Intelligent diagnosis method for partial discharge of transformer based on multi-dimensional data fusion

Abstract

The application provides a transformer partial discharge intelligent diagnosis method based on multi-dimensional data fusion, relates to the technical field of intelligent diagnosis of power equipment, and specifically comprises the following steps: obtaining transformer multi-source response data and equipment inherent attributes and constructing a transformer multi-source response window; performing partial discharge pulse response segmentation and discharge frequency band response aggregation on the transformer multi-source response window, and constructing time-frequency collaborative representation; forming window discharge fusion representation and window diagnosis representation based on the time-frequency collaborative representation; evaluating an operation correlation offset and outputting a defect category, a discharge severity and an operation risk, and generating a transformer partial discharge comprehensive diagnosis result.

Description

Technical Field

[0001] This invention relates to the field of intelligent diagnostic technology for power equipment, and in particular to an intelligent diagnostic method for partial discharge of transformers based on multi-dimensional data fusion. Background Technology

[0002] Currently, transformers, as key transmission and transformation equipment in power systems, directly affect the reliability and operational safety of power grids. Partial discharge, as an important characterizing signal in the process of transformer insulation degradation, local defect propagation, and abnormal evolution of internal insulation structure, is crucial for accurately acquiring, effectively characterizing, and stably diagnosing partial discharge responses. This is essential for identifying potential transformer faults, assessing the degree of insulation degradation, determining risk development status, and improving the operational stability of power systems. With the increasing application of transformers under complex load fluctuations, changing operating conditions, and multi-source sensing conditions, partial discharge responses exhibit pulse bursts, frequency band distribution differences, multi-channel coupling, and correlations with operating states. Therefore, there is an urgent need to establish a processing method capable of unified modeling and intelligent diagnosis of multi-source response information.

[0003] Publication No. CN119104848A discloses a method and system for diagnosing transformer partial discharge faults. This method analyzes the ultrasonic signals generated during partial discharge in a transformer using a trained ultrasonic detection and analysis model to determine the current partial discharge data and obtain a primary judgment result. Simultaneously, it identifies and processes the dissolved gas content data in transformer oil using a trained acetylene characteristic gas detection and analysis model to obtain acetylene characteristic gas content data and form a secondary judgment result. Finally, it combines the primary and secondary judgment results to determine the partial discharge state of the transformer. Publication No. CN118152866A discloses a method and system for diagnosing transformer faults based on multi-source data fusion. This method collects oil chromatography analysis data, high-voltage bushing infrared detection spectra, discharge ultrasonic detection spectra, and ultra-high frequency partial discharge detection spectra. It extracts features from different types of data and forms multi-source fused feature vectors, and then uses a diagnostic model to diagnose transformer faults.

[0004] However, existing technologies mainly focus on the joint judgment of ultrasonic signals and characteristic gases, the fusion of oil chromatography and image-based detection data, and the identification of single-type fault states or the classification of multi-source feature splicing. They have not fully considered the unified modeling requirements of pulse segment expression, frequency band response distribution, inherent equipment attribute constraints, multi-monitoring channel collaborative relationships, cross-segment evolution paths, and normal operation correlation offsets of transformer partial discharge under multi-source response windows. As a result, existing methods have problems such as incomplete basic characterization of partial discharge, insufficient utilization of channel collaborative information, insufficient characterization of state correlation changes, low accuracy in defect category determination, and insufficient stability in risk assessment in application scenarios with complex working conditions, enhanced response disturbances, and continuous defect evolution. Summary of the Invention

[0005] This invention proposes an intelligent diagnostic method for transformer partial discharge based on multi-dimensional data fusion. Addressing the issues of significant pulse variations, dispersed frequency band responses, complex channel coordination relationships, unstable temporal evolution paths, and difficulty in quantifying operational correlation offsets in transformer partial discharge responses under multi-source monitoring conditions, this invention establishes a multi-stage processing method. This method includes constructing a multi-source response window for the transformer, segmenting the partial discharge pulse response, aggregating the discharge frequency band response, constructing a time-frequency collaborative characterization, enhancing the collaborative performance of multiple monitoring channels, forming a window discharge fusion characterization, extracting short-range discharge evolution information, extracting cross-segment discharge correlation information, forming a window diagnostic characterization, evaluating operational correlation offsets, and outputting comprehensive diagnostic results. This enables a unified diagnostic output of the transformer's partial discharge defect category, discharge severity, and operational risk status.

[0006] A transformer partial discharge intelligent diagnosis method based on multi-dimensional data fusion is characterized by:

[0007] S1. Obtain transformer multi-source response data and inherent equipment attributes, and construct transformer multi-source response window;

[0008] S2. Perform partial discharge pulse response segmentation and discharge frequency band response aggregation on the multi-source response window of the transformer, and construct a time-frequency collaborative characterization in combination with the inherent attributes of the equipment;

[0009] S3. Calculate the contribution weight of the monitoring channel based on the time-frequency collaborative characterization and perform multi-monitoring channel collaborative enhancement to form a window discharge fusion characterization;

[0010] S4. Based on the window discharge fusion characterization, extract short-range discharge evolution information and cross-segment discharge correlation information in sequence, and perform long-range and short-range response coordination fusion to form a window diagnostic characterization.

[0011] S5. Based on the window diagnostic characterization, construct the current operation correlation map and the corresponding current operation correlation coefficient, construct the normal operation correlation benchmark map and the corresponding normal operation correlation coefficient, evaluate the operation correlation offset, and combine the operation correlation offset to output the defect category judgment result, the discharge severity judgment result and the operation risk judgment result, and generate the transformer partial discharge comprehensive diagnostic result.

