weight of 0.25. The sum of all feature dynamic attention weights is 0.4 + 0.35 + 0.25 = 1.0. Then, the probability of the meta-evaluation model selecting feature a is 0.4 ÷ 1.0 = 0.4, the probability of selecting feature b is 0.35 ÷ 1.0 = 0.35, and the probability of selecting feature c is 0.25 ÷ 1.0 = 0.25.

[0082] Then, based on the integration selection probability, multiple meta-evaluation models need to be driven to iteratively select multiple heterogeneous fusion features until the meta-evaluation model has a corresponding heterogeneous fusion feature.

[0083] Based on the integration selection probability, each model in the meta-evaluation model set selects all heterogeneous fusion features in turn. Each time a feature with a higher selection probability is selected, the feature is more likely to be selected. This process is iterated until each meta-evaluation model is assigned at least one heterogeneous fusion feature.

[0084] For example, the meta-evaluation model set contains copies of Model A (Support Vector Machine), Model B (Decision Tree), and Model C (Neural Network). Feature selection is performed with probabilities of 0.4 for feature a, 0.35 for feature b, and 0.25 for feature c. In its first iteration, Model A randomly selects feature a, completing the assignment. Model B does not select any feature in its first iteration, but selects feature b in its second iteration. Model C selects feature c in its first iteration. Ultimately, the three models correspond to features a, b, and c respectively, and the iteration stops. If a model fails to select a feature in multiple iterations, the selection probability thresholds can be adjusted appropriately, or the number of iterations can be increased to ensure that all models have a corresponding feature.

[0085] Finally, it is necessary to configure the integrated output layer of multiple meta-evaluation models as a weighted integration, and to define the integrated output weights of the weighted integration as equal weights by default.

[0086] This involves setting a unified integration output layer for all models in the meta-evaluation model set, specifying that the fusion method of the output layer is weighted integration, meaning that the output results of each model will be weighted and summed according to their corresponding weights to obtain the final evaluation result. Before initialization based on multimodal data fusion rules, the integration output weights of all meta-evaluation models are uniformly set to equal weights to ensure that the influence of each model's output is consistent in the initial state.

[0087] For example, the meta-evaluation model set contains 15 models. After configuring the ensemble output layer as weighted ensemble, the default ensemble output weight for each model is 1 ÷ 15 ≈ 0.067. Assuming the outputs of the 15 models are Y1, Y2...Y15, then the final output of the meta-evaluation model in the initial state is 0.067Y1 + 0.067Y2 + ... + 0.067Y15.

[0088] In step S300 of this embodiment, the integrated output weights of the degradation state assessment model are initialized according to the multimodal data fusion rules, and iterative training is performed, including: Determine the real-time maintenance cycle and the neighborhood maintenance cycle of the target substation equipment, wherein the neighborhood maintenance cycle includes at least one of a first neighborhood maintenance cycle with a cycle greater than the real-time maintenance cycle and a second neighborhood maintenance cycle with a cycle less than the real-time maintenance cycle. Based on the multimodal prior database, a training sample dataset is constructed by combining the data paradigms of the real-time maintenance cycle and the neighborhood maintenance cycle. Based on the multimodal data fusion rules, the integrated output weights of the degradation state assessment model are initialized according to the real-time maintenance cycle and the neighborhood maintenance cycle, respectively, to obtain a first integrated output layer and several second integrated output layers; Based on the training sample dataset, the degradation state assessment model is subjected to iterative supervised learning, and after each iteration, the output is obtained by passing the first integrated output layer and several second integrated output layers respectively to obtain a heterogeneous output set. The heterogeneous output set is traversed to evaluate accuracy verification, and supervised learning is performed iteratively until each output in the heterogeneous output set meets the accuracy requirements of the target substation equipment.

[0089] In this embodiment of the application, the purpose of the above steps in step S300 is to initialize the integrated output weights of the degradation state assessment model according to the multimodal data fusion rules and perform iterative training. The core purpose is to make the model adapt to the assessment scenarios of real-time maintenance cycle and adjacent maintenance cycle. Through iterative training driven by multi-cycle data, the model weight parameters are optimized to ensure that the model can output accurate degradation assessment results in different cycle-related scenarios.

