Power grid harmonic monitoring method based on multi-sensor fusion

Abstract

The invention discloses a power grid harmonic monitoring method based on multi-sensor fusion, and belongs to the field of power grid harmonic monitoring, and the method comprises the following steps: carrying out the preprocessing of harmonic data and non-electrical data collected by multiple sensors, and obtaining the preprocessing data; carrying out dimension reduction processing on the preprocessed data to extract harmonic features and equipment state features; constructing a nonlinear mapping model based on the harmonic characteristics and the equipment state characteristics to obtain mapped harmonic characteristic data; carrying out feature level fusion on the mapped harmonic feature data and the preprocessed non-electrical quantity data, and screening a high-credibility sensor data set to obtain a sensor credibility weight and a sensor harmonic feature parameter; obtaining optimized harmonic information based on the credibility weight of the sensor and the harmonic characteristic parameters of the sensor; and performing parameter optimization on the nonlinear mapping model based on the optimized harmonic information to update the incidence relation between the harmonic characteristics and the equipment state.

Description

Technical Field

[0001] The present invention belongs to the technical field of power grid harmonic monitoring, and particularly relates to a power grid harmonic monitoring method based on multi-sensor fusion. Background Art

[0002] In the practical application of power grid harmonic monitoring, due to the complex and changeable power grid environment, traditional single-sensor monitoring methods are difficult to comprehensively capture harmonic information. There are multiple harmonic sources in the power grid, and the generated harmonic signals have non-stationary characteristics, and there may be mutual interference between different harmonic sources. When multi-sensors are used for monitoring, the harmonic data collected by each sensor often has a high degree of redundancy, resulting in low data processing efficiency. At the same time, there is heterogeneity between the electrical and non-electrical quantity data collected by different sensors, and how to effectively fuse these multi-modal data has become a key issue. Especially when the operating state of the equipment changes, the correlation between harmonic characteristics and equipment state may change non-linearly, which makes it difficult for traditional linear analysis methods to accurately capture this complex relationship. In addition, when there are conflicts in the harmonic information provided by multiple sensors, how to reasonably weigh the credibility of each sensor and make decision fusion based on prior knowledge has become the core challenge to improve the accuracy of harmonic monitoring. In such a complex scenario, it is difficult to comprehensively solve the above problems by relying solely on a certain technical means, and it is necessary to consider the synergistic effects of multiple links such as data dimensionality reduction, feature extraction, multi-modal fusion, and decision optimization. Summary of the Invention

[0003] To solve the above technical problems, the present invention proposes a power grid harmonic monitoring method based on multi-sensor fusion to solve the problems existing in the above prior art.

[0004] To achieve the above object, the present invention provides a power grid harmonic monitoring method based on multi-sensor fusion, including:

[0005] Preprocessing the harmonic data and non-electrical quantity data collected by multi-sensors to obtain preprocessed data;

[0006] Performing dimensionality reduction processing on the preprocessed data to extract harmonic characteristics and equipment state characteristics;

[0007] Constructing a non-linear mapping model based on the harmonic characteristics and the equipment state characteristics, and obtaining the mapped harmonic characteristic data based on the non-linear mapping model;

[0008] Performing feature-level fusion on the mapped harmonic characteristic data and the preprocessed non-electrical quantity data to generate a multi-modal feature vector;

[0009] Based on the multi-modal feature vector, screening a high-credibility sensor data set to obtain sensor credibility weights and sensor harmonic characteristic parameters;

[0010] Obtain the optimized harmonic information based on the sensor credibility weight and the sensor harmonic characteristic parameters;

[0011] Optimize the parameters of the non - linear mapping model based on the optimized harmonic information to obtain the optimized non - linear mapping model parameters;

[0012] Determine the correlation between harmonic characteristics and equipment status based on the optimized non - linear mapping model parameters.

