Ring main unit partial discharge on-line detection method based on multi-feature fusion

Abstract

The application relates to the technical field of power equipment insulation monitoring, in particular to a ring main unit partial discharge online detection method based on multi-feature fusion, which comprises the following steps: deploying multiple types of sensing devices at key insulation components of a ring main unit, synchronously collecting multi-source heterogeneous partial discharge monitoring signals, extracting multiple initial feature vectors after enhancing and aligning the signals, constructing a multi-level feature fusion architecture to realize cross-domain fusion, and generating comprehensive feature expression containing time domain, frequency domain and time-frequency domain information. The comprehensive features are input into a trained deep learning model to identify the discharge mode, the three-dimensional spatial positioning point of the discharge is calculated in combination with the physical layout position of the sensor, and the discharge severity level and potential risk are evaluated according to the discharge mode recognition result and the positioning spatial distribution. The method optimizes the ring main unit partial discharge online detection process through multi-feature fusion and three-dimensional positioning, and improves the accuracy of the detection result.

Description

Technical Field

[0001] This invention relates to the field of insulation monitoring technology for power equipment, and in particular to an online detection method for partial discharge in ring main units based on multi-feature fusion. Background Technology

[0002] Existing online partial discharge detection technologies for ring main units mostly employ single-type sensors to acquire signals. The acquired monitoring signals undergo only routine preprocessing, and the extracted features are limited to a single signal dimension. Discharge pattern recognition is achieved through conventional deep learning networks, and discharge location determination relies solely on coarse interval division based on signal features, without considering the actual sensor deployment locations for precise spatial calculations. Furthermore, multi-source heterogeneous monitoring signals lack synchronous enhancement and alignment processing. Feature extraction from different sensor signals is independent, failing to form a unified feature processing system. The time-domain, frequency-domain, and real-time frequency-domain information of the discharge signals cannot be correlated, resulting in fragmented feature information.

[0003] Existing detection methods lack a multi-level feature fusion architecture, making it difficult to achieve cross-domain fusion of multiple initial feature vectors and form a comprehensive feature representation covering multi-dimensional information. Discharge pattern recognition is susceptible to on-site interference factors, compromising the accuracy of the recognition results. Discharge activity cannot be precisely located in the three-dimensional space of the ring main unit, the spatial correlation between the discharge location and key insulation components is unclear, and the severity level and potential risk assessment of partial discharge events lack spatial distribution data support. Therefore, the detection results cannot fully reflect the actual situation of partial discharge inside the ring main unit. Insufficient feature fusion and inaccurate spatial positioning directly affect the actual effectiveness of online detection. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and propose an online detection method for partial discharge of ring main units based on multi-feature fusion.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: an online detection method for partial discharge of ring main units based on multi-feature fusion, comprising: Multiple types of sensing devices are deployed at key insulating components inside the ring main unit to simultaneously collect multi-source heterogeneous monitoring signals reflecting partial discharge phenomena. The acquired multi-source heterogeneous monitoring signals were enhanced and aligned to extract various initial feature vectors related to partial discharge activity. A multi-level feature fusion architecture is constructed to fuse multiple initial feature vectors across domains, generating a comprehensive feature representation that includes time domain, frequency domain, and time-frequency domain information. Historical partial discharge case data are used to train a deep learning network model, and the comprehensive feature representation is input into the trained deep learning network model for discharge pattern recognition. For the identified discharge pattern, combined with the physical layout of the sensors in the ring main unit, the location point of the discharge activity in the three-dimensional space of the equipment is calculated; Based on the results of discharge pattern recognition and the spatial distribution of location points, the severity level and potential risks of the current partial discharge event are assessed.

[0006] As a further aspect of the present invention, the step of performing signal enhancement and alignment processing on the acquired multi-source heterogeneous monitoring signals to extract various initial feature vectors related to partial discharge activity specifically includes: Ultrasonic signals are received from deployed ultrasonic sensors, bandpass filtered to suppress ambient noise, and an adaptive thresholding method is used to detect pulse events in the ultrasonic signals. The UHF electromagnetic wave signal is received from the UHF sensor, demodulated, and its envelope waveform is obtained. The peak value and rise time of the envelope waveform are then calculated. The transient ground voltage signal is received from the transient ground voltage sensor, the transient ground voltage signal is processed to remove baseline drift, and its pulse repetition frequency and pulse amplitude sequence are extracted. The detected ultrasonic pulse events, ultra-high frequency envelope waveform events, and transient ground voltage pulse events are strictly aligned according to their timestamps to ensure that signals belonging to the same discharge event are matched in time. Waveform statistical parameters, spectral component amplitudes, and wavelet energy spectrum features are extracted from the aligned ultrasonic pulse signal, UHF envelope signal, and transient ground voltage signal, respectively, to form various initial feature vectors.

[0007] As a further aspect of the present invention, the construction of a multi-level feature fusion architecture, which fuses multiple initial feature vectors across domains to generate a comprehensive feature representation containing time-domain, frequency-domain, and time-frequency-domain information, specifically includes: Multiple initial feature vectors characterizing the time-domain waveform are input into the feature concatenation layer, and simple vector concatenation is performed to form primary fusion features; Multiple initial feature vectors representing frequency domain characteristics are input into a feature cross-attention network, the correlation weights between different feature vectors are calculated, and a weighted aggregated frequency domain fusion feature is generated based on the weights. The primary fusion features and the frequency domain fusion features are simultaneously input into a bidirectional long short-term memory network, which captures the long-distance temporal dependency between the primary fusion features and the frequency domain fusion features. The last hidden state of the bidirectional long short-term memory network is concatenated with the intermediate states of the temporal dependencies to form the final comprehensive feature representation.

