A partial discharge detection method based on multi-source data fusion

Abstract

The application discloses a partial discharge detection method based on multi-source data fusion, and particularly relates to the technical field of discharge detection; the method is characterized by the following steps: collecting multi-modal partial discharge signals of a partial discharge target device, performing unified time reference alignment and multi-modal data structure normalization processing, and generating multi-source partial discharge observation data; constructing a cross-modal discharge event response sequence, identifying response time delay and amplitude difference between signals of various modes, and extracting inter-modal response coupling feature data; constructing a nonlinear feature alignment mapping function, performing time domain and frequency domain joint mapping on the multi-source observation data, obtaining discharge feature multi-dimensional tensor data after nonlinear coupling compensation, and performing inter-modal weight reconstruction and feature redistribution to generate a partial discharge fusion feature spectrum; through a classification discrimination model, accurate identification of partial discharge types and spatial positions is realized, and a partial discharge detection result is output; and the accuracy of partial discharge detection is effectively improved.

Description

Technical Field

[0001] This invention relates to the field of discharge detection technology, and more specifically, to a partial discharge detection method based on multi-source data fusion. Background Technology

[0002] In the operation and maintenance of power equipment, accurate detection of partial discharge plays a crucial role in ensuring the safety and reliability of the equipment.

[0003] Existing partial discharge detection methods mostly employ single-mode signals for discharge signal analysis. However, signals generated by partial discharge often exhibit multimodal characteristics during actual detection, and complex interference coupling relationships exist between different modes. Because current technologies typically fail to effectively integrate and process the complex relationships between multimodal signals, the accuracy of detection results in determining discharge type and spatial location is limited, hindering the improvement of partial discharge detection accuracy and impacting the safe operation and maintenance efficiency of power equipment. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art, embodiments of the present invention provide a partial discharge detection method based on multi-source data fusion to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A partial discharge detection method based on multi-source data fusion includes the following steps:

[0007] S1. Acquire multi-modal partial discharge signals from the partial discharge target device, perform unified time reference alignment and multi-modal data structure normalization processing, and output multi-source partial discharge observation data.

[0008] S2. Based on multi-source partial discharge observation data, construct cross-modal discharge event response sequences, identify the relative time delay and response amplitude differences between each modal signal in the partial discharge event response, and output inter-modal response coupling characteristic data;

[0009] S3. Based on the intermodal response coupling characteristic data, extract the coupling phase drift relationship between multi-source signals and construct a nonlinear feature alignment mapping function;

[0010] S4. Based on the nonlinear feature alignment mapping function, the time-domain and frequency-domain joint mapping processing of multi-source partial discharge observation data is performed to generate nonlinearly coupled compensated multidimensional tensor data of discharge features.

[0011] S5. Perform intermodal weight reconstruction and feature redistribution on the multidimensional tensor data of discharge characteristics, and output the partial discharge fusion feature map.

[0012] S6. Construct a classification and discrimination model, identify the type and spatial location of partial discharge based on the partial discharge fusion feature map, and output the partial discharge detection results.

[0013] In a preferred embodiment, S1 specifically refers to:

[0014] Acquire multimodal partial discharge signals from the target device;

[0015] High-speed sampling and synchronous sampling processing are performed on the multimodal partial discharge signals respectively;

[0016] The sampled and processed multimodal partial discharge signals are aligned with a unified time reference.

[0017] After the multimodal partial discharge signals are aligned with a unified time reference, they are normalized according to the same multimodal data structure to output multi-source partial discharge observation data.

[0018] In a preferred embodiment, S2 specifically refers to:

[0019] Based on multi-source partial discharge observation data, multi-modal partial discharge signals are correlated and arranged according to time sequence and signal type to construct a cross-modal discharge event response sequence;

[0020] The time response characteristics and amplitude response characteristics of modal partial discharge signals in the cross-modal discharge event response sequence are compared and analyzed to identify the relative time delay and response amplitude differences of each modal partial discharge signal in the same discharge event response process.

[0021] The relative time delay and response amplitude differences between the partial discharge signals of each mode are quantized according to a fixed multimodal feature structure to generate intermodal response coupling feature data.

[0022] In a preferred embodiment, S3 specifically refers to:

[0023] Based on intermodal response coupling feature data, the coupling phase drift relationship of different modal partial discharge signals in response to the same partial discharge event is extracted;

[0024] Calculate the phase drift difference between different modes of partial discharge signals based on the coupling phase drift relationship;

[0025] A nonlinear regression fitting method was used to fit the phase drift difference values ​​to obtain a nonlinear feature alignment mapping function applicable to all modal partial discharge signals;

[0026] The intermodal response coupling feature data is mapped and corrected based on a nonlinear feature alignment mapping function.

[0027] In a preferred embodiment, S4 specifically refers to:

[0028] Based on the nonlinear feature alignment mapping function, time-domain feature transformation and frequency-domain feature decomposition are performed on the observation data of multi-source partial discharge to obtain the feature representation of each mode of partial discharge signal in the time domain and frequency domain, respectively.

[0029] The feature representations of each modal partial discharge signal in the time and frequency domains are fused to form a high-dimensional data structure that can simultaneously represent the feature correlations within and between modes.

[0030] Based on a high-dimensional data structure, the nonlinear coupling relationship between partial discharge signals of different modes is compensated, and multidimensional tensor data of discharge characteristics after nonlinear coupling compensation is generated.