[0013] Optionally, the process of extracting harmonic characteristics and equipment status characteristics includes:

[0014] Use the principal component analysis method to classify the pre - processed data to obtain the standardized pre - processed data and several feature vectors;

[0015] Project the standardized pre - processed data onto several principal component directions to obtain the reduced - dimension data matrix;

[0016] Determine harmonic characteristics based on the principal component distribution in the reduced - dimension data matrix;

[0017] Combine the principal components in the reduced - dimension data matrix and equipment operation parameters to determine equipment status characteristics; where the equipment operation parameters include: equipment temperature, equipment vibration.

[0018] Optionally, the process of constructing a non - linear mapping model and obtaining the mapped harmonic characteristic data based on the non - linear mapping model includes:

[0019] Use the Gaussian kernel function to map the reduced - dimension data matrix to a high - dimensional feature space to obtain a feature matrix;

[0020] Use the support vector machine algorithm to classify the feature matrix to obtain the equipment operation status;

[0021] When the equipment operation status is abnormal, use the random forest algorithm to predict the abnormal status and determine the abnormal type;

[0022] Analyze the abnormal causes using the decision tree algorithm based on the abnormal type to generate a list of abnormal causes;

[0023] Extract the corresponding harmonic characteristics based on the list of abnormal causes and update the feature matrix to obtain the mapped harmonic characteristic data.

[0024] Optionally, the process of generating a multi - modal feature vector by performing feature - level fusion on the mapped harmonic characteristic data and the pre - processed non - electrical quantity data includes:

[0025] Perform element - level addition on the mapped harmonic characteristic data and the pre - processed non - electrical quantity data to obtain a first feature map;

[0026] Divide the first feature map into several groups along the channel dimension, and perform interaction operations on the features within each group to obtain an attention mask that captures the feature correlation between channels;

[0027] Normalize the attention mask to obtain an initial fusion feature map;

[0028] Separate the initial fusion feature map to obtain strong features, weak features, and strong-weak feature weight information;

[0029] Compare the strong-weak feature weight information to obtain a strong feature attention map and a weak feature attention map;

[0030] Concatenate and fuse the strong feature attention map and the weak feature attention map to generate a multi-modal feature vector.

[0031] Optionally, the process of screening a high-confidence sensor data set based on the multi-modal feature vector to obtain the sensor credibility weight includes:

[0032] Construct a sensor credibility evaluation model using a Bayesian network based on the device state characteristics and the conditional probability distribution of the multi-modal feature vector;

[0033] The sensor credibility evaluation model calculates the conditional probability value of the sensor based on prior knowledge and real-time data;

[0034] Compare the conditional probability value of the sensor with a preset threshold to judge the credibility weight of each sensor.

[0035] Optionally, the process of optimizing the harmonic information includes: using a weighted fusion algorithm to fuse and calculate the sensor credibility weight and the sensor harmonic feature parameters to obtain the optimized harmonic information.

[0036] Optionally, the process of obtaining the optimized non-linear mapping model parameters includes:

[0037] Extract the feature parameters of the optimized harmonic information to obtain harmonic feature parameters;

[0038] Use a support vector machine model to classify the harmonic feature parameters to obtain an initial classification result;

[0039] Judge the device operating state based on the initial classification result;

[0040] When the device is in an abnormal state, use a clustering analysis method to perform pattern recognition on the optimized harmonic information to obtain a pattern recognition result;

[0041] Update the classification criterion of the support vector machine model based on the pattern recognition result, and output the final classification result;

[0042] Input the obtained final classification result into the non-linear mapping model to obtain the optimized parameters of the non-linear mapping model.