[0008] As a further aspect of the present invention, the step of training a deep learning network model using historical partial discharge case data, and inputting the comprehensive feature representation into the trained deep learning network model for discharge pattern recognition, specifically includes: We collected historical partial discharge case data including various discharge types such as surface discharge, internal discharge, and corona discharge, and manually annotated the comprehensive feature representation corresponding to each case data. Construct a deep learning network model containing convolutional layers, attention mechanism layers, and fully connected layers to process comprehensive feature representations; Using labeled historical partial discharge case data and their comprehensive feature representation, and with the cross-entropy loss function as the optimization objective, the deep learning network model is iteratively trained. Monitor the recognition accuracy of the deep learning network model on the validation set, and stop training when the accuracy no longer improves significantly, and save the network parameters. The comprehensive feature representation generated in the online detection process is input into the trained deep learning network model. Through the forward propagation calculation of the model, the probability distribution of its belonging to various preset discharge modes is output.

[0009] As a further aspect of the present invention, the construction of a deep learning network model comprising convolutional layers, attention mechanism layers, and fully connected layers specifically includes: One-dimensional convolutional kernels are used to perform local perception on the comprehensive feature representation, extract its local detail patterns, and output convolutional feature maps; The convolutional feature map is input into the attention mechanism layer, which learns the importance scores of different feature dimensions in the convolutional feature map for the discharge pattern recognition task. The convolutional feature map is reweighted based on its importance score to highlight key features and suppress redundant features. The weighted feature map is flattened into a feature vector and then input into multiple stacked fully connected layers for nonlinear transformation. After the last fully connected layer, a classification output layer is connected to output the recognition probability of various discharge modes.

[0010] As a further aspect of the present invention, the step of calculating the location point of the discharge activity in the three-dimensional space of the equipment based on the identified discharge mode and the physical layout of the sensors within the ring main unit specifically includes: Based on the time difference between the discharge pulse signals received by each sensor in the ultrasonic sensor array, the spatial coordinates of the discharge sound source relative to the sensor array are calculated using a time difference positioning algorithm. Based on the amplitude attenuation characteristics or signal arrival angle of the UHF sensor array received signal, spatial spectrum estimation is performed to preliminarily delineate the area where the discharge activity occurs. Spatial consistency verification was performed between the time difference positioning results based on ultrasonic signals and the spatial region estimation results based on UHF signals. If the spatial consistency check passes, the centroid of the two positioning results will be used as the positioning point of the discharge activity in the three-dimensional space of the device. If the spatial consistency check fails, the result obtained by the positioning technology with higher confidence will be used as the positioning point, and the inconsistency information will be recorded.

[0011] As a further aspect of the present invention, the calculation of the spatial coordinates of the discharge sound source relative to the sensor array using a time-difference positioning algorithm specifically includes: The arrival times of the same discharge event on at least four ultrasonic sensors are read from the aligned multi-source signals. Using the position of one of the ultrasonic sensors as a reference, calculate the arrival time difference of the signals received by the other sensors and the reference sensor. Establish a sound wave propagation speed model and determine the current sound wave propagation speed value based on the type and temperature of the gas medium inside the ring main unit; Based on the three-dimensional spatial coordinates, time difference of arrival, and sound wave propagation velocity of each ultrasonic sensor, a system of hyperbolic equations is established. Solving the hyperbolic equations yields the spatial coordinate solution of the discharge sound source that satisfies all time difference constraints.

[0012] As a further aspect of the present invention, the step of assessing the severity level and potential risk of the current partial discharge event based on the results of discharge pattern recognition and the spatial distribution of location points specifically includes: Establish a standard mapping table between discharge mode, discharge quantity, and degree of hazard; Based on the results of discharge pattern identification, the baseline discharge quantity range and typical hazard description corresponding to the discharge pattern are found in the standard mapping table; Analyze the distance between the location point and the core insulating components and conductors inside the ring main unit. The closer the distance, the higher the potential risk assessment. Based on the combined baseline discharge range and the level of potential risk, the current partial discharge event is classified into several preset severity levels; Record the discharge mode, location coordinates, and final severity level determined during the assessment process.

[0013] As a further aspect of the present invention, the step of classifying the current partial discharge event into multiple preset severity levels specifically includes: Set discharge threshold and risk distance threshold as the basic basis for classifying severity levels; If the baseline discharge level corresponding to the current event is in the high range and the location point is less than the risk distance threshold from the core component, then it is classified as the highest severity level. If the reference discharge quantity is in the high range but the location point is far from the core component, or if the reference discharge quantity is in the medium range but the location point is close to the core component, it is classified as a moderately severe level. If the baseline discharge level is in the low range and the location point is far from the core component, it is classified as a low severity level. Establish a mapping table between severity levels and recommended maintenance measures and response times.

[0014] As a further aspect of the present invention, the method further includes: integrating the discharge mode, location point, and severity level to form a standardized partial discharge detection report, and generating status warning information based on the content of the partial discharge detection report, specifically including: Enter the specific type of discharge pattern recognition, confidence probability, three-dimensional coordinate data of the location point, and severity level determination results into the preset report format template; Retrieve component information for the corresponding location area from the ring main unit equipment information database and add it to the inspection report; The report includes the signal acquisition time, duration, and the version of the feature fusion and recognition algorithm used in this test at the end. Based on the final severity level, select the corresponding warning level and recommended action from the mapping table between severity levels and recommended maintenance measures; The text information, which includes the warning level and recommended actions, is associated and encapsulated with a standardized partial discharge detection report to generate a complete status warning information package.