[0090] To achieve the above objectives, firstly, it is necessary to determine the real-time maintenance cycle and the neighborhood maintenance cycle of the target substation equipment. The neighborhood maintenance cycle includes at least one of a first neighborhood maintenance cycle with a cycle longer than the real-time maintenance cycle and a second neighborhood maintenance cycle with a cycle shorter than the real-time maintenance cycle.

[0091] First, the real-time maintenance cycle for assessing the deterioration of the target equipment is determined. Then, maintenance cycles adjacent to this cycle are selected as neighboring maintenance cycles. Neighboring maintenance cycles may include one or both of the following: a first neighboring maintenance cycle with a longer cycle and a second neighboring maintenance cycle with a shorter cycle.

[0092] For example, if the real-time maintenance cycle of the target substation equipment is quarterly (90 days), then the first neighboring maintenance cycle, which has a cycle longer than the real-time maintenance cycle, can be set to annual (365 days); the second neighboring maintenance cycle, which has a cycle shorter than the real-time maintenance cycle, can be set to monthly (30 days). The final determined cycle combination is the quarterly real-time maintenance cycle, and the annual and monthly neighboring maintenance cycles.

[0093] Next, based on the multimodal prior database, a training sample dataset is constructed by combining the data paradigms of the real-time maintenance cycle and the neighborhood maintenance cycle.

[0094] First, the data paradigms for the real-time maintenance cycle and the neighborhood maintenance cycle are clearly defined. The data paradigms include data acquisition frequency, data type selection criteria, data format, etc. Then, multimodal sensor data and operation and maintenance management data that conform to the data paradigms of each cycle are extracted from the multimodal prior database, and organized according to the format requirements of the training samples to form a training sample dataset.

[0095] For example, the data paradigm for the quarterly real-time maintenance cycle is to collect data every 7 days, including temperature, vibration, power, and maintenance records; the data paradigm for the annual maintenance cycle in the first neighborhood is to collect data every 30 days, including temperature, voltage, insulation resistance, and fault handling reports; and the data paradigm for the monthly maintenance cycle in the second neighborhood is to collect data every 3 days, including vibration, current, humidity, and maintenance logs. Data conforming to the above paradigms from the past 5 years were extracted from the multimodal prior database. Each data point includes feature values ​​and the corresponding actual equipment degradation level label, resulting in 1000 samples to form the training sample dataset.

[0096] Then, based on the multimodal data fusion rules, the integrated output weights of the degradation state assessment model are initialized according to the real-time maintenance cycle and the neighborhood maintenance cycle, respectively, to obtain the first integrated output layer and several second integrated output layers.

[0097] That is, from the multimodal data fusion rules, the dynamic attention weights of the heterogeneous fusion features corresponding to the real-time maintenance cycle and the maintenance cycles of each neighborhood are extracted respectively. These weights are used as the initial weights of the integrated output layer of the degradation state assessment model, and a first integrated output layer adapted to the real-time maintenance cycle and several second integrated output layers adapted to each neighborhood maintenance cycle are generated.

[0098] For example, in the multimodal data fusion rules, the weights corresponding to the quarterly cycle are: feature a (temperature + vibration + power) 0.4, feature b (voltage + humidity) 0.35, and feature c (pressure + insulation resistance) 0.25; the weights corresponding to the annual cycle are: feature a 0.3, feature b 0.4, and feature c 0.3; and the weights corresponding to the monthly cycle are: feature a 0.45, feature b 0.3, and feature c 0.25. These weights are used to initialize the model integration output layers, resulting in a first integration output layer adapted to the quarter with initial weights of 0.4, 0.35, and 0.25. Two second integration output layers are adapted to the annual and monthly cycles, respectively, with initial weights of 0.3, 0.4, and 0.3, and 0.45, 0.3, and 0.25, respectively.