[0043] Compared with the prior art, the present invention has the following advantages and technical effects:

[0044] The present invention discloses a power grid harmonic monitoring method based on multi-sensor fusion. For the heterogeneous data collected by multi-sensors, dimension reduction is performed through preprocessing and principal component analysis to extract harmonic features and equipment status features. A non-linear mapping model is constructed by using the kernel function method to capture the relationship between harmonics and equipment status. The harmonic data and non-electrical quantity data are fused by using the feature-level fusion method to generate a multi-modal feature vector. The credibility of sensors is evaluated based on the Bayesian network, and the weighted fusion method is used to optimize the harmonic information. Finally, the optimized harmonic information is classified by using a support vector machine to judge whether the equipment operation status is abnormal and dynamically adjust the model parameters. The present invention realizes the effective fusion and analysis of multi-source heterogeneous data, improves the accuracy and reliability of harmonic anomaly detection, and provides a new technical means for power equipment status monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] The drawings constituting a part of this application are used to provide a further understanding of this application. The schematic embodiments and descriptions thereof of this application are used to explain this application and do not constitute an improper limitation to this application. In the drawings:

[0046] Figure 1 It is a flowchart of the power grid harmonic monitoring method based on multi-sensor fusion according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0047] It should be noted that, without conflict, the embodiments in this application and the features in the embodiments can be combined with each other. The following will refer to the drawings and combine the embodiments to detail this application.

[0048] It should be noted that the steps shown in the flowchart of the drawings can be executed in a computer system such as a set of computer executable instructions, and although the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in a different order than here.

[0049] Embodiment 1

[0050] As Figure 1 shown, in this embodiment, a power grid harmonic monitoring method based on multi-sensor fusion is provided, including the following steps:

[0051] Step S101, preprocess the harmonic data and non-electrical quantity data collected by multi-sensors to obtain preprocessed data.

[0052] Further, obtain the harmonic data and non-electrical quantity data collected by multiple sensors, and perform preprocessing on the heterogeneous data characteristics to unify the data format and dimension.

[0053] Obtain the harmonic data and non-electrical quantity data collected by multiple sensors through a preset data acquisition interface. For the obtained harmonic data and non-electrical quantity data, use a data cleaning algorithm to remove noise and outliers. According to the characteristics of heterogeneous data, use a data conversion algorithm to convert data in different formats into a unified format. If the dimensions are inconsistent, use a data normalization algorithm to unify the data to the same dimension. Store the processed data with a unified format and dimension through a preset data storage interface.

[0054] As a specific implementation manner of this embodiment, in the sensor data acquisition stage, real-time acquisition of sensor voltage harmonic data and non-electrical quantity data such as bearing vibration is realized. The voltage harmonic probe collects data such as fundamental wave, third harmonic, and fifth harmonic output by the distribution cabinet, and at the same time, the vibration sensor collects the operating state data of the motor bearing. The data cleaning link preprocesses the collected data. The original voltage harmonic data may have abnormalities such as mutations and jumps, and is screened by setting a normal value range. For example, the effective value range of the fundamental wave voltage is between 90% and 110% of the rated value. The median filtering method is used to remove impact noise from the vibration data. In the data format conversion stage, the voltage harmonic data is stored in a time series form, including timestamps, amplitudes, and phases. The vibration sensor data is stored in the form of a waveform file. When converting, the amplitude values are extracted from the waveform data according to the sampling interval and unified into the time series data format with the harmonic data. During the normalization process, the fundamental wave amplitude is in the kilovolt order of magnitude, the harmonic amplitude is in the volt order of magnitude, and the vibration amplitude is in the millimeter order of magnitude. Map various types of data to a unified interval from zero to one to avoid the impact of dimension differences on subsequent analysis. The data storage uses a distributed database, and a data table is established to store the unified time series data. The table structure includes fields such as time, measurement point, and value, and supports fast retrieval and analysis.

[0055] Step S102, perform dimensionality reduction processing on the preprocessed data to extract harmonic features and equipment state features.

[0056] Further, use the principal component analysis method to perform dimensionality reduction on the preprocessed data, eliminate redundant information, and extract harmonic features and equipment state features.

