Broadband feature coupling and sparse low-rank combined imaging method, product and equipment

Abstract

The invention discloses a broadband feature coupling and sparse low-rank joint imaging method, product and equipment, and belongs to the technical field of electrical equipment monitoring. The method comprises the following steps: S1, establishing a pure acoustic propagation model of a partial discharge sound source, and fusing partial discharge pulse waveforms and sound wave propagation characteristics; s2, constructing a physically constrained over-complete sparse dictionary based on the time domain damped oscillation of the partial discharge signal and the physical fingerprints with concentrated frequency domain and high energy; s3, constructing a joint optimization objective function by combining the characteristics of sparse time domain and low space rank of the partial discharge signal, and iteratively solving a space-time signal matrix; s4, designing a lightweight deep network fused with physical prior, and considering both precision and real-time performance; and S5, realizing data acquisition and imaging real-time processing based on a GPU stream processing architecture in combination with dual-threshold detection and pre-calculation optimization. The broadband transient characteristic and the space-time sparsity of the partial discharge pulse can be embedded into the imaging model, and a traditional general acoustic imaging frame is replaced.

Description

Technical Field

[0001] This invention belongs to the field of electrical equipment monitoring technology, specifically relating to a broadband feature coupling and sparse low-rank joint imaging method, product and equipment. Background Technology

[0002] While phased array-based positioning and imaging technologies are rapidly developing in the radar field, their application in industrial partial discharge monitoring has significant limitations. Traditional ultrasonic imaging methods such as plane wave coherent synthesis and compressed sensing are typically based on uniform sampling strategies and universal acoustic inversion models. These methods are insufficiently adaptable to the microsecond-level transient characteristics of 20-200kHz broadband partial discharge pulses, resulting in a high rate of invalid data acquisition in actual monitoring. Furthermore, existing technologies neglect the unique physical fingerprint characteristics inherent in partial discharge ultrasonic signals. Their non-stationary time-frequency distribution typically exhibits a Morlet wavelet decay oscillation mode, which is naturally sparse in the spatiotemporal domain and has significant single-event point source characteristics. However, current mainstream positioning schemes such as MUSIC algorithms and MV beamforming still generally follow the universal array signal processing framework. These methods mainly rely on post-processing filtering operations to suppress background noise, failing to deeply integrate the physical fingerprint characteristics of partial discharge signals into the core processing mechanism of their imaging algorithms. This makes it difficult for traditional methods to fully utilize the sparse representation of signals and the coherence information between multi-channel signals in complex industrial monitoring environments with strong interference. Consequently, the methods suffer from problems such as large computational load for positioning, poor real-time performance, and susceptibility to false positioning and resolution loss.

[0003] While recent studies have attempted to accelerate ultrasound imaging using deep learning, end-to-end reconstruction schemes using CNNs in the medical field suffer from poor physical interpretability due to their black-box nature and difficulty adapting to the variable conditions in industrial settings. Existing compressed sensing frameworks can achieve rapid localization, but they do not fully incorporate the time-frequency characteristics of partial discharge pulses, thus sacrificing imaging localization accuracy while maintaining a high compression ratio. In summary, current technologies have not yet fundamentally established a closed-loop theoretical framework covering physical modeling, sparse representation, and rapid imaging. Summary of the Invention

[0004] The present invention aims to at least partially solve one of the technical problems in the aforementioned related technologies.

[0005] Therefore, the purpose of this invention is to provide a broadband feature coupling and sparse low-rank joint imaging method, product and device, which can embed the broadband transient characteristics and spatiotemporal sparsity of partial discharge pulses into the imaging model, construct a feature-driven joint optimization objective function, and replace the traditional general acoustic imaging framework.

[0006] To solve the above-mentioned technical problems, the present invention is implemented as follows: This invention provides a method for joint imaging of broadband feature coupling and sparse low-rank imaging, the method comprising: S1. Broadband pulse characteristic coupling modeling: Establish a pure acoustic propagation model of the partial discharge source, integrate the partial discharge pulse waveform and sound wave propagation characteristics, and overcome the defect of the point source model in ignoring directionality and high-frequency information. S2. Construction of Time-Frequency Sparse Dictionary: Based on the physical fingerprint of the time-domain decay oscillation and high-energy concentration of the partial discharge signal, a physically constrained overcomplete sparse dictionary is constructed to replace the traditional general basis function, thereby improving signal discrimination and anti-interference capability. S3. Low-rank-sparse joint optimization reconstruction: Combining the characteristics of temporal sparsity and spatial low-rank of partial discharge signals, a joint optimization objective function is constructed, and the spatiotemporal signal matrix is ​​solved through ADMM iteration to improve noise resistance and positioning accuracy; S4. Physics-guided deep learning acceleration: Design a lightweight deep network that integrates physical priors to achieve millisecond-level signal reconstruction, balancing accuracy and real-time performance; S5. Real-time Imaging Engine Design: Based on a GPU stream processing architecture, combined with dual-threshold detection and pre-computation optimization, it realizes real-time processing of the entire data acquisition-imaging visualization chain, meeting the needs of industrial edge deployment.

