GIS partial discharge positioning method based on deep learning and acoustoelectric combination

Abstract

The invention provides a GIS partial discharge positioning method based on deep learning and acoustoelectric combination, and relates to the field of data processing. The method comprises the following steps: firstly, acquiring an original ultrahigh frequency signal and an original ultrasonic signal of a target detection entity, constructing an acoustoelectric signal matrix, and performing short-time Fourier transform to obtain a complex spectrum tensor and a power spectrum tensor; and then, extracting a high-dimensional feature embedding tensor by using a double-branch deep learning network, and generating a time-frequency masking tensor and a source confidence tensor through a multi-source sub-ion network to complete source-by-source spectrum reconstruction. And a time difference of arrival matrix is calculated by using a phase transformation generalized cross-correlation method, a sensor space coordinate is combined to be input into the depth regression positioning sub-network, and a discharge source is output to predict three-dimensional coordinates and regression uncertainty. And finally, fusing the regression uncertainty and the source confidence tensor to construct a confidence interval, and generating a visual positioning result with an ellipsoid confidence boundary. By implementing the technical scheme provided by the invention, the accuracy of partial discharge positioning is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, and particularly relates to a GIS partial discharge positioning method based on deep learning and sound-electricity combination. BACKGROUND

[0002] In the operation process of a gas insulated metal enclosed device, partial discharge detection is an important means to ensure the insulation state and operation safety of the device. The existing partial discharge positioning method usually relies on arranging sensors at multiple positions, collecting the propagation signals of partial discharge in the device shell and gas medium, and combining the time difference or waveform characteristics for positioning.

[0003] However, in actual engineering applications, especially when there are multiple insulation defects in the device and multiple partial discharge sources occur at close time, the signals generated by each discharge source at the sensor end will overlap in the time domain. Due to the influence of the multi-path effect of the metal shell, the dispersion characteristics of the gas medium and the environmental electromagnetic noise in the propagation process of the signals, the signals of different discharge sources partially coincide in the time domain waveform and frequency domain characteristics, so that the traditional positioning method based on a single physical quantity or simple filtering separation is difficult to accurately distinguish the positions of each discharge source, resulting in low positioning accuracy.

[0004] Therefore, there is an urgent need for a GIS partial discharge positioning method based on deep learning and sound-electricity combination. SUMMARY

[0005] The present application provides a GIS partial discharge positioning method based on deep learning and sound-electricity combination, which facilitates to improve the accuracy of partial discharge positioning.

[0006] In a first aspect of the present application, a GIS partial discharge positioning method based on deep learning and acoustic-electric combination is provided, the method comprising: acquiring original ultra-high frequency signals and original ultrasonic signals for a target detection entity, and generating an acoustic-electric signal matrix; performing short-time Fourier transform on the acoustic-electric signal matrix to obtain a complex spectrum tensor and a power spectrum tensor; inputting the complex spectrum tensor and the power spectrum tensor as inputs, inputting an ultra-high frequency signal branch and an ultrasonic signal branch into a deep learning network for double-branch feature extraction, and outputting a high-dimensional feature embedding tensor; generating a time-frequency mask tensor and a source confidence tensor of each potential partial discharge source based on the high-dimensional feature embedding tensor through a multi-source separation sub-network, and performing source-by-source spectral reconstruction based on the time-frequency mask tensor and the source confidence tensor to obtain a source-by-source time-domain waveform sequence; calculating the source-by-source time-domain waveform sequence using a phase transform generalized cross-correlation method to obtain a time difference of arrival matrix, and inputting the time difference of arrival matrix and the sensor space coordinates of the detection sensors corresponding to the target detection entity into a deep regression positioning sub-network to output a discharge source predicted three-dimensional coordinate and a regression uncertainty; constructing a confidence interval and outputting a visual positioning result with an ellipsoid confidence boundary through the discharge source predicted three-dimensional coordinate, the regression uncertainty, and the source confidence tensor.

[0007] Optionally, the acquiring original ultra-high frequency signals and original ultrasonic signals for a target detection entity, and generating an acoustic-electric signal matrix specifically comprises: acquiring original ultra-high frequency signals and original ultrasonic signals sent by ultra-high frequency sensors and ultrasonic sensors pre-deployed on the target detection entity; attaching high-precision time stamps, sensor space coordinates, and channel identifiers to corresponding sampling points of the original ultra-high frequency signals and the original ultrasonic signals; time stamping the original ultra-high frequency signals and the original ultrasonic signals, constructing a multi-channel data structure with sampling time as row index, sensor channel number as column index, and matrix unit value as corresponding amplitude, and determining the multi-channel data structure as the acoustic-electric signal matrix.

[0008] Optionally, the short-time Fourier transform is performed on the acoustic-electric signal matrix to obtain a complex spectrum tensor and a power spectrum tensor, specifically comprising: before performing the short-time Fourier transform, binding the channel set belonging to the ultrahigh frequency signal and the channel set belonging to the ultrasonic signal in the acoustic-electric signal matrix with corresponding sampling rates, sensor calibration coefficients and channel gain factors, respectively; performing band-pass anti-aliasing filtering and resampling processing on channels with different sampling rates to align to a unified sampling rate and keep the channel index, spatial coordinates and time axis consistent; configuring window length, frame shift, zero padding length and window function type for different types of signals respectively, and using an overlap windowing strategy to frame, window and fast Fourier transform to obtain a complex spectrum tensor containing amplitude information and phase information; performing window energy correction, channel gain correction and sensor sensitivity correction on the complex spectrum tensor to generate the power spectrum tensor.

[0009] Optionally, the complex spectrum tensor and the power spectrum tensor are input into a deep learning network as inputs, and the ultrahigh frequency signal branch and the ultrasonic signal branch are input into the deep learning network for double-branch feature extraction, and a high-dimensional feature embedding tensor is output, specifically comprising: performing normalization processing, geometric prior encoding and early convolution feature representation on the complex spectrum tensor and the power spectrum tensor in the ultrahigh frequency signal branch and the ultrasonic signal branch, and introducing a spatial coordinate embedding generated by the sensor spatial coordinates in each branch to retain spatial position information; in the ultrahigh frequency signal branch, phase correlation features are extracted through complex convolution, complex normalization and complex nonlinear units, and the phase correlation features and the power spectrum tensor are fused to form a first intermediate feature tensor of the ultrahigh frequency signal branch; in the ultrasonic signal branch, narrowband ringing and echo tail features are extracted by combining time series convolution and multi-head self-attention, and the narrowband ringing and echo tail features and the power spectrum tensor are fused to form a second intermediate feature tensor of the ultrasonic signal branch; in the multi-scale representation stage, different spread widths of time-frequency patterns are captured by using a hollow time-frequency convolution, and channel attention mechanisms are used to re-label the multi-scale features to obtain re-labeled features; the re-labeled features, the first intermediate feature tensor and the second intermediate feature tensor are input into a residual reshaping network to generate the high-dimensional feature embedding tensor.

[0010] Optionally, the time-frequency mask tensor and source confidence tensor of each potential partial discharge source are generated from the high-dimensional feature embedding tensor by a multi-source separation sub-network, and source-by-source spectral reconstruction is performed based on the time-frequency mask tensor and the source confidence tensor to obtain a source-by-source time-domain waveform sequence, specifically including: using the shared backbone and source-specific output structure of the multi-source separation sub-network to encode the high-dimensional feature embedding tensor to generate a time-frequency mask tensor containing real part mask, imaginary part mask and amplitude mask; using the uncertainty output structure of the multi-source separation sub-network to combine mask estimation and network intermediate features to generate a source confidence tensor; in the source-by-source spectral reconstruction stage, combining the complex spectrum tensor and the power spectrum tensor with the time-frequency mask tensor and the source confidence tensor to perform consistency weighted reconstruction in the complex domain to obtain complex spectrum estimation of each potential partial discharge source, and restore it to the source-by-source time-domain waveform sequence through inverse short-time Fourier transform.

[0011] Optionally, the phase transform generalized cross-correlation method is used to calculate the source-by-source time-domain waveform sequence to obtain a time difference of arrival matrix, and the time difference of arrival matrix and the sensor spatial coordinates of the detection sensor corresponding to the target detection entity are input into a deep regression positioning sub-network to output discharge source predicted three-dimensional coordinates and regression uncertainty, specifically including: performing pre-emphasis, band-pass filtering and amplitude normalization processing on the source-by-source time-domain waveform sequence to generate a time difference of arrival candidate signal sequence; calculating cross-channel mutual power spectrum from the time difference of arrival candidate signal sequence and applying phase transform weight to obtain a generalized cross-correlation function, obtaining a sub-sampling precision time difference of arrival estimate value by detecting peak position and combining a parabolic interpolation method, and extracting peak height and peak width as time difference of arrival quality degree; weighting and fusing the time difference of arrival estimate value and the time difference of arrival quality degree with the source confidence tensor to construct a consistent time difference of arrival matrix; encoding the consistent time difference of arrival matrix and the sensor spatial coordinates of the detection sensor into geometric embedding, and forming the input of the deep regression positioning sub-network together with the time difference of arrival value; in the deep regression positioning sub-network, end-to-end regression is performed based on time delay consistency and geometric residual error constraint to output the discharge source predicted three-dimensional coordinates and the corresponding regression uncertainty.

[0012] Optionally, the constructing the confidence interval and outputting the visual positioning result with the ellipsoid confidence boundary through the discharge source predicted three-dimensional coordinates, the regression uncertainty and the source confidence tensor specifically comprises: fusing the regression uncertainty and the source confidence tensor to generate a fused covariance matrix; performing eigenvalue decomposition on the fused covariance matrix to obtain ellipsoid principal axis directions and principal axis radii, and combining the ellipsoid principal axis directions and the principal axis radii with the discharge source predicted three-dimensional coordinates to form a three-dimensional ellipsoid parameter set; mapping the three-dimensional ellipsoid parameter set into a two-dimensional ellipse visual element according to a view matrix and a projection matrix of a terminal visual interface, and calculating color parameters and transparency parameters in combination with the source confidence tensor and the regression uncertainty; judging the spatial relationship inside and outside the ellipsoid confidence boundary through Mahalanobis distance, and generating a smooth trajectory and an ellipsoid confidence boundary changing over time in a continuous time window by using the discharge source predicted three-dimensional coordinates and the fused covariance matrix, to obtain the visual positioning result.

