Hyperspectral data transfer learning state prediction method and system for power line insulator

Abstract

The application discloses a power line insulator hyperspectral data transfer learning state prediction method, system, device and medium, and belongs to the technical field of transfer learning state prediction, and comprises the following steps: dividing source domain labeled samples into a plurality of degradation state intervals according to degradation degrees, constructing a feature effectiveness mapping table for spectral band classification marking, iteratively updating target domain sample state estimation until convergence and recording a historical state estimation sequence; performing feature alignment based on the inclusion relationship of the feature change direction sets of the source domain and the target domain; classifying feedback samples according to maintenance work order operation types, triggering synchronous updating of the feature effectiveness mapping table and feature alignment; querying the historical state estimation sequence for boundary sample, identifying the sources of prediction uncertainty of the insufficient coverage type and the degradation transition period type, and respectively performing prediction state adjustment. The application takes feature physical source differentiation as the core, and realizes physical constraint driving of feature alignment operation, feedback data utilization and boundary sample decision.

Description

Technical Field

[0001] This invention relates to the field of transfer learning state prediction technology, specifically to a method and system for predicting the state of insulators of power lines using hyperspectral data transfer learning. Background Technology

[0002] Transfer learning methods apply the adjusted source domain pre-trained model to the target domain data through domain adaptation techniques. Existing solutions perform uniform alignment of all spectral band features as statistical objects when performing domain adaptation, without distinguishing spectral bands based on the physical source of the features.

[0003] When there are differences in environmental conditions between the target domain and the source domain, the spectral bands driven by environmental factors and the spectral bands reflecting the intrinsic degradation state of the equipment cannot be separated under uniform alignment operations. The distribution deviation of environmental interference bands is simultaneously introduced into the migration results, and the state prediction error accumulates with changes in environmental conditions. Existing solutions also fail to differentiate maintenance feedback data based on the physical meaning of maintenance operations. Different maintenance operations have different mechanisms of action on equipment state, and treating them the same results in the high-value state information carried in the feedback data failing to be specifically transformed into improved model accuracy.

[0004] Furthermore, when the predicted value is near the boundary of adjacent state intervals, the existing scheme does not distinguish the physical source of the prediction uncertainty, and the insufficient coverage is confused with the actual state change, which affects the reliability of subsequent maintenance decision recommendations. Summary of the Invention

[0005] In view of the above-mentioned problems, the present invention provides a method, system, device and medium for predicting the state of hyperspectral data transfer learning of insulators of power lines.

[0006] Therefore, the technical problem solved by this invention is that existing state prediction schemes based on transfer learning do not distinguish between features reflecting the intrinsic degradation state of equipment and features affected by environmental factors during the domain adaptation process. This leads to the feature alignment operation introducing the distribution deviation of environmental interference components into the transfer results. The state prediction error continues to accumulate as the target domain environmental conditions change. The intrinsic state information carried in the maintenance feedback data cannot be effectively utilized due to the lack of a targeted processing path. The mixed sources of prediction uncertainty in boundary samples lead to inaccurate maintenance decision recommendations.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a method for predicting the state of insulators of power collection lines using hyperspectral data transfer learning, comprising, The source domain labeled samples are divided into several deterioration state intervals according to the degree of deterioration. A source domain behavior benchmark for each spectral band is established for each deterioration state interval. The response value distribution characteristics of the target domain samples are compared with the source domain behavior benchmark interval by interval and band by band. The comparison results are used to classify and label the spectral bands to construct a feature validity mapping table. The state estimation of the target domain samples is iteratively updated with the feature validity mapping table as input until convergence. The historical state estimation sequence is recorded during the iterative convergence process. Based on the feature validity mapping table, feature alignment is performed according to the inclusion relationship between the feature change direction sets of the source domain and the target domain for each deterioration state interval; The maintenance feedback samples are classified according to the operation type of the maintenance work order. The classification results are used to perform a local update on the feature validity mapping table. The local update triggers a synchronous update of the feature change direction set and feature alignment. The target domain samples are predicted using the synchronized updated feature alignment results. For target domain samples whose predicted values ​​fall at the boundary of adjacent deterioration state intervals, the historical state estimation sequence is queried to identify the source of prediction uncertainty as insufficient coverage or deterioration transition period. Predictive state adjustment is performed for insufficient coverage samples and deterioration transition period samples respectively.

[0008] This invention provides a hyperspectral data transfer learning state prediction system for insulators of power lines.

[0009] To solve the above technical problems, the present invention provides the following technical solution: a hyperspectral data transfer learning state prediction system for insulators of power lines, comprising: a mapping table construction module, used to divide source domain labeled samples into several deterioration state intervals according to the degree of deterioration, establish a source domain behavior benchmark for each deterioration state interval, compare the response value distribution characteristics of target domain samples with the source domain behavior benchmark interval by interval and band by band, classify and label the spectral bands with the comparison results to construct a feature validity mapping table, iteratively update the state estimation of target domain samples with the feature validity mapping table as input until convergence, and record the historical state estimation sequence during the iterative convergence process; The feature alignment module is used to perform feature alignment based on the feature validity mapping table, using the inclusion relationship between the feature change direction sets of the source domain and the target domain of each deteriorated state interval as the basis; The feedback processing module is used to classify maintenance feedback samples according to the operation type of the maintenance work order, and to perform a local update on the feature validity mapping table based on the classification results. The local update triggers a synchronous update of the feature change direction set and feature alignment. The state prediction module is used to perform state prediction on target domain samples with synchronously updated feature alignment results. For target domain samples whose predicted values ​​fall at the boundary of adjacent deterioration state intervals, the module queries the historical state estimation sequence to identify the source of prediction uncertainty as insufficient coverage or deterioration transition period. For samples with insufficient coverage and samples with deterioration transition period, the module performs prediction state adjustment respectively.

[0010] The present invention provides a computer device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the method for predicting the state of hyperspectral data of insulators of power lines.

[0011] The present invention provides 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 predicting the state of hyperspectral data of insulators of power lines.

[0012] The beneficial effects of this invention are as follows: This invention uses the interval-by-interval and spectral band classification label of the feature validity mapping table as a constraint, and limits the feature alignment operation to the spectral band subspace corresponding to the effective migration band set. This isolates the pollution of the migration results by the distribution deviation of the hyperspectral response value of the environmental interference band, and improves the stability of the migration model in state prediction under the changing environmental conditions of the target domain.

[0013] By using historical state estimation sequences as a basis, the sources of uncertainty in the prediction of inadequate coverage and deterioration transition period are distinguished, so that the two types of boundary samples can obtain differentiated prediction state adjustments that are consistent with their physical sources, thereby reducing the false alarm rate and false negative rate of maintenance decision recommendations.

[0014] The differentiated feedback processing path based on the physical meaning of maintenance operations transforms the band-by-band response difference sequence before and after cleaning, the degradation final state label, and the time series change into targeted update inputs for the feature validity mapping table, thereby accelerating the accuracy convergence speed of the transfer model as field data accumulates. Attached Figure Description

[0015] 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.

[0016] Figure 1 This is a flowchart illustrating the overall process of a method for predicting the state of a collector line insulator using hyperspectral data transfer learning, as provided in one embodiment of the present invention.

[0017] Figure 2The present invention provides a flowchart of the feature alignment process for a hyperspectral data transfer learning state prediction method for insulators of power lines, as provided in one embodiment of the present invention.

[0018] Figure 3 The flowchart illustrates the classification of maintenance feedback samples and the execution of local updates in a hyperspectral data transfer learning state prediction method for insulators of power lines, as provided in one embodiment of the present invention. Detailed Implementation

[0019] To make the present invention more apparent and understandable, the 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, 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.

[0020] Example 1, referring to Figures 1-3 This is one embodiment of the present invention, which provides a method for predicting the state of a current collector insulator using hyperspectral data transfer learning, comprising: S1. Divide the source domain labeled samples into several deterioration state intervals according to the degree of deterioration. Establish the source domain behavior benchmark for each spectral band in each deterioration state interval. Compare the response value distribution characteristics of the target domain samples with the source domain behavior benchmark interval by interval and band by band. Use the comparison results to classify and label the spectral bands and construct a feature validity mapping table. Use the feature validity mapping table as input to iteratively update the state estimation of the target domain samples until convergence. Record the historical state estimation sequence during the iterative convergence process.

[0021] In some embodiments, the specific implementation of constructing a feature validity mapping table for spectral band classification labels in step S1 includes steps S11 to S14: S11, when the deviation between the target domain response value distribution characteristics and the source domain behavior reference is less than or equal to the dispersion parameter of the source domain behavior reference and the dispersion parameter of the target domain response value set is greater than or equal to the dispersion parameter of the source domain behavior reference, the spectral band is marked as an effective migration band.

[0022] S12, when the deviation is less than or equal to the dispersion parameter of the source domain behavior reference and the dispersion parameter of the target domain response value set is less than the dispersion parameter of the source domain behavior reference, the spectral band is marked as a concentrated distribution band.

[0023] S13. When the deviation is greater than the dispersion parameter of the source domain behavior reference and the deviation direction belongs to the known spectral response deviation direction of the pollution area type where the target line is located, the spectral band is marked as an environmental interference band.

[0024] S14. When the deviation is greater than the dispersion parameter of the source domain behavior reference and the deviation direction does not belong to the known spectral response deviation direction of the pollution area type where the target line is located, the spectral band is marked as an unknown deviation band.

