Generator partial discharge pattern recognition method and device based on multi-scale perception, computer equipment, readable storage medium and program product

Abstract

The application relates to a generator partial discharge mode recognition method and device based on multi-scale perception, computer equipment, a computer readable storage medium and a computer program product, and relates to the technical field of electric power detection, and can improve the detection efficiency of the generator partial discharge mode. The method comprises the following steps: acquiring a partial discharge original signal of a generator stator and converting the partial discharge original signal into a pulse time sequence signal and a partial discharge spectrum. Time sequence features of multiple time scales of the pulse time sequence signal and pixel features of multiple space scales of the spectrum are extracted respectively, feature fusion weights of the features are determined, weighted time sequence features and pixel features are subjected to mode recognition respectively, first and second mode recognition results are obtained, and then a fusion mode recognition result is determined. The feature fusion weights are dynamically adjusted based on the fusion mode recognition result, and the mode recognition step is executed until the fusion mode recognition result meets a preset condition, so that a final partial discharge target mode recognition result is output.

Description

Technical Field

[0001] This application relates to the field of power detection technology, and in particular to a generator partial discharge pattern recognition method, device, computer equipment, computer-readable storage medium, and computer program product based on multi-scale sensing. Background Technology

[0002] Partial discharge caused by generator stator insulation degradation is a major cause of equipment failure, and accurate pattern recognition is crucial for insulation condition monitoring. However, traditional partial discharge pattern recognition methods often rely on single-dimensional feature analysis, resulting in incomplete information representation. Furthermore, due to the complex and variable operating conditions and strong background noise interference, existing recognition methods have weak generalization and adaptability to complex conditions such as multiple defects coexisting, which can easily lead to missed or false alarms.

[0003] Therefore, improving the recognition accuracy of generator partial discharge mode recognition under strong noise and complex and variable operating conditions is a technical problem that urgently needs to be solved. Summary of the Invention

[0004] Therefore, it is necessary to provide a generator partial discharge pattern recognition method, device, computer equipment, computer-readable storage medium, and computer program product based on multi-scale sensing to address the above-mentioned technical problems.

[0005] In a first aspect, this application provides a generator partial discharge pattern recognition method based on multi-scale sensing, including:

[0006] Obtain the original partial discharge signal of the generator stator, and obtain the pulse timing signal and partial discharge spectrum corresponding to the original partial discharge signal;

[0007] Extract the temporal features at different time scales from the pulse timing signal and the pixel features at different spatial scales from the partial discharge spectrum, and determine the feature fusion weights for the temporal features at each time scale and the pixel features at each spatial scale.

[0008] For the temporal features and pixel features after being assigned the feature fusion weights, pattern recognition is performed respectively to obtain a first pattern recognition result and a second pattern recognition result;

[0009] Based on the first pattern recognition result and the second pattern recognition result, a fused pattern recognition result is determined;

[0010] Based on the fusion pattern recognition result, the feature fusion weight is adjusted, and based on the adjusted feature fusion weight, the process returns to perform pattern recognition on the temporal features and pixel features after assigning the feature fusion weight, respectively, to obtain the first pattern recognition result and the second pattern recognition result, until the fusion pattern recognition result meets the preset conditions. Based on the fusion pattern recognition result that meets the preset conditions, the target pattern recognition result of partial discharge is determined.

[0011] In one embodiment, the feature fusion weights include temporal fusion weights and spatial fusion weights; adjusting the feature fusion weights based on the fusion pattern recognition result includes:

[0012] Based on the fusion pattern recognition results, the temporal fusion weights corresponding to the temporal features at different time scales are determined;

[0013] Based on the fusion pattern recognition results, spatial fusion weights for pixel features corresponding to different spatial scales are determined;

[0014] The feature fusion weights are determined based on the temporal fusion weights and the spatial fusion weights.

[0015] In one embodiment, determining the temporal fusion weights for temporal features corresponding to different time scales based on the fusion pattern recognition result includes:

[0016] Based on the fusion pattern recognition results, feature scale perception adjustment is performed through a network architecture search algorithm to determine the temporal fusion weights of temporal features corresponding to different time scales.

[0017] The step of determining the spatial fusion weights of pixel features corresponding to different spatial scales based on the fusion pattern recognition results includes:

[0018] Based on the fusion pattern recognition results, feature scale perception adjustment is performed through dynamic convolution algorithm to determine the spatial fusion weights of pixel features corresponding to different spatial scales.

[0019] In one embodiment, determining the fused pattern recognition result based on the first pattern recognition result and the second pattern recognition result includes:

[0020] The fused pattern recognition result is obtained by fusing the preset result fusion weights, the first pattern recognition result and the second pattern recognition result.

[0021] The step of adjusting the feature fusion weights based on the fusion pattern recognition result, and then, based on the adjusted feature fusion weights, performing pattern recognition on the temporal features and pixel features after assigning the feature fusion weights, respectively, to obtain a first pattern recognition result and a second pattern recognition result, includes:

[0022] Based on the fusion pattern recognition result, adjust the feature fusion weight and the result fusion weight;

[0023] Based on the adjusted feature fusion weights, pattern recognition is performed on the temporal features and pixel features given the adjusted feature fusion weights to obtain an updated first pattern recognition result and an updated second pattern recognition result.

[0024] Based on the adjusted result fusion weights, the updated first pattern recognition result and the updated second pattern recognition result are weighted and fused to obtain an updated fused pattern recognition result, until the updated fused pattern recognition result meets the preset conditions.

[0025] In one embodiment, extracting the temporal features at different time scales in the pulse timing signal and the pixel features at different spatial scales in the partial discharge spectrum includes:

[0026] The pulse timing signal is divided into multiple time-resolution segments according to a preset multi-level time observation window to obtain multiple time-series signals corresponding to different time resolutions.

[0027] Based on the multiple time series signals corresponding to different time resolutions, the temporal characteristics of the different time scales are determined;

[0028] The partial discharge pattern is divided into multiple partial discharge patterns with different grid resolutions;

[0029] The image size of each of the partial discharge electron spectra is resized to obtain a resized partial discharge electron spectra with the same image size as the partial discharge spectra.

[0030] Based on the partial discharge spectrum and the reorganized partial discharge spectrum, the pixel features at different spatial scales are determined.

[0031] In one embodiment, extracting the temporal features at different time scales in the pulse timing signal and the pixel features at different spatial scales in the partial discharge spectrum includes:

[0032] A multi-scale temporal convolutional network is used to extract high-dimensional features from the pulse time-series signal to obtain the temporal features at different time scales;

[0033] A hierarchical visual attention network is used to extract pixel features from the partial discharge map to obtain pixel features at different spatial scales;

[0034] The step of performing pattern recognition on the temporal features and pixel features after assigning the feature fusion weights includes:

[0035] A state-space-based sequence modeling network performs pattern recognition on temporal features after assigning fusion weights to the features;

[0036] Pattern recognition is performed on pixel features after the feature fusion weights are assigned based on a graph-aware attention network.

[0037] Secondly, this application also provides a generator partial discharge pattern recognition device based on multi-scale sensing, comprising:

[0038] The data acquisition module is used to acquire the original partial discharge signal of the generator stator, and to acquire the pulse timing signal and partial discharge spectrum corresponding to the original partial discharge signal;

[0039] The feature extraction module is used to extract the temporal features at different time scales in the pulse timing signal and the pixel features at different spatial scales in the partial discharge spectrum, and to determine the feature fusion weights of the temporal features at each time scale and the pixel features at each spatial scale.