[0015] Compared with the prior art, the advantages and positive effects of the present invention are as follows: The multi-level feature fusion architecture can perform cross-domain fusion of multiple initial feature vectors. Based on the feature information extracted after enhancement and alignment of multi-source heterogeneous monitoring signals, it integrates time domain, frequency domain and time-frequency domain information to form a comprehensive feature expression. Discharge features of different dimensions complement each other, completely retaining various characterization information of partial discharge signals, weakening the information loss problem caused by a single feature dimension, reducing the impact of environmental interference on discharge feature discrimination, and enabling discharge-related features to fully reflect the real state of partial discharge. This allows deep learning network models to rely on more comprehensive feature information to perform calculations when identifying discharge patterns, reducing misjudgments in the pattern recognition process, improving the detail of discharge pattern differentiation, and making the recognition results more consistent with the actual characteristics of partial discharge inside the ring main unit.

[0016] By combining the physical deployment of sensors within the ring main unit to calculate the three-dimensional spatial location points of discharge activities, a correspondence can be established between the discharge pattern recognition results and the internal spatial coordinates of the ring main unit. This clarifies the specific spatial location of the discharge activity within the equipment, clearly demonstrates the spatial relationship between the discharge point and key insulating components, refines the spatial description accuracy of the discharge location, and changes the ambiguity of traditional positioning methods. This allows discharge location information to directly participate in the assessment of the severity level and potential risks of partial discharge events, ensuring that the risk assessment results match the internal structural layout of the ring main unit. Relying on precise spatial positioning information improves the analytical dimensions of discharge events, making the severity level and risk assessment results more closely aligned with the actual operating state of the internal insulating components of the ring main unit. This enhances the accuracy and reliability of discharge risk judgment and provides more complete spatial reference information for the state analysis of discharge events. Attached Figure Description

[0017] Figure 1 The flowchart shows the online detection method for partial discharge of ring main unit based on multi-feature fusion as described in this invention. Figure 2 A flowchart for signal enhancement and alignment processing and initial feature vector extraction; Figure 3 This is a flowchart for training a deep learning network model and recognizing discharge patterns. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0019] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0020] See Figure 1 This invention provides an online detection method for partial discharge in ring main units based on multi-feature fusion, and its overall implementation scheme is as follows: Multiple types of sensing devices, including ultrasonic sensors, UHF sensors, and transient ground voltage sensors, are deployed at key insulating components inside the ring main unit. These sensors simultaneously acquire multi-source heterogeneous monitoring signals reflecting partial discharge phenomena. The acquired multi-source heterogeneous monitoring signals undergo signal enhancement and alignment processing to extract various initial feature vectors related to partial discharge activity. A multi-level feature fusion architecture is constructed to fuse these initial feature vectors across domains, generating a comprehensive feature representation containing time-domain, frequency-domain, and real-time frequency-domain information. A pre-built deep learning network model is trained using historical partial discharge case data. During the online detection phase, the real-time comprehensive feature representation is input into the trained deep learning network model for discharge pattern identification. For the identified discharge patterns, combined with the physical deployment of sensors within the ring main unit, an algorithm calculates the location points of the discharge activity in the three-dimensional space of the equipment. Based on the results of discharge pattern identification and the spatial distribution information of the location points, the severity level and potential risk of the current partial discharge event are assessed.

[0021] In one embodiment of the present invention, see [reference] Figure 2Ultrasonic sensors deployed at key insulating components of the ring main unit receive ultrasonic signals. These signals are bandpass filtered to suppress environmental noise, and an adaptive threshold method is used to detect pulse events within the ultrasonic signal. The core of this adaptive threshold method lies in the fact that the threshold can be dynamically adjusted according to the background noise level, rather than being a fixed value. The specific mechanism and calculation method are as follows: First, the bandpass-filtered ultrasonic signal is segmented into frames, dividing the continuous signal into multiple overlapping time windows. For each current signal frame to be analyzed, several consecutive signal frames preceding it that do not contain obvious discharge pulses are selected as the noise background estimation interval. The standard deviation of the signal amplitude at all sampling points within this noise background estimation interval is calculated and denoted as the noise level estimate. The adaptive threshold is defined as this noise level estimate multiplied by a preset scaling factor. This scaling factor is preset based on the typical noise characteristics of the ring main unit and the amplitude characteristics of the ultrasonic discharge pulse. During detection, starting from the beginning of the ultrasonic signal sequence, the absolute value of the instantaneous amplitude of each sampling point is compared with the dynamically calculated adaptive threshold corresponding to the current signal frame. When the number of sampling points whose instantaneous amplitude continuously exceeds the adaptive threshold reaches the preset minimum pulse width, it is determined as the starting point of a valid pulse event. The peak amplitude of the pulse event and the timestamp of reaching the peak are recorded. As the signal processing progresses, the noise background estimation interval is continuously updated, thereby realizing the dynamic adaptive adjustment of the threshold. This method aims to effectively distinguish between real partial discharge ultrasonic pulses and random background noise, improving the robustness of pulse event detection. A UHF sensor receives UHF electromagnetic wave signals, demodulates the UHF electromagnetic wave signals to obtain their envelope waveform, and calculates the peak value and rise time of the envelope waveform. A transient ground voltage sensor receives transient ground voltage signals, performs baseline drift removal processing on the transient ground voltage signals, and extracts their pulse repetition frequency and pulse amplitude sequence. In some embodiments, the detected ultrasonic pulse events, ultra-high frequency envelope waveform events, and transient ground voltage pulse events are strictly aligned according to their timestamps to ensure that signals belonging to the same discharge event are matched in time. Optionally, waveform statistical parameters, spectral component amplitudes, and wavelet energy spectrum features are extracted from the aligned ultrasonic pulse signal, ultra-high frequency envelope signal, and transient ground voltage signal, respectively, to construct various initial feature vectors.