[0013] In a second aspect of the present application, a GIS partial discharge positioning device based on deep learning and acoustic-electric combination is provided. The device comprises an acquisition module and a processing module. The acquisition module is configured to acquire original ultra-high frequency signals and original ultrasonic signals of a target detection entity, and generate an acoustic-electric signal matrix. The processing module is configured to perform short-time Fourier transform on the acoustic-electric signal matrix to obtain a complex spectrum tensor and a power spectrum tensor. The processing module is further configured to input the complex spectrum tensor and the power spectrum tensor as inputs, input an ultra-high frequency signal branch and an ultrasonic signal branch into a deep learning network respectively for double-branch feature extraction, and output a high-dimensional feature embedding tensor. The processing module is further configured to generate a time-frequency mask tensor and a source confidence tensor of each potential partial discharge source by a multi-source separation sub-network according to the high-dimensional feature embedding tensor, and perform source-by-source spectral reconstruction based on the time-frequency mask tensor and the source confidence tensor to obtain a source-by-source time-domain waveform sequence. The processing module is further configured to calculate the source-by-source time-domain waveform sequence by a phase transformation generalized cross-correlation method to obtain a time difference of arrival matrix, and input the time difference of arrival matrix and sensor space coordinates of a detection sensor corresponding to the target detection entity into a deep regression positioning sub-network to output discharge source predicted three-dimensional coordinates and regression uncertainty. The processing module is further configured to construct a confidence interval and output a visual positioning result with an ellipsoid confidence boundary through the discharge source predicted three-dimensional coordinates, the regression uncertainty and the source confidence tensor.

[0014] In a third aspect of the present application, an electronic device is provided. The electronic device comprises a processor, a memory, a user interface and a network interface. The memory is configured to store instructions. The user interface and the network interface are configured to communicate with other devices. The processor is configured to execute the instructions stored in the memory, so that the electronic device performs the method described above.

[0015] A computer-readable storage medium is provided in a fourth aspect of the present application, and the computer-readable storage medium stores instructions which, when executed, perform the method described above.

[0016] To sum up, the one or more technical solutions provided in the present application have at least the following technical effects or advantages:

[0017] Through the continuous association of technical features such as acoustic-electric combined acquisition, double-branch deep feature extraction, multi-source separation with physical constraints, source-by-source time difference calculation, deep regression positioning and uncertainty visualization, a complete closed-loop partial discharge positioning process is established. This scheme can stably distinguish and accurately locate multiple discharge sources under complex working conditions with multiple defects active at the same time and serious signal overlap, while outputting positioning results with ellipsoid confidence boundaries, which facilitates improving the accuracy of partial discharge positioning. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 A flowchart of a GIS partial discharge positioning method based on deep learning and acoustic-electric combination is provided for the embodiments of the present application;

[0019] Figure 2 A module schematic diagram of a GIS partial discharge positioning device based on deep learning and acoustic-electric combination is provided for the embodiments of the present application;

[0020] Figure 3 A structural schematic diagram of an electronic device is provided for the embodiments of the present application.

[0021] Explanation of reference signs: 21, acquisition module; 22, processing module; 31, processor; 32, communication bus; 33, user interface; 34, network interface; 35, memory. DETAILED DESCRIPTION

[0022] In order for those skilled in the art to better understand the technical solutions in the present specification, the technical solutions in the embodiments of the present specification will be described clearly and completely below in conjunction with the drawings in the embodiments of the present specification. Obviously, the described embodiments are only some of the embodiments of the present application, not all.

[0023] In the description of the embodiments of the present application, the words such as “for example” or “for instance” are used to represent an example, illustration or description. Any embodiment or design scheme described as “for example” or “for instance” in the embodiments of the present application should not be interpreted as more preferred or more advantageous than other embodiments or design schemes. Rather, the words such as “for example” or “for instance” are intended to present the relevant concept in a specific manner.

[0024] In the description of the embodiments of the present application, the term "a plurality of" means two or more. For example, a plurality of systems means two or more systems, and a plurality of screen terminals means two or more screen terminals. In addition, the terms "first", "second" are only for the purpose of description, and cannot be understood as indicating or implying relative importance or implicitly indicating the indicated technical features. Therefore, the features defined with "first", "second" can be explicitly or implicitly included one or more of the features. The terms "include", "contain", "have" and their variants mean "include but not limited to", unless otherwise specifically emphasized.

[0025] To solve the above technical problems, the present application provides a GIS partial discharge positioning method based on deep learning and sound-electricity combination. Referring to Figure 1 , Figure 1 A flowchart of a GIS partial discharge positioning method based on deep learning and sound-electricity combination provided by the embodiments of the present application. The method is applied to a server, which includes steps S110 to S160, and the above steps are as follows:

[0026] S110, obtaining original ultra-high frequency signals and original ultrasonic signals for a target detection entity, and generating a sound-electricity signal matrix.

[0027] Specifically, the server refers to a computing platform unit that undertakes data collection, time synchronization, data storage, preprocessing, model inference, and result publishing. It can be a single industrial cabinet server, or a cluster composed of multiple high-availability nodes. The server has a high-speed data interface with ultra-high frequency sensors and ultrasonic sensors, as well as unified time-granting capability. For example, a cabinet-level server deployed in a GIS control room can issue a time reference to all sensors through a synchronous Ethernet time grant, and continuously receive waveform data through a high-throughput channel. The target detection entity refers to a specific unit of equipment within the monitoring range that needs to complete partial discharge positioning, or a spatial sub-area of the unit. The target detection entity includes circuit breaker cabin sections, disconnecting switch cabin sections, busbar cabin sections, bushing end sections, transformer joint sections, etc. It can also be a functional area composed of multiple adjacent cabin sections. For example, the area where the busbar and circuit breaker of the same interval are connected can be regarded as a target detection entity.

[0028] Raw UHF signal refers to the discrete time-domain waveform of electromagnetic radiation from partial discharge, which is directly collected by UHF sensors after propagating in the metal shell and mixed gas medium. The signal is not filtered, down-sampled, interpolated, or gain-corrected. The typical frequency band range covers hundreds of megahertz to gigahertz. The waveform features include narrow pulse trains, fast decay envelope, and multi-path echoes, etc. For example, when a partial discharge occurs on a busbar support insulator, two adjacent UHF sensors may record pulse clusters with different amplitudes within a time difference of microseconds. Raw ultrasonic signal refers to the discrete time-domain waveform of mechanical shock and gas acoustic pressure disturbance caused by partial discharge, which is directly collected by ultrasonic sensors after propagating in the gas medium and solid support. The signal is not subjected to any digital processing. The typical frequency band range covers twenty kilohertz to several hundred kilohertz. The waveform features include narrowband ringing, echo tailing, and structure resonance, etc. For example, when a micro-discharge occurs near the contact of a circuit breaker, an ultrasonic sensor close to the flange may record a ringing segment lasting for several milliseconds.

[0029] Acoustic-electric signal matrix refers to the multi-channel data structure constructed by the server for unified management of raw UHF signals and raw ultrasonic signals. The matrix organizes all channel samples in time axis alignment. The row index corresponds to the sampling time, and the column index corresponds to the sensor channel number. The matrix unit records the amplitude of the channel at the sampling time. At the same time, the accompanying metadata of the matrix retains the channel index, sensor spatial coordinates, unified time axis, sampling rate identifier, and time accuracy identifier. For example, after four UHF sensors and four ultrasonic sensors continuously collect for ten seconds, an acoustic-electric signal matrix with a sample point number multiplied by eight can be formed, which is one-to-one corresponding to eight three-dimensional coordinate entries.

[0030] Acquisition refers to the process by which the server receives and stores raw UHF signals and raw ultrasonic signals in real time under a unified time reference through a high-speed acquisition interface connected to the sensors at a set sampling rate. The acquisition process simultaneously records the channel number, sensor spatial coordinates, timestamp, and trigger marker. For example, the server receives data from each UHF channel at a rate of ten million samples per second; receives data from each ultrasonic channel at a rate of five hundred thousand samples per second, and adds nanosecond-level timestamps to all samples. Generation refers to the process by which the server, after completing timestamp alignment, necessary resampling, packet filling, and channel ordering verification, splices the two types of raw waveforms onto the same time axis and matches the unified index, thereby forming an acoustic-electric signal matrix. For example, after the server performs alignment based on time reference for all channels, the samples are arranged on a common time grid, and the channel gain identifier and spatial coordinate validity are verified, finally outputting an acoustic-electric signal matrix that can directly enter the short-time Fourier transform step.

[0031] In a possible implementation, the original ultrahigh frequency signal and the original ultrasonic signal for the target detection entity are acquired, and an acoustic-electric signal matrix is generated, specifically including: acquiring the original ultrahigh frequency signal and the original ultrasonic signal sent by the ultrahigh frequency sensor and the ultrasonic sensor pre-deployed on the target detection entity; attaching a high-precision timestamp, a sensor spatial coordinate, and a channel identifier to each corresponding sampling point of the original ultrahigh frequency signal and the original ultrasonic signal; time stamping the original ultrahigh frequency signal and the original ultrasonic signal, constructing a multi-channel data structure with the sampling time as the row index, the sensor channel number as the column index, and the matrix unit value as the corresponding amplitude, and determining the multi-channel data structure as the acoustic-electric signal matrix.

[0032] Specifically, in the acquisition stage, the server issues a unified time reference and sampling parameters to the ultrahigh frequency sensor and the ultrasonic sensor pre-deployed on the target detection entity and starts synchronous collection. The ultrahigh frequency sensor works at a sampling frequency and bandwidth and dynamic range that meet the observation requirements of the frequency band of 100 MHz to 1 GHz. The ultrasonic sensor works at a sampling frequency and bandwidth and dynamic range that meet the observation requirements of the frequency band of 20 kHz to several hundred kHz. An anti-aliasing filter and a low-noise preamplifier are used in the front end. The collected original ultrahigh frequency signal and the original ultrasonic signal are real-time returned to the ring buffer of the server through a high-speed link and are branched and buffered according to the channel identifier. At the same time, the reference time pulse and the trigger mark are recorded to ensure cross-channel synchronization. In order to form a traceable time index, a high-precision timestamp is generated for each sampling point of each channel. The timestamp is determined by the unified time reference, the channel local sampling frequency, and the residual clock offset. The calculation relationship is:

[0033]

[0034] Where τ m，j represents the timestamp of the jth sampling point of the mth channel, t0 represents the time of the last time reference, b m represents the residual clock offset of the mth channel relative to the time reference, F s，m represents the sampling frequency of the mth channel, and ∈ m，j represents a small random error introduced by sampling jitter. This step outputs the original ultrahigh frequency signal and the original ultrasonic signal with timestamps and channel identifiers as inputs for subsequent additional spatial information and time alignment.