[0040] The pattern recognition module is used to perform pattern recognition on the temporal features and the pixel features after the feature fusion weights are assigned, respectively, to obtain a first pattern recognition result and a second pattern recognition result;

[0041] The recognition result fusion module is used to determine the fused pattern recognition result based on the first pattern recognition result and the second pattern recognition result;

[0042] The dynamic recognition module is used to adjust the feature fusion weights based on the fusion pattern recognition result, and based on the adjusted feature fusion weights, return to the step of performing pattern recognition on the temporal features and the pixel features after assigning the feature fusion weights, respectively, to obtain a first pattern recognition result and a second pattern recognition result, until the fusion pattern recognition result meets a preset condition, and based on the fusion pattern recognition result that meets the preset condition, determine the target pattern recognition result of partial discharge.

[0043] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0044] Obtain the original partial discharge signal of the generator stator, and obtain the pulse timing signal and partial discharge spectrum corresponding to the original partial discharge signal;

[0045] Extract the temporal features at different time scales from the pulse timing signal and the pixel features at different spatial scales from the partial discharge spectrum, and determine the feature fusion weights for the temporal features at each time scale and the pixel features at each spatial scale.

[0046] For the temporal features and pixel features after being assigned the feature fusion weights, pattern recognition is performed respectively to obtain a first pattern recognition result and a second pattern recognition result;

[0047] Based on the first pattern recognition result and the second pattern recognition result, a fused pattern recognition result is determined;

[0048] Based on the fusion pattern recognition result, the feature fusion weight is adjusted, and based on the adjusted feature fusion weight, the process returns to perform pattern recognition on the temporal features and pixel features after assigning the feature fusion weight, respectively, to obtain the first pattern recognition result and the second pattern recognition result, until the fusion pattern recognition result meets the preset conditions. Based on the fusion pattern recognition result that meets the preset conditions, the target pattern recognition result of partial discharge is determined.

[0049] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the following steps:

[0050] Obtain the original partial discharge signal of the generator stator, and obtain the pulse timing signal and partial discharge spectrum corresponding to the original partial discharge signal;

[0051] Extract the temporal features at different time scales from the pulse timing signal and the pixel features at different spatial scales from the partial discharge spectrum, and determine the feature fusion weights for the temporal features at each time scale and the pixel features at each spatial scale.

[0052] For the temporal features and pixel features after being assigned the feature fusion weights, pattern recognition is performed respectively to obtain a first pattern recognition result and a second pattern recognition result;

[0053] Based on the first pattern recognition result and the second pattern recognition result, a fused pattern recognition result is determined;

[0054] Based on the fusion pattern recognition result, the feature fusion weight is adjusted, and based on the adjusted feature fusion weight, the process returns to perform pattern recognition on the temporal features and pixel features after assigning the feature fusion weight, respectively, to obtain the first pattern recognition result and the second pattern recognition result, until the fusion pattern recognition result meets the preset conditions. Based on the fusion pattern recognition result that meets the preset conditions, the target pattern recognition result of partial discharge is determined.

[0055] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, performs the following steps:

[0056] Obtain the original partial discharge signal of the generator stator, and obtain the pulse timing signal and partial discharge spectrum corresponding to the original partial discharge signal;

[0057] Extract the temporal features at different time scales from the pulse timing signal and the pixel features at different spatial scales from the partial discharge spectrum, and determine the feature fusion weights for the temporal features at each time scale and the pixel features at each spatial scale.

[0058] For the temporal features and pixel features after being assigned the feature fusion weights, pattern recognition is performed respectively to obtain a first pattern recognition result and a second pattern recognition result;

[0059] Based on the first pattern recognition result and the second pattern recognition result, a fused pattern recognition result is determined;

[0060] Based on the fusion pattern recognition result, the feature fusion weight is adjusted, and based on the adjusted feature fusion weight, the process returns to perform pattern recognition on the temporal features and pixel features after assigning the feature fusion weight, respectively, to obtain the first pattern recognition result and the second pattern recognition result, until the fusion pattern recognition result meets the preset conditions. Based on the fusion pattern recognition result that meets the preset conditions, the target pattern recognition result of partial discharge is determined.

[0061] The aforementioned generator partial discharge pattern recognition method, apparatus, computer equipment, computer-readable storage medium, and computer program product based on multi-scale perception acquire the original partial discharge signal of the generator stator and the corresponding pulse time-series signal and partial discharge spectrum; extract time-series features at different time scales from the pulse time-series signal and pixel features at different spatial scales from the partial discharge spectrum, and determine the feature fusion weights for the time-series features at each time scale and the pixel features at each spatial scale; perform pattern recognition on the time-series features and pixel features after assigning feature fusion weights, respectively, to obtain a first pattern recognition result and a second pattern recognition result; determine the fused pattern recognition result based on the first and second pattern recognition results; adjust the feature fusion weights based on the fused pattern recognition result, and based on the adjusted feature fusion weights, return to the steps of performing pattern recognition on the time-series features and pixel features after assigning feature fusion weights, respectively, to obtain the first and second pattern recognition results, until the fused pattern recognition result meets preset conditions; and determine the target pattern recognition result of the partial discharge based on the fused pattern recognition result that meets the preset conditions. In this application, by simultaneously extracting multi-timescale features of one-dimensional pulse timing signals and multi-spatial-scale pixel features of two-dimensional partial discharge maps, complementary fusion and comprehensive perception of multi-modal and multi-scale partial discharge information are achieved. On this basis, by using the preliminary candidate pattern recognition results as feedback control signals, an adaptive dynamic adjustment closed-loop iterative mechanism is established, enabling the feature fusion weights to adaptively optimize and correct their representation according to the dynamic changes of the actual partial discharge signals. This completely breaks through the bottleneck of traditional methods, which suffer from poor generalization ability under complex conditions such as strong noise and multiple defects due to single scale or fixed fusion weights, thus improving the accuracy of generator partial discharge pattern recognition. Attached Figure Description

[0062] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0063] Figure 1 This is an application environment diagram of a generator partial discharge mode recognition method based on multi-scale sensing in one embodiment;

[0064] Figure 2 This is a flowchart illustrating a generator partial discharge mode recognition method based on multi-scale sensing in one embodiment.

[0065] Figure 3This is a flowchart illustrating a generator partial discharge pattern recognition method based on multi-scale sensing in another embodiment.

[0066] Figure 4 This is a structural block diagram of a generator partial discharge pattern recognition device based on multi-scale perception in one embodiment;

[0067] Figure 5 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0068] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0069] It should be noted that the terms "first," "second," etc., used in this application can be used to describe various objects, but these objects are not limited by these terms. These terms are only used to distinguish the first object from the second object. The terms "comprising" and "having," and any variations thereof, used in this application, are intended to cover non-exclusive inclusion. The term "multiple" used in this application refers to two or more. The term "and / or" used in this application refers to one of the embodiments, or any combination of multiple embodiments.

[0070] The generator partial discharge pattern recognition method based on multi-scale sensing provided in this application can be applied to, for example... Figure 1 The application environment shown depicts a scenario where the terminal communicates with the server via a network. Data storage can store data that the server needs to process. This data storage can be integrated onto the server or hosted in the cloud or on other network servers. The terminal can receive raw partial discharge signals collected by relevant sensors or testing equipment deployed at the generator end when the generator stator bars are in operation or under partial discharge pressure testing conditions.

[0071] In one embodiment, the executing entity can be a local monitoring terminal, edge computing node, or cloud server device that is connected to the signal acquisition side via a wired or wireless communication network. After acquiring the acquired partial discharge raw signal, it can convert the raw signal into a pulse time series signal and a PRPD map. Relying on its internally deployed deep learning computing architecture, it performs multi-scale segmentation, feature extraction, feature adaptive perception fusion, and closed-loop weight adjustment and pattern recognition on the signal. Finally, it outputs the target pattern recognition result of the generator stator bar defect type, thereby realizing online monitoring or offline status assessment of the generator equipment's operational safety and reliability.