[0025] It should be noted that in steps S11 to S14, the deviation refers to the absolute value of the difference between the central tendency parameter of the target domain response value set and the central tendency parameter of the source domain behavioral benchmark. The deviation direction is represented by the sign of the difference obtained by subtracting the central tendency parameter of the source domain behavioral benchmark from the central tendency parameter of the target domain response value set in each spectral band. A positive difference indicates a positive deviation, and a negative difference indicates a negative deviation. In the feature space composed of multispectral bands, the deviation direction is represented by a vector composed of the signs of the deviations in each spectral band. Steps S11 and S12 use whether the dispersion parameter of the target domain response value set is less than the dispersion parameter of the source domain behavioral benchmark as the sole distinguishing criterion. Steps S13 and S14 use whether the deviation direction belongs to the known spectral response deviation direction of the pollution area type where the target line is located as the sole distinguishing criterion. In steps S11 to S14, both the target domain response value set and the source domain behavioral benchmark are calculated based on the response values ​​after min-max normalization to ensure that they are compared in the same feature space.

[0026] It is understandable that the concentrated distribution band set is a subset of the effective migration band set; the dispersion parameter of the spectral bands in the concentrated distribution band set in the target domain is smaller than the dispersion parameter of the source domain behavioral benchmark. Using the concentrated distribution band set as the anchoring starting point for feature alignment can reduce the possibility of introducing environmental interference in the initial stage of the alignment operation.

[0027] Furthermore, in some embodiments, the method for determining whether the distributions belong to the same family in steps S11 to S14 is as follows: A goodness-of-fit test is performed on the set of response values ​​in the spectral bands of the target domain's deterioration state interval and the distribution family to which the source domain's behavioral benchmark belongs. When the test conclusion is that the set of response values ​​does not reject the distribution family to which the source domain's behavioral benchmark belongs, it is determined that the two belong to the same family. The Kolmogorov-Smirnov test is preferably used for the goodness-of-fit test, as it does not rely on the prior estimation of distribution parameters and is suitable for the actual situation where the sample size of the hyperspectral response value set of the collector line insulator is limited. The significance level of the goodness-of-fit test is predetermined based on the richness of the historical inspection data of the target line before step S1 is executed, and is determined during the execution of step S1. The significance level remains unchanged; when historical inspection data is abundant, the preferred significance level is 0.05, and when historical inspection data is limited, the preferred significance level is 0.10; whether historical inspection data is abundant is judged by whether the number of historical inspection batches of the target line is not less than 10. The insulators of the collector line are inspected once every quarter to half a year according to the operation and maintenance procedures. 10 inspection batches correspond to 2.5 to 5 years of historical data accumulation. 10 inspection batches cover all seasonal collection nodes of the collector line insulators within the complete seasonal change cycle of the pollution area. Using 10 inspection batches as the judgment criterion can ensure that the historical data covers the complete seasonal change pattern of the pollution area. The significance level setting is matched with the actual data accumulation status of the target line.

[0028] Further, in some embodiments, the specific implementation of establishing the source domain behavior benchmark for each spectral band in step S1 is as follows: Minimum-maximum normalization is performed on the spectral response values ​​of all labeled samples in the source domain. The minimum and maximum values ​​of the response values ​​of all labeled samples in the source domain at each spectral band are used as the normalization benchmark, and the response value of each spectral band is normalized to the range of 0 to 1. The normalization benchmark is determined before step S1 is executed, and the same normalization operation is performed on the corresponding spectral band response values ​​of the labeled samples in the source domain and the samples in the target domain. When the difference between the maximum and minimum values ​​of the normalization benchmark for a certain spectral band is zero, the spectral band... The minimum-maximum normalization process cannot be executed. The normalized response values ​​of the spectral bands are uniformly set to 0.5, which is the midpoint of the range from 0 to 1. The contribution of the spectral bands to the distance metric calculation for each degradation state interval depends only on the difference between the source domain behavior benchmark central tendency parameter and 0.5, and is not affected by the difference in the target domain sample response values. In the spectral band behavior comparison in steps S11 to S14, the spectral bands are directly marked as deviating from unknown bands, and no goodness-of-fit test is performed. The value of the corresponding component of the spectral band in the known spectral response deviation direction vector is 0, indicating that the spectral band does not contribute deviation direction information. When the maximum number of iterations is reached, the transition operation in step S14 is not executed. The deviation of the unknown band in the spectral band is marked unchanged in the feature validity mapping table. Subsequent steps treat the spectral band in the same way as other deviations of the unknown band. Since the normalized response value of the spectral band is uniformly 0.5, the variance contribution of the spectral band in the principal component analysis is zero. It does not appear in the dominant spectral band of any principal component direction and does not affect the direction judgment. All labeled samples in the source domain are arranged in ascending order of labeled state value. The all labeled samples in the source domain are divided into several deterioration state intervals according to the numerical range of the labeled state value. Adjacent deterioration intervals are separated into adjacent intervals. The boundary values ​​between degradation state intervals are determined by the constraint that the number of source domain samples in adjacent intervals is not less than the minimum statistical sample size. The first degradation state interval corresponds to the initial state, and the last degradation state interval corresponds to the final degradation state, denoted as the final degradation state interval. If the total number of labeled samples in the source domain is insufficient to divide the required total number of degradation state intervals while satisfying the minimum statistical sample size constraint, the total number of degradation state intervals is reduced, with the priority being to ensure that the number of source domain samples in each degradation state interval is not less than the minimum statistical sample size. The minimum statistical sample size achieves the expected test power of 0 at the set significance level before step S1 is executed.The minimum sample size required for 80 is determined and remains unchanged during step S1. For each spectral band in each degradation state interval, the normalized response value set of all source domain samples in the spectral band is extracted. The distribution characteristics of the normalized response value set are recorded as the source domain behavior benchmark for the spectral band in the degradation state interval. The source domain behavior benchmark includes three distribution characteristics: central tendency parameter, dispersion parameter, and distribution family. The central tendency parameter is the median after sorting the normalized response value set; the dispersion parameter is the interquartile range (interquartile range) between the third quartile and the first quartile of the normalized response value set; and the distribution family is determined by performing a Kolmogorov-Smirnov test on the normalized response value set.

[0029] It is understandable that the known spectral response deviation direction of the contaminated area type of the target line is predetermined and remains unchanged before step S1 is executed. The method for determining the known spectral response deviation direction is as follows: clean samples and dirty samples within the same deterioration state range are extracted from the historical inspection data of the target line. The clean samples are historical inspection samples in the inspection records where the salt density and ash density values ​​do not exceed the upper limit of the contamination level corresponding to the contamination level of the target line and no maintenance recommendation has been triggered in the current period. The dirty samples are historical inspection samples in the inspection records where cleaning is confirmed to be required and cleaning operations have been performed. The central tendency parameter of the normalized response value set of the dirty samples and the normalized response value set of the clean samples are calculated band by band. The difference in the central tendency parameter is used to construct a deviation direction vector with the sign of the difference in each spectral band. The corresponding component in the deviation direction vector for spectral bands with a difference of zero is set to 0. This deviation direction is predetermined and remains unchanged before the execution of step S1, based on the known spectral response deviation direction of the pollution area type where the target line is located. The pollutant chemical components of different pollution areas produce directional absorption characteristic shifts in the hyperspectral band. The physical properties of the deviation direction, rather than the statistical magnitude of the deviation, are used as the basis for distinguishing between steps S13 and S14, so that the classification results are directly bound to the spectroscopic characteristics of the pollutant chemical components. This solves the technical problem of existing methods relying solely on statistical distance judgment, which leads to confusion between environmental interference characteristics and material degradation characteristics.

[0030] Further, in some embodiments, the specific implementation of using the pre-trained model to generate initial state estimates for all target domain samples in step S1 is as follows: The pre-trained model takes the full-band hyperspectral response value vector of the source domain labeled samples as input. Before inputting into the pre-trained model, the response value of each spectral band is subjected to min-max normalization processing to normalize the response value of each spectral band to the range of 0 to 1, so as to eliminate the influence of the difference in the dimensions of response values ​​of different spectral bands on model training. The dimension of the normalized response value vector is equal to the total number of spectral bands. The corresponding deterioration state label value is used as the output target, and the output is a continuous value that falls within the numerical range of the corresponding deterioration state interval. The pre-trained model is preferably a support vector regression model or a random forest model. The support vector regression model preferably uses a radial basis function kernel. The regularization parameter C and the kernel function parameter γ are determined by performing 5-fold cross-validation on the source domain labeled samples. The search range of C is preferably 0.1 to 1000, and the search range of γ is preferably 0.001 to 10. The random forest model preferably contains 100 to 500 decision trees, and the maximum depth of each decision tree does not exceed 20 layers. The specific quantity and depth are determined by performing 5-fold cross-validation on the source domain labeled samples. The pre-trained model uses root mean square error as the model evaluation index, and selects the parameter combination with the smallest root mean square error in the 5-fold cross-validation as the final parameters. After the pre-trained model is trained on the source domain labeled samples, the model parameters are fixed and not updated on the target domain data. The full-band hyperspectral response value vector of each target domain sample is subjected to the same min-max normalization process as the source domain labeled samples and then input into the pre-trained model to obtain the initial state estimate of each target domain sample. Each target domain sample is assigned to the corresponding deterioration state interval according to the comparison result between its initial state estimate and the boundary value of the deterioration state interval. Target domain samples whose initial state estimate is greater than or equal to the lower boundary value of the deterioration state interval and less than the upper boundary value of the deterioration state interval are assigned to the deterioration state interval. For target domain samples whose initial state estimate is equal to the boundary values ​​of two adjacent deterioration state intervals, they are simultaneously assigned to the two deterioration state intervals on both sides of the boundary. In the spectral band behavior comparison in steps S11 to S14, the target domain samples participate in the calculation in both deterioration state intervals.