[0035] In the stage of adding high-precision timestamps, sensor spatial coordinates, and channel identifiers, the aforementioned timestamps and channel identifiers are written into the sample metadata, and the sensor spatial coordinates are calculated according to the installation calibration record. The sensor spatial coordinates are represented by the three-dimensional device coordinate system of the target detection entity and are obtained by rigid transformation mapping of the installation pose calibration. The calculation relationship is:

[0036]

[0037] where p m denotes the three-dimensional sensor space coordinate of the mth channel, denotes the local three-dimensional coordinate of the channel on the mounting base, R denotes the three-dimensional rotation matrix from the mounting base coordinate system to the device coordinate system, and t denotes the corresponding three-dimensional translation vector. The server also establishes a one-way mapping of the channel identifier to the column index to guarantee the fixed column sequence of the subsequent data structure. If a packet is found to be missing or duplicated, the missing packet marking and duplicate removal are performed based on the timestamp as the primary key, thereby obtaining the normalized data stream of the original ultra-high frequency signal and the original ultrasonic signal with the timestamp, the sensor space coordinate, and the channel identifier.

[0038] In the timestamp alignment and multi-channel data structure construction stage, the unified sampling rate and the unified time axis are first determined, and the offset and drift correction of the timestamps of each channel are performed. The unified time axis defines a discrete time grid with the starting time and the unified sampling rate, and is denoted as t n The equally spaced sequence is calculated as follows:

[0039]

[0040] where n m，j denotes the row index of the jth sampling point of the mth channel on the unified time axis, F s denotes the unified sampling rate, denotes the channel residual offset correction quantity estimated by the time service protocol or the calibration process, denotes the channel clock drift correction coefficient estimated by the time service protocol or the calibration process. Subsequently, the non-isochronous time samples of each channel are mapped to the unified time grid by an interpolation operator to obtain the aligned amplitude sequence and assemble it into a matrix X, which is defined as:

[0041]

[0042] where X[n, m] denotes the matrix element value of the acoustoelectric signal matrix at the nth row and the mth column, and is the amplitude of the corresponding channel at the nth sampling time on the unified time axis, denotes the amplitude of the mth channel on the unified time axis after offset and drift correction and interpolation. The row index of the matrix is the sampling time, and the column index is the sensor channel number. The accompanying metadata saves the channel identifier, the sensor space coordinate, the unified time axis, the unified sampling rate, and the time service accuracy identifier. Thus, the multi-channel data structure is determined as the acoustoelectric signal matrix, which is the only input for the subsequent short-time Fourier transform and double-branch feature extraction, thereby realizing strict connection and semantic consistency of the technical features before and after the acquisition and additional metadata, time alignment, and matrix construction at the data stream level.

[0043] S120, performing a short-time Fourier transform on the acoustoelectric signal matrix to obtain a complex spectrum tensor and a power spectrum tensor.

[0044] Specifically, the acoustic-electric signal matrix is a multi-channel data structure for managing original ultrahigh frequency signals and original ultrasonic signals in time axis alignment and channel axis grouping, with row index being sampling time, column index being sensor channel number, and matrix cell value being amplitude of corresponding channel at corresponding sampling time, while having accompanying metadata to record sensor spatial coordinates, unified time axis, unified sampling rate, and time accuracy identification. An engineering example is an eight-channel matrix, four ultrahigh frequency channels and four ultrasonic channels, with a unified sampling rate of one million samples per second, and a one-second data block is sent to the server's short-time Fourier transform pipeline in units of continuous one-second data blocks for windowing processing.

[0045] Short-time Fourier transform is a processing step of performing local spectral analysis on time-domain waveforms within a sliding time window, by framing, windowing, and discrete Fourier transform on time-domain sequences, to obtain a local spectral representation with time index and frequency index. The complex spectrum tensor is a tensor structure formed by stacking the complex spectrum values of each channel at each time frame and each frequency point according to channel index, frequency index, and time index, which completely retains amplitude information and phase information as well as three-axis index mapping relationship, and is used for subsequent phase-sensitive time difference of arrival estimation and cross-modal feature alignment. The power spectrum tensor is an energy distribution representation obtained by amplitude squaring and normalization correction with the complex spectrum tensor as input, which is used for robust energy prior and noise suppression evaluation, and is strictly aligned with the complex spectrum tensor in index.

[0046] In one possible implementation, the short-time Fourier transform is performed on the acoustic-electric signal matrix to obtain the complex spectrum tensor and the power spectrum tensor, specifically including: before performing the short-time Fourier transform, binding the channel set belonging to the ultrahigh frequency signal and the channel set belonging to the ultrasonic signal in the acoustic-electric signal matrix to corresponding sampling rate, sensor calibration coefficient, and channel gain factor respectively; performing bandpass anti-aliasing filtering and resampling processing on channels with different sampling rates to align to a unified sampling rate and keep the channel index, spatial coordinates, and time axis consistent; configuring window length, frame shift, zero padding length, and window function type for different types of signals respectively, and using overlapping windowing strategy for framing, windowing, and fast Fourier transform to obtain the complex spectrum tensor containing amplitude information and phase information; performing window energy correction, channel gain correction, and sensor sensitivity correction on the complex spectrum tensor to generate the power spectrum tensor.

[0047] Specifically, the accompanying metadata of the acoustic-electric signal matrix is read by the server when binding the sampling rate, the sensor calibration coefficient and the channel gain factor, and three sets of parameters are written respectively for the channel set belonging to the ultrasonic signal and the channel set belonging to the ultrasonic signal, the sampling rate is used to define the time discrete grid, the sensor calibration coefficient is used to convert the voltage amplitude or digital count to the physical quantity amplitude, and the channel gain factor is used to offset the fixed amplification multiple of the front-end amplification link and the difference of the digital-analog conversion scale. After completing the parameter binding, a unified channel index, a unified sensor spatial coordinate and a unified time axis mapping are established on all channels to ensure that the subsequent processing always runs with a consistent index system. To ensure the traceability of subsequent calculations, the server also solidifies the parameter snapshot, and writes the parameter version number and the time reference label in the header of each channel.

[0048] In the band-pass anti-aliasing filtering and resampling processing, for channels with different sampling rates, band-pass anti-aliasing filtering is first performed at the respective original sampling rates to limit the effective frequency band, and then interpolation alignment is performed according to the unified sampling rate. The discrete time-domain implementation of the band-pass anti-aliasing filtering is:

[0049]

[0050] wherein represents the filtered time-domain sequence of the mth channel, x m represents the original time-domain sequence of the mth channel, h m represents the band-pass anti-aliasing filter impulse response of the mth channel, A m represents the length of the impulse response, and n represents the discrete time index.

[0051] The filtered sequence is mapped to the unified sampling rate F s Resampling is achieved by using band-windowed sinc interpolation:

[0052]

[0053] wherein represents the time-domain sample of the mth channel at the unified sampling rate, F s，m represents the original sampling rate of the mth channel, sinc(·) represents the normalized sinc interpolation kernel, w(·) represents the finite support window function kernel, j represents the original sampling grid index, and n represents the unified sampling grid index. After completing the resampling, the channel index, the sensor spatial coordinate and the unified time axis are kept consistent, and the time stamp is used as the primary key to check whether there is a missing packet or a duplicate packet.

[0054] The window length, frame shift, zero padding length and window function type are set for different types of signals in the window configuration and frame, and the window function type is set for different types of signals in the window configuration and frame. The continuous frame column is formed by using the overlapping window strategy. The short-time Fourier transform is performed on the mth channel under the unified sampling rate, and the complex spectrum is defined as:

[0055]

[0056] Where X m [k,n] represents the complex spectrum value of the kth frequency point of the nth frame of the mth channel, represents the time domain sequence after resampling, w m [t] represents the window function of the mth channel, L m represents the window length, H m represents the frame shift, N m represents the fast Fourier transform length, k represents the frequency index, n represents the time frame index, t represents the window time sample index, and j represents the imaginary unit.

[0057] The complex spectrum tensor is obtained by stacking all the X m [k,n] according to the channel index, frequency index and time index, and the physical index mapping is given:

[0058]

[0059] Where f m，k represents the physical frequency corresponding to the kth frequency point of the mth channel, τ m，n represents the center time of the nth frame of the mth channel, F s represents the unified sampling rate. The overlap rate of the overlapping window strategy is defined as and is used to constrain the time overlap ratio of adjacent frames, so as to balance the time resolution and frequency resolution.

[0060] When generating the power spectrum tensor, the complex spectrum tensor is subjected to window energy correction, channel gain correction and sensor sensitivity correction, and the amplitude energy is converted into physical consistent power dimension. The window energy correction coefficient is defined as:

[0061]

[0062] Where G m represents the window power gain of the mth channel, w m [t] represents the window sample of the mth channel, L m represents the window length. The channel gain correction factor is denoted as A m and is used to offset the fixed amplification, and the sensor sensitivity correction coefficient is denoted as S m and is used to convert the voltage dimension or the quantity dimension to the physical dimension, and the power spectrum tensor element is defined as:

[0063]

[0064] where P m [k, n] represents the power spectrum tensor element of the mth channel, the nth frame, and the kth frequency bin, |X m [k, n] represents the complex spectrum modulus, Θ m represents the comprehensive correction coefficient and is given by A m represents the channel gain factor, S m represents the sensor calibration coefficient. Via this definition, a power spectrum tensor strictly aligned with the channel index, the frequency index, and the time index is obtained from the complex spectrum tensor, which is used to retain the amplitude information and the phase information, and the power spectrum tensor is used to provide the energy prior and the noise evaluation, both of which serve as the upstream input of the subsequent double-branch deep feature extraction and multi-source separation, and maintain a mapping relationship consistent with the sound and electrical signal matrix at the index level to ensure data flow continuity and physical semantic consistency.

[0065] S130, taking the complex spectrum tensor and the power spectrum tensor as inputs, inputting the ultrahigh frequency signal branch and the ultrasonic signal branch into a deep learning network for double-branch feature extraction, and outputting a high-dimensional feature embedding tensor.