[0099] Furthermore, the degradation state assessment model is subjected to iterative supervised learning based on the training sample dataset, and after each iteration, the output is obtained by passing the first integrated output layer and several second integrated output layers respectively to obtain a heterogeneous output set.

[0100] The training sample dataset is proportionally divided into a training set and a validation set. The degradation state assessment model is trained iteratively using the training set as input. After each iteration, the evaluation results are output through the first ensemble output layer and all second ensemble output layers, respectively. These evaluation results from different output layers are then integrated to form a heterogeneous output set.

[0101] For example, 1000 training samples are divided into an 8:2 ratio: 800 training samples and 200 validation samples. In the first iteration, after training the model with the 800 training samples, the first ensemble output layer outputs the quarterly evaluation results of the 200 validation samples. Two second ensemble output layers output the annual and monthly evaluation results, respectively, which are then integrated into a heterogeneous output set containing 600 evaluation results. This process is repeated in each subsequent iteration, updating the heterogeneous output set.

[0102] Finally, the heterogeneous output set is traversed to evaluate accuracy verification, and supervised learning is performed iteratively until each output in the heterogeneous output set meets the accuracy requirements of the target substation equipment.

[0103] This involves verifying the actual degradation level of each evaluation result in the heterogeneous output set against the corresponding sample, and calculating indicators such as evaluation accuracy and error value to validate accuracy. If any output result does not meet the accuracy requirements, such as an accuracy rate lower than a preset threshold, the model is iteratively trained using the training set, and the weight parameters are updated; if all output results meet the preset threshold, training is stopped.

[0104] For example, the preset accuracy requirement is an evaluation result accuracy of no less than 90% and an error value of no more than 5%. After the first iteration, the accuracy of the quarterly cycle evaluation results in the heterogeneous output set was 82%, the annual cycle accuracy was 85%, and the monthly cycle accuracy was 83%, all of which did not meet the requirements, so iterative training continued. After 20 iterations, it was verified again that the quarterly cycle evaluation accuracy was 92%, the annual cycle accuracy was 91%, and the monthly cycle accuracy was 93%, all output results met the accuracy requirements, iterative training stopped, and the deterioration state evaluation model training was completed.

[0105] In step S400 of this embodiment, multimodal maintenance data is collected in real time and input into the degradation state assessment model to obtain a quantitative degradation trend. The degradation level is then determined through integral calculation, and the output is the degradation state assessment result, including: Acquire the multimodal maintenance data under the real-time maintenance cycle; The multimodal maintenance data is standardized according to the preset neighborhood maintenance cycle data paradigm. Using the maintenance cycle as an index, the standardized processing results are input into the degradation state assessment model, wherein the degradation state assessment model includes multiple state assessment sub-models with differentiated maintenance cycles; Obtain the first degradation trend under the real-time maintenance cycle and several second degradation trends under the neighboring maintenance cycle; Calculate and integrate the residual distributions of several second degradation trends and the first degradation trend respectively to obtain the verification parameter. If the verification parameter satisfies the preset confidence level verification value, then output the first degradation trend as the quantitative degradation trend. The cumulative degradation amount is obtained by integrating the historical quantitative degradation trend with the quantitative degradation trend, and the degradation level is determined by matching it with a preset degradation grading rule. The degradation level and the quantitative degradation trend are combined and output as the degradation status assessment result.

[0106] In this embodiment, the purpose of step S400 is to accurately obtain the current quantitative degradation trend and cumulative degradation level of the equipment based on the trained model, combined with real-time data and neighborhood periodic data verification. Standardization ensures data compatibility with model input requirements, residual integral verification guarantees the reliability of the quantitative degradation trend, and integral calculations combining historical data comprehensively reflect the cumulative degree of equipment degradation. This ensures the evaluation results reflect both real-time status and long-term variation patterns, providing an accurate basis for equipment maintenance decisions.

[0107] To achieve the above objectives, it is first necessary to obtain the multimodal maintenance data under the real-time maintenance cycle.

