Method, system, device and medium for intelligent evaluation of pole-mounted switch health state based on feature attention and multi-task learning

Abstract

The application discloses a method, system, device and medium for intelligent evaluation of health state of pole-mounted switch based on feature attention and multi-task learning, and belongs to the technical field of power equipment state monitoring. The method comprises the following steps: acquiring multi-source operating parameter data of the pole-mounted switch, and constructing a multi-dimensional heterogeneous feature preprocessing model; performing deviation elimination and dimensionless normalization through the multi-dimensional heterogeneous feature preprocessing model to obtain a standard monitoring feature sequence; and extracting local space-time dependent features through the standard monitoring feature sequence combined with multi-scale mapping logic to obtain a multi-dimensional enhanced feature vector. The application provides high-quality data support for subsequent evaluation through multi-dimensional heterogeneous preprocessing and deviation elimination, dimensionless normalization of multi-source operating parameter data, extracts local space-time dependent features through multi-scale mapping to generate a multi-dimensional enhanced feature vector, realizes feature recombination and weighting combined with adaptive weight distribution, strengthens key information representation, and improves the feature description ability for the health state.

Description

Technical Field

[0001] This invention relates to the field of power equipment condition monitoring technology, specifically to a method, system, device, and medium for intelligent assessment of the health status of pole-mounted switches based on feature attention and multi-task learning. Background Technology

[0002] With the increasing prevalence of smart grids and the growing number of electrical devices, the distribution network, as the core of power transmission and distribution, has seen 10 kV overhead lines become the mainstay of urban distribution networks, handling over 60% of urban power supply. Pole-mounted switches are key equipment on these lines, primarily responsible for isolating affected areas and transferring electrical loads during faults. Their proper functioning directly determines how long it takes for power to be restored and whether power supply can be sustained after a grid failure.

[0003] As power distribution network maintenance becomes increasingly intelligent, various monitoring technologies are now available, such as Hall effect sensors, wireless temperature measurement, and vibration monitoring. These technologies are applied to the status monitoring of pole-mounted switches, collecting data from various sources, including online operation data, offline test data, and historical maintenance records, providing a data foundation for assessing equipment health. However, extracting useful information from this vast amount of diverse and complex data remains a significant challenge. The core issue is: how to remove irrelevant and interfering information from this multi-dimensional monitoring data, highlighting key fault-related characteristics, and automatically and accurately determining the health status of the pole-mounted switch as "good / normal / abnormal / critical" without overly relying on expert experience. Without solving this problem, intelligent maintenance of the power distribution network cannot be advanced.

[0004] To address this need, the industry has conducted research in various aspects, including evaluation model optimization, weight determination, hardware support, and algorithm application, but none of these have solved the core challenges mentioned above. Specific problems include: In evaluating models and determining weights, commonly used methods such as the Analytic Hierarchy Process (AHP), entropy weight method combined with AHP, and local penalty incentive-based variable weight fuzzy comprehensive evaluation, even if some schemes use historical data to correct the judgment matrix or use the objectivity of the entropy weight method to compensate for the subjectivity of the AHP, still essentially rely on expert experience. How to allocate weights and determine relevant parameters all depend on experts, making them unsuitable for large-scale maintenance scenarios with varying equipment conditions. Furthermore, these methods use fixed weights, which cannot be adjusted according to changes in the equipment's state throughout its entire lifespan. When encountering complex situations where multiple characteristics interact, the accuracy of the judgment will be significantly reduced. Summary of the Invention

[0005] To address the aforementioned technical challenges, a method for intelligent health status assessment of pole-mounted switches based on feature attention and multi-task learning is proposed. This method includes: acquiring multi-source operational parameter data of the pole-mounted switches and constructing a multi-dimensional heterogeneous feature preprocessing model; performing bias removal and dimensional normalization through the multi-dimensional heterogeneous feature preprocessing model to obtain a standard monitoring feature sequence; extracting local spatiotemporal dependent features by combining the standard monitoring feature sequence with multi-scale mapping logic to obtain a multi-dimensional enhanced feature vector; analyzing the contribution differences of different monitoring dimensions using the multi-dimensional enhanced feature vector to obtain an adaptive weight distribution, and performing feature recombination and weighting; based on the recombined and weighted feature sequence, combining deep nonlinear mapping and a multi-task collaborative architecture to obtain continuous-state health evaluation indicators and discrete-state health level distributions; performing correlation verification by combining the continuous-state health evaluation indicators and discrete-state health level distributions to output a comprehensive evaluation result of the pole-mounted switches.

[0006] As a preferred embodiment of the intelligent assessment method for the health status of pole-mounted switches based on feature attention and multi-task learning described in this invention, the construction of the multidimensional heterogeneous feature preprocessing model includes: constructing an appropriate outlier identification mechanism based on the attribute features of the multidimensional heterogeneous monitoring parameters to be processed; performing filtering processing on abnormal data by combining the outlier identification mechanism with data distribution characteristics; and verifying the correlation evolution law between multidimensional heterogeneous features based on the filtering processing results to obtain the multidimensional heterogeneous feature preprocessing model.

[0007] As a preferred embodiment of the intelligent assessment method for the health status of pole-mounted switches based on feature attention and multi-task learning described in this invention, the method includes: performing deviation elimination and dimensional normalization through the multidimensional heterogeneous feature preprocessing model to obtain a standard monitoring feature sequence; performing threshold determination on multi-source operating parameter data through the multidimensional heterogeneous feature preprocessing model to identify abnormal fluctuation values; performing numerical interpolation to fill information gaps in abnormal fluctuation values ​​through correlation weighted logic; and performing statistical distribution standardization processing on the completed data sequence to obtain the standard monitoring feature sequence.

[0008] As a preferred embodiment of the intelligent assessment method for the health status of a pole-mounted switch based on feature attention and multi-task learning described in this invention, the method involves: extracting local spatiotemporal dependent features by combining the standard monitoring feature sequence with multi-scale mapping logic to obtain a multidimensional enhanced feature vector, including: extracting fluctuation features at different time scales based on the standard monitoring feature sequence to construct a state assessment model; generating multiple parallel feature subspaces through the state assessment model; and performing numerical compression and dimensional concatenation based on the feature subspaces to generate a multidimensional enhanced feature vector.

[0009] The beneficial effects of this preferred technical solution are as follows: By leveraging multi-scale mapping logic to mine fluctuation characteristics across different time dimensions, it helps the state assessment model to more comprehensively perceive changes in the operating state of the pole-mounted switch, improving the completeness and accuracy of feature perception and providing a more representative basis for subsequent assessments. Through the construction of parallel feature subspaces, it effectively captures potential correlation patterns between monitoring parameters, breaking the limitations of single feature analysis and enhancing the ability to mine state characteristics under complex operating conditions. The multi-dimensional enhanced feature vector generated through numerical compression and dimensional concatenation can integrate multi-dimensional effective information, reduce redundant interference, and improve the reliability of subsequent health status assessments.

[0010] As a preferred embodiment of the intelligent assessment method for the health status of a pole-mounted switch based on feature attention and multi-task learning described in this invention, the following steps are taken: combining the multi-dimensional enhanced feature vector, analyzing the contribution differences of different monitoring dimensions to obtain an adaptive weight distribution, and performing feature reorganization weighting, including: performing nonlinear semantic transformation on the multi-dimensional enhanced feature vector to generate an initial feature matrix reflecting the feature interaction capability; constructing a weight scoring logic based on a feedforward architecture to perform importance quantification on each dimension in the initial feature matrix; locking the weight probability distribution of each dimension through scoring normalization processing to obtain an adaptive weight distribution; and performing element-wise multiplication operation between the adaptive weight distribution as an adjustment factor and the multi-dimensional enhanced feature vector to achieve feature reorganization weighting by dynamically adjusting the amplitude ratio of each feature component.