[0031] In a preferred embodiment, S5 specifically refers to:

[0032] The multidimensional tensor data of discharge characteristics after nonlinear coupling compensation is divided into multiple feature subspace data corresponding to local discharge signals of different modes;

[0033] Each feature subspace data is processed independently by weight calculation to obtain feature weight coefficients that represent the importance of features of each modal partial discharge signal.

[0034] Based on the feature weight coefficients, all feature subspace data are re-weighted and combined to complete the weight reconstruction among the features of each modal partial discharge signal.

[0035] Feature redistribution is performed on the feature subspace data after weight reconstruction to generate a partial discharge fusion feature map.

[0036] In a preferred embodiment, S6 specifically refers to:

[0037] Feature fusion processing is performed on historical multimodal partial discharge signal data with pre-labeled partial discharge type and spatial location information to generate a partial discharge feature training dataset for training a classification and discrimination model.

[0038] The classification and discrimination model is trained and optimized using a partial discharge feature training dataset to obtain a classification and discrimination model for identifying the type and spatial location of partial discharges;

[0039] The partial discharge fusion feature map is input into the classification and discrimination model to identify the type of partial discharge and divide its spatial location, and output the partial discharge detection results containing partial discharge type information and spatial location information.

[0040] The technical effects and advantages of the partial discharge detection method based on multi-source data fusion of the present invention are as follows:

[0041] By unifying the time reference and normalizing the structure of multimodal signals, the differences in timing and amplitude between different detection channels can be eliminated, ensuring the consistency and comparability of multi-source observation data. A cross-modal discharge event response sequence is constructed, and the time delay and amplitude coupling features between signals are extracted, enabling in-depth correlation analysis of information from each mode. A nonlinear feature alignment mapping function is extracted based on modal coupling phase drift, providing a foundation for signal compensation. Multidimensional tensor data after nonlinear coupling compensation is generated under joint mapping in the time and frequency domains, improving the feature extraction accuracy of partial discharge signals. Through intermodal weight reconstruction and feature redistribution, the expression of multi-source features is optimized, interference components are suppressed, and the identification of key signals is enhanced. By constructing a partial discharge fusion feature map and combining it with a classification and discrimination model, the identification of discharge type and spatial location is finally achieved, improving the reliability and accuracy of partial discharge detection and ensuring the safety and stability of power equipment operation. Attached Figure Description

[0042] Figure 1 This is a schematic diagram of a partial discharge detection method based on multi-source data fusion according to the present invention. Detailed Implementation

[0043] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0044] Example

[0045] Figure 1 This invention presents a partial discharge detection method based on multi-source data fusion, which includes the following steps:

[0046] S1. Acquire multimodal partial discharge signals from the partial discharge target device, perform unified time reference alignment and multimodal data structure normalization processing, and output multi-source partial discharge observation data.

[0047] S2. Based on multi-source partial discharge observation data, construct cross-modal discharge event response sequences, identify the relative time delay and response amplitude differences between each modal signal in the partial discharge event response, and output inter-modal response coupling characteristic data;

[0048] S3. Based on the intermodal response coupling characteristic data, extract the coupling phase drift relationship between multi-source signals and construct a nonlinear feature alignment mapping function;

[0049] S4. Based on the nonlinear feature alignment mapping function, the time-domain and frequency-domain joint mapping processing of multi-source partial discharge observation data is performed to generate nonlinearly coupled compensated multidimensional tensor data of discharge features.

[0050] S5. Perform intermodal weight reconstruction and feature redistribution on the multidimensional tensor data of discharge characteristics, and output the partial discharge fusion feature map.

[0051] S6. Construct a classification and discrimination model, identify the type and spatial location of partial discharge based on the partial discharge fusion feature map, and output the partial discharge detection results.

[0052] S1. Acquire multimodal partial discharge signals from the target device, perform unified time reference alignment and multimodal data structure normalization, and output multi-source partial discharge observation data, including:

[0053] Acquire multimodal partial discharge signals from the target device;

[0054] Partial discharge target equipment refers to electrical equipment that may experience partial discharge faults, such as oil-filled transformers, gas-insulated switchgear, switch cabinets, cable joints, and other power equipment. Multimodal partial discharge signals include ultra-high frequency (UHF) partial discharge signals, ultrasonic partial discharge signals, transient ground voltage (TGF-T) partial discharge signals, and infrared image partial discharge signals. UHF partial discharge signals are electromagnetic wave signals with frequencies typically between 300 MHz and 3 GHz generated inside or on the surface of power equipment when partial discharge occurs. For example, when a small discharge occurs inside a gas-insulated switchgear, UHF electromagnetic waves are formed and propagate inside the switchgear. Ultrasonic partial discharge signals are mechanical wave signals with frequencies above 20 kHz generated during partial discharge. For example, when insulation defects in an oil-filled transformer cause partial discharge, ultrasonic waves will be generated. Transformer oil generates detectable ultrasonic vibration signals; transient ground voltage partial discharge signals refer to the voltage pulses with small amplitude and short duration that are instantaneously generated on the grounding wire or equipment casing when the equipment discharges; for example, when a partial discharge fault occurs at a cable joint, the grounding wire will generate a measurable millivolt-level instantaneous voltage change signal; infrared image partial discharge signals are infrared radiation generated by the local temperature rise caused by partial discharge of the equipment, which is manifested as infrared image signals collected by an infrared thermal imager. For example, when a partial discharge occurs at the bus joint inside the switch cabinet, the infrared image captured by the infrared thermal imager will show a clear temperature anomaly area; therefore, multimodal partial discharge signals can be collected by deploying different types of sensor devices such as ultra-high frequency sensors, ultrasonic sensors, transient ground voltage sensors, and infrared thermal imagers on the partial discharge target equipment.