[0108] This involves clearly defining the current real-time maintenance cycle for the target substation equipment, and using various sensors deployed on the equipment to collect multimodal sensor data related to the equipment's operating status in real time during this cycle. Simultaneously, operation and maintenance management data during this cycle are also collected and integrated to form multimodal maintenance data.

[0109] For example, the real-time maintenance cycle is 90 days per quarter. Through temperature sensors, vibration sensors, voltage sensors, etc., multimodal sensor data such as temperature data, vibration amplitude data, and voltage data of the equipment are collected in real time during these 90 days. At the same time, maintenance management data such as equipment maintenance records and component replacement logs within the quarter are collected, which together constitute multimodal maintenance data.

[0110] Next, the multimodal maintenance data needs to be standardized according to the preset neighborhood maintenance cycle data paradigm.

[0111] First, the data paradigm of the preset neighborhood maintenance cycle is retrieved. The data paradigm includes standards such as data acquisition frequency, data format, and numerical range. Then, the real-time collected multimodal maintenance data is uniformly adjusted according to this paradigm to ensure that the data format and dimensions are consistent with the neighborhood cycle data and meet the multi-cycle data input requirements of the model.

[0112] For example, if the maintenance cycle for a neighborhood is 30 days per month and 365 days per year, the data paradigm specifies a data collection frequency of once per day, with values ​​retained to two decimal places, and temperature data uniformly converted to degrees Celsius. In the real-time collected multimodal maintenance data, temperature data is collected every 3 hours, and voltage data is retained to one decimal place. During standardization, the temperature data is adjusted to a daily collection frequency by averaging the data daily, and the voltage data is retained to two decimal places, ultimately yielding standardized data that conforms to the paradigm.

[0113] Then, the standardized processing results need to be input into the degradation state assessment model, which includes multiple state assessment sub-models with differentiated maintenance cycles, using the maintenance cycle as an index.

[0114] The degradation condition assessment model comprises multiple condition assessment sub-models adapted to different maintenance cycles, with each sub-model corresponding to a maintenance cycle index. Standardized multimodal maintenance data is input into the corresponding condition assessment sub-models according to the indices of real-time maintenance cycle and neighborhood maintenance cycle.

[0115] For example, the degradation status assessment model includes monthly, quarterly, and annual sub-models. Standardized multimodal maintenance data is input into the quarterly sub-model using quarterly indexes, the monthly sub-model using monthly indexes, and the annual sub-model using annual indexes, ensuring that the data is compatible with the sub-model cycles.

[0116] Furthermore, it is necessary to obtain the first degradation trend under the real-time maintenance cycle and several second degradation trends under the neighborhood maintenance cycle.

[0117] Each condition assessment sub-model processes the standardized input data and outputs a trend curve or numerical sequence of equipment degradation status over time for the corresponding period. Specifically, the sub-model output for the real-time maintenance period represents the first degradation trend, while the sub-model output for each neighboring maintenance period represents the second degradation trend.

[0118] For example, the quarterly sub-model outputs the first degradation trend, which is a sequence in which the equipment degradation value gradually increases from the initial 1.2 to 2.8 within the quarter; the monthly sub-model outputs the second degradation trend, which is a sequence in which the monthly degradation values ​​within the three months included in the quarter are 1.1, 1.9, and 2.8 respectively; the annual sub-model outputs another second degradation trend, which is a curve showing the change in the equipment degradation value within the year including the quarter.

[0119] Furthermore, it is necessary to calculate and integrate the residual distributions of several second degradation trends and the first degradation trend to obtain the verification parameters. If the verification parameters meet the preset confidence verification value, the first degradation trend is output as the quantitative degradation trend.

[0120] The residual distribution refers to the distribution formed by the difference between the corresponding values ​​of the second degradation trend and the first degradation trend at each time point. The residual distribution of each second degradation trend and the first degradation trend is integrated, and the result is the verification parameter. The verification parameter is compared with a preset confidence level. If the verification parameter is less than or equal to the confidence level, the first degradation trend is considered reliable and is thus identified as the quantitative degradation trend.

