Power equipment safety partial discharge multi-modal sensing online monitoring method

Abstract

The application relates to the technical field of power equipment safety monitoring, and discloses an online monitoring method for multi-modal sensing of partial discharge safety of power equipment, which collects data through ultrasonic wave, ultrahigh frequency and electromagnetic interference sensors, obtains predicted values of each data at a prediction time through a prediction model, and corrects the predicted values of the ultrasonic wave data in combination with correlation degree analysis; after three-dimensional feature points are clustered, the similarity of different sequence lengths is analyzed, and historical predicted values are obtained in combination with historical discharge characteristic values; finally, partial discharge characteristic predicted values are obtained according to the prediction correction values, the historical predicted values and the similarity, and discharge state monitoring is realized through threshold discrimination. The method fuses multi-modal data, analyzes the correlation degree and the similarity of historical data, improves the accuracy and anti-interference capability of partial discharge monitoring, and can effectively guarantee the safety of power equipment.

Description

Technical Field

[0001] This invention relates to the field of power equipment safety monitoring technology, specifically to a multi-modal sensing online monitoring method for partial discharge in power equipment. Background Technology

[0002] Partial discharge of power equipment is a significant contributing factor to equipment failure during power system operation. Early monitoring and accurate warning of partial discharge are crucial for ensuring the safe and stable operation of the power system. As power equipment develops towards higher voltage and larger capacity, its operating environment becomes increasingly complex, posing numerous challenges to traditional partial discharge monitoring methods.

[0003] In existing technologies, single-mode partial discharge monitoring methods are common, such as relying solely on ultrasonic or ultra-high frequency sensors for data acquisition and analysis. However, these methods, which can only acquire discharge characteristic information in a single dimension, struggle to comprehensively and accurately reflect the true state of partial discharge. They also exhibit weak anti-interference capabilities in complex electromagnetic environments, easily leading to biased or even misjudged monitoring results. Furthermore, traditional methods often lack in-depth analysis of the correlations between different types of data when processing multi-source data, failing to fully utilize the complementary information between ultrasonic data, ultra-high frequency data, and electromagnetic interference data, thus affecting the accuracy and reliability of partial discharge characteristic prediction.

[0004] Existing partial discharge monitoring methods typically employ fixed-time data acquisition strategies. This approach fails to adaptively adjust to the actual operating cycle of power equipment and the fluctuation frequency of historical partial discharge data, potentially leading to low data acquisition efficiency or the omission of crucial information. Regarding predictive model construction, current technologies largely rely on single data sequences for prediction, lacking comprehensive consideration of the correlation between multimodal data and failing to effectively correct the prediction results. Consequently, they are ill-suited to the complex and ever-changing operating conditions of power equipment.

[0005] Traditional methods for utilizing historical data often fail to fully exploit the similarities between data points, and cannot effectively categorize historical data with similar discharge characteristics through cluster analysis or other means. Consequently, it is difficult to leverage the patterns in historical data to improve the accuracy of current discharge state predictions. Furthermore, existing technologies lack effective interference identification and suppression mechanisms when facing external electromagnetic interference, significantly impacting the stability and reliability of the monitoring system. In summary, existing methods for monitoring partial discharge in power equipment have significant shortcomings in multimodal data fusion, anti-interference capabilities, prediction accuracy, and data acquisition efficiency, urgently requiring a more advanced and reliable monitoring method to meet practical engineering needs. Summary of the Invention

[0006] The purpose of this invention is to provide an online monitoring method for multi-modal sensing of partial discharge in power equipment to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a multi-modal sensing online monitoring method for partial discharge safety in power equipment, the method comprising:

[0008] Ultrasonic data, ultra-high frequency data, and electromagnetic interference data of the external environment of power equipment partial discharge are collected using different types of sensors.

[0009] The next moment after the current moment is recorded as the prediction moment. For a preset number of moments before the current moment, the prediction values ​​of ultrasonic data, UHF data, and electromagnetic interference data at the prediction moment are obtained through the prediction model. Based on the correlation between ultrasonic data and UHF data, the correlation between ultrasonic data and electromagnetic interference data, and the prediction values ​​of UHF data and ultrasonic data at the prediction moment, different partial discharge characteristic prediction values ​​are obtained. The prediction values ​​of ultrasonic data at the prediction moment are corrected by combining the respective proportions of the correlation between ultrasonic data and UHF data and the correlation between ultrasonic data and electromagnetic interference data with the different partial discharge characteristic prediction values, and the prediction correction value is obtained.

[0010] The ultrasonic signal, UHF signal, and electromagnetic interference signal at each moment are used to construct three-dimensional feature points, and all three-dimensional feature points are clustered. For the cluster where the current moment is located, the three-dimensional feature points are constructed into different sequences according to the ultrasonic signal, UHF signal, and electromagnetic interference signal. The similarity between three-dimensional feature points under different sequence lengths is obtained by analyzing the sequences of different lengths. Then, the actual discharge feature values ​​are combined to form weights to obtain the historical predicted values ​​of partial discharge characteristics.

[0011] Partial discharge feature prediction values ​​are obtained based on the predicted correction values, historical prediction values, and the similarity between three-dimensional feature points; the discharge state is monitored by threshold discrimination based on the partial discharge feature prediction values.

[0012] Preferably, the method for collecting ultrasonic data, ultra-high frequency data, and electromagnetic interference data of the external environment for partial discharge of power equipment using different types of sensors is as follows:

[0013] Ultrasonic sensors and ultra-high frequency sensors are installed on the surface of the power equipment, and electromagnetic interference sensors are installed in the surrounding environment. Each sensor collects data once every preset time, and a number of data are collected. The data collected at all times are recorded as a sequence, and ultrasonic sequence, ultra-high frequency sequence and electromagnetic interference sequence are obtained respectively.

[0014] Preferably, the method for obtaining different partial discharge characteristic prediction values ​​based on the correlation between ultrasonic data and UHF data, the correlation between ultrasonic data and electromagnetic interference data, and the predicted values ​​of UHF data and ultrasonic data at the prediction time is as follows:

[0015] Calculate the correlation degree between the sequence values ​​of the ultrasonic sequence and the UHF sequence at the same time, and form a sequence of correlation degrees at all times, which is called the audio correlation sequence; calculate the correlation degree between the sequence values ​​of the ultrasonic sequence and the electromagnetic interference sequence at the same time, and form a sequence of correlation degrees at all times, which is called the electromagnetic interference correlation sequence.

[0016] The average value of all sequence values ​​in the audio frequency correlation sequence is recorded as the first correlation degree, and the average value of all sequence values ​​in the acoustic disturbance correlation sequence is recorded as the second correlation degree.

[0017] The predicted values ​​of the second and third partial discharge characteristics are obtained based on the first correlation degree, the second correlation degree, and the predicted values ​​of the UHF sequence and the electromagnetic interference sequence at the prediction time.