[0055] High-speed sampling and synchronous sampling processing are performed on the multimodal partial discharge signals respectively;

[0056] Ultra-high frequency partial discharge signals and transient ground voltage partial discharge signals have fast waveform changes and short durations, so they need to be sampled at high speed to ensure high-precision reconstruction of each tiny discharge event. Ultrasonic partial discharge signals and infrared image partial discharge signals fluctuate relatively slowly and have low change frequencies, so synchronous sampling is used. High-speed sampling and synchronous sampling together ensure the consistency of different modal partial discharge signals on the time axis.

[0057] The sampled and processed multimodal partial discharge signals are aligned with a unified time reference.

[0058] Unified time reference alignment refers to aligning the sampled partial discharge signals of various modes according to a unified time coordinate to eliminate time errors; establishing a unified high-precision time reference, such as the GPS clock or the BeiDou satellite clock, and using the unified clock as the time reference; processing the original timestamps carried by each set of sampled data through interpolation or time difference compensation methods to align the acquired signals of each mode on a unified time axis, thereby ensuring the consistency of partial discharge signals of different modes.

[0059] After the multimodal partial discharge signals are aligned with a unified time reference, they are normalized according to the same multimodal data structure to output multi-source partial discharge observation data.

[0060] A unified two-dimensional or multi-dimensional data array is defined. Ultra-high frequency partial discharge signals and transient ground voltage partial discharge signals are stored in a time-amplitude two-dimensional array, ultrasonic partial discharge signals are stored in a time-frequency-value three-dimensional array, and infrared image partial discharge signals are converted into a two-dimensional pixel-temperature array or a three-dimensional time-pixel-temperature array. A normalization method (e.g., maximum-minimum amplitude normalization or Z-fraction normalization) is used to standardize all signal amplitude data, ensuring that all modal signals have the same amplitude scale and a completely consistent data structure format, facilitating subsequent data fusion operations. The data after unified data structure normalization is the multi-source partial discharge observation data output by this method, possessing a unified multi-modal characteristic structure, which can be used for subsequent feature analysis and fault diagnosis.

[0061] S2. Based on multi-source partial discharge observation data, construct a cross-modal discharge event response sequence, identify the relative time delay and response amplitude differences between each modal signal in the partial discharge event response, and output inter-modal response coupling characteristic data, including:

[0062] Based on multi-source partial discharge observation data, multi-modal partial discharge signals are correlated and arranged according to time sequence and signal type to construct a cross-modal discharge event response sequence;

[0063] Constructing a cross-modal discharge event response sequence refers to arranging and processing multiple modal data, after unified alignment with a time reference, according to the time sequence of acquisition. This involves arranging ultra-high frequency partial discharge signals, ultrasonic partial discharge signals, transient ground voltage partial discharge signals, and infrared image partial discharge signals on a unified time axis based on timestamps. For example, taking a partial discharge event occurring in a gas-insulated switchgear as an example, at the time of the event, in the unified multi-modal characteristic structure, various signals exhibit corresponding characteristic responses at the same or close time points. Based on the unified time axis coordinates, the electromagnetic wave amplitude and occurrence time measured by the ultra-high frequency partial discharge signal, the vibration amplitude and occurrence time measured by the ultrasonic partial discharge signal, the voltage amplitude and occurrence time measured by the transient ground voltage partial discharge signal, and the infrared image temperature anomaly location and occurrence time recorded by the infrared image partial discharge signal are all uniformly recorded. This forms a cross-modal discharge event response sequence with temporal characteristics, reflecting the temporal order of occurrence of each partial discharge event on different modal signals.

[0064] The time response characteristics and amplitude response characteristics of modal partial discharge signals in the cross-modal discharge event response sequence are compared and analyzed to identify the relative time delay and response amplitude differences of each modal partial discharge signal in the same discharge event response process.

[0065] Under a unified time sequence, the time response characteristics of each modal partial discharge signal are compared, i.e., the occurrence time and duration of each signal in response to a partial discharge event, and the amplitude response characteristics, i.e., the amplitude magnitude and variation law of each signal in the partial discharge event. For example, when partial discharge occurs in an oil-filled transformer, the ultra-high frequency partial discharge signal usually appears fastest and lasts for a very short time, the ultrasonic partial discharge signal appears slightly later and lasts for a longer time, the transient ground voltage partial discharge signal is relatively synchronous with the ultra-high frequency partial discharge signal and has a relatively small amplitude, and the infrared image partial discharge signal will lag behind the temperature rise for a long time due to the slow temperature change. Through comparative analysis, the relative time delay difference of each modal partial discharge signal in the process of responding to the same partial discharge event can be accurately identified, i.e., the order of occurrence of the response event and the delay time between each modal signal, and the amplitude difference exhibited by different modal signals in response to the event can also be identified. The time delay and amplitude differences reflect the differences in the response characteristics of each modal signal to the partial discharge event.

[0066] The relative time delay and response amplitude differences between various modal partial discharge signals are quantized according to a fixed multimodal feature structure to generate intermodal response coupling feature data.