[0066] Specifically, the ultrahigh frequency signal branch is a calculation path in the deep learning network that is specifically designed to process the characteristics of ultrahigh frequency signals. This branch focuses on capturing pulse-type distribution, rapid phase change, and multi-path interference patterns, and usually uses a combination of complex convolution units, time-frequency convolution units, and channel attention units, taking the real part and the imaginary part of the complex spectrum tensor and the amplitude of the power spectrum tensor as multi-channel input. An example is that four ultrahigh frequency channels output a set of intermediate feature tensors aligned with the channel index, the frequency index, and the time index after being processed by this branch, which are used for cross-modal fusion. The ultrasonic signal branch is a calculation path in the deep learning network that is specifically designed to process the characteristics of ultrasonic signals. This branch focuses on capturing narrowband ringing, echo tail, and structural resonance patterns, and usually uses a combination of time series convolution units and multi-head self-attention units, taking the real part and the imaginary part of the complex spectrum tensor and the amplitude of the power spectrum tensor as multi-channel input. An example is that another set of intermediate feature tensors aligned with the channel index, the frequency index, and the time index with the ultrahigh frequency signal branch are output after being processed by this branch.

[0067] The deep learning network is an end-to-end model composed of dual-branch feature extraction, cross-modal fusion and residual shaping. After the two branches complete early representation and normalization respectively, cross-modal fusion is performed based on frequency index and time index, and index alignment and feature stability are maintained through residual shaping. An example is a model containing a complex convolution dry end, an attention fusion middle end and a light residual head, with input of complex spectrum tensor and power spectrum tensor, and output of high-dimensional feature embedding tensor. Dual-branch feature extraction refers to the process of parallel feature learning of the same batch of complex spectrum tensor and power spectrum tensor by the ultrahigh frequency signal branch and the ultrasonic signal branch. The two branches share aligned channel index, frequency index and time index, but use differentiated convolution kernels and attention mechanisms to match the physical characteristics of their respective signals. An example is that both branches stack three to five time-frequency modules, outputting intermediate feature tensors that can be directly spliced at the same time index and frequency index.

[0068] The high-dimensional feature embedding tensor is a unified feature representation formed after cross-modal fusion and residual shaping of the deep learning network. The tensor is strictly aligned with the input in channel index, frequency index and time index, but is significantly upscaled in channel dimension or feature dimension to carry multiple physical and statistical patterns; meanwhile, it retains geometric position encoding and channel identity encoding, facilitating direct consumption by downstream multi-source separation and time difference of arrival estimation. An example is a four-dimensional tensor with a size of eight by twenty-five hundred and sixty by four thousand by one hundred and twenty-eight, where one hundred and twenty-eight is the feature dimension, which is used as the only input for the multi-source separation subnetwork.

[0069] In one possible implementation, the complex spectrum tensor and the power spectrum tensor are taken as input, and the ultrahigh frequency signal branch and the ultrasonic signal branch are input into the deep learning network for dual-branch feature extraction, outputting a high-dimensional feature embedding tensor, specifically including: performing normalization processing, geometric prior encoding and early convolution feature representation on the complex spectrum tensor and the power spectrum tensor in the ultrahigh frequency signal branch and the ultrasonic signal branch, and introducing a spatial coordinate embedding generated by the sensor spatial coordinates in each branch to retain spatial position information; in the ultrahigh frequency signal branch, phase correlation features are extracted through complex convolution, complex normalization and complex nonlinear units, and the phase correlation features are fused with the power spectrum tensor to form a first intermediate feature tensor of the ultrahigh frequency signal branch; in the ultrasonic signal branch, narrowband ringing and echo tail features are extracted using a method combining time series convolution and multi-head self-attention, and the narrowband ringing and echo tail features are fused with the power spectrum tensor to form a second intermediate feature tensor of the ultrasonic signal branch; in the multi-scale representation stage, different spread width time-frequency patterns are captured through hollow time-frequency convolution, and channel attention mechanism is used to re-label the multi-scale features to obtain re-labeled features; the re-labeled features, the first intermediate feature tensor and the second intermediate feature tensor are input into a residual shaping network to generate a high-dimensional feature embedding tensor.

[0070] Specifically, in the normalization processing and geometric prior encoding and early convolution feature representation stage, the server performs intra-channel standardization and index alignment on the complex spectrum tensor and the power spectrum tensor in the ultrahigh frequency signal branch and the ultrasonic signal branch, and constructs a spatial coordinate embedding to retain spatial position information. The real part and the imaginary part of the complex spectrum tensor are respectively subjected to intra-branch instance normalization, and the calculation relationship is:

[0071]

[0072] wherein and are the real part and the imaginary part of the complex spectrum tensor of the mth channel, the nth frame and the kth frequency point, and are the mean values of all time indexes and frequency indexes on the channel, and are the corresponding standard deviations, and ∈ is a numerical stability constant. The power spectrum tensor is subjected to order-of-magnitude compression and zero-mean unit-variance normalization, and the calculation relationship is:

[0073]

[0074] wherein P m [k, n] is the power spectrum tensor element, δ is a small lower limit constant, and are the mean value and the standard deviation in the logarithmic domain, respectively. The geometric prior encoding constructs a spatial coordinate embedding with the sensor spatial coordinates and the channel index, and the calculation relationship is

[0075] E geo (m) = φ([p m , m]) = W2σ(W1[p m , m] + b1) + b2

[0076] wherein p m is the three-dimensional sensor spatial coordinates of the mth channel, m is the numerical representation of the channel index, φ(·) is a two-layer perceptron mapping, W1 and W2 are weight matrices, b1 and b2 are biases, and σ(·) is a nonlinear function. E geo (m) is broadcast to all frequency indexes and time indexes by channel, and is concatenated with and and In the channel dimension, small kernel time-frequency convolution is used as early convolution feature representation to extract local texture while keeping the channel index and the frequency index and the time index consistent.

[0077] In the phase correlation feature extraction and fusion stage of the ultrahigh frequency signal branch, complex convolution, complex normalization and complex nonlinear unit are used to represent phase coupling and multipath interference, and the calculation relationship of the complex convolution is:

[0078]

[0079] where W R and W I are real and imaginary weights of the complex kernel, b R and b I are biases, * is two-dimensional convolution. Complex normalization adopts joint amplitude and phase standardization, with the computational relationship:

[0080]

[0081] where and are real and imaginary parts, and and are branch-wise statistics, γ R and γ I are learnable scales, β R and β I are learnable translations. Complex nonlinearities adopt modReLU with length gating, with the computational relationship:

[0082]

[0083] where is a complex representation, is a complex length, and b is a learnable threshold. The length of the phase-related feature is gated with the power spectrum tensor to suppress weak-energy phase noise, with the computational relationship:

[0084]

[0085] where F UHF is the first intermediate feature tensor of the ultrahigh frequency signal branch, and are real and imaginary parts, σ(·) is Sigmoid, ⊙ is element-wise multiplication, W g and b g are gating mappings. The first intermediate feature tensor maintains a one-to-one correspondence with the input in the channel index, frequency index, and time index.

[0086] In the narrowband ringing and echo tail feature extraction and fusion stage of the ultrasonic signal branch, short-term time-domain patterns are extracted by time convolution, and long-range dependencies are represented by multi-head self-attention, and then energy priori fusion is performed with the power spectrum tensor. The computational relationship of the time convolution is:

[0087]

[0088] where Conv t is one-dimensional or two-dimensional small kernel convolution along the time index, and U is a short-term representation. The computational relationship of the multi-head self-attention is:

[0089]

[0090] where Q = U W Q and K = U W K and V = U W V is the query and key-value mapping, W Q and W K and W V is the learnable weight, d k is the key vector dimension, B τ is the relative position bias matrix generated by the time index. The calculation relationship of energy prior fusion is:

[0091]

[0092] where F US is the second intermediate feature tensor of the ultrasonic signal branch, W e and b e is the energy gating mapping, σ(·) is Sigmoid, and is element-wise multiplication. The second intermediate feature tensor and the first intermediate feature tensor are strictly aligned in the channel index, frequency index, and time index.

[0093] In the multi-scale representation and channel attention re-scaling stage, the first intermediate feature tensor and the second intermediate feature tensor are concatenated in the channel dimension and then sent to a set of dilated time-frequency convolution groups to capture time-frequency patterns of different spreading widths, and the multi-scale features are re-scaled by a channel attention mechanism to obtain re-scaled features. The calculation relationship of dilated time-frequency convolution is:

[0094]

[0095] where Conv ds is a two-dimensional convolution with a dilation rate d s in the time index and the frequency index, ψ(·) is a nonlinearity, U (s) is the output of the s-th dilation rate branch. The calculation relationship of channel attention is:

[0096] α = σ(W2δ(W1GAP(Concat s U (s) ))), F CAL = α ⊙ Concat s U (s)

[0097] where GAP(·) is global average pooling, δ(·) is an intermediate layer nonlinearity, W1 and W2 are fully connected weights, σ(·) is Sigmoid, α is the channel attention weight vector, is channel-wise scaling, F CALThe re-calibration feature is indexed to align with the two-branch intermediate features, ensuring splicing and traceability.

[0098] In the residual shaping and output high-dimensional feature embedding tensor stage, the re-calibration feature is input into the residual shaping network after being spliced with the first intermediate feature tensor and the second intermediate feature tensor in the channel dimension. The residual shaping unit is composed of small kernel time-frequency convolution, normalization and nonlinearity, and maintains feature stability with an identity shortcut branch. The calculation relationship is:

[0099]

[0100] wherein is a residual mapping composed of several layers of convolution, normalization and nonlinearity, and H is the shaped time-frequency feature tensor. Finally, a linear dimension lifting mapping is used to output the high-dimensional feature embedding tensor, and the calculation relationship is:

[0101] T = HW proj +b proj

[0102] wherein T is the high-dimensional feature embedding tensor and is strictly aligned with the input in the channel index, frequency index and time index, W proj is the dimension lifting projection weight, and b proj is the projection bias. The high-dimensional feature embedding tensor is the only input of the subsequent multi-source separation sub-network, directly driving the generation of the time-frequency masking tensor and the source confidence tensor, and fully inheriting the physical semantics and index relationship of the complex spectrum tensor and the power spectrum tensor at the data flow level.

[0103] S140, according to the high-dimensional feature embedding tensor, generating the time-frequency masking tensor and the source confidence tensor of each potential partial discharge source through the multi-source separation sub-network, and performing source-by-source spectrum reconstruction based on the time-frequency masking tensor and the source confidence tensor, to obtain the source-by-source time-domain waveform sequence.