[0057] A set of harmonic data and non - electrical quantity data collected by multiple sensors. After data pre - processing, an n×m matrix X is obtained, where n is the number of samples and m is the number of features. In this embodiment, PCA - like methods are directly used to complete the dimensionality reduction of the data to obtain pre - processed data and several eigenvectors after standardization, that is, several principal components. The pre - processed data after standardization is projected onto several principal component directions to obtain a dimensionality - reduced data matrix. Harmonic features and equipment status features are extracted based on the dimensionality - reduced data matrix. Among them, harmonic features are obtained by analyzing the distribution and change trend of principal components to identify features such as the frequency and amplitude of harmonics. For example, the principal components may reflect the change in harmonic intensity at a specific frequency or the harmonic correlation between different sensors. Equipment status features are combined with equipment operation parameters and historical data to extract features reflecting the health status of the equipment from the principal components. For example, some principal components may be related to status parameters such as the temperature and vibration of the equipment and can be used to evaluate the operating health of the equipment.

[0058] Step S103, construct a non - linear mapping model based on the harmonic features and the equipment status features, and obtain the mapped harmonic feature data based on the non - linear mapping model.

[0059] Furthermore, according to the harmonic features and the equipment status features, a non - linear mapping model is constructed using the kernel function method to capture the non - linear relationship between the harmonic features and the equipment status. According to the non - linear mapping model, harmonic features and equipment status features are extracted to construct a feature matrix. For the data in the feature matrix, a support vector machine algorithm is used for classification to judge the equipment operating status. If the equipment operating status is abnormal, a random forest algorithm is used to predict the abnormal status to determine the type of abnormality. According to the type of abnormality, a decision tree algorithm is used to analyze the cause of the abnormality to generate a list of causes of the abnormality. For the key factors in the list of causes of the abnormality, the corresponding harmonic features are extracted, and the equipment status feature matrix is updated to obtain the mapped harmonic feature data. According to the mapped harmonic feature data, a non - linear mapping model is reconstructed to complete the equipment status monitoring and optimization process.

[0060] Furthermore, as a specific implementation of this embodiment, the data matrix after dimensionality reduction is mapped to a high-dimensional feature space using the Gaussian kernel function, which can better capture this non-linear relationship. Through kernel function mapping, features such as the content of the third harmonic and the content of the fifth harmonic can be transformed into a form that is easier to analyze. The construction of the feature matrix needs to comprehensively consider multiple dimensions. In the monitoring of power equipment, indicators such as the harmonic distortion rate, load rate, and temperature can be used as basic features. These features form a new feature matrix after kernel function mapping, which is more conducive to subsequent analysis. For example, when a certain transformer is in operation, with the third harmonic content being 3% and the temperature being 40 degrees Celsius, the feature vector obtained after kernel function mapping can more accurately reflect the equipment status. When using the support vector machine algorithm for state classification, the radial basis kernel function can be selected to construct a classifier. The classification model obtained through training with historical data can identify abnormal states. For example, when the harmonic distortion rate exceeds 5% and the load rate is lower than 60%, the system will determine it as an abnormal state. When using the random forest algorithm to predict the type of abnormality, multiple decision trees can be established. Each tree makes predictions based on different combinations of features, and finally determines the type of abnormality through voting. For example, when the system detects an abnormal increase in the harmonic distortion rate, combined with indicators such as the load rate, it can be judged whether the abnormality is caused by an increase in non-linear loads. When using the decision tree algorithm to analyze the cause of the abnormality, it can start from the root node and analyze layer by layer. Taking the overheating of a transformer as an example, first, it is judged whether there is excessive harmonic, then it is analyzed whether the load rate is abnormal, and finally, it is concluded that it may be a large number of non-linear loads causing harmonic heating. Through this progressive analysis, an abnormal list containing multiple possible causes is generated. Key factors need to be focused on when updating the feature matrix. If it is found that the abnormal harmonic content causes the equipment to overheat, relevant harmonic features need to be extracted with emphasis. The updated feature matrix will contain more harmonic-related indicators, enabling the non-linear mapping model to more accurately reflect the equipment status. This dynamic optimization process can continuously improve the monitoring accuracy of the system and provide a strong guarantee for the safe and stable operation of the equipment.