[0072] In one exemplary embodiment, such as Figure 2 As shown, a generator partial discharge pattern recognition method based on multi-scale sensing is provided, which is then applied to... Figure 1 Taking the server in the example, the explanation includes the following steps S201 to S205. Wherein:

[0073] Step S201: Obtain the original partial discharge signal of the generator stator, and obtain the pulse timing signal and partial discharge spectrum corresponding to the original partial discharge signal.

[0074] The original partial discharge signal can be a high-dimensional physical quantity waveform or time-domain voltage sequence directly captured by electromagnetic, acoustic, or optical sensing elements under aging conditions where generator insulation deteriorates in the early stages, without deep processing. It is mainly used to characterize the true dynamics of discharge pulses caused by stator bar insulation deterioration. It can also be transient discharge data generated during partial discharge pressure testing of slot discharge defects, end discharge defects, internal discharge defects, and interphase discharge defects. To facilitate subsequent intelligent analysis, this original partial discharge signal can be further deconstructed and transformed into a pulse timing signal and a partial discharge spectrum through digital sampling and analog-to-digital conversion. The pulse timing signal can be a time series recording the occurrence time and amplitude fluctuations of partial discharge pulses in a continuous time dimension, while the partial discharge spectrum can be a two-dimensional or three-dimensional image reflecting the statistical distribution characteristics between discharge phase, discharge amplitude, and discharge frequency, such as a phase-analyzed partial discharge distribution map under power frequency cycles.

[0075] Specifically, the original partial discharge signal is acquired from the generator stator bars. After acquiring the original partial discharge signal, it undergoes multi-dimensional nonlinear deconstruction and transformation processing to obtain the corresponding pulse time-series signal and partial discharge spectrum in parallel. In this process, the time-domain dynamic sequence representing the characteristics of single or continuous pulse waveforms in the original signal is extracted as the pulse time-series signal. At the same time, the original signal is subjected to multi-dimensional mapping statistics by combining power frequency phase information and statistical frequency information to draw and derive a partial discharge spectrum that can represent the statistical spatial characteristics of the discharge. This provides independent time-series data streams and image spatial data streams in the subsequent feature perception extraction process.

[0076] Optionally, after acquiring the original partial discharge signal, it can be transformed into a joint time-frequency distribution map using short-time Fourier transform or instantaneous time-frequency analysis, serving as the partial discharge map. Simultaneously, amplitude envelope extraction and downsampling transformation can be directly performed on the original discharge pulse to obtain an amplitude fluctuation sequence reflecting the low-frequency macroscopic fluctuation trend, which can then be used as the pulse time-series signal. Furthermore, during the signal acquisition stage, multiple sensors positioned at different observation nodes on the generator stator can simultaneously acquire multidimensional composite original signals from different channels. Through multi-channel joint alignment processing, a composite pulse time-series signal can be generated, providing more diverse alternative basic feature inputs for subsequent feature processing.

[0077] Step S202: Extract the temporal features at different time scales in the pulse timing signal and the pixel features at different spatial scales in the partial discharge spectrum, and determine the feature fusion weights of the temporal features at each time scale and the pixel features at each spatial scale.

[0078] Among them, the time-series features at different time scales can be high-dimensional abstract representations of transient waveform details or long-term evolution laws captured from partial discharge pulse sequences under various time observation ranges of different sizes. They are mainly used to characterize the dynamic evolution trajectory of partial discharge over time and pulse time-series dependent information.

[0079] Pixel features at different spatial scales can be texture or morphological topology information extracted from discharge images at various resolution levels, reflecting the aggregation state of discharge amplitude and phase distribution. This information is used to characterize the structural properties of partial discharge in two-dimensional or three-dimensional spatial distribution.

[0080] Feature fusion weights can be used to quantify and balance the relative importance of the aforementioned temporal features and pixel features in jointly representing a specific defect pattern. They can be represented as a set of dynamically initialized weight matrices, scalar coefficients, or attention score vectors, used to guide the adaptive perception and splicing fusion of multi-source heterogeneous features.

[0081] Specifically, after acquiring the pulse time-series signal and the partial discharge map, multi-scale sequence analysis and processing are performed on the pulse time-series signal to extract temporal features at different time scales. Simultaneously, multi-scale spatial analysis and processing are performed on the partial discharge map to extract pixel features at different spatial scales. After extracting these multi-scale heterogeneous features, feature-level interactive evaluation is further performed on the extracted sets of temporal features at different time dimensions and sets of pixel features at different spatial dimensions. The feature fusion weights for each time scale and each spatial scale are then determined, providing a crucial parameter basis for subsequent dual-modal feature fusion and pattern recognition.

[0082] Optionally, when extracting temporal features at different time scales, sliding time windows of varying lengths and widths can be applied to the pulse time-series signal to extract multiple subsequences with different time spans (e.g., millisecond, second, or minute-level subsequences), and then the hidden layer features in the time dimension can be extracted separately through a sequence encoder. When extracting pixel features at different spatial scales, image pooling pyramids or hierarchical image grid segmentation mechanisms with different receptive field sizes can be constructed to obtain hierarchical spatial pixel features from the microscopic clustering regions of the local discharge phase to the global discharge contour in parallel. Furthermore, when determining feature fusion weights, a prior static weight allocation mechanism based on feature variance or information entropy can be used, or a dynamic attention evaluation network based on the correlation calculation of the inner product of feature vectors can be used to adaptively assign initial weight coefficient values ​​to the temporal features and pixel features of each dimension.

[0083] Step S203: For the temporal features and pixel features after being assigned feature fusion weights, perform pattern recognition respectively to obtain the first pattern recognition result and the second pattern recognition result.

[0084] The first pattern recognition result can be an initial category determination based on the dynamic changes of the pulse sequence in the time domain. It is used to characterize the probability distribution of partial discharge defect types predicted based on the pulse time sequence evolution law.

[0085] The second pattern recognition result can be a parallel category determination based on the discharge distribution pattern in the spatial dimension, which is used to characterize the probability distribution of local discharge defect types predicted based on image texture and phase distribution features.

[0086] Specifically, temporal features and pixel features, weighted by feature fusion, are acquired and then independently analyzed and classified in a parallel-deployed pattern recognition architecture. During this process, temporal categories are inferred for temporal features carrying patterns at different time scales, yielding a first pattern recognition result; simultaneously, spatial categories are inferred for pixel features reflecting structures at different spatial scales, yielding a second pattern recognition result. Subsequently, based on the extracted first and second pattern recognition results, joint verification and result reconstruction are performed at the decision level to determine the fused pattern recognition result that integrates the two types of feature information.

[0087] Step S204: Determine the fused pattern recognition result based on the first pattern recognition result and the second pattern recognition result.

[0088] The fusion pattern recognition result is used to comprehensively evaluate the global decision output of the above dual-modal independent judgment. It represents the final category confidence given after comprehensively considering the temporal continuity and spatial topology, such as the output comprehensive judgment label representing specific insulation defects such as slot discharge and end discharge.

[0089] Optionally, when performing pattern recognition for temporal features and pixel features separately, independent classification branches based on multilayer perceptrons can be constructed to reduce the dimensionality of the weighted high-dimensional feature vectors and map them to initial probability scores for different defect categories. In determining the fused pattern recognition result, the first and second pattern recognition results can be integrated by averaging through a pre-defined majority voting mechanism. Alternatively, an additional shallow aggregation classifier can be introduced to concatenate and perform a quadratic linear mapping on the two previously obtained independent result vectors. This post-decision fusion method generates a more inclusive fusion decision output, which serves as the evaluation basis for the partial discharge type.

