Method and apparatus for determining a pollution level of an insulating surface

Abstract

The application relates to the field of power equipment state monitoring, and in particular to a method and device for determining the contamination level of an insulating surface. The method comprises: collecting spectral data, event stream data and environmental data of a target insulating surface; performing alignment and fusion processing on the spectral data and the event stream data to determine multi-modal features; wherein the multi-modal features represent spectral feature components and event stream feature components; performing dynamic weight distribution on the multi-modal features according to the environmental data to obtain weighted fusion features; inputting the weighted fusion features into a classification model to obtain a classification result; wherein the classification result represents the contamination level of the target insulating surface. The method is used to improve the accuracy and stability of the determination of the contamination level of the insulating surface.

Description

Technical Field

[0001] This application relates to the field of power equipment condition monitoring, and in particular to a method and apparatus for determining the pollution level of an insulating surface. Background Technology

[0002] In power systems, pollution flashover poses a significant threat to safe operation. With increasing emissions of industrial dust and smoke, this pollution can adhere to the surface of insulators. In humid weather conditions such as fog or drizzle, a conductive water film forms on the insulator surface, creating a conductive path that reduces the flashover voltage and leads to pollution flashover. Insulator pollution level assessment typically involves on-site sampling and laboratory testing, or non-contact online detection methods such as infrared, ultraviolet, and hyperspectral imaging to determine the degree of surface pollution and subsequently implement appropriate anti-flashover maintenance measures.

[0003] However, the above-mentioned methods have long laboratory testing cycles, which are difficult to meet the needs of online real-time evaluation. In non-contact methods, some means are difficult to characterize the composition of pollutants and are easily affected by environmental factors. They are also prone to insufficient accuracy in distinguishing adjacent pollution levels, which limits the stability of determining the pollution level of the insulation surface.

[0004] Therefore, improving the accuracy and stability of determining the contamination level of insulation surfaces under complex environmental conditions has become an urgent technical problem to be solved. Summary of the Invention

[0005] The method and apparatus for determining the contamination level of an insulating surface provided in this application can improve the accuracy and stability of determining the contamination level of an insulating surface.

[0006] In a first aspect, embodiments of this application provide a method for determining the contamination level of an insulating surface, including:

[0007] Spectral data, event flow data, and environmental data of the target insulating surface are collected. Among them, spectral data characterizes the spectral radiation characteristics of plasma during the discharge process, event flow data characterizes the dynamic changes of plasma optical signals in the spatiotemporal dimension, and environmental data characterizes the environmental conditions of the target insulating surface.

[0008] The spectral data and event stream data are aligned and fused to determine multimodal features; the multimodal features represent the spectral feature components and the event stream feature components.

[0009] Based on environmental data, dynamic weight allocation is performed on multimodal features to obtain weighted fusion features;

[0010] The weighted fusion features are input into the classification model to obtain the classification results; the classification results characterize the pollution level of the target insulating surface.

[0011] Optionally, as described above, the spectral data and event stream data are aligned and fused to determine multimodal features, including:

[0012] The spectral data is subjected to a first feature extraction process to obtain spectral features, and the event stream data is subjected to a second feature extraction process to obtain event stream features.

[0013] By mapping spectral features and event flow features to the same preset dimensional space, corresponding spectral alignment features and event flow alignment features are obtained.

[0014] The spectral alignment features and event flow alignment features are fused to obtain multimodal features.

[0015] Optionally, as described above, the spectral data undergoes a first feature extraction process to obtain spectral features, including:

[0016] The spectral data is preprocessed to obtain preprocessed data; wherein, the preprocessing includes at least one of baseline correction processing, noise reduction processing, and normalization processing;

[0017] Based on a preset dimensionality reduction algorithm, the preprocessed data is dimensionality reduced to obtain spectral features; wherein the preset dimensionality reduction algorithm includes at least one of linear dimensionality reduction algorithm and nonlinear dimensionality reduction algorithm; the linear dimensionality reduction algorithm includes at least one of principal component analysis (PCA) algorithm and independent component analysis (ICA) algorithm; the nonlinear dimensionality reduction algorithm includes the locally linear embedding (LLE) algorithm.

[0018] Optionally, as described above, the event stream data is event stream information based on plasma optical signals acquired by a dynamic visual sensor; the event stream data undergoes a second feature extraction process to obtain event stream features, including:

[0019] The event stream data is divided into a first image set and a second image set; wherein, the first image in the first image set represents the spatial distribution of the brightness enhancement region in the plasma optical signal within the corresponding time window, and the second image in the second image set represents the spatial distribution of the brightness reduction region in the plasma optical signal within the corresponding time window;

[0020] Each image frame in the first image set and the second image set is filtered to obtain a filtered image sequence.

[0021] Geometric morphological features and event statistical features of plasma are extracted from the filtered image sequence to obtain event flow features; among them, geometric morphological features include shape features and size features, and event statistical features include event quantity features.

[0022] Optionally, as described above, the spectral features and event flow features are mapped to the same preset dimensional space to obtain corresponding spectral alignment features and event flow alignment features, including:

[0023] Based on the physical constraint rules of plasma radiation, the physical correlation between spectral features and event flow features is determined; among them, the physical constraint rules characterize the physical correspondence between plasma spectral intensity attenuation and plasma spatial scale contraction.

[0024] Based on the physical correlation, the spectral features and event flow features are matched and aligned under physical constraints to obtain physically aligned spectral features and event flow features;

[0025] Physically aligned spectral features and physically aligned event flow features are mapped to the same preset dimensional space to obtain spectral aligned features and event flow aligned features.

[0026] Optionally, as described above, based on environmental data, dynamic weights are assigned to the multimodal features to obtain weighted fusion features, including:

[0027] Multimodal features are input into the attention mechanism network to generate initial dynamic weight coefficients. The attention mechanism network is used to perform cross-attention calculation based on spectral feature components and event flow feature components as query vectors and key vectors to obtain the initial dynamic weight coefficients. The initial dynamic weight coefficients include initial spectral feature weights and initial event flow feature weights.

[0028] Based on environmental data, correction factors are determined; whereby the correction factors characterize the degree of influence of environmental disturbances on spectral characteristic components and event flow characteristic components, respectively.

[0029] Based on the correction factor, the allocation ratio of each weight in the initial dynamic weight coefficient is adjusted to obtain the target dynamic weight coefficient.

[0030] Based on the target dynamic weight coefficient, the multimodal features are weighted and fused to obtain the weighted fused features.

[0031] Optionally, as described above, the environmental data includes illumination parameters and laser parameters, where the illumination parameters characterize the ambient background light intensity and the laser parameters characterize the fluctuation amplitude of the laser output energy used to generate the plasma; based on the environmental data, a correction factor is determined, including:

[0032] Based on the illumination parameters, the first signal quality parameter corresponding to the event stream feature component is determined; wherein, the first signal quality parameter is inversely correlated with the illumination parameters;

[0033] Based on the laser parameters, the second signal quality parameter corresponding to the spectral characteristic component is determined; wherein, the second signal quality parameter is inversely correlated with the laser parameters;

[0034] The first signal quality parameter and the second signal quality parameter are determined as correction factors.

[0035] Optionally, as described above, the classification model includes multiple fully connected layers, batch normalization layers, and activation function layers; weighted fusion features are input into the classification model to obtain classification results, including:

[0036] The weighted fusion features are input into multiple fully connected layers, and then pass through the feature transformation of each fully connected layer, the normalization of the corresponding batch normalization layer, and the nonlinear activation of the corresponding activation function layer. Finally, the classification score corresponding to the number of pollution levels is output through the last fully connected layer.

[0037] The classification scores are converted into a probability distribution, and the category with the highest probability value is determined as the classification result.

[0038] Optionally, the method described above further includes:

[0039] Obtain the training dataset; the training dataset includes multimodal data of multiple contamination levels of insulating surfaces collected under different environmental conditions. The multimodal data includes sample spectral data, sample event stream data, and sample environmental data.

[0040] Alignment and fusion processing is performed on sample spectral data and sample event stream data to determine the multimodal features of the samples;

[0041] Based on the sample environment data, dynamic weight allocation is performed on the multimodal features of the samples to obtain the sample weighted fusion features;

[0042] Based on the weighted fusion features of the samples and their corresponding pollution level labels, the initial model is iteratively trained to obtain a classification model.

[0043] Secondly, embodiments of this application provide a device for determining the contamination level of an insulating surface, comprising:

[0044] The data acquisition module is used to collect spectral data, event stream data, and environmental data of the target insulating surface. Among them, the spectral data characterizes the spectral radiation characteristics of the plasma during the discharge process, the event stream data characterizes the dynamic changes of the plasma optical signal in the spatiotemporal dimension, and the environmental data characterizes the environmental conditions of the target insulating surface.

[0045] The alignment and fusion module is used to align and fuse spectral data and event stream data to determine multimodal features; where multimodal features represent spectral feature components and event stream feature components.

[0046] The weight allocation module is used to dynamically allocate weights to multimodal features based on environmental data to obtain weighted fusion features;

[0047] The classification module is used to input weighted fusion features into the classification model to obtain classification results; the classification results characterize the pollution level of the target insulation surface.

[0048] Thirdly, embodiments of this application provide a device for determining the contamination level of an insulating surface, including: a memory and a processor;

[0049] The memory stores the instructions that the computer executes;

[0050] The processor executes computer execution instructions stored in memory, causing the processor to perform the first aspect and / or various possible implementations of the first aspect as described above.

[0051] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the first aspect and / or various possible implementations of the first aspect.

[0052] Fifthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the first aspect and / or various possible implementations of the first aspect.