[0067] Quantization according to a fixed multimodal feature structure refers to quantifying the identified relative time delay differences and response amplitude differences according to a predetermined unified feature format. That is, it involves converting the feature differences existing in the form of continuous waveforms or images into numerical feature vectors. For example, taking a partial discharge event in a switch cabinet as an example, the relative time delay between the ultra-high frequency partial discharge signal and the transient ground voltage partial discharge signal is measured to be 50 nanoseconds, the time delay of the ultrasonic partial discharge signal relative to the ultra-high frequency partial discharge signal is measured to be 2 microseconds, and the delay of the infrared image partial discharge signal relative to the ultra-high frequency partial discharge signal is measured to be 500 milliseconds. At the same time, the measured amplitude differences are quantized into numerical features, namely, the electromagnetic wave amplitude is 200 millivolts, the transient ground voltage amplitude is 30 millivolts, the ultrasonic signal amplitude is 60 millivolts, and the infrared image signal temperature difference is 2 degrees Celsius. All of the above feature data are uniformly stored as digital feature vectors, ultimately forming intermodal response coupling feature data.

[0068] S3. Based on the intermodal response coupling characteristic data, extract the coupling phase drift relationship between multi-source signals and construct a nonlinear feature alignment mapping function, including:

[0069] Based on intermodal response coupling feature data, the coupling phase drift relationship of different modal partial discharge signals in response to the same partial discharge event is extracted;

[0070] Intermodal response coupling characteristic data is a feature vector formed by quantifying the relative time delay and amplitude differences of partial discharge signals of different modes. It can reflect the differences in response characteristics between different modes of partial discharge signals of the target device when the same partial discharge event occurs. The coupling phase drift relationship is that during the response to a partial discharge event, the partial discharge signal of each mode not only shows differences in time delay and amplitude, but also has a certain phase drift relationship. That is, the waveform phase of the partial discharge signal of different modes changes relatively with time. For example, taking the partial discharge event of gas-insulated switchgear as an example, when the equipment experiences partial discharge, the waveform of the ultra-high frequency partial discharge signal changes. The first type of partial discharge signal may appear first, and its waveform phase can be used as a reference phase. The transient ground voltage partial discharge signal and the ultrasonic partial discharge signal appear successively, and the waveform phase of the transient ground voltage partial discharge signal and the ultrasonic partial discharge signal may have a drift relationship with the phase of the ultra-high frequency signal. The temperature change of the infrared image partial discharge signal is manifested on a longer time scale. Therefore, it is necessary to extract the feature vectors of different modal partial discharge signals in the intermodal response coupling feature data, and extract the waveform phase change curves of different modal partial discharge signals relative to the reference signal through, for example, phase correlation analysis or Fourier transform, so as to obtain the coupling phase drift relationship of different modal partial discharge signals in response to the same partial discharge event.

[0071] Calculate the phase drift difference between different modes of partial discharge signals based on the coupling phase drift relationship;

[0072] The extracted coupling phase drift relationship is reflected as the relative phase change curves between multiple modal partial discharge signals. In order to accurately evaluate the coupling relationship between signals, it is necessary to calculate the phase drift difference between each modal partial discharge signal based on the phase drift relationship data. First, a modal partial discharge signal is selected as the reference phase. For example, the first UHF partial discharge signal is often used as the reference. The phase difference between the waveforms of other modal partial discharge signals and the reference signal waveform is measured and recorded to obtain the phase difference value of each modal partial discharge signal relative to the reference signal. For example, in a partial discharge fault occurring in an oil-filled transformer, the phase drift of the ultrasonic partial discharge signal relative to the UHF partial discharge signal is 45 degrees, and the corresponding transient ground voltage partial discharge signal relative to the UHF partial discharge signal is 30 degrees.

[0073] A nonlinear regression fitting method was used to fit the phase drift difference values ​​to obtain a nonlinear feature alignment mapping function applicable to all modal partial discharge signals;

[0074] After obtaining the phase drift differences between different modal partial discharge signals, in order to establish a unified feature alignment processing rule applicable to all modal signals, a nonlinear regression fitting method is needed to analyze the phase drift differences. Specifically, by selecting a regression function with nonlinear expressive power, such as an exponential function, a higher-order polynomial function, or a logarithmic function, and using the phase drift difference as input data, the parameters of the function are determined according to the principle of minimizing prediction error, and a nonlinear feature alignment mapping function is established. For example, taking a partial discharge event occurring in a switchgear as an example, nonlinear regression calculation is performed on the phase drift data of ultrasonic partial discharge signals and transient ground voltage partial discharge signals relative to ultra-high frequency partial discharge signals, and finally a mapping function similar to the following form is obtained: Y=a×exp(b×X)+c, where X represents the original phase difference between a certain modal partial discharge signal and the reference signal, Y represents the feature-aligned phase difference after fitting, and a, b, and c are all function parameters determined by nonlinear regression calculation. The nonlinear feature alignment mapping function gives a unified conversion rule for the phase drift between arbitrary modal partial discharge signals, and can map the feature data of all different modal partial discharge signals to achieve feature data alignment.

[0075] Mapping and correcting intermodal response coupling feature data based on a nonlinear feature alignment mapping function;

[0076] A nonlinear feature alignment mapping function is applied to intermodal response coupling feature data to achieve mapping correction of response features between different modal partial discharge signals. Specifically, the original feature vectors with phase drift differences in the intermodal response coupling feature data are sequentially input into the nonlinear feature alignment mapping function to obtain mapped, standardized feature vectors. For example, after inputting the original feature vector of the ultrasonic partial discharge signal with a phase difference of 45 degrees into the mapping function, the phase feature difference obtained after mapping correction is 40 degrees, and the phase difference of the transient ground voltage partial discharge signal is 30 degrees, which is corrected to 28 degrees. The intermodal response coupling feature data after mapping correction has higher accuracy and consistency. Finally, the nonlinear feature alignment mapping function and the intermodal response coupling feature data mapped and corrected based on the nonlinear feature alignment mapping function are output.