[0022] In practical implementation, when constructing a multi-level feature fusion architecture, multiple initial feature vectors representing time-domain waveform characteristics are input into the feature concatenation layer for simple vector concatenation to form primary fused features. Multiple initial feature vectors representing frequency-domain characteristics are input into the feature cross-attention network to calculate the correlation weights between different feature vectors. It can be understood that the feature cross-attention network generates a weighted aggregated frequency-domain fused feature based on the weights. The calculation of the correlation weights is expressed by the following formula: in: Indicates the first The initial eigenvector in the frequency domain and the first Attention weights among the initial feature vectors in the frequency domain and It is the initial feature vector in the frequency domain of the input. It is the p-th eigenvector. This represents the similarity calculation function based on the dot product. This represents the total number of initial feature vectors in the frequency domain. In some embodiments, the primary fusion features and the frequency domain fusion features are simultaneously input into a bidirectional long short-term memory (LSTM) network. The LSM network captures the long-distance temporal dependencies between the primary fusion features and the frequency domain fusion features. Specifically, the primary fusion features and the frequency domain fusion features are input into the LSM network simultaneously using a feature concatenation and combination method. Specifically, before being input into the LSM network, the primary fusion feature vector and the frequency domain fusion feature vector are concatenated end-to-end along the feature dimension to form a combined feature vector. Let the primary fusion feature vector be... ,in Let the dimension of the primary fusion feature be denoted as . Let the frequency domain fusion feature vector be . ,in This represents the dimension of the frequency domain fusion features. Then the combined feature vector... It is formed through the following splicing operations: Among the symbols This represents the concatenation operation of vectors along the feature dimension, resulting in a combined feature vector. This combined feature vector fully encompasses both the time-domain waveform characteristics and the weighted, aggregated frequency-domain characteristics. Subsequently, This serves as the input to the bidirectional Long Short-Term Memory (LSTM) network at each time step. The bidirectional LSM network consists of two LSM layers, forward and backward, which encode the temporal evolution patterns of the combined feature vectors from the forward and backward directions of the sequence, respectively, thereby capturing their long-distance temporal dependencies. This concatenation and combination method achieves deep fusion of the two types of features at the input level, providing a unified and information-rich input for subsequent temporal modeling.

[0023] Optionally, the last hidden state of the bidirectional long short-term memory network can be concatenated with the intermediate states of the temporal dependencies to form the final comprehensive feature representation.

[0024] In one embodiment of the present invention, see [reference] Figure 3Historical partial discharge case data, including surface discharge, internal discharge, and corona discharge, was collected, and the comprehensive feature representations generated by a multi-level feature fusion architecture for each case data were manually labeled. A deep learning network model for processing the comprehensive feature representations was constructed, comprising convolutional layers, attention mechanism layers, and fully connected layers. Using the labeled historical partial discharge case data and their corresponding comprehensive feature representations, the deep learning network model was iteratively trained with the cross-entropy loss function as the optimization objective. In some embodiments, the cross-entropy loss function is used to measure the difference between the probability distribution predicted by the deep learning network model and the true labeled categories, guiding the update of model parameters. The recognition accuracy of the deep learning network model was monitored on the validation set. Training was stopped when the accuracy no longer improved significantly, and the network parameters of the deep learning network model at this point were saved. The comprehensive feature representations generated in the online detection stage were input into the trained deep learning network model, and the probability distributions belonging to various preset discharge modes were output through the forward propagation calculation of the deep learning network model.

[0025] In practice, the constructed deep learning network model uses one-dimensional convolutional kernels to perform local perception on the comprehensive feature representation of the input, extracting local detail patterns from the comprehensive feature representation and outputting a convolutional feature map. This convolutional feature map is then input into an attention mechanism layer, which learns the importance scores of different feature dimensions in the convolutional feature map for the discharge pattern recognition task. The importance score reflects the contribution of each feature dimension in distinguishing different discharge patterns, and its calculation process can be expressed by the following formula: in: Represents the first convolutional feature map. Importance scores for each feature dimension Indicates the first The feature vector corresponding to each feature dimension This represents the feature vector corresponding to the j-th feature dimension. It is a learnable mapping function implemented through a fully connected layer. This represents the total number of feature dimensions. Optionally, the feature dimensions of the convolutional feature map are reweighted based on the calculated importance scores. Key features are highlighted and redundant features are suppressed by multiplying the original feature values ​​by their corresponding importance scores. The weighted feature map is flattened into a one-dimensional feature vector, which is then input into multiple stacked fully connected layers for non-linear transformation. A classification output layer is connected after the last fully connected layer. This classification output layer typically uses a softmax function and outputs a comprehensive feature representation representing the recognition probability of belonging to each discharge mode. In some embodiments, the category with the highest recognition probability is the discharge mode type determined by the deep learning network model.