[0053] The method and apparatus for determining the pollution level of an insulating surface provided in this application collect spectral data, event stream data, and environmental data of the target insulating surface. The spectral data and event stream data are then aligned and fused to determine multimodal features. Further, based on the environmental data, dynamic weights are assigned to the multimodal features to obtain weighted fused features. These weighted fused features are then input into a classification model to obtain a classification result. Spectral data characterizes the spectral radiation characteristics of plasma during discharge, event stream data characterizes the dynamic changes of plasma optical signals in the spatiotemporal dimension, environmental data characterizes the environmental conditions of the target insulating surface, multimodal features characterize spectral feature components and event stream feature components, and the classification result characterizes the pollution level of the target insulating surface. The method for determining the pollution level of an insulating surface in this application enriches the feature dimensions of pollution level assessment by collecting both spectral data and event stream data. By introducing environmental data for dynamic weight allocation of multimodal features, the model's adaptability to environmental disturbances is enhanced. Furthermore, the alignment and fusion of spectral data and event stream data improves the semantic consistency between different modal features. The method described in this application improves the accuracy and stability of determining the contamination level of insulating surfaces. Attached Figure Description

[0054] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0055] Figure 1 A flowchart illustrating a method for determining the pollution level of an insulating surface provided in this application. Figure 1 ;

[0056] Figure 2 A flowchart illustrating a method for determining the pollution level of an insulating surface provided in this application. Figure 2 ;

[0057] Figure 3 A flowchart illustrating a model training method provided in this application;

[0058] Figure 4 A schematic diagram of the dimensionality reduction result provided in this application;

[0059] Figure 5a A schematic diagram of a weight distribution provided in this application Figure 1 ;

[0060] Figure 5b A schematic diagram of a weight distribution provided in this application Figure 2 ;

[0061] Figure 5c A schematic diagram of a weight distribution provided in this application Figure 3 ;

[0062] Figure 5d A schematic diagram of a weight distribution provided in this application Figure 4 ;

[0063] Figure 6 A schematic diagram of the architecture of a classification model based on a convolutional neural network provided in this application;

[0064] Figure 7 This application provides a schematic diagram illustrating the effect of a different model.

[0065] Figure 8 A matrix diagram illustrating the classification effect provided in this application;

[0066] Figure 9 A schematic diagram of a system for determining the contamination level of an insulating surface provided in this application;

[0067] Figure 10 A schematic diagram of a device for determining the contamination level of an insulating surface provided in this application;

[0068] Figure 11 This is a schematic diagram of a device for determining the contamination level of an insulating surface, provided in an embodiment of this application.

[0069] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0070] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0071] In power systems, pollution flashover poses a significant threat to safe operation. With increasing emissions of industrial dust and smoke, this pollution can adhere to the surface of insulators. In humid weather conditions such as fog or drizzle, a conductive water film forms on the insulator surface, creating a conductive path that reduces the flashover voltage and leads to pollution flashover. Insulator pollution level assessment typically involves on-site sampling and laboratory testing, or non-contact online detection methods such as infrared, ultraviolet, and hyperspectral imaging to determine the degree of surface pollution and subsequently implement appropriate anti-flashover maintenance measures.

[0072] However, the above-mentioned methods have long laboratory testing cycles, which are difficult to meet the needs of online real-time evaluation. In non-contact methods, some means are difficult to characterize the composition of pollutants and are easily affected by environmental factors. They are also prone to insufficient accuracy in distinguishing adjacent pollution levels, which limits the stability of determining the pollution level of the insulation surface.

[0073] Specifically, when using laser-induced breakdown spectroscopy (LIBS) to detect contamination on insulating surfaces, although LIBS can distinguish contamination levels by analyzing the differences in the types and contents of elements in the contamination layer, the LIBS spectral signal is easily affected by factors such as laser energy fluctuations, plasma evolution states, and changes in ambient light, resulting in insufficient classification accuracy and robustness of a single spectral mode under complex working conditions.

[0074] Therefore, improving the accuracy and stability of determining the contamination level of insulation surfaces under complex environmental conditions has become an urgent technical problem to be solved.

[0075] The method and apparatus for determining the pollution level of insulating surfaces disclosed in this application enrich the feature dimensions of pollution level assessment by acquiring two modal information: spectral data and event stream data. By introducing environmental data to dynamically assign weights to multimodal features, the model's adaptability to environmental disturbances is enhanced. Furthermore, by aligning and fusing spectral data and event stream data, the semantic consistency between different modal features is improved. The method of this application improves the accuracy and stability of determining the pollution level of insulating surfaces.

[0076] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0077] Figure 1 A flowchart illustrating a method for determining the pollution level of an insulating surface provided in this application. Figure 1 ,like Figure 1 As shown, the method may include:

[0078] S101. Collect spectral data, event flow data, and environmental data of the target insulating surface; among which, spectral data characterizes the spectral radiation characteristics of plasma during the discharge process, event flow data characterizes the dynamic changes of plasma optical signals in the spatiotemporal dimension, and environmental data characterizes the environmental conditions of the target insulating surface.

[0079] S102. Align and fuse the spectral data and event stream data to determine multimodal features; wherein, the multimodal features characterize the spectral feature components and the event stream feature components;

[0080] S103. Based on environmental data, perform dynamic weight allocation on multimodal features to obtain weighted fusion features;

[0081] S104. Input the weighted fusion features into the classification model to obtain the classification result; wherein, the classification result characterizes the pollution level of the target insulating surface.

[0082] In step S101, the target insulation surface is the exposed surface of the outer insulation to be evaluated for the level of contamination.

[0083] Spectral data, as input data characterizing the spectral radiation characteristics of plasma during the discharge process, can reflect the distribution of spectral lines, continuous spectra, or spectral bands during the pollution discharge process.

[0084] Event stream data, as input data characterizing the dynamic changes of plasma optical signals in the spatiotemporal dimension, can reflect the emergence, expansion, migration, and decay processes of the discharge luminescence region;

[0085] Environmental data is used to characterize the environmental conditions of the target insulating surface.

[0086] In practice, the aforementioned spectral data, event flow data, and environmental data can be acquired using appropriate acquisition devices and mapped to the same monitoring process on the same target insulating surface. If necessary, the acquired data can also undergo routine processing to serve as input for subsequent processing.

[0087] It can be seen that this step acquires spectral data, event flow data, and environmental data through the same monitoring process around the same target insulating surface, so that subsequent processing no longer depends on a single detection source.

[0088] In step S102, multimodal features can refer to intermediate feature representations used to characterize the fusion result of spectral feature components and event stream feature components. The core of this approach lies in corresponding and jointly expressing data with different sampling formats and dimensional structures. Since spectral data and event stream data differ in data structure and sampling rhythm, they need to be aligned before fusion.

[0089] In practical implementation, the entity executing this method can align and fuse spectral data and event stream data based on the data correspondence formed during the acquisition process to determine the multimodal features corresponding to the current monitoring period. These multimodal features can simultaneously reflect spectral radiative characteristics and spatiotemporal dynamic change characteristics, and serve as input for subsequent dynamic weight allocation.

[0090] As can be seen, this step establishes a correspondence between spectral data and event stream data, and performs alignment and fusion processing on the two, thereby enabling the plasma radiation composition information and discharge dynamic behavior information to be jointly represented in the same feature expression. For adjacent level samples with similar levels of contamination, relying solely on single detection information is prone to insufficient differentiation. However, through the above alignment and fusion processing, multimodal features can simultaneously preserve spectral line differences and discharge morphology differences, providing an operable feature basis for subsequent environmental adaptive weighting.

[0091] In step S103, the environmental data serves as a quantitative description of the environmental conditions of the target insulating surface and is used to adjust the feature contribution ratio of the multimodal features under the current monitoring conditions.

[0092] Weighted fusion features are output features after dynamic weight allocation. They serve as direct input to the classification model and reflect a more reliable combination of multimodal information under current environmental disturbances.

[0093] Since environmental changes in outdoor scenarios can affect the ability of different information to indicate the level of pollution, it is necessary to dynamically assign weights to multimodal features based on environmental data.

[0094] In practice, the implementer of this method determines weight information appropriate to the current monitoring conditions based on environmental data, and adjusts the multimodal features based on the weight information to obtain weighted fusion features. Compared with unweighted multimodal features, the weighted features can better reflect stable and effective pollution characterization information under the current scenario.

[0095] As can be seen, this step, by introducing environmental data and adjusting the weights of multimodal features accordingly, enables the feature representation of the same target insulating surface under different environmental conditions to have environmental adaptability. This processing step redistributes the reliability within the multimodal features based on environmental parameters, thereby reducing the impact of environmental disturbances on the consistency of subsequent pollution level judgments.

[0096] In step S104, the classification model is used to classify and map the weighted fusion features, and output the contamination level of the target insulating surface.

[0097] The classification results can be represented by discrete level labels, such as multiple level intervals corresponding to the pollution classification standards of the power industry.

[0098] The classification model is the discriminative execution unit of this method. Its input is the weighted fusion feature obtained in the previous step, and its output is the pollution level corresponding result of the target insulating surface.

[0099] In practice, the classification model can be deployed on a local processor or a backend processing device. The processor inputs the weighted fused features into the classification model, which then outputs the corresponding classification result. The output classification result can be further sent to the operation and maintenance platform for generating alarms, scheduling cleaning, or assessing the operational risks of insulators.

[0100] The classification results characterize the pollution level of the target insulation surface, which is essentially a judgment of the pollution state of the insulation surface by using the multimodal features after environmental adaptation.

[0101] Because the input features already incorporate the spectral radiation characteristics, spatiotemporal dynamics, and environmental conditions of plasma discharge, the classification model can establish clearer decision boundaries for adjacent pollution levels in the feature space. Thus, even under complex outdoor monitoring conditions, it can still output a pollution level result corresponding to the current state of the target insulation surface.

[0102] This embodiment provides a method for determining the pollution level of an insulating surface. By incorporating spectral data, event stream data, and environmental data into the same monitoring link and establishing an alignment and fusion relationship between spectral radiation information and spatiotemporal dynamic information, the contribution ratio of the fused features is dynamically corrected using environmental data. Finally, a weighted fusion feature adapted to the current operating conditions is formed and the pollution level classification is completed. This enhances the separability of insulating surfaces with similar pollution levels in the classification model, while maintaining high consistency of the classification results with environmental disturbances. This can improve the accuracy and stability of determining the pollution level of an insulating surface.