[0061] Step S104, perform feature-level fusion on the mapped harmonic feature data and the preprocessed non-electrical quantity data to generate a multi-modal feature vector.

[0062] Furthermore, to enhance the correlation between features and semantic representation ability, an advanced multi-feature fusion technology - Grouped Feature Focus Unit (GFF) is introduced. The core idea of GFF is to further strengthen the semantic representation of features by focusing on the spatial context information between different features. First, the low-resolution features are upsampled and the number of channels is adjusted through 1×1 convolution to make it consistent with the number of channels of the high-resolution feature map. Subsequently, the low-resolution feature map and the high-resolution feature map are added element-wise to obtain a first feature map. This process aims to initially integrate the information of features with different resolutions and lay a foundation for subsequent feature interaction. Next, the first feature map is divided into multiple groups along the channel dimension, and feature interaction operations are performed on the features within each group to generate an attention mask that can capture the feature correlation between channels. This attention mechanism can automatically identify and strengthen important features while suppressing redundant information, thereby improving the quality and usability of features. Finally, the attention mask is normalized through the normalization layer of Multilevel Feature Fusion (MFF) to obtain the initial fusion feature map, further enhancing the spatial information of the features. This normalization step can ensure the comparability of different features during the fusion process and avoid biases caused by differences in dimension or numerical range.

[0063] After the initial fusion and enhancement of features, to further reduce the redundant features that may be generated during the feature fusion process while retaining more information about small targets or details, Multilevel Feature Reconstruction Module (MFR) is introduced. The core of MFR lies in the refined processing of feature maps at different stages. First, strong features and weak features are separated from the initial fusion feature map, and strong and weak feature weight information is generated based on these features. By comparing the strong and weak feature weight information, a strong feature attention map and a weak feature attention map can be obtained. Subsequently, the strong feature attention map and the weak feature attention map are cascaded and fused to generate the final feature map. This fused feature map not only contains detailed feature information but also further enhances the feature expression ability through cross-channel information exchange.

[0064] By introducing GFF and MFR technologies, the mapped harmonic feature data can be deeply fused with the preprocessed non-electrical quantity data to generate a multi-modal feature vector. This multi-modal feature vector not only integrates harmonic features but also fuses the information of non-electrical quantity data (such as temperature, humidity, equipment status, etc.), thus being able to more comprehensively reflect the operating state of the power grid. This fusion method not only makes full use of the advantages of different types of data but also optimizes the quality and expression ability of features through advanced technical means, providing more reliable data support for subsequent analysis and decision-making.

[0065] Step S105: Based on the multi-modal feature vectors, screen the high-confidence sensor data sets to obtain the sensor credibility weights and the sensor harmonic feature parameters.

[0066] Further, construct a sensor credibility evaluation model based on the Bayesian network, and judge the credibility weights of each sensor according to the prior knowledge and real-time data. Use the Bayesian network model to establish the conditional probability distribution relationship between the sensor states and the multi-modal feature vectors. According to the prior knowledge and real-time data, calculate the conditional probability values of each sensor. If the conditional probability value is lower than the preset threshold, judge that the sensor credibility is low and update its weight value. According to the weight values, screen the high-confidence sensor data and determine the sensor harmonic feature parameters.

[0067] Step S106: Obtain the optimized harmonic information based on the sensor credibility weights and the sensor harmonic feature parameters.

[0068] Further, adopt a weighted fusion algorithm to fuse and calculate the sensor credibility weight values and the sensor harmonic feature parameters. If the deviation between the harmonic feature parameter of a certain sensor and the fusion result exceeds the preset threshold, judge that the sensor data is abnormal. For the abnormal sensor data, update its credibility weight value and perform the fusion calculation again. Through multiple iterative fusions, obtain the optimized multi-sensor harmonic information.