[0011] As a preferred embodiment of the intelligent health status assessment method for pole-mounted switches based on feature attention and multi-task learning described in this invention, the method involves obtaining continuous-state health evaluation indicators and discrete-state health level distributions based on the recombined and weighted feature sequence, combined with deep nonlinear mapping and a multi-task collaborative architecture. This includes: inputting the recombined and weighted feature sequence into a multi-level fully connected architecture to perform hierarchical mapping, and extracting abstract representations through layer-by-layer dimensional compression; constructing multi-task parallel branches, and inputting the abstract representations into regression evaluation branches and classification evaluation branches respectively; performing full-range linear regression mapping through the regression evaluation branch, and outputting continuous-state health evaluation indicators by combining numerical constraint logic; and performing state probability mapping through the classification evaluation branch, and outputting discrete-state health level distributions by combining discrete decision logic.

[0012] As a preferred embodiment of the intelligent assessment method for the health status of pole-mounted switches based on feature attention and multi-task learning described in this invention, the following steps are taken: Performing correlation verification by combining the continuous-state health evaluation index with the discrete-state health level distribution to output a comprehensive assessment result for the pole-mounted switches includes: performing a mapping consistency comparison based on the continuous-state health evaluation index and the discrete-state health level distribution to trigger a reliability confirmation logic for the assessment result; determining the physical matching degree between the classification result and the regression value through the reliability confirmation logic to obtain health status representation information; and combining the health status representation information with historical operation and maintenance benchmarks to encapsulate the assessment conclusion and output the comprehensive assessment result for the pole-mounted switches.

[0013] The beneficial effects of this preferred technical solution are as follows: By comparing the consistency of the mapping between continuous and discrete health indicators, the reliability verification logic is triggered, which can screen out biased information in the evaluation process in advance, ensuring the credibility of subsequent evaluation results. By using the reliability verification logic to determine the physical matching degree of classification and regression results, core health status characterization information is accurately identified, avoiding interference from invalid information and improving the pertinence of health status description. Combining historical operation and maintenance benchmarks with the encapsulated evaluation conclusions makes the output comprehensive evaluation results more closely match the needs of actual operation and maintenance scenarios.

[0014] As a preferred embodiment of the intelligent health status assessment system for pole-mounted switches based on feature attention and multi-task learning described in this invention, the system is characterized by comprising: a multi-dimensional feature preprocessing module for acquiring multi-source operating parameter data of the pole-mounted switch and constructing a multi-dimensional heterogeneous feature preprocessing model; a standard feature generation module for performing deviation elimination and dimensional normalization through the multi-dimensional heterogeneous feature preprocessing model to obtain a standard monitoring feature sequence; a feature enhancement extraction module for extracting local spatiotemporal dependent features by combining the standard monitoring feature sequence with multi-scale mapping logic to obtain a multi-dimensional enhanced feature vector; an adaptive weighting module for combining the multi-dimensional enhanced feature vector to analyze the contribution differences of different monitoring dimensions, obtain an adaptive weight distribution, and perform feature recombination and weighting; a multi-task collaborative assessment module for obtaining continuous-state health evaluation indicators and discrete-state health level distributions based on the recombined and weighted feature sequence, combined with deep nonlinear mapping and a multi-task collaborative architecture; and a comprehensive assessment output module for performing correlation verification by combining the continuous-state health evaluation indicators and discrete-state health level distributions to output the comprehensive assessment result of the pole-mounted switch.

[0015] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of the method for intelligent assessment of the health status of a pole-mounted switch based on feature attention and multi-task learning.

[0016] A computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the method for intelligent assessment of the health status of a pole-mounted switch based on feature attention and multi-task learning.

[0017] The beneficial effects of this invention are: by preprocessing multidimensional heterogeneous data from multiple sources, removing biases, and normalizing dimensions, the reliability and consistency of monitoring features are improved, providing high-quality data support for subsequent evaluation.

[0018] By extracting local spatiotemporal dependent features through multi-scale mapping and generating multi-dimensional enhanced feature vectors, and combining them with adaptive weight distribution to achieve feature reorganization and weighting, the representation of key information is strengthened and the ability of features to characterize health status is improved.

[0019] By relying on deep nonlinear mapping and multi-task collaborative architecture to obtain dual health assessment results, and outputting comprehensive conclusions after correlation verification, the system balances the detail and accuracy of the assessment, providing a precise basis for the health management of pole-mounted switches. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 The flowchart illustrates the overall computational process of a method for intelligent assessment of the health status of a pole-mounted switch based on feature attention and multi-task learning, as provided in an embodiment of the present invention.

[0022] Figure 2 The graph shows the loss and accuracy of a method for intelligent assessment of the health status of a pole-mounted switch based on feature attention and multi-task learning, provided in an embodiment of the present invention.

[0023] Figure 3 This is a comprehensive graph showing the health index prediction and classification performance of a method for intelligent assessment of the health status of a pole-mounted switch based on feature attention and multi-task learning, provided in an embodiment of the present invention.

[0024] Figure 4 This is a health status classification prediction confusion matrix diagram of a system for intelligent assessment of the health status of a pole-mounted switch based on feature attention and multi-task learning, provided as an embodiment of the present invention.

[0025] Figure 5 This is a systematic analysis diagram of the prediction error of a system for intelligent assessment of the health status of a pole-mounted switch based on feature attention and multi-task learning, provided as an embodiment of the present invention. Detailed Implementation

[0026] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0027] Example 1, referring to Figure 1 This is the first embodiment of the present invention, which provides a method for intelligent assessment of the health status of a pole-mounted switch based on feature attention and multi-task learning, including: S1: Obtain multi-source operating parameter data of the pole-mounted switch and construct a multi-dimensional heterogeneous feature preprocessing model.

[0028] S2: By performing bias elimination and dimensional normalization through a multidimensional heterogeneous feature preprocessing model, a standard monitoring feature sequence is obtained.

[0029] S3: By combining standard monitoring feature sequences with multi-scale mapping logic, local spatiotemporal dependent features are extracted to obtain multi-dimensional enhanced feature vectors.

[0030] S4: Combine multidimensional enhanced feature vectors to analyze the differences in contribution of different monitoring dimensions, obtain adaptive weight distribution, and perform feature reorganization and weighting.

[0031] S5: Based on the recombined weighted feature sequence, combined with deep nonlinear mapping and multi-task collaborative architecture, the continuous health evaluation index and discrete health level distribution are obtained.

[0032] S6: Combine the continuous state health evaluation index with the discrete state health level distribution to perform correlation verification and output the comprehensive evaluation result of the column switch.

[0033] It should be noted that in existing operation and maintenance scenarios where current and signals are intertwined, the current path for assessing the health status of pole-mounted switches mostly relies on experienced personnel to set feature weights. This setting process is affected by individual cognitive differences and has adaptation barriers when facing a large number of operation and maintenance tasks. The feature processing stage has not formed an adaptive mechanism that can be adjusted autonomously. Feature engineering operations completed by manual intervention or feature extraction behavior that only focuses on a single dimension are prone to loss of weak fault-related features and cannot achieve effective separation of interfering information.

[0034] Therefore, to address the aforementioned issues, this invention, through steps S1-S6, acquires multi-source operational parameter data of pole-mounted switches and simultaneously constructs a multi-dimensional heterogeneous feature preprocessing model. This model is used to eliminate data biases and normalize dimensions, ultimately obtaining monitoring features with standardized attributes. Multi-scale mapping logic is combined to extract local spatiotemporal dependent features, generating multi-dimensional enhanced feature vectors. By analyzing the differences in the contribution of monitoring dimensions, an adaptive weight distribution is obtained. Based on this, feature reorganization and weighting operations are performed to enhance the expression of key fault features. Relying on deep nonlinear mapping and a multi-task collaborative architecture, continuous-state health evaluation indicators and discrete-state health level distributions are simultaneously acquired. This achieves adaptation to the demand for accurate assessment of the health status of pole-mounted switches during the large-scale operation and maintenance of smart grids.