[0077] S4. Based on the nonlinear feature alignment mapping function, perform joint time-domain and frequency-domain mapping processing on multi-source partial discharge observation data to generate nonlinearly coupled compensated multidimensional tensor data of discharge features, including:

[0078] Based on the nonlinear feature alignment mapping function, time-domain feature transformation and frequency-domain feature decomposition are performed on the observation data of multi-source partial discharge to obtain the feature representation of each mode of partial discharge signal in the time domain and frequency domain, respectively.

[0079] Multi-source partial discharge observation data are mapped using a nonlinear feature alignment mapping function to correct phase differences between different modal signals. For example, when a partial discharge event occurs in a gas-insulated switchgear, the ultra-high frequency (UHF) partial discharge signal is used as the reference signal. After the phase differences of other modal partial discharge signals relative to the UHF partial discharge signal are calculated and corrected using the mapping function, the phase difference of the ultrasonic partial discharge signal is corrected from 45 degrees to 40 degrees, and the phase difference of the transient ground voltage partial discharge signal is corrected from 30 degrees to 28 degrees. The infrared image partial discharge signal does not directly participate in the phase correction. The mapped and corrected data are then used for time-domain feature transformation and frequency-domain feature decomposition. Time-domain feature transformation involves extracting waveform features related to the partial discharge event from the waveform of each modal partial discharge signal, such as peak amplitude, waveform slope, rise time, and duration. Frequency-domain feature decomposition involves using Fourier transform and short-time Fourier transform. Signal processing methods such as wavelet transform are used to transform each modal partial discharge signal from the time domain to the frequency domain for analysis. Multiple frequency domain characteristic parameters, such as peak frequency, center frequency, dominant frequency distribution, and band energy, are extracted from the spectrum. For example, for partial discharge events occurring in oil-filled transformers, the oscillation attenuation amplitude can be obtained after time-domain transformation of the ultra-high frequency partial discharge signal, and the dominant frequency characteristic with a center frequency of 1 GHz can be obtained after frequency-domain decomposition. The ultrasonic partial discharge signal yields a vibration signal packet lasting 2 milliseconds after time-domain transformation, and the spectral characteristic with a dominant frequency of 50 kHz can be obtained after frequency-domain decomposition. The transient ground voltage partial discharge signal's time-domain characteristic is manifested as an instantaneous amplitude of 30 millivolts, and its frequency-domain characteristic is a broadband feature within a range of several hundred kilohertz. The infrared image partial discharge signal's time-domain characteristic transformation shows a gradually increasing temperature rise lasting about 1 second. After processing using the above methods, each modal partial discharge signal obtains characteristic representations in both the time and frequency domains.

[0080] The feature representations of each modal partial discharge signal in the time and frequency domains are fused to form a high-dimensional data structure that can simultaneously represent the feature correlations within and between modes.

[0081] After time-domain feature transformation and frequency-domain feature decomposition, each modal partial discharge signal yields multiple time-domain and frequency-domain feature parameters. To fully reflect the intrinsic connections between different modal signals during partial discharge events and the correlations between features within each modality, a unified feature dimension fusion process is required for the time-domain and frequency-domain feature parameters of each modal signal. Feature dimension fusion refers to using a specific multidimensional matrix or tensor representation method to arrange the time-domain and frequency-domain feature parameters obtained from each modal partial discharge signal into a unified multidimensional feature matrix. For example, for ultra-high frequency partial discharge signals, the time-domain and frequency-domain feature parameters together constitute a two-dimensional feature matrix. The time-domain features and frequency-domain features form another set of two-dimensional feature matrices. The feature representation of transient ground voltage partial discharge signals is also stored using a similar two-dimensional matrix. For infrared image partial discharge signals, since frequency-domain features are not extracted, only time-domain features are used to form a two-dimensional feature matrix. The two-dimensional feature matrices of the above different modes are integrated into a higher-dimensional feature tensor, and an additional dimension is introduced to represent the mode category. In this way, the feature matrices of the four modes of ultra-high frequency, ultrasound, transient ground voltage and infrared image are stacked and combined to form a high-dimensional data structure. The high-dimensional data structure has the correspondence between time-frequency domain features within a mode and the correlation between features between modes, which fully reflects all the multimodal feature information contained in the partial discharge event.

[0082] Based on a high-dimensional data structure, the nonlinear coupling relationship between partial discharge signals of different modes is compensated to generate multidimensional tensor data of discharge characteristics after nonlinear coupling compensation.