[0007] In addition, the broadband feature coupling and sparse low-rank joint imaging method according to the present invention may also have the following additional technical features: In some implementations, step S1 includes: Sound pressure field convolution modeling: The partial discharge sound pressure field is represented as the convolution of the partial discharge pulse signal and the sound propagation Green's function, which accurately describes the propagation process of the sound signal from the source to the sensor; Time-frequency domain decoupling: Short-time Fourier transform is used to convert the time-domain signal to the time-frequency domain, decoupling the spectral components and propagation terms to obtain the frequency domain expression, capturing the transient characteristics of broadband pulses in a specific frequency band; Frequency-varying directivity function embedding: Introducing the frequency-varying directivity function of the simulated partial discharge source dipole antenna radiation to construct an optimized objective function, thereby enhancing the model's characterization of signal directivity.

[0008] In some implementations, step S2 includes: Time-domain atomic design: Morlet wavelet basis is used to simulate the decaying oscillation of partial discharge pulse, taking into account both time-domain locality and frequency-domain selectivity; Spatial atomic design: Employing a time delay function to simulate sound wave propagation delay, accurately matching spatial location information; Tensor dictionary construction: By using the tensor product of the temporal Morlet wavelet dictionary and the spatial time delay dictionary, a three-dimensional overcomplete sparse dictionary is generated, which represents the multi-channel observation signal as a formula related to the observation signal, sparse dictionary, sparse coefficients and noise. Sparse coding solution: Orthogonal matching pursuit or minimum absolute shrinkage and selection operator algorithm is used to select the optimal atom combination from the redundant atom library and extract the real partial discharge signal that excludes noise and reflected waves.

[0009] In some implementations, step S3 includes: Spatiotemporal signal matrix construction: Assuming multiple sensors and multiple time sampling points for partial discharge, the multi-channel signals are organized into a spatiotemporal signal matrix; Joint optimization objective function: By reconstructing the joint optimization objective function of error, row sparsity constraint, low-rank constraint of nuclear norm, and coupling term correlation, a coordinated expression of physical prior is achieved; ADMM is solved by block iterative method, including sparse term update, low-rank term update and multiplier update.

[0010] In some implementations, the sparsity term is updated by fixing the low-rank matrix and using a hybrid norm shrinkage algorithm to update the sparsity coefficients. Low-rank term update: With fixed sparse coefficients, the low-rank matrix is ​​updated using the singular value thresholding algorithm; Multiplier update: Update the Lagrange multipliers to maintain consistency between sparse and low-rank components and avoid optimization divergence.

[0011] In some implementations, step S4 includes: Input preprocessing: The multi-channel raw signal is fed into the STFT layer to generate the time spectrum, and combined with the frequency-varying directivity function in S1 to generate a directional mask, guiding the network to focus on the main energy region of the partial discharge. Two-stream network structure: The sparse branch uses 1D CNN to encode the STFT results, learns the transient sparse activation pattern of the signal, and outputs a sparse estimate; a differentiable singular value decomposition layer is introduced in the low-rank branch to deconstruct the signal or intermediate representation and output a low-rank estimate. Physical constraint loss function: Define a multi-objective loss function related to reconstruction error, sparse regularization, low-rank regularization, consistency constraint, and temporal gradient sparsity term to avoid poor generalization of black box models and ensure that the output conforms to physical laws.

[0012] In some implementations, step S5 includes: Dual-threshold event detection: Employs dual triggering with a time-domain amplitude threshold and a frequency-domain energy threshold to reduce invalid calculations and false alarms; Feature coupling calculation: After triggering, the STFT is executed to generate the spectrum, and the pre-calculated tensor dictionary pseudo-inverse matrix is ​​loaded to directly map the signal to the sparse coefficient space, reducing the online solution time; Joint solution and imaging localization: The sparse / low-rank components are solved in parallel using a deep network trained with S4. The location of the discharge source is determined by the maximum activation position of the sparse coefficients, and the results are corrected by the low-rank components. GPU parallel architecture: Deployed on CUDA-enabled GPUs, it achieves multi-module parallelism through asynchronous streaming.