[0121] For example, the preset confidence level is 0.5. The residual distributions of the first second degradation trend and the first degradation trend are 0.1, 0.2, -0.1, and 0.3. Integrating this distribution yields the validation parameter 0.3. The residual distributions of the second second degradation trend and the first degradation trend are 0.2, -0.1, 0.1, and 0.2. Integrating this distribution yields the validation parameter 0.4. Both validation parameters are less than 0.5, satisfying the confidence level requirement, and the first degradation trend is output as a quantitative degradation trend.

[0122] Then, it is necessary to perform an integral calculation by combining the historical quantitative degradation trend with the quantitative degradation trend to obtain the cumulative degradation amount, and determine the degradation level by matching it with the preset degradation grading rules.

[0123] This involves retrieving historical quantitative degradation trend data of the equipment, integrating the historical data with the currently determined quantitative degradation trend in chronological order, and performing an integral calculation on the integrated complete trend data. The integral result is the cumulative degradation amount of the equipment since it was put into use. A preset degradation grading rule is established, dividing the equipment into different grades based on the numerical range of the cumulative degradation amount. The calculated cumulative degradation amount is then matched with the rule to determine the corresponding degradation grade.

[0124] For example, the integral result corresponding to the historical quantitative degradation trend is 5.6, and the integral result of the current quantitative degradation trend is 2.8. After integration, the cumulative degradation amount is 8.4. The preset degradation classification rules are: a cumulative degradation amount of 0-3 is Level 1 slight degradation, 3-7 is Level 2 moderate degradation, 7-10 is Level 3 severe degradation, and above 10 is Level 4 high-risk degradation. The cumulative degradation amount of 8.4 falls within the range of 7-10, and the degradation level is determined to be Level 3 severe degradation.

[0125] Finally, the degradation level and the quantitative degradation trend are combined and output as the degradation status assessment result.

[0126] The determined degradation levels and quantitative degradation trends are integrated and organized according to a unified output format to form a complete degradation status assessment result, which is provided to equipment operation and maintenance personnel as a basis for maintenance decisions.

[0127] For example, the merged degradation status assessment result shows that the quantitative degradation trend is that the equipment degradation value increased from 1.2 to 2.8 within the quarter, and the degradation level is Level 3 severe degradation. The output clearly presents the current degradation trend and cumulative degradation degree of the equipment, making it easier for maintenance personnel to determine whether immediate maintenance needs to be arranged.

[0128] Example 2, as Figure 2 As shown, based on the same inventive concept as the substation equipment degradation status assessment method using multimodal data fusion provided in Embodiment 1, this embodiment of the invention also provides a substation equipment degradation status assessment system using multimodal data fusion, comprising: The multimodal prior database acquisition module 11 is used to construct a multimodal prior database for target substation equipment; The multimodal data fusion rule acquisition module 12 is used to analyze and determine heterogeneous fusion features based on the multimodal prior database, and generate feature dynamic attention weights under differentiated maintenance cycles to form multimodal data fusion rules. The degradation state assessment model acquisition module 13 is used to construct a degradation state assessment model based on ensemble learning, initialize the ensemble output weights of the degradation state assessment model according to the multimodal data fusion rules, and perform iterative training. The degradation status assessment result acquisition module 14 is used to collect multimodal maintenance data in real time and input it into the degradation status assessment model to obtain quantitative degradation trends, calculate the degradation level by integration, and output the degradation status assessment result.

[0129] Furthermore, the multimodal data fusion rule acquisition module 12 includes the following execution steps: Prior features are extracted by traversing the multimodal prior database, and the prior features include basic sensing features and prior fusion features; Based on preset dimensional constraints, the basic sensing features and the prior fusion features are randomly combined and dimensional deduplication is performed to obtain multiple candidate fusion features, wherein the candidate fusion features contain feature dimension differentiation configurations. Based on principal component analysis and combined with the multimodal prior database, the degradation assessment contribution of multiple candidate fusion features is analyzed, and multiple heterogeneous fusion features are selected accordingly.