[0083] While high-dimensional data structures fully express the feature relationships within and between modes, the nonlinear coupling inherent in partial discharge signals—that is, under certain partial discharge conditions, signal features of different modes can influence or interfere with each other—may lead to distortion or falsification of certain mode signal features during data fusion. To eliminate feature distortion, compensation is needed for the nonlinear coupling relationships between modes reflected in the high-dimensional data structure. Nonlinear coupling compensation refers to feature correction of the mutual influence parts of different modes of partial discharge signals in the high-dimensional data structure. For example, this can be achieved by establishing a nonlinear coupling model between multimodal signals, such as a neural network mapping model or a nonlinear least squares model, to address the interference between different modes of partial discharge signals within the high-dimensional data structure. Using the characteristic parameters as input data and the ideal state of the corresponding signal characteristics as the target output, the model is trained to obtain nonlinear compensation parameters. The compensation parameters are used to compensate and correct the nonlinear coupling relationship in the high-dimensional data structure. For example, the characteristic components of the UHF partial discharge signal affected by the ultrasonic signal are compensated to make the characteristics closer to the real state of the partial discharge event. The characteristics of the transient ground voltage partial discharge signal affected by the UHF signal are also compensated using a similar method. After the compensation and correction process, the final discharge characteristic multidimensional tensor data after nonlinear coupling compensation is formed. The discharge characteristic multidimensional tensor data reflects the real characteristic information of each of the multimodal partial discharge signals, and effectively eliminates the distortion or error caused by the nonlinear coupling between the signals.

[0084] S5. Perform inter-modal weight reconstruction and feature redistribution on the multi-dimensional tensor data of discharge characteristics, and output a partial discharge fusion feature map, including:

[0085] The multidimensional tensor data of discharge characteristics after nonlinear coupling compensation is divided into multiple feature subspace data corresponding to local discharge signals of different modes;

[0086] The discharge feature multidimensional tensor data simultaneously expresses multidimensional feature information of ultra-high frequency partial discharge signals, ultrasonic partial discharge signals, transient ground voltage partial discharge signals, and infrared image partial discharge signals. To more meticulously analyze and identify the unique fault information carried by each modal partial discharge signal and improve the accuracy of diagnostic analysis, the discharge feature multidimensional tensor data needs to be divided into multiple feature subspaces corresponding to different modal partial discharge signals. This is done by splitting the discharge feature multidimensional tensor data according to its modal dimension, that is, decomposing the entire data tensor into multiple subspaces each containing only single-modal information. For example, taking a partial discharge event occurring in a gas-insulated switchgear as an example, the high-dimensional data tensor contains feature information of four modes. After partitioning, it generates feature subspaces containing only ultra-high frequency partial discharge signal feature parameters, feature subspaces containing only ultrasonic partial discharge signal feature parameters, feature subspaces containing only transient ground voltage partial discharge signal feature parameters, and feature subspaces containing only infrared image partial discharge signal feature parameters. This partitioning method ensures that the feature information of each mode is expressed independently.

[0087] Each feature subspace data is processed independently by weight calculation to obtain feature weight coefficients that represent the importance of features of each modal partial discharge signal.

[0088] After dividing the multidimensional tensor data of discharge characteristics after nonlinear coupling compensation into various feature subspaces, it is necessary to calculate the feature importance weights for each modal partial discharge signal feature subspace data separately to quantitatively evaluate the importance of each feature in characterizing partial discharge events. First, feature weight calculation methods are defined, such as analysis of variance, principal component analysis, entropy weighting, or statistical contribution analysis based on partial discharge diagnostic accuracy, to calculate the feature weights for each modal partial discharge signal feature subspace data. For example, for ultra-high frequency partial discharge signal feature subspace data, the time-domain amplitude characteristics and frequency-domain principal components are calculated. The variance or entropy of features such as frequency characteristics are used to obtain feature weight coefficients that reflect the importance of each feature parameter to diagnose the fault. Similarly, for ultrasonic partial discharge signal feature subspace data, feature weight coefficients for vibration amplitude and dominant vibration frequency are calculated. For transient ground voltage partial discharge signal feature subspace data, feature weight coefficients for instantaneous amplitude and frequency band energy characteristics are calculated. For infrared image partial discharge signal feature subspace data, feature weight coefficients for local temperature rise amplitude and temperature continuous change characteristics are calculated. Through the above independent calculation methods, each feature parameter of each modal partial discharge signal obtains a feature weight coefficient representing its importance.

[0089] Based on the feature weight coefficients, all feature subspace data are re-weighted and combined to complete the weight reconstruction among the features of each modal partial discharge signal.

[0090] The feature weighting coefficients reflect the importance of different feature parameters in each modal signal. To better achieve intermodal fusion, it is necessary to re-weight and combine the feature subspace data of each modality according to the feature weighting coefficients, i.e., to achieve weight reconstruction. First, each feature parameter of each feature subspace data is multiplied by the corresponding feature weighting coefficient for re-weighting. For example, in the feature subspace data of ultra-high frequency partial discharge signal, the original amplitude feature parameter is 200 mV and the feature weighting coefficient is 0.8, then the re-weighted feature parameter becomes 160 mV; the original amplitude feature parameter of ultrasonic partial discharge signal feature subspace data... The original amplitude feature of the transient ground voltage partial discharge signal is 30 mV, with a weighting coefficient of 0.6, resulting in a weighted value of 21 mV. The original temperature feature parameter of the infrared image partial discharge signal is 2 degrees Celsius, with a weighting coefficient of 0.5, resulting in a weighted value of 1 degree Celsius. After completing the above individual weighting processing, all the reweighted feature subspace data are combined together according to a unified multimodal feature structure, thereby realizing the weight reconstruction between the features of each modal partial discharge signal. That is, the reconstructed feature parameters reflect the differences in the importance of modal features.

[0091] The feature subspace data after weight reconstruction is reassigned to generate a partial discharge fusion feature map.