[0012] Preferably, in the transformer multi-source response dataset construction process in step S1, firstly, partial discharge response channel data and operating condition auxiliary channel data formed under normal operation, load fluctuation, and partial discharge development states of the transformer are collected. The partial discharge response channel data includes high-frequency current channel response, ultra-high frequency envelope channel response, ultrasonic channel response, and bushing end screen current channel response. The operating condition auxiliary channel data includes oil temperature channel response, winding temperature channel response, load current channel response, and operating voltage channel response. Then, a unified time reference configuration and sampling position alignment are performed on all channels, and the samples are truncated according to the preset sampling length to form a transformer multi-source response window. Next, attribute data used to characterize the transformer insulation structure state and basic operating conditions are extracted to form the inherent attributes of the equipment. Finally, the transformer multi-source response window, inherent attributes of the equipment, window time index, and operating status label are bound and organized to construct a transformer multi-dimensional response dataset for subsequent partial discharge diagnosis tasks.

[0013] Preferably, in step S2, the transformer multi-source response window is segmented for partial discharge pulse response and aggregated for discharge frequency band response, including:

[0014] The transformer multi-source response window is divided into multiple partial discharge pulse response segments according to a preset segment length;

[0015] By mapping each partial discharge pulse response segment, pulse response characterization quantities are obtained;

[0016] Perform spectral transformation on each partial discharge pulse response segment to obtain local spectral coefficients;

[0017] The local spectral coefficients are aggregated according to a preset frequency band range to obtain the frequency band response intensity;

[0018] A frequency band response vector is constructed from the response intensity of each frequency band, and the frequency band response vector is mapped to obtain the frequency band response characterization quantity;

[0019] Map the inherent attributes of the device to obtain the inherent attribute guidance quantity;

[0020] A time-frequency co-characteristic is constructed based on the impulse response characterization, the band response characterization, and the intrinsic attribute guidance.

[0021] Furthermore, addressing the pulse burst nature of transformer partial discharge response on a short time scale, the banded distribution differences in the frequency dimension, and the attribute dependence under different equipment conditions, this invention proposes a time-frequency collaborative characterization mechanism combining partial discharge pulse response segmentation, discharge frequency band response aggregation, and equipment inherent attribute guidance. First, the transformer's multi-source response window is divided into multiple partial discharge pulse response segments according to a preset segment length. Each partial discharge pulse response segment is mapped to obtain a pulse response characterization quantity. Then, a spectral transformation is performed on each partial discharge pulse response segment to obtain local spectral coefficients. These local spectral coefficients are then aggregated according to a preset frequency band interval to form a frequency band response intensity and a frequency band response vector. The frequency band response vector is then mapped to obtain a frequency band response characterization quantity. Based on this, the equipment inherent attributes are mapped to obtain an inherent attribute guidance quantity. Finally, a time-frequency collaborative characterization is constructed based on the pulse response characterization quantity, the frequency band response characterization quantity, and the inherent attribute guidance quantity. This achieves the collaborative fusion of partial discharge pulse morphology information, frequency band response information, and equipment inherent condition information within a unified characterization space, improving the completeness and stability of the basic characterization of partial discharge.

[0022] Preferably, in step S3, the contribution weights of the monitoring channels are calculated based on the time-frequency collaborative characterization, and multi-monitoring channel collaborative enhancement is performed, including:

[0023] The channel contribution score of each monitoring channel in each segment is calculated based on the time-frequency collaborative characterization and the inherent attribute guidance quantity.

[0024] The channel contribution scores are normalized to obtain the monitoring channel contribution weights;

[0025] Channel coordination coefficients are constructed from the time-frequency coordination characteristics of each monitoring channel on the same segment.

[0026] Based on the channel coordination coefficient, the time-frequency coordination characterization of each monitoring channel is enhanced by multi-monitoring channel coordination to obtain the coordination enhancement characterization;

[0027] The collaborative enhancement characterization is weighted and aggregated according to the contribution weight of the monitoring channel to form a window discharge fusion characterization.

[0028] Furthermore, addressing the issues of varying sensitivities to discharge states among different response channels during transformer partial discharge monitoring, collaborative changes among multiple channels within the same segment, and susceptibility of single-channel representation to local interference, this invention proposes a window discharge fusion construction mechanism combining monitoring channel contribution assessment and multi-channel collaborative enhancement. First, based on time-frequency collaborative representation and inherent attribute guidance, the channel contribution score of each monitoring channel in each segment is calculated, and the channel contribution score is normalized to obtain the monitoring channel contribution weight, representing the degree of contribution of different channels to the current partial discharge diagnostic task. Then, a channel collaboration coefficient is constructed from the time-frequency collaborative representation of each monitoring channel in the same segment, and multi-channel collaborative enhancement is performed on the time-frequency collaborative representation of each monitoring channel based on the channel collaboration coefficient, forming a collaborative enhancement representation. Finally, the collaborative enhancement representation is weighted and aggregated according to the monitoring channel contribution weight to form a window discharge fusion representation, thereby achieving contribution screening, collaborative compensation, and fusion compression of multi-source response information, improving the effectiveness and anti-disturbance capability of window-level diagnostic input.

[0029] Preferably, in step S4, based on the window discharge fusion characterization, short-range discharge evolution information and cross-segment discharge correlation information are extracted sequentially, and long- and short-range response coordination fusion is performed, including:

[0030] The initial evolutionary state and initial memory state are constructed based on the inherent attribute guidance quantity;

[0031] The window discharge fusion characterization corresponding to each segment is input into the discharge timing collaborative diagnosis process in chronological order, and the write gate, retention gate, output gate and candidate evolution quantity are updated segment by segment to form segment memory state and discharge short-range evolution information.

[0032] Based on the short-range evolution information of discharge in each segment, the segment retrieval quantity, segment comparison quantity, and segment carrying capacity are constructed, and the segment association weight between segments is calculated to obtain cross-segment discharge association information.

[0033] Based on short-range discharge evolution information, cross-segment discharge correlation information, and inherent attribute guidance quantities, long-range and short-range responses are coordinated and fused to form segmented diagnostic characterization.

[0034] The importance of each segmental diagnostic characterization is aggregated to obtain the window diagnostic characterization.