[0013] In some implementations, the dual triggering mechanism of time-domain amplitude threshold + frequency-domain energy threshold is as follows: When both conditions are met simultaneously, such as the amplitude of any channel signal exceeding a set level and the energy of a specific frequency band exceeding a threshold, subsequent feature extraction and imaging are triggered.

[0014] This invention also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the broadband feature coupling and sparse low-rank joint imaging method as described in any of the preceding claims.

[0015] This invention also provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the broadband feature coupling and sparse low-rank joint imaging method as described in any of the preceding embodiments.

[0016] Compared with the prior art, the present invention has at least the following beneficial effects: In this embodiment of the invention, the broadband feature coupling and sparse low-rank joint imaging method establishes a low-rank-sparse joint optimization theory and proposes a novel objective function that integrates row sparsity constraints, nuclear norm low-rank constraints and consistency coupling terms. This can achieve a collaborative expression of spatiotemporal physical priors and significantly improve the theoretical completeness of the reconstruction. In this embodiment of the invention, the broadband feature coupling and sparse low-rank joint imaging method hard-codes differentiable physical operators in the network structure and designs a dedicated loss function, pioneering the fusion learning of physical models and data-driven approaches, and providing a fast imaging solution for edge computing scenarios.

[0017] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0018] Figure 1 This is a flowchart of a broadband feature coupling and sparse low-rank joint fast imaging method considering the time-frequency characteristics of partial discharge pulse ultrasound, as disclosed in one embodiment of the present invention. Figure 2 This is a schematic diagram illustrating the construction of ultrasonic time-frequency characteristic parameters according to an embodiment of the present invention; Figure 3 This is an atomic construction diagram of a time-frequency sparse dictionary disclosed in one embodiment of the present invention; Figure 4Here is a flowchart of a low-rank-sparse joint optimization algorithm disclosed in one embodiment of the present invention; Figure 5 This is a diagram of a physics-guided deep learning network architecture disclosed in one embodiment of the present invention; Figure 6 This is a block diagram of a real-time imaging engine design disclosed in one embodiment of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings and specific examples and application scenarios.

[0021] In some embodiments of the present invention, a broadband feature coupling and sparse low-rank joint fast imaging method considering the time-frequency characteristics of partial discharge pulse ultrasound is provided. First, a pure acoustic propagation physical model is constructed; second, a time-frequency sparse dictionary based on Morlet wavelets and acoustic delay is designed to achieve a physically constrained sparse representation of the pulse signal; furthermore, a method including... The low-rank-sparse joint optimization objective function of norm, kernel norm, and coupling terms is used to solve the spatiotemporal signal matrix through ADMM iteration. Then, a physics-guided deep learning network is further developed, embedding SVD decomposition and sparse constraints into the loss function to accelerate reconstruction. Finally, a dual-threshold event detector and GPU streaming architecture are combined to achieve high frame rate real-time imaging. This scheme deeply integrates the broadband transient characteristics and spatiotemporal sparsity of partial discharge pulses into the entire imaging process, reducing the amount of data used compared to traditional methods and significantly improving the localization calculation speed. It can maintain high-precision and rapid localization imaging performance even in noisy environments.

[0022] Please see Figure 1 As shown, the broadband feature coupling and sparse low-rank joint fast imaging method of the present invention, which considers the time-frequency characteristics of partial discharge pulse ultrasound, includes the following steps: Step 1: Establish a purely acoustic propagation physical constraint model for the partial discharge source. First, express the sound pressure field as an ultrasonic audio partial discharge pulse signal. Convolution with Green's function Then, a short-time Fourier transform is used to convert the signal to the time-frequency domain, decoupling the spectral terms. and dissemination items And introduce frequency-varying directivity function Then construct the constraint objective function. .

[0023] Step 2: Construct a time-frequency sparse dictionary based on Morlet wavelets and acoustic delay. First, a complete atom library was designed, with the time domain based on Morlet wavelet bases. Matching pulse decay oscillation, spatial domain adopts Match the acoustic wave delay; then construct a tensor dictionary. The sparse coefficient vector is solved using the OMP algorithm. .