[0069] Step S107: Optimize the parameters of the non-linear mapping model based on the optimized harmonic information to obtain the optimized non-linear mapping model parameters; determine the correlation between the harmonic features and the device states based on the optimized non-linear mapping model parameters.

[0070] Obtain the optimized harmonic information and extract its feature parameters. Classify the harmonic feature parameters according to the support vector machine model. If the classification result deviates from the preset normal state range by more than the threshold, judge that the device operating state is abnormal. For the abnormal state, use the clustering analysis method to perform pattern recognition on the optimized harmonic information to obtain the pattern recognition result. Through the pattern recognition result, update the classification standard of the support vector machine model to output the final classification result. Input the final classification result into the non-linear mapping model to obtain the optimized non-linear mapping model parameters. Determine the correlation between the harmonic features and the device states based on the optimized non-linear mapping model parameters.

[0071] As a specific implementation of this embodiment, for an operating motor, its current waveform may contain characteristic parameters such as a third harmonic percentage of 15% and a fifth harmonic percentage of 8% after decomposition. These parameters constitute a characteristic vector reflecting the operating state of the device. In a preset classification model, the characteristic parameters under the normal operating state of the device usually have a relatively stable range. For example, when a certain type of transformer is working normally, its third harmonic content should be maintained between 10% and 20%. If it exceeds this range, it may indicate that there is an abnormality in the device. Classification processing is to compare the measured characteristic parameters with these preset ranges. When a large deviation is found between the harmonic characteristics and the preset state, the non-linear mapping model needs to be adjusted. This situation may be caused by factors such as equipment aging and load changes. For example, for a motor that was originally operating normally, its fifth harmonic content suddenly rises to 15%, far exceeding the normal range of 5% to 10%. At this time, the parameter adjustment mechanism needs to be activated. The adaptive algorithm continuously iteratively optimizes the parameters of the mapping model. For example, the least squares method is used to update the mapping relationship according to the newly collected sample data. Suppose it was originally considered that when the fifth harmonic content exceeds 10%, it is judged as abnormal, but under certain working conditions, this threshold may need to be increased to 12% to be more reasonable. The updated correlation relationship can more accurately reflect the current operating characteristics of the device. For example, considering the influence of seasonal temperature changes on the device, the harmonic content is generally 2% to 3% higher in summer than in winter, and this rule will be included in the updated mapping relationship. When finally judging the device state, not only should a single harmonic characteristic be considered, but also the ratio relationship between multiple harmonics should be comprehensively considered. For example, if the third harmonic of a certain device rises to 18%, but the fifth and seventh harmonics remain normal, and the ratio relationship of each harmonic meets the expectation, this situation may still be within the normal range. Through this multi-dimensional analysis, the accuracy of state judgment can be improved and false alarms can be avoided. The final classification result should consider specific factors such as device type and operating environment. For example, for the same type of transformer, the allowable harmonic content range is different under different load rates. The acceptable harmonic level at a load rate of 80% may require key attention during low-load operation. This differential judgment can provide a more targeted state assessment result.

[0072] The above is only a preferred specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed by the present application should be covered by the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A power grid harmonic monitoring method based on multi-sensor fusion, characterized in that: The following steps are involved: Preprocessing the harmonic data and non-electrical quantity data collected by multiple sensors to obtain preprocessed data; Performing dimensionality reduction processing on the pre-processed data to extract harmonic features and equipment status features; Building a nonlinear mapping model based on the harmonic characteristics and the device state characteristics, and obtaining mapped harmonic characteristic data based on the nonlinear mapping model; Performing feature-level fusion on the mapped harmonic feature data and the preprocessed non-electrical quantity data to generate a multimodal feature vector; Filtering a high-confidence sensor data set based on the multimodal feature vector to obtain a sensor credibility weight and a sensor harmonic feature parameter; Obtaining optimized harmonic information based on the sensor credibility weight and the sensor harmonic characteristic parameter; Performing parameter optimization on the nonlinear mapping model based on the optimized harmonic information to obtain optimized nonlinear mapping model parameters; The correlation between the harmonic characteristics and the equipment status is determined based on the optimized nonlinear mapping model parameters.