[0035] Furthermore, addressing the issues of strong local variation persistence, complex cross-segment correlation paths, and the inability of a single time scale to fully represent the state transition trajectory in the temporal evolution of transformer partial discharge response, this invention proposes a window diagnostic representation formation mechanism that combines short-range evolution extraction, cross-segment correlation modeling, and long-short-range coordinated fusion. First, an initial evolution state and an initial memory state are constructed based on inherent attribute guidance quantities. Then, the window discharge fusion representations corresponding to each segment are input into the discharge timing collaborative diagnostic process in chronological order. The write gate, retention gate, output gate, and candidate evolution quantities are updated segment by segment to form segmented memories. The system first obtains state and short-range discharge evolution information; then, based on the short-range discharge evolution information of each segment, it constructs segment retrieval quantity, segment comparison quantity, and segment carrying capacity, calculates the segment association weight between segments, and obtains cross-segment discharge association information; on this basis, it performs long- and short-range response coordination and fusion based on discharge short-range evolution information, cross-segment discharge association information, and inherent attribute guidance quantity to form segmented diagnostic representations, and aggregates the importance of each segmented diagnostic representation to obtain window diagnostic representations, thereby realizing unified modeling of the short-range dynamics of partial discharge and cross-segment evolution relationship, and improving the coherence and discrimination ability of time-series diagnostic representations.

[0036] Preferably, in step S5, the current operational correlation map and corresponding current operational correlation coefficients are constructed based on the window diagnostic characterization, the normal operation correlation benchmark map and corresponding normal operation correlation coefficients are constructed, and the operational correlation offset is evaluated, including:

[0037] Construct the current running association coefficients in the current running association graph based on the window diagnostic characterization;

[0038] Based on the statistical analysis of the operational correlation results from multiple normal operation windows, the normal operation correlation coefficients are constructed in the normal operation correlation benchmark map.

[0039] The operational correlation offset is evaluated based on the current operational correlation coefficient and the normal operation correlation coefficient.

[0040] Preferably, in step S5, the defect category determination result, discharge severity determination result, and operational risk determination result of the operation-related offset output are combined to generate a comprehensive diagnostic result for transformer partial discharge, including:

[0041] By inputting the window diagnostic characterization and the operational correlation offset into the diagnostic output process, the defect category probability, discharge severity value, and operational risk value are obtained.

[0042] The defect category determination result is determined based on the defect category probability.

[0043] The severity of the discharge is determined based on the aforementioned discharge severity value.

[0044] Based on the stated operational risk value, the operational risk assessment result is determined, and a comprehensive diagnostic result for transformer partial discharge is generated.

[0045] Furthermore, addressing the issues of the sensitivity of transformer partial discharge states to changes in correlation structure between normal operating conditions and abnormal development conditions, the difficulty in stably distinguishing different defect states with a single amplitude threshold, and the need for diagnostic results to simultaneously cover category determination and risk assessment, this invention proposes a comprehensive partial discharge diagnostic mechanism based on operational correlation offset. First, it constructs the current operational correlation coefficient in the current operational correlation map based on window diagnostic representations. Then, it statistically constructs the normal operating correlation coefficient in the normal operating correlation benchmark map based on the operational correlation results of multiple normal operating windows. Finally, it assesses the operational correlation offset based on the current operational correlation coefficient and the normal operating correlation coefficient. Subsequently, it inputs both the window diagnostic representation and the operational correlation offset into the diagnostic output process to obtain the defect category probability, discharge severity value, and operational risk value. Further, it determines the defect category determination result based on the defect category probability, the discharge severity determination result based on the discharge severity value, and the operational risk determination result based on the operational risk value, generating a comprehensive diagnostic result for transformer partial discharge. This achieves a unified output of partial discharge state correlation offset, defect category identification, severity estimation, and risk level assessment.

[0046] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0047] 1. This invention constructs an intelligent diagnostic method for transformer partial discharge based on multi-dimensional data fusion. It proposes a time-frequency characterization mechanism that collaboratively constructs partial discharge pulse response segments, discharge frequency band response aggregation, and equipment inherent attributes. Addressing the problems in traditional transformer partial discharge monitoring, such as relying solely on single time-domain pulse information, insufficient utilization of frequency domain response, and failure to participate in diagnosis with equipment condition information, this invention first divides the multi-source response window into partial discharge pulse response segments. Then, it further extracts the discharge frequency band response intensity and jointly constructs a time-frequency collaborative characterization with the equipment inherent attributes. This step can simultaneously characterize the differences in partial discharge in the pulse morphology dimension, frequency band distribution dimension, and equipment state dimension, breaking through the limitations of traditional methods that rely heavily on a single response mode. It achieves a unified expression of the basic characteristics of partial discharge, significantly improving the completeness and stability of partial discharge characterization.

[0048] 2. This invention addresses the problems in traditional transformer partial discharge analysis, such as unclear contribution levels of different monitoring channels, insufficient utilization of inter-channel collaborative relationships, and susceptibility to interference in single-channel responses, by establishing a monitoring channel contribution assessment and multi-monitoring channel collaborative enhancement mechanism, combined with the window discharge fusion characterization formation process. First, it calculates the channel contribution score of each monitoring channel in each segment and normalizes it to obtain the monitoring channel contribution weight. Then, it constructs channel collaborative coefficients based on the time-frequency collaborative characterization in the same segment, performs collaborative enhancement on the characterization of each monitoring channel, and then performs weighted aggregation. This mechanism can effectively screen response channels that are more sensitive to partial discharge diagnosis, strengthen the collaborative compensation capability between multiple channels, and improve the anti-disturbance capability and discrimination accuracy of window-level fusion characterization.

[0049] 3. This invention is based on the extraction of short-range discharge evolution information, the extraction of cross-segment discharge correlation information, and the operational correlation offset diagnosis output mechanism. It combines the comparison process between the current operational correlation map and the normal operation correlation benchmark map. Addressing the shortcomings of traditional partial discharge diagnosis methods, such as insufficient modeling of state evolution continuity, inadequate characterization of correlation structure changes, and a single risk assessment dimension, this invention first extracts short-range evolution information and cross-segment correlation information through a discharge time-series collaborative diagnosis process. Then, it performs long-range and short-range response coordination and fusion to form a window diagnostic representation. Furthermore, it assesses the operational correlation offset by comparing the current operational correlation coefficient with the normal operation correlation coefficient, and outputs the defect category judgment result, discharge severity judgment result, and operational risk judgment result. This design achieves unified diagnosis of partial discharge time-series dynamics, cross-segment correlation structure, and operational risk status, significantly improving the accuracy, continuity, and engineering applicability of comprehensive partial discharge diagnosis for transformers. Attached Figure Description

[0050] Figure 1 This is a flowchart of the intelligent diagnostic method for transformer partial discharge based on multi-dimensional data fusion provided by the present invention.