[0018] Preferably, the method for obtaining the second and third partial discharge feature prediction values ​​based on the first correlation degree, the second correlation degree, and the predicted values ​​of the UHF sequence and the electromagnetic interference sequence at the prediction time is as follows:

[0019] The predicted value of the UHF sequence at the prediction time is multiplied by the first correlation degree to obtain the second partial discharge characteristic prediction value, and the predicted value of the electromagnetic interference sequence at the prediction time is multiplied by the second correlation degree to obtain the third partial discharge characteristic prediction value.

[0020] Preferably, the method for obtaining the prediction correction value by combining the respective proportions of the correlation between ultrasonic data and UHF data, and the correlation between ultrasonic data and electromagnetic interference data, with different partial discharge characteristic prediction values ​​to correct the predicted value of ultrasonic data at the prediction time is as follows:

[0021] Calculate the standard deviation of all sequence values ​​in the audio frequency correlation sequence, and use the reciprocal of the standard deviation as the correlation between the ultrasonic sequence and the UHF sequence;

[0022] Calculate the standard deviation of all sequence values ​​in the acoustic interference correlation sequence, and use the reciprocal of the standard deviation as the correlation between the ultrasonic sequence and the electromagnetic interference sequence;

[0023] The predicted value of the ultrasonic data obtained through the prediction model at the prediction time is recorded as the first partial discharge characteristic prediction value.

[0024] A weighted value is preset for the first partial discharge feature prediction value, and the weighted values ​​of the second and third partial discharge feature prediction values ​​are obtained based on the correlation ratio.

[0025] The predicted correction value is obtained by weighting and summing the predicted values ​​of the first partial discharge feature, the second partial discharge feature, the third partial discharge feature, and their respective weighted values.

[0026] Preferably, the method for constructing different sequences of three-dimensional feature points within the current cluster based on ultrasonic signals, ultra-high frequency signals, and electromagnetic interference signals, analyzing the similarity between three-dimensional feature points of different sequence lengths, and then combining the actual discharge feature values ​​to construct weights to obtain historical predicted values ​​of partial discharge features is as follows:

[0027] The three-dimensional feature point corresponding to the current moment is recorded as the target feature point. The cluster in which the target feature point belongs is obtained, and all feature points in this cluster are recorded as class feature points.

[0028] The ultrasonic, UHF, and electromagnetic interference sequences corresponding to the feature points are obtained by selecting ultrasonic, UHF, and electromagnetic interference signals at preset times.

[0029] The similarity between 3D feature points is obtained by analyzing the similarity between the corresponding sequences of class feature points and target feature points, and similar feature points are determined.

[0030] Historical predicted values ​​of partial discharge characteristics are obtained by using the discharge characteristic values ​​of similar feature points corresponding to different sequence lengths at the next time step.

[0031] Preferably, the method for obtaining the similarity between three-dimensional feature points based on the similarity of the corresponding sequences of class feature points and target feature points, and determining similar feature points, is as follows:

[0032] For each class of feature points, calculate the similarity between its corresponding ultrasonic sequence, UHF sequence, and electromagnetic interference sequence and the ultrasonic sequence, UHF sequence, and electromagnetic interference sequence at the current moment;

[0033] The similarity between the class feature points and the target feature points is calculated by averaging the similarity of the three sequences and then linearly normalized to obtain the similarity between the class feature points and the target feature points.

[0034] Obtain the similarity between all class feature points and the target feature point. For the similarity between all class feature points and the target feature point, classify all similarities using a threshold segmentation algorithm. Obtain the mean similarity of each class after classification. Record the class feature points in the class with the larger mean similarity as similar feature points.

[0035] Preferably, the method for obtaining historical predicted values ​​of partial discharge characteristics based on the discharge characteristic values ​​of similar feature points corresponding to different sequence lengths at the next time step is as follows:

[0036] For each sequence length, the mean of the discharge characteristic values ​​of all similar feature points at the next adjacent time step is used as the prediction reference value for the prediction time step; for this sequence length, the variance of the discharge characteristic values ​​of all similar feature points at the next time step is calculated; the reciprocal of the variance is used as the reliability parameter; the ratio of each reliability parameter to the sum of all reliability parameters is used as the period weight.

[0037] Historical predicted values ​​are obtained by summing the predicted reference values ​​for all sequence lengths using period weights.

[0038] Preferably, the method for obtaining the predicted value of partial discharge features based on the predicted correction value, historical predicted values, and the similarity between three-dimensional feature points is as follows:

[0039] Get all similar feature points for the current sequence length, and get the predicted correction value and historical prediction value of the similar feature points;

[0040] The predicted correction value and historical prediction value of each similar feature point are subtracted from the actual discharge feature value at the next moment to form the correction difference sequence and the historical difference sequence.

[0041] For each differential sequence, the product of the variance of the differential sequence and the mean of all sequence values ​​in the differential sequence is used as the weighting factor;

[0042] The ratio of the weight factor corresponding to the corrected difference sequence to the sum of the weight factors is used as the weight coefficient of the predicted corrected value, and the ratio of the weight factor corresponding to the historical difference sequence to the sum of the weight factors is used as the weight coefficient of the historical predicted value.

[0043] The predicted value and the historical predicted value are weighted and added together using their respective weighting coefficients to obtain the predicted value of partial discharge characteristics.

[0044] Preferably, the method for determining the preset time in the process of each sensor collecting data once every preset time is as follows: based on the operating cycle of the power equipment and the fluctuation frequency of historical partial discharge data, the optimal value of the data collection time interval is obtained through statistical analysis; the standard collection cycle corresponding to the current equipment type and the optimal value are averaged to obtain the final preset time.

[0045] Compared with the prior art, the beneficial effects of the present invention are:

[0046] By employing different types of sensors to simultaneously acquire ultrasonic data, ultra-high frequency data, and electromagnetic interference data, collaborative monitoring of multi-source heterogeneous data was achieved, enhancing the comprehensive capture capability of partial discharge characteristics. The fusion analysis of multimodal data can fully utilize the complementary information between different data types, effectively overcoming the limitations of single-modal monitoring, thereby improving the accuracy and reliability of partial discharge monitoring.

[0047] In terms of data prediction and correction, a prediction model is constructed to obtain the predicted values ​​of various types of data at the prediction time. The prediction is then corrected by combining the correlation between ultrasonic data, UHF data, and electromagnetic interference data. This approach not only considers the changing trends of the data itself but also delves into the intrinsic relationships between different data sequences. This correlation-based correction mechanism effectively suppresses the influence of external electromagnetic interference, improves the anti-interference capability of the prediction results, and enables the monitoring system to maintain stable operation even in complex electromagnetic environments.

[0048] The introduction of cluster analysis and similarity calculation enables efficient utilization of historical data. By clustering three-dimensional feature points and analyzing the similarity under different sequence lengths, patterns similar to the current discharge state can be extracted from historical data. These patterns are then combined with actual discharge feature values ​​to form weights for obtaining historical prediction values. This fully leverages the guiding role of historical experience in current predictions, further improving the accuracy and reliability of partial discharge feature prediction.