[0026] In one embodiment of the present invention, the spatial coordinates of the discharge sound source relative to the ultrasonic sensor array are calculated using a time-difference localization algorithm based on the time difference between the discharge pulse signals received by each ultrasonic sensor in the ultrasonic sensor array. Specifically, the precise arrival times of at least four ultrasonic sensors are read from the multi-source signals after signal alignment processing, indicating the same discharge event. The precise arrival time is the arrival time of the initial ultrasonic pulse wave generated by the discharge. The identification and localization method is as follows: For each ultrasonic sensor's received, aligned ultrasonic pulse signal, the first obvious pulse peak in the signal is located. Using the pulse peak position as a reference point, a backtracking search is performed on the leading edge signal waveform. During the backtracking process, the short-time energy and short-time average zero-crossing rate of the signal waveform are calculated. When the short-time energy first drops below a certain proportion of the short-time energy of the pulse peak, and the short-time average zero-crossing rate shows a trend of changing from stable to abrupt, this point is determined as the starting point of the initial electrical pulse. The moment when the signal amplitude first exceeds the dynamic threshold calculated by the aforementioned adaptive threshold method from this starting point is recorded as the arrival time of the initial electrical pulse of the discharge event on that sensor. .set up This is the absolute start time of the signal frame. This represents the index position of the starting point in the signal sequence. If the signal sampling frequency is , then the arrival time is . The calculation formula is: in, It is a timestamp recorded by the data acquisition system. It is the starting point index. This refers to the sampling frequency. The above-described starting point identification algorithm is applied to at least four ultrasonic sensors to obtain a set of precise arrival times. This method aims to accurately extract feature points from pulse signals that represent the initial propagation time of the sound source signal, thereby reducing time reading errors caused by pulse waveform width and noise interference, and providing accurate input for solving the hyperbolic equations based on time difference.

[0027] Using the position of one ultrasonic sensor as a reference, the arrival time difference of the signals received by the other ultrasonic sensors and this reference ultrasonic sensor is calculated. A sound wave propagation velocity model is established, and the current sound wave propagation velocity value is determined based on the specific gas medium type inside the ring main unit and the real-time monitored temperature. In essence, based on the pre-calibrated three-dimensional spatial coordinates of each ultrasonic sensor inside the ring main unit, the calculated arrival time difference, and the determined sound wave propagation velocity value, a system of hyperbolic equations is established with the coordinates of the discharge sound source as unknowns. Solving the established hyperbolic equations yields the spatial coordinate solution of the discharge sound source that satisfies all time difference geometric constraints; this solution is the direct output of the time difference localization algorithm.

[0028] In practical implementation, the specific calculation process of the time difference positioning algorithm is described by the following formula, where the path difference between the i-th ultrasonic sensor and the reference ultrasonic sensor satisfies the equation: in: This represents the path difference between the i-th ultrasonic sensor and the reference ultrasonic sensor. It is the currently determined value of the speed of sound propagation. and These are the arrival times of the signals recorded by the i-th ultrasonic sensor and the reference ultrasonic sensor, respectively. These are the spatial coordinates of the discharge sound source to be determined. and These are the known three-dimensional coordinates of the i-th ultrasonic sensor and the reference ultrasonic sensor, respectively. In some embodiments, spatial spectrum estimation is performed based on the amplitude attenuation characteristics of the UHF electromagnetic wave signal received by the UHF sensor array or the signal arrival angle estimated using array signal processing technology to preliminarily delineate the possible area where the discharge activity may occur. This spatial spectrum estimation is implemented using an algorithm based on Multiple Signal Classification (MUSIC). Specifically, for a receiving array composed of M UHF sensor array elements, the covariance matrix of the array received signal is first calculated using the UHF signal sampling data received by each array element. The covariance matrix is ​​then subjected to eigenvalue decomposition. Based on the distribution of eigenvalues, the eigenvectors are divided into a signal subspace and a noise subspace. The signal subspace consists of eigenvectors corresponding to the larger eigenvalues ​​corresponding to the number of UHF partial discharge sources, while the noise subspace consists of eigenvectors corresponding to the remaining smaller eigenvalues. A spatial spectrum function is constructed using the orthogonality between the signal steering vector and the noise subspace. This function is calculated as follows: in, For the corresponding azimuth angle and pitch angle Spatial spectral values, For array in The directional steering vector, the specific form of which is determined by the geometric layout of the sensor array and the signal wavelength, Let be the noise subspace matrix, and its superscript... This represents the conjugate transpose operation. It involves searching for spectral peaks across all possible angles within the three-dimensional space inside the ring main unit, identifying the peak values ​​corresponding to the spectral function peaks. The angle is an estimate of the direction of arrival of the discharge signal. Based on the estimated directions of arrival from at least two separately arranged UHF subarrays or multiple different spatial locations, the intersection area of ​​each direction line is calculated using the triangulation positioning principle, thereby initially delineating the area where the discharge activity is located.

[0029] Optionally, the spatial coordinates calculated by the time difference positioning algorithm based on ultrasonic signals and the area range determined by the spatial spectrum estimation based on UHF signals are spatially consistent. The consistency verification usually includes determining whether the coordinate points are located within the estimated area and whether the orientations indicated by the two methods are within the preset tolerance range.