[0103] Figure 2 A flowchart illustrating a method for determining the pollution level of an insulating surface provided in this application. Figure 2 , Figure 2 Implementation examples in Figure 1 Based on the embodiments, the method for determining the contamination level of the insulating surface is described in detail, such as... Figure 2 As shown, the method may include:

[0104] S201. Collect spectral data, event flow data, and environmental data of the target insulating surface; among which, spectral data characterizes the spectral radiation characteristics of plasma during the discharge process, event flow data characterizes the dynamic changes of plasma optical signals in the spatiotemporal dimension, and environmental data characterizes the environmental conditions of the target insulating surface.

[0105] S202. Perform the first feature extraction process on the spectral data to obtain the spectral features.

[0106] Among them, spectral features can refer to the feature representations extracted from the original spectral data that can effectively characterize the composition and content information of pollutant elements and have discriminative properties, which are used for subsequent alignment and fusion with event stream features.

[0107] In an optional example, step S202 may include:

[0108] S2021. Preprocess the spectral data to obtain preprocessed data; wherein, the preprocessing includes at least one of baseline correction processing, noise reduction processing, and normalization processing.

[0109] Baseline correction processing can refer to the use of methods such as polynomial fitting and morphological filtering to eliminate baseline drift in the spectral signal caused by factors such as background radiation and detector dark current, so that the spectral signal returns to the zero baseline level and highlights the true spectral peak information.

[0110] Denoising can refer to methods such as smoothing filtering, wavelet transform threshold denoising, or noise removal based on statistical models to reduce the random noise component in the spectral signal and improve the signal-to-noise ratio and detection limit of the spectral peak.

[0111] Normalization processing can refer to scaling the intensity values ​​of spectral data according to preset rules. For example, it can include total intensity normalization (i.e., dividing the spectral intensity by the total spectral intensity), maximum intensity normalization (i.e., dividing the spectral intensity by the maximum intensity of the spectral peak), or standard normal variable transformation, etc., to eliminate the overall scaling difference of spectral intensity caused by factors such as laser energy fluctuations and changes in detection distance, and enhance the comparability of spectral data.

[0112] In practice, one or more of the above preprocessing methods can be selected and combined according to the quality of the spectral data and the actual application requirements.

[0113] S2022. Based on a preset dimensionality reduction algorithm, the preprocessed data is subjected to dimensionality reduction processing to obtain spectral features; wherein, the preset dimensionality reduction algorithm includes at least one of linear dimensionality reduction algorithm and nonlinear dimensionality reduction algorithm; the linear dimensionality reduction algorithm includes at least one of principal component analysis (PCA) algorithm and independent component analysis (ICA) algorithm; the nonlinear dimensionality reduction algorithm includes the locally linear embedding (LLE) algorithm.

[0114] Among them, the preset dimensionality reduction algorithm can refer to the mathematical transformation method that maps high-dimensional spectral data to a low-dimensional feature space while retaining the most discriminative information in the data, in order to reduce the redundant dimensions of spectral data and reduce computational complexity.

[0115] Linear dimensionality reduction algorithms refer to dimensionality reduction methods that project high-dimensional data into a low-dimensional subspace through matrix operations based on the assumption of linear transformation. They preserve the global structural relationship of the data and are suitable for scenarios where the data distribution is approximately linear.

[0116] Nonlinear dimensionality reduction algorithms refer to dimensionality reduction methods that embed high-dimensional data into a low-dimensional manifold space based on the assumption of nonlinear mapping and through methods such as manifold learning. They can reveal the local geometric structure of the data and are suitable for scenarios where the data distribution has a complex nonlinear structure.

[0117] PCA (Principal Component Analysis) algorithm refers to an unsupervised linear dimensionality reduction method that calculates the covariance matrix of the data and performs eigenvalue decomposition, selects the top principal components with the largest variance contribution rate as new feature bases, and projects the original high-dimensional data into a low-dimensional subspace composed of these principal components, thereby achieving data dimensionality reduction while retaining the maximum variance information.

[0118] The ICA (Independent Component Analysis) algorithm refers to the process of decomposing observed multivariate data into a linear combination of several independent source signals by finding statistically independent non-Gaussian components, in order to separate the independent feature components hidden in the mixed data. It is suitable for blind source separation scenarios where the source signals are mutually independent.

[0119] LLE (Locally Linear Embedding) algorithm refers to a nonlinear dimensionality reduction method that assumes that the data have a linear relationship in the local neighborhood and maps high-dimensional data to a low-dimensional space while preserving the local geometric structure by keeping the linear reconstruction weight of each sample in its neighborhood unchanged.

[0120] In practice, a suitable dimensionality reduction algorithm can be selected based on the distribution characteristics of the spectral data and the requirements of the classification task. For example, when the main variation directions of the spectral data are concentrated and exhibit an approximately linear distribution, the PCA algorithm can be chosen; when the spectral data is composed of a mixture of multiple independent source signals, the ICA algorithm can be chosen; and when the spectral data has a complex nonlinear manifold structure, the LLE algorithm can be chosen.

[0121] It is understandable that by performing dimensionality reduction processing on high-dimensional spectral data to compress the original spectral vector into a low-dimensional feature representation, redundant dimensions and noise interference in the spectral data can be reduced, the computational complexity of the classification model can be reduced, and the classification accuracy of subsequent pollution levels can be improved.

[0122] S203. Perform a second feature extraction process on the event stream data to obtain event stream features.

[0123] Event flow features refer to the feature representations extracted from plasma optical signal event flow data acquired by dynamic vision sensors, which can effectively characterize the spatiotemporal evolution characteristics and dynamic behavior patterns of plasma. For example, event flow features characterize the geometric features reflecting the spatial distribution morphology of plasma and the quantitative features reflecting the dynamic changes in plasma luminescence intensity, which are used for subsequent alignment and fusion with spectral features.

[0124] In an optional example, step S203 may include:

[0125] S2031. Divide the event stream data into a first image set and a second image set; wherein, the first image in the first image set represents the spatial distribution of the brightness enhancement region in the plasma optical signal within the corresponding time window, and the second image in the second image set represents the spatial distribution of the brightness reduction region in the plasma optical signal within the corresponding time window.

[0126] Among them, event stream data can refer to the asynchronous event sequence output by the Dynamic Vision Sensor (DVS) during plasma discharge. Each event contains a timestamp, pixel coordinates and event polarity information. It is triggered when the light intensity change detected by the pixel exceeds a preset threshold. Positive polarity events (representing On events) indicate light intensity enhancement, and negative polarity events (representing Off events) indicate light intensity reduction.

[0127] The first image set can refer to the image sequence obtained by accumulating and restoring the brightness enhancement event stream (On event stream) according to a preset time window using the event frame method. Each image frame represents the spatial distribution of the plasma brightness enhancement area within the corresponding time window, and the pixel value reflects the cumulative number or density of On events in that area.

[0128] The second image set can refer to the image sequence obtained by accumulating and restoring the brightness reduction event stream (Off event stream) according to a preset time window using the event frame method. Each image frame represents the spatial distribution of the plasma brightness reduction area within the corresponding time window, and the pixel value reflects the cumulative number or density of Off events in that area.

[0129] In practice, the time window length of the event frame method can be dynamically adjusted according to the evolution speed of the plasma. For example, a shorter time window (e.g., 1 to 5 milliseconds) can be used for rapidly evolving plasmas to capture fine dynamic processes, while a longer time window (e.g., 10 to 20 milliseconds) can be used for slowly evolving plasmas to improve the signal-to-noise ratio. Overlapping regions can be set between time windows to ensure temporal continuity.

[0130] S2032. Filter each image frame in the first image set and the second image set to obtain the filtered image sequence.

[0131] Among them, filtering can refer to suppressing and smoothing random noise and isolated outliers in the event frame reconstructed image (i.e., a two-dimensional pixel grid image obtained by accumulating and mapping asynchronous event streams through the event frame method) in order to improve image quality, enhance the contrast between plasma regions and background, and at the same time preserve plasma edge and detail information as much as possible.

[0132] In one possible implementation, filtering can be achieved using methods such as Kalman filtering, particle filtering, or Gaussian filtering. For example, Kalman filtering can be applied to each image frame in the first and second image sets, using the estimated plasma region state (including centroid position, region area, and contour shape) from the previous image frame as prior information. This is combined with the observations from the current image frame to predict and update the state, smoothing random fluctuations during event accumulation and suppressing sudden noise, resulting in a filtered image frame. Alternatively, particle filtering can be applied to each image frame in the first and second image sets, propagating and resampling a set of random samples (particles) representing possible plasma region states on the current image frame. Weights are adjusted based on the matching degree between each particle and the observation, approximating the posterior probability distribution and estimating the optimal state, resulting in a filtered image frame. Alternatively, Gaussian filtering can be applied to each image frame in the first and second image sets, convolving a two-dimensional Gaussian kernel function with the current image frame to smooth random noise using a weighted average of neighboring pixel values, resulting in a filtered image frame.

[0133] It should be understood that the above three filtering methods can be used individually or in combination. The selection of filtering methods should be determined comprehensively based on plasma evolution characteristics, noise type, and real-time requirements. No restrictions are placed on the specific selection of filtering methods and parameter configurations here.

[0134] S2033. Extract the geometric morphological features and event statistical features of the plasma from the filtered image sequence to obtain event flow features; wherein, the geometric morphological features include shape features and size features, and the event statistical features include event quantity features.

[0135] Among them, geometric morphological features can refer to the set of features extracted from the plasma region of each image frame in the filtered image sequence, reflecting the spatial distribution morphology and geometric structure of the plasma, which can include shape features and size features.

[0136] Shape features can refer to characteristic parameters that describe the geometric shape of the plasma region's outline. These can include, for example, the curvature of the outline, the number of corner points, the proportion of straight line segments, symmetry, and the direction of the principal axis, and are used to distinguish the differences in plasma shape under different levels of contamination.