[0130] The dimensional constraints include: Dimension number constraint, used to limit the maximum number of feature dimensions contained in the candidate fusion features; Dimension repetition constraint is used to limit the candidate fusion features from containing intuitively repetitive feature dimensions. Dimension combination constraints are used to exclude specific feature dimension combinations that are determined to be invalid or inefficient based on prior knowledge. The dimension combination constraints are determined based on a preset prior knowledge base.

[0131] The process of selecting and determining multiple heterogeneous fusion features includes: Based on the degradation assessment contribution, multiple candidate fusion features are sorted in descending order to obtain a list of sequence features; Starting from the top of the sequence feature list, the number of basic sensing feature dimensions involved in multiple candidate fusion features is accumulated and calculated to obtain the accumulated dimension. The cumulative dimension number is calculated iteratively. If the cumulative dimension number in the Nth iteration is greater than or equal to the preset cumulative dimension number constraint, then the multiple candidate fusion features included in the N-1th iteration are determined to be multiple heterogeneous fusion features.

[0132] Obtain the differentiated maintenance cycle of the target scenario, and extract data from the multimodal prior database according to the differentiated maintenance cycle to generate multiple cycle prior databases; By traversing multiple prior databases of the aforementioned cycles and using multiple heterogeneous fusion features as the analysis targets, the contribution of the degradation cycle is evaluated and obtained based on the principal component analysis method, wherein each contribution of the degradation cycle has a cycle label. Using the period marker as the normalization range, the contribution of multiple degradation periods is normalized to obtain the feature dynamic attention weight of each heterogeneous fusion feature. Based on the periodic markers, multiple dynamic attention weights of the features are output in a structured manner to form the multimodal data fusion rules.

[0133] Furthermore, the degradation state assessment model acquisition module 13 includes the following execution steps: Multiple meta-evaluation models are configured differently, and the multiple meta-evaluation models are randomly copied to obtain a meta-evaluation model set; Based on the feature dynamic attention weights of multiple heterogeneous fusion features, the integration selection probability of the meta-evaluation model for each heterogeneous fusion feature is defined; Based on the integration selection probability, multiple meta-evaluation models are driven to iteratively select multiple heterogeneous fusion features until the meta-evaluation model has a corresponding heterogeneous fusion feature. The integrated output layer of multiple meta-evaluation models is configured as a weighted integration, and the integrated output weights of the weighted integration are defaulted to equal weights.

[0134] Determine the real-time maintenance cycle and the neighborhood maintenance cycle of the target substation equipment, wherein the neighborhood maintenance cycle includes at least one of a first neighborhood maintenance cycle with a cycle greater than the real-time maintenance cycle and a second neighborhood maintenance cycle with a cycle less than the real-time maintenance cycle. Based on the multimodal prior database, a training sample dataset is constructed by combining the data paradigms of the real-time maintenance cycle and the neighborhood maintenance cycle. Based on the multimodal data fusion rules, the integrated output weights of the degradation state assessment model are initialized according to the real-time maintenance cycle and the neighborhood maintenance cycle, respectively, to obtain a first integrated output layer and several second integrated output layers; Based on the training sample dataset, the degradation state assessment model is subjected to iterative supervised learning, and after each iteration, the output is obtained by passing the first integrated output layer and several second integrated output layers respectively to obtain a heterogeneous output set. The heterogeneous output set is traversed to evaluate accuracy verification, and supervised learning is performed iteratively until each output in the heterogeneous output set meets the accuracy requirements of the target substation equipment.

[0135] Furthermore, the degradation status assessment result acquisition module 14 includes the following execution steps: Acquire the multimodal maintenance data under the real-time maintenance cycle; The multimodal maintenance data is standardized according to the preset neighborhood maintenance cycle data paradigm. Using the maintenance cycle as an index, the standardized processing results are input into the degradation state assessment model, wherein the degradation state assessment model includes multiple state assessment sub-models with differentiated maintenance cycles; Obtain the first degradation trend under the real-time maintenance cycle and several second degradation trends under the neighboring maintenance cycle; Calculate and integrate the residual distributions of several second degradation trends and the first degradation trend respectively to obtain the verification parameter. If the verification parameter satisfies the preset confidence level verification value, then output the first degradation trend as the quantitative degradation trend. The cumulative degradation amount is obtained by integrating the historical quantitative degradation trend with the quantitative degradation trend, and the degradation level is determined by matching it with a preset degradation grading rule. The degradation level and the quantitative degradation trend are combined and output as the degradation status assessment result.