[0031] It should be noted that if no target domain sample is assigned within a certain deterioration state interval, the spectral band behavior comparison of the deterioration state interval in steps S11 to S14 is skipped, and all spectral bands of the deterioration state interval are marked as bands to be confirmed. In step S2, the deterioration state interval is uniformly processed in the same way as the processing method for missing directions of the effective migration band set. After the subsequent batch inspection data enables the deterioration state interval to obtain at least one target domain sample assignment, the spectral band behavior comparison of the deterioration state interval in steps S11 to S14 is re-executed.

[0032] Furthermore, in some embodiments, in step S14, when the cumulative number of iterations in step S1 reaches the upper limit of the number of iterations, if the spectral band still deviates from the unknown band mark, the combination of the spectral band and the deteriorated state interval is first recorded in the environmental feature queue before the incomplete convergence processing is executed, and then the environmental interference band marking processing is carried out according to step S13; after the transition operation is completed, no new iteration rounds are added, and the incomplete convergence processing is executed directly; the calculation of the two most recent iterations in the incomplete convergence processing takes the results of the last two iterations before the upper limit of the number of iterations is triggered, and the transition operation itself is not included in the number of iterations; the upper limit of the number of iterations is preferably 2 to 3 times the total number of deteriorated state intervals. The basis of this range is that each deteriorated state interval needs to undergo an average of 2 to 3 comparison updates during the iteration process to reach a stable mark. It is determined before the execution of step S1 based on the total number of deteriorated state intervals of the target line and the scale of historical inspection data, and remains unchanged during the execution of step S1.

[0033] Further, in some embodiments, the specific implementation of step S1, which iteratively updates the state estimation of the target domain sample with the feature validity mapping table as input until convergence, is as follows: In each iteration, based on the spectral bands corresponding to the effective migration band sets of each deterioration state interval in the feature validity mapping table generated in the current iteration, the absolute value of the difference between the normalized response value of each target domain sample and the central tendency parameter of the source domain behavior benchmark for each deterioration state interval is calculated for each spectral band. The sum of the absolute values ​​on the spectral bands in all effective migration band sets is used as the distance metric of the target domain sample belonging to the deterioration state interval. The midpoint value of the deterioration state interval with the smallest distance metric is used as the state estimation value of the target domain sample in this iteration. The midpoint value is half of the sum of the upper and lower boundary values ​​of the deterioration state interval. When there are multiple deterioration state intervals with equal distance metrics and all of them being the minimum value, the midpoint value of the deterioration state interval with the smallest index among the multiple deterioration state intervals is taken as the state estimation value, based on the insulator deterioration... The degradation state interval with the smallest sequence number corresponds to the slightest degree of degradation. Taking the interval with the smallest sequence number can avoid overestimation of the degree of degradation due to the distance metric being the smallest. The interval assignment of the target domain response value sample set is updated, and the next iteration proceeds to steps S11 to S14 for spectral band behavior comparison. When the feature validity mapping table generated by two consecutive iterations is completely consistent in all spectral band label types in all degradation state intervals, the iteration terminates, and the feature validity mapping table and its version number are output for use in subsequent steps. If the cumulative iteration count in step S1 reaches the upper limit of the iteration count but the convergence condition is not met, the transfer operation of the unknown band is executed first. After the transfer operation is completed, no new iteration rounds are added. The intersection of the effective migration band sets in the last two iterations before the upper limit of the iteration count is triggered is taken as the final effective migration band set, and the union of the environmental interference band sets in the last two iterations is taken as the final environmental interference band set. The feature validity mapping table is output and marked as not fully converged, and the version number is not incremented.

[0034] It should be noted that the sum of the absolute values ​​of the differences calculated for each target domain sample per spectral band during the S1 iteration is a band-by-band distance metric for a single sample, which has a different meaning from the deviation calculated for the entire set of target domain response values ​​in steps S11 to S14. The two are respectively applicable to the operations at their respective computational levels. The target domain sample response values ​​used in the S1 iteration have undergone the same min-max normalization process as the source domain labeled samples. The central tendency parameter of the source domain behavioral benchmark is also calculated based on the source domain labeled sample response values ​​after min-max normalization to ensure that the distance metric calculation and the initial state estimate are performed in the same feature space. The midpoint value of the interval ensures that the state estimate maintains a continuous numerical form during the iteration process, consistent with the initial state estimate output by the pre-trained model in terms of numerical type.

[0035] Furthermore, in some embodiments, the specific implementation of recording the historical state estimation sequence during the iterative convergence process in step S1 is as follows: when the iteration terminates, for all target domain samples in the current batch, the state estimation value at the termination of this iteration and the current inspection batch identifier are appended and recorded to the historical state estimation sequence of the target domain samples; the historical state estimation sequence is organized according to the sample identifier, each sample corresponds to one sequence, and each record in the sequence contains two items: the state estimation value and the inspection batch identifier, arranged in chronological order of the inspection batch time.

[0036] It should be noted that the sample processing status record set is organized by sample identifier, with each sample corresponding to one record. The record content is the historical number of times the sample meets the adjustment conditions in the prediction status adjustment and the corresponding inspection batch identifier. It is independent of the historical status estimation sequence and is not stored in the historical status estimation sequence. All records are cleared when the sample is removed from the waiting queue. The waiting queue is a data set of target domain samples that need to be further confirmed by subsequent inspection data for prediction status, and it is organized by sample identifier.

[0037] Through step S1, the present invention realizes the independent performance judgment of spectral band migration based on the deterioration state interval. Based on the physical properties of the known spectral response deviation direction of the target line's pollution area type, the environmental interference band and the deviation unknown band are distinguished and marked. This solves the technical problem of existing transfer learning methods relying solely on statistical distance judgment, which leads to confusion between environmental interference features and material deterioration features and accumulation of state prediction errors.

[0038] S2, based on the feature validity mapping table, performs feature alignment according to the inclusion relationship between the feature change direction sets of the source domain and the target domain of each deteriorated state interval.

[0039] In some embodiments, the specific implementation of performing feature alignment in step S2 includes steps S21 to S23: S21. For directions that exist in the source domain feature change direction set but are missing in the target domain feature change direction set, examine whether the dominant spectral band of the missing direction belongs to the environmental interference band.

[0040] S22, when the dominant spectral band belongs to the non-environmental interference band in the effective migration band set, the missing direction is recorded to the insufficient coverage direction record set, and feature alignment from the source domain to the target domain is performed on the effective migration bands in the deterioration state interval.

[0041] Further, in some embodiments, the specific implementation of performing feature alignment from the source domain to the target domain on the effective migration bands of the deteriorated state interval in step S22 is as follows: using the spectral band subspace of the deteriorated state interval as the feature space, the maximum mean difference between the sample distribution in the source domain deteriorated state interval and the sample distribution in the target domain deteriorated state interval is calculated in the feature space. Minimizing the maximum mean difference is the optimization objective. A linear transformation matrix is ​​solved to map the eigenvalue vectors corresponding to the effective migration band set of the target domain samples to the feature space aligned with the source domain sample distribution. The maximum mean difference is calculated using a radial basis function kernel. The kernel function bandwidth parameter is taken as the median of the Euclidean distance between each pair of samples in the source domain deteriorated state interval to ensure that the kernel function bandwidth matches the actual scale of the sample distribution in the deteriorated state interval. The linear transformation matrix is ​​solved by gradient descent iteration on the gradient of the maximum mean difference with respect to the transformation matrix. The initial transformation matrix is ​​the identity matrix, and the iteration step size is preferably 0.01 to 0.1. The termination condition is that the maximum mean difference between the target domain samples and the source domain samples in the transformed feature space is less than a convergence threshold, which is preferably 1×10⁻⁶. -4 The linear transformation matrix is ​​fixed after step S22 is completed and is uniformly applied to the target domain samples in the current batch, and is not updated when the target domain data arrives.

[0042] It should be noted that the maximum mean difference minimization method is a standard feature alignment technique in the field of domain adaptation. Its core idea is to achieve feature distribution alignment by minimizing the mean difference between the source domain and target domain samples in the regenerating kernel Hilbert space. This invention limits the method to be performed within the spectral band subspace corresponding to the effective migration band set, so that the feature alignment operation only acts on the spectral bands related to the intrinsic degradation characteristics of the material, eliminating the influence of environmental interference bands and deviations from unknown bands on the alignment results.

[0043] S23, when the dominant spectral band belongs to the environmental interference band, the missing direction is recorded in the verification set of the environmental feature queue, and feature alignment is not performed.

[0044] It should be noted that, in step S21, the dominant spectral band refers to the spectral band corresponding to the component with the largest absolute value in the principal component vector corresponding to the missing direction; the method for determining the dominant spectral band is as follows: extract the principal component vector corresponding to the missing direction in the spectral band subspace, calculate the absolute value of each component in the principal component vector for each spectral band, and take the spectral band corresponding to the component with the largest absolute value as the dominant spectral band of the missing direction; in step S2, the spectral band subspace refers to the feature subspace composed of spectral bands in the effective migration band set of the deteriorated state interval; the spectral bands in the environmental interference band set and the deviation unknown band set of each deteriorated state interval in the feature validity mapping table do not participate in the construction of the spectral band subspace and do not participate in any alignment operations in steps S21 to S23.