[0137] Size characteristics can refer to characteristic parameters that describe the size of the plasma region, such as region area, equivalent diameter, length and width of the circumscribed rectangle, and perimeter of the outline, which are used to reflect the variation of plasma spatial scale under different pollution levels.

[0138] Event statistical features refer to a set of numerical features calculated statistically from raw event stream data or event frame images, reflecting the intensity of plasma dynamic evolution and the distribution patterns of events. These features may include, for example, the total number of events, event density, event triggering frequency, and event spatial distribution entropy. Specifically, event statistical features include event quantity features, which refer to the total number of events counted within a preset time window or spatial region, including the number of On events, the number of Off events, and the ratio or difference between the two, used to quantify the severity of plasma brightness changes and the intensity of dynamic evolution.

[0139] In practical implementation, geometric features can be automatically extracted using image processing algorithms. For example, the Canny or Sobel operator can be used to detect plasma contours, and morphological operations (i.e., operations based on set theory to process the shape of the image, such as filling holes inside the contour, smoothing boundary burrs, and connecting broken areas) can be used to fill the region and calculate its area and perimeter. Event statistical features can be directly obtained from the pixel values ​​of event stream data or event frame images. For image sequences, features can be extracted from each frame and then averaged or maximized over time to obtain a stable feature representation.

[0140] It should be understood that the purpose of extracting geometric morphological features and event statistical features is to comprehensively characterize the physical properties of plasma from different dimensions: geometric morphological features reflect the spatial distribution pattern of plasma and are related to the physical properties of the contamination layer (e.g., thickness, density, and composition distribution); event statistical features reflect the dynamic evolution intensity of plasma and are related to the intensity of the interaction between laser and contamination and the energy coupling efficiency. These two types of features complement and verify each other, jointly constituting the discriminative feature representation of the event flow mode. No specific algorithms or parameter combinations for feature extraction are limited here; those skilled in the art can select appropriate image processing algorithms and statistical methods according to actual application requirements.

[0141] It is understandable that by performing polarity separation, event frame restoration, and filtering on the event stream data, the asynchronous and sparse event stream can be converted into a synchronous and dense image sequence and noise interference can be removed. This can preserve the spatiotemporal evolution information of plasma, thereby achieving effective characterization of event stream features and improving the accuracy of subsequent pollution level classification.

[0142] S204. Map the spectral features and event flow features to the same preset dimension space to obtain the corresponding spectral alignment features and event flow alignment features.

[0143] The same preset dimension space can refer to a unified feature dimension determined according to the data format requirements of subsequent fusion processing (such as one-dimensional feature format or two-dimensional feature format), so that the spectral features and event stream features are consistent in the number of dimensions, which facilitates splicing, weighting or fusion operations on the feature dimensions.

[0144] For example, spectral alignment features and event flow alignment features are projected onto the same preset dimensional space through independent linear mapping layers (learnable parameter matrices). For instance, spectral features are projected to 512 dimensions through a first linear mapping layer, and event flow features are projected to 512 dimensions through a second linear mapping layer, ensuring that the two are fused while maintaining dimensional consistency.

[0145] In an optional implementation, step S204 may include:

[0146] S2041. Based on the physical constraint rules of plasma radiation, determine the physical correlation between spectral features and event flow features; wherein, the physical constraint rules characterize the physical correspondence between plasma spectral intensity attenuation and plasma spatial scale contraction.

[0147] Among them, physical constraint rules can refer to mathematical or logical constraints established based on the principles of plasma physics, describing the intrinsic relationship between the macroscopic radiation characteristics and the microscopic spatiotemporal evolution characteristics of plasma. Specifically, during the cooling process, the spectral radiation intensity of plasma decays exponentially with time, while the spatial scale of plasma (e.g., the area of ​​the luminescent region, equivalent diameter) shrinks with time. Both are related to the decay process of plasma temperature and electron density. Therefore, there is a physical correspondence between the rate of spectral intensity decay and the rate of spatial scale shrinkage.

[0148] The physical correlation between spectral features and event flow features can refer to the establishment of a corresponding mapping relationship between spectral features (characterizing plasma spectral radiation characteristics) and event flow features (characterizing plasma spatiotemporal evolution characteristics) at the same physical evolution time using the above physical constraint rules, so that the two heterogeneous modal features can characterize the same physical state.

[0149] Specifically, physical correlations can be determined by: identifying the plasma generation time (starting point of the spectral signal), the peak time of the spectral intensity, and the time when the spectral signal decays to background noise in the spectral features; identifying the starting time of plasma brightness enhancement, the peak time of brightness, and the ending time of brightness reduction in the event flow features; matching and aligning the above key time points to establish a physical correspondence between the spectral features and the event flow features in the time dimension.

[0150] S2042. Based on the physical correlation, match and align the spectral features and event flow features under physical constraints to obtain physically aligned spectral features and event flow features.

[0151] Among them, the matching alignment under physical constraints can refer to aligning the spectral features and event flow features in the same physical evolution stage under the physical association constraints established in step S2041, so that the aligned spectral features and event flow features characterize the physical state of the plasma at the same moment, thereby enhancing the semantic consistency between the two modal features.

[0152] Specifically, the feature sequences in the spectral features corresponding to the entire process from plasma generation to cooling can be aligned with the feature sequences in the event flow features corresponding to the same physical evolution stage on the time axis. For two modal features with different time resolutions, interpolation or resampling methods can be used to unify the time resolution, or a segmented correspondence can be established by anchoring key time points.

[0153] It should be understood that the core of matching and alignment under physical constraints lies in using the laws of plasma physical evolution to establish semantic relationships between two modal features, ensuring that the fused multimodal features have physical interpretability and discriminative consistency.

[0154] S2043. Map the physically aligned spectral features and physically aligned event flow features to the same preset dimension space to obtain spectral aligned features and event flow aligned features.

[0155] In this context, mapping to the same preset dimension space can refer to projecting the physically aligned spectral features and event flow features onto a feature space of the same dimension through an independent linear mapping layer (which can be a learnable parameter matrix that is pre-trained or randomly initialized and optimized through end-to-end training).

[0156] Specifically, spectral features are projected onto a preset dimension through a first linear mapping layer, and event flow features are projected onto the same preset dimension through a second linear mapping layer. The selection of the preset dimension is determined according to the requirements of subsequent fusion processing. For example, when using a one-dimensional feature format for fusion, the preset dimension can be 256 dimensions; when using a two-dimensional feature format for fusion, the preset dimension can be 512 dimensions.

[0157] It is understandable that by establishing a physical correlation between spectral features and event flow features based on physical constraint rules of plasma radiation, and using the physical correspondence between plasma spectral intensity attenuation and spatial scale contraction as constraints for matching and alignment, it is possible to ensure that the two heterogeneous modal features represent the plasma state at the same physical evolution stage. Then, by linear mapping, the physically aligned features are projected onto the same preset dimensional space to achieve semantic consistency and dimensional consistency between modes, thereby improving the accuracy and robustness of subsequent contamination level determination.

[0158] S205. The spectral alignment features and event flow alignment features are fused to obtain multimodal features.

[0159] Among them, fusion processing can refer to combining the aligned spectral features and event flow features in the feature dimension to form a unified feature representation containing information from both modalities.

[0160] Specifically, the fusion processing method can be determined based on the data format requirements of subsequent classification models. For example, when the subsequent classification model uses a one-dimensional feature format, the spectral aligned features and event flow aligned features are concatenated or weighted along the feature dimension to obtain a one-dimensional multimodal feature vector. When the subsequent classification model uses a two-dimensional feature format, the spectral aligned features are rearranged into a two-dimensional matrix and concatenated with the event flow aligned features along the channel dimension or feature map level to obtain a two-dimensional multimodal feature matrix. For example, multimodal features are implemented using both one-dimensional vectors (e.g., 512-dimensional) and two-dimensional features (e.g., a 512×1 matrix). The input layer of the classification model selects the appropriate processing method based on the feature format. For example, if a one-dimensional vector is used, the input layer directly receives 512-dimensional features; if two-dimensional features are used, the input layer processes them through channel concatenation (e.g., 512 channels × 1 pixel) to ensure compatibility with subsequent network layers.

[0161] It should be understood that the core of fusion processing lies in enabling multimodal features to simultaneously contain information on the chemical composition and physical state of the plasma, thereby enhancing the features' ability to distinguish pollution levels and their representational completeness. One-dimensional feature formats are suitable for classification models such as support vector machines, K-nearest neighbors, random forests, or one-dimensional convolutional neural networks; two-dimensional feature formats are suitable for two-dimensional convolutional neural networks, capable of capturing spatial local correlations using convolution operations. The specific fusion processing method (e.g., concatenation, weighted fusion, channel concatenation, or feature map concatenation) can be flexibly selected based on the input requirements of the classification model and the needs of practical applications. No restrictions are placed on the specific choice of fusion method or parameter configuration here.

[0162] It is understandable that by fusing spectral alignment features and event flow alignment features to obtain multimodal features, the fused features can simultaneously contain information on the chemical composition and physical state of the plasma, thereby enhancing the ability of the fused features to distinguish the level of contamination and improving the accuracy and reliability of the contamination level determination.

[0163] S206. Input the multimodal features into the attention mechanism network to generate initial dynamic weight coefficients. The attention mechanism network is used to perform cross-attention calculations based on the spectral feature components and event flow feature components as query vectors and key vectors to obtain the initial dynamic weight coefficients. The initial dynamic weight coefficients include initial spectral feature weights and initial event flow feature weights.

[0164] The attention mechanism network can refer to a neural network module built based on the multi-head self-attention mechanism or cross-attention mechanism in the Transformer architecture. It is used to calculate the correlation strength between spectral feature components and event flow feature components, and to generate the initial weight assignment of each modal component based on the correlation strength.

[0165] In practice, the attention mechanism network calculates the attention score of spectral features on event flow features by performing a dot product operation with spectral feature components as query vectors and event flow feature components as key vectors; at the same time, it calculates the attention score of event flow features on spectral features by performing a dot product operation with event flow feature components as query vectors and spectral feature components as key vectors; the attention scores from the two directions are then fused and normalized to output the initial dynamic weight coefficients.