[0136] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0137] Those skilled in the art will understand that embodiments of the present invention can provide methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0138] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0139] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0140] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0141] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Clearly, those skilled in the art can make various alterations and variations to the invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the invention and its equivalents, the invention is also intended to include these modifications and variations.

Claims

1. A method for assessing the deterioration status of substation equipment using multimodal data fusion, characterized in that, include: Construct a multimodal prior database for target substation equipment; Based on the multimodal prior database, heterogeneous fusion features are analyzed and determined, and dynamic attention weights for features under differentiated maintenance cycles are generated to form multimodal data fusion rules. A degradation state assessment model based on ensemble learning is constructed, and the ensemble output weights of the degradation state assessment model are initialized according to the multimodal data fusion rules, and iterative training is performed. Real-time multimodal maintenance data is collected and input into the degradation status assessment model to obtain quantitative degradation trends, and the degradation level is determined by integral calculation, and the degradation status assessment result is output. Based on the aforementioned multimodal prior database, heterogeneous fusion features are analyzed and determined, including: Prior features are extracted by traversing the multimodal prior database, and the prior features include basic sensing features and prior fusion features; Based on preset dimensional constraints, the basic sensing features and the prior fusion features are randomly combined and dimensional deduplication is performed to obtain multiple candidate fusion features, wherein the candidate fusion features contain feature dimension differentiation configurations. Based on the principal component analysis method and combined with the multimodal prior database, the degradation assessment contribution of multiple candidate fusion features is analyzed, and multiple heterogeneous fusion features are selected and determined accordingly. The heterogeneous fusion features were selected and identified through screening, including: Based on the degradation assessment contribution, multiple candidate fusion features are sorted in descending order to obtain a list of sequence features; Starting from the top of the sequence feature list, the number of basic sensing feature dimensions involved in multiple candidate fusion features is accumulated and calculated to obtain the accumulated dimension. The cumulative dimension number is calculated iteratively. If the cumulative dimension number in the Nth iteration is greater than or equal to the preset cumulative dimension number constraint, then the multiple candidate fusion features included in the N-1th iteration are determined to be multiple heterogeneous fusion features. Generate dynamic attention weights for features under differentiated maintenance cycles to form multimodal data fusion rules, including: Obtain the differentiated maintenance cycle of the target scenario, and extract data from the multimodal prior database according to the differentiated maintenance cycle to generate multiple cycle prior databases; By traversing multiple prior databases of the aforementioned cycles and using multiple heterogeneous fusion features as the analysis targets, the contribution of the degradation cycle is evaluated and obtained based on the principal component analysis method, wherein each contribution of the degradation cycle has a cycle label. Using the period marker as the normalization range, the contribution of multiple degradation periods is normalized to obtain the feature dynamic attention weight of each heterogeneous fusion feature. Based on the periodic markers, multiple dynamic attention weights of the features are output in a structured manner to form the multimodal data fusion rules.

2. The substation equipment degradation status assessment method based on multimodal data fusion as described in claim 1, characterized in that, The multimodal prior database includes multimodal sensing data and operation and maintenance management data.

3. The substation equipment degradation status assessment method based on multimodal data fusion as described in claim 1, characterized in that, Constructing a degradation state assessment model based on ensemble learning, including: Multiple meta-evaluation models are configured differently, and the multiple meta-evaluation models are randomly copied to obtain a meta-evaluation model set; Based on the feature dynamic attention weights of multiple heterogeneous fusion features, the integration selection probability of the meta-evaluation model for each heterogeneous fusion feature is defined; Based on the integration selection probability, multiple meta-evaluation models are driven to iteratively select multiple heterogeneous fusion features until the meta-evaluation model has a corresponding heterogeneous fusion feature. The integrated output layer of multiple meta-evaluation models is configured as a weighted integration, and the integrated output weights of the weighted integration are defaulted to equal weights.