[0049] Regarding the data acquisition strategy, the optimal data acquisition time interval is determined through statistical analysis based on the operating cycle of power equipment and the fluctuation frequency of historical partial discharge data. This interval is then optimized in conjunction with the standard acquisition cycle corresponding to the equipment type, achieving adaptive adjustment of data acquisition. This approach ensures both the integrity and validity of the data while improving data acquisition efficiency and reducing the system's operating load.

[0050] Furthermore, by combining the predicted correction values, historical prediction values, and the similarity between three-dimensional feature points, a more scientific and reasonable partial discharge characteristic prediction model was constructed. This model can comprehensively consider the prediction correction results of current data and the patterns of historical data, and through dynamic adjustment of weight factors, it achieves accurate prediction of partial discharge characteristics, providing a more reliable basis for the condition assessment and fault early warning of power equipment. Attached Figure Description

[0051] Figure 1 This is a schematic diagram illustrating the working principle of the online monitoring method for multi-modal partial discharge safety of power equipment described in this invention.

[0052] Figure 2 A schematic diagram illustrating the working principle of obtaining the predicted correction value by correcting the predicted value of ultrasonic data at the predicted time.

[0053] Figure 3 A schematic diagram illustrating the working principle for obtaining historical predicted values ​​of partial discharge characteristics;

[0054] Figure 4 This diagram illustrates the working principle of obtaining historical predicted values ​​based on the discharge characteristic values ​​of similar feature points. Detailed Implementation

[0055] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0056] Please see Figures 1-4 The present invention relates to a multi-modal online monitoring method for partial discharge safety of power equipment, the specific implementation steps of which are as follows:

[0057] Ultrasonic data, ultra-high frequency data, and electromagnetic interference data of the external environment for partial discharge of power equipment are collected using different types of sensors.

[0058] The next moment after the current moment is designated as the prediction moment. For a preset number of moments before the current moment, the predicted values ​​of ultrasonic data, UHF data, and electromagnetic interference data at the prediction moment are obtained through the prediction model. Based on the correlation between ultrasonic data and UHF data, the correlation between ultrasonic data and electromagnetic interference data, and the predicted values ​​of UHF data and ultrasonic data at the prediction moment, different predicted values ​​of partial discharge characteristics are obtained. The predicted values ​​of ultrasonic data at the prediction moment are corrected by combining the respective proportions of the correlation between ultrasonic data and UHF data and the correlation between ultrasonic data and electromagnetic interference data with the different predicted values ​​of partial discharge characteristics, and the corrected prediction value is obtained.

[0059] The ultrasonic signal, UHF signal, and electromagnetic interference signal at each moment are used to construct three-dimensional feature points, and all three-dimensional feature points are clustered. For the cluster where the current moment is located, the three-dimensional feature points are used to construct different sequences based on the ultrasonic signal, UHF signal, and electromagnetic interference signal. The similarity between three-dimensional feature points under different sequence lengths is obtained by analyzing the sequences of different lengths. Then, the actual discharge feature values ​​are combined to form weights to obtain the historical predicted values ​​of partial discharge features.

[0060] Partial discharge feature prediction values ​​are obtained based on the predicted correction values, historical prediction values, and the similarity between three-dimensional feature points; the discharge state is monitored by threshold discrimination based on the partial discharge feature prediction values.

[0061] Example 1:

[0062] When implementing this online monitoring method for partial discharge in power equipment using multimodal sensing, the specific implementation method for collecting ultrasonic data, ultra-high frequency (UHF) data, and electromagnetic interference (EMI) data of partial discharge from the power equipment using different types of sensors is as follows. First, ultrasonic sensors and UHF sensors need to be installed on the surface of the power equipment, while EMI sensors are installed in the surrounding environment. The placement of these sensors is crucial. Ultrasonic sensors should be closely fitted to the surface of the power equipment to ensure effective capture of ultrasonic signals generated by partial discharge, avoiding signal attenuation or distortion due to improper installation. The placement of UHF sensors must consider the structural characteristics of the power equipment, selecting a location that optimally receives UHF signals from partial discharge. They are typically installed near insulating components or areas where discharge may occur to ensure the validity of the collected UHF data. The placement of EMI sensors needs to consider the distribution of EMI sources in the surrounding environment. They should be kept away from strong EMI sources while still comprehensively monitoring the EMI environment around the equipment. They are generally placed at a certain distance from the equipment and can cover the main EMI areas around it.

[0063] Each sensor collects data once every preset time interval, collecting a certain number of data points. This preset time interval is not fixed but needs to be determined based on the actual situation. The specific determination method is as follows: First, the operating cycle of the power equipment must be fully considered. Different types of power equipment have different operating cycles; for example, the operating cycles of transformers and high-voltage cables may differ, requiring detailed analysis of the specific equipment's operating cycle. Simultaneously, the fluctuation frequency of historical partial discharge data must be analyzed. Through statistical analysis and research of a large amount of historical data, the temporal variation patterns and fluctuation characteristics of partial discharge data can be understood. Based on the operating cycle of the power equipment and the fluctuation frequency of historical partial discharge data, statistical analysis methods, such as time series analysis and frequency distribution analysis, are used to obtain the optimal value of the data acquisition interval. This optimal value ensures that the sensor collects data within a suitable time interval, avoiding both missed important discharge information due to excessively long intervals and excessively short intervals that generate a large amount of redundant data, increasing the burden on data processing.

[0064] After obtaining the optimal data acquisition interval, it is also necessary to consider the standard acquisition cycle corresponding to the current equipment type. Different types of power equipment typically have corresponding standard acquisition cycles for their respective industries or enterprises, which are formulated based on the characteristics of the equipment and safety operation requirements. The average value obtained earlier, calculated by averaging the standard acquisition cycle corresponding to the current equipment type with this average value, becomes the final preset time. Determining the preset time in this way ensures that the equipment's standard requirements are met while also taking into account the actual operating conditions and historical data characteristics, making the preset time more reasonable and scientific.

[0065] After a preset time is determined, the sensor collects data according to that preset time. Each time data is collected, the sensor converts the collected signal into an electrical signal and performs preliminary amplification and filtering to improve data quality. The collected data is temporarily stored in the sensor's built-in buffer or transmitted in real time to the data processing system via a data transmission line.

[0066] After collecting a number of data points, these data need to be processed. The data from all collected moments are recorded as a sequence, and ultrasonic, ultra-high frequency (UHF), and electromagnetic interference (EMI) sequences are obtained separately. Specifically, for the data collected by the ultrasonic sensor, the ultrasonic data collected at each moment are arranged in chronological order to form an ultrasonic sequence; similarly, the data collected by the UHF sensor are arranged in chronological order to form an UHF sequence; and the data collected by the EMI sensor are arranged in chronological order to form an EMI sequence.