[0051] Figure 2 This is a structural diagram of the time-frequency co-representation provided by the present invention.

[0052] Figure 3 This is a structural diagram of the window discharge fusion characterization provided by the present invention.

[0053] Figure 4 This is a structural diagram of the formation window diagnostic characterization provided by the present invention.

[0054] Figure 5 This is a structural diagram of the comprehensive diagnostic results for partial discharge in transformers provided by the present invention.

[0055] Figure 6 This is a graph showing the distribution results of the runtime associated offset provided by the present invention.

[0056] Figure 7This is a graph showing the comparison between the actual severity and the predicted severity provided by this invention. Detailed Implementation

[0057] This invention proposes an intelligent diagnostic method for transformer partial discharge based on multi-dimensional data fusion. Addressing the issues of significant pulse variations, dispersed frequency band responses, complex channel coordination relationships, unstable temporal evolution paths, and difficulty in quantifying operational correlation offsets in transformer partial discharge responses under multi-source monitoring conditions, this invention establishes a multi-stage processing method. This method includes constructing a multi-source response window for the transformer, segmenting the partial discharge pulse response, aggregating the discharge frequency band response, constructing a time-frequency collaborative characterization, enhancing the collaborative performance of multiple monitoring channels, forming a window discharge fusion characterization, extracting short-range discharge evolution information, extracting cross-segment discharge correlation information, forming a window diagnostic characterization, evaluating operational correlation offsets, and outputting comprehensive diagnostic results. This enables a unified diagnostic output of the transformer's partial discharge defect category, discharge severity, and operational risk status.

[0058] S1. Obtain transformer multi-source response data and inherent equipment attributes, and construct transformer multi-source response window.

[0059] Please refer to Figure 1. The intelligent diagnostic method for transformer partial discharge based on multi-dimensional data fusion in this embodiment of the application includes the following steps: First, partial discharge response channel data and operating condition auxiliary channel data during transformer operation are collected. In this embodiment, the total number of monitoring channels is uniformly set to [value missing]. The partial discharge response channel comprises a high-frequency current channel, an ultra-high-frequency envelope channel, an ultrasonic channel, and a bushing end-screen current channel. The oil temperature channel, winding temperature channel, load current channel, and operating voltage channel constitute the operating condition auxiliary channel. The partial discharge response channel and the operating condition auxiliary channel are synchronously acquired under a unified clock reference, and their representation dimensions are uniformly set as follows: The normal number of windows is The number of diagnostic categories is The number of diagnostic categories is determined by a preset set of defect categories. In this embodiment, normal, internal discharge, surface discharge, tip discharge, and floating discharge are coded into five diagnostic categories, with the category index corresponding to the normal state denoted as [index missing]. To ensure that the sampled values ​​from different monitoring channels correspond to the same time reference axis, let the unified first... The sampling time corresponding to each sampling location is ,in, , For the first The sampling time corresponding to each sampling position The window start time, To ensure a uniform sampling interval, a uniform sampling interval is set in this embodiment. It is 0.1 microseconds. The window sampling length is determined based on the minimum complete observation time of the partial discharge pulse response and the synchronous observation time of the auxiliary channel under operating conditions. The value range is set to 1024 to 4096. In this embodiment, it is set as follows: This ensures that a single window can simultaneously cover the partial discharge pulse change process and the corresponding operating condition change process. After synchronous acquisition, alignment is performed under a unified time reference. A fixed sampling interval reconstruction method is used to construct the time reference. The continuous sampled values ​​of the partial discharge response channel are reconstructed to a unified sampling time by linear interpolation of adjacent sampling points, and the continuous sampled values ​​of the operating condition auxiliary channel are reconstructed to a unified sampling time by the most recent sample-and-hold method. This yields the synchronous response sequence of each monitoring channel on a unified time reference axis. After unifying the time reference, the synchronous response sequences of each monitoring channel on the unified time reference axis are truncated at the same sampling position to form a transformer multi-source response window matrix. ,in, Represents the set of real numbers. For the first The monitoring channel is in the first Channel response values ​​at each sampling location , , To monitor the number of channels, The window sampling length is set in this embodiment. , While constructing the transformer multi-source response window matrix, attribute data characterizing the transformer insulation structure state and basic operating conditions are extracted and organized to form an inherent attribute vector of the equipment. ,in This is a vector of inherent attributes of the device. For the first Each inherent attribute component As an inherent attribute dimension, in this embodiment It consists of the equipment model, insulation medium type, insulation structure parameters, rated voltage level and service life.

[0060] S2. Perform partial discharge pulse response segmentation and discharge frequency band response aggregation on the multi-source response window of the transformer, and construct a time-frequency collaborative characterization in combination with the inherent attributes of the equipment.