[0166] The initial dynamic weight coefficients include initial spectral feature weights and initial event flow feature weights. These weights can be represented as scalar values, probability distributions, or normalized weight vectors. Specifically, the initial spectral feature weights are numerical measures of the relative importance of spectral feature components in multimodal fusion, with a value ranging from [0,1]. A larger value indicates a higher contribution of the spectral feature component to the classification task. Similarly, the initial event flow feature weights are numerical measures of the relative importance of event flow feature components in multimodal fusion, also ranging from [0,1]. A larger value indicates a higher contribution of the event flow feature component to the classification task.

[0167] For example, when the cross-attention score between the spectral feature components and the event flow feature components shows that the two are highly correlated and the spectral features have stronger discriminative power, the attention mechanism network can generate initial dynamic weight coefficients with an initial spectral feature weight of 0.6 and an initial event flow feature weight of 0.4; conversely, when the event flow features have stronger discriminative power, it can generate initial dynamic weight coefficients with an initial spectral feature weight of 0.4 and an initial event flow feature weight of 0.6.

[0168] It should be understood that the initial dynamic weight coefficients generated by the attention mechanism network are determined only based on the content relevance of the multimodal features themselves, without considering the impact of external environmental factors on the quality of the modal signals. Therefore, it is necessary to introduce environmental data in subsequent steps to correct the weights in order to achieve environmentally adaptive dynamic weight allocation.

[0169] S207. Based on environmental data, determine the correction factor; whereby the correction factor characterizes the degree of influence of environmental disturbances on spectral characteristic components and event flow characteristic components, respectively.

[0170] Among them, the correction factor can refer to a set of numerical parameters that quantify the impact of environmental disturbances on the reliability of each modal component based on the physical correlation between environmental parameters and modal signal quality. It is used to adjust the initial dynamic weight coefficients generated by the attention mechanism network so that the final weights reflect the actual situation under environmental disturbances, thereby improving the classification accuracy of subsequent pollution levels.

[0171] In an optional implementation, the environmental data includes illumination parameters and laser parameters, whereby the illumination parameters characterize the ambient background light intensity and the laser parameters characterize the amplitude of laser output energy fluctuations used to generate plasma; step S207 may include:

[0172] S2071. Determine the first signal quality parameter corresponding to the event stream feature component based on the illumination parameters; wherein the first signal quality parameter is inversely correlated with the illumination parameters.

[0173] Among them, illumination parameters can refer to physical quantities that characterize the intensity of background light in the detection scene, obtained by ambient light sensors or dynamic vision sensors, and their unit is lux.

[0174] The first signal quality parameter can be a numerical measure of the reliability and effectiveness of the event stream characteristic components under the current illumination conditions.

[0175] In one possible implementation, the deviation of the current illumination parameters from the optimal operating illumination range of the dynamic vision sensor can be calculated (e.g., it can be represented as a normalized deviation index). The reciprocal of the normalized deviation index or a negative mapping can be used as the first signal quality parameter, where the greater the deviation of the illumination parameters, the lower the first signal quality parameter.

[0176] S2072. Based on the laser parameters, determine the second signal quality parameter corresponding to the spectral characteristic component; wherein the second signal quality parameter is inversely correlated with the laser parameters.

[0177] The laser parameters characterize the output energy fluctuation characteristics of the laser pulse used to generate plasma through laser-induced breakdown spectroscopy. For example, the laser parameters may include the average value, standard deviation, coefficient of variation, or variance of the laser energy.

[0178] The second signal quality parameter can be a numerical measure of the reliability and stability of spectral feature components under the current laser energy conditions, and is used to assess the degree of influence of laser energy fluctuations on spectral signal quality.

[0179] In one possible implementation, the coefficient of variation of the laser pulse energy within a preset time window (e.g., the ratio of the standard deviation to the average value of the laser energy) can be calculated. The reciprocal or negative of the coefficient of variation can then be used as the second signal quality parameter. The greater the fluctuation in laser energy (the larger the coefficient of variation), the lower the second signal quality parameter. For example, when the coefficient of variation of the laser energy is 0.1, the second signal quality parameter is 0.9; when the coefficient of variation is 0.3, the second signal quality parameter is 0.7.

[0180] S2073. Determine the first signal quality parameter and the second signal quality parameter as correction factors.

[0181] For example, the correction factor can be represented as a two-dimensional vector containing the first signal quality parameter and the second signal quality parameter, or as a diagonal matrix with the first signal quality parameter and the second signal quality parameter as diagonal elements. The specific representation of the correction factor can be flexibly selected according to the calculation requirements of subsequent weight correction.

[0182] It is understandable that by determining the first signal quality parameter of the event flow characteristic component based on the ambient light intensity (which is inversely correlated with the light intensity), and the second signal quality parameter of the spectral characteristic component based on the laser energy fluctuation amplitude (which is inversely correlated with the laser energy fluctuation), the influence of environmental disturbances on the signal quality of the two modes can be objectively quantified, thereby improving the stability and accuracy of pollution level assessment under complex environmental conditions.

[0183] S208. Based on the correction factor, adjust the distribution ratio of each weight in the initial dynamic weight coefficient to obtain the target dynamic weight coefficient.

[0184] Among them, adjusting the initial dynamic weight coefficients can refer to using the first signal quality parameter and the second signal quality parameter in the correction factor to perform weighted correction or renormalization on the initial spectral feature weights and initial event flow feature weights generated by the attention mechanism network, so that the corrected weights reflect the actual modal reliability under environmental disturbances.

[0185] In one possible implementation, the initial spectral feature weights can be multiplied by the second signal quality parameter, and the initial event flow feature weights can be multiplied by the first signal quality parameter to obtain the corrected intermediate weights. The corrected intermediate weights are then normalized so that the sum of the target spectral feature weights and the target event flow feature weights is 1, thus obtaining the target dynamic weight coefficients.

[0186] S209. Based on the target dynamic weight coefficient, the multimodal features are weighted and fused to obtain the weighted fused features.

[0187] Weighted fusion can refer to multiplying the spectral feature components and event flow feature components in multimodal features by the corresponding target dynamic weight coefficients, and then performing weighted summation or weighted concatenation to form a unified feature representation after fusion.

[0188] In one possible implementation, the spectral feature components can be multiplied by the target spectral feature weights, the event flow feature components can be multiplied by the target event flow feature weights, and then the two weighted feature components can be summed along the feature dimension to obtain the weighted fusion feature; or the two weighted feature components can be concatenated along the feature dimension to obtain the dimension-extended weighted fusion feature.

[0189] It is understandable that by adjusting the initial dynamic weight coefficients generated by the attention mechanism using a correction factor, the fusion weights can be redistributed according to the modal reliability under environmental disturbances. This allows modes with high signal quality to receive higher weights and modes with low signal quality to receive lower weights, thereby suppressing the interference of environmental noise on the fusion results and achieving environmentally adaptive robustness.

[0190] S210. Input the weighted fusion features into the classification model to obtain the classification result; wherein, the classification result characterizes the pollution level of the target insulating surface.

[0191] In an optional implementation, the classification model includes multiple fully connected layers, batch normalization layers, and activation function layers; step S210 may include:

[0192] S2101. The weighted fusion features are input into a multi-layer fully connected layer, and then pass through the feature transformation processing of each fully connected layer, the normalization processing of the corresponding batch normalization layer, and the nonlinear activation processing of the corresponding activation function layer. Finally, the classification score corresponding to the number of pollution levels is output through the last fully connected layer.

[0193] Among them, a fully connected layer can refer to a linear transformation layer in which each neuron in a neural network establishes a connection relationship with all neurons in the previous layer. It performs linear mapping on the input features through learnable weight matrices and bias vectors, thereby realizing the transformation of feature dimensions and the projection of feature space.

[0194] In one possible implementation, for each fully connected layer, the output features of the previous fully connected layer are linearly transformed through the weight matrix and then superimposed with the bias vector to obtain a linear output result. This linear output result is then input into the next fully connected layer. Specifically, for the preceding fully connected layers except the last one, the linear output result is sequentially processed by the batch normalization layer and the non-linear activation process of the ReLU activation function layer to obtain the activation output features of that layer. For the last fully connected layer, the linear output result is directly used as the classification score output without batch normalization and non-linear activation processing.

[0195] In another possible implementation, the multi-layer fully connected layer includes an input layer, a first hidden layer, a second hidden layer, a third hidden layer, a fourth hidden layer, and an output layer. The number of neurons in the input layer corresponds to the dimension of the weighted fused feature, for example, 512 dimensions. The first hidden layer has 256 neurons, the second hidden layer has 128 neurons, the third hidden layer has 64 neurons, and the fourth hidden layer has 64 neurons. The number of neurons in the output layer corresponds to the number of smudge levels, for example, 5 (corresponding to levels 1, 2, 3, 4, and 5 smudge levels). Specifically, the input layer receives the 512-dimensional weighted fused feature, which is transformed into 256-dimensional features by the first hidden layer, 128-dimensional features by the second hidden layer, 64-dimensional features by the third hidden layer, and retains the 64-dimensional features by the fourth hidden layer. Finally, the output layer maps these features to a 5-dimensional classification score, which corresponds to the confidence score for level 1, 2, 3, 4, and 5 smudge levels, respectively. A higher score indicates a greater probability of that smudge level.

[0196] It should be understood that the channel size, layer configuration, and number of neurons in each layer of the above-mentioned multi-layer fully connected layers can be flexibly adjusted according to the actual application scenario and task complexity. For example, for scenarios with more detailed filth level classification, the number of neurons in the output layer can be increased, and for scenarios with higher feature dimensions, the number of neurons in the input layer and hidden layer can be expanded accordingly. This application does not limit this.

[0197] S2102. Convert the classification scores into a probability distribution and determine the category with the highest probability value as the classification result.