[0067] During the entire data acquisition process, the following points should be noted: First, ensure the normal operation of the sensors. Regularly calibrate and maintain the sensors, checking their performance and identifying any faults or anomalies. Second, ensure the stability of data transmission, avoiding data loss, errors, or delays during transmission. Reliable data transmission protocols and media should be used. Third, monitor the acquired data in real time, observing trends and promptly identifying and addressing any anomalies. Fourth, establish a comprehensive data storage and management system, classifying and storing all acquired data for easy retrieval and analysis later.

[0068] The detailed implementation methods described above ensure accurate and reliable acquisition of ultrasonic data, ultra-high frequency data, and electromagnetic interference data from the external environment related to partial discharge in power equipment using different types of sensors. This provides a high-quality data foundation for subsequent partial discharge characteristic prediction and discharge status monitoring. This data acquisition method fully considers the actual conditions of the equipment and the characteristics of the data, possessing strong practicality and scientific validity, and can meet the monitoring needs for the safe operation of power equipment.

[0069] Example 2:

[0070] After acquiring ultrasonic, UHF, and electromagnetic interference data, these data need to be processed in depth to obtain predicted values ​​for different partial discharge characteristics. Specifically, correlation calculations are performed on the ultrasonic and UHF sequences. The ultrasonic sequence is composed of ultrasonic data of partial discharge from power equipment collected by an ultrasonic sensor at preset times, arranged chronologically. Similarly, the UHF sequence is the corresponding data sequence collected by the UHF sensor. During the calculation, correlation analysis is performed on the sequence values ​​at the same time point for each of the two sequences. For example, for the nth acquisition time, the nth data point in the ultrasonic sequence and the nth data point in the UHF sequence are extracted, and the correlation degree between the two data points at that time is determined using a specific correlation calculation method (such as a time-series-based correlation analysis method). The correlation calculation results for all times (from the first acquisition time to the current time) are arranged chronologically to form a new sequence, denoted as the audio frequency correlation sequence.

[0071] The correlation between ultrasonic sequences and electromagnetic interference (EMI) sequences is calculated. The EMI sequence is formed by arranging data collected by EMI sensors in the surrounding environment of the equipment in chronological order. Using the same correlation calculation logic as described above, the correlation between ultrasonic sequence values ​​and EMI sequence values ​​at the same time is calculated, and the calculation results at all times are used to construct an acoustic interference correlation sequence. The construction of these two correlation sequences can intuitively reflect the changes in the correlation characteristics between ultrasonic data and UHF data and EMI data in the time dimension.

[0072] Statistical analysis was performed on the audio frequency correlation sequence and the acoustic interference correlation sequence to obtain key parameters. For the audio frequency correlation sequence, the average value of all sequence values ​​was calculated, and this average value was denoted as the first correlation degree. Specifically, the correlation degree values ​​at each time point in the audio frequency correlation sequence were summed and then divided by the total length of the sequence (i.e., the total number of acquisition times). The resulting average value characterizes the overall correlation between the ultrasonic sequence and the UHF sequence. Similarly, the average value of all sequence values ​​in the acoustic interference correlation sequence was calculated, and the resulting average value was denoted as the second correlation degree, which reflects the overall correlation level between the ultrasonic sequence and the electromagnetic interference sequence.

[0073] After obtaining the first and second correlation scores, the predicted values ​​of the UHF sequence and the electromagnetic interference sequence at the prediction time are combined to further calculate the predicted values ​​of the second and third partial discharge characteristics. Here, the prediction time refers to the time following the current time. The predicted values ​​of the UHF sequence and the electromagnetic interference sequence at the prediction time are obtained by processing data from a predetermined number of time points prior to the current time using a pre-established prediction model. For example, the prediction model might employ a time series prediction algorithm from machine learning, predicting the UHF data value at the prediction time based on the time series characteristics of historical UHF sequence data; similarly, the predicted value of the electromagnetic interference sequence at the prediction time is obtained.

[0074] The specific calculation logic is as follows: the predicted value of the UHF sequence at the prediction time is multiplied by the first correlation degree, and the product is the predicted value of the second partial discharge feature. The principle behind this calculation is that the first correlation degree reflects the correlation strength between the ultrasonic data and the UHF data. When the predicted value of the UHF data at the prediction time is known, the predicted value of the ultrasonic feature associated with that UHF predicted value can be inferred through the correlation degree. Similarly, the predicted value of the electromagnetic interference sequence at the prediction time is multiplied by the second correlation degree to obtain the predicted value of the third partial discharge feature, which reflects the correlation prediction result between the predicted value of the electromagnetic interference data and the ultrasonic data feature.

[0075] Throughout the implementation process, the following key points should be noted: First, the choice of correlation calculation method must be combined with the data characteristics to ensure accurate reflection of the correlation between different sequences. For example, if the data exhibits linear correlation characteristics, methods such as the Pearson correlation coefficient can be used; if nonlinear correlation exists, more complex correlation calculation methods may be required. Second, the establishment of the prediction model requires training and optimization based on a large amount of historical data to ensure the accuracy of the predicted values ​​for UHF sequences and electromagnetic interference sequences. The quality and quantity of historical data directly affect the performance of the prediction model; therefore, it is necessary to ensure the integrity and reliability of the historical data. Furthermore, when calculating the average value, all sequence values ​​must be accurately processed to avoid deviations in the first and second correlation degrees due to data omissions or calculation errors.

[0076] Meanwhile, data time alignment is a crucial step. When calculating the correlation between sequence values ​​at the same time, it is essential to ensure that the timestamps of the ultrasonic sequence, UHF sequence, and electromagnetic interference sequence strictly correspond to avoid distortion of the correlation calculation results due to time misalignment. In practice, synchronous acquisition mechanisms or timestamp calibration algorithms can be used to ensure the time consistency of data collected by each sensor.

[0077] For the predicted value at the predicted time, the specific value of the preset number of time points needs to be clearly defined. The setting of this number must comprehensively consider the temporal relevance of the data and computational efficiency. If the preset number is too small, it may be impossible to capture sufficient historical data features; if it is too large, it will increase computational complexity. A reasonable preset number can usually be determined through experiments or analysis based on equipment operating characteristics.

[0078] Example 3:

[0079] After obtaining predicted values ​​of different partial discharge characteristics, the predicted values ​​of the ultrasonic data at the predicted time need to be corrected to obtain predicted correction values ​​that better reflect the actual situation. Specifically, the standard deviation of the audio-frequency correlation sequence is calculated first. The audio-frequency correlation sequence is a sequence composed of the correlation between the ultrasonic sequence and the UHF sequence at the same time. Each value in this sequence reflects the degree of correlation between the two sequence data at the corresponding time. To calculate the standard deviation, the average value of all sequence values ​​in the audio-frequency correlation sequence is first calculated, then the difference between each sequence value and the average value is calculated, these differences are squared and averaged, and finally the square root of this average value is taken. The result is the standard deviation of the audio-frequency correlation sequence. Since the standard deviation reflects the dispersion of the sequence values, the smaller the standard deviation, the more concentrated the sequence values ​​are, and vice versa. The reciprocal of this standard deviation is used as the correlation between the ultrasonic sequence and the UHF sequence. After this processing, the smaller the standard deviation, the greater the correlation, indicating that the correlation between the two sequences is more stable.