[0035] Example 2, refer to Figure 1 - Figure 2 This is the second embodiment of the present invention, which differs from the first embodiment in that: the method for intelligent assessment of the health status of a pole-mounted switch based on feature attention and multi-task learning further includes: Specifically, in step S1, multi-source operating parameter data of the pole-mounted switch are obtained, and a multi-dimensional heterogeneous feature preprocessing model is constructed, including the following steps A1-A3: A1: Based on the attribute characteristics of the multidimensional heterogeneous monitoring parameters to be processed, an adapted outlier identification mechanism is constructed. A2: By combining the outlier identification mechanism with the data distribution characteristics, outlier data is filtered. A3: Based on the filtering results, the correlation and evolution law among multidimensional heterogeneous features is verified, and a multidimensional heterogeneous feature preprocessing model is obtained.

[0036] In this embodiment, step A1 involves constructing a recognition mechanism based on statistical criteria by combining the attribute differences and data distribution characteristics of the 17-dimensional monitoring features of the pole-mounted switch. This includes the following steps A111-A113: A111: Collect multi-source operating parameter data of pole-mounted switches to form a raw dataset containing online monitoring data, offline test data and historical ledger data. The online monitoring data includes real-time parameters such as contact temperature and control box temperature and humidity. The offline test data includes test parameters such as insulation resistance. The historical ledger data records statistical parameters such as the number of annual faults and the number of operations. A112: Perform data distribution characteristic analysis on the 17-dimensional feature parameters in the original dataset, determine the statistical indicators such as the mean and standard deviation of each feature parameter, and clarify the normal fluctuation range of different parameters. Among them, the distribution pattern of fault-sensitive parameters such as insulation resistance and opening / closing coil current needs to be analyzed in detail. A113: Based on the distribution characteristics of each feature parameter, an outlier identification mechanism with the 3σ criterion as its core is constructed. An outlier judgment threshold is set for each parameter. This threshold is calculated by adding or subtracting three times the standard deviation of each parameter and is used for the initial screening of subsequent outlier data.

[0037] In an optional embodiment, the construction method can also be a collaborative identification mechanism combining parameter correlation, including the following steps A121-A123: A121: Conduct correlation analysis on the 17-dimensional monitoring feature parameters to identify feature combinations with strong correlations, such as parameter pairs with coupled relationships, such as ambient temperature and contact temperature, and control box temperature and humidity. A122: Based on the inherent mapping relationship of strongly correlated feature combinations, a collaborative verification model is established. When the value of a certain parameter exceeds the single parameter 3σ threshold, the value status of its associated parameters is used for auxiliary judgment. A123: Integrating the results of single-parameter statistical judgment and correlation parameter collaborative verification, the final outlier identification rules are generated, improving the accuracy of outlier identification and reducing misjudgments in extreme environments.

[0038] In another alternative embodiment, the construction method can also be to introduce a composite identification mechanism for data integrity verification, including the following steps A131-A134: A131: Perform an integrity scan on the collected multi-source operational parameter data, calculate the missing rate of each characteristic parameter, and mark the time node and device number of the missing data. A132: For parameters with a missing rate lower than a preset threshold, an improved KNN interpolation method is used for preprocessing. This method improves the interpolation accuracy by adaptively adjusting the weights of nearest neighbor samples through correlation weighting and local density. A133: Apply the 3σ criterion to the interpolated and completed dataset for preliminary outlier identification to obtain a candidate outlier dataset. A134: Combine historical fault records of the equipment to perform secondary verification on the candidate abnormal data set, eliminate false abnormal data caused by temporary failures in the data acquisition link, and finally determine the output result of the outlier identification mechanism.

[0039] It should be noted that the outlier identification mechanism needs to be adapted to the heterogeneous characteristics of multi-source data from the pole-mounted switch. It must not only cover the distribution differences of different types of parameters, but also consider actual acquisition scenarios such as data loss and link interference, so as to ensure the accuracy of subsequent data preprocessing.

[0040] Furthermore, in step S2, bias removal and dimensional normalization are performed using a multidimensional heterogeneous feature preprocessing model to obtain a standard monitoring feature sequence, including the following steps B1-B3: B1: Use a multi-dimensional heterogeneous feature preprocessing model to perform threshold determination on multi-source operating parameter data to identify abnormal fluctuation values; B2: To fill in the information gaps in abnormal fluctuation values, numerical interpolation is performed using correlation-weighted logic. B3: Perform statistical distribution standardization processing on the completed data sequence to obtain the standard monitoring feature sequence.

[0041] In this embodiment, step B3 involves processing the data by using Z-Score normalization to eliminate the influence of different dimensions, including the following steps B311-B313: B311: Divide the data sequence after step B2 into batches. Based on the 5-year operation and maintenance data of 237 pole-mounted switches in a power grid, divide it into training set, validation set and test set according to equipment number and time dimension. The training set is used to calculate the statistical parameters required for standardization. B312: Calculate the mean and standard deviation of each of the 17 feature parameters in the training set. Subtract the corresponding mean from the original value of each feature parameter and then divide it by the corresponding standard deviation to obtain the standardized feature value. B313: Apply the mean and standard deviation calculated from the training set to the standardization process of the validation and test sets to ensure that all datasets adopt a unified standardization benchmark and generate standard monitoring feature sequences with uniform dimensions and consistent units.

[0042] In an optional embodiment, the processing can also be performed by using a min-max normalization method to achieve interval mapping of feature values, including the following steps B321-B323: B321: Calculate the maximum and minimum values ​​of each feature parameter in the training set to determine the range of numerical variation for each parameter; B322: The original values ​​of each feature parameter are transformed to the range [0,1] by a linear mapping formula. The mapping formula is the original value minus the minimum value of the parameter, divided by the difference between the maximum and minimum values. B323: Perform extreme value verification on the mapped feature sequence to eliminate mapping deviations caused by abnormal extreme values ​​in the original data, ensuring that the standardized data sequence is evenly distributed and providing stable input for subsequent feature extraction.

[0043] In another alternative embodiment, the processing can also be carried out by combining a parameter type differentiation standardization strategy, including the following steps B331-B334: B331: The 17-dimensional characteristic parameters are divided into two categories: continuous parameters (such as contact temperature and insulation resistance) and discrete parameters (such as pollution level and annual number of failures). B332: The Z-Score standardization method is used to eliminate the influence of dimensions for continuous parameters, and the one-hot encoding method is used for standardization of discrete parameters, converting the category information into a numerical vector; B333: Concatenate the dimensions of the encoded discrete parameter vector and the standardized continuous parameter sequence to form a fused feature matrix; B334: Normalize the fused feature matrix to unify the numerical range of all elements in the matrix to the [-1,1] interval, thereby improving the convergence speed of subsequent model training.

[0044] It should be noted that the core objective of standardization is to eliminate the differences in the dimensions and numerical ranges of different monitoring parameters, and to ensure that each feature parameter has an equal weight contribution basis in model training. The selection of different standardization methods should be determined according to the type and distribution characteristics of the feature parameters, taking into account both processing efficiency and model adaptability.