[0198] In one possible implementation, converting the classification scores into a probability distribution can include: performing an exponential normalization operation on the 5-dimensional classification scores output by the output layer using a normalized exponential function (Softmax function), mapping the classification scores of each dimension to probability values, so that the sum of the probability values ​​corresponding to all dirt levels is 1.

[0199] Furthermore, the probability values ​​corresponding to each pollution level are compared, and the pollution level with the highest probability value is determined as the pollution level of the target insulation surface, which is then used as the classification result.

[0200] It should be understood that the Softmax function described above is only an exemplary way to convert classification scores into a probability distribution. Other probability normalization methods may also be used in other possible implementations, and this application does not limit them.

[0201] It is understandable that by inputting weighted fusion features into a multi-layer fully connected classification network and performing feature transformation layer by layer, high-level discrimination features related to the pollution level can be extracted step by step, and then the probability distribution of each pollution level can be output and the optimal classification result can be determined, thus achieving accurate and reliable determination of the pollution level of the insulation surface.

[0202] In this embodiment, based on the above embodiments, on the one hand, the technical problem of semantic inconsistency in heterogeneous modal features is solved by aligning spectral features and event flow features based on plasma physics constraint rules; on the other hand, the model's adaptability and robustness under complex environmental conditions are further enhanced by introducing environmental data to correct the initial weights generated by the attention mechanism. Thus, the accuracy and reliability of pollution level determination are improved.

[0203] Figure 3 This is a flowchart illustrating a model training method provided in this application. Figure 3 Based on the foregoing embodiments, the examples provide a detailed description of the training methods for the classification models used in the foregoing embodiments, such as... Figure 3 As shown, the method includes:

[0204] S301. Obtain the training dataset; wherein, the training dataset includes multimodal data of multiple pollution level samples of insulating surfaces collected under different environmental conditions, and the multimodal data includes sample spectral data, sample event stream data and sample environmental data.

[0205] In one possible implementation, the process of constructing the training dataset may include:

[0206] In a laboratory environment, insulating surface samples with different pollution levels (e.g., levels one to five) were simulated. Different combinations of environmental conditions (e.g., different light intensities, temperatures, and humidity) were set using the controlled variable method. Spectral data of the samples were acquired using a LIBS acquisition device, event stream data of the samples were acquired using a DVS acquisition device, and environmental data of the samples were acquired using environmental sensors. Each set of multimodal data was labeled with the corresponding pollution level to form labeled training samples. Multiple training samples were divided into training set, validation set, and test set according to a preset ratio (e.g., 8:1:1). The training set was used for model parameter updates, the validation set was used for hyperparameter tuning and early stopping judgment, and the test set was used for final model performance evaluation.

[0207] S302. Align and fuse the sample spectral data and sample event stream data to determine the multimodal characteristics of the samples.

[0208] For example, the specific steps for aligning and fusing sample spectral data and sample event stream data can be referred to steps S202 to S205 in the foregoing embodiments, and will not be repeated here.

[0209] For example, if only sample spectral data is used as a single modal feature for analysis, diatomaceous earth and sodium chloride are used to prepare contaminated samples, and LIBS tests are performed on samples with 5 contamination levels. The sample information is shown in Table 1.

[0210] Table 1:

[0211]

[0212] Among them, Equivalent Salt Deposit Density (ESDD) is a parameter that measures the content of conductive substances in sewage as equivalent to the content of sodium chloride, and its unit can be milligrams per square centimeter. Equivalent Ash Deposit Density (NSDD) is a parameter that measures the content of insoluble inert substances in sewage, and its unit can be milligrams per square centimeter.

[0213] Furthermore, taking t-distribution stochastic neighbor embedding (t-SNE) as an example of dimensionality reduction method, Figure 4 A schematic diagram of the dimensionality reduction result provided in this application is shown below. Figure 4 As shown, Figure 4The results of the t-SNE two-dimensional embedding of the sample spectral data are shown. The horizontal axis represents the first dimension of t-SNE, and the vertical axis represents the second dimension of t-SNE. The categories 0, 1, 2, 3, and 4 in the figure represent level 1, level 2, level 3, level 4, and level 5 pollution, respectively. The higher the pollution level, the more severe the pollution.

[0214] from Figure 4 As can be seen, the scatter distribution of the sample spectral data shows that each sample forms several local clusters in the low-dimensional space. However, there is still some overlap between different contamination levels, indicating that spectral data has a certain ability to distinguish contamination levels, but its inter-class separability is limited. Therefore, it is necessary to introduce event stream data and spectral data for multimodal fusion. By leveraging the complementary information of the two modal features, the characterization ability of contamination levels can be enhanced, which can improve inter-class separability and classification accuracy, thus overcoming the problem of insufficient discrimination ability of single modal features under complex contamination conditions.

[0215] S303. Based on the sample environment data, dynamically assign weights to the multimodal features of the samples to obtain the weighted fusion features of the samples.

[0216] For example, the specific steps for dynamically assigning weights to the multimodal features of the samples based on the sample environment data can be referred to steps S206 to S209 in the aforementioned embodiments, and will not be repeated here.

[0217] For example, Figure 5a A schematic diagram of a weight distribution provided in this application Figure 1 , Figure 5b A schematic diagram of a weight distribution provided in this application Figure 2 , Figure 5c A schematic diagram of a weight distribution provided in this application Figure 3 , Figure 5d A schematic diagram of a weight distribution provided in this application Figure 4 , Figure 5a , Figure 5b , Figure 5c , Figure 5d Both indicate that the attention mechanism network is used to calculate the initial spectral feature weights and initial event flow feature weights by using spectral feature components and event flow feature components as query vectors and key vectors respectively.

[0218] It can be seen that, Figure 5a The weight distribution shown represents the case where the spectral feature components and event flow feature components are strongly correlated, and the event flow features are dominant in discriminative power. Figure 5b The weight distribution shown represents the case where the correlation between spectral feature components and event flow feature components is weak, and the event flow feature information is abundant. Figure 5cThe weight distribution shown represents the case where the spectral feature components are strongly correlated with the event flow feature components and the spectral features are dominant in terms of discriminative power. Figure 5d The weight distribution shown represents cases where the correlation between spectral feature components and event flow feature components is weak, but the spectral feature information is rich. Different modal feature correlations and discriminative differences result in different weight distributions, demonstrating that the attention mechanism network can adaptively allocate fusion weights based on the content association of multimodal features themselves.

[0219] It is understandable that by dynamically calculating the initial weight coefficients through the attention mechanism network, the correlation strength between spectral feature components and event flow feature components can be quantified, thus providing a basis for subsequent weight correction based on environmental data and realizing preliminary weight allocation based on content relevance.

[0220] S304. Based on the weighted fusion features of the samples and their corresponding pollution level labels, the initial model is iteratively trained to obtain a classification model.

[0221] The initial model can refer to an untrained, multi-layer fully connected neural network with preset network structure parameters, whose network structure is consistent with the classification model structure in the aforementioned embodiments.

[0222] The classification model is a multi-layer fully connected classification network that has been iteratively optimized using training data and is used in the aforementioned embodiments to determine the pollution level of the target insulating surface.

[0223] In one possible implementation, step S304 may include:

[0224] The weighted fusion features of the samples are input into the initial model, and the predicted probability distribution of each level of contamination is calculated through forward propagation. Based on the predicted probability distribution and the contamination level label corresponding to the sample, the cross-entropy loss function value is calculated. Based on the cross-entropy loss function value, the gradient of the parameters of each layer of the initial model is calculated through the backpropagation algorithm. Using the Adam optimizer (the parameters of the Adam optimizer can include: initial learning rate of 0.0005, batch size of 32), the weight matrix and bias vector of the initial model are updated according to the calculated gradient. The iterative process of forward propagation, loss calculation, backpropagation and parameter update is repeated until the preset maximum number of training rounds (e.g., 100 rounds) is reached or the accuracy of the validation set does not improve for multiple consecutive rounds (e.g., 5 rounds), triggering the early stopping mechanism to obtain the trained classification model.

[0225] In one possible implementation, Dropout regularization is also used during training. After the activation output of the preceding fully connected layer, some neurons are randomly dropped with a preset probability (e.g., 0.3) to prevent the model from overfitting. Dropout is turned off during the validation phase to evaluate the model performance with the complete network structure.

[0226] In one possible implementation Figure 6 The schematic diagram of the architecture of a classification model based on a convolutional neural network provided in this application is as follows: Figure 6 As shown, the architecture includes convolutional layers, pooling layers, and fully connected layers.

[0227] The sample spectral data and sample event stream data are used as input data. After preprocessing and alignment fusion, the multimodal features of the samples are determined. The convolutional layer performs local feature extraction and channel expansion on the multimodal features of the samples to determine the convolutional feature map. The convolutional feature map is then input into the pooling layer for downsampling and feature compression to determine the pooling feature representation. This pooling feature representation is then input into the fully connected layer for feature transformation and classification decision, and the predicted probability distribution of each pollution level is output.

[0228] For example, the convolutional layer receives multimodal features or weighted fusion features of samples as input, performs local perception on the input features through sliding convolutional kernels, extracts local pattern features and edge information in the multimodal data, and expands the feature dimension through multiple output channels to enhance the expressive power of the features; the pooling layer downsamples the convolutional feature map, retains significant feature responses and reduces the feature dimension, enhancing the model's translation invariance and noise resistance; the fully connected layer receives the pooling feature representation, performs layer-by-layer nonlinear transformation through a combination structure of multiple fully connected sub-layers, batch normalization layers and activation function layers, and finally maps it to an output dimension consistent with the number of pollution level categories, outputs classification scores and converts them into a probability distribution through the Softmax function.

[0229] It is understandable that by using convolutional neural networks, local discriminative patterns and multi-scale feature representations in multimodal features of samples can be automatically learned, thereby enhancing the classification model's ability to represent and generalize features related to pollution levels, and improving the accuracy and robustness of pollution level determination.