[0045] Further, in some embodiments, in step S2, when the source domain feature change direction set is the same as all independent change directions contained in the target domain feature change direction set, forward alignment is not performed on the deterioration state interval; simultaneously, it is checked whether there is a subset of source domain samples whose feature values ​​fall outside the coverage of the target domain feature change direction set within the source domain deterioration state interval; the coverage of the target domain feature change direction set refers to the super-rectangular region formed by the minimum to maximum value of the projected components of all samples in the target domain response value sample set in the feature subspace spanned by all principal component directions in the target domain feature change direction set, in each principal component direction. The range is greater than or equal to the actual distribution range of the target domain response value sample set in the principal component space, that is, the hyperrectangular region is an approximation of the outer bound of the actual distribution range of the target domain response value sample set; when the projection component value of a source domain sample in any principal component direction within the source domain degradation state interval is less than the minimum value or greater than the maximum value of the projection component of the target domain response value sample set in the principal component direction, it is determined that the source domain sample falls outside the coverage range of the target domain feature change direction set; if there is a subset of source domain samples, the participation weight of the source domain sample subset in subsequent alignment operations is reset to zero; the insufficient coverage direction record set of the degradation state interval is an empty set, and the processing of the next degradation state interval begins.

[0046] Furthermore, in some embodiments, in step S2, when there is a missing direction in the source domain feature change direction set and a target domain-specific direction exists in the target domain feature change direction set, each missing direction is processed according to steps S21 to S23, and each target domain-specific direction is processed according to the target domain-specific direction processing rules. The two types of processing are executed independently and do not interfere with each other.

[0047] Furthermore, in some embodiments, the method for establishing the source domain feature change direction set for each deterioration state interval in step S2 is as follows: The samples within the source domain deterioration state interval are confined to the spectral band subspace of the deterioration state interval. Principal component analysis is used to decompose the source domain sample distribution in the spectral band subspace. The principal components are arranged from largest to smallest eigenvalue. The set of principal component directions corresponding to when the cumulative variance contribution reaches a preset proportion of the response value data is extracted, and this set is used as the source domain feature change direction set for the deterioration state interval. The preset proportion of the response value data is based on the historical inspection data of the target line before step S2 is executed. The results of the spectral band correlation analysis are determined and remain unchanged during the execution of step S2; preferably, the value is 0.85 to 0.95. The basis for this range is that in the hyperspectral data of the insulator of the collector line, the characteristic distribution differences between adjacent deterioration state intervals are mainly reflected in the first 5 to 15 principal component directions. When the cumulative variance contribution reaches 0.85, the extracted principal component direction set has covered the main change directions that can distinguish adjacent deterioration state intervals. Setting the upper limit to 0.95 is to avoid introducing noise directions with too small variance contribution, thereby achieving a balance between retaining the main change directions and eliminating noise directions.

[0048] Further, in some embodiments, the method for establishing the target domain feature change direction set for each degradation state interval in step S2 is as follows: Project the samples from the target domain response value sample set of the degradation state interval onto the spectral band subspace of the degradation state interval; perform decomposition using principal component analysis; arrange the principal components in descending order of eigenvalues; extract the principal component direction set corresponding to when the cumulative variance contribution reaches a preset proportion of the response value data; and denote this as the response value direction set of the degradation state interval; project the samples from the time-series change sample set of the degradation state interval onto the spectral band subspace of the degradation state interval; perform decomposition using principal component analysis; and extract the principal component direction set corresponding to when the cumulative variance contribution reaches a preset proportion of the time-series change data. The set of principal component directions corresponding to the preset proportion of time-series change data is denoted as the set of change direction of the deterioration state interval; the union of the response value direction set and the change direction set is taken as the set of change direction of the target domain feature of the deterioration state interval; the preset proportion of response value data and the preset proportion of time-series change data are independent of each other and are determined according to the statistical characteristics of their respective data types, and cannot be mixed; the variance of response value data reflects the dispersion of the deterioration state distribution among insulators, while the variance of time-series change data reflects the variability of the deterioration rate. The two have fundamentally different dimensional meanings, and using the same threshold will lead to the incorrect selection or omission of the principal directions of a type of data due to differences in absolute magnitude; time-series change The preset ratio of the time-series change data is determined before step S2 based on the spectral band correlation analysis results of the historical inspection data of the target line, and remains unchanged during the execution of step S2; preferably, it is between 0.80 and 0.90. The lower limit of this range is lower than the lower limit of the preset ratio of the response value data. This is because the time-series change data reflects the change in response value between adjacent inspection cycles, and its numerical range is constrained by the deterioration rate. Under normal operation and maintenance conditions, the deterioration rate of the insulators of the collector line is relatively slow, resulting in the total variance of the time-series change data being less than the total variance of the target domain response value sample set. Under the same cumulative variance contribution threshold, the number of principal components extracted from the time-series change data will be less than that from the response value data, leading to dynamic... The direction of dynamic change was missed. The lower limit of the preset ratio of time-series change data was set to 0.80, which can ensure the complete coverage of the dynamic change direction under the condition that the sample size of the target line inspection data is limited. The time-series change sample collection stores the difference of the band-by-band response values ​​of the samples of two adjacent inspection cycles, reflecting the dynamic change direction of the target domain samples in the degradation process. The target domain response value sample collection stores the full-band hyperspectral response values ​​of the samples of each inspection cycle, reflecting the absolute distribution range of the target domain samples in the degradation state. The two types of data are extracted separately and then the union is taken, so that the set of target domain characteristic change directions simultaneously covers the static distribution characteristics and dynamic change characteristics of the degradation state.

[0049] Furthermore, in some embodiments, in step S22, when the set of concentrated distribution bands in the deterioration state interval is not empty, the spectral bands in the concentrated distribution band set are aligned before the remaining spectral bands in the effective migration band set. The alignment result of the concentrated distribution band set serves as the anchoring constraint for the alignment operation of the remaining spectral bands. The anchoring constraint is based on the fact that the spectral bands in the concentrated distribution band set satisfy the definition condition of step S12, that is, the dispersion parameter of the target domain response value set is less than the dispersion parameter of the source domain behavior benchmark. Using this as the alignment starting anchor point can ensure that the alignment operation starts from the spectral band with the highest degree of conformity to the source domain behavior to establish a reference benchmark. When the set of concentrated distribution bands in the deterioration state interval is empty, all spectral bands in the effective migration band set are aligned directly without anchoring constraints.

[0050] Furthermore, in some embodiments, when the missing direction is recorded in the set to be verified in the environmental feature queue in step S23, the confirmation status is marked as environmental source to be verified. After the subsequent maintenance feedback sample arrives, the verification will be performed. After verification and confirmation, it will be moved from the set to be verified in the environmental feature queue to the environmental feature queue for formal recording.

[0051] Further, in some embodiments, in step S2, when the target domain feature change direction set includes all directions of the source domain feature change direction set and there is at least one target domain-specific direction not appearing in the source domain feature change direction set, each target domain-specific direction is recorded in the unconfirmed direction record set of the deterioration state interval; at the same time, it is checked whether the target domain-specific direction is consistent with the deviation direction already recorded in the environmental feature queue; when the target domain-specific direction is consistent with the deviation direction already recorded in the environmental feature queue, the target domain-specific direction is recorded in the unverified set of the environmental feature queue and marked as the confirmed status as environmental source unverified; when the target domain-specific direction is inconsistent with the deviation direction already recorded in the environmental feature queue, the target domain-specific direction is marked as the material unconfirmed direction and stored in the unsupplemented direction set of the deterioration state interval. The material unconfirmed direction is retained in the unsupplemented direction set and will be confirmed and verified during subsequent maintenance feedback sample processing.

[0052] It should be noted that after step S2 is completed, the projection judgment parameters corresponding to each direction in the insufficient coverage direction record set and the unconfirmed direction record set are recalculated based on the updated target domain response value sample set. The projection judgment parameter is defined as the third quartile of the set of signed projection component values ​​of all samples in the target domain response value sample set in the direction. The signed projection component value takes a positive value when the projection direction is consistent with the direction vector of the direction and a negative value when it is opposite. The signed projection component values ​​of all samples are sorted from smallest to largest, and the value located at the 75th quartile is taken as the third quartile. The updated projection judgment parameter is recorded in the corresponding projection judgment parameter storage location and remains unchanged until the next update in step S2. After step S2 is completed, the version number of the feature validity mapping table corresponding to the completion of this update in step S2 is recorded as the latest update version number of step S2.

[0053] Through step S2, the present invention realizes the differential processing of missing directions based on the physical source difference between the intrinsic characteristics of materials and the coupling characteristics of the environment, and solves the technical problem that the existing subspace alignment method performs alignment indiscriminately on the missing directions of the target domain, resulting in the introduction of environmental interference directions into the migration results.

[0054] S3 classifies maintenance feedback samples according to the operation type of maintenance work orders, and performs a local update on the feature validity mapping table based on the classification results. The local update triggers a synchronous update of the feature change direction set and feature alignment.

[0055] In some embodiments, the specific implementation of classifying maintenance feedback samples and performing partial updates in step S3 includes steps S31 to S34: S31 classifies maintenance feedback samples into replacement type, cleaning type, and inspection-untreated type.