[0024] Step 3: Establish a low-rank-sparse joint optimization objective function and solve the spatiotemporal signal matrix using ADMM iteration. First, establish the joint optimization objective function. Then, the spatiotemporal signal matrix is ​​solved using the ADMM three-stage iterative method.

[0025] Step 4: Construct a physics-guided deep learning network and accelerate reconstruction. First, construct a two-stream network combining sparse and low-rank branches, where the sparse branch uses a one-dimensional CNN to learn sparse coefficient vectors. The time-domain model is obtained by decomposing the low-rank branch into U, Σ, and V through a differentiable SVD layer; then physical constraints are embedded in the loss function. .

[0026] Step 5: Design a real-time imaging engine by combining a dual-threshold event detector with a GPU streaming architecture. First, combine temporal amplitude thresholds. With frequency domain energy threshold Implement dual-threshold event triggering, and then pre-calculate. The pseudo-inverse matrix is ​​obtained, and matrix multiplication and SVD are performed in parallel during the online stage to achieve GPU stream processing.

[0027] Example 1:

[0028] The broadband feature coupling and sparse low-rank joint fast imaging method considering the time-frequency characteristics of partial discharge pulse ultrasound in this embodiment includes the following steps: Step 1: Broadband Pulse Feature Coupling Modeling During partial discharge, energy is excited at insulation defects, and the resulting transient strong electric field causes mechanical deformation of the local medium through electrostriction, thereby generating longitudinal sound waves propagating along the medium. These sound waves reflect the transient dynamic characteristics of partial discharge, and their spectral characteristics are significantly modulated by the discharge current waveform. Therefore, using only the traditional pure acoustic point source model cannot effectively explain the amplitude-frequency changes of the signal received by the sensor in actual measurements, especially in complex structures or monitoring environments with varying angles, where the monitoring accuracy significantly decreases. To address this, this invention couples the discharge pulse waveform and sound wave propagation characteristics in partial discharge to establish a physically constrained model of partial discharge sound source radiation.

[0029] like Figure 2 As shown, multi-channel ultrasonic signals are first acquired, and then processed by a sound pressure field convolution model; the sound pressure field generated by partial discharge is represented as the convolution of the discharge pulse waveform and the Green's function of sound propagation:

[0030] In the formula, To put on a show of force and to indicate the time... and spatial location The sound pressure signal received at the location; For at any time The discharge pulse generated at the partial discharge source; Let be the Green's function for sound propagation, representing the sound source from... To the receiving point Between, through System response after time propagation; Let be the spatial position vector of the power source; This is the spatial position vector of the receiving sensor.

[0031] Because partial discharge generates microsecond-level transient acoustic pulses, their spectral energy is non-uniformly distributed in the range of 20-200kHz. The high-frequency band carries the directional information of the discharge source location, while the low-frequency band reflects the material properties of the insulation structure. To effectively capture these transient characteristics, this invention employs the Short-Time Fourier Transform (STFT) method to convert the signal to the time-frequency domain, performing frequency domain decoupling to separate the spectral components and propagation terms, constructing a frequency domain representation in the following form:

[0032] In the formula, Discharge signal Frequency domain representation; Let be the propagation function, representing the sound source from... To the receiving point Frequency response; It is noise.

[0033] To further enhance the model's characterization of the directionality of partial discharge signals, a frequency-varying directivity function is introduced. This reflects the radiation characteristics of a partial discharge source, similar to a dipole antenna. Specifically, the function has a maximum response in the direction of the discharge channel and tends to zero in the vertical direction. Considering the stronger directivity of high-frequency components and the diffraction effect caused by the geometry of the discharge source, the function... It can effectively describe the energy decay behavior at different angles, and its far-field normalized directivity is defined as:

[0034] Where a is the equivalent aperture radius, Let be the angle between the source spindle and the "source → sensor" direction, and c be the speed of sound in air. First-order Bessel function The specific usage process is as follows: for the first... One channel, with known sensor axial unit vector. Source to sensor direction Find the polar angle Then calculate Then multiply by the propagation operator. ,in , This is frequency-dependent atmospheric attenuation, related to temperature and humidity, and can be calibrated offline or looked up in tables. It can be solved through "parameter optimization". Figure 2 )get (as well as ), and can put The generated directional mask is passed to the network branch of S4 as a physical prior. Parameter optimization mainly refers to the optimization of the source location after step S1. Main spindle direction And, if necessary, perform least-squares estimation of the source spectrum or its low-dimensional parameters.

