Partial discharge identification method, device, equipment, medium and program product

Abstract

The invention provides a partial discharge identification method, device and equipment, a medium and a program product, and particularly relates to the technical field of power equipment monitoring. The method comprises the following steps: inputting a phase resolution partial discharge diagram or a phase resolution partial discharge signal diagram into a preset double-branch identification model; performing feature extraction on the input phase resolution partial discharge diagram or the phase resolution partial discharge signal diagram through a self-adaptive long-short-term memory module in a preset double-branch identification model to obtain a first output representation; performing feature extraction on the input phase resolution partial discharge diagram or the phase resolution partial discharge signal diagram through a self-adaptive linear state space module in a preset double-branch identification model to obtain a second output representation; and carrying out fusion processing on the first output representation and the second output representation to obtain a partial discharge identification result. The method is used for achieving the effect of improving the classification precision and robustness of partial discharge identification.

Description

Technical Field

[0001] This application relates to the field of power equipment monitoring technology, and in particular to a partial discharge identification method, device, equipment, medium, and program product. Background Technology

[0002] In power systems, high-voltage equipment such as oil-immersed transformers and shunt reactors are subjected to high voltage and thermal stress for extended periods. Insulation defects can gradually evolve from partial discharge (PD) into breakdown accidents. Partial discharge refers to a microscopic discharge phenomenon occurring in localized areas within or on the surface of insulating materials, accompanied by charge migration and electromagnetic, acoustic, optical, and thermal effects. If not identified and intervened in a timely manner, partial discharge can lead to erosion of the insulating dielectric structure, localized temperature rise, carbonization, and even fire, ultimately causing power outages and equipment damage. Therefore, power systems need to monitor the type and progression of partial discharges in real time and accurately