[0045] Based on the completion of S1-S2, in order to identify the health status of the pole-mounted switch, a scientific status assessment index system needs to be constructed, including two parts: a feature generation model and a status grading standard, as detailed below: This study's data generation is based on statistical analysis of five years of operation and maintenance data from 237 pole-mounted switches in a provincial power grid. Through analysis, the baseline value ranges of each key characteristic parameter under healthy and typical fault conditions were determined, and a linear interpolation model was constructed based on this. This model assumes that the equipment characteristic values ​​change linearly within the aforementioned baseline range. This method simplifies complex physical processes while ensuring the generated data has good physical rationality and can effectively capture the basic trajectory of equipment degradation, thus providing high-quality, large-scale training data for subsequent machine learning models. The generation model for each monitoring feature is as follows: (1) in, For the first The specific numerical value of each feature This serves as a baseline value for the characteristics of the device when it is in perfect health. This serves as a baseline value for the characteristics of a complete equipment failure. It is Gaussian noise. For health, , Indicates complete health. This indicates a complete failure.

[0046] The four-level status grading standard, referencing the "Q / GDW645-2011 Distribution Network Equipment Status Evaluation Guidelines" and combined with operational and maintenance practice requirements, includes the continuous health index.S The system is categorized into four discrete levels: "Good / Normal / Abnormal / Severe". The grading thresholds were determined using the statistical distribution of five years' worth of maintenance data from 237 devices, ensuring that the failure rate for each level matches the maintenance response cost. The grading model is as follows: (2) In this classification system, 1, 2, 3, and 4 represent good, normal, abnormal, and severe, respectively. This four-level classification standard not only considers the numerical distribution of health status but also emphasizes the engineering response measures corresponding to each level, achieving effective integration from status assessment to operational and maintenance decisions.

[0047] Combining the preprocessing logic of steps S1 and S2, to eliminate dimensional differences and monitoring noise, a two-step preprocessing process of "outlier removal and standardization" needs to be completed before inputting the model. Outlier removal uses the 3σ criterion to identify abrupt changes caused by sensor malfunctions and extreme environments. For missing values, an improved KNN interpolation is used. (The improvement of the "improved KNN interpolation" method is mainly reflected in the optimization of weight calculation. The problem with standard KNN interpolation is that when filling in missing values, the traditional KNN interpolation method usually finds the K nearest samples based on Euclidean distance and then simply takes their mean or a weighted average based on the inverse of the distance. This approach assumes that the contributions of all nearest neighbor samples are homogeneous and fails to fully consider the correlation between different features.) The differences and local distribution characteristics of the samples. Improvement scheme: Relevance weighting: When calculating the distance between samples, we not only consider Euclidean distance, but also combine the correlation coefficient between features to adjust the weights. Nearby samples that are more relevant to the missing sample in terms of features will be given greater weights, because they may contain more relevant information for interpolation. Local density adaptation: We consider the local distribution density of nearby samples. For nearby points in sparse regions, their information may be more unique and important, so their weights will be appropriately increased to avoid samples in high-density regions "drowning" key information from sparse regions. Standardization processing: Z-Score standardization is used to eliminate the influence of dimensions, and the formula is as follows: (3) in, These are the standardized eigenvalues; Features The mean on the training set; for The standard deviation on the training set.

[0048] Furthermore, in step S3, local spatiotemporal dependent features are extracted by combining the standard monitoring feature sequence with multi-scale mapping logic to obtain a multi-dimensional enhanced feature vector, including the following steps C1-C3: C1: Extract fluctuation features at different time scales based on standard monitoring feature sequences and construct a state assessment model; C2: Generates multiple parallel feature subspaces through a state evaluation model; C3: Performs numerical compression and dimension concatenation based on the feature subspace to generate multidimensional enhanced feature vectors.

[0049] In this embodiment, step C1 involves extraction, which is performed by using multi-scale convolutional kernels in parallel to capture feature patterns across different time dimensions, including the following steps C111-C113: C111: Divide the standard monitoring feature sequence into continuous time-series feature segments according to the time window length. Each segment contains 17-dimensional feature parameters corresponding to a preset number of time steps, and constructs a time-series input matrix. C112: Construct a feature extraction structure with three parallel convolutional branches. Each branch uses a one-dimensional convolutional kernel with different temporal widths, namely 3, 5, and 7, to capture short-term transient features, mid-range fluctuation features, and long-range trend features, respectively. For example, for parameters with transient characteristics such as the opening and closing coil current, a convolution kernel with a time width of 3 can effectively capture the instantaneous current change characteristics during opening and closing; for parameters with slow-changing characteristics such as insulation resistance, a convolution kernel with a time width of 7 can better extract its long-term degradation trend.

[0050] C113: Each convolutional branch performs feature mapping on the temporal input matrix through convolution operations, generating feature maps at the corresponding time scales, and completing the initial extraction of fluctuation features at different time scales.

[0051] In an optional embodiment, for extraction, the extraction method can also be an extraction method combining sliding window and multi-resolution analysis, including the following steps C121-C123: C121: Set three different sizes of sliding windows, with window lengths corresponding to short, medium, and long time scales respectively. The sliding step size is determined according to the data sampling frequency to ensure that the overlap rate between windows meets the feature continuity requirements. C122: By traversing and sampling the standard monitoring feature sequence through a sliding window, a set of subsequences at different time scales is obtained. For each subsequence, statistical characteristics such as variance, peak value, and mean are calculated to form a statistical feature matrix. C123: Wavelet transform is used to perform multi-resolution analysis on subsequences at various time scales, extracting low-frequency trend components and high-frequency fluctuation components at different scales, and fusing statistical feature matrices with wavelet transform features to obtain a multi-scale fluctuation feature set.

[0052] In another alternative embodiment, for extraction, the extraction method can also be a multi-scale feature extraction method that introduces a temporal attention mechanism, including the following steps C131-C134: C131: Construct a basic multi-scale convolutional feature extraction structure, using convolutional kernels with time widths of 3, 5, and 7 in parallel to extract initial multi-scale features; C132: Introduces a temporal attention module to calculate the attention weight of each time step in features at different scales. The weight value reflects the importance of that time step to the assessment of the device's health status. C133: Weighted fusion of attention weights with initial features at the corresponding scale enhances feature representation at key time steps and suppresses interference from irrelevant time steps; C134: Standardize the features at each scale after weighted fusion to eliminate the amplitude differences between features at different scales and obtain the final multi-scale fluctuation feature sequence.

[0053] It should be noted that the fluctuation characteristics at different time scales correspond to different types of fault modes of pole-mounted switches. Short-term transient characteristics are mostly related to sudden faults such as mechanical jamming, while long-term trend characteristics correspond to progressive faults such as insulation aging. Multi-scale feature extraction can comprehensively capture the evolution information of equipment health status, providing rich feature support for subsequent feature weighting and evaluation models.

[0054] The overall architecture of the state assessment model is as follows: Figure 1 As shown, the input is a standardized 17-dimensional monitoring feature vector, which first enters the convolutional feature extraction module. This module uses one-dimensional convolution operations to capture the local dependencies between features and initially extracts feature representations with structural information. The feature sequence is then input to the feature adaptive weighting module, which dynamically learns the importance weights of each feature dimension through an attention mechanism to strengthen key features and weaken the contribution of redundant or noisy features, thereby optimizing the representation capability of the feature space.

[0055] The weighted feature vectors are fed into the MLP backbone network module, which consists of two fully connected layers. Each layer includes linear transformation, batch normalization, ReLU activation function and Dropout regularization operation. Through nonlinear mapping, the features are deeply abstracted and information is fused to extract a high-level semantic representation that describes the health status of the device.

[0056] The network ultimately outputs a 64-dimensional feature vector, which is processed in parallel by a multi-task output module: one branch performs health status classification (discrete output), and the other branch performs health index regression (continuous output). This multi-task learning framework promotes knowledge complementarity between the two tasks by sharing feature representations, which is beneficial to improving the model's generalization performance.

[0057] In summary, the AttentionMLP model constructs an end-to-end mapping from raw features to comprehensive evaluation through the coordinated operation of the four modules mentioned above. The model is trained using a multi-task objective function that combines classification cross-entropy loss and regression mean squared error loss, enabling joint optimization of local feature extraction, attention weight allocation, feature abstraction, and decision output, thereby improving the accuracy and systematic nature of state evaluation.