[0230] For example, Figure 7 This application provides a schematic diagram illustrating the effect of a different model, such as... Figure 7As shown, taking different classification models using Convolutional Neural Network (CNN), K-Nearest Neighbor (KNN), and Support Vector Machine (SVM) as examples, and training data including sample spectral data and / or sample event stream data as examples, the following are the comparison results of the classification performance of different models with different training data. The vertical axis represents "Fusion, CNN", indicating that a CNN model is trained using sample spectral data and sample event stream data; "Fusion, KNN", indicating that a KNN model is trained using features obtained by fusing sample spectral data and sample event stream data; "Fusion, SVM", indicating that an SVM model is trained using features obtained by fusing sample spectral data and sample event stream data; "DVS, CNN", indicating that a CNN model is trained using only sample event stream data; "DVS, SVM", indicating that an SVM model is trained using only sample event stream data; "LIBS, CNN", indicating that a CNN model is trained using only sample spectral data; and "LIBS, SVM", indicating that an SVM model is trained using only sample spectral data.

[0231] It can be seen that when using only the LIBS single modality, the accuracy of SVM and KNN is only 47.50% and 49.00%, respectively, with corresponding precision and recall both below 51%. This indicates that the single spectral modality has limited discriminative power in a one-dimensional feature space and is difficult to reliably identify the level of contamination. In contrast, the DVS single modality exhibits stronger feature representation capabilities: under the same one-dimensional feature dimension, the accuracy of SVM increases to 77.50%, while the DVS-based CNN model achieves approximately 94.5% in accuracy, precision, and recall. Furthermore, when the dimensionality-reduced LIBS and DVS features are fused in a one-dimensional format, the classification performance of all three algorithms is significantly improved compared to the single LIBS modality. Specifically, the accuracy of SVM further increases from 77.50% to 93.50%, KNN to 85.50%, and the accuracy of CNN reaches 95.00%, with precision and recall also approaching 95%, making it the best result among all combinations. This indicates that LIBS and DVS bimodal features are highly complementary in pollution level assessment. In particular, when combined with the strong discriminative power of CNN under the DVS modality, the model's classification ability and robustness for different pollution levels can be significantly enhanced.

[0232] Specifically, based on training a CNN model using sample spectral data and sample event stream data to obtain a classification model, Figure 8 A matrix diagram illustrating the classification effect provided in this application, such as... Figure 8As shown, Figure 8 The horizontal axis represents the predicted category (represented as 0, 1, 2, 3, 4), i.e., the predicted level of contamination, and the vertical axis represents the true category (represented as 0, 1, 2, 3, 4), i.e., the actual level of contamination; from Figure 8 It can be seen that the model's diagonal element values ​​are higher than its off-diagonal element values ​​for all five levels of contamination. Specifically, categories 0, 1, and 4 all achieved completely correct predictions for 40 samples; in category 2, 36 samples were correctly identified, with only 4 misclassified; and in category 3, 34 samples were correctly classified, but 6 samples were still misclassified, with 5 of them misclassified as adjacent category 2 samples. This demonstrates the model's high accuracy across all categories.

[0233] This embodiment provides a model training method. By constructing a training dataset containing multimodal data under different environmental conditions, and using an alignment fusion and dynamic weight allocation process consistent with the inference stage to generate sample weighted fusion features, the method trains a multi-layer fully connected classification network end-to-end. This enables the classification model to learn an environment-adaptive multimodal feature fusion strategy and a pollution level discrimination mode. Thus, in practical applications, it can accurately and reliably determine the pollution level of the target insulation surface based on real-time multimodal data and environmental conditions.

[0234] Figure 9 A schematic diagram of a system for determining the pollution level of an insulating surface provided in this application is shown below. Figure 9 As shown, the system includes a host, a delay controller, a laser, a dynamic vision sensor, and a spectrometer.

[0235] The host is used to control the coordinated operation of various hardware components, receive and process data collected by various sensors, and can execute the method for determining the contamination level of the insulating surface in any of the aforementioned embodiments.

[0236] The delay controller is used to send synchronous trigger signals to the laser and the dynamic vision sensor respectively according to the preset delay parameters to achieve time synchronization of spectral data acquisition and event stream data acquisition. The delay parameters of the delay controller can be adaptively adjusted according to the laser response time and the dynamic vision sensor start-up time to ensure synchronous acquisition of the two modes of data during the plasma discharge process.

[0237] The laser is used to emit high-energy pulsed laser light towards the target insulating surface after receiving a trigger signal from the delay controller, and generate plasma on the insulating surface through laser-induced breakdown effect. The output energy, pulse width and repetition frequency of the laser can be adjusted according to the material and degree of contamination of the insulating surface to ensure stable plasma generation and effective excitation of spectral signals.

[0238] The dynamic vision sensor is used to acquire spatial dynamic evolution information of the plasma discharge process in an asynchronous event-driven manner after receiving the trigger signal from the delay controller, and outputs event stream data containing pixel position, timestamp and polarity information; specifically, the dynamic vision sensor can capture the rapid dynamic change process of plasma.

[0239] A spectrometer is used to collect the spectral signals emitted by plasma during discharge and output spectral data containing wavelength and intensity information. The spectrometer's detection range can cover the ultraviolet to near-infrared band and can acquire plasma characteristic spectral line information for analyzing the elemental composition and content of contaminants on insulating surfaces.

[0240] In one possible implementation, the system also includes an environmental monitoring module, which includes a light sensor, a temperature sensor, and a humidity sensor. This module is used to collect environmental data such as light intensity, temperature, and humidity of the environment in which the target insulating surface is located, and transmits the environmental data to the host to participate in the dynamic weight allocation process.

[0241] In one possible implementation, the host is connected to the delay controller, laser, dynamic vision sensor, spectrometer, and environmental monitoring module via wired or wireless communication. The arrangement of each component can be flexibly adjusted according to the actual detection scenario. The optical axes of the spectrometer and dynamic vision sensor are aligned with the same detection area on the target insulating surface to ensure spatial consistency.

[0242] The system for determining the contamination level of the insulating surface in this embodiment can execute the technical solution in the above method. Its specific implementation process and technical principle are the same, and will not be repeated here.

[0243] Figure 10 A schematic diagram of a device for determining the pollution level of an insulating surface provided in this application is shown below. Figure 10 As shown, the contamination level determination device 100 for insulating surfaces includes: a data acquisition module 1001, an alignment and fusion module 1002, a weight allocation module 1003, and a level determination module 1004.

[0244] The data acquisition module 1001 is used to acquire spectral data, event stream data and environmental data of the target insulating surface; among them, the spectral data characterizes the spectral radiation characteristics of the plasma during the discharge process, the event stream data characterizes the dynamic changes of the plasma optical signal in the spatiotemporal dimension, and the environmental data characterizes the environmental conditions of the target insulating surface.

[0245] Alignment and fusion module 1002 is used to perform alignment and fusion processing on spectral data and event stream data to determine multimodal features; wherein, the multimodal features represent spectral feature components and event stream feature components;

[0246] The weight allocation module 1003 is used to dynamically allocate weights to multimodal features based on environmental data to obtain weighted fusion features;

[0247] The grade determination module 1004 is used to input the weighted fusion features into the classification model to obtain the classification result; wherein, the classification result characterizes the pollution level of the target insulation surface.

[0248] In an optional example, the alignment and blending module 1002 is also used for:

[0249] The spectral data is subjected to a first feature extraction process to obtain spectral features, and the event stream data is subjected to a second feature extraction process to obtain event stream features. The spectral features and event stream features are mapped to the same preset dimension space to obtain corresponding spectral alignment features and event stream alignment features. The spectral alignment features and event stream alignment features are fused to obtain multimodal features.

[0250] In an optional example, the alignment and blending module 1002 is also specifically used for:

[0251] The spectral data is preprocessed to obtain preprocessed data; the preprocessing includes at least one of baseline correction, denoising, and normalization; based on a preset dimensionality reduction algorithm, the preprocessed data is dimensionality reduced to obtain spectral features; the preset dimensionality reduction algorithm includes at least one of linear dimensionality reduction and nonlinear dimensionality reduction; the linear dimensionality reduction algorithm includes at least one of principal component analysis (PCA) and independent component analysis (ICA); the nonlinear dimensionality reduction algorithm includes the locally linear embedding (LLE) algorithm.

[0252] In an optional example, the event stream data is event stream information based on plasma optical signals acquired by a dynamic vision sensor; the alignment and fusion module 1002 is also specifically used for:

[0253] The event stream data is divided into a first image set and a second image set. The first image in the first image set represents the spatial distribution of the brightness enhancement region in the plasma optical signal within the corresponding time window, and the second image in the second image set represents the spatial distribution of the brightness reduction region in the plasma optical signal within the corresponding time window. Each image frame in the first and second image sets is filtered to obtain a filtered image sequence. The geometric morphological features and event statistical features of the plasma are extracted from the filtered image sequence to obtain the event stream features. The geometric morphological features include shape features and size features, and the event statistical features include event quantity features.

[0254] In an optional example, the alignment and blending module 1002 is also specifically used for:

[0255] Based on the physical constraint rules of plasma radiation, the physical correlation between spectral features and event flow features is determined. The physical constraint rules characterize the physical correspondence between plasma spectral intensity attenuation and plasma spatial scale contraction. According to the physical correlation, the spectral features and event flow features are matched and aligned under physical constraints to obtain physically aligned spectral features and event flow features. The physically aligned spectral features and physically aligned event flow features are mapped to the same preset dimension space to obtain spectral aligned features and event flow aligned features.

[0256] In an optional example, the weight allocation module 1003 is also used for:

[0257] Multimodal features are input into an attention mechanism network to generate initial dynamic weight coefficients. The attention mechanism network is used to perform cross-attention calculations based on spectral feature components and event flow feature components, which are used as query vectors and key vectors respectively, to obtain the initial dynamic weight coefficients. These initial dynamic weight coefficients include initial spectral feature weights and initial event flow feature weights. A correction factor is determined based on environmental data, representing the degree of influence of environmental disturbances on the spectral feature components and event flow feature components respectively. Based on the correction factor, the allocation ratio of each weight in the initial dynamic weight coefficients is adjusted to obtain the target dynamic weight coefficients. Based on the target dynamic weight coefficients, the multimodal features are weighted and fused to obtain weighted fused features.