[0080] Then, the standard deviation of the acoustic interference correlation sequence is calculated. The acoustic interference correlation sequence is a sequence composed of the correlation between the ultrasonic sequence and the electromagnetic interference sequence at the same time. Using the same standard deviation calculation steps, the standard deviation of the acoustic interference correlation sequence is obtained, and its reciprocal is taken as the correlation between the ultrasonic sequence and the electromagnetic interference sequence. The calculation of these two correlations reflects the correlation stability between the ultrasonic data and the other two types of data from different dimensions.

[0081] The predicted value of the first partial discharge characteristic is obtained. This predicted value is obtained by processing the ultrasonic data through a prediction model. Specifically, the prediction model is based on the ultrasonic data sequence of a preset number of time steps prior to the current time. By analyzing the time series characteristics, trend changes, and periodic patterns of historical data, the model predicts the ultrasonic data value at the next time step (i.e., the time step after the current time). The result is recorded as the predicted value of the first partial discharge characteristic. The selection of the prediction model should be combined with the characteristics of the data. Models such as autoregressive models, moving average models, or more complex machine learning prediction algorithms can be used to ensure that the prediction results can reflect the changing trend of the ultrasonic data as accurately as possible.

[0082] After obtaining the first partial discharge feature prediction value, a weighting value needs to be preset for it. The setting of this weighting value should comprehensively consider factors such as the historical performance of the prediction model and the stability of the data. It can usually be initially set based on actual application scenarios and experience, and then adjusted and optimized through subsequent data verification. Simultaneously, based on the previously calculated correlation between the ultrasonic sequence and the UHF sequence, and the correlation between the ultrasonic sequence and the electromagnetic interference sequence, the ratio between the two is calculated. For example, assuming the correlation between ultrasonic and UHF is A, and the correlation with electromagnetic interference is B, then the correlation ratio is A / B (or B / A, depending on the setting of the calculation logic). Based on this ratio, the weighting values ​​for the second and third partial discharge feature prediction values ​​are determined. Here, the second partial discharge feature prediction value is the product of the UHF sequence prediction value and the first correlation, and the third partial discharge feature prediction value is the product of the electromagnetic interference sequence prediction value and the second correlation. Their weighting values ​​need to be allocated according to the correlation ratio to reflect the weight differences of each prediction value under different correlations.

[0083] The predicted values ​​of the first, second, and third partial discharge characteristics, along with their corresponding weighted values, are weighted and summed to obtain the prediction correction value. Specifically, each predicted value is multiplied by its corresponding weighted value, and all products are summed to obtain the prediction correction value. This process, by comprehensively considering prediction values ​​from different sources and their weights, corrects the predicted values ​​of the ultrasonic data, making the results closer to the actual discharge characteristics.

[0084] The following key points should be noted during implementation: The calculation of standard deviation must ensure data accuracy. If outliers exist in the audio frequency correlation sequence or acoustic interference correlation sequence, data cleaning and preprocessing are necessary to avoid significant impacts on the standard deviation calculation results. For example, statistical methods can be used to identify and remove outlier correlation values ​​that significantly deviate from the normal range before calculating the standard deviation. Training and optimization of the prediction model are crucial. A large amount of historical data should be used to train the model, and model parameters should be adjusted through methods such as cross-validation to improve the accuracy of the predicted first partial discharge characteristic. The time span and coverage of the historical data must be sufficiently broad to ensure that the model can capture the ultrasonic data characteristics under different operating conditions.

[0085] Furthermore, the mapping relationship between the correlation ratio and the weighting value needs to be set appropriately. For example, if the correlation ratio is large, it indicates a closer correlation between ultrasound and UHF data, and the weighting value of the predicted second partial discharge characteristic can be increased accordingly to highlight the influence of this predicted value. This mapping relationship can be achieved by establishing rules or functions to ensure that the allocation of weighting values ​​can truly reflect the degree of correlation between data. At the same time, the adjustment of the preset weighting value needs to be combined with the actual application effect. By comparing the deviation between the predicted value and the actual discharge characteristic value before and after correction, the setting of the weighting value can be gradually optimized to improve the accuracy of prediction correction.

[0086] The time synchronization of data is also crucial throughout the process. The construction of audio correlation sequences and acoustic interference correlation sequences, as well as the input data for the prediction model, all require consistent timestamps to avoid deviations in correlation calculations and prediction results due to time misalignment. In practice, a unified clock synchronization mechanism or a time alignment algorithm during data preprocessing can ensure the time correspondence of each sequence of data.

[0087] Example 4:

[0088] After constructing three-dimensional feature points from the ultrasonic signals, UHF signals, and electromagnetic interference signals at each moment and completing clustering, a deep analysis is needed for the cluster to which the current moment belongs to obtain historical predicted values ​​of partial discharge characteristics. In practice, the relationship between the target feature point and the cluster must first be clarified. The three-dimensional feature point corresponding to the current moment is denoted as the target feature point, which is composed of the ultrasonic signal value, UHF signal value, and electromagnetic interference signal value at the current moment. After clustering all three-dimensional feature points using clustering algorithms (such as K-means, DBSCAN, etc.), the target feature point will be assigned to a specific cluster. All feature points within this cluster are denoted as class feature points, which have high similarity to the target feature point in the feature space.

[0089] A corresponding signal sequence is constructed for each class of feature points. For each class of feature points, ultrasonic, UHF, and electromagnetic interference signals at preset times are selected and arranged in chronological order to form the ultrasonic, UHF, and electromagnetic interference sequences corresponding to the class of feature points. The selection of preset times needs to comprehensively consider the temporal correlation of the data and computational efficiency. Typically, several historical times adjacent to the current time can be selected, or a reasonable time window can be determined based on the time decay characteristics of the equipment discharge characteristics to ensure that the signals at the selected times can effectively reflect the historical change patterns of the class of feature points.

[0090] The similarity between the sequences corresponding to class-specific feature points and target feature points is calculated to determine similar feature points. Specifically, for each class-specific feature point, the similarity is calculated between its corresponding ultrasonic sequence and the current ultrasonic sequence of the target feature point, its UHF sequence and the current UHF sequence of the target feature point, and its electromagnetic interference sequence and the current electromagnetic interference sequence of the target feature point. Similarity calculation methods can include Dynamic Time Warping (DTW), Euclidean distance, and cosine similarity, with the most suitable method selected based on the characteristics of the sequence. For example, if the sequences have temporal misalignment but similar shapes, the DTW algorithm can more accurately measure similarity; if the sequence is a standardized numerical sequence, Euclidean distance or cosine similarity can be calculated more efficiently.