[0058] To meet the multi-scale feature extraction requirements of step S3 and effectively capture the local temporal features and cross-parameter correlation patterns in the multi-source monitoring data of the pole-mounted switch, a one-dimensional convolutional neural network is introduced as a feature extractor before the feature adaptive weighting module. This module performs local feature learning on 17-dimensional feature parameters, aiming to enhance the model's sensitivity and expressive ability to weak fault features. Its mathematical model is as follows: Suppose the input is a standardized feature sequence, Where T is the length of the time window. Convolution kernel Perform sliding computation over the time dimension to output the feature map. : (4) In the formula k The convolution kernel's time width, C This refers to the number of output channels. This is achieved through parallel operations using multi-scale convolutional kernels (…). k =3,5,7), the model simultaneously captures short-term transient, medium-range fluctuation and long-range trend characteristics, improving the ability to characterize fault modes at different time scales such as the transient process of opening and closing and the slow degradation of insulation.

[0059] The outputs of convolutions at each scale, after dimensionality reduction via max pooling, are concatenated along the feature channel dimensions as follows: (5) Features after fusion As input to the subsequent attention module, The time dimension is represented by pooling. This design preserves multi-scale feature information while enhancing feature robustness through pooling operations, aiming to provide a clearer and more stable basis for dynamically assigning weights to different features in the model.

[0060] Furthermore, in step S4, the contribution differences of different monitoring dimensions are analyzed by combining the multidimensional enhanced feature vectors to obtain an adaptive weight distribution, and feature reorganization and weighting are performed, including the following steps D1-D4: D1: Perform nonlinear semantic transformation on the multidimensional enhanced feature vector to generate an initial feature matrix that reflects the feature interaction capability; D2: Construct a weighted scoring logic based on a feedforward architecture, and perform importance quantification on each dimension in the initial feature matrix; D3: By normalizing the scores, the probability distribution of the weights for each dimension is locked, resulting in an adaptive weight distribution; D4: The adaptive weight distribution is used as an adjustment factor to perform element-wise multiplication with the multidimensional enhanced feature vector. Feature reorganization and weighting are achieved by dynamically adjusting the magnitude ratio of each feature component.

[0061] In this embodiment, step D2, regarding the construction, is carried out by using a shallow feedforward network to quantify feature importance, including the following steps D211-D213: D211: Construct a shallow feedforward network with one hidden layer. The number of neurons in the hidden layer is determined according to the dimension of the multidimensional enhanced feature vector. The activation function is the tanh function to enhance the representation ability of feature interactions. D212: The initial feature matrix after nonlinear semantic transformation is input into a shallow feedforward network. Through linear transformation and activation function operation of the hidden layer, the initial importance score of each feature dimension is generated. For example, for feature dimensions that are closely related to the health status of equipment, such as insulation resistance and contact temperature, the network will generate a higher initial importance score through learning; for feature dimensions that are more disturbing, such as environmental humidity, a lower initial importance score will be generated.

[0062] D213: A linear layer is set in the output layer of the shallow feedforward network to adjust the dimension of the initial importance score of the hidden layer output, ensuring that the dimension of the output score is consistent with the dimension of the multidimensional enhanced feature vector, so as to provide a suitable input for subsequent normalization processing.

[0063] In an optional embodiment, the construction can also be carried out using a scoring logic that combines a multilayer perceptron with an attention mechanism, including the following steps D221-D223: D221: Construct a multilayer perceptron with two hidden layers. The first hidden layer uses the ReLU activation function, and the second hidden layer uses the sigmoid activation function to enhance the network's ability to fit nonlinear features. D222: Introduces a multi-head attention mechanism, which divides the initial feature matrix into multiple subspaces and calculates feature importance scores in each subspace, thereby improving the accuracy of the scores through multi-view evaluation; D223: Weighted fusion of the scoring results from multiple subspaces is performed, with the fusion weights dynamically adjusted based on the evaluation accuracy of each subspace, to generate the final feature importance score sequence.

[0064] In another alternative embodiment, the construction can also be carried out by guiding the scoring logic with prior knowledge of the failure mechanism, including the following steps D231-D234: D231: Based on the fault mechanism analysis of pole-mounted switches, the key feature sets corresponding to the three core faults of insulation aging, mechanical degradation and electrical performance deterioration are determined, providing prior guidance for the scoring logic; D232: Construct a basic feedforward scoring network to generate initial feature importance scores; D233: Introduce a priori weight adjustment factor to appropriately amplify the scores corresponding to the core fault key feature set and constrain the scores of non-key features, so that the scoring results are more in line with the understanding of fault mechanism. D234: By using gradient backpropagation to collaboratively optimize the parameters of the feedforward network and the prior weight adjustment factor, the scoring logic can both conform to the prior knowledge of the domain and adapt to the characteristics of data distribution.

[0065] It should be noted that the core advantage of adaptive weight distribution lies in replacing the traditional manual weighting and constant weight strategy. It can dynamically adjust the weight ratio of each dimension according to the specific situation of the input feature sequence, strengthen the expression of key fault features such as insulation resistance and opening and closing coil current, weaken the interference of redundant or noise features, and improve the accuracy and generalization ability of the subsequent evaluation model.

[0066] The adaptive weight allocation requirement in step S4 is that, in the equipment status assessment task, the ability of each monitoring feature to represent the health status varies significantly.

[0067] Traditional methods typically apply equal weights to features or rely on expert experience to pre-set fixed weights, making it difficult to adaptively capture the dynamic changes in feature importance.

[0068] To overcome this limitation, an attention mechanism is used to dynamically learn and assign feature weights. The adaptive weighting module is implemented by a shallow feedforward network. Its processing flow is as follows: The standardized input features undergo a nonlinear transformation, and the tanh activation function is used to enhance the representational power of feature interactions; an initial importance score is generated for each feature via a linear layer; the scores are normalized to a probability distribution using a softmax function to obtain the attention weights for each feature. The weight values ​​reflect the relative importance of the corresponding feature in the current evaluation. The weighted feature vector is obtained by multiplying the original features by their corresponding weights and summing the results. Its mathematical model can be expressed as: (6) in, , This is the weight matrix; It is the bias vector; For the first Attention weights for each feature, satisfying ; This is the weighted feature vector of the output.

[0069] Furthermore, in step S5, based on the recombined weighted feature sequence, combined with deep nonlinear mapping and a multi-task collaborative architecture, the continuous-state health evaluation index and the discrete-state health level distribution are obtained, including the following steps E1-E4: E1: Input the recombined and weighted feature sequence into a multi-level fully connected architecture to perform hierarchical mapping, and extract the abstract representation by compressing and refining it layer by layer. E2: Construct multi-task parallel branches, and input the abstract representation into the regression evaluation branch and the classification evaluation branch respectively; E3: Perform full-range linear regression mapping through the regression evaluation branch, and output continuous health evaluation indicators by combining numerical constraint logic; E4: By classifying and evaluating the probability mapping of branch execution states, and combining discrete decision logic, output the discrete state health level distribution.

[0070] In this embodiment, step E2, for constructing a multi-task parallel branch, is constructed using a parallel architecture with a shared feature layer and independent output heads, including the following steps E211-E213: E211: Construct a shared feature layer, which reuses the 64-dimensional high-level abstract feature vector output by the multi-level fully connected architecture to achieve feature sharing between regression and classification tasks and promote knowledge complementarity between tasks; E212: Construct two independent output heads: a regression evaluation branch and a classification evaluation branch. The regression evaluation branch consists of a fully connected layer with an output dimension of 1, used to output continuous health evaluation indicators. The classification evaluation branch consists of a fully connected layer and a softmax function with an output dimension of 4, corresponding to four health levels: good, normal, abnormal, and severe. For example, when the high-level abstract feature vector of a certain pole-mounted switch is input into the parallel branch, the regression evaluation branch outputs a continuous state health evaluation index of 0.92, and the classification evaluation branch outputs a probability distribution of [0.95, 0.03, 0.01, 0.01], indicating that the probability that the device is in good health is 95%.