[0090] Step S205: Based on the fusion pattern recognition result, adjust the feature fusion weight, and based on the adjusted feature fusion weight, return to the step of performing pattern recognition on the temporal features and pixel features after assigning feature fusion weights, respectively, to obtain the first pattern recognition result and the second pattern recognition result, until the fusion pattern recognition result meets the preset conditions. Based on the fusion pattern recognition result that meets the preset conditions, determine the target pattern recognition result of partial discharge.

[0091] Among them, the preset conditions can be evaluation benchmarks used to measure the confidence of the current model decision, the effectiveness of feature fusion, or the convergence status of network training, and can be used to determine whether to terminate the closed-loop iterative process of adaptive adjustment of feature weights and feature recognition. These conditions can be set based on the expected value of the partial discharge pattern recognition accuracy, the maximum number of iteration optimization rounds, or the convergence threshold of the loss function.

[0092] The target pattern recognition result can be the final output of the partial discharge defect type after multiple rounds of adaptive optimization of feature weights. It is used to characterize the most likely real discharge defect category of the generator stator bar under the current multi-scale sensing architecture. It can be represented as a stable and convergent high-dimensional probability distribution vector or a direct defect classification label, such as the specific warning instruction representing "internal discharge defect" or "end discharge defect" in the final output.

[0093] Specifically, after determining the fusion pattern recognition result, it is compared and analyzed with preset conditions. If the current fusion pattern recognition result does not meet the preset conditions, the aforementioned feature fusion weights are directionally adjusted based on the deviation information fed back by the fusion pattern recognition result. After obtaining the adjusted feature fusion weights, the updated weights are reassigned to the previously extracted temporal features and pixel features, and the process returns to perform pattern recognition on the temporal features and pixel features respectively to obtain updated first and second pattern recognition results, and then the new fusion pattern recognition result is evaluated again. The above process of feature weighting, dual-path recognition, and result evaluation constitutes an adaptive closed-loop iterative loop, which is continuously executed until a generated fusion pattern recognition result meets the preset conditions. At this point, the iterative adjustment process is terminated, and the target pattern recognition result of partial discharge is determined and output based on the fusion pattern recognition result that meets the preset conditions.

[0094] Optionally, during the closed-loop process of adjusting feature fusion weights based on the fusion pattern recognition results, a backpropagation feedback path based on the network loss function can be constructed. By calculating the deviation between the classification prediction distribution and the expected distribution corresponding to the current fusion recognition result, the initial weight ratios of features at each time and spatial scale are updated in reverse. Alternatively, a reinforcement learning optimization strategy based on heuristic search can be introduced to dynamically adjust these weights. When evaluating whether the preset conditions are met, the change in the divergence of the feature weight matrix in two adjacent closed-loop iterations, or the convergence slope of the overall evaluation loss, can be used as a comprehensive stopping condition. For example, when the decrease in network loss in several consecutive iterations is less than a certain minimum value, or when the fluctuation of the feature fusion weights tends to level off, the preset conditions can be considered met, and the current comprehensive judgment state can be solidified and output as the final target pattern recognition result.

[0095] In this embodiment, by simultaneously extracting multi-timescale features of one-dimensional pulse timing signals and multi-spatial-scale pixel features of two-dimensional partial discharge maps, complementary fusion and comprehensive perception of multi-modal and multi-scale partial discharge information are achieved. Based on this, by using the preliminary candidate pattern recognition results as feedback control signals, an adaptive dynamic adjustment closed-loop iterative mechanism is established, enabling the feature fusion weights to adaptively optimize and correct their representation according to the dynamic changes of the actual partial discharge signals. This completely breaks through the bottleneck of traditional methods, which suffer from poor generalization ability under complex conditions such as strong noise and multiple defects due to single scale or fixed fusion weights, thus improving the accuracy of generator partial discharge pattern recognition.

[0096] In one embodiment, extracting temporal features at different time scales from the pulse timing signal and pixel features at different spatial scales from the partial discharge spectrum includes:

[0097] The pulse time series signal is segmented into multiple time series signals with different time resolutions according to a preset multi-level time observation window. Based on the multiple time series signals with different time resolutions, the time series characteristics at different time scales are determined. The partial discharge spectrum is divided into multiple partial discharge electron spectra with different grid resolutions. The image size of each partial discharge electron spectra is resized to obtain a resized partial discharge electron spectra with the same image size as the partial discharge spectrum. Based on the partial discharge spectrum and the resized partial discharge electron spectra, the pixel features at different spatial scales are determined.

[0098] Among them, the multi-level time observation window can be a time boundary range with different length spans used to hierarchically extract the continuously evolving partial discharge data stream. It is mainly used to realize multi-level dynamic capture from extremely short transient waveforms to long-term macroscopic trends. It can be obtained based on multi-time resolution segmentation of pulse timing signals, such as observation intervals corresponding to millisecond-level pulses or minute-level periods.

[0099] Local discharge electron spectra with different grid resolutions can be subsets of images with differentiated field-of-view granularity obtained by spatially dividing the original global discharge image. They are used to characterize the microscopic aggregation morphology of discharge phase and amplitude in local regions.

[0100] Image resizing is a pixel scaling or interpolation transformation process used to uniformly map sub-maps of different physical segment sizes to the same standard dimension. It can provide dimensionally aligned resized local electron emission maps for subsequent homogeneous perception in the network.

[0101] Specifically, when extracting spatiotemporal multi-scale features, the acquired pulse time-series signal is segmented into multiple time-resolution segments according to a preset multi-level time observation window, thereby obtaining multiple time-series signals corresponding to different time resolutions in parallel. Subsequently, based on these multiple time-series signals corresponding to different time resolutions, which contain different waveform evolution patterns, time-domain abstraction representations are performed to determine the time-series features at different time scales. Simultaneously, when processing image spatial features, the original partial discharge pattern is segmented into multiple partial discharge electron patterns with different grid resolutions on a two-dimensional plane. Then, each segmented partial discharge electron pattern is independently resized to obtain a resized partial discharge electron pattern with the same image size as the original partial discharge pattern. Finally, joint feature mapping is performed based on the original partial discharge pattern and the resized partial discharge electron pattern to determine pixel features covering different spatial scales.

[0102] For example, for the segmentation of multi-scale time-series signals, the acquired original pulse sequence signal can be segmented into time-series signals with four spans: a single pulse, a power frequency cycle, 1 second, and 1 minute. This allows for the simultaneous monitoring of high-frequency transient discharge waveforms and long-term insulation degradation trends. For the spatial segmentation and reshaping of partial discharge maps, they can be sequentially segmented into sub-maps of different granularities, such as 8×8, 4×4, and 2×2 grids. These sub-maps are then combined with the unsegmented map that retains global features to construct a multi-pixel-scale feature space. After segmentation, image scaling algorithms such as bilinear interpolation are used to uniformly reshape the partial discharge segmented maps of different grid sizes into images with the same pixel size as the original partial discharge map, facilitating seamless input into subsequent visual perception networks.

[0103] In this embodiment, by employing multi-temporal resolution segmentation and image segmentation and reshaping operations with different grid resolutions, the subsequent deep neural network can perform lossless alignment of feature extraction on microscopic local clusters and macroscopic global contours using a consistent parallel computation graph.

[0104] In one embodiment, the feature fusion weights include temporal fusion weights and spatial fusion weights; based on the fusion pattern recognition result, the feature fusion weights are adjusted, including:

[0105] Based on the fusion pattern recognition results, the temporal fusion weights of temporal features corresponding to different time scales are determined; based on the fusion pattern recognition results, the spatial fusion weights of pixel features corresponding to different spatial scales are determined; based on the temporal fusion weights and spatial fusion weights, the feature fusion weights are determined.

[0106] Among them, the time fusion weight can be a numerical evaluation index used to measure the influence of the partial discharge time series features extracted at each independent time observation scale on decision-making, and is used to dynamically adjust the importance ratio of time series patterns of different lengths in the overall feature representation.