[0056] S32, extract the band-by-band response difference sequence before and after cleaning from the feedback sample of cleaning type, and check whether the deviation of the response value before and after cleaning of the non-environmental interference band in the effective migration band set is greater than the dispersion parameter of the source domain behavior benchmark.

[0057] In some embodiments, the processing method after checking the deviation of the response values ​​before and after cleaning in step S32 includes steps S321 to S322: S321, when the deviation of the response value before and after cleaning is less than or equal to the dispersion parameter of the source domain behavior benchmark, the sample after cleaning is stored in the target domain response value sample set of the deterioration state interval to which the sample before cleaning belongs, triggering a local update of the feature validity mapping table.

[0058] S322, when the deviation of the response value before and after cleaning is greater than the dispersion parameter of the source domain behavior benchmark, the attribution judgment is performed on the response value of the cleaned sample in the non-environmental interference band in the effective migration band set for each deterioration state interval, and the cleaned sample is stored in the target domain response value sample set of the attribution deterioration state interval, triggering a local update of the feature validity mapping table.

[0059] In some embodiments, the step S3 of performing a local update on the feature validity mapping table based on the classification results further includes steps S33 to S34: S33, for the continuous inspection sequence consisting of feedback samples of untreated types of inspections, for three consecutive periods of samples in the continuous inspection sequence where the state estimate of the second period sample belongs to a deterioration state interval number that is less than the state estimate of the first period sample and less than the state estimate of the third period sample, the three consecutive periods of samples are marked as state estimate unstable samples and excluded from the continuous inspection sequence, without triggering a local update of the feature validity mapping table.

[0060] S34. For the remaining sequence after excluding unstable samples in the state estimation, store the remaining samples into the target domain response value sample set of the corresponding deterioration state interval, triggering a local update of the feature validity mapping table.

[0061] It should be noted that if the work order operation field in step S31 does not belong to the three categories of replacement type, cleaning type, and inspection unhandled type, the feedback sample will be marked as type pending confirmation, and no local update operation will be performed. After the operation type is confirmed, it will be added to the corresponding type queue for re-execution.

[0062] Further, in some embodiments, the processing method for changing the feedback sample type in step S31 is as follows: the hyperspectral sample before replacement is stored in the spectral band behavior comparison dataset of the deteriorated final state interval, and step S1 is performed for local iteration of the deteriorated final state interval; the local iteration only performs the spectral band behavior comparison of steps S11 to S14 on the deteriorated final state interval, and updates the effective migration band set, environmental interference band set, concentrated distribution band set and deviation unknown band set corresponding to the deteriorated final state interval in the feature validity mapping table with the comparison results, without modifying the corresponding sets of other deteriorated state intervals; if the set corresponding to the deteriorated final state interval changes, the version number of the feature validity mapping table is incremented; after the local iteration is completed, a convergence test is performed, and the judgment is only made on the set change of the deteriorated final state interval; when the local iteration terminates, the state estimate value of the target domain sample in the deteriorated final state interval and the current inspection batch identifier are appended to the historical state estimate sequence of the corresponding sample.

[0063] Furthermore, in some embodiments, the change of type feedback sample in step S31 also requires a confirmation check of the material direction to be confirmed in the direction set to be supplemented: the hyperspectral sample before replacement is stored in the target domain response value sample set of the deterioration final state interval; when extracting the direction set of changes in the deterioration final state interval, the material direction to be confirmed is included as a candidate direction in the feature vector set of principal component analysis; it is checked whether the material direction to be confirmed appears in the principal component direction set extracted according to the preset proportion of time-series change data; if the material direction to be confirmed appears in the re-extracted direction set of changes in the deterioration final state interval, then it is determined that the direction is in the target domain sample set of the deterioration final state interval. This process involves moving the material's unconfirmed direction from the set of directions to be supplemented to the set of source domain feature change directions within the original degradation state interval to which the material's unconfirmed direction belonged, updating the version number of the set, and removing the direction from the set of unconfirmed direction records within the original degradation state interval to which the material's unconfirmed direction belonged. If the material's unconfirmed direction does not appear in the newly extracted set of change directions within the degradation final state interval, then the material's unconfirmed direction is moved from the set of directions to be supplemented to the set of unverified directions in the environmental feature queue, its confirmation status is marked as "environmental source unverified," and the status of the direction in the set of unconfirmed direction records within the original degradation state interval to which the material's unconfirmed direction belonged is updated to "environmental source unverified."

[0064] Furthermore, in some embodiments, the specific operation of extracting the band-by-band response difference sequence before and after cleaning in step S32 is as follows: from the data pair before and after cleaning, extract the difference in response values ​​of the sample before cleaning and the sample after cleaning in the spectral bands to obtain the band-by-band response difference sequence before and after cleaning; the data pair before and after cleaning requires that the two hyperspectral samples before and after cleaning are acquired under the same lighting conditions and acquisition angle, that is, the cleaning operation is completed on the same day of a single inspection, and the two hyperspectral samples before and after cleaning are acquired under the same weather conditions, the same inspection equipment, and the same acquisition location, so as to ensure that the band-by-band response difference sequence before and after cleaning only reflects the spectral contribution of surface dirt. The sampling process does not introduce interference caused by differences in acquisition conditions; when the above acquisition conditions cannot be met, the cleaning type feedback sample is downgraded to type pending confirmation processing; when step S32 is executed, the components of the spectral bands corresponding to the environmental interference band set in the band-by-band response difference sequence before and after cleaning are first stored in the behavior record of the corresponding entry in the environmental feature queue, which is used to update the benchmark behavior description of the environmental interference band part in the source domain behavior benchmark of the spectral band in the deterioration state interval. After the benchmark behavior of the environmental interference band is updated, the components of the non-environmental interference band in the effective migration band set in the band-by-band response difference sequence before and after cleaning are tested for the deviation of the response value before and after cleaning.

[0065] It should be noted that the non-environmental interference bands in the effective migration band set in step S32, i.e., the intrinsic bands of the material, refer to the spectral bands corresponding to the intrinsic characteristics of the material determined by the intrinsic chemical structure of the silicone rubber material. The deviation of the response values ​​before and after cleaning in step S32 refers to the absolute value of the corresponding component of the intrinsic band in the band-by-band response difference sequence before and after cleaning. This has a different meaning from the absolute value of the difference between the central tendency parameter of the target domain response value set and the central tendency parameter of the source domain behavior benchmark in steps S11 to S14. The two are respectively applicable to the judgment operation in their respective steps. Step S32, by examining the deviation of the response values ​​of the intrinsic bands of the material before and after cleaning, uses the constraint that the cleaning operation only removes surface dirt without changing the intrinsic aging state of the material to determine whether the intrinsic band response values ​​of the sample before cleaning are affected by dirt obscuring interference, thereby determining the true deterioration state of the sample after cleaning.

[0066] Further, in some embodiments, the specific method for performing attribution judgment on the response values ​​of the cleaned sample in the non-environmental interference bands of the effective migration band set in step S322, which involves checking whether the response values ​​of the cleaned sample in the non-environmental interference bands of the effective migration band set fall within the interval formed by subtracting the dispersion parameter from the central tendency parameter of the source domain behavior benchmark of the material intrinsic band corresponding to the degradation state interval, to the interval formed by adding the central tendency parameter to the dispersion parameter; when the distribution family of the source domain behavior benchmark is symmetrical, the interval can accurately cover the main distribution range of the source domain behavior benchmark; when the distribution family is asymmetrical, the interval provides a conservative attribution judgment basis, avoiding attribution omissions due to asymmetry; the degradation state interval where the response values ​​of the cleaned sample in all tested material intrinsic bands meet the conditions is determined. The candidate degradation state interval is determined. If there are multiple candidate degradation state intervals, the candidate degradation state interval with the smallest serial number is selected as the degradation state interval based on the monotonically irreversible degradation of the silicone rubber material of the insulator. If there are no candidate degradation state intervals, the degradation state interval with the smallest absolute value of the difference between the response value of the cleaned sample in each material intrinsic band and the source domain behavior benchmark central tendency parameter of each degradation state interval is selected as the degradation state interval. At the same time, the correspondence between the degradation state interval to which the sample belonged before cleaning and the newly determined degradation state interval is recorded as the pollution shading correction record of the insulator. When the degradation state interval determination of the sample of the same type of insulator before cleaning is triggered in step S322, the correspondence in the pollution shading correction record is used as the reference for the candidate degradation state interval.

[0067] Furthermore, in some embodiments, after step S32 is completed, it is also necessary to check whether there are entries in the environmental feature queue with the confirmed status of "environmental source pending verification". If so, the components of the corresponding spectral bands in the band-by-band response difference sequence before and after cleaning are compared with the direction to be verified: when there are components in the band-by-band response difference sequence before and after cleaning that are consistent with the direction to be verified, the confirmation status of the entry is updated to "environmental source confirmed", and the entry is moved from the set to be verified in the environmental feature queue to the formal record in the environmental feature queue; when there are no components in the band-by-band response difference sequence before and after cleaning that are consistent with the direction to be verified, the entry continues to be retained in the set to be verified in the environmental feature queue, the confirmation status remains "environmental source pending verification", and it waits for subsequent cleaning type feedback samples to continue verification.