[0071] E213: Set a multi-task loss function, which is a weighted sum of classification cross-entropy loss and regression mean squared error loss. By adjusting the loss weights of the two tasks through hyperparameters, the collaborative optimization of the two tasks can be achieved.

[0072] In an optional embodiment, for constructing multi-task parallel branches, the construction method can also be a parallel branch architecture with adaptive feature allocation, including the following steps E221-E223: E221: After the shared feature layer, a feature allocation module is introduced. This module calculates the contribution of each feature dimension to the regression and classification tasks through an attention mechanism. E222: Based on contribution, the high-level abstract feature vector is divided into three parts: regression-specific features, classification-specific features, and shared features. The regression evaluation branch only accepts regression-specific features and shared features, and the classification evaluation branch only accepts classification-specific features and shared features. E223: Adaptive network structures are designed for the two branches. The regression branch uses a fully connected layer with L2 regularization, while the classification branch uses a structure combining a fully connected layer and a dropout layer to improve the specialization and generalization ability of each task.

[0073] In another alternative embodiment, for constructing multi-task parallel branches, the construction method can also be to introduce a dynamic parallel architecture with task priorities, including the following steps E231-E234: E231: Set task priorities based on operation and maintenance decision requirements, and dynamically adjust the priority weights of regression tasks and classification tasks at different stages of equipment operation and maintenance. Increase the priority of classification tasks during high-fault periods and increase the priority of regression tasks during stable equipment operation periods. E232: Constructs a basic multi-task parallel branching architecture, including a shared feature layer and two independent output heads; E233: Introduces a dynamic weight adjustment module, which adjusts the loss weights of two tasks in the multi-task loss function in real time according to the set task priority, so that the model training is tilted towards the high-priority task. E234: During the model inference phase, a comprehensive evaluation conclusion is generated by combining the current task priority with the output results of the two branches. The output results of high-priority tasks have a higher weight in the conclusion determination.

[0074] It should be noted that the multi-task collaborative architecture achieves the synchronous output of continuous health evaluation indicators and discrete health level distributions through shared feature representations and collaborative optimization. It can quantify the continuous evolution of equipment health status and clarify the discrete health level to which the equipment belongs, providing a comprehensive basis for operation and maintenance decisions that combines qualitative and quantitative methods. At the same time, it improves the generalization performance of the model through knowledge complementarity between tasks.

[0075] Combining the deep mapping requirements of step E1, the main task of the MLP backbone network is to learn the nonlinear mapping relationship required for state evaluation from weighted features. It consists of a two-layer fully connected network with a funnel-shaped design, progressively mapping the feature dimensions from 17 to 128, and then compressing them to 64. This design helps the network gradually condense information and extract higher-level, more abstract feature representations. Each layer includes operations such as linear transformation, batch normalization (BN), ReLU activation, and Dropout regularization. Batch normalization accelerates training and improves stability, the ReLU activation function introduces nonlinearity into the model, and Dropout effectively suppresses overfitting by randomly masking some neurons during training. The model is as follows: (7) in, This is the weight matrix; It is the bias vector; This represents the batch normalization operation, and its calculation formula is: ; final output This refers to the extracted high-level feature vector.

[0076] To provide more comprehensive evaluation information, the model employs multi-task learning, performing classification and regression tasks in parallel in the final stage. High-level feature vectors The data is simultaneously fed into two independent output heads. The classification head outputs the raw scores of the device's four health states, which are then transformed into a probability distribution using the softmax function. The regression head outputs a continuous health index prediction value, which is constrained to the (0,1) interval using the sigmoid function, where 1 represents complete health and 0 represents complete failure.

[0077] The model's final decision is to select the category with the highest probability as the health status classification result, and combine it with the health index obtained from regression to form a comprehensive assessment conclusion that combines qualitative and quantitative methods. The model is as follows: (8) in, ; For class probabilities; For predicted health index; For the first The final predicted state category for each feature value.

[0078] The model is trained by optimizing all parameters by minimizing a multi-task loss function. This loss function is a weighted sum of the cross-entropy loss for the classification task and the mean squared error loss for the regression task, quantified by hyperparameters. and This design balances the importance of the two tasks. It allows the model to simultaneously ensure classification accuracy and regression precision during training. The model is as follows: (9) in, Cross-entropy loss; For the mean squared error loss of the regression task; For loss; This refers to the number of samples in the batch. Indicates the first The true label of each sample; For the model to predict the first Each sample belongs to category c The probability of; For the first The true health index of each sample; For the first Health index of each sample; These are the weighting coefficients for the classification loss; These are the weighting coefficients for the regression loss. The model is trained using the backpropagation algorithm and the AdamW optimizer.

[0079] Furthermore, in step S6, a correlation verification is performed by combining the continuous-state health evaluation index with the discrete-state health level distribution to output the comprehensive evaluation result of the on-pole switch, including the following steps F1-F3: F1: Based on the continuous state health evaluation index and the discrete state health level distribution, perform a mapping consistency comparison to trigger the logic for confirming the reliability of the evaluation results; F2: By verifying the physical match between the classification results and the regression values ​​through reliability confirmation logic, health status characterization information is obtained; F3: Combining health status representation information with historical operation and maintenance benchmarks, encapsulate the evaluation conclusions and output the comprehensive evaluation results of the pole-mounted switch.

[0080] In this embodiment, step F1 involves performing a mapping consistency comparison. The comparison method is to establish a mapping rule between continuous indicators and discrete levels and perform bidirectional verification, including the following steps F111-F113: F111: Based on the statistical distribution of 5 years of operation and maintenance data of multiple pole-mounted switches, establish a mapping rule between continuous health evaluation indicators and discrete health levels. For example, a health index of [0.9, 1.0] corresponds to a good level, [0.7, 0.9) corresponds to a normal level, [0.4, 0.7) corresponds to an abnormal level, and [0, 0.4) corresponds to a severe level. F112: Based on the mapping rules, the continuous health evaluation index is converted into the corresponding theoretical discrete level, and the actual discrete level with the highest probability in the discrete health level distribution is extracted. For example, if the continuous health evaluation index is 0.85, it corresponds to the normal level according to the mapping rule. If the discrete health level distribution is [0.02, 0.96, 0.01, 0.01], and the actual discrete level is the normal level, then the two mappings are consistent.

[0081] F113: Perform two-way verification. On the one hand, it verifies whether the theoretical level of the continuous index conversion is consistent with the actual discrete level. On the other hand, it calculates the degree of deviation between the index range corresponding to the actual discrete level and the continuous index. If the degree of deviation is less than the preset threshold, it is determined that the mapping is consistent.

[0082] In an optional embodiment, for performing a mapping consistency comparison, the comparison method can also be a probability-weighted mapping consistency verification method, including the following steps F121-F123: F121: Preserves the complete probabilistic information of the discrete health level distribution, not only extracting the level with the highest probability, but also considering the probability proportion of other levels; F122: Determine the theoretical level interval corresponding to the continuous health evaluation index according to the mapping rule, calculate the probability proportion of the discrete level corresponding to the interval in the distribution, and at the same time calculate the weighted distance between the continuous index and the center value of each level interval, with the weight being the probability of the corresponding level. F123: Set probability percentage threshold and weighted distance threshold. When the probability percentage of the theoretical level is higher than the threshold and the weighted distance is lower than the threshold, it is determined that the mapping is consistent; otherwise, it is determined that the mapping is inconsistent.