[0107] Spatial fusion weights can be spatial allocation coefficients used to quantify the topological information of the discharge structure captured at different image segmentation resolutions. They are used to characterize the difference in contributions between micro and macro regions in the partial discharge spectrum.

[0108] Specifically, based on the fusion pattern recognition results, the feature fusion weights are adjusted with refined feedback at the decoupled dimension. In this evaluation process, the update path of the feature fusion weights is split into two parallel feedback control branches: temporal and spatial. First, based on the current prediction deviation represented by the fusion pattern recognition results, the temporal fusion weights corresponding to different time scales of temporal features are independently calculated and determined. Simultaneously, based on the fusion pattern recognition results, the spatial fusion weights corresponding to different spatial scales of pixel features are independently calculated and determined. After obtaining the updated temporal and spatial feedback indices, the temporal and spatial fusion weights are numerically combined and reconstructed to determine the overall feature fusion weights, thus returning to execute the next round of feature weighting and dual-path pattern recognition iterations.

[0109] For example, for the time-dimensional feedback branch, the cross-entropy loss between the current fusion pattern recognition result and the expected distribution can be extracted and backpropagated along the gradients of neurons in each layer of the time-series multi-scale extractor. This allows for the calculation of four temporal gradient update rates corresponding to four different time-resolution scales: a single pulse, a power frequency cycle, 1 second, and 1 minute. These rates are then normalized to a local temporal fusion weight vector summing to 1 using the Softmax activation function. For the spatial-dimensional feedback branch, the recognition bias loss can be input into the parameter estimator. For the four sets of multi-scale pixel map features corresponding to 8×8 pixel grids, 4×4 pixel grids, 2×2 pixel grids, and globally unsegmented features, their attention scores under the current recognition decision path are calculated, serving as the spatial fusion weight vector. Finally, the temporal and spatial fusion weight vectors can be combined into a high-dimensional joint weight diagonal matrix using tensor concatenation or matrix diagonalization. This matrix serves as the final determined feature fusion weight, enabling precise numerical scaling and importance control of features at various spatiotemporal scales.

[0110] In this embodiment, by dividing the overall feature fusion weight into temporal fusion weight and spatial fusion weight, and determining and adjusting them independently based on the fusion pattern recognition results, the decoupled fine optimization of the dual-modal heterogeneous feature weights of the partial discharge signal is achieved. This not only effectively avoids gradient interference and mutual masking caused by the difference in dimensionality between temporal fluctuation features and spatial clustering features during mixed optimization, but also enables the allocation of more sensitive and accurate scale-aware weights for specific discharge defects (such as slot discharge or internal discharge) with significant transient abrupt changes or local spatial clustering. This improves the convergence efficiency of feature feedback adjustment and the classification robustness of target pattern recognition results.

[0111] In one embodiment, based on the fusion pattern recognition results, the temporal fusion weights corresponding to temporal features at different time scales are determined, including:

[0112] Based on the fusion pattern recognition results, feature scale perception adjustment is performed through a network architecture search algorithm to determine the temporal fusion weights of temporal features corresponding to different time scales.

[0113] Based on the fusion pattern recognition results, spatial fusion weights for pixel features corresponding to different spatial scales are determined, including:

[0114] Based on the fusion pattern recognition results, a dynamic convolution algorithm is used to adjust the feature scale perception and determine the spatial fusion weights of pixel features corresponding to different spatial scales.

[0115] Among them, network architecture search algorithms can be intelligent optimization mechanisms that automatically find the optimal feature combination configuration in a preset operation search space or branch path. They are mainly used to adaptively derive and determine the most favorable weight allocation strategy for multi-timescale features based on the current classification performance. This can be generated based on gradient analysis and continuous calculation of the continuous evolution law of time-series pulses. For example, in specific scenarios, it can manifest as a differentiable architecture search mechanism. Dynamic convolution algorithms can be adaptive perception techniques in the field of image topology processing that can dynamically adjust their own filter kernel parameters according to the input data content. They are used to give flexible perceptual attention to different pixel grid resolutions and spatial aggregation positions when fusing partial discharge spectrum features. This can be obtained through feature aggregation and attention weight generation modules, such as dynamic convolutional layers driven by image content.

[0116] Specifically, when independently evaluating the weights of the temporal and spatial dimensions based on the fusion pattern recognition results, for the temporal dimension adjustment branch, based on the prediction bias fed back by the fusion pattern recognition results, a deep feature scale-aware adjustment is performed by calling a network architecture search algorithm. During this process, adaptive path evolution calculations are performed at the network layers of the time series to determine the temporal fusion weights of temporal features corresponding to different time scales. Simultaneously, for the spatial dimension adjustment branch, based on the fusion pattern recognition results, a dynamic convolution algorithm is called to perform feature scale-aware adjustment of local spatial features of the image, enabling the model to dynamically generate spatial scaling coefficients according to the grid resolution differences of the spectral features, thereby determining the spatial fusion weights of pixel features corresponding to different spatial scales.

[0117] For example, in practical applications, the aforementioned adjustment process for multi-scale time-series signals can employ a partial discharge pulse (PD) temporal feature scale fusion weight adjustment strategy based on Differentiable Architecture Search (DARTS). Four temporal feature inputs at different time scales are constructed into a hybrid operation space. Utilizing the continuous relaxation technique in the DARTS algorithm, the discrete time scale selection problem is transformed into a continuous weight optimization problem. The optimal distribution coefficients of each time series branch are directly calculated via backpropagation as the temporal fusion weights. For multi-pixel scale PRPD spectral signals, a PRPD pixel feature fusion weight adjustment strategy based on Dynamic Convolution can be used. The attention module within this mechanism calculates a set of dynamic convolution kernel combination parameters in real time based on the currently input PRPD pixel feature tensor. These parameters are then directly applied to weighted fusion operations at different grid scales (such as 8×8, 4×4, 2×2, and global features), serving as the determined spatial fusion weights to achieve efficient dynamic overlay of image features.

[0118] In this embodiment, feature scale-aware adjustments in the temporal and spatial dimensions are made through network architecture search algorithm and dynamic convolution algorithm, respectively, realizing automated and highly nonlinear optimization of bimodal feature weights. This ensures that when the model faces complex partial discharge spectra with different phase aggregation patterns or strong electromagnetic noise interference, it can flexibly and dynamically focus on the most discriminative feature subspace, thereby significantly improving the convergence efficiency of the feature weight adjustment process and the robustness of the final decision.

[0119] In one embodiment, determining the fused pattern recognition result based on the first pattern recognition result and the second pattern recognition result includes:

[0120] The results of the first and second pattern recognitions are fused based on the preset result fusion weights to obtain the fused pattern recognition result.

[0121] Based on the fusion pattern recognition results, the feature fusion weights are adjusted, and based on the adjusted feature fusion weights, the process of performing pattern recognition on the temporal features and pixel features after assigning feature fusion weights is returned to obtain the first pattern recognition result and the second pattern recognition result includes:

[0122] Based on the fusion pattern recognition results, the feature fusion weights and result fusion weights are adjusted. Based on the adjusted feature fusion weights, pattern recognition is performed on the temporal features and pixel features with the adjusted feature fusion weights to obtain updated first pattern recognition results and updated second pattern recognition results. Based on the adjusted result fusion weights, the updated first pattern recognition results and updated second pattern recognition results are weighted and fused to obtain updated fusion pattern recognition results, until the updated fusion pattern recognition results meet the preset conditions.

[0123] Among them, the preset result fusion weight can be a numerical allocation coefficient used in the decision-making stage to balance the contributions of the temporal recognition branch decision conclusion and the spatial recognition branch decision conclusion. It can be used to coordinate the first mode recognition result and the second mode recognition result when the dual-path decision results conflict. It can be represented as a set of weighted scalars for different modal recognition channels.