[0104] Specifically, the multi-source separation sub-network is a deep learning sub-network that takes the high-dimensional feature embedding tensor as input and simultaneously performs time-frequency layer separation estimation for multiple potential partial discharge sources. It uses a shared backbone and source-specific output structure, and utilizes gated time-frequency convolution, attention mechanism and physical consistency constraints to output the time-frequency masking tensor and the source confidence tensor. The goal of the multi-source separation sub-network is to give a distinguishable representation of each potential partial discharge source in the time-frequency plane even when the signals overlap severely, the noise is significant and the multipath is prominent. For example, in the case of two defects discharging simultaneously in the same frame, the multi-source separation sub-network can give different masking assignments for the two sources in the overlapping frequency band.

[0105] The time-frequency masking tensor is the selective weight estimated by the multi-source separation subnetwork for each potential partial discharge source in the time-frequency plane. The common forms include complex domain masking and amplitude domain masking, and the numerical range is limited to ensure numerical stability. The time-frequency masking tensor corresponds to the complex spectrum tensor one by one in the channel index, frequency index, and time index, and is used to perform element-wise selective enhancement and suppression on the complex spectrum tensor, so that the spectral components of each potential partial discharge source are distinguished in the overlapping area. For example, when two sources have similar energy in the same frequency band but obvious phase difference, the complex domain masking can achieve effective distinction based on the phase difference. For engineering implementation, the complex domain masking can be recorded as two tensors of real part masking and imaginary part masking, and the amplitude domain masking can be recorded as one tensor of amplitude masking.

[0106] The source confidence tensor is the reliability quantization result of the time-frequency masking tensor in the multi-source separation subnetwork in each channel, each frequency point, and each time frame, and the value is between zero and one, which is used to represent the confidence degree of the corresponding masking in the time-frequency unit. The source confidence tensor is usually obtained by fusing classification confidence and uncertainty estimation, and can be used as a time-frequency weight in subsequent source-by-source spectrum reconstruction and time difference of arrival estimation. For example, in a strong interference frame or a sensor saturation frame, the source confidence tensor will significantly decrease in the corresponding time-frequency unit, thereby reducing the influence of this area on the reconstruction result. The source-by-source spectrum reconstruction is a process of weighting the complex spectrum tensor by using the time-frequency masking tensor and the source confidence tensor to obtain the complex spectrum estimation of each potential partial discharge source in each channel, and restoring it to the time domain through inverse short-time Fourier transform.

[0107] The source-by-source time-domain waveform sequence is a set of time-domain waveforms obtained by inverse short-time Fourier transform of the channel-by-channel complex spectrum estimation, which is organized according to the channel index, aligned on the same time axis, and consistent with the time index of the source confidence tensor. The source-by-source time-domain waveform sequence directly serves the accurate estimation of the time difference of arrival matrix and subsequent geometric positioning, and can be used for engineering processing such as abnormal frame rejection and multipath identification. For example, in the case of two defects discharging at the same time and partially overlapping, the source-by-source time-domain waveform sequence can respectively represent two groups of pulse trains with consistent phase and stable time difference of arrival, thereby providing clean input for subsequent phase transformation generalized cross-correlation.

[0108] In one possible implementation, according to the high-dimensional feature embedding tensor, the time-frequency mask tensor and the source confidence tensor of each potential partial discharge source are generated by the multi-source separation sub-network, and the source-by-source spectral reconstruction is performed based on the time-frequency mask tensor and the source confidence tensor to obtain the source-by-source time-domain waveform sequence, specifically comprising: using the shared backbone and the source-specific output structure of the multi-source separation sub-network, the high-dimensional feature embedding tensor is encoded to generate the time-frequency mask tensor containing the real part mask, the imaginary part mask and the amplitude mask; using the uncertainty output structure of the multi-source separation sub-network, the mask estimation and the intermediate features in the network are combined to generate the source confidence tensor; in the source-by-source spectral reconstruction stage, the complex spectrum tensor and the power spectrum tensor are combined with the time-frequency mask tensor and the source confidence tensor to perform consistent weighted reconstruction in the complex domain to obtain the complex spectrum estimation of each potential partial discharge source, and the source-by-source time-domain waveform sequence is restored through the inverse short-time Fourier transform.

[0109] Specifically, in the encoding stage of the shared backbone and the source-specific output structure, the server sends the high-dimensional feature embedding tensor into the shared backbone of the multi-source separation sub-network to extract the public representation across channels, frequencies and time, and then sends it into the source-specific output structure of each potential partial discharge source to generate three components of the time-frequency mask tensor, i.e. the real part mask, the imaginary part mask and the amplitude mask. Taking the kth potential partial discharge source as an example, the source-specific output structure performs linear transformation or small kernel convolution on the high-dimensional feature embedding tensor and applies bounded nonlinearity to obtain:

[0110]

[0111] wherein, is the real part mask, is the imaginary part mask, is the amplitude mask; m is the channel index; f is the physical frequency corresponding to the frequency index; τ is the frame center time corresponding to the time index; is the unbound mask estimation tensor (generated by the source-specific output structure); T>0 is the temperature parameter for adjusting the mask smoothness; tanh(·) is the hyperbolic tangent function with the output range [-1, 1]; σ(·) is the Sigmoid function with the output range [0, 1]. The output of this step is strictly aligned with the complex spectrum tensor and the power spectrum tensor in the channel index, the frequency index and the time index, and keeps consistent with the spatial coordinate embedding of the high-dimensional feature embedding tensor. The engineering example is that the shared backbone uses the gated time-frequency convolution and the layer normalization, and the source-specific output structure is a three-way one-multiplication convolution head to generate the real part mask, the imaginary part mask and the amplitude mask respectively.

[0112] In the uncertainty output structure generation source confidence tensor stage, the server jointly encodes the masking prediction and the network intermediate feature to obtain a reliability measure for each time-frequency unit, thereby forming a source confidence tensor. Taking the kth potential partial discharge source as an example, the source confidence tensor is defined as

[0113]

[0114] wherein C k ∈ [0, 1] is an element of the source confidence tensor; γ ∈ [0, 1] is a fusion weight; H(·) is an information entropy operator; P k is a class distribution or quality distribution obtained based on the intermediate feature; C is a discrete state number, used for normalizing the information entropy; U k is a prediction uncertainty based on Monte Carlo dropout or feature variance; β > 0 is an uncertainty attenuation coefficient. This structure unifies the “classification confidence” and “model uncertainty” into a scalar weight, which is used to weight the time-frequency units in the subsequent source-by-source spectrum reconstruction. An engineering example is: performing multiple random dropout sampling during inference, and calculating the variance of the masking output to obtain U k , and mapping the intermediate feature to P k through a fully connected layer.

[0115] In the complex domain consistency weighted source-by-source spectrum reconstruction stage, the server combines the complex spectrum tensor and the power spectrum tensor, and the time-frequency masking tensor and the source confidence tensor, generates a source-by-source complex spectrum estimate for each potential partial discharge source, and restores it to a source-by-source time-domain waveform sequence through inverse short-time Fourier transform. Taking the kth potential partial discharge source and channel m as an example, the complex domain consistency weighted reconstruction is defined as:

[0116]

[0117] wherein X is a source-by-source complex spectrum estimate; X m (f, τ) is an element of the complex spectrum tensor in channel m; α ≥ 0 is an amplitude compensation weight, used to balance the phase correction and amplitude gating; ε > 0 is a numerical stability constant; |·| is a complex modulus; j is an imaginary unit. Then, the inverse short-time Fourier transform is performed using the window function, window length and frame shift completely consistent with the forward short-time Fourier transform, to obtain a source-by-source time-domain waveform sequence:

[0118]

[0119] wherein x is a source-by-source time-domain waveform sequence; ISTFT (L，H，w)denotes the inverse short-time Fourier transform operator using window length L, frame shift H, window function w; t is the continuous time or discrete time index. The stage output is strictly aligned with the original data in channel index and time axis, which can be directly input into the phase-transformed generalized cross-correlation to form the time-difference-of-arrival matrix. Engineering example: reconstructing the per-source complex spectrum estimation of two to three potential partial discharge sources in parallel for eight-channel data, and recovering the time domain in overlap-add manner, dynamically adjusting a and T to stabilize the waveform quality when the reconstruction residual exceeds the threshold.

[0120] S150, using a phase-transformed generalized cross-correlation method to calculate the per-source time-domain waveform sequence, obtaining a time-difference-of-arrival matrix, and inputting the time-difference-of-arrival matrix and the sensor spatial coordinates of the detection sensors corresponding to the target detection entity into a deep regression positioning sub-network to output the predicted three-dimensional coordinates of the discharge source and the regression uncertainty.

[0121] Specifically, the phase-transformed generalized cross-correlation method refers to amplitude normalization (phase transformation) in the frequency domain on the cross-channel mutual power spectrum, and then inverse transformation to obtain the time-difference-of-arrival estimation method of the time-domain cross-correlation function. Its core is to suppress the amplitude dominant term and highlight the phase consistency, so as to maintain robust peak positioning under multi-path, amplitude fluctuation and noise conditions. Engineering example: forming a channel pair with a reference channel and a target channel, performing short-time Fourier transform on the per-source time-domain waveform sequence corresponding to each potential partial discharge source, calculating the phase-transformed mutual power spectrum and inverse transforming to obtain the cross-correlation function, and reading the main peak position as the time-difference-of-arrival estimation of a single channel pair. The time-difference-of-arrival matrix refers to a two-dimensional array formed by stacking the time-difference-of-arrival between the reference channel and other channels according to the channel index for each potential partial discharge source, which is used for geometric solving and input of the deep regression positioning sub-network. The common organization method is "source dimension x channel pair dimension", among which the channel pair dimension can be selected as "reference channel pair other channel" for compact representation to reduce redundancy.

[0122] The detection sensor refers to a set of sensors installed on the periphery or shell of the target detection entity for collecting partial discharge signals, including ultra-high frequency sensors and ultrasonic sensors, and having a unique channel identifier and calibrated sensor spatial coordinates in the acquisition link. Engineering example: four ultra-high frequency sensors and four ultrasonic sensors, all connected to the server through unified time synchronization. The sensor spatial coordinates refer to the three-dimensional coordinates in the device coordinate system of the target detection entity after installation pose calibration and coordinate transformation, which are used to establish a mapping relationship between the time-difference-of-arrival and the geometric distance difference, and serve as the geometric prior of the deep regression positioning sub-network. Engineering example: each detection sensor records a three-dimensional coordinate vector, which is one-to-one corresponding to the channel index for network input coding and visualization projection multiplexing.