[0035] The following optimization objective function is then constructed to model the directionality of discharge energy radiation and the sound propagation process:

[0036] In the formula, For the frequency domain representation of the measured signal; It is the Frobenius norm.

[0037] By combining the partial discharge pulse waveform with the acoustic propagation model Coupling can overcome the shortcomings of traditional methods that neglect directionality and high-frequency information, and improve the robustness and positioning accuracy of imaging systems in complex industrial environments.

[0038] Step 2: Construction of Time-Frequency Sparse Dictionary Partial discharge signals exhibit typical non-stationary transient characteristics, manifesting as rapidly decaying oscillations in the time domain and a locally concentrated high-energy structure in the frequency domain. This characteristic allows partial discharge signals to possess good sparse representation capabilities under specific sparse substrates. Therefore, this invention proposes a physically constrained, overcomplete sparse dictionary construction method to replace traditional uniform grid search and general basis function methods, thereby improving reconstruction accuracy and anti-interference capability.

[0039] like Figure 3As shown, firstly, the decaying oscillation behavior of the partial discharge pulse is simulated in the time domain using the Morlet wavelet basis function. The Morlet wavelet possesses both time-domain locality and frequency-domain selectivity, effectively matching the transient characteristics of the partial discharge signal. Its expression is as follows:

[0040] In the formula, The Gaussian envelope is used to simulate pulse decay. This is an oscillation term used to preserve signal phase information; The center frequency is used to match the main frequency of the partial discharge pulse; This is a scaling parameter used to control the width of the wavelet envelope; The wavelet center time corresponds to the pulse arrival time.

[0041] Secondly, in the spatial dimension, adopt A time delay function is used to accurately simulate the propagation delay of sound waves from the source to each receiving sensor. This spatial atomic form is as follows:

[0042] In the formula, To discharge power source position To the receiver location The delay in propagation Let be the Dirac function, representing the response of an ideal point source.

[0043] Based on this, a three-dimensional sparse dictionary is constructed from temporal and spatial atoms through tensor product:

[0044] in:

[0045]

[0046] In the formula, A time-domain sparse dictionary composed of multiple Morlet wavelets with different parameters; It is a spatially sparse dictionary composed of propagation delay functions of source points at different locations.

[0047] Finally, the received multi-channel signals are represented as a sparse combination:

[0048] In the formula, For the observed signal; for the constructed sparse dictionary; It is a sparse coefficient vector. The location and timing of the stimulus representing the real source; This is background noise.

[0049] Sparse coding is implemented using algorithms such as Orthogonal Matching Pursuit (OMP) or Least Absolute Shrinkage and Selection Operator (LASO) to select the atom combinations that best match the characteristics of the actual signal from a high-dimensional redundant atom library. Compared to traditional methods, the physically guided sparse dictionary can automatically adapt to the changing characteristics of partial discharge pulses in the time-frequency domain, effectively distinguishing real signals from noise or reflected waves, and improving positioning accuracy. Simultaneously, it preserves the coherent structure of the signal, providing high-quality foundational data for subsequent low-rank optimization and deep network modeling.

[0050] Step 3: Low-rank-sparse joint optimization reconstruction Partial discharge signals in multi-channel receivers exhibit a significant coexistence of temporal sparsity and spatial low-rank characteristics. Sparsity indicates that each partial discharge event excites a finite region for only a very short time, thus exhibiting a sparse distribution along the time axis. Low-rank indicates that signals received by multiple sensors originate from the same discharge event, possessing high coherence, and the channel signals can be represented as a linear combination of two basis vectors. To comprehensively model the temporal sparsity and spatial low-rank of partial discharge signals, this invention proposes a joint optimization reconstruction framework that integrates sparse representation and low-rank modeling to maximize the extraction of effective information and suppress multi-channel redundancy and noise interference.

[0051] like Figure 4 As shown, the spatiotemporal signal matrix of partial discharge is first constructed, assuming the following layout: Each sensor collects data. Each time sampling point can organize all channel signals into a spatiotemporal matrix. :

[0052] The following objective function is introduced to simultaneously minimize reconstruction error, constrain signal sparsity, and inter-channel low rank:

[0053] In the formula, This is the reconstruction error term, used to ensure that the reconstructed signal output by the model is close to the observed signal in order to improve the model fidelity; To implement sparse constraints, which are used to ensure that impulse events are activated only at a few moments; The nuclear norm is used to ensure that the reconstruction results have a low-rank structure in space, that is, the signals between multiple channels are highly coherent. This is a coupling constraint between sparse and low-rank terms, used to improve the consistency of solutions and avoid optimization divergence or overfitting. This is a regularization parameter used to control the degree of sparsity; is a coupling parameter used to control the consistency strength between sparse and low-rank components.