[0030] In specific implementation, if the spatial consistency check passes, the geometric centroid of the coordinate points of the time-difference positioning result based on ultrasonic signals and the spatial region estimation result region based on UHF signals is calculated, and the calculated centroid coordinates are used as the final positioning point of the discharge activity in the three-dimensional space of the device. In some embodiments, if the spatial consistency check fails, the result obtained by the positioning technology with higher confidence is selected as the final positioning point. The confidence level can be evaluated based on the signal-to-noise ratio or the stability of historical positioning results. When the confidence level is quantified based on the signal-to-noise ratio, the calculation process is as follows: First, a pulse event is selected from the ultrasonic sensor signals involved in the positioning calculation. Centered on the pulse peak point, a signal segment containing the pulse body is extracted forward as the effective signal interval, and a signal segment containing only background noise is extracted backward as the noise interval. The root mean square value of the amplitude of the sampling points within the effective signal interval is calculated. and the root mean square value of the sampling point amplitude within the noise interval Then the signal-to-noise ratio The calculation is defined as follows Subsequently, the signal-to-noise ratio (SNR) value is mapped to a confidence level value in the range of 0 to 1. The mapping function used is an sigmoid function, the expression of which is: in, The confidence level is derived based on the signal-to-noise ratio. This is the calculated signal-to-noise ratio value. A scaling factor to control the steepness of the mapping curve, This is the signal-to-noise ratio offset threshold parameter. This is the base of the natural logarithm. When the signal-to-noise ratio is much higher than... When the confidence level approaches 1, the signal-to-noise ratio is much lower than 1. When the confidence level approaches 0, the confidence level is calculated for each sensor signal involved in the positioning process, and the minimum value is taken as the overall confidence level of the positioning result.

[0031] Optionally, while outputting the final location point coordinates, the system will record events where spatial consistency verification fails and related inconsistency information, such as the distance deviation between the time difference location coordinates and the center point of the UHF estimation area.

[0032] In one embodiment of the present invention, a standard mapping table is established between discharge modes, discharge quantities, and hazard levels. This standard mapping table is predefined based on historical cases and industry standards. Based on the discharge mode identification results, the baseline discharge quantity range and typical hazard description corresponding to the identified discharge mode are retrieved from the standard mapping table. The distance between the location point coordinates and the three-dimensional coordinates of the core insulating components and conductors inside the ring main unit is analyzed. It is understood that the closer the distance, the higher the potential risk assessment. The distance is calculated using the three-dimensional Euclidean distance formula. Combining the baseline discharge quantity range obtained from the standard mapping table with the potential risk level obtained through distance analysis, the current partial discharge event is divided into multiple preset severity levels. The discharge mode type used in the assessment process, the three-dimensional coordinates of the location point, and the final severity level are recorded to form an assessment log.

[0033] In practical implementation, discharge quantity thresholds and risk distance thresholds are set as the basic criteria for classifying severity levels. The discharge quantity thresholds are divided into high, medium, and low range boundary values, while the risk distance thresholds are set based on the internal structural dimensions of the ring main unit and the importance of its components. If the baseline discharge quantity corresponding to the current event is in the high range and the location point is less than the risk distance threshold, it is classified as the highest severity level. If the baseline discharge quantity is in the high range but the location point is far from the core component, or if the baseline discharge quantity is in the medium range but the location point is close to the core component, it is classified as a medium severity level. If the baseline discharge quantity is in the low range and the location point is far from the core component, it is classified as a low severity level. In some embodiments, the "close" and "far" distances are determined by comparing the calculated actual distance with the preset risk distance threshold. It is calculated using the following formula: in: These are the three-dimensional coordinates of the discharge location point. These are the three-dimensional coordinates of the nearest core insulating component or conductor. This is the Euclidean distance between the two. Establish a mapping table between severity level and recommended maintenance measures and response times to guide subsequent operation and maintenance. Optional mapping tables are shown in Table 1. Table 1: Severity Level and Maintenance Response Mapping Table In practical implementation, the determination of the baseline discharge level range relies on a standard mapping table. Examples of possible mapping relationships are as follows: surface discharge modes correspond to medium to high baseline discharge levels, internal discharge modes correspond to high baseline discharge levels, and corona discharge modes correspond to low baseline discharge levels. In some embodiments, the assessment of potential risk considers not only distance but also the specific component type where the location point is situated; for example, the risk is higher near a cable termination than in an open cabinet location. It can be understood that the final severity level is a classification result derived by combining the discharge intensity represented by the baseline discharge level range and the spatial threat represented by distance.

[0034] In one embodiment of the present invention, the specific type of discharge pattern recognition, the confidence probability of discharge pattern recognition, the three-dimensional coordinate data of the location point, and the severity level determination result are filled into a preset report format template, which includes fixed fields for the aforementioned data items. Component information for the corresponding location point area is retrieved from the ring main unit equipment information database. Component information includes component name, component model, and component number, and is added to the inspection report. The signal acquisition time, signal duration, and the version of the feature fusion and recognition algorithm used in this inspection are appended at the end of the report. In some embodiments, the report format template exists in the form of a structured document or database record to ensure that each inspection result is saved in a uniform format. Based on the final severity level determination, the corresponding warning level and recommended operation are selected from the severity level and recommended maintenance action mapping table. Warning levels and severity levels usually have a direct correspondence. The text information containing the warning level and recommended operation is associated and encapsulated with the standardized partial discharge inspection report to generate a complete status warning information package, which can be sent to the superior monitoring system.