[0124] The updated fusion pattern recognition result can be a more accurate probability distribution of partial discharge defect types, which is recalculated after the optimization of feature fusion weights and the dynamic adjustment of result fusion weights. It is used to characterize the classification state achieved by the model after multiple rounds of feature correction and decision route reconstruction.

[0125] Specifically, when determining the fusion decision based on the first and second pattern recognition results, a preset result fusion weight is invoked, and this result fusion weight is comprehensively weighted and calculated with the extracted first and second pattern recognition results to obtain the fused pattern recognition result in the initial state. When initiating the subsequent error feedback and adaptive adjustment process, the aforementioned feature fusion weight and result fusion weight are simultaneously updated and adjusted based on the prediction state reflected by the fused pattern recognition result. After obtaining the adjusted feature fusion weight, the temporal features and pixel features assigned the updated weights are subjected to pattern recognition operations again to derive the updated first and second pattern recognition results. Subsequently, based on the synchronously obtained adjusted result fusion weight, this set of updated dual-path pattern recognition results is weighted and fused to obtain the updated fused pattern recognition result. The above-mentioned dual weight closed-loop adjustment process for the feature extraction layer and the decision fusion layer will continue to be executed until the updated fused pattern recognition result meets the set preset conditions.

[0126] For example, during the aforementioned weighted fusion and dynamic adjustment of results, an adaptive feature fusion (AdaFusion) module and a dynamic routing mechanism can be deployed to achieve deep decision-making linkage. In the initial result fusion stage, the AdaFusion algorithm framework can be used to perform multi-dimensional weighted combination of the recognition output vectors of multi-scale temporal signals and multi-scale PRPD maps to form a global recognition output. When entering the closed-loop iterative circuit, not only are architecture search and dynamic convolution invoked to adjust the underlying feature scale weights, but the dynamic routing mechanism is also triggered. This dynamic routing mechanism can dynamically calculate and allocate routing communication probabilities between the temporal recognition branch and the pixel recognition branch based on the backpropagation of the current classification error, using these probabilities as the adjusted result fusion weights. Through this repeated allocation of routing probabilities and correction of underlying weights, the network model continuously approaches the target accuracy.

[0127] In this embodiment, by introducing result fusion weights on the basis of feature fusion layer, and performing synchronous closed-loop iteration and dynamic adjustment of these two types of weights based on the fusion pattern recognition results, the whole-link dual collaborative optimization from the bottom feature perception space to the top logic decision space is realized, which enhances the anti-interference ability, decision confidence and classification accuracy of the generator partial discharge pattern recognition model under changing operating conditions.

[0128] In one embodiment, extracting temporal features at different time scales from the pulse timing signal and pixel features at different spatial scales from the partial discharge spectrum includes:

[0129] A multi-scale temporal convolutional network is used to extract high-dimensional features from the pulse time-series signal to obtain temporal features at different time scales; a hierarchical visual attention network is used to extract pixel features from the partial discharge spectrum to obtain pixel features at different spatial scales.

[0130] For the temporal features and pixel features after feature fusion weighting, pattern recognition is performed separately, including:

[0131] A state-space-based sequence modeling network performs pattern recognition on temporal features after feature fusion weighting; a graph-aware attention network performs pattern recognition on pixel features after feature fusion weighting.

[0132] Among them, multi-scale temporal convolutional networks can be a one-dimensional deep learning topology framework with parallel or cascaded receptive field structures, used to nonlinearly map high-dimensional hidden layer representations that take into account both high-frequency transient fluctuations and low-frequency long-term decay from discrete partial discharge time-series waveforms. They can be constructed based on dilated convolution or multi-branch feature extraction operators.

[0133] Hierarchical visual attention network can be a visual extraction model designed for two-dimensional partial discharge maps, with local window self-attention and cross-layer spatial aggregation capabilities. It is used to perceive the two-dimensional texture clustering pattern of discharge phase and frequency at different resolution levels, and can be represented as a visual converter network with a pyramid structure.

[0134] State-space sequence modeling networks are mainly used to capture the nonlinear dependencies between distant time nodes in ultra-long sequences. Through the discretized mathematical expression of continuous time, global pattern recognition of the temporal evolution trajectory of partial discharges can be achieved.

[0135] Graph-aware attention networks are mapping networks used to transform discharge features scattered in a two-dimensional pixel grid into a structured graph node sequence for higher-order classification. They can rely on the attention interaction between nodes to output defect type identification results based on spatial topology.

[0136] Specifically, when extracting features from the deconstructed bimodal signal, a multi-scale temporal convolutional network is used to extract high-dimensional features from the pulse time-series signal, thereby obtaining the aforementioned temporal features at different time scales. Simultaneously, a hierarchical visual attention network is used in parallel to extract deep pixel features from the segmented and reconstructed partial discharge map, obtaining pixel features at different spatial scales. After adaptive weighting of the features, for the temporal features with assigned feature fusion weights, a state-space-based sequence modeling network performs independent pattern recognition to derive its temporal dimension prediction results; while for the pixel features with assigned feature fusion weights, a graph-aware attention network performs independent pattern recognition to derive its spatial dimension prediction results.

[0137] For example, in practical industrial deployment scenarios, the extraction of the aforementioned multi-scale temporal features can be specifically implemented using the MTNet (Multi-Task / Multi-Scale Time Series Network), which can extract feature vectors in parallel for sequences ranging from single pulses to 1 minute in length. For pixel feature extraction, the Swin Transformer can be introduced as a hierarchical visual attention network. It effectively performs local and global self-attention feature perception on the reorganized PRPD multi-scale map while limiting computational complexity through a shifted window mechanism. In the pattern recognition stage, the S4 (Structured State Space Model) can be used as the sequence modeling network for the state space. Leveraging its superior ability to handle long-distance dependencies, it performs high-precision classification of defect types for highly weighted multi-scale pulse temporal features. Simultaneously, the Graphormer architecture is used as a graph-aware attention network to transform the pixel features of the PRPD map into a graph structure for centrality and spatial distance encoding, thereby achieving high-dimensional category calculation of the map's pixel features.

[0138] In this embodiment, feature extraction is performed using a multi-scale temporal convolutional network and a hierarchical visual attention network, and pattern recognition is carried out independently by combining a state-space sequence modeling network and a graph-aware attention network. This achieves a deep network adaptation that is highly targeted to the multimodal characteristics of generator partial discharge signals. It not only overcomes the gradient vanishing problem of traditional recurrent neural networks when processing long-term partial discharge sequences, but also breaks through the local receptive field limitation of conventional convolutional neural networks when perceiving global discharge patterns. This enables the accurate capture of the spatiotemporal evolution of insulation defects under complex electromagnetic interference, thereby greatly improving the final partial discharge pattern recognition accuracy.

[0139] Specifically, this embodiment provides a complete execution scheme for multi-scale sensing generator partial discharge mode recognition, such as... Figure 3 As shown, it specifically includes:

[0140] First, the data acquisition and multi-scale segmentation process is initiated in the first stage. For four pre-prepared typical partial discharge defects in generator stator bars (including slot discharge defects, end discharge defects, internal discharge defects, and interphase discharge defects), partial discharge pressure tests are conducted to acquire raw partial discharge signals under different defect types. Subsequently, these raw partial discharge signals are deconstructed into two parallel data streams: a pulse time-series signal and a PRPD map. For the pulse time-series signal, a multi-time-scale time-series signal segmentation operation is performed, dividing it into four time-series signals containing a single pulse, one power frequency cycle, 1 second duration, and 1 minute duration. Simultaneously, for the two-dimensional PRPD map, a multi-pixel-scale PRPD map segmentation operation is performed, dividing it into 8×8, 4×4, and 2×2 sub-maps, as well as an unsegmented map retaining global features. Image resizing technology is then used to uniformly resize the segmented maps with different grid resolutions into images of the same size as the original PRPD map, thus completing the preprocessing and construction of the multi-scale heterogeneous data.