[0123] The deep regression positioning sub-network refers to a neural network subsystem that directly regresses the predicted three-dimensional coordinates of the discharge source and the regression uncertainty under geometric consistency constraints, with the time difference of arrival matrix and the spatial coordinates of the sensors as inputs. It contains time difference of arrival encoding, geometric position encoding, and multi-layer perceptron or attention modules. During training, the time delay consistency and geometric residual are used as the main loss, and the uncertainty regularization is used as the auxiliary loss, to obtain robust coordinates and credibility evaluation. The engineering example is to concatenate each column of the time difference of arrival and the corresponding sensor spatial coordinates of the channel into a fixed-length feature at the input end, and after several layers of non-linear mapping, output the three-dimensional coordinates and the diagonal variance or complete covariance parameters. The predicted three-dimensional coordinates of the discharge source refer to the spatial position estimation of the discharge source output by the deep regression positioning sub-network in the device coordinate system, usually represented as a three-dimensional vector, which can be directly used for visualization and maintenance positioning. The engineering example is to output two three-dimensional coordinate solutions simultaneously under the condition of multiple concurrent defects, which fall near the bus support insulator and the disconnecting switch contact, respectively, and form a stable trajectory with the historical samples. The regression uncertainty refers to the confidence quantization result given by the deep regression positioning sub-network for the predicted three-dimensional coordinates of the discharge source, which is usually parameterized as a diagonal variance or a covariance matrix, used to construct an ellipsoidal confidence boundary and downstream alarm grading. It reflects both the inherent noise level of the time difference of arrival matrix and the adaptation degree of the network to the current sample distribution. The engineering example is that in periods with high noise and significant multipath, the regression uncertainty increases, the ellipsoidal confidence boundary becomes larger, and the alarm strategy correspondingly increases the trigger threshold.

[0124] In one possible implementation, the phase transform generalized cross-correlation method is used to calculate the time-domain waveform sequence of each source to obtain the time difference of arrival matrix, and the time difference of arrival matrix and the sensor spatial coordinates of the detection sensors corresponding to the target detection entity are input into the deep regression positioning sub-network to output the predicted three-dimensional coordinates of the discharge source and the regression uncertainty, specifically including: performing pre-emphasis, band-pass filtering and amplitude normalization processing on the time-domain waveform sequence of each source to generate a time difference of arrival candidate signal sequence; calculating the cross-channel mutual power spectrum according to the time difference of arrival candidate signal sequence and applying a phase transform weight to obtain a generalized cross-correlation function, obtaining a sub-sampling precision time difference of arrival estimate value by detecting the peak position and combining a parabolic interpolation method, and extracting the peak height and peak width as the time difference of arrival quality degree; weighting and fusing the time difference of arrival estimate value and the time difference of arrival quality degree with the source confidence tensor to construct a consistent time difference of arrival matrix; encoding the consistent time difference of arrival matrix and the sensor spatial coordinates of the detection sensors into geometric embedding, and forming the input of the deep regression positioning sub-network together with the time difference of arrival value; in the deep regression positioning sub-network, end-to-end regression is performed based on the time delay consistency and geometric residual constraints to output the predicted three-dimensional coordinates of the discharge source and the corresponding regression uncertainty.

[0125] Specifically, in the processing stage of the per-source time-domain waveform sequence, firstly, a pre-emphasis processing is performed, the per-source time-domain waveform sequence is input into a pre-emphasis filter to increase the amplitude of the high-frequency transient component, so that the high-frequency pulse related to the partial discharge is more prominent, while the interference of the low-frequency background noise is suppressed. Then, a band-pass filtering processing is performed, according to the effective frequency range of the target detection entity in the ultrahigh frequency signal and ultrasonic signal domain, the upper and lower limits of the corresponding passband frequency are selected, and the frequency components exceeding the passband range in the per-source time-domain waveform sequence are filtered out, so as to retain the frequency components useful for positioning. Subsequently, an amplitude normalization processing is performed, the amplitude of the per-source time-domain waveform sequence of each channel is scaled to a dimensionless interval, so as to eliminate the gain difference of the cross-channel sensors and maintain the relative energy proportion between signals.

[0126] In the time difference of arrival estimation stage, firstly, a cross-channel mutual power spectrum is calculated, and a phase transformation weight is applied to obtain a generalized cross-correlation function, so as to improve the resolution of peak positioning. Then, the maximum peak position of the generalized cross-correlation function is detected, and a parabolic interpolation method is combined to obtain a sub-sampling precision time difference of arrival estimation value. At the same time, the peak height and peak width are calculated as the time difference of arrival quality degree index, which is used to measure the reliability of the time difference of arrival estimation.

[0127] In order to improve the robustness of the time difference of arrival matrix in the complex scene of superimposed multi-source signals, the following weighted peak confidence fusion formula is used for weighted calculation of the time difference of arrival of multiple potential partial discharge sources:

[0128]

[0129] Wherein, τ w (i, j) represents the weighted time difference of arrival of the detection sensor channel i and channel j, K represents the number of potential partial discharge sources, represents the time difference of arrival estimation value of the kth potential partial discharge source, represents the time difference of arrival quality degree corresponding to the kth potential partial discharge source, represents the source confidence corresponding to the kth potential partial discharge source.

[0130] In the consistent time difference of arrival matrix construction stage, the weighted time difference of arrival is filled into the matrix according to the index order and spatial layout relationship of the detection sensor channel, so as to keep the matrix structure consistent with the geometric topology of the sensor array. Subsequently, in the geometric embedding coding stage, the spatial coordinates of the detection sensor are coded into a geometric embedding vector through a geometric mapping function, and are concatenated with the consistent time difference of arrival matrix in the feature dimension, to form a joint feature representation containing time delay information and spatial geometric information.

[0131] Finally, the joint feature representation is input into a deep regression positioning subnetwork with time-delay consistency constraint and geometric residual constraint, and the discharge source predicted 3D coordinates and regression uncertainty are output in an end-to-end regression inference process, providing input data support for subsequent spatial confidence interval construction and visual positioning.

[0132] S160, by the discharge source predicted 3D coordinates, regression uncertainty and source confidence tensor, constructing the confidence interval and outputting the visual positioning result with ellipsoid confidence boundary.

[0133] Specifically, the confidence interval represents the probability boundary description of the discharge source true position falling into a certain region under a given confidence level, which is usually obtained by fusing the covariance matrix according to the regression uncertainty and combining the weight modulation of the source confidence tensor, and then determining the size of the spatial region for reliability interpretation and alarm grading in engineering. For example, by selecting a confidence level of ninety-five percent, a spatial region with a coverage probability of ninety-five percent can be obtained. The ellipsoid confidence boundary represents the geometricization of the confidence interval in three-dimensional space in the shape of an ellipsoid, the principal axis direction of which is determined by the covariance principal direction of the regression uncertainty, and the principal axis length is positively correlated with the confidence level, which is used for intuitive presentation and quantitative quantification on the three-dimensional model of the device. For example, the ellipsoid confidence boundary corresponding to the discharge source predicted 3D coordinates near the flange is more "flat" in the flange normal direction and more "long" in the direction along the pipeline, reflecting the anisotropic measurement uncertainty.

[0134] The visual positioning result represents the graphical output generated by the server based on the discharge source predicted 3D coordinates, regression uncertainty and source confidence tensor, including 3D coordinate labels, ellipsoid confidence boundaries, color and transparency mapping, time label and channel index mapping, etc., which can be displayed on a three-dimensional device model or a two-dimensional engineering drawing, and supports historical playback and alarm linkage. For example, the system represents the source confidence tensor in color and the regression uncertainty in transparency, and highlights the ellipsoid confidence boundary and generates a maintenance task when the alarm threshold is triggered.

[0135] In a possible implementation, the three-dimensional coordinates predicted by the discharge source, the regression uncertainty, and the source confidence tensor are fused to construct a confidence interval and output a visual positioning result with an ellipsoid confidence boundary, specifically including: fusing the regression uncertainty and the source confidence tensor to generate a fused covariance matrix; performing eigenvalue decomposition on the fused covariance matrix to obtain an ellipsoid principal axis direction and a principal axis radius, and combining the ellipsoid principal axis direction and the principal axis radius with the three-dimensional coordinates predicted by the discharge source to form a three-dimensional ellipsoid parameter set; mapping the three-dimensional ellipsoid parameter set to a two-dimensional ellipse visual element according to a view matrix and a projection matrix of a terminal visual interface, and calculating color parameters and transparency parameters in combination with the source confidence tensor and the regression uncertainty; judging the spatial relationship inside and outside the ellipsoid confidence boundary by Mahalanobis distance, and generating a smooth trajectory and an ellipsoid confidence boundary changing over time by using the three-dimensional coordinates predicted by the discharge source and the fused covariance matrix in a continuous time window to obtain the visual positioning result.

[0136] Specifically, in the stage of fusing the regression uncertainty and the source confidence tensor and generating the fused covariance matrix, first, the source confidence tensor of each potential partial discharge source in the peak neighborhood is performed weighted aggregation to obtain a corresponding confidence weight, and then the confidence weight is used for scale modulation of the regression uncertainty to form the fused covariance matrix, and the specific calculation is as follows:

[0137]

[0138] Wherein, η k represents the scalar confidence weight of the kth discharge source, C k represents the regression uncertainty (covariance matrix) of the kth discharge source, and Ω k (f, τ) represents the source confidence tensor element of the kth discharge source at (f, τ). k k k k

[0139] In the stage of performing eigenvalue decomposition on the fused covariance matrix and combining the ellipsoid principal axis direction and the principal axis radius with the three-dimensional coordinates predicted by the discharge source to form a three-dimensional ellipsoid parameter set, the scale factor corresponding to the significance level is calculated first, and then the characteristic decomposition is performed, and the specific calculation is as follows:

[0140]

[0141] Wherein, V represents the principal axis direction matrix composed of eigenvectors, Λ represents the eigenvalue diagonal matrix, c represents the confidence level 1-α corresponding to the chi-square distribution quantile, and a represents the ellipsoid principal axis radius vector. k k α k k ​​​​​​​​​​This represents the parameter set of a three-dimensional ellipsoid.

[0142] When mapping the 3D ellipsoid parameter set to 2D elliptical visualization elements based on the viewpoint matrix and projection matrix of the terminal visualization interface, the image domain covariance is obtained through projection linearization, and 2D elliptical parameters are generated. Simultaneously, color and transparency parameters are calculated based on uncertainty and confidence levels. Specifically, the calculations are as follows:

[0143]

[0144] Among them, u k Let π(·) represent the 2D pixel center, π(·) represent the perspective projection function, K represent the camera intrinsic parameter matrix, R and t represent the rotation matrix and translation vector, respectively, and J(r k ) indicates that in r k The projection Jacobian at the location, U k S represents the two-dimensional principal direction matrix. k ρ represents a two-dimensional eigenvalue diagonal matrix. k This represents the length of the semi-axis of a two-dimensional ellipse.