[0091] After obtaining the similarity of three sequences between each class feature point and the target feature point, the average of these three similarities is calculated to obtain the comprehensive similarity between the class feature point and the target feature point. In order to map the similarity values ​​to a uniform scale, the comprehensive similarity needs to be linearly normalized so that its value range is usually normalized to [0,1], which facilitates subsequent comparison and classification.

[0092] After obtaining the normalized similarity between all class feature points and the target feature points, these similarities are classified using a threshold segmentation algorithm. The threshold segmentation algorithm can employ Otsu's method, adaptive thresholding, or a manually set threshold based on domain knowledge. The purpose of classification is to distinguish class feature points with high similarity from those with low similarity. Specifically, all similarity values ​​are compared to a threshold; values ​​greater than the threshold are classified into one class, and values ​​less than the threshold into another. The mean similarity value for each class is calculated, and the class feature points in the class with the highest mean similarity value are recorded as similar feature points. These feature points have stronger similarity to the target feature points in the feature space, and their corresponding discharge characteristics are more valuable for reference.

[0093] After identifying similar feature points, historical predicted values ​​need to be obtained based on the discharge feature values ​​of similar feature points under different sequence lengths. For each preset sequence length (e.g., 10 time points, 20 time points, etc.), the discharge feature values ​​of all similar feature points under that sequence length at the next adjacent time point are first collected. For example, if the sequence length is L, then for each similar feature point, its signal sequence from the current time point backwards by L time points is taken, and the actual discharge feature value of that similar feature point at time point L+1 is obtained. The average of these discharge feature values ​​is then calculated to obtain the predicted reference value under that sequence length.

[0094] After obtaining the predicted reference value, the variance of the discharge characteristic values ​​of all similar feature points within the sequence length at the next time step is calculated. Variance reflects the dispersion of the data; the smaller the variance, the more concentrated the discharge characteristic values ​​of similar feature points, and the higher the reliability of the predicted reference value. The reciprocal of the variance is used as the reliability parameter, thus the reliability parameter is inversely proportional to the variance; the smaller the variance, the larger the reliability parameter.

[0095] The ratio of the reliability parameter for each sequence length to the sum of the reliability parameters for all sequence lengths is calculated. This ratio serves as the period weight, which measures the weight of the predicted reference value for that sequence length. Finally, the predicted reference values ​​for all sequence lengths are multiplied by their corresponding period weights and summed to obtain the historical predicted values ​​of the partial discharge characteristics. This weighted summation method fully utilizes the prediction information for different sequence lengths, automatically assigning weights based on reliability, thus making the historical predicted values ​​more reasonable.

[0096] The following key points should be noted during implementation: The selection and parameter adjustment of the clustering algorithm directly affect the quality of cluster division. A suitable clustering algorithm should be selected based on the distribution characteristics of the 3D feature points, and clustering parameters should be optimized using indicators such as the silhouette coefficient to ensure the rationality of the cluster feature points. The accuracy of the similarity calculation method is crucial. The similarity measurement method should be selected or adjusted according to the characteristics of different signal sequences (such as the pulse characteristics of ultrasound and the broadband characteristics of ultra-high frequency signals) to avoid misjudging similar feature points due to inappropriate methods.

[0097] The threshold setting for the threshold segmentation algorithm needs to be combined with the characteristics of the device's historical discharge data and the similarity distribution. The optimal threshold can be determined through multiple experiments or statistical analysis to ensure that similar feature points can fully reflect the characteristics of the target feature points without including too many irrelevant feature points. The selection of the preset sequence length needs to cover the possible periodic characteristics of the device's discharge. Multiple different sequence lengths, such as short period, medium period, and long period, can be set according to the device type and the fluctuation period of historical data to capture the discharge patterns at different time scales.

[0098] Furthermore, data preprocessing is indispensable. This involves denoising, filtering, and normalizing ultrasonic signals, UHF signals, and electromagnetic interference signals to prevent noise and outliers from interfering with feature point construction, similarity calculation, and subsequent prediction results. In practical applications, it is also necessary to regularly update clusters and similar feature points. As new data is continuously collected, the clustering results and similarity calculations should be adjusted in a timely manner to ensure that historical prediction values ​​reflect changes in the equipment's discharge state in real time.

[0099] Example 5:

[0100] After obtaining the predicted correction value and historical predicted value, these two types of predicted values ​​need to be fused by combining the similarity between three-dimensional feature points to obtain the final partial discharge feature prediction value, thereby achieving effective monitoring of the discharge state. In specific implementation, it is necessary to first determine all similar feature points under the current sequence length. The current sequence length here is a specific length set when analyzing different sequence lengths in Example 4, and needs to be selected according to actual processing needs. For example, a sequence length with higher weight in the calculation of historical predicted values ​​can be selected. The process of determining similar feature points needs to be based on the clustering analysis and similarity calculation method described in Example 4, that is, by calculating the sequence similarity between the class feature points and the target feature points, and after threshold segmentation, a set of feature points with high similarity to the target feature points is obtained. These feature points are the similar feature points under the current sequence length.

[0101] After obtaining similar feature points, it is necessary to extract the predicted correction value and historical predicted value corresponding to each similar feature point. The predicted correction value is obtained using the correction method described in Example 3, which is the result of weighted correction of the predicted values ​​of the ultrasonic data based on the correlation between the ultrasonic sequence and UHF and electromagnetic interference sequences. The historical predicted value is obtained by weighted summation of the historical discharge characteristic values ​​of similar feature points using the historical predicted value calculation method described in Example 4. These data must correspond one-to-one with the similar feature points to ensure the accuracy and correlation of the data.

[0102] The differences between the predicted correction value, historical predicted value, and actual discharge characteristic value at the next time step are calculated for similar feature points. For each similar feature point, the difference between its predicted correction value and the actual discharge characteristic value at the next time step is used to obtain a correction difference value; similarly, the difference between its historical predicted value and the actual discharge characteristic value at the next time step is used to obtain a historical difference value. The correction difference values ​​of all similar feature points are arranged in order to form a correction difference sequence; all historical difference values ​​form a historical difference sequence. These two difference sequences reflect the degree of deviation between the predicted correction value and the historical predicted value and the actual discharge characteristic value.

[0103] After obtaining the variance sequences, the weighting factor for each sequence needs to be calculated. For the corrected variance sequence, first calculate the variance of the sequence, which reflects the dispersion of the corrected variance values, i.e., the deviation and fluctuation between the predicted and actual values. Then calculate the average of all variance values ​​in the sequence, which reflects the overall deviation trend of the predicted and corrected values. Multiply the variance and the average, and the product is used as the weighting factor for the corrected variance sequence. Similarly, perform the same calculation for the historical variance sequences to obtain their weighting factors. The magnitude of the weighting factor comprehensively reflects the degree of deviation and fluctuation characteristics between the corresponding predicted and actual values, and is used for subsequent weight allocation.