[0061] Furthermore, in step S2, a time-frequency co-representation is constructed, the process of which is as follows: Figure 2 As shown, the specific steps for constructing the time-frequency co-representation are as follows: First, considering the characteristic of partial discharge signals exhibiting sudden pulse changes on a short time scale, the transformer multi-source response window matrix is... According to the preset segment length The window is divided into multiple partial discharge pulse response segments, with the number of segments within a single window being [number missing]. ,in, The number of segments within a single window. For window sampling length, The segment length is... This indicates a floor function, where the segment length is determined based on the partial discharge pulse duration, sampling frequency, and the integrity of the pulse response after segmentation. The value range is set to 128 to 512. In this embodiment, the value is... Thus, in this embodiment, When the window sampling length is not divisible by the segment length, the tail sampling points that are less than a complete segment are discarded or padded with zeros; at the same time, the first... The monitoring channel is in the first Partial discharge pulse response segments on each segment ,in, For the first The monitoring channel is in the first Partial discharge pulse response segments on each segment, This is the sampling position index within the segment. ; response segment to partial discharge pulse By performing a linear mapping, the impulse response characterization quantity is obtained. ,in, For the first The monitoring channel is in the first Impulse response characterization quantity on each segment, The impulse response mapping matrix, This is the impulse response bias vector. To characterize the dimension, the dimension is determined based on a trade-off between the impulse response characterization capability and the model computational complexity. In this embodiment... Subsequently, the partial discharge pulse response segment was analyzed. Perform a discrete spectrum transform to obtain the local spectrum coefficients. ,in, The imaginary unit, For frequency location index, The segment length in this embodiment refers to the number of frequency positions in the positive frequency portion of the spectrum. The number of frequency bands is denoted as The number of frequency bands is determined based on the resolution of the positive frequency portion of the spectrum and the frequency band differentiation requirements. The value range is set to 4 to 16, and in this embodiment, we take... ; will the Each frequency position is divided into equal-width positions. The frequency band is obtained to get the first frequency band. Starting index of each frequency band With Termination Index ,in, Further obtained the first The monitoring channel number The segment in the first Band response intensity of each frequency band ,in, Local spectral coefficients The real part, Local spectral coefficients The imaginary part; the total frequency band response intensity of the same segment. Composition of frequency band response vector By performing a linear mapping on the frequency band response vector, the frequency band response characterization quantity is obtained. ,in, This is the frequency band response mapping matrix. Frequency band response bias vector; subsequently, the device intrinsic attribute vector. Perform a linear mapping to obtain the inherent attribute guidance value. ,in, This is a guiding value for inherent attributes. This is an intrinsic attribute mapping matrix. This is the bias vector of inherent properties. The dimension is the inherent property; based on the impulse response characterization. Frequency band response characterization quantity With inherent attribute guidance quantity Calculate the time-frequency coordination coefficient ,in, For time-frequency coordination strength term, This indicates that the impulse response characterization quantity Frequency band response characterization quantity With inherent attribute guidance quantity A 3D vector formed by concatenating columns. For time-frequency co-readout vector, for transpose, For time-frequency coordinated bias scalar, For the first The monitoring channel number The time-frequency coordination coefficients on each segment, based on the time-frequency coordination coefficients Constructing time-frequency co-characterization ,in, For the first The monitoring channel number Time-frequency co-representation quantities on each segment.

[0062] S3. Calculate the contribution weight of the monitoring channel based on the time-frequency collaborative characterization and perform multi-monitoring channel collaborative enhancement to form a window discharge fusion characterization.

[0063] Furthermore, in step S3, a window discharge fusion characterization is formed, the process of which is as follows: Figure 3 As shown, the specific steps are as follows: First, considering the different contributions of multiple monitoring channels to partial discharge diagnosis, based on the time-frequency co-characterization quantity... With inherent attribute guidance quantity Calculate the first The monitoring channel number Channel contribution score in each segment ,in, To contribute the readout vector, For piecewise mapping matrix, The modulation matrix is ​​an inherent property. This is the contribution bias vector; subsequently, the channel contribution scores of each monitoring channel within the same segment are normalized to obtain the first... The monitoring channel number The contribution weight of monitoring channels in each segment ,in, For the first The monitoring channel number The contribution weights of monitoring channels in each segment are calculated; the collaborative response relationships between different monitoring channels within the same segment are then constructed. The first segment The monitoring channel and the first Channel coordination coefficient between monitoring channels ,in, For the first The monitoring channel number Time-frequency co-representation quantity on each segment For the first The monitoring channel number Time-frequency co-representation quantity on each segment , To characterize the dimension, based on the channel coordination coefficient , with the first The monitoring channel number Time-frequency co-characterization of each segment For the target representation, for the first Time-frequency co-characterization of each monitoring channel within each segment Perform weighted aggregation to obtain the first... The monitoring channel number Collaborative enhancement characterization on each segment ,in, , For the first The monitoring channel number Time-frequency co-representation quantity on each segment For the first The monitoring channel and the first The monitoring channel is in the first Channel coordination coefficients on each segment For the first On the monitoring channel, the first The collaborative enhancement characterization quantity on each segment; based on the contribution weight of the monitoring channel. For collaborative enhancement characterization Weighted aggregation is performed to obtain the window discharge fusion characterization quantity. ,in For the first Window discharge fusion characterization on each segment.

[0064] S4. Based on the window discharge fusion characterization, extract short-range discharge evolution information and cross-segment discharge correlation information in sequence, and perform long-range and short-range response coordination fusion to form a window diagnostic characterization.

[0065] Furthermore, in step S4, a window diagnostic characterization is formed, the process of which is as follows: Figure 4 As shown, the specific steps are as follows: First, guide the quantity based on the inherent attributes. Constructing the initial evolutionary state Compared with the initial memory state ,in, , To initialize the matrix, , To initialize the bias vector; after completing the initial state construction, the window discharge fusion characterization of each segment. The discharge timing collaborative diagnostic process will be input in chronological order, and the gate vector will be updated segment by segment. Preserving gate vectors Output gate vector Candidate evolution quantity Segmented memory state and short-range evolution information of discharge ,in, For the first The write gate vector for each segment, For the first The segmented gate vectors, For the first Each segment of the output gate vector, For the first The number of candidate evolutions for each segment. For the first Segmented memory state of each segment, For the first The short-range evolution information of the discharge in each segment, symbolized by " " indicates the operation of multiplying corresponding components. For the gate mapping matrix, The state mapping matrix, The gate bias vector is used; based on the short-range discharge evolution information corresponding to each segment, the first... Short-range evolution information of discharge segments As input for fragment retrieval, take the first... Short-range evolution information of discharge segments Fragment retrieval metrics were constructed using fragment comparison input and fragment carrying input, respectively. Fragment control group and fragment carrying capacity ,in, , , For the correlation mapping matrix, Based on fragment retrieval volume Fragment control group and fragment carrying capacity Calculate fragment association weights ,in, For the first The segment and the first The segment association weights between segments are used to further obtain cross-segment discharge association information. Subsequently, based on short-range discharge evolution information... Cross-segment discharge correlation information With inherent attribute guidance quantity First, construct the long-range and short-range consistent strength terms. Deviation from long and short distances constraint terms ,in, To coordinate the readout vector, To reconcile the bias scalar, This indicates the short-range evolution information of the discharge. Cross-segment discharge correlation information With inherent attribute guidance quantity The 3D vector formed by concatenating columns is based on the long-short range consistency strength term. and long-range and short-range deviation constraints Calculate the long-range and short-range coordination coefficients Based on this, according to the long-range and short-range coordination coefficients Constructing segmented diagnostic characterization ,in, For the first Segmental diagnostic characteristics of each segment, For the first Short-range evolution information of discharge in each segment, For the first Cross-segment discharge correlation information for each segment, For the first The long-range and short-range coordination coefficients of each segment were determined, and then, based on the diagnostic characteristics of each segment... Calculate the first Importance weight of each segment ,in, Segmented summation index, For the first Segmental diagnostic characteristics of each segment, For the first The importance weight of each segment Read out the vector for importance. As an importance-biased scalar, The number of segments within a single window, based on the importance weight of each segment. Diagnostic characteristics of each segment Weighted aggregation is performed to obtain the window diagnostic characterization. ,in, This serves as a diagnostic representation for the window.