[0068] Furthermore, in some embodiments, in step S33, for the continuous inspection sequence consisting of unprocessed feedback samples, when the sequence length of the continuous inspection sequence is 2, only the processing of step S34 is performed, and the identification of three consecutive phases of samples in step S33 is not performed; when the sequence length of the continuous inspection sequence is not less than 3, the identification of three consecutive phases of samples in step S33 is performed first, and then the processing of step S34 is performed; the identification criterion for three consecutive phases of samples in step S33 is: the index of the deterioration state interval to which the state estimate of the second phase sample belongs is less than the index of the deterioration state interval to which the state estimate of the first phase sample belongs and is smaller than the index of the deterioration state interval to which the state estimate of the first phase sample belongs. The state estimate of the third-period sample belongs to the deterioration state interval number, which is the same as the state estimate of the second-period sample belonging to the deterioration state interval number, and forms a local minimum between the two adjacent periods. Based on the monotonic irreversibility of the deterioration of the silicone rubber material of the insulator, the back-and-forth fluctuations reflect the instability of the prediction rather than the actual state reversal. Therefore, the samples of the three consecutive periods are marked as unstable state estimates and excluded from the continuous inspection sequence. The unstable state estimates do not perform local updates of the feature validity mapping table, do not store the target domain response value sample set, and do not store the time series change sample set. The sub-case judgment will be re-executed after the new period sample is obtained in the next inspection cycle.

[0069] Further, in some embodiments, step S34 performs the following operations on the remaining sequence after excluding unstable samples in the state estimation: for each remaining sample, its hyperspectral response value is stored in the target domain response value sample set of the deterioration state interval to which the state estimation value of the remaining sample belongs; for each pair of adjacent period samples in the remaining sequence, the difference between the response value of the next period sample and the previous period sample is extracted band by band to obtain the band-by-band response change of adjacent periods, and the band-by-band response change is stored in the time-series change sample set of the deterioration state interval to which the state estimation value of the next period sample belongs; after the above storage operation is completed, the feature validity mapping table is locally updated; when the state estimation values ​​of all samples in the remaining sequence belong to the same deterioration state interval, only the processing of step S2 is performed, and the local iteration of step S1 is not performed; when there are two adjacent period samples in the remaining sequence, the next period If the state estimate of a sample belongs to a deterioration state interval number that is strictly greater than the state estimate of the previous period's sample, and the state estimate of all remaining samples in the remaining sequence after these two periods belongs to a deterioration state interval number that is greater than or equal to the deterioration state interval number of the next period, then the data pairs of the two adjacent periods' samples are stored in the spectral band behavior comparison dataset of the two related deterioration state intervals in step S1. Step S1 is then performed for the local iteration of the two related deterioration state intervals. When the local iteration terminates, the state estimate of the target domain sample in the affected deterioration state interval and the current inspection batch identifier are appended to the historical state estimate sequence of the corresponding sample. When there is a change in the deterioration state interval number in the remaining sequence but the above monotonicity condition is not met, only the processing of step S2 is performed, and the local iteration of step S1 is not performed.

[0070] It should be noted that after the local iteration of step S1 is executed, the convergence recovery is performed in the following order after the arbitrary processing operation in step S3: After the local iteration of step S1 is completed, it is checked whether the corresponding set of the feature validity mapping table of the affected deterioration state interval has changed; if it has changed, the version number of the feature validity mapping table is incremented, and the updated contents of the effective migration band set, environmental interference band set, concentrated distribution band set, and deviation unknown band set in the feature validity mapping table of the deterioration state interval that has changed are transmitted. Step S2 re-executes steps S21 to S23 for the deterioration state interval. After step S2 is completed, the projection judgment parameters of the affected direction are updated synchronously according to the calculation rules of the projection judgment parameters in step S2, and the version number of the feature validity mapping table corresponding to the completion of this update of step S2 is recorded as the latest update version number of step S2; if the corresponding set of the feature validity mapping table has not changed, the convergence recovery is completed, and step S2 update is not executed; after step S2 update is completed, the updated source domain feature change direction set and target domain feature change direction set are recorded in the current version of the feature validity mapping table for use in the next batch.

[0071] Through step S3, this invention achieves differentiated processing of different types of feedback samples based on the physical meaning of maintenance operations. It corrects the sample assignment error caused by dirt occlusion by utilizing the physical constraint that cleaning operations only remove surface dirt without changing the intrinsic aging state of the material. It updates the feature validity mapping table of the degradation final state interval by utilizing the characteristic of the replacement operation to confirm the degradation final state. It identifies and excludes unstable samples of state estimation with back-and-forth jumping by utilizing the monotonic irreversibility of insulator degradation. This solves the technical problem of existing methods processing maintenance feedback data indiscriminately, which leads to environmental interference information being mixed into the migration model.

[0072] S4. Perform state prediction on the target domain samples using the synchronously updated feature alignment results. For target domain samples whose predicted values ​​fall at the boundary of adjacent deterioration state intervals, query the historical state estimation sequence to identify the source of prediction uncertainty as insufficient coverage or deterioration transition period. Perform prediction state adjustment for insufficient coverage samples and deterioration transition period samples respectively.

[0073] In some embodiments, the specific implementation of identifying the source of prediction uncertainty and performing prediction state adjustment in step S4 includes steps S41 to S45: S41, check whether the feature value of the target domain sample falls in the coverage-deficient direction of the record set. When the feature value of the target domain sample falls in any coverage-deficient direction of the record set, the target domain sample is identified as coverage-deficient.

[0074] S42, when the feature value of the target domain sample does not fall in any of the insufficient coverage directions in the insufficient coverage direction record set, query the historical state estimation sequence, record the boundary between the two adjacent deterioration state intervals where the state prediction value of the target domain sample is located as the current boundary, and record the deterioration state interval that is adjacent to the current boundary and whose sequence number is greater than the sequence number of another adjacent deterioration state interval as the first adjacent deterioration state interval. When the state estimation values ​​of two consecutive historical inspection cycles both fall within the first adjacent deterioration state interval and are not marked as unstable state estimation samples in two consecutive historical inspection cycles, the target domain sample is identified as a deterioration transition period type.

[0075] It should be noted that in step S42, the current boundary refers to the boundary between two adjacent deterioration state intervals where the state prediction value of the target domain sample is located; the two deterioration state intervals adjacent to the current boundary are respectively denoted as the first adjacent deterioration state interval and the second adjacent deterioration state interval, wherein the sequence number of the first adjacent deterioration state interval is greater than the sequence number of the second adjacent deterioration state interval; the basis for two consecutive historical inspection cycles in step S42 is that the gradual nature of insulator deterioration makes it possible that a single state estimate biased towards a more severe interval may come from random prediction fluctuations, and two consecutive biases towards the same side are necessary to initially distinguish between random fluctuations and systematic deterioration trends, and two are the minimum number of historical records required to satisfy the distinction purpose.

[0076] S43, target domain samples that do not meet the identification conditions for insufficient coverage and deterioration transition period are identified as trend pending confirmation type.

[0077] S44, for samples in the deterioration transition period, the predicted state is adjusted to the first adjacent deterioration state interval. When the adjusted predicted state falls within the maintenance suggestion trigger range, a maintenance suggestion is generated.

[0078] S45: For samples with insufficient coverage, calculate the proportion of samples with insufficient coverage in the current batch to the total number of samples in the current batch. When the proportion is less than the threshold for the proportion of samples with insufficient coverage boundary, query the historical state estimation sequence. When the state estimation values ​​of two consecutive historical inspection cycles both fall within the first adjacent deterioration state interval and the samples have not been marked as unstable state estimation samples in two consecutive historical inspection cycles, adjust the predicted state to the first adjacent deterioration state interval. When the proportion is greater than or equal to the threshold for the proportion of samples with insufficient coverage boundary, delay until the feature validity mapping table is updated and then re-execute source identification. When the feature validity mapping table is marked as not fully converged, all samples with insufficient coverage in the current batch are uniformly processed as if the proportion is greater than or equal to the threshold for the proportion of samples with insufficient coverage boundary, and delay until the feature validity mapping table is updated and then re-execute source identification.

[0079] It should be noted that the method for identifying target domain samples whose predicted values ​​fall at the boundaries of adjacent deterioration state intervals in step S4 is as follows: when the absolute value of the difference between the predicted state value of the target domain sample and the boundary value between two adjacent deterioration state intervals is less than the interval width of each of the two adjacent deterioration state intervals, the predicted value of the target domain sample is determined to fall at the boundary of the adjacent deterioration state interval, and the target domain sample is identified as a sample near the boundary; the interval width refers to the difference between the upper and lower boundary values ​​of the deterioration state interval; target domain samples that do not meet the conditions are directly output with the original predicted state and do not proceed to the subsequent processing in steps S41 to S45; the threshold for the proportion of samples with insufficient coverage of the boundary is determined before the execution of step S4 based on the target line. Historical inspection data is determined as follows: the proportion of samples near the boundary of each inspection batch in the historical inspection data of the target line to the total number of samples in the current batch is statistically analyzed batch by batch to obtain a proportion series. The median of the proportion series is taken as the threshold for the proportion of samples with insufficient coverage. When the number of historical inspection batches of the target line is insufficient to perform median statistics, the threshold for the proportion of samples with insufficient coverage is taken as a preset initial value. The preset initial value is determined based on the historical statistical data of the same type of line before step S4 is executed, and is replaced by the actual statistical median after accumulating sufficient batch data. The threshold for the proportion of samples with insufficient coverage remains unchanged during the execution of step S4. The maintenance suggestion trigger range is predetermined before step S4 is executed and remains unchanged.

[0080] It should be noted that when the number of historical inspection batches of the target line is no less than 10, the preset initial value is replaced by the actual statistical median. The number of 10 inspection batches is consistent with the batch number standard used in the significance level judgment in step S1, so as to ensure that the parameter update triggering condition and the basis for judging the richness of data are consistent.