[0104] Calculate the weight coefficients for the predicted correction value and the historical prediction value. The weight factor of the corrected difference sequence is compared to the sum of the weight factors of the corrected difference sequence and the historical difference sequence; the ratio obtained is used as the weight coefficient of the predicted correction value. Similarly, the weight factor of the historical difference sequence is compared to the sum of the weight factors of the two sequences to obtain the weight coefficient of the historical prediction value. The weight coefficients are set according to the principle of "the smaller the deviation, the greater the weight." That is, if the weight factor of the corrected difference sequence is small, it indicates that the overall deviation of the predicted correction value from the actual value is small and the fluctuation is not significant; therefore, its weight coefficient is larger, and vice versa.

[0105] Finally, the predicted correction value and the historical prediction value are weighted and summed using their respective weighting coefficients to obtain the predicted partial discharge characteristic value. Specifically, the predicted correction value is multiplied by its weighting coefficient, and the historical prediction value is multiplied by its weighting coefficient; the sum of these two products is the final predicted partial discharge characteristic value. This predicted value combines the advantages of both the predicted correction value and the historical prediction value, and dynamically adjusts the weights based on the deviations between the two and the actual values, making the result closer to the actual discharge characteristics.

[0106] After obtaining the predicted values ​​of partial discharge characteristics, threshold discrimination is required to monitor the discharge status. Different threshold ranges for different discharge states are pre-defined, such as normal state threshold, warning state threshold, and fault state threshold. The calculated predicted values ​​of partial discharge characteristics are compared with these thresholds: if the predicted value is less than the normal state threshold, the equipment is determined to be in a normal discharge state; if the predicted value is between the normal state threshold and the warning state threshold, the equipment is determined to be in a state of potential discharge risk, requiring enhanced monitoring; if the predicted value is greater than the warning state threshold, the equipment may have abnormal discharge, requiring further investigation of the fault.

[0107] The following points should be noted during implementation: The accuracy of similar feature points directly affects the calculation of the difference sequence. Therefore, it is necessary to ensure that the screening process for similar feature points in Example 4 is rigorous to avoid including irrelevant feature points in the calculation, which would lead to distortion of the difference sequence. The acquisition of actual discharge characteristic values ​​needs to be achieved through reliable measurement methods to ensure the authenticity and timeliness of the data. Calibration can be performed using high-precision discharge measurement instruments or by combining the measurement results of multiple sensors.

[0108] The processing of differential sequences requires attention to the impact of outliers. If significant outliers exist in the differential sequence or historical differential sequences, data cleaning is necessary first. For example, outliers can be identified and removed using the 3σ principle to avoid them significantly interfering with the calculation of variance and mean, thus affecting the accuracy of weighting factors and weighting coefficients. The calculation logic of weighting factors and weighting coefficients must be consistent to ensure the rationality of weight allocation. For example, when calculating weighting factors, a larger product of variance and mean indicates a greater overall deviation and volatility of the corresponding predicted value, and its weighting coefficient should be smaller; conversely, a smaller product should be larger.

[0109] Threshold setting needs to consider equipment type, operating conditions, and historical discharge data, and determine a reasonable threshold range through statistical analysis and domain knowledge. For example, the normal discharge thresholds for high-voltage transformers and high-voltage cables may differ and need to be set separately. Furthermore, the thresholds need to be updated and optimized regularly based on actual equipment operating data to ensure the effectiveness of threshold discrimination.

[0110] The temporal consistency of the data is crucial throughout the process. The timestamps of the predicted correction values, historical predicted values, and actual discharge characteristic values ​​must be strictly aligned to ensure that the difference calculation is based on the predicted and actual values ​​at the same time, avoiding errors in difference calculation due to time misalignment. In practice, data temporal consistency can be guaranteed through a unified timestamp management system or a time alignment algorithm in data preprocessing.

[0111] Through the above implementation method, the predicted correction value and historical predicted value can be dynamically weighted and fused, making full use of the complementarity of multi-dimensional prediction information. Simultaneously, the weights are adaptively adjusted according to the deviation between the predicted and actual values, improving the accuracy of partial discharge characteristic prediction. Combined with a scientifically sound threshold discrimination mechanism, real-time monitoring and early warning of the discharge status of power equipment can be achieved, providing strong technical support for the safe operation of the equipment. This process is interconnected, and the accuracy and rigor of each step jointly ensure the reliability and effectiveness of discharge status monitoring.

[0112] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0113] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for online monitoring of partial discharge safety in power equipment using multi-modal sensing, characterized in that, The method includes the following steps: Ultrasonic data, ultra-high frequency data, and electromagnetic interference data of the external environment of power equipment partial discharge are collected using different types of sensors. The next moment after the current moment is recorded as the prediction moment. For a preset number of moments before the current moment, the prediction values ​​of ultrasonic data, UHF data, and electromagnetic interference data at the prediction moment are obtained through the prediction model. Based on the correlation between ultrasonic data and UHF data, the correlation between ultrasonic data and electromagnetic interference data, and the prediction values ​​of UHF data and ultrasonic data at the prediction moment, different partial discharge characteristic prediction values ​​are obtained. The prediction values ​​of ultrasonic data at the prediction moment are corrected by combining the respective proportions of the correlation between ultrasonic data and UHF data and the correlation between ultrasonic data and electromagnetic interference data with the different partial discharge characteristic prediction values, and the prediction correction value is obtained. The ultrasonic signal, UHF signal, and electromagnetic interference signal at each moment are used to construct three-dimensional feature points, and all three-dimensional feature points are clustered. For the cluster where the current moment is located, the three-dimensional feature points are constructed into different sequences according to the ultrasonic signal, UHF signal, and electromagnetic interference signal. The similarity between three-dimensional feature points under different sequence lengths is obtained by analyzing the sequences of different lengths. Then, the actual discharge feature values ​​are combined to form weights to obtain the historical predicted values ​​of partial discharge characteristics. Partial discharge feature prediction values ​​are obtained based on the predicted correction values, historical prediction values, and the similarity between three-dimensional feature points; the discharge state is monitored by threshold discrimination based on the partial discharge feature prediction values; The method for obtaining partial discharge feature prediction values ​​based on predicted correction values, historical prediction values, and the similarity between three-dimensional feature points is as follows: Get all similar feature points for the current sequence length, and get the predicted correction value and historical prediction value of the similar feature points; The predicted correction value and historical prediction value of each similar feature point are subtracted from the actual discharge feature value at the next moment to form the correction difference sequence and the historical difference sequence. For each differential sequence, the product of the variance of the differential sequence and the mean of all sequence values ​​in the differential sequence is used as the weighting factor; The ratio obtained by comparing the weight factor of the corrected difference sequence with the sum of the weight factors of the corrected difference sequence and the weight factors of the historical difference sequence is used as the weight coefficient of the predicted corrected value; the weight factor of the historical difference sequence is compared with the sum of the weight factors of the two to obtain the weight coefficient of the historical predicted value. The predicted value and the historical predicted value are weighted and added together using their respective weighting coefficients to obtain the predicted value of partial discharge characteristics.