[0035] In practical implementation, the filling of the report format template is an automated process. This means that the specific type field for discharge pattern recognition is filled with text such as "surface discharge," "internal discharge," or "corona discharge," the confidence probability field for discharge pattern recognition is filled with a value between 0 and 1, and the three-dimensional coordinate data field for the location point is filled with floating-point values ​​for X, Y, and Z. The process of retrieving component information from the ring main unit equipment information database is achieved through spatial queries. The database stores the spatial coordinate range of each component model within the ring main unit. When the coordinates of a location point fall within the coordinate range of a component, it is determined to be information for that component. In some embodiments, the signal acquisition time is accurate to milliseconds, the signal duration is recorded in seconds, and the feature fusion and recognition algorithm version is recorded in string form, such as "MF-CNN-ATTN-V2.1." Optionally, the mapping table between severity levels and recommended maintenance measures is stored in the system in the form of a lookup table. By inputting severity level keywords such as "high," "medium," and "low," the corresponding warning level text and recommended operation text are output. The status warning information package can be encapsulated in JSON or XML format, containing a link to the storage location of the standardized partial discharge detection report, as well as the core content of the warning level and recommended operation.

[0036] In practice, the process of selecting the warning level and recommended action is deterministic. For example, when the input severity level is "high," the output warning level is "Level 1 Warning" and the recommended action is "Immediately arrange on-site inspection." It can be understood that the association encapsulation operation integrates discrete data elements into a data object with a specific structure. This data object contains a global event ID that uniquely identifies the detected event. The global event ID is generated using the following formula: in: It is the generated global event ID string. It is the selected cryptographic hash function, and it is the unique device identifier string of the ring main unit. It is a precise timestamp string indicating the signal acquisition time, with symbols... This indicates a string concatenation operation. The global event ID will serve as a common index identifier for the status warning information packet and the corresponding standardized partial discharge detection report. In some embodiments, the generated complete status warning information packet will be uploaded to the remote status monitoring master station system via a specified communication protocol. Optionally, after the status warning information packet is generated, the system will trigger different local prompts according to the warning level, such as screen flashing or sounds of different frequencies.

[0037] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for online detection of partial discharge in ring main units based on multi-feature fusion, characterized in that, Includes the following steps: Multiple types of sensing devices are deployed at key insulating components inside the ring main unit to simultaneously collect multi-source heterogeneous monitoring signals reflecting partial discharge phenomena. The acquired multi-source heterogeneous monitoring signals were enhanced and aligned to extract various initial feature vectors related to partial discharge activity. A multi-level feature fusion architecture is constructed to fuse multiple initial feature vectors across domains, generating a comprehensive feature representation that includes time domain, frequency domain, and time-frequency domain information. Historical partial discharge case data are used to train a deep learning network model, and the comprehensive feature representation is input into the trained deep learning network model for discharge pattern recognition. For the identified discharge pattern, combined with the physical layout of the sensors in the ring main unit, the location point of the discharge activity in the three-dimensional space of the equipment is calculated; Based on the results of discharge pattern recognition and the spatial distribution of location points, the severity level and potential risks of the current partial discharge event are assessed.

2. The online detection method for partial discharge of ring main unit based on multi-feature fusion according to claim 1, characterized in that, The process of enhancing and aligning the acquired multi-source heterogeneous monitoring signals to extract various initial feature vectors related to partial discharge activity specifically includes: Ultrasonic signals are received from deployed ultrasonic sensors, bandpass filtered to suppress ambient noise, and an adaptive thresholding method is used to detect pulse events in the ultrasonic signals. The UHF electromagnetic wave signal is received from the UHF sensor, demodulated, and its envelope waveform is obtained. The peak value and rise time of the envelope waveform are then calculated. The transient ground voltage signal is received from the transient ground voltage sensor, the transient ground voltage signal is processed to remove baseline drift, and its pulse repetition frequency and pulse amplitude sequence are extracted. The detected ultrasonic pulse events, ultra-high frequency envelope waveform events, and transient ground voltage pulse events are strictly aligned according to their timestamps to ensure that signals belonging to the same discharge event are matched in time. Waveform statistical parameters, spectral component amplitudes, and wavelet energy spectrum features are extracted from the aligned ultrasonic pulse signal, UHF envelope signal, and transient ground voltage signal, respectively, to form various initial feature vectors.

3. The online detection method for partial discharge of ring main unit based on multi-feature fusion according to claim 1, characterized in that, The construction of a multi-level feature fusion architecture involves cross-domain fusion of multiple initial feature vectors to generate a comprehensive feature representation containing time-domain, frequency-domain, and time-frequency-domain information. Specifically, this includes: Multiple initial feature vectors characterizing the time-domain waveform are input into the feature concatenation layer, and simple vector concatenation is performed to form primary fusion features; Multiple initial feature vectors representing frequency domain characteristics are input into a feature cross-attention network, the correlation weights between different feature vectors are calculated, and a weighted aggregated frequency domain fusion feature is generated based on the weights. The primary fusion features and the frequency domain fusion features are simultaneously input into a bidirectional long short-term memory network, which captures the long-distance temporal dependency between the primary fusion features and the frequency domain fusion features. The last hidden state of the bidirectional long short-term memory network is concatenated with the intermediate states of the temporal dependencies to form the final comprehensive feature representation.