[0092] The reconstructed feature subspace data has expressed the importance distribution of each modal feature. To improve the accuracy of partial discharge diagnosis, it is necessary to redistribute the reconstructed features to form a fused feature map. The feature redistribution method is to rearrange the reconstructed feature parameters of each modality according to the predetermined feature dimension positions, based on a unified multimodal data feature space structure, such as a multidimensional tensor structure, to form a unified fused feature map. For example, the reconstructed feature parameters of UHF, ultrasound, transient ground voltage, and infrared images are arranged in specific dimensions of the multidimensional tensor, with each dimension representing a modality and each position representing a feature, forming a fused feature map. For example, taking the partial discharge event of gas-insulated switchgear as an example, the fused feature map places the UHF feature parameter 160 mV, the ultrasound feature parameter 36 mV, the transient ground voltage feature parameter 21 mV, and the infrared image feature parameter 1 degree Celsius in specific positions, while keeping the order and dimension of all features fixed. The final generated partial discharge fused feature map reflects the important feature information of each modal signal.

[0093] S6. Construct a classification and discrimination model to identify the type and spatial location of partial discharges based on the partial discharge fusion feature map, and output the partial discharge detection results, including:

[0094] Feature fusion processing is performed on historical multimodal partial discharge signal data with pre-labeled partial discharge type and spatial location information to generate a partial discharge feature training dataset for training a classification and discrimination model.

[0095] By collecting data on various known types of partial discharge events and their corresponding locations from historical events, a multimodal historical dataset containing complete type and spatial location information is formed. This dataset includes UHF partial discharge signals, ultrasonic partial discharge signals, transient ground voltage partial discharge signals, and infrared image partial discharge signals. Each data record corresponds to a specific partial discharge type and spatial location. For example, taking a common oil-filled transformer in a power system as an example, the historical dataset records different types of partial discharge faults, such as partial discharges caused by aging insulation paper, partial discharges caused by air bubbles in transformer oil, and partial discharges caused by the presence of metallic foreign objects. Each fault event is clearly marked with its discharge location, such as the upper, middle, or lower region of the high-voltage winding. To ensure data accuracy, the annotation of historical data is recorded and confirmed by professionals based on historical equipment disassembly analysis, on-site testing, and measurements from professional testing equipment. The collected data is not just single-modal data, but rather a combination of UHF, ultrasonic, and infrared partial discharge signals. Transient ground voltage, infrared image, and four other modal data are collected, processed, and labeled uniformly. The feature fusion process involves uniform high-speed or synchronous sampling of historical data, time reference alignment, data structure normalization, phase drift feature alignment, time-domain and frequency-domain feature extraction, nonlinear coupling compensation, feature dimension fusion, and feature weight reconstruction and redistribution. Ultimately, all historical data are processed into a unified multimodal fusion feature structure. For example, after processing, historical data of oil-filled transformers generates a unified fusion feature map for each historical event. To train the classification and discrimination model, all processed historical data and known labeled information are uniformly matched to form a partial discharge feature training dataset suitable for model training. Each data sample in the training dataset includes fused multimodal features and corresponding known partial discharge types and spatial location labels. The scale and types of data samples can cover all possible discharge types and locations in practical application scenarios, thus ensuring the sufficiency and effectiveness of model training.

[0096] The classification and discrimination model is trained and optimized using a partial discharge feature training dataset to obtain a classification and discrimination model for identifying the type and spatial location of partial discharges;

[0097] A classification and discrimination model is used to train the pre-processed partial discharge feature training dataset. The chosen model can be a support vector machine, random forest, or deep learning network model, suitable for multimodal signal classification and recognition. Taking a deep learning network model as an example, the deep learning network model consists of an input layer, several hidden layers, and an output layer. The input layer receives the fused multimodal feature data from the partial discharge feature training dataset. The hidden layers consist of multiple neuron nodes. Through layer-by-layer nonlinear transformation, the deep features required for partial discharge type and spatial location information are automatically extracted from the fused feature data. The output layer is designed as a classification and discrimination layer, and the number of nodes in the output layer is determined according to the classification requirements of partial discharge type and spatial location. For example, for an oil-filled transformer, assuming three typical partial discharge types and three typical spatial locations, the output layer can... The model comprises six classification nodes, representing the combined classification results of type and location. The training and optimization process involves continuously adjusting the network model parameters using training samples from historical datasets. For example, cross-entropy loss and stochastic gradient descent or adaptive moment estimation algorithms are employed to train and optimize the model parameters. In each training iteration, the model receives fused feature data as input, compares the output classification result with known annotation information, calculates the loss function value based on the error magnitude, and adjusts the network parameters using backpropagation. As the number of training iterations increases, the accuracy of the network model's classification gradually improves. The training process stops when the model's classification accuracy on the independent validation set reaches a preset requirement, such as above 95%. Through these training and optimization methods, a classification model that can be used to identify different types of partial discharges and delineate spatial locations is finally obtained.

[0098] The partial discharge fusion feature map is input into the classification and discrimination model to perform partial discharge type discrimination and spatial location division, and outputs partial discharge detection results containing partial discharge type information and spatial location information.