[0141] Next, the second stage involves deep learning-driven multi-scale feature extraction and initial fusion. For the temporal dimension, MTNet (Multi-Task / Multi-Scale Temporal Network) is used to extract high-dimensional features in parallel from the four pulse sequence signals obtained from the previous segmentation at different time scales, yielding multi-scale temporal features of the partial discharge pulses. For the spatial dimension, SwinTransformer (Hierarchical Visual Attention Network) is used to perform deep pixel feature perception on the four reshaped PRPD maps at different pixel scales, yielding multi-scale pixel features of the PRPD map. After independent extraction of the dual-path features, the multi-scale pulse temporal features and PRPD map pixel features are input together into a fusion unit based on DAFM (Dynamic Attention Fusion Module). This DAFM module performs adaptive perceptual fusion of the multi-path features and assigns and adjusts initial feature fusion weights for each temporal and spatial scale based on a built-in dynamic weighting mechanism, ultimately outputting multi-dimensional joint features with added feature fusion weights.

[0142] Finally, the third stage involves multi-scale feature fusion and a deep learning-based closed-loop optimization process for pattern recognition. For the temporal features output from the previous stage, weighted by feature fusion, independent defect type inference is performed based on the S4 model (Structured StateSpace Model), outputting the first pattern recognition result (multi-scale temporal feature recognition result). Simultaneously, for the pixel features weighted by feature fusion, independent defect type inference is performed based on Graphermer (Graph Aware Attention Network), outputting the second pattern recognition result (multi-scale pixel feature recognition result). Subsequently, based on the AdaFusion mechanism and combined with preset result fusion weights, the above dual-path recognition results are weighted and fused to obtain the initial fused pattern recognition result.

[0143] To further improve recognition accuracy and generalization ability, an adaptive feature scale-aware adjustment and dynamic routing closed-loop mechanism is triggered based on the deviation between the current fusion pattern recognition result and the target expectation. On the one hand, feature scale aware adjustment is performed at the lower layer: an algorithm based on DARTS (Differentiable Architecture Search) is invoked to update the scale fusion weights (i.e., temporal fusion weights) of the partial discharge pulse temporal features in reverse; simultaneously, an algorithm based on Dynamic Convolution is invoked to adaptively reconstruct the scale fusion weights (i.e., spatial fusion weights) of the PRPD pixel features. On the other hand, dynamic routing adjustment is performed at the upper layer: based on the updated feature fusion weights, the aforementioned S4 model and Graphormer model are controlled to output new multi-scale pulse temporal feature recognition results and multi-scale PRPD pixel feature recognition results again, and a dynamic routing mechanism is introduced to dynamically reallocate and adjust the result fusion weights of these two new results (dynamic weighting mechanism readjustment). The above feedback optimization process will be repeatedly executed multiple times until the latest fusion pattern recognition result meets the preset conditions (e.g., reaching the set target accuracy or loss function). Finally, the current optimal combination of network parameters is solidified, and based on the judgment state when the preset conditions are met, the final target identification result of the generator stator bar defect type is output.

[0144] In this embodiment, multi-type partial discharge (PD) signals are comprehensively collected and analyzed by combining typical defect simulation experiments with PD data under actual operating conditions. By constructing a multi-scale representation of pulse time-series signals and PRPD maps, and introducing a multi-scale feature extraction network based on MTNet and Swin Transformer, and integrating dynamic attention mechanisms and adaptive weighting strategies, efficient joint modeling and deep recognition of PD signal temporal and image features are achieved. This method can effectively improve the recognition accuracy of PD modes for different defect types, is suitable for offline detection and online monitoring of generator insulation status, and has significant engineering practical value and broad application prospects.

[0145] It should be understood that although the steps in the flowcharts of the above embodiments are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the above embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages in other steps. It is understood that the steps in different embodiments can be freely combined as needed, and all non-contradictory solutions formed by such combinations are within the scope of protection of this application.

[0146] Based on the same inventive concept, this application also provides a generator partial discharge pattern recognition device based on multi-scale sensing for implementing the generator partial discharge pattern recognition method based on multi-scale sensing described above. The solution provided by this device is similar to the solution described in the above method. Therefore, the specific limitations in one or more embodiments of the generator partial discharge pattern recognition device based on multi-scale sensing provided below can be found in the limitations of the generator partial discharge pattern recognition method based on multi-scale sensing described above, and will not be repeated here.

[0147] In one exemplary embodiment, such as Figure 4 As shown, a generator partial discharge pattern recognition device based on multi-scale sensing is provided, comprising: a data acquisition module 410, a feature extraction module 420, a pattern recognition module 430, a recognition result fusion module 440, and a dynamic recognition module 450, wherein:

[0148] The data acquisition module 410 is used to acquire the original partial discharge signal of the generator stator, and to acquire the pulse timing signal and partial discharge spectrum corresponding to the original partial discharge signal;

[0149] The feature extraction module 420 is used to extract the temporal features at different time scales in the pulse timing signal and the pixel features at different spatial scales in the partial discharge spectrum, and to determine the feature fusion weights of the temporal features at each time scale and the pixel features at each spatial scale.

[0150] The pattern recognition module 430 is used to perform pattern recognition on the temporal features and the pixel features after the feature fusion weights are assigned, respectively, to obtain a first pattern recognition result and a second pattern recognition result;

[0151] The recognition result fusion module 440 is used to determine the fused pattern recognition result based on the first pattern recognition result and the second pattern recognition result;

[0152] The dynamic recognition module 450 is used to adjust the feature fusion weights based on the fusion pattern recognition result, and based on the adjusted feature fusion weights, return to the step of performing pattern recognition on the temporal features and the pixel features after being given the feature fusion weights, respectively, to obtain a first pattern recognition result and a second pattern recognition result, until the fusion pattern recognition result meets a preset condition, and based on the fusion pattern recognition result when the preset condition is met, determine the target pattern recognition result of partial discharge.

[0153] In one embodiment, the feature extraction module 420 is further configured to:

[0154] The pulse timing signal is segmented into multiple time-resolution segments according to a preset multi-level time observation window to obtain multiple time-series signals corresponding to different time resolutions; based on the multiple time-series signals corresponding to different time resolutions, the temporal characteristics of the different time scales are determined; the partial discharge spectrum is segmented into multiple partial discharge electron spectra with different grid resolutions; the image size of each partial discharge electron spectra is resized to obtain a resized partial discharge electron spectra with the same image size as the partial discharge spectrum; based on the partial discharge spectrum and the resized partial discharge electron spectra, the pixel features of the different spatial scales are determined.

[0155] In one embodiment, the feature extraction module 420 is further configured to:

[0156] A multi-scale temporal convolutional network is used to extract high-dimensional features from the pulse time-series signal to obtain temporal features at different time scales; a hierarchical visual attention network is used to extract pixel features from the partial discharge spectrum to obtain pixel features at different spatial scales.

[0157] In one embodiment, the pattern recognition module 430 is further configured to:

[0158] A state-space-based sequence modeling network performs pattern recognition on temporal features after assigning fusion weights to the features; a graph-aware attention network performs pattern recognition on pixel features after assigning fusion weights to the features.

[0159] In one embodiment, the recognition result fusion module 440 is further configured to:

[0160] The fused pattern recognition result is obtained by fusing the preset result fusion weights, the first pattern recognition result, and the second pattern recognition result.