2. The method according to claim 1, characterized in that: The process of extracting harmonic features and equipment status features includes: Using a principal component analysis method to classify the preprocessed data to obtain standardized preprocessed data and a number of feature vectors; Project the standardized preprocessed data onto several principal component directions to obtain the reduced-dimensional data matrix; Determine harmonic features based on the principal component distribution in the reduced-dimensional data matrix; The device status characteristics are determined by combining the principal components in the reduced-dimensional data matrix and the device operating parameters; wherein the device operating parameters include: device temperature and device vibration.

3. The method according to claim 2, characterized in that The process of constructing a nonlinear mapping model and obtaining mapped harmonic characteristic data based on the nonlinear mapping model includes: Mapping the dimension-reduced data matrix to a high-dimensional feature space using a Gaussian kernel function to obtain a feature matrix; Using a support vector machine algorithm to classify the feature matrix to obtain the equipment operation status; When the equipment is in an abnormal state, the random forest algorithm is used to predict the abnormal state and determine the abnormal type; Based on the abnormality type, a decision tree algorithm is used to analyze the abnormality cause and generate an abnormality cause list; The corresponding harmonic features are extracted based on the abnormal cause list, and the feature matrix is ​​updated to obtain mapped harmonic feature data.

4. The method according to claim 3, characterized in that: The process of performing feature-level fusion of the mapped harmonic feature data and the preprocessed non-electrical quantity data to generate a multimodal feature vector includes: Adding the mapped harmonic characteristic data and the preprocessed non-electrical quantity data at element level to obtain a first characteristic graph; The first feature map is divided into several groups along the channel dimension, and the features in each group are interactively operated to obtain an attention mask that captures the correlation of features between channels; Performing normalization masking on the attention mask to obtain an initial fused feature map; Separating the initial fusion feature map to obtain strong features, weak features and strong and weak feature weight information; Comparing the strong and weak feature weight information to obtain a strong feature attention map and a weak feature attention map; The strong feature attention map and the weak feature attention map are cascaded and fused to generate a multimodal feature vector.

5. The method according to claim 4, characterized in that The process of screening the high-credibility sensor data set based on the multimodal feature vector to obtain the sensor credibility weight includes: A sensor credibility assessment model is constructed using a Bayesian network based on the conditional probability distribution of the device state characteristics and the multimodal feature vector; The sensor credibility assessment model calculates the conditional probability value of the sensor based on prior knowledge and real-time data; The conditional probability value of the sensor is compared with a preset threshold to determine the credibility weight of each sensor.

6. The method according to claim 5, characterized in that The process of optimizing the harmonic information includes: using a weighted fusion algorithm to fuse the sensor credibility weight and the sensor harmonic characteristic parameters to obtain the optimized harmonic information.

7. The method according to claim 6, characterized in that The process of obtaining the optimized nonlinear mapping model parameters includes: Extracting characteristic parameters of the optimized harmonic information to obtain harmonic characteristic parameters; Using a support vector machine model to classify the harmonic characteristic parameters to obtain an initial classification result; Determine the operating status of the device based on the initial classification result; When the equipment is in an abnormal state, the cluster analysis method is used to perform pattern recognition on the optimized harmonic information to obtain the pattern recognition result; Updating the classification criteria of the support vector machine model based on the pattern recognition result, and outputting the final classification result; The final classification result is input into the nonlinear mapping model to obtain optimized nonlinear mapping model parameters.

Images