[0083] In another optional embodiment, for performing a mapping consistency comparison, the comparison method can also be a consistency verification method combining the historical state evolution trajectory of the device, including the following steps F131-F134: F131: Extract the historical health status assessment data of the pole-mounted switch, construct the historical state evolution trajectory, and analyze the continuity and regularity of its state changes; F132: Theoretical health status at the current moment is predicted based on historical trajectory, including theoretical continuous indicators and theoretical discrete levels; F133: Compare the consistency between the current continuous health evaluation index and the theoretical continuous index, and between the current discrete health level distribution and the theoretical discrete level, to obtain consistency scores in two dimensions. F134: Weighted fusion of consistency scores between two dimensions. The fusion weight is dynamically adjusted based on the reliability of historical trajectories. If the fusion score is higher than the preset benchmark, it is determined that the overall mapping is consistent.

[0084] It should be noted that the core function of mapping consistency comparison is to verify the reliability of continuous-state evaluation results and discrete-state evaluation results, avoiding erroneous evaluation conclusions caused by deviations in the output of a single task. When the two mappings are consistent, health status characterization information can be directly obtained; when the mappings are inconsistent, a secondary verification process needs to be triggered to re-examine the feature extraction, weight allocation, and model inference processes to ensure that the final comprehensive evaluation result is accurate and reliable, providing credible technical support for distribution network condition-based maintenance.

[0085] Example 3, referring to Figure 2 - Figure 5 This is the third embodiment of the present invention, which differs from the previous two embodiments in that: the method for intelligent assessment of the health status of a pole-mounted switch based on feature attention and multi-task learning further includes, in order to verify and explain the technical effects used in this method, comparing experimental results with scientific demonstration methods to verify the real effect of this method.

[0086] The experimental data came from five years of operation and maintenance data of 237 pole-mounted switches in a provincial power grid, containing 25,000 samples. Each sample consisted of 17-dimensional monitoring features, a continuous health index, and discrete state labels. To evaluate the generalization ability of the model in the technical solution, the experimental data was randomly divided into training, validation, and test sets in a ratio of 7:1.5:1.5. The training set was used for learning model parameters; the validation set was used for model hyperparameter tuning and monitoring of the training process; and the test set served as independent data for the final performance evaluation of the model, ensuring the fairness of the evaluation results.

[0087] The health status assessment of pole-mounted switches requires the establishment of a scientific system encompassing "mechanism adaptation, data support, and standard compliance." This paper, based on equipment failure mechanisms (mechanical degradation, insulation aging, and electrical performance deterioration) and failure mode analysis, and in conjunction with the common technical requirements of the "DL / T593-2016 Standard for High Voltage Switchgear and Control Equipment," constructs an assessment system containing 17 dimensions, as shown in Table 1. The selection of indicators strictly adheres to the principles of comprehensiveness, measurability, and sensitivity.

[0088] Table 1 – 17-Dimensional Feature Table

[0089] Model training dynamics are a core technical feature that reflects its learning and generalization abilities. Figure 2 This document records the changes in loss and accuracy during the training of the AttentionMLP-based state evaluation model. The loss curves for both the training and validation sets exhibit a pattern of initial rapid decline followed by a plateau, with a high degree of fit between the two curves and the difference remaining consistently low. This indicates that the model possesses stable learning performance, effectively extracting general technical patterns from the training data without overfitting. The validation loss value stabilized after 25 training epochs, confirming the rationality of the model structure design and hyperparameter settings, and demonstrating the effectiveness of the employed regularization techniques.

[0090] In terms of model classification performance, both training accuracy and validation accuracy showed a steady upward trend; for example... Figure 2 The validation accuracy eventually stabilized at 95.47%, with a difference of 0.46 percentage points from the training accuracy (95.01%), further confirming that the model's generalization ability meets technical expectations. The accuracy curve remained stable overall without drastic fluctuations, indicating the synergistic adaptability of the AdamW optimizer and the learning rate scheduling strategy, and demonstrating the stability of the model optimization process. These training dynamic characteristics prove that the AttentionMLP model possesses stable and efficient learning characteristics, and its convergence behavior meets the preset technical requirements, making it suitable for subsequent system performance evaluation.

[0091] Accurate prediction of health index is a core technical task in equipment condition assessment, and the performance of the AttentionMLP model in this task has been verified through multi-faceted experiments. Figure 3 (a) Records the correspondence between model predictions and actual values, with data points closely distributed near the diagonal, and the coefficient of determination R. 2 The value reached 0.9911, indicating that the model can accurately capture the continuous changes in the health index. Samples of different health states were evenly distributed in the scatter plot, with no obvious regional clustering deviations, proving that the model maintains stable predictive performance across all health intervals. Statistical analysis of the prediction residuals further validated the model's reliability.

[0092] Figure 3 (b) Record the residual distribution data. The residuals approximately follow a normal distribution with a mean of -0.0033 and a standard deviation of 0.0267. The near-zero mean and small standard deviation statistically confirm the high accuracy and unbiasedness of the model prediction. Figure 3 (c) The recorded classification accuracy data reveals the differences in classification performance of the model under different health conditions, with an accuracy of 96.3% for the "good" state and 94.9% for the "normal" state. These results are consistent with the domain cognition of extreme state feature patterns.

[0093] State classification performance is directly related to the effectiveness of operational and maintenance decisions. Figure 4The recorded confusion matrix data shows that the number of true cases on the main diagonal of the model is significantly higher than that on the off-diagonal elements, with the "good" and "severe" endpoints showing the best classification performance. The main misclassifications occur between adjacent categories; this "nearest neighbor misclassification" pattern aligns with the physical laws governing the continuous evolution of device states and is acceptable in engineering applications.

[0094] The distribution characteristics of prediction errors are a key technical indicator for evaluating model reliability. Figure 5 The recorded error analysis data shows that the distribution of prediction errors across different health index intervals exhibits a regularity: the error band width increases slightly in the middle region of the health index, while the error band narrows relatively at both ends. This phenomenon reflects the changing characteristics of the model's prediction uncertainty for different state intervals. Each error band includes a zero error line, and the range is controlled within a preset reasonable interval, further confirming the unbiasedness of the model's predictions.

[0095] Residual analysis provides in-depth technical support for model performance evaluation. The scatter plot of residuals versus predicted values ​​shows that data points are randomly and evenly distributed on both sides of the zero line, without obvious heteroscedasticity or trend patterns. This random distribution characteristic indicates that the model prediction error is mainly composed of unpredictable random noise, and there is no systematic overestimation or underestimation bias.

[0096] The interpretability of the attention mechanism is a core technical link connecting model decision-making and domain cognition. Analysis of the attention weight distribution characteristics shows that the model can autonomously learn and identify key feature dimensions highly correlated with equipment condition assessment. Experimental results indicate that the model tends to assign higher weights to core monitoring features closely related to equipment health status. This result is consistent with the domain prior knowledge of power equipment condition assessment, demonstrating the model's ability to learn condition assessment logic that conforms to physical laws from data.

[0097] Attention weights exhibit significant dynamic evolution characteristics, allowing the model to generate differentiated weight distributions for different input samples. This adaptive focusing mechanism enables the model to dynamically adjust the level of attention given to different features based on the device's current state; for example, when the device exhibits a certain abnormal state, the model can automatically increase the attention weight of the feature most relevant to that state. This dynamic decision-making mechanism is the core technological foundation for the model to achieve accurate and personalized evaluation, while also enhancing the transparency and reliability of the model's decision-making process.