[0081] Further, in some embodiments, the specific method for checking whether the feature value of the target domain sample falls on any insufficient coverage direction in the insufficient coverage direction record set in step S41 is as follows: project the feature value vector corresponding to the effective migration band set of the target domain sample onto the insufficient coverage direction, calculate the signed projection component value, wherein the signed projection component value takes a positive value when the projection direction is consistent with the direction vector of the insufficient coverage direction, and takes a negative value when they are opposite; when the signed projection component value is greater than the projection judgment parameter corresponding to the insufficient coverage direction, it is determined that the feature value of the target domain sample falls on the insufficient coverage direction; the projection judgment parameter is calculated and recorded according to the calculation rules of the projection judgment parameter in step S2 after being updated in step S2, and remains unchanged before the next update in step S2; the verification is performed one by one on all directions recorded in the insufficient coverage direction record set and the direction record set to be confirmed, and the target domain sample is identified as insufficient coverage type when any direction meets the condition.

[0082] Furthermore, in some embodiments, the specific method for adjusting the predicted state of the deterioration transition type sample in step S44 is as follows: the predicted state of the deterioration transition type sample is adjusted to the first adjacent deterioration state interval; when the adjusted predicted state is within the maintenance suggestion trigger range, a maintenance suggestion is directly generated and not stored in the pending queue; when the adjusted predicted state is not within the maintenance suggestion trigger range, the adjusted predicted state is directly output and not stored in the pending queue.

[0083] Further, in some embodiments, when performing predictive state adjustment on under-coverage samples in step S45, the following two scenarios are handled based on whether there are records in the sample processing state record set of under-coverage samples that previously met the predictive state adjustment conditions: If there are no records in the sample processing state record set that previously met the predictive state adjustment conditions, and the adjusted predictive state does not fall within the maintenance suggestion trigger range, it is directly output; if the adjusted predictive state falls within the maintenance suggestion trigger range, it is stored in the pending queue, and the mark indicating that the conditions were met and the current inspection batch identifier are recorded in the sample processing state record set; if there are records in the sample processing state record set that previously met the predictive state adjustment conditions, it is directly output; if the adjusted predictive state falls within the maintenance suggestion trigger range, it is stored in the pending queue, and the mark indicating that the conditions were met and the current inspection batch identifier are recorded in the sample processing state record set. If there is a record that previously met the predicted state adjustment conditions, if the adjusted predicted state is not within the range of maintenance suggestion triggering, it will be output directly. If the adjusted predicted state is within the range of maintenance suggestion triggering, it will be determined based on the monotonic irreversibility of insulator deterioration that two consecutive trend confirmations are sufficient to support the generation of maintenance suggestions, and maintenance suggestions will be generated directly. In step S4, the sample processing state record set is organized according to the sample identifier. Each sample corresponds to one record. The record content is the historical number of times the sample met the predicted state adjustment conditions and the corresponding inspection batch identifier. It is independent of the historical state estimation sequence and is not stored in the historical state estimation sequence. All records are cleared when the sample is removed from the waiting queue.

[0084] Furthermore, in some embodiments, the insufficiently covered samples stored in the processing queue in step S45 are processed in the same way as unprocessed feedback samples after a new batch of data is obtained in subsequent inspection cycles. The source identification and judgment in steps S41 to S45 are re-executed each time new data arrives, until one of the following four termination conditions is met: First termination scenario: When the degradation transition period identification condition is met after re-executing steps S41 to S45 and the adjusted predicted state falls within the maintenance suggestion trigger range, the sample is removed from the pending queue and a maintenance suggestion is generated. Second termination scenario: If, after re-executing steps S41 to S45, there is a record in the sample processing status record set that previously met the prediction status adjustment conditions and the adjusted prediction status falls within the range of maintenance suggestion triggering, the sample is removed from the pending queue and a maintenance suggestion is generated. The third termination scenario: If, after the source identification and judgment in steps S41 to S45 are completed, the final predicted state in step S4 of two consecutive inspection cycles falls within the same deterioration state range, and the deterioration state range is not within the scope of maintenance suggestion triggering, and the sample is not marked as an unstable sample in state estimation within two consecutive inspection cycles, the sample is removed from the pending queue, and all records of the sample in the sample processing state record set are cleared. No maintenance suggestion is generated, and normal inspection cycle processing resumes. The final predicted state in step S4 refers to the predicted state value after the source identification of the sample in steps S41 to S45 and the prediction state adjustment when necessary. If the prediction state adjustment is not performed in steps S41 to S45, the final predicted state in step S4 is equal to the original predicted state. Fourth termination scenario: When the cumulative retention time of a sample in the pending queue exceeds the longest observation period, the sample is transferred to manual review and removed from the pending queue. At the same time, all records of the sample in the sample processing status record set are cleared. The longest observation period is determined by the target line operation and maintenance procedures, and is predetermined and kept unchanged before the execution of step S4. It is preferably the time interval between two adjacent periodic maintenance of the target line. When the manual review conclusion is that maintenance is required, the historical hyperspectral data corresponding to the sample and the manual review confirmation record are stored in the corresponding type queue for processing according to the feedback type corresponding to the specific maintenance operation type determined by the manual review. When the manual review conclusion is that the state is stable and maintenance is not required, the hyperspectral response value of the sample in the most recent inspection is directly stored in the target domain response value sample set of the deterioration state interval to which the sample state estimate belongs. Step S2 is triggered to re-execute steps S21 to S23 for the deterioration state interval. After the new data is obtained in the next inspection cycle, the sequence composed of the expected data and the first inspection data after manual review will re-enter the inspection untreated type feedback sample processing process.

[0085] Through step S4, this invention distinguishes between inadequate coverage prediction uncertainty and deterioration transition period prediction uncertainty based on historical state estimation sequences. For deterioration transition period samples, it directly performs prediction state adjustment and generates maintenance suggestions based on the monotonic irreversibility of insulator deterioration. For inadequate coverage samples, it generates maintenance suggestions after two consecutive trend confirmations by constraining the historical number of sample processing state record sets. This solves the technical problem of existing methods failing to distinguish the sources of prediction uncertainty for samples near the boundary, leading to the coexistence of missed and false alarms.

[0086] Furthermore, in some embodiments, after step S4 is completed, a cross-step version consistency verification is also required: This involves checking whether the version number of the feature validity mapping table used during the execution of steps S41 to S45 is consistent with the most recently updated version number of step S2; if the version numbers are consistent, the current batch processing is complete; if the version numbers are inconsistent, it is determined that the processing operation of step S3 triggered a local iteration of step S1 during the execution of steps S41 to S45, causing the version number of the feature validity mapping table to increment, while the corresponding local update of step S2 has not yet been completed, and the most recently updated version of step S2 is used during the execution of steps S41 to S45. The information in the under-coverage direction record set corresponding to this number is outdated. Under-coverage samples in the current batch that have completed prediction status adjustments are marked as samples to be reviewed. After the convergence and recovery process in step S3 is completed, the local update in step S2 is completed, and the latest update version number in step S2 is updated to match the version number of the feature validity mapping table, the verification in step S41 is re-executed for all samples to be reviewed using the latest version of the under-coverage direction record set and the direction record set to be confirmed in step S2. After re-executing step S41, if the sample to be reviewed no longer falls in either the under-coverage direction record set or the direction record set to be confirmed, the verification process continues. Step S42 identifies the source of degradation transition. Processing is done according to the results of steps S42 to S44. If step S44 generates a repair suggestion, the sample is simultaneously removed from the pending processing queue, and all records in the sample processing status record set are cleared. If step S42 determines the trend to be confirmed, the sample remains in the pending processing queue. After re-executing step S41, if the sample to be reviewed still falls in either the insufficient coverage direction record set or the direction to be confirmed record set, and the processing results of re-executing steps S41 to S45 are the same as the original processing results, then the original processing results are maintained. Step S44 is then re-executed. After step S41, if the sample to be reviewed still falls in either the insufficient coverage direction record set or the direction to be confirmed record set, and there is a record in the sample processing status record set that previously met the prediction status adjustment conditions and the adjusted prediction status falls within the maintenance suggestion trigger range, then the sample is removed from the pending queue and a maintenance suggestion is generated. After re-executing step S41, if the sample to be reviewed still falls in either the insufficient coverage direction record set or the direction to be confirmed record set, and the new processing result is a trend to be confirmed, then the status of the sample in the pending queue remains unchanged, and the first condition-met mark in the sample processing status record set is cleared.

[0087] Example 2 is an embodiment of the present invention, which provides a hyperspectral data transfer learning state prediction system for insulators of power lines, comprising: The mapping table construction module is used to divide the source domain labeled samples into several deterioration state intervals according to the degree of deterioration. For each deterioration state interval, a source domain behavior benchmark for spectral bands is established. The response value distribution characteristics of the target domain samples are compared with the source domain behavior benchmarks interval by interval and band by band. The comparison results are used to classify and label the spectral bands to construct a feature validity mapping table. The feature validity mapping table is used as input to iteratively update the state estimation of the target domain samples until convergence. During the iterative convergence process, the historical state estimation sequence is recorded. The feature alignment module is used to perform feature alignment based on the feature validity mapping table, using the inclusion relationship between the feature change direction sets of the source domain and the target domain of each deteriorated state interval as the basis; The feedback processing module is used to classify maintenance feedback samples according to the operation type of the maintenance work order, and to perform a local update on the feature validity mapping table based on the classification results. The local update triggers a synchronous update of the feature change direction set and feature alignment. The state prediction module is used to perform state prediction on target domain samples with synchronously updated feature alignment results. For target domain samples whose predicted values ​​fall at the boundary of adjacent deterioration state intervals, the module queries the historical state estimation sequence to identify the source of prediction uncertainty as insufficient coverage or deterioration transition period. For samples with insufficient coverage and samples with deterioration transition period, the module performs prediction state adjustment respectively.