[0054] Since this optimization problem contains multiple non-differentiable terms and coupled variables, the Alternating Direction Method of Multipliers (ADMM) is used for a block-based iterative solution. The core steps are as follows: (1) Sparse Term Update Using a mixed norm shrinkage algorithm, fix Update sparsity coefficients :

[0055] In the formula, For the mixed norm contraction operator; It is a Lagrange multiplier.

[0056] (2) Low-rank term update Employing a singular value thresholding algorithm, fixing Update the low-rank matrix :

[0057] In the formula, This is a singular value threshold shrinkage operator.

[0058] (3) Lagrange multiplier update Maintaining consistency between two variables by updating the Lagrange multipliers:

[0059] pass and The joint optimization achieves collaborative modeling of three physical priors: sparsity, low rank, and consistency. Sparsity ensures signal reconstruction focuses solely on the impulse events themselves, suppressing background noise; low rank leverages redundancy among multi-channel signals to improve anti-interference capabilities; and consistency constraints enhance model stability and improve the physical interpretability of the solution. Compared to traditional compressed sensing methods that only apply... The method uses the norm, or only performs low-rank decomposition, to more completely describe the multidimensional structural characteristics of partial discharge signals in industrial environments, especially in complex scenarios with high noise or severe reflection, and has higher robustness and positioning accuracy.

[0060] Step 4: Physics-guided Deep Learning Acceleration While the proposed ADMM iterative optimization method achieves high accuracy in solving low-rank sparse models, its high computational cost makes it difficult to meet the millisecond-level response speed requirements of industrial applications. Therefore, this invention further designs a lightweight deep learning network that integrates physical priors and data-driven approaches to rapidly reconstruct sparse coefficients and low-rank structures, thereby significantly accelerating the entire imaging process. Figure 5 As shown, the designed network structure is as follows: (1) Input and preprocessing The network input is the raw time-domain signal matrix acquired from a multi-channel ultrasound array. To enhance the representation of time-frequency features, the data is first fed into a Short-Time Fourier Transform (STFT) layer to generate a time-frequency spectrum. This is used for subsequent spectral masking guidance.

[0061] Among them, the frequency-varying directivity function constructed in step one is used. Generate time-frequency mask This guides the network to focus on the main energy region of the partial discharge signal, thereby enhancing the network's physical interpretability.

[0062] (2) Design of two-stream network To balance the sparsity and low-rank characteristics of the signal, the network adopts a two-branch structure with sparse branches and low-rank branches.

[0063] The sparse branch uses a one-dimensional convolutional neural network (1D CNN) to encode the STFT results, learn the transient sparse activation patterns of the signal, and output a sparse estimate. .

[0064] The low-rank branch introduces a differentiable singular value decomposition (DSVD) layer, which... Alternatively, the intermediate representation can be deconstructed to output a low-rank estimate. .

[0065] (3) Design of physical constraint loss function To ensure that the network output satisfies both physical laws and reconstruction accuracy during training, the following loss function is defined:

[0066] In the formula, This is for signal reconstruction error, used to ensure that the output signal reconstructs the observed signal; This is a sparse regularization term used to achieve sparse representation of pulses; This is a low-rank regularization term used to maintain the consistency of the spatial structure; Consistency constraints for sparse and low-rank solutions; This is a sparse term for the temporal gradient, used to prevent smoothing artifacts from appearing in the network output; , , , , Weights for loss items; This represents the time gradient.

[0067] By explicitly encoding the aforementioned physical laws into the loss function, the network training process becomes more controllable, and the problems of instability and poor generalization of "black box" models are effectively avoided.

[0068] Step 5: Real-time Imaging Engine Design To meet the stringent requirements of power equipment partial discharge monitoring for real-time response, low latency, and edge deployment, a millisecond-level fast imaging engine based on a GPU stream processing architecture was designed. Combined with a dual-threshold event detector and a pre-computation optimization mechanism, it achieves end-to-end accelerated processing from data acquisition to imaging visualization. Figure 6 As shown, the imaging engine structure is as follows: (1) System input and overall process The system input is a continuously acquired multi-channel raw ultrasound signal stream. The overall processing flow mainly includes five modules: event detector, feature coupling calculation module, sparse-low-rank joint reconstruction module, imaging and localization module, and visualization and output interface. Each module is connected through a GPU parallel pipeline to ensure minimal communication overhead and waiting time.