4. The substation equipment degradation status assessment method based on multimodal data fusion as described in claim 1, characterized in that, The integrated output weights of the degradation state assessment model are initialized according to the multimodal data fusion rules, and iterative training is performed, including: Determine the real-time maintenance cycle and the neighborhood maintenance cycle of the target substation equipment, wherein the neighborhood maintenance cycle includes at least one of a first neighborhood maintenance cycle with a cycle greater than the real-time maintenance cycle and a second neighborhood maintenance cycle with a cycle less than the real-time maintenance cycle. Based on the multimodal prior database, a training sample dataset is constructed by combining the data paradigms of the real-time maintenance cycle and the neighborhood maintenance cycle. Based on the multimodal data fusion rules, the integrated output weights of the degradation state assessment model are initialized according to the real-time maintenance cycle and the neighborhood maintenance cycle, respectively, to obtain a first integrated output layer and several second integrated output layers; Based on the training sample dataset, the degradation state assessment model is subjected to iterative supervised learning, and after each iteration, the output is obtained by passing the first integrated output layer and several second integrated output layers respectively to obtain a heterogeneous output set. The heterogeneous output set is traversed to evaluate accuracy verification, and supervised learning is performed iteratively until each output in the heterogeneous output set meets the accuracy requirements of the target substation equipment.

5. The substation equipment degradation status assessment method based on multimodal data fusion as described in claim 1, characterized in that, Real-time multimodal maintenance data is collected and input into the degradation state assessment model to obtain quantitative degradation trends. The degradation level is determined through integral calculation, and the output is the degradation state assessment result, including: Acquire the multimodal maintenance data under the real-time maintenance cycle; The multimodal maintenance data is standardized according to the preset neighborhood maintenance cycle data paradigm. Using the maintenance cycle as an index, the standardized processing results are input into the degradation state assessment model, wherein the degradation state assessment model includes multiple state assessment sub-models with differentiated maintenance cycles. Obtain the first degradation trend under the real-time maintenance cycle and several second degradation trends under the neighboring maintenance cycle; Calculate and integrate the residual distributions of several second degradation trends and the first degradation trend respectively to obtain the verification parameter. If the verification parameter satisfies the preset confidence level verification value, then output the first degradation trend as the quantitative degradation trend. The cumulative degradation amount is obtained by integrating the historical quantitative degradation trend with the quantitative degradation trend, and the degradation level is determined by matching it with a preset degradation grading rule. The degradation level and the quantitative degradation trend are combined and output as the degradation status assessment result.

6. The substation equipment degradation status assessment method based on multimodal data fusion as described in claim 1, characterized in that, The dimensional constraints include: Dimension number constraint, used to limit the maximum number of feature dimensions contained in the candidate fusion features; Dimension repetition constraint is used to limit the candidate fusion features from containing intuitively repetitive feature dimensions. Dimension combination constraints are used to exclude specific feature dimension combinations that are determined to be invalid or inefficient based on prior knowledge. The dimension combination constraints are determined based on a preset prior knowledge base.

7. A substation equipment degradation status assessment system based on multimodal data fusion, characterized in that, The system is used to perform the method according to any one of claims 1-6, the system comprising: The multimodal prior database acquisition module is used to construct a multimodal prior database for target substation equipment; The multimodal data fusion rule acquisition module is used to analyze and determine heterogeneous fusion features based on the multimodal prior database, and generate feature dynamic attention weights under differentiated maintenance cycles to form multimodal data fusion rules. The degradation state assessment model acquisition module is used to construct a degradation state assessment model based on ensemble learning, initialize the ensemble output weights of the degradation state assessment model according to the multimodal data fusion rules, and perform iterative training. The degradation status assessment result acquisition module is used to collect multimodal maintenance data in real time and input it into the degradation status assessment model to obtain quantitative degradation trends, calculate the degradation level by integration, and output the degradation status assessment result.

Images