[0066] S5. Construct the current operation correlation map and the normal operation correlation benchmark map based on the window diagnostic characterization, evaluate the operation correlation offset, and output the defect category judgment result, discharge severity judgment result and operation risk judgment result in combination with the operation correlation offset.

[0067] Furthermore, in step S5, a comprehensive diagnostic result for partial discharge of the transformer is generated, the process of which is as follows: Figure 5 As shown, the specific steps are as follows: First, based on the segmented diagnostic characterization... Construct the current running association coefficients in the current running association graph, for the th Segmented diagnostic characteristics With the Segmental diagnostic characteristics Calculate the current running correlation coefficient ,in, The first one in the current window to be diagnosed The segment and the first The current operational correlation coefficient between each segment , It should be noted that the historical samples of normal operation are selected from the historical multi-source response windows collected during the continuous 30-day normal operation of the target transformer. The historical multi-source response windows are consistent with the current transformer to be diagnosed in terms of equipment model, insulation medium type, rated voltage level, and operating conditions. After the current operation correlation coefficients are calculated from multiple normal operation windows, the current operation correlation coefficients at the corresponding positions are statistically averaged to form the normal operation correlation coefficients in the normal operation correlation benchmark map. ,in, For the first The current running correlation coefficient corresponding to each normally operating window For the first The first normal operating window Segmental diagnostic characteristics of each segment, For the first The first normal operating window Segmental diagnostic characteristics of each segment, This represents the normal number of windows, ranging from 200 to 1000. In this embodiment, we take [value missing]. ; based on the current operating correlation coefficient Correlation coefficient with normal operation Evaluate the associated offset during operation , window diagnostic characterization Offset associated with operation The common input diagnostic output process yields a defect category probability vector, a discharge severity value, and an operational risk value; where the defect category probability vector is denoted as... ,in, For the first Each diagnostic category corresponds to a defect category probability component. In this embodiment, the number of diagnostic categories is set to [number]. The number of diagnostic categories is determined by a preset set of defect categories. In this embodiment, normal, internal discharge, surface discharge, tip discharge, and floating discharge are coded into five diagnostic categories, with the category index corresponding to the normal state denoted as [index missing]. ; For the first Category readout vectors for each category, For the first Offset modulation coefficients for each category, For the first The category bias scalar for each category. Summation index for diagnostic categories, based on defect category probability vectors. The diagnostic category corresponding to the highest probability component is used to determine the defect category; further, based on the window diagnostic representation quantity... Offset associated with operation Calculate the severity value of discharge ,in, For severity enhancement items, Read out the severity vector. The severity modulation coefficient, As a severity bias scalar, in this embodiment, when At that time, it was determined to be a mild discharge. At that time, it was determined to be a moderate discharge. At that time, it was determined to be a severe discharge. At that time, it was determined to be an extremely severe discharge; subsequently, the operational risk value was defined as... ,in, The index is for the normal category, determined according to the diagnostic category coding rules. This embodiment sets... , This is a risk balancing coefficient used to adjust the relative weight of the discharge severity value and the deviation from the normal category in the operational risk calculation. Its value is determined based on the stability of risk assessment on the validation samples, and is set to a range of 0.4 to 0.8. In this embodiment, it is taken as... In this embodiment, when When the operational risk result is determined to be low risk, When the operational risk result is determined to be a medium-risk state, When the operational risk result is determined to be high-risk, the aforementioned discharge severity classification threshold and operational risk classification threshold are determined based on the statistical distribution of historical labeled samples on the validation set and the consistency requirements of classification judgment. The defect category judgment result, discharge severity judgment result, and operational risk judgment result together constitute the transformer partial discharge. Overall diagnostic results.

[0068] Furthermore, in one embodiment, the trainable parameters in this invention include an impulse response mapping matrix. Impulse response bias vector Frequency band response mapping matrix Frequency band response bias vector Inherent property mapping matrix Inherent property bias vector Time-frequency co-readout vector Time-frequency co-bias scalar Piecewise mapping matrix Inherent property modulation matrix Contribution readout vector Contribution bias vector Initialization matrix and Initialize the bias vector and Gate mapping matrix State mapping matrix Gate bias vector Association mapping matrix Coordinated readout vector Coordinated bias scalar Importance Readout Vector Importance bias scalar Category readout vector Offset modulation coefficient Category bias scalar Severity readout vector Severity modulation coefficient and severity bias scalar All matrix parameters are initialized using a Xavier uniform distribution, all bias vectors and bias scalars are initialized to 0, all readout vectors are initialized using a zero-mean uniform distribution, and all modulation coefficients are initialized to 0. The trainable parameters are jointly updated in batches during training based on the gradient backpropagation results of the total loss function. This ensures that the trainable parameters in impulse response representation, band response representation, intrinsic attribute guidance, channel contribution evaluation, temporal evolution modeling, cross-segment correlation modeling, and diagnostic output processes form a consistent coupled mapping relationship under the same optimization objective.