[0099] After obtaining the trained and optimized classification and discrimination model, the fused feature map of the partial discharge target device to be detected is input into the trained classification and discrimination model to classify and discriminate the partial discharge type and spatial location. For example, in the actual field detection of a gas-insulated switchgear, the fused feature map indicates that the ultra-high frequency partial discharge characteristic parameter is 160 mV, the ultrasonic partial discharge characteristic is 36 mV, the transient ground voltage characteristic is 21 mV, and the infrared image characteristic is 1 degree Celsius. The fused feature map is input into the classification and discrimination model, and after processing through the input layer and hidden layer, the partial discharge type discrimination result is output at the output layer. For example, the discrimination result is... The system identifies the type of internal micro-discharge and provides spatial location information, such as "upper busbar connection area." The classification model outputs information that includes the type and location of the partial discharge, i.e., the partial discharge detection result. The detection result is presented not only in numerical or textual form but also visually on the equipment structure diagram. For example, the software interface can automatically mark the detection result of "internal micro-discharge occurring in the upper busbar connection area" on the equipment structure diagram. This allows the system to identify the type and location of the partial discharge in the early stages of minor equipment failure, enabling timely maintenance or repair and preventing the failure from escalating.

[0100] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. The semiconductor medium can be a solid-state drive.

[0101] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0102] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and modules described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0103] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or modules may be electrical, mechanical, or other forms.

[0104] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0105] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0106] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0107] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0108] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A partial discharge detection method based on multi-source data fusion, characterized in that, Includes the following steps: S1. Acquire multi-modal partial discharge signals from the partial discharge target device, perform unified time reference alignment and multi-modal data structure normalization processing, and output multi-source partial discharge observation data. S2. Based on multi-source partial discharge observation data, construct cross-modal discharge event response sequences, identify the relative time delay and response amplitude differences between each modal signal in the partial discharge event response, and output inter-modal response coupling characteristic data; S3. Based on the intermodal response coupling characteristic data, extract the coupling phase drift relationship between multi-source signals and construct a nonlinear feature alignment mapping function; S4. Based on the nonlinear feature alignment mapping function, the multi-source partial discharge observation data are subjected to joint time-domain and frequency-domain mapping processing to generate nonlinearly coupled compensated multidimensional tensor data of discharge features, specifically: Based on the nonlinear feature alignment mapping function, time-domain feature transformation and frequency-domain feature decomposition are performed on the observation data of multi-source partial discharge to obtain the feature representation of each mode of partial discharge signal in the time domain and frequency domain, respectively. The feature representations of each modal partial discharge signal in the time and frequency domains are fused to form a high-dimensional data structure that can simultaneously represent the feature correlations within and between modes. Based on a high-dimensional data structure, the nonlinear coupling relationship between partial discharge signals of different modes is compensated to generate multidimensional tensor data of discharge characteristics after nonlinear coupling compensation. S5. Perform intermodal weight reconstruction and feature redistribution on the multidimensional tensor data of discharge characteristics, and output the partial discharge fusion feature map. S6. Construct a classification and discrimination model, identify the type and spatial location of partial discharge based on the partial discharge fusion feature map, and output the partial discharge detection results.

2. The partial discharge detection method based on multi-source data fusion according to claim 1, characterized in that, S1, specifically: Acquire multimodal partial discharge signals from the target device; High-speed sampling and synchronous sampling processing are performed on the multimodal partial discharge signals respectively; The sampled and processed multimodal partial discharge signals are aligned with a unified time reference. After the multimodal partial discharge signals are aligned with a unified time reference, they are normalized according to the same multimodal data structure to output multi-source partial discharge observation data.

3. The partial discharge detection method based on multi-source data fusion according to claim 2, characterized in that, S2, specifically: Based on multi-source partial discharge observation data, multi-modal partial discharge signals are correlated and arranged according to time sequence and signal type to construct a cross-modal discharge event response sequence; The time response characteristics and amplitude response characteristics of modal partial discharge signals in the cross-modal discharge event response sequence are compared and analyzed to identify the relative time delay and response amplitude differences of each modal partial discharge signal in the same discharge event response process. The relative time delay and response amplitude differences between the partial discharge signals of each mode are quantized according to a fixed multimodal feature structure to generate intermodal response coupling feature data.

4. The partial discharge detection method based on multi-source data fusion according to claim 3, characterized in that, S3, specifically: Based on intermodal response coupling feature data, the coupling phase drift relationship of different modal partial discharge signals in response to the same partial discharge event is extracted; Calculate the phase drift difference between different modes of partial discharge signals based on the coupling phase drift relationship; A nonlinear regression fitting method was used to fit the phase drift difference values ​​to obtain a nonlinear feature alignment mapping function applicable to all modal partial discharge signals; The intermodal response coupling feature data is mapped and corrected based on a nonlinear feature alignment mapping function.

5. The partial discharge detection method based on multi-source data fusion according to claim 4, characterized in that, S5, specifically: The multidimensional tensor data of discharge characteristics after nonlinear coupling compensation is divided into multiple feature subspace data corresponding to local discharge signals of different modes; Each feature subspace data is processed independently by weight calculation to obtain feature weight coefficients that represent the importance of features of each modal partial discharge signal. Based on the feature weight coefficients, all feature subspace data are re-weighted and combined to complete the weight reconstruction among the features of each modal partial discharge signal. Feature redistribution is performed on the feature subspace data after weight reconstruction to generate a partial discharge fusion feature map.

6. The partial discharge detection method based on multi-source data fusion according to claim 5, characterized in that, S6, specifically: Feature fusion processing is performed on historical multimodal partial discharge signal data with pre-labeled partial discharge type and spatial location information to generate a partial discharge feature training dataset for training a classification and discrimination model. The classification and discrimination model was trained and optimized using a partial discharge feature training dataset to obtain a classification and discrimination model for identifying the type and spatial location of partial discharges. The partial discharge fusion feature map is input into the classification and discrimination model to identify the type of partial discharge and divide its spatial location, and output the partial discharge detection results containing partial discharge type information and spatial location information.

Images