[0145] The calculation of color and transparency parameters is as follows:

[0146]

[0147] Among them, κ k Indicates the color parameter, α k The transparency parameters are represented by ω1, ω2, and ξ, which are hyperparameters.

[0148] When using Mahalanobis distance to determine whether a spatial point lies inside the confidence boundary of an ellipsoid, the calculation formula is as follows:

[0149]

[0150] in, Let x represent the squared Mahalanobis distance, and let x represent the spatial detection point.

[0151] When generating smooth trajectories and time-varying ellipsoidal confidence boundaries within a continuous time window, exponential smoothing is applied to the three-dimensional coordinates and covariance, specifically as follows:

[0152]

[0153] in, Represents the smoothed three-dimensional coordinates. Let represent the smoothed fusion covariance matrix, λ represent the smoothing factor, and t represent the time index.

[0154] After the above steps, a smooth and visually defined positioning result with dynamic confidence boundaries is finally obtained, which can be presented as a dynamic graphic with ellipsoidal uncertainty boundaries on the terminal.

[0155] In summary, the embodiment of the present application aims at the problem that partial discharge sources are difficult to be accurately located in complex environments. By introducing multi-modal fusion of acoustic-electric signals, complex spectrum tensor and power spectrum tensor construction of short-time Fourier transform, high-dimensional feature embedding of double-branch deep learning network, and time-frequency masking tensor and source confidence tensor output of multi-source separation network, the potential partial discharge sources are realized to be separated and reconstructed source by source. Compared with the traditional detection method relying on a single signal domain or based on threshold determination, the present application can more effectively separate multi-source interference and restore the real waveform of the partial discharge source, thereby improving the signal quality and source feature distinguishability.

[0156] In the positioning phase, the phase transformation generalized cross-correlation method is used to accurately estimate the time difference of arrival between multiple sensors, which is input into the deep regression positioning sub-network together with the spatial coordinates of the sensors. Combining the time delay consistency and geometric residual constraints, an end-to-end regression is realized, which significantly improves the accuracy and robustness of the discharge source spatial positioning. Compared with the positioning method relying only on geometric inversion, the present application can resist the deviation caused by strong noise and multipath interference to the time difference of arrival estimation, ensuring the credibility of the positioning result.

[0157] In terms of visualized result output, the fusion covariance matrix is constructed based on the predicted three-dimensional coordinates of the discharge source, the regression uncertainty and the source confidence tensor, and the principal axis parameters of the ellipsoid are obtained through eigenvalue decomposition, and then the visualized ellipsoid with confidence boundary is formed. The inside and outside of the boundary are determined by Mahalanobis distance, and the confidence level is expressed by color and transparency parameters, which can intuitively present the uncertainty distribution and dynamic evolution trajectory of the discharge source position, solving the problem that the traditional method can only give single point positioning result but lacks uncertainty representation.

[0158] The present application also provides a GIS partial discharge positioning device based on deep learning and acoustic-electric combination, which is described with reference to Figure 2 , Figure 2A module schematic diagram of a GIS partial discharge positioning device based on deep learning and sound-electricity combination provided by the embodiment of the application is provided. The device is a server. The server includes an acquisition module 21 and a processing module 22. The acquisition module 21 is configured to acquire original ultra-high frequency signals and original ultrasonic signals for a target detection entity, and generate a sound-electricity signal matrix. The processing module 22 is configured to perform short-time Fourier transform on the sound-electricity signal matrix to obtain a complex spectrum tensor and a power spectrum tensor. The processing module 22 is further configured to take the complex spectrum tensor and the power spectrum tensor as input, input an ultra-high frequency signal branch and an ultrasonic signal branch into a deep learning network for double-branch feature extraction, and output a high-dimensional feature embedding tensor. The processing module 22 is further configured to generate a time-frequency mask tensor and a source confidence tensor of each potential partial discharge source through a multi-source separation sub-network according to the high-dimensional feature embedding tensor, and perform source-by-source spectrum reconstruction based on the time-frequency mask tensor and the source confidence tensor to obtain a source-by-source time-domain waveform sequence. The processing module 22 is further configured to calculate the source-by-source time-domain waveform sequence by using a phase transform generalized cross-correlation method to obtain a time difference of arrival matrix, and input the time difference of arrival matrix and sensor space coordinates of a detection sensor corresponding to the target detection entity into a deep regression positioning sub-network to output a discharge source predicted three-dimensional coordinate and a regression uncertainty. The processing module 22 is further configured to construct a confidence interval and output a visual positioning result with an ellipsoid confidence boundary through the discharge source predicted three-dimensional coordinate, the regression uncertainty and the source confidence tensor.

[0159] It should be noted that the device provided in the above embodiment is only used as an example to illustrate the division of the above functional modules. In actual application, the above functions can be completed by different functional modules according to needs, that is, the internal structure of the device is divided into different functional modules to complete all or part of the functions described above. In addition, the device and method embodiments provided in the above embodiments belong to the same concept, and the specific implementation process is detailed in the method embodiments, which will not be described here.

[0160] The application further provides an electronic device, referring to Figure 3 , Figure 3 A structural schematic diagram of an electronic device provided by the embodiment of the application is provided. The electronic device can include at least one processor 31, at least one network interface 34, a user interface 33, a memory 35, and at least one communication bus 32.

[0161] The communication bus 32 is used to realize the connection and communication between the components.

[0162] The user interface 33 can include a display screen (Display) and a camera (Camera). Optionally, the user interface 33 can further include a standard wired interface and a wireless interface.

[0163] The network interface 34 can optionally include a standard wired interface, a wireless interface (such as a Wi-Fi interface).

[0164] The processor 31 can include one or more processing cores. The processor 31 connects various parts within the server through various interfaces and lines, performs various functions of the server and processes data by running or executing instructions, programs, code sets or instruction sets stored in the memory 35, and calling data stored in the memory 35. Optionally, the processor 31 can be implemented in at least one of a hardware form of a digital signal processing (DSP), a field-programmable gate array (FPGA), and a programmable logic array (PLA). The processor 31 can be integrated with a combination of one or more of a central processing unit (CPU), a graphics processing unit (GPU), and a modem. Among them, the CPU mainly processes the operating system, user interface, and application programs; the GPU is responsible for rendering and drawing the content to be displayed on the display screen; and the modem is used for processing wireless communication. It can be understood that the above-mentioned modem can also not be integrated into the processor 31, but can be implemented by a separate chip.

[0165] The memory 35 can include a random access memory (RAM) and a read-only memory (ROM). Optionally, the memory 35 includes a non-transitory computer-readable storage medium. The memory 35 can be used to store instructions, programs, codes, code sets or instruction sets. The memory 35 can include a program storage area and a data storage area, wherein the program storage area can store instructions for implementing an operating system, instructions for at least one function (such as a touch function, a sound playing function, an image playing function, etc.), instructions for implementing the above-mentioned various method embodiments, etc.; the data storage area can store data involved in the above-mentioned various method embodiments, etc. The memory 35 can optionally be at least one storage device located away from the aforementioned processor 31. As shown, the memory 35 as a computer storage medium can include an operating system, a network communication module, a user interface module, and an application program of a GIS partial discharge positioning method based on deep learning and sound-electricity combination. Figure 3

[0166] In​Figure 3 In the electronic device shown, the user interface 33 is mainly used to provide an interface for the user to input, and obtain data input by the user; and the processor 31 can be used to invoke an application program of a GIS partial discharge positioning method based on deep learning and acoustic-electric combination stored in the memory 35, which, when executed by one or more processors, causes the electronic device to perform the method of one or more of the above-described embodiments.

[0167] It should be noted that, for the foregoing method embodiments, in order to simply describe, they are all described as a series of action combinations, but those skilled in the art should know that the present application is not limited to the order of the actions described, because according to the present application, certain steps can be performed in other order or at the same time. Secondly, those skilled in the art should know that the embodiments described in the specification all belong to preferred embodiments, and the actions and modules involved are not necessarily required by the present application.

[0168] The present application also provides a computer-readable storage medium, which stores instructions. When executed by one or more processors, the electronic device performs the method of one or more of the above-described embodiments.

[0169] In the above embodiments, the description of each embodiment has its own focus, and the parts not described in detail in a certain embodiment can be referred to the relevant description of other embodiments.

[0170] In several embodiments provided by the present application, it should be understood that the disclosed device can be implemented in other ways. For example, the device embodiments described above are only schematic. The division of the units is only a logical function division. There can be another division manner in actual implementation, for example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the displayed or discussed mutual couplings or direct couplings or communication connections between the units can be indirect couplings or communication connections through some interfaces, devices or units, and can be electrical or other forms.

[0171] The units described as separate components can or can not be physically separate, and the components displayed as units can or can not be physical units, that is, they can be located in one place, or can be distributed on a plurality of network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the embodiment scheme.

[0172] In addition, each of the functional units in the various embodiments of the present application can be integrated in one processing unit, or each unit can be physically present separately, or two or more units can be integrated in one unit. The integrated unit can be realized in the form of hardware or in the form of a software functional unit.

[0173] If the integrated unit is realized in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer readable memory. Based on such understanding, the technical solutions of the present application, essentially or in part, or all or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a memory and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods in the various embodiments of the present application. The memory where the software is stored includes: U disk, mobile hard disk, magnetic disk or optical disk, and various other media that can store program codes.

[0174] The above is only exemplary embodiments of the present disclosure, and cannot limit the scope of the present disclosure. That is, any equivalent changes and modifications made in accordance with the teachings of the present disclosure are still within the scope of the present disclosure. Other embodiments of the present disclosure will be readily apparent to those skilled in the art upon considering the specification and practicing the true principles of the present disclosure. The present application is intended to cover any variations, uses, or adaptive changes of the present disclosure that follow the general principles of the present disclosure and include common knowledge or conventional technical means in the technical field not disclosed by the present disclosure. The specification and examples are only considered as exemplary, and the scope and spirit of the present disclosure are defined by the claims.