[0069] Furthermore, addressing the pulse burstiness, frequency band distribution differences, cross-segment correlation variations, and operational correlation shift sensitivity exhibited by transformer partial discharge response signals under multi-monitoring channel conditions, this invention constructs an intelligent diagnostic model for transformer partial discharge, using the transformer multi-source response window matrix constructed in the aforementioned steps. Device inherent attribute vector As input, the following steps are executed sequentially: partial discharge pulse response segmentation, pulse response characterization construction, discharge frequency band response aggregation, intrinsic attribute guidance, time-frequency collaborative characterization construction, monitoring channel contribution weight calculation, multi-monitoring channel collaborative enhancement, window discharge fusion characterization formation, discharge short-range evolution information extraction, cross-segment discharge correlation information extraction, long- and short-range response coordination and fusion, current operation correlation map construction, normal operation correlation benchmark map construction, and operation correlation offset evaluation, to obtain a defect category probability vector. Severity of discharge Operating risk value Under labeled sample conditions, the total loss function is constructed by combining the defect category discrimination loss and the discharge severity loss. ,in To determine the loss based on the defect category, To account for the severity of the discharge loss, This is the severity loss weight, used to adjust the relative influence of defect category discrimination loss and discharge severity loss on the total loss function. In this embodiment, it is taken as... , In this embodiment, the number of diagnostic categories is... The number of diagnostic categories is determined by a preset set of defect categories. In this embodiment, normal discharge, internal discharge, surface discharge, tip discharge, and floating discharge are coded as five diagnostic categories. For category indexing, For the first The true label components for each category are constructed using one-hot encoding. For the first Defect category probability components for each category, This represents the true severity value of the discharge corresponding to the training sample, which is mapped to an interval based on the severity level labeling results of the training sample. The model is obtained internally; based on this, gradient backpropagation and iterative optimization are performed on the aforementioned trainable parameters according to the total loss function, so that the model gradually learns the mapping law between partial discharge pulse segments, discharge frequency band response, multi-monitoring channel collaborative relationship, cross-segment evolution relationship and operation correlation offset, and finally forms a transformer partial discharge intelligent diagnostic model that can simultaneously output defect category judgment result, discharge severity judgment result and operation risk judgment result.

[0070] Furthermore, the intelligent diagnostic model for transformer partial discharge proposed in this invention is implemented using the Python 3.10 programming language and developed based on the PyTorch 2.1 deep learning framework. The experimental environment is configured with an Ubuntu 22.04 operating system, a CUDA 12.1 computing platform, an NVIDIA RTX 4090 graphics processor with 24 GB of video memory, an Intel Core i9-13900K central processing unit, and 64 GB of RAM. Training samples, validation samples, and test samples are divided in an 8:1:1 ratio. The optimizer used is Adam, the training batch size is set to 32, the number of training epochs is set to 120, the initial learning rate is set to 0.001, and the learning rate decay strategy adopts a step decay method of multiplying by 0.8 every 20 epochs. Severity loss weights are used. The parameter is set to 0.5. The other fixed parameters, including input construction, segmentation rules, frequency band division, category set, number of normal windows, normal category index, risk balance coefficient, and graded threshold interval, are all set as described in S1 to S5 above. This ensures that the experimental environment configuration is consistent with the parameter system in the specific implementation and meets the requirements for the model training and diagnostic output process.

[0071] Furthermore, the acquired transformer multi-source response window matrix and equipment inherent attribute vector are input into the constructed transformer partial discharge intelligent diagnostic model for processing. The distribution results of the operational correlation offset are shown in Figure 6. It can be seen from the figure that the operational correlation offset of the normal samples is close to zero and the distribution is concentrated, indicating that the current operational correlation map and the normal operation correlation benchmark map maintain a high degree of consistency under normal operation. In contrast, the operational correlation offsets of internal discharge, surface discharge, tip discharge, and floating discharge samples are significantly higher than those of normal samples. Among them, the median value and dispersion of the surface discharge samples are the most significant, indicating that this type of defect causes the strongest offset in the coupling relationship of multiple monitoring channels and the fragment correlation structure. Although the internal discharge, tip discharge, and floating discharge samples have some overlap in distribution range, their overall level is still clearly distinguishable from that of normal samples. This indicates that the present invention can effectively characterize the degree of structural anomaly under transformer partial discharge state through the construction of the current operational correlation map, the construction of the normal operation correlation benchmark map, and the evaluation of operational correlation offset. The comparison results of the actual severity and the predicted severity are shown in Figure 7. As shown in the figure, the horizontal axis represents the sample number sorted by the actual severity, and the vertical axis represents the severity score. It can be seen from the figure that the predicted severity curve changes synchronously with the actual severity curve, maintaining good following ability in the low, medium, and high severity ranges. This indicates that a stable mapping relationship has been established between the window diagnostic representation, operational correlation offset, and diagnostic output process constructed in this invention. Although there are some fluctuations at local sample levels, the predicted results remain consistent with the actual results in terms of overall trend, segmented upward process, and high-value range position. This demonstrates that the partial discharge can effectively reflect the process of development from mild to severe. The experimental results verify the effectiveness and stability of this invention in identifying transformer partial discharge defects, assessing discharge severity, and characterizing operational status.

[0072] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept of the present invention, and these modifications and improvements all fall within the protection scope of the present invention.

Claims

1. A method for intelligent diagnosis of transformer partial discharge based on multi-dimensional data fusion, characterized in that: Acquire transformer multi-source response data and inherent device attributes, and construct a transformer multi-source response window; The transformer's multi-source response window is segmented for partial discharge pulse response and aggregated for discharge frequency band response, and a time-frequency collaborative characterization is constructed by combining the inherent properties of the device. The contribution weights of the monitoring channels are calculated based on the time-frequency collaborative characterization, and multi-monitoring channel collaborative enhancement is performed to form a window discharge fusion characterization. Based on the window discharge fusion characterization, short-range discharge evolution information and cross-segment discharge correlation information are extracted sequentially, and long- and short-range response coordination and fusion are performed to form a window diagnostic characterization. Based on the window diagnostic characterization, a current operation correlation map and corresponding current operation correlation coefficients are constructed. A normal operation correlation benchmark map and corresponding normal operation correlation coefficients are constructed. The operation correlation offset is evaluated. Based on the operation correlation offset, the defect category judgment result, the discharge severity judgment result, and the operation risk judgment result are output to generate a comprehensive diagnostic result for transformer partial discharge.