2. The online monitoring method for partial discharge safety of power equipment using multi-modal sensing as described in claim 1, characterized in that, The method for collecting ultrasonic data, ultra-high frequency data, and electromagnetic interference data of the external environment for partial discharge of power equipment using different types of sensors is as follows: Ultrasonic sensors and ultra-high frequency sensors are installed on the surface of the power equipment, and electromagnetic interference sensors are installed in the surrounding environment. Each sensor collects data once every preset time, and a number of data are collected. The data collected at all times are recorded as a sequence, and ultrasonic sequence, ultra-high frequency sequence and electromagnetic interference sequence are obtained respectively.

3. The online monitoring method for partial discharge safety of power equipment using multi-modal sensing as described in claim 2, characterized in that, The method for obtaining different partial discharge characteristic prediction values ​​based on the correlation between ultrasonic data and UHF data, the correlation between ultrasonic data and electromagnetic interference data, and the predicted values ​​of UHF data and ultrasonic data at the prediction time is as follows: Calculate the correlation degree between the sequence values ​​of the ultrasonic sequence and the UHF sequence at the same time, and form a sequence of correlation degrees at all times, which is called the audio correlation sequence; calculate the correlation degree between the sequence values ​​of the ultrasonic sequence and the electromagnetic interference sequence at the same time, and form a sequence of correlation degrees at all times, which is called the electromagnetic interference correlation sequence. The average value of all sequence values ​​in the audio frequency correlation sequence is recorded as the first correlation degree, and the average value of all sequence values ​​in the acoustic disturbance correlation sequence is recorded as the second correlation degree. The predicted values ​​of the second and third partial discharge characteristics are obtained based on the first correlation degree, the second correlation degree, and the predicted values ​​of the UHF sequence and the electromagnetic interference sequence at the prediction time.

4. The online monitoring method for partial discharge safety of power equipment using multi-modal sensing as described in claim 3, characterized in that, The method for obtaining the second and third partial discharge feature prediction values ​​based on the first correlation degree, the second correlation degree, and the predicted values ​​of the UHF sequence and the electromagnetic interference sequence at the prediction time is as follows: The predicted value of the UHF sequence at the prediction time is multiplied by the first correlation degree to obtain the second partial discharge characteristic prediction value, and the predicted value of the electromagnetic interference sequence at the prediction time is multiplied by the second correlation degree to obtain the third partial discharge characteristic prediction value.

5. The online monitoring method for partial discharge safety of power equipment using multi-modal sensing as described in claim 2, characterized in that, The method for obtaining the prediction correction value by combining the correlation between ultrasonic data and UHF data, the respective proportions of the correlation between ultrasonic data and electromagnetic interference data, and different partial discharge characteristic prediction values ​​to correct the prediction value of ultrasonic data at the prediction time is as follows: Calculate the standard deviation of all sequence values ​​in the audio frequency correlation sequence, and use the reciprocal of the standard deviation as the correlation between the ultrasonic sequence and the UHF sequence; Calculate the standard deviation of all sequence values ​​in the acoustic interference correlation sequence, and use the reciprocal of the standard deviation as the correlation between the ultrasonic sequence and the electromagnetic interference sequence; The predicted value of the ultrasonic data obtained through the prediction model at the prediction time is recorded as the first partial discharge characteristic prediction value. A weighted value is preset for the first partial discharge feature prediction value, and the weighted values ​​of the second and third partial discharge feature prediction values ​​are obtained based on the correlation ratio. The predicted correction value is obtained by weighting and summing the predicted values ​​of the first partial discharge feature, the second partial discharge feature, the third partial discharge feature, and their respective weighted values.

6. The online monitoring method for partial discharge safety of power equipment using multi-modal sensing as described in claim 2, characterized in that, The method for obtaining historical predicted values ​​of partial discharge characteristics by constructing different sequences of three-dimensional feature points based on ultrasonic signals, ultra-high frequency signals, and electromagnetic interference signals for the current cluster, analyzing the similarity between three-dimensional feature points of different sequence lengths, and then combining the actual discharge characteristic values ​​to form weights is as follows: The three-dimensional feature point corresponding to the current moment is recorded as the target feature point. The cluster in which the target feature point belongs is obtained, and all feature points in this cluster are recorded as class feature points. The ultrasonic, UHF, and electromagnetic interference sequences corresponding to the feature points are obtained by selecting ultrasonic, UHF, and electromagnetic interference signals at preset times. The similarity between 3D feature points is obtained by analyzing the similarity between the corresponding sequences of class feature points and target feature points, and similar feature points are determined. Historical predicted values ​​of partial discharge characteristics are obtained by using the discharge characteristic values ​​of similar feature points corresponding to different sequence lengths at the next time step.

7. The online monitoring method for partial discharge safety of power equipment using multi-mode sensing as described in claim 6, characterized in that, The method for obtaining the similarity between 3D feature points based on the similarity of the corresponding sequences of class feature points and target feature points, and determining similar feature points, is as follows: For each class of feature points, calculate the similarity between its corresponding ultrasonic sequence, UHF sequence, and electromagnetic interference sequence and the ultrasonic sequence, UHF sequence, and electromagnetic interference sequence at the current moment; The similarity between the class feature points and the target feature points is calculated by averaging the similarity of the three sequences and then linearly normalized to obtain the similarity between the class feature points and the target feature points. Obtain the similarity between all class feature points and the target feature point. For the similarity between all class feature points and the target feature point, classify all similarities using a threshold segmentation algorithm. Obtain the mean similarity of each class after classification. Record the class feature points in the class with the larger mean similarity as similar feature points.

8. The online monitoring method for partial discharge safety of power equipment using multi-mode sensing as described in claim 6, characterized in that, The method for obtaining historical predicted values ​​of partial discharge characteristics based on the discharge characteristic values ​​of similar feature points corresponding to different sequence lengths at the next time step is as follows: For each sequence length, the mean of the discharge characteristic values ​​of all similar feature points at the next adjacent time step is used as the prediction reference value for the prediction time step; for this sequence length, the variance of the discharge characteristic values ​​of all similar feature points at the next time step is calculated; the reciprocal of the variance is used as the reliability parameter; the ratio of each reliability parameter to the sum of all reliability parameters is used as the period weight. Historical predicted values ​​are obtained by summing the predicted reference values ​​for all sequence lengths using period weights.

9. The online monitoring method for partial discharge safety of power equipment using multi-modal sensing as described in claim 2, characterized in that, The method for determining the preset time in which each sensor collects data once every preset time is as follows: based on the operating cycle of the power equipment and the fluctuation frequency of historical partial discharge data, the optimal value of the data collection time interval is obtained through statistical analysis; the standard collection cycle corresponding to the current equipment type and the optimal value are averaged to obtain the final preset time.

Images