[0161] In one embodiment, the dynamic recognition module 450 is further configured to:

[0162] Based on the fusion pattern recognition results, the temporal fusion weights corresponding to temporal features at different time scales are determined; based on the fusion pattern recognition results, the spatial fusion weights corresponding to pixel features at different spatial scales are determined; based on the temporal fusion weights and the spatial fusion weights, the feature fusion weights are determined.

[0163] In one embodiment, the dynamic recognition module 450 is further configured to:

[0164] Based on the fusion pattern recognition results, feature scale perception adjustment is performed through a network architecture search algorithm to determine the temporal fusion weights of temporal features corresponding to different time scales; based on the fusion pattern recognition results, feature scale perception adjustment is performed through a dynamic convolution algorithm to determine the spatial fusion weights of pixel features corresponding to different spatial scales.

[0165] In one embodiment, the dynamic recognition module 450 is further configured to:

[0166] Based on the fusion pattern recognition result, the feature fusion weight and the result fusion weight are adjusted; based on the adjusted feature fusion weight, pattern recognition is performed on the temporal feature and the pixel feature with the adjusted feature fusion weight to obtain an updated first pattern recognition result and an updated second pattern recognition result; based on the adjusted result fusion weight, the updated first pattern recognition result and the updated second pattern recognition result are weighted and fused to obtain an updated fusion pattern recognition result, until the updated fusion pattern recognition result meets the preset condition.

[0167] Each module in the aforementioned generator partial discharge pattern recognition device based on multi-scale sensing can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0168] In one exemplary embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 5 As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operating system and computer programs stored in the non-volatile storage media. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. When the computer program is executed by the processor, it implements a generator partial discharge pattern recognition method based on multi-scale sensing.

[0169] Those skilled in the art will understand that Figure 5 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0170] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0171] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0172] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0173] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0174] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0175] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0176] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for generator partial discharge pattern recognition based on multiscale perception, characterized in that, The method includes: Obtain the original partial discharge signal of the generator stator, and obtain the pulse timing signal and partial discharge spectrum corresponding to the original partial discharge signal; The pulse timing signal is divided into multiple time-resolution segments according to a preset multi-level time observation window to obtain multiple time-series signals corresponding to different time resolutions. Based on the multiple time series signals corresponding to different time resolutions, a multi-scale temporal convolutional network is used to extract high-dimensional features to obtain temporal features at different time scales. The partial discharge pattern is divided into multiple partial discharge electron patterns with different grid resolutions. The image size of each partial discharge electron pattern is resized to obtain a resized partial discharge electron pattern with the same image size as the partial discharge pattern. Based on the partial discharge spectrum and the reorganized partial discharge spectrum, a hierarchical visual attention network is used to extract pixel features to obtain pixel features at different spatial scales. Determine the feature fusion weights for temporal features at each time scale and pixel features at each spatial scale; A state-space-based sequence modeling network performs pattern recognition on the temporal features after assigning fusion weights to the features to obtain a first pattern recognition result, and a graph-aware attention network performs pattern recognition on the pixel features after assigning fusion weights to the features to obtain a second pattern recognition result. Based on the first pattern recognition result and the second pattern recognition result, a fused pattern recognition result is determined; Based on the fusion pattern recognition result, the feature fusion weights are adjusted, and based on the adjusted feature fusion weights, the state-space-based sequence modeling network is used to perform pattern recognition on the temporal features after the feature fusion weights are applied to obtain a first pattern recognition result. Then, based on the graph-aware attention network, the pixel features after the feature fusion weights are applied to obtain a second pattern recognition result. This process continues until the fusion pattern recognition result meets a preset condition. Based on the fusion pattern recognition result that meets the preset condition, the target pattern recognition result for partial discharge is determined.

2. The method of claim 1, wherein, The feature fusion weights include temporal fusion weights and spatial fusion weights; The step of adjusting the feature fusion weights based on the fusion pattern recognition result includes: Based on the fusion pattern recognition results, the temporal fusion weights corresponding to the temporal features at different time scales are determined; Based on the fusion pattern recognition results, spatial fusion weights for pixel features corresponding to different spatial scales are determined; The feature fusion weights are determined based on the temporal fusion weights and the spatial fusion weights.

3. The method of claim 2, wherein, The step of determining the temporal fusion weights for temporal features corresponding to different time scales based on the fusion pattern recognition results includes: Based on the fusion pattern recognition results, feature scale perception adjustment is performed through a network architecture search algorithm to determine the temporal fusion weights of temporal features corresponding to different time scales. The step of determining the spatial fusion weights of pixel features corresponding to different spatial scales based on the fusion pattern recognition results includes: Based on the fusion pattern recognition results, feature scale perception adjustment is performed through dynamic convolution algorithm to determine the spatial fusion weights of pixel features corresponding to different spatial scales.

4. The method of claim 1, wherein, The step of determining the fused pattern recognition result based on the first pattern recognition result and the second pattern recognition result includes: The fused pattern recognition result is obtained by fusing the preset result fusion weights, the first pattern recognition result and the second pattern recognition result. The steps of the state-space-based sequence modeling network performing pattern recognition on the temporal features after assigning the feature fusion weights to obtain a first pattern recognition result, and performing pattern recognition on the pixel features after assigning the feature fusion weights based on a graph-aware attention network to obtain a second pattern recognition result, include: Based on the fusion pattern recognition result, adjust the feature fusion weight and the result fusion weight; Based on the adjusted feature fusion weights, pattern recognition is performed on the temporal features and pixel features given the adjusted feature fusion weights to obtain an updated first pattern recognition result and an updated second pattern recognition result. Based on the adjusted result fusion weights, the updated first pattern recognition result and the updated second pattern recognition result are weighted and fused to obtain an updated fused pattern recognition result, until the updated fused pattern recognition result meets the preset conditions.

5. A multi-scale perception based generator partial discharge pattern recognition apparatus, characterized in that, The device includes: The data acquisition module is used to acquire the original partial discharge signal of the generator stator, and to acquire the pulse timing signal and partial discharge spectrum corresponding to the original partial discharge signal; The feature extraction module is used to segment the pulse time-series signal into multiple time-resolution segments according to a preset multi-level time observation window, obtaining multiple time-series signals corresponding to different time resolutions; based on the multiple time-series signals corresponding to different time resolutions, a multi-scale temporal convolutional network is used to perform high-dimensional feature extraction to obtain temporal features at different time scales; the partial discharge spectrum is segmented into multiple partial discharge electron spectra with different grid resolutions, and the image size of each partial discharge electron spectra is resized to obtain a resized partial discharge electron spectra with the same image size as the partial discharge spectrum; based on the partial discharge spectrum and the resized partial discharge electron spectra, a hierarchical visual attention network is used to extract pixel features to obtain pixel features at different spatial scales; and the feature fusion weights of the temporal features at each time scale and the pixel features at each spatial scale are determined. The pattern recognition module is used to perform pattern recognition on the temporal features after the feature fusion weights are assigned based on a state-space sequence modeling network to obtain a first pattern recognition result, and to perform pattern recognition on the pixel features after the feature fusion weights are assigned based on a graph-aware attention network to obtain a second pattern recognition result. The recognition result fusion module is used to determine the fused pattern recognition result based on the first pattern recognition result and the second pattern recognition result; The dynamic recognition module is used to adjust the feature fusion weights based on the fusion pattern recognition result, and based on the adjusted feature fusion weights, return the state-space-based sequence modeling network to perform pattern recognition on the temporal features after assigning the feature fusion weights to obtain a first pattern recognition result, and perform pattern recognition on the pixel features after assigning the feature fusion weights based on a graph-aware attention network to obtain a second pattern recognition result, until the fusion pattern recognition result meets a preset condition, and based on the fusion pattern recognition result when the preset condition is met, determine the target pattern recognition result of partial discharge.

6. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 4.

8. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 4.

Images