[0098] Example 4, the fourth embodiment of the present invention, differs from the previous three embodiments in that: a system for intelligent assessment of the health status of a pole-mounted switch based on feature attention and multi-task learning includes a multi-dimensional feature preprocessing module for acquiring multi-source operating parameter data of the pole-mounted switch and constructing a multi-dimensional heterogeneous feature preprocessing model; a standard feature generation module for performing deviation elimination and dimensional normalization through the multi-dimensional heterogeneous feature preprocessing model to obtain a standard monitoring feature sequence; a feature enhancement extraction module for extracting local spatiotemporal dependent features by combining the standard monitoring feature sequence with multi-scale mapping logic to obtain a multi-dimensional enhanced feature vector; an adaptive weighting module for combining the multi-dimensional enhanced feature vector to analyze the contribution differences of different monitoring dimensions, obtain an adaptive weight distribution, and perform feature recombination and weighting; a multi-task collaborative assessment module for obtaining continuous-state health evaluation indicators and discrete-state health level distributions based on the recombined and weighted feature sequence, combined with deep nonlinear mapping and a multi-task collaborative architecture; and a comprehensive assessment output module for performing correlation verification by combining the continuous-state health evaluation indicators and discrete-state health level distributions to output the comprehensive assessment result of the pole-mounted switch.

[0099] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0100] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0101] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0102] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented in combination with any of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0103] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for intelligent assessment of the health status of a pole-mounted switch based on feature attention and multi-task learning, characterized in that: include, Acquire multi-source operating parameter data of pole-mounted switches and construct a multi-dimensional heterogeneous feature preprocessing model; A standard monitoring feature sequence is obtained by performing bias removal and dimensional normalization through a multidimensional heterogeneous feature preprocessing model; By combining standard monitoring feature sequences with multi-scale mapping logic, local spatiotemporal dependent features are extracted to obtain multidimensional enhanced feature vectors; By combining multidimensional enhanced feature vectors, the contribution differences of different monitoring dimensions are analyzed to obtain an adaptive weight distribution, and feature reorganization and weighting are performed. Based on the recombined weighted feature sequence, combined with deep nonlinear mapping and multi-task collaborative architecture, we obtain the continuous health evaluation index and the discrete health level distribution. By combining continuous-state health evaluation indicators with discrete-state health level distribution, a correlation verification is performed, and the comprehensive evaluation results of the column switch are output.

2. The method for intelligent assessment of the health status of a pole-mounted switch based on feature attention and multi-task learning as described in claim 1, characterized in that: Constructing the multidimensional heterogeneous feature preprocessing model includes, Based on the attribute characteristics of the multidimensional heterogeneous monitoring parameters to be processed, an adapted outlier identification mechanism is constructed. By combining an outlier identification mechanism with data distribution characteristics, outlier data is filtered. Based on the filtering results, the correlation and evolution law among multidimensional heterogeneous features is verified, and a multidimensional heterogeneous feature preprocessing model is obtained.

3. The method for intelligent assessment of the health status of a pole-mounted switch based on feature attention and multi-task learning as described in claim 2, characterized in that: By performing bias removal and dimensional normalization through the aforementioned multidimensional heterogeneous feature preprocessing model, the standard monitoring feature sequence is obtained, including: A multidimensional heterogeneous feature preprocessing model is used to perform threshold determination on multi-source operating parameter data to identify abnormal fluctuation values. To fill the information gaps in abnormal fluctuation values, numerical interpolation is performed using correlation-weighted logic. Statistical distribution standardization is performed on the completed data sequence to obtain the standard monitoring feature sequence.

4. The method for intelligent assessment of the health status of a pole-mounted switch based on feature attention and multi-task learning as described in claim 3, characterized in that: By combining the standard monitoring feature sequence with multi-scale mapping logic, local spatiotemporal dependent features are extracted to obtain a multi-dimensional enhanced feature vector, including... Based on standard monitoring feature sequences, fluctuation characteristics at different time scales are extracted to construct a state assessment model; Multiple parallel feature subspaces are generated through a state evaluation model; Numerical compression and dimensional concatenation are performed based on the feature subspace to generate multidimensional enhanced feature vectors.

5. The method for intelligent assessment of the health status of a pole-mounted switch based on feature attention and multi-task learning as described in claim 4, characterized in that: By combining the aforementioned multidimensional enhanced feature vectors, the contribution differences of different monitoring dimensions are analyzed to obtain an adaptive weight distribution, and feature reorganization and weighting are performed. Perform nonlinear semantic transformation on the multidimensional enhanced feature vectors to generate an initial feature matrix that reflects the feature interaction capability; Construct a weighted scoring logic based on a feedforward architecture to perform importance quantification on each dimension of the initial feature matrix; By normalizing the scores, the probability distribution of the weights for each dimension is locked, resulting in an adaptive weight distribution. The adaptive weight distribution is used as an adjustment factor to perform element-wise multiplication with the multidimensional enhanced feature vector, and the feature reorganization weighting is achieved by dynamically adjusting the magnitude ratio of each feature component.

6. The method for intelligent assessment of the health status of a pole-mounted switch based on feature attention and multi-task learning as described in claim 5, characterized in that: Based on the recombined and weighted feature sequences, combined with deep nonlinear mapping and a multi-task collaborative architecture, the continuous-state health evaluation index and the discrete-state health level distribution are obtained, including... The recombined and weighted feature sequence is input into a multi-level fully connected architecture to perform hierarchical mapping, and the abstract representation is extracted and refined through layer-by-layer dimensional compression. Construct multi-task parallel branches and input the abstract representation into the regression evaluation branch and the classification evaluation branch respectively; The full-range linear regression mapping is performed through the regression evaluation branch, and the continuous state health evaluation index is output by combining numerical constraint logic. By classifying and evaluating the probability mapping of branch execution states, and combining it with discrete decision-making logic, a discrete health level distribution is output.

7. The method for intelligent assessment of the health status of a pole-mounted switch based on feature attention and multi-task learning as described in claim 6, characterized in that: By performing correlation verification between the continuous-state health evaluation index and the discrete-state health level distribution, the output of the comprehensive evaluation result of the on-board switch includes: Based on the consistency comparison of the mapping between continuous health evaluation indicators and discrete health level distribution, the logic for confirming the reliability of evaluation results is triggered. By verifying the physical match between the classification results and the regression values ​​through reliability confirmation logic, health status characterization information is obtained. By combining health status information with historical operation and maintenance benchmarks, the evaluation conclusions are encapsulated and the comprehensive evaluation results of the pole-mounted switches are output.

8. A system for intelligent assessment of the health status of a pole-mounted switch based on feature attention and multi-task learning, employing the method for intelligent assessment of the health status of a pole-mounted switch based on feature attention and multi-task learning as described in any one of claims 1 to 7, characterized in that, include: The multi-dimensional feature preprocessing module is used to acquire multi-source operating parameter data of the pole-mounted switch and construct a multi-dimensional heterogeneous feature preprocessing model. The standard feature generation module is used to perform bias removal and dimensional normalization through a multidimensional heterogeneous feature preprocessing model to obtain a standard monitoring feature sequence; The feature enhancement extraction module is used to extract local spatiotemporal dependent features by combining standard monitoring feature sequences with multi-scale mapping logic to obtain multi-dimensional enhanced feature vectors; The adaptive weighting module is used to combine multi-dimensional enhanced feature vectors, analyze the differences in contribution of different monitoring dimensions, obtain an adaptive weight distribution, and perform feature reorganization and weighting. The multi-task collaborative evaluation module is used to obtain continuous health evaluation indicators and discrete health level distributions based on the recombined weighted feature sequences, combined with deep nonlinear mapping and multi-task collaborative architecture. The comprehensive evaluation output module is used to perform correlation verification by combining continuous state health evaluation indicators with discrete state health level distribution, and output the comprehensive evaluation results of the column switch.

9. A computer device comprising a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the method for intelligent assessment of the health status of a pole-mounted switch based on feature attention and multi-task learning as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, the computer program being executed by a processor to implement the steps of the method for intelligent assessment of the health status of a pole-mounted switch based on feature attention and multi-task learning as described in any one of claims 1 to 7.

Images