[0069] (2) Design of a dual-threshold event detector Partial discharge events lack stable periodicity and exhibit strong randomness. To reduce invalid calculations and false alarm rates, this system employs a dual-threshold event triggering mechanism using both time-domain amplitude thresholds and frequency-domain energy thresholds. The time-domain amplitude threshold is expressed as follows: Triggered when the amplitude of any channel signal exceeds a set level; the frequency domain energy threshold is expressed as... After STFT analysis, if the energy in a specific frequency band exceeds a threshold, a partial discharge signal is considered to be present. Only when both conditions are met simultaneously will the subsequent feature extraction and imaging process proceed. This method can significantly reduce spurious calculations caused by background noise or interference.

[0070] (3) Feature Coupling Calculation Module After the event detection is triggered, the system starts the feature calculation module and processes the signal as follows: First, it performs STFT on the input signal, combined with the frequency-varying directivity function. Generate a directional mask; then extract the sparse dictionary and the propagation delay matching matrix; finally, use the preloaded tensor dictionary pseudo-inverse matrix. The signal is directly mapped to the sparse coefficient space. By pre-compiling the pseudo-inverse of the key matrix, the online solution time can be significantly reduced, thus providing fast feature initialization for subsequent solvers.

[0071] (4) Joint solver and imaging localization Using the deep network model trained in step four, solve the sparse components in parallel. With low-rank structure Then based on the sparsity coefficient The location of the largest activation is used to determine the most likely discharge source, and low-rank components are used to correct the structure consistency of the results. If multiple activation sources exist, the maximum cluster center or energy weight averaging method is used for fusion.

[0072] (5) GPU pipelined parallel processing architecture The entire system is deployed on a CUDA-enabled GPU, achieving multi-module parallelism through asynchronous streaming operations. Specifically, STFT and directional mask generation utilize the CUDA FFT library; sparse reconstruction and SVD acceleration employ GPU computation in PyTorch / TensorRT; and the imaging module directly generates heatmaps and outputs them to an industrial UI interface.

[0073] Any part of this invention not described in detail can be referred to in the prior art or in the art known to those skilled in the art. This embodiment does not limit such part and will not describe it in detail here.

[0074] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of the present invention.

Claims

1. A broadband feature coupling and sparse low-rank joint imaging method, characterized in that, The method includes: S1. Broadband pulse characteristic coupling modeling: Establish a pure acoustic propagation model of the partial discharge source, integrate the partial discharge pulse waveform and sound wave propagation characteristics, and overcome the defect of the point source model in ignoring directionality and high-frequency information. S2. Construction of Time-Frequency Sparse Dictionary: Based on the physical fingerprint of the time-domain decay oscillation and high-energy concentration of the partial discharge signal, a physically constrained overcomplete sparse dictionary is constructed to replace the traditional general basis function, thereby improving signal discrimination and anti-interference capability. S3. Low-rank-sparse joint optimization reconstruction: Combining the characteristics of temporal sparsity and spatial low-rank of partial discharge signals, a joint optimization objective function is constructed, and the spatiotemporal signal matrix is ​​solved through ADMM iteration to improve noise resistance and positioning accuracy; S4. Physics-guided deep learning acceleration: Design a lightweight deep network that integrates physical priors to achieve millisecond-level signal reconstruction, balancing accuracy and real-time performance; S5. Real-time Imaging Engine Design: Based on a GPU stream processing architecture, combined with dual-threshold detection and pre-computation optimization, it realizes real-time processing of the entire data acquisition-imaging visualization chain, meeting the needs of industrial edge deployment.

2. The broadband feature coupling and sparse low-rank joint imaging method according to claim 1, characterized in that, Step S1 includes: Sound pressure field convolution modeling: The partial discharge sound pressure field is represented as the convolution of the partial discharge pulse signal and the sound propagation Green's function, which accurately describes the propagation process of the sound signal from the source to the sensor; Time-frequency domain decoupling: Short-time Fourier transform is used to convert the time-domain signal to the time-frequency domain, decoupling the spectral components and propagation terms to obtain the frequency domain expression, capturing the transient characteristics of broadband pulses in a specific frequency band; Frequency-varying directivity function embedding: Introducing the frequency-varying directivity function of the simulated partial discharge source dipole antenna radiation to construct an optimized objective function, thereby enhancing the model's characterization of signal directivity.