[0258] In an optional example, the environmental data includes illumination parameters and laser parameters, where the illumination parameters characterize the ambient background light intensity and the laser parameters characterize the fluctuation amplitude of the laser output energy used to generate plasma; the weight allocation module 1003 is also specifically used for:

[0259] Based on the illumination parameters, a first signal quality parameter corresponding to the event flow feature component is determined; wherein, the first signal quality parameter is inversely correlated with the illumination parameters; based on the laser parameters, a second signal quality parameter corresponding to the spectral feature component is determined; wherein, the second signal quality parameter is inversely correlated with the laser parameters; the first signal quality parameter and the second signal quality parameter are determined as correction factors.

[0260] In an optional example, the classification model includes multiple fully connected layers, batch normalization layers, and activation function layers; the ranking determination module 1004 is also used for:

[0261] The weighted fusion features are input into multiple fully connected layers, and then sequentially undergo feature transformation processing in each fully connected layer, normalization processing in the corresponding batch normalization layer, and nonlinear activation processing in the corresponding activation function layer. Finally, the last fully connected layer outputs the classification score corresponding to the number of pollution levels. The classification score is converted into a probability distribution, and the category with the highest probability value is determined as the classification result.

[0262] In an optional example, the contamination level determination device 100 for insulating surfaces further includes a model training module, which is used for:

[0263] Obtain a training dataset; the training dataset includes multimodal data of multiple contamination levels of insulating surfaces collected under different environmental conditions. The multimodal data includes sample spectral data, sample event stream data, and sample environmental data. Align and fuse the sample spectral data and sample event stream data to determine the sample multimodal features. Based on the sample environmental data, dynamically assign weights to the sample multimodal features to obtain sample weighted fusion features. Iteratively train the initial model based on the sample weighted fusion features and their corresponding contamination level labels to obtain a classification model.

[0264] The device for determining the contamination level of the insulating surface in this embodiment can execute the technical solution in the above method. Its specific implementation process and technical principle are the same, and will not be described again here.

[0265] Figure 11 A schematic diagram of a device for determining the pollution level of an insulating surface is provided in an embodiment of this application, as shown below. Figure 11 As shown, the contamination level determination device 110 for insulating surfaces provided in this embodiment includes at least one processor 1101 and a memory 1102. Optionally, the contamination level determination device 110 for insulating surfaces further includes a communication component 1103. The processor 1101, memory 1102, and communication component 1103 are connected via a bus 1104.

[0266] In a specific implementation, at least one processor 1101 executes computer execution instructions stored in memory 1102, causing at least one processor 1101 to perform the above-described method.

[0267] The specific implementation process of processor 1101 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.

[0268] In the above embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.

[0269] The memory may include random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage device.

[0270] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0271] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the above-described method.

[0272] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0273] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as magnetic disks or optical disks.

[0274] Finally, it should be noted that other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein, and is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. A method for determining the contamination level of an insulating surface, characterized in that, include: Spectral data, event flow data, and environmental data of the target insulating surface are collected; wherein, the spectral data characterizes the spectral radiation characteristics of the plasma during the discharge process, the event flow data characterizes the dynamic changes of the plasma optical signal in the spatiotemporal dimension, and the environmental data characterizes the environmental conditions of the target insulating surface. The spectral data and the event stream data are aligned and fused to determine multimodal features; wherein, the multimodal features characterize spectral feature components and event stream feature components; Based on the environmental data, the multimodal features are dynamically weighted to obtain weighted fusion features; The weighted fusion features are input into the classification model to obtain the classification result; wherein the classification result characterizes the contamination level of the target insulating surface.

2. The method according to claim 1, characterized in that, The step of aligning and fusing the spectral data and the event stream data to determine multimodal features includes: The spectral data is subjected to a first feature extraction process to obtain spectral features, and the event stream data is subjected to a second feature extraction process to obtain event stream features; The spectral features and the event flow features are mapped to the same preset dimension space to obtain the corresponding spectral alignment features and event flow alignment features; The spectral alignment feature and the event flow alignment feature are fused together to obtain the multimodal feature.

3. The method according to claim 2, characterized in that, The first feature extraction process on the spectral data to obtain spectral features includes: The spectral data is preprocessed to obtain preprocessed data; wherein the preprocessing includes at least one of baseline correction processing, noise reduction processing, and normalization processing. Based on a preset dimensionality reduction algorithm, the preprocessed data is subjected to dimensionality reduction processing to obtain the spectral features; wherein, the preset dimensionality reduction algorithm includes at least one of linear dimensionality reduction algorithm and nonlinear dimensionality reduction algorithm; the linear dimensionality reduction algorithm includes at least one of principal component analysis (PCA) algorithm and independent component analysis (ICA) algorithm; the nonlinear dimensionality reduction algorithm includes the locally linear embedding (LLE) algorithm.

4. The method according to claim 2, characterized in that, The event stream data is event stream information based on plasma optical signals acquired by a dynamic visual sensor; the second feature extraction process on the event stream data to obtain event stream features includes: The event stream data is divided into a first image set and a second image set; wherein, the first image in the first image set represents the spatial distribution of the brightness enhancement region in the plasma optical signal within the corresponding time window, and the second image in the second image set represents the spatial distribution of the brightness reduction region in the plasma optical signal within the corresponding time window; Each image frame in the first image set and the second image set is filtered to obtain a filtered image sequence. The event flow features are obtained by extracting the geometric morphology features and event statistical features of the plasma from the filtered image sequence; wherein the geometric morphology features include shape features and size features, and the event statistical features include event quantity features.

5. The method according to claim 2, characterized in that, The step of mapping the spectral features and the event flow features to the same preset dimension space to obtain corresponding spectral alignment features and event flow alignment features includes: Based on the physical constraint rules of plasma radiation, the physical correlation between the spectral features and the event flow features is determined; wherein, the physical constraint rules characterize the physical correspondence between plasma spectral intensity attenuation and plasma spatial scale contraction; Based on the physical association, the spectral features and the event flow features are matched and aligned under physical constraints to obtain physically aligned spectral features and event flow features; The physically aligned spectral features and the physically aligned event flow features are mapped to the same preset dimension space to obtain the spectral alignment features and the event flow alignment features.

6. The method according to claim 1, characterized in that, The step of dynamically assigning weights to the multimodal features based on the environmental data to obtain weighted fusion features includes: The multimodal features are input into an attention mechanism network to generate initial dynamic weight coefficients; wherein, the attention mechanism network is used to perform cross-attention calculation based on spectral feature components and event flow feature components as query vectors and key vectors to obtain initial dynamic weight coefficients, and the initial dynamic weight coefficients include initial spectral feature weights and initial event flow feature weights; Based on the environmental data, a correction factor is determined; wherein the correction factor characterizes the degree of influence of environmental disturbances on spectral feature components and event flow feature components, respectively. Based on the correction factor, the allocation ratio of each weight in the initial dynamic weight coefficient is adjusted to obtain the target dynamic weight coefficient; The multimodal features are weighted and fused according to the target dynamic weight coefficients to obtain the weighted fused features.

7. The method according to claim 6, characterized in that, The environmental data includes illumination parameters and laser parameters. The illumination parameters characterize the ambient background light intensity, and the laser parameters characterize the fluctuation amplitude of the laser output energy used to generate plasma. The step of determining the correction factor based on the environmental data includes: Based on the illumination parameters, a first signal quality parameter corresponding to the event stream feature component is determined; wherein, the first signal quality parameter is inversely correlated with the illumination parameters; Based on the laser parameters, a second signal quality parameter corresponding to the spectral feature component is determined; wherein, the second signal quality parameter is inversely correlated with the laser parameters; The first signal quality parameter and the second signal quality parameter are determined as the correction factor.

8. The method according to claim 1, characterized in that, The classification model includes multiple fully connected layers, batch normalization layers, and activation function layers; the step of inputting the weighted fusion features into the classification model to obtain the classification result includes: The weighted fusion features are input into the multi-layer fully connected layer, and then sequentially pass through the feature transformation processing of each fully connected layer, the normalization processing of the corresponding batch normalization layer, and the nonlinear activation processing of the corresponding activation function layer. Finally, the classification score corresponding to the number of pollution levels is output through the last fully connected layer. The classification scores are converted into a probability distribution, and the category with the highest probability value is determined as the classification result.

9. The method according to any one of claims 1-8, characterized in that, Also includes: Obtain a training dataset; wherein the training dataset includes multimodal data of multiple contamination levels of insulating surfaces collected under different environmental conditions, and the multimodal data includes sample spectral data, sample event stream data, and sample environmental data; The sample spectral data and the sample event stream data are aligned and fused to determine the multimodal features of the samples; Based on the sample environment data, the dynamic weight allocation is performed on the multimodal features of the samples to obtain the sample weighted fusion features; Based on the weighted fusion features of the samples and their corresponding pollution level labels, the initial model is iteratively trained to obtain the classification model.

10. A device for determining the contamination level of an insulating surface, characterized in that, include: The data acquisition module is used to acquire spectral data, event stream data, and environmental data of the target insulating surface; wherein, the spectral data characterizes the spectral radiation characteristics of the plasma during the discharge process, the event stream data characterizes the dynamic changes of the plasma optical signal in the spatiotemporal dimension, and the environmental data characterizes the environmental conditions of the target insulating surface. An alignment and fusion module is used to perform alignment and fusion processing on the spectral data and the event stream data to determine multimodal features; wherein, the multimodal features characterize spectral feature components and event stream feature components; The weight allocation module is used to dynamically allocate weights to the multimodal features based on the environmental data to obtain weighted fusion features; The grade determination module is used to input the weighted fusion features into the classification model to obtain the classification result; wherein the classification result characterizes the pollution level of the target insulating surface.

Images