[0088] This embodiment also provides an electronic device applicable to a method for predicting the state of hyperspectral data transfer learning of insulators for collector lines, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the method for predicting the state of hyperspectral data transfer learning of insulators for collector lines as proposed in the above embodiment.

[0089] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements a method for predicting the state of hyperspectral data of insulators of power lines as proposed in the above embodiments.

[0090] The storage medium proposed in this embodiment belongs to the same inventive concept as the method for predicting the state of hyperspectral data transfer learning of insulators for collector lines proposed in the above embodiments. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0091] Based on the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of the various embodiments of the present invention.

[0092] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not 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 predicting the state of insulators of a current collector line using hyperspectral data transfer learning, characterized in that, include: The source domain labeled samples are divided into several deterioration state intervals according to the degree of deterioration. A source domain behavior benchmark for each spectral band is established for each deterioration state interval. The response value distribution characteristics of the target domain samples are compared with the source domain behavior benchmark interval by interval and band by band. The comparison results are used to classify and label the spectral bands to construct a feature validity mapping table. The state estimation of the target domain samples is iteratively updated with the feature validity mapping table as input until convergence. The historical state estimation sequence is recorded during the iterative convergence process. The step of constructing a feature validity mapping table by classifying and labeling spectral bands based on the comparison results includes: when the deviation between the target domain response value distribution characteristics and the source domain behavioral benchmark is less than or equal to the dispersion parameter of the source domain behavioral benchmark and the dispersion parameter of the target domain response value set is greater than or equal to the dispersion parameter of the source domain behavioral benchmark, the spectral band is labeled as a valid migration band; when the deviation is less than or equal to the dispersion parameter of the source domain behavioral benchmark and the dispersion parameter of the target domain response value set is less than the dispersion parameter of the source domain behavioral benchmark, the spectral band is labeled as a concentrated distribution band; when the deviation is greater than the dispersion parameter of the source domain behavioral benchmark and the deviation direction belongs to the known spectral response deviation direction of the pollution area type where the target line is located, the spectral band is labeled as an environmental interference band; when the deviation is greater than the dispersion parameter of the source domain behavioral benchmark and the deviation direction does not belong to the known spectral response deviation direction of the pollution area type where the target line is located, the spectral band is labeled as a deviation unknown band. Based on the feature validity mapping table, feature alignment is performed according to the inclusion relationship between the feature change direction sets of the source domain and the target domain for each deterioration state interval; The feature alignment based on the inclusion relationship between the feature change direction sets of the source and target domains in each deterioration state interval includes: for directions present in the source domain feature change direction set but missing in the target domain feature change direction set, checking whether the dominant spectral band of the missing direction belongs to an environmental interference band; when the dominant spectral band belongs to a non-environmental interference band in the effective migration band set, recording the missing direction to the insufficient coverage direction record set, and performing feature alignment from the source domain to the target domain for the effective migration bands of the deterioration state interval; when the dominant spectral band belongs to an environmental interference band, recording the missing direction to the unverified set of the environmental feature queue, and not performing feature alignment. The maintenance feedback samples are classified according to the operation type of the maintenance work order. The classification results are used to perform a local update on the feature validity mapping table. The local update triggers a synchronous update of the feature change direction set and feature alignment. The classification of maintenance feedback samples based on the operation type of the maintenance work order includes classifying maintenance feedback samples into replacement type, cleaning type, and unhandled inspection type. The process of locally updating the feature validity mapping table based on the classification results includes: extracting the band-by-band response difference sequence before and after cleaning from the cleaning type feedback samples, and checking whether the deviation of the response values ​​before and after cleaning in the non-environmental interference bands in the effective migration band set is greater than the dispersion parameter of the source domain behavior benchmark; when the deviation is less than or equal to the dispersion parameter of the source domain behavior benchmark, storing the cleaned samples in the target domain response value sample set of the deterioration state interval to which the samples belonged before cleaning, triggering a local update of the feature validity mapping table; when the deviation is greater than the dispersion parameter of the source domain behavior benchmark, performing a classification judgment on the response values ​​of the cleaned samples in the non-environmental interference bands in the effective migration band set for each deterioration state interval, and storing the cleaned samples in the deterioration state interval. The target domain response value sample set of the interval triggers a local update of the feature validity mapping table; for a continuous inspection sequence consisting of feedback samples of untreated types, for three consecutive periods of samples in the continuous inspection sequence where the state estimate of the second period sample belongs to a deterioration state interval number that is less than the state estimate of the first period sample and less than the state estimate of the third period sample, the three consecutive periods of samples are marked as state estimation unstable samples and excluded from the continuous inspection sequence without triggering a local update of the feature validity mapping table; for the remaining sequence after excluding state estimation unstable samples, the remaining samples are stored in the target domain response value sample set of the corresponding deterioration state interval, triggering a local update of the feature validity mapping table; The target domain samples are predicted using the synchronized updated feature alignment results. For target domain samples whose predicted values ​​fall at the boundary of adjacent deterioration state intervals, the historical state estimation sequence is queried to identify the source of prediction uncertainty as insufficient coverage or deterioration transition period. Predicted state adjustments are performed for insufficient coverage samples and deterioration transition period samples respectively. The identification of uncertainty sources as either insufficient coverage or degradation transition period includes: checking whether the feature value of the target domain sample falls on any insufficient coverage direction recorded in the insufficient coverage direction record set; when the feature value of the target domain sample falls on any insufficient coverage direction in the insufficient coverage direction record set, the target domain sample is identified as insufficient coverage; when the feature value of the target domain sample does not fall on any insufficient coverage direction in the insufficient coverage direction record set, the historical state estimation sequence is queried, and the boundary between two adjacent degradation state intervals where the state prediction value of the target domain sample is located is recorded as the current boundary; the degradation state interval adjacent to the current boundary and whose sequence number is greater than the sequence number of another adjacent degradation state interval is recorded as the first adjacent degradation state interval; when the state estimation values ​​of two consecutive historical inspection cycles both fall within the first adjacent degradation state interval and the sample is not marked as an unstable state estimation sample in the two consecutive historical inspection cycles, the target domain sample is identified as degradation transition period; target domain samples that do not meet the insufficient coverage identification condition and do not meet the degradation transition period identification condition are identified as trend pending confirmation type.

2. The method for predicting the state of a current collector insulator using hyperspectral data transfer learning as described in claim 1, characterized in that, The process of performing prediction state adjustments for under-covered samples and samples in the deterioration transition phase includes: For samples in the deterioration transition period, the predicted state is adjusted to the first adjacent deterioration state interval. When the adjusted predicted state falls within the maintenance suggestion trigger range, a maintenance suggestion is generated. For samples with insufficient coverage, the proportion of samples with insufficient coverage in the current batch to the total number of samples in the current batch is calculated. When the proportion is less than the threshold for the proportion of samples with insufficient coverage boundary, the historical state estimation sequence is queried. When the state estimation values ​​of two consecutive historical inspection cycles both fall within the first adjacent deterioration state interval and the samples are not marked as unstable state estimation samples in the two consecutive historical inspection cycles, the predicted state is adjusted to the first adjacent deterioration state interval. When the proportion is greater than or equal to the threshold for the proportion of samples with insufficient coverage boundary, the source identification is re-executed after the feature validity mapping table is updated.

3. A hyperspectral data transfer learning state prediction system for collector line insulators, employing the hyperspectral data transfer learning state prediction method for collector line insulators as described in any one of claims 1 or 2, characterized in that, include: The mapping table construction module is used to divide the source domain labeled samples into several deterioration state intervals according to the degree of deterioration. For each deterioration state interval, a source domain behavior benchmark for spectral bands is established. The response value distribution characteristics of the target domain samples are compared with the source domain behavior benchmarks interval by interval and band by band. The comparison results are used to classify and label the spectral bands to construct a feature validity mapping table. The feature validity mapping table is used as input to iteratively update the state estimation of the target domain samples until convergence. During the iterative convergence process, the historical state estimation sequence is recorded. The feature alignment module is used to perform feature alignment based on the feature validity mapping table, using the inclusion relationship between the feature change direction sets of the source domain and the target domain of each deteriorated state interval as the basis; The feedback processing module is used to classify maintenance feedback samples according to the operation type of the maintenance work order, and to perform a local update on the feature validity mapping table based on the classification results. The local update triggers a synchronous update of the feature change direction set and feature alignment. The state prediction module is used to perform state prediction on target domain samples with synchronously updated feature alignment results. For target domain samples whose predicted values ​​fall at the boundary of adjacent deterioration state intervals, the module queries the historical state estimation sequence to identify the source of prediction uncertainty as insufficient coverage or deterioration transition period. For samples with insufficient coverage and samples with deterioration transition period, the module performs prediction state adjustment respectively.

4. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the hyperspectral data transfer learning state prediction method for collector line insulators according to any one of claims 1 or 2.

5. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the hyperspectral data transfer learning state prediction method for collector line insulators as described in claim 1 or 2.

Images