2. The intelligent diagnostic method for transformer partial discharge based on multi-dimensional data fusion according to claim 1, characterized in that, Acquire transformer multi-source response data and inherent device attributes, and construct a transformer multi-source response window, including: Collect transformer partial discharge response channel data and operating condition auxiliary channel data; The partial discharge response channel data and the operating condition auxiliary channel data are unified in time reference, aligned in sampling position and truncated with a preset sampling length to form a transformer multi-source response window; Extract attribute data to characterize the insulation structure and basic operating conditions of the transformer to obtain the inherent attributes of the equipment.

3. The intelligent diagnostic method for transformer partial discharge based on multi-dimensional data fusion according to claim 2, characterized in that, The transformer multi-source response window is segmented for partial discharge pulse response and aggregated for discharge frequency band response, including: The transformer multi-source response window is divided into multiple partial discharge pulse response segments according to a preset segment length; By mapping each partial discharge pulse response segment, pulse response characterization quantities are obtained; Perform spectral transformation on each partial discharge pulse response segment to obtain the local spectral coefficients.

4. The intelligent diagnostic method for transformer partial discharge based on multi-dimensional data fusion according to claim 3, characterized in that, The process of segmenting the partial discharge pulse response and aggregating the discharge frequency band response for the multi-source response window of the transformer also includes: The local spectral coefficients are aggregated according to a preset frequency band range to obtain the frequency band response intensity; A frequency band response vector is constructed from the response intensity of each frequency band, and the frequency band response vector is mapped to obtain the frequency band response characterization quantity; Map the inherent attributes of the device to obtain the inherent attribute guidance quantity; A time-frequency co-characteristic is constructed based on the impulse response characterization, the band response characterization, and the intrinsic attribute guidance.

5. The intelligent diagnostic method for transformer partial discharge based on multi-dimensional data fusion according to claim 4, characterized in that, The contribution weights of monitoring channels are calculated based on time-frequency collaborative characterization, and multi-monitoring channel collaborative enhancement is performed, including: The channel contribution score of each monitoring channel in each segment is calculated based on the time-frequency collaborative characterization and the inherent attribute guidance quantity. The contribution scores of the channels are normalized to obtain the contribution weights of the monitoring channels.

6. The intelligent diagnostic method for transformer partial discharge based on multi-dimensional data fusion according to claim 5, characterized in that, The calculation of monitoring channel contribution weights based on time-frequency collaborative characterization and the implementation of multi-monitoring channel collaborative enhancement also include: Channel coordination coefficients are constructed from the time-frequency coordination characteristics of each monitoring channel on the same segment. Based on the channel coordination coefficient, the time-frequency coordination characterization of each monitoring channel is enhanced by multi-monitoring channel coordination to obtain the coordination enhancement characterization; The collaborative enhancement characterization is weighted and aggregated according to the contribution weight of the monitoring channel to form a window discharge fusion characterization.

7. The intelligent diagnostic method for transformer partial discharge based on multi-dimensional data fusion according to claim 6, characterized in that, Based on the window discharge fusion characterization, short-range discharge evolution information and cross-segment discharge correlation information are extracted sequentially, and long- and short-range response coordination fusion is performed, including: The initial evolutionary state and initial memory state are constructed based on the inherent attribute guidance quantity; The window discharge fusion characterization corresponding to each segment is input into the discharge timing collaborative diagnosis process in chronological order. The write gate, retention gate, output gate and candidate evolution quantity are updated segment by segment to form segment memory state and discharge short-range evolution information.

8. The intelligent diagnostic method for transformer partial discharge based on multi-dimensional data fusion according to claim 7, characterized in that, Based on the window discharge fusion characterization, short-range discharge evolution information and cross-segment discharge correlation information are extracted sequentially, and long- and short-range response coordination fusion is performed. This also includes: Based on the short-range evolution information of discharge in each segment, the segment retrieval quantity, segment comparison quantity, and segment carrying capacity are constructed, and the segment association weight between segments is calculated to obtain cross-segment discharge association information. Based on short-range discharge evolution information, cross-segment discharge correlation information, and inherent attribute guidance quantities, long-range and short-range responses are coordinated and fused to form segmented diagnostic characterization. The importance of each segmental diagnostic characterization is aggregated to obtain the window diagnostic characterization.

9. The intelligent diagnostic method for transformer partial discharge based on multi-dimensional data fusion according to claim 8, characterized in that, Based on the window diagnostic characterization, a current operational correlation map and corresponding current operational correlation coefficients are constructed; a normal operation correlation baseline map and corresponding normal operation correlation coefficients are constructed; and operational correlation offsets are evaluated, including: Construct the current running association coefficients in the current running association graph based on the window diagnostic characterization; Based on the statistical analysis of the operational correlation results from multiple normal operation windows, the normal operation correlation coefficients are constructed in the normal operation correlation benchmark map. The operational correlation offset is evaluated based on the current operational correlation coefficient and the normal operation correlation coefficient.

10. The intelligent diagnostic method for transformer partial discharge based on multi-dimensional data fusion according to claim 9, characterized in that, Combining the defect category determination results, discharge severity determination results, and operational risk determination results from the operational correlation offset output, a comprehensive diagnostic result for transformer partial discharge is generated, including: By inputting the window diagnostic characterization and the operational correlation offset into the diagnostic output process, the defect category probability, discharge severity value, and operational risk value are obtained. The defect category determination result is determined based on the defect category probability. The severity of the discharge is determined based on the aforementioned discharge severity value. Based on the stated operational risk value, the operational risk assessment result is determined, and a comprehensive diagnostic result for transformer partial discharge is generated.

Images