Claims

1. A GIS partial discharge localization method based on deep learning and acoustic-electrical integration, characterized in that, The method includes: The original ultra-high frequency signal and the original ultrasonic signal of the target entity are acquired, and an acoustic-electric signal matrix is ​​generated. Perform a short-time Fourier transform on the acoustic-electric signal matrix to obtain the complex spectral tensor and the power spectral tensor; Using the complex spectrum tensor and the power spectrum tensor as inputs, the ultra-high frequency signal branch and the ultrasonic signal branch are respectively input into the deep learning network for dual-branch feature extraction, and a high-dimensional feature embedding tensor is output. Based on the high-dimensional feature embedding tensor, a time-frequency masking tensor and a source confidence tensor for each potential local discharge source are generated through a multi-source separation sub-network. Then, source-by-source spectral reconstruction is performed based on the time-frequency masking tensor and the source confidence tensor to obtain a source-by-source time-domain waveform sequence. The phase transformation generalized cross-correlation method is used to calculate the source-by-source time-domain waveform sequence to obtain the arrival time difference matrix. The arrival time difference matrix and the sensor spatial coordinates of the detection sensor corresponding to the target detection entity are input into the depth regression localization subnetwork, and the output discharge power source prediction three-dimensional coordinates and regression uncertainty are output. The confidence interval is constructed and a visualized positioning result with ellipsoidal confidence boundaries is output by using the predicted three-dimensional coordinates of the discharge source, the regression uncertainty, and the source confidence tensor.

2. The GIS partial discharge localization method based on deep learning and acoustic-electrical integration according to claim 1, characterized in that, The process of acquiring the original ultra-high frequency signal and the original ultrasonic signal for the target detection entity, and generating an acoustic-electric signal matrix, specifically includes: Acquire the original ultra-high frequency signals and original ultrasonic signals transmitted by the ultra-high frequency sensors and ultrasonic sensors pre-deployed on the target detection entity; Add high-precision timestamps, sensor spatial coordinates, and channel identifiers to the sampling points corresponding to the original ultra-high frequency signal and the original ultrasonic signal; Align the original ultra-high frequency signal with the original ultrasonic signal timestamp, construct a multi-channel data structure with sampling time as row index, sensor channel number as column index, and matrix unit value as corresponding amplitude, and determine the multi-channel data structure as the acoustic-electric signal matrix.

3. The GIS partial discharge localization method based on deep learning and acoustic-electrical integration according to claim 1, characterized in that, The step of performing a short-time Fourier transform on the acoustic-electric signal matrix to obtain the complex spectral tensor and the power spectral tensor specifically includes: Before performing the short-time Fourier transform, the channel set belonging to ultra-high frequency signals and the channel set belonging to ultrasonic signals in the acoustic-electric signal matrix are respectively bound to the corresponding sampling rate, sensor calibration coefficient and channel gain factor; Bandpass anti-aliasing filtering and resampling are performed on channels with different sampling rates to align them to a uniform sampling rate and keep the channel index, spatial coordinates and time axis consistent. For different types of signals, the window length, frame shift, zero-padding length and window function type are configured respectively, and the overlapping windowing strategy is used for framing, windowing and fast Fourier transform to obtain complex spectral tensors containing amplitude and phase information; The complex spectral tensor is subjected to window energy correction, channel gain correction, and sensor sensitivity correction to generate the power spectral tensor.

4. The GIS partial discharge localization method based on deep learning and acoustic-electrical integration according to claim 1, characterized in that, The process of using the complex spectral tensor and the power spectral tensor as inputs, inputting the ultra-high frequency signal branch and the ultrasonic signal branch into a deep learning network for dual-branch feature extraction, and outputting a high-dimensional feature embedding tensor, specifically includes: Within the ultra-high frequency signal branch and the ultrasonic signal branch, the complex spectrum tensor and the power spectrum tensor are normalized, geometrically prior encoded, and characterized by early convolutional features. In each branch, spatial coordinate embedding generated by the sensor spatial coordinates is introduced to preserve spatial location information. In the ultra-high frequency signal branch, phase correlation features are extracted by complex convolution, complex normalization and complex nonlinear units, and the phase correlation features are fused with the power spectrum tensor to form the first intermediate feature tensor of the ultra-high frequency signal branch; In the ultrasonic signal branch, narrowband ringing and echo tail features are extracted using a combination of temporal convolution and multi-head self-attention. The narrowband ringing and echo tail features are then fused with the power spectrum tensor to form the second intermediate feature tensor of the ultrasonic signal branch. In the multi-scale representation stage, time-frequency patterns with different spread widths are captured by dilated time-frequency convolution, and the multi-scale features are recalibrated using the channel attention mechanism to obtain recalibrated features. The recalibrated features, the first intermediate feature tensor, and the second intermediate feature tensor are input into the residual shaping network to generate the high-dimensional feature embedding tensor.

5. The GIS partial discharge localization method based on deep learning and acoustic-electrical integration according to claim 1, characterized in that, The process involves generating a time-frequency masking tensor and a source confidence tensor for each potential local discharge source using a multi-source splitter network based on the high-dimensional feature embedding tensor, and performing source-by-source spectral reconstruction based on the time-frequency masking tensor and the source confidence tensor to obtain a source-by-source time-domain waveform sequence. Specifically, this includes: The high-dimensional feature embedding tensor is encoded using the shared backbone and source-specific output structure of the multi-source separation sub-network to generate a time-frequency masking tensor containing real part masking, imaginary part masking, and magnitude masking. By utilizing the uncertainty output structure of the multi-source separation sub-network, the masking estimation is combined with the intermediate features of the network to generate the source confidence tensor; In the source-by-source spectrum reconstruction stage, the complex spectrum tensor and the power spectrum tensor are combined with the time-frequency masking tensor and the source confidence tensor to perform consistent weighted reconstruction in the complex domain to obtain the complex spectrum estimate of each potential local discharge source, and then restored to the source-by-source time-domain waveform sequence through inverse short-time Fourier transform.

6. The GIS partial discharge localization method based on deep learning and acoustic-electrical integration according to claim 1, characterized in that, The step involves using a phase-transform generalized cross-correlation method to calculate the source-by-source time-domain waveform sequence, obtaining the arrival time difference matrix. This arrival time difference matrix, along with the sensor spatial coordinates of the detection sensor corresponding to the target detection entity, is then input into the depth regression localization subnetwork. The output includes the predicted three-dimensional coordinates of the discharge source and the regression uncertainty. Specifically, this includes: The source-by-source time-domain waveform sequence is subjected to pre-emphasis, bandpass filtering and amplitude normalization to generate a candidate signal sequence of arrival time difference; Based on the candidate signal sequence of arrival time difference, the cross-channel cross power spectrum is calculated and a phase transformation weight is applied to obtain the generalized cross-correlation function. By detecting the peak position and combining it with the parabolic interpolation method, the arrival time difference estimate with subsampling accuracy is obtained. At the same time, the peak height and peak width are extracted as arrival time difference quality scores. The estimated arrival time difference and the quality score of arrival time difference are weighted and fused with the source confidence tensor to construct a consistent arrival time difference matrix; The consistent arrival time difference matrix and the sensor spatial coordinates of the detection sensor are encoded into a geometric embedding, which together with the arrival time difference value forms the input of the depth regression localization sub-network; In the deep regression localization subnetwork, end-to-end regression is performed based on time delay consistency and geometric residual constraints, and the predicted three-dimensional coordinates of the discharge source and the corresponding regression uncertainty are output.

7. The GIS partial discharge localization method based on deep learning and acoustic-electrical integration according to claim 1, characterized in that, The process of constructing a confidence interval and outputting a visualized positioning result with an ellipsoidal confidence boundary by using the predicted three-dimensional coordinates from the discharge source, the regression uncertainty, and the source confidence tensor specifically includes: The regression uncertainty is fused with the source confidence tensor to generate a fused covariance matrix; Eigenvalue decomposition is performed on the fused covariance matrix to obtain the principal axis direction and principal axis radius of the ellipsoid, which are then combined with the predicted three-dimensional coordinates of the discharge source to form a three-dimensional ellipsoid parameter set. Based on the viewpoint matrix and projection matrix of the terminal visualization interface, the three-dimensional ellipsoid parameter set is mapped into two-dimensional elliptical visualization elements, and the color parameters and transparency parameters are calculated by combining the source confidence tensor and the regression uncertainty. The spatial relationship between the inside and outside of the ellipsoidal confidence boundary is determined by Mahalanobis distance, and the three-dimensional coordinates are predicted by the discharge power source and the fused covariance matrix within a continuous time window to generate a smooth trajectory and the time-varying ellipsoidal confidence boundary, so as to obtain the visualized positioning result.

8. A GIS partial discharge localization device based on deep learning and acoustic-electrical integration, characterized in that, The device includes an acquisition module (21) and a processing module (22), wherein, The acquisition module (21) is used to acquire the original ultra-high frequency signal and the original ultrasonic signal for the target detection entity and generate an acoustic-electric signal matrix; The processing module (22) is used to perform a short-time Fourier transform on the acoustic signal matrix to obtain a complex spectral tensor and a power spectral tensor. The processing module (22) is also used to take the complex spectrum tensor and the power spectrum tensor as inputs, input the ultra-high frequency signal branch and the ultrasonic signal branch into the deep learning network respectively for dual-branch feature extraction, and output a high-dimensional feature embedding tensor; The processing module (22) is further configured to generate a time-frequency masking tensor and a source confidence tensor for each potential local discharge source through a multi-source separation sub-network based on the high-dimensional feature embedding tensor, and perform source-by-source spectrum reconstruction based on the time-frequency masking tensor and the source confidence tensor to obtain a source-by-source time-domain waveform sequence. The processing module (22) is also used to calculate the source-by-source time-domain waveform sequence using the phase transformation generalized cross-correlation method to obtain the arrival time difference matrix, and input the arrival time difference matrix and the sensor spatial coordinates of the detection sensor corresponding to the target detection entity into the depth regression positioning subnetwork, and output the predicted three-dimensional coordinates and regression uncertainty of the discharge power source. The processing module (22) is also used to predict the three-dimensional coordinates, the regression uncertainty and the source confidence tensor through the discharge power source, construct the confidence interval and output the visualized positioning result with ellipsoidal confidence boundary.

9. An electronic device, characterized in that, The electronic device includes a processor (31), a memory (35), a user interface (33), and a network interface (34). The memory (35) is used to store instructions. The user interface (33) and the network interface (34) are both used to communicate with other devices. The processor (31) is used to execute the instructions stored in the memory (35) to cause the electronic device to perform the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions that, when executed, perform the method as described in any one of claims 1 to 7.

Images