3. The broadband feature coupling and sparse low-rank joint imaging method according to claim 1, characterized in that, Step S2 includes: Time-domain atomic design: Morlet wavelet basis is used to simulate the decaying oscillation of partial discharge pulse, taking into account both time-domain locality and frequency-domain selectivity; Spatial atomic design: Employing a time delay function to simulate sound wave propagation delay, accurately matching spatial location information; Tensor dictionary construction: By using the tensor product of the temporal Morlet wavelet dictionary and the spatial time delay dictionary, a three-dimensional overcomplete sparse dictionary is generated, which represents the multi-channel observation signal as a formula related to the observation signal, sparse dictionary, sparse coefficients and noise. Sparse coding solution: Orthogonal matching pursuit or minimum absolute shrinkage and selection operator algorithm is used to select the optimal atom combination from the redundant atom library and extract the real partial discharge signal that excludes noise and reflected waves.

4. The broadband feature coupling and sparse low-rank joint imaging method according to claim 1, characterized in that, Step S3 includes: Spatiotemporal signal matrix construction: Assuming multiple sensors and multiple time sampling points for partial discharge, the multi-channel signals are organized into a spatiotemporal signal matrix; Joint optimization objective function: By reconstructing the joint optimization objective function of error, row sparsity constraint, low-rank constraint of nuclear norm, and coupling term correlation, a coordinated expression of physical prior is achieved; ADMM is solved by block iterative method, including sparse term update, low-rank term update and multiplier update.

5. The broadband feature coupling and sparse low-rank joint imaging method according to claim 4, characterized in that, Sparse term update: With the low-rank matrix fixed, the sparse coefficients are updated using a mixed norm shrinkage algorithm; Low-rank term update: With fixed sparse coefficients, the low-rank matrix is ​​updated using the singular value thresholding algorithm; Multiplier update: Update the Lagrange multipliers to maintain consistency between sparse and low-rank components and avoid optimization divergence.

6. The broadband feature coupling and sparse low-rank joint imaging method according to claim 1, characterized in that, Step S4 includes: Input preprocessing: The multi-channel raw signal is fed into the STFT layer to generate the time spectrum, and combined with the frequency-varying directivity function in S1 to generate a directional mask, guiding the network to focus on the main energy region of the partial discharge. Two-stream network structure: The sparse branch uses 1D CNN to encode the STFT results, learns the transient sparse activation pattern of the signal, and outputs a sparse estimate; a differentiable singular value decomposition layer is introduced in the low-rank branch to deconstruct the signal or intermediate representation and output a low-rank estimate. Physical constraint loss function: Define a multi-objective loss function related to reconstruction error, sparse regularization, low-rank regularization, consistency constraint, and temporal gradient sparsity term to avoid poor generalization of black box models and ensure that the output conforms to physical laws.

7. The broadband feature coupling and sparse low-rank joint imaging method according to claim 1, characterized in that, Step S5 includes: Dual-threshold event detection: Employs dual triggering with a time-domain amplitude threshold and a frequency-domain energy threshold to reduce invalid calculations and false alarms; Feature coupling calculation: After triggering, the STFT is executed to generate the spectrum, and the pre-calculated tensor dictionary pseudo-inverse matrix is ​​loaded to directly map the signal to the sparse coefficient space, reducing the online solution time; Joint solution and imaging localization: The sparse / low-rank components are solved in parallel using a deep network trained with S4. The location of the discharge source is determined by the maximum activation position of the sparse coefficients, and the results are corrected by the low-rank components. GPU parallel architecture: Deployed on CUDA-enabled GPUs, it achieves multi-module parallelism through asynchronous streaming.

8. The broadband feature coupling and sparse low-rank joint imaging method according to claim 7, characterized in that, The dual triggering mechanism of time-domain amplitude threshold + frequency-domain energy threshold is as follows: When both conditions are met simultaneously, such as the amplitude of any channel signal exceeding a set level and the energy of a specific frequency band exceeding a threshold, subsequent feature extraction and imaging are triggered.

9. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the steps of the broadband feature coupling and sparse low-rank joint imaging method as described in any one of claims 1-8.

10. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the steps of the broadband feature coupling and sparse low-rank joint imaging method according to any one of claims 1-8.

Images