4. The online detection method for partial discharge of ring main unit based on multi-feature fusion according to claim 1, characterized in that, The process of training a deep learning network model using historical partial discharge case data, and then inputting the comprehensive feature representation into the trained deep learning network model for discharge pattern recognition, specifically includes: We collected historical partial discharge case data including various discharge types such as surface discharge, internal discharge, and corona discharge, and manually annotated the comprehensive feature representation corresponding to each case data. Construct a deep learning network model containing convolutional layers, attention mechanism layers, and fully connected layers to process comprehensive feature representations; Using labeled historical partial discharge case data and their comprehensive feature representation, and with the cross-entropy loss function as the optimization objective, the deep learning network model is iteratively trained. Monitor the recognition accuracy of the deep learning network model on the validation set, and stop training when the accuracy no longer improves significantly, and save the network parameters. The comprehensive feature representation generated in the online detection process is input into the trained deep learning network model. Through the forward propagation calculation of the model, the probability distribution of its belonging to various preset discharge modes is output.

5. The online detection method for partial discharge of ring main unit based on multi-feature fusion according to claim 4, characterized in that, The construction of a deep learning network model including convolutional layers, attention mechanism layers, and fully connected layers specifically includes: One-dimensional convolutional kernels are used to perform local perception on the comprehensive feature representation, extract its local detail patterns, and output convolutional feature maps; The convolutional feature map is input into the attention mechanism layer, which learns the importance scores of different feature dimensions in the convolutional feature map for the discharge pattern recognition task. The convolutional feature map is reweighted based on its importance score to highlight key features and suppress redundant features. The weighted feature map is flattened into a feature vector and then input into multiple stacked fully connected layers for nonlinear transformation. A classification output layer is connected after the last fully connected layer to output the recognition probability of various discharge modes.

6. The online detection method for partial discharge of ring main unit based on multi-feature fusion according to claim 1, characterized in that, The process of calculating the location point of the discharge activity in the three-dimensional space of the equipment based on the identified discharge pattern and the physical layout of the sensors within the ring main unit includes: Based on the time difference between the discharge pulse signals received by each sensor in the ultrasonic sensor array, the spatial coordinates of the discharge sound source relative to the sensor array are calculated using a time difference positioning algorithm. Based on the amplitude attenuation characteristics or signal arrival angle of the UHF sensor array received signal, spatial spectrum estimation is performed to preliminarily delineate the area where the discharge activity occurs. Spatial consistency verification is performed between time difference positioning results based on ultrasonic signals and spatial region estimation results based on UHF signals; If the spatial consistency check passes, the centroid of the two positioning results will be used as the positioning point of the discharge activity in the three-dimensional space of the device. If the spatial consistency check fails, the result obtained by the positioning technology with higher confidence will be used as the positioning point, and the inconsistency information will be recorded.

7. The online detection method for partial discharge of ring main unit based on multi-feature fusion according to claim 6, characterized in that, The calculation of the spatial coordinates of the discharge sound source relative to the sensor array using the time-difference positioning algorithm specifically includes: The arrival times of the same discharge event on at least four ultrasonic sensors are read from the aligned multi-source signals. Using the position of one of the ultrasonic sensors as a reference, calculate the arrival time difference of the signals received by the other sensors and the reference sensor. Establish a sound wave propagation speed model and determine the current sound wave propagation speed value based on the type and temperature of the gas medium inside the ring main unit; Based on the three-dimensional spatial coordinates, time difference of arrival, and sound wave propagation velocity of each ultrasonic sensor, a system of hyperbolic equations is established. Solving the hyperbolic equations yields the spatial coordinates of the discharge sound source that satisfy all time difference constraints.

8. The online detection method for partial discharge of ring main unit based on multi-feature fusion according to claim 1, characterized in that, The assessment of the severity level and potential risk of the current partial discharge event based on the results of discharge pattern recognition and the spatial distribution of location points specifically includes: Establish a standard mapping table between discharge mode, discharge quantity, and degree of hazard; Based on the results of discharge pattern identification, the baseline discharge quantity range and typical hazard description corresponding to the discharge pattern are found in the standard mapping table; Analyze the distance between the location point and the core insulating components and conductors inside the ring main unit. The closer the distance, the higher the potential risk assessment. Based on the combined baseline discharge range and the level of potential risk, the current partial discharge event is classified into several preset severity levels; Record the discharge mode, location coordinates, and final severity level determined during the assessment process.

9. The online detection method for partial discharge of ring main unit based on multi-feature fusion according to claim 8, characterized in that, The process of classifying the current partial discharge event into multiple preset severity levels specifically includes: Set discharge threshold and risk distance threshold as the basic basis for classifying severity levels; If the baseline discharge level corresponding to the current event is in the high range and the location point is less than the risk distance threshold from the core component, then it is classified as the highest severity level. If the reference discharge quantity is in the high range but the location point is far from the core component, or if the reference discharge quantity is in the medium range but the location point is close to the core component, it is classified as a moderately severe level. If the baseline discharge level is in the low range and the location point is far from the core component, it is classified as a low severity level. Establish a mapping table between severity levels and recommended maintenance measures and response times.

10. The online detection method for partial discharge of ring main unit based on multi-feature fusion according to claim 9, characterized in that, The method further includes: integrating discharge modes, location points, and severity levels to form a standardized partial discharge detection report, and generating status warning information based on the content of the partial discharge detection report, specifically including: Enter the specific type of discharge pattern recognition, confidence probability, three-dimensional coordinate data of the location point, and severity level determination results into the preset report format template; Retrieve component information for the corresponding location area from the ring main unit equipment information database and add it to the inspection report; The report includes the signal acquisition time, duration, and the version of the feature fusion and recognition algorithm used in this test at the end. Based on the final severity level, select the corresponding warning level and recommended action from the mapping table between severity levels and recommended maintenance measures; The text information, which includes the warning level and recommended actions, is associated and encapsulated with a standardized partial discharge detection report to generate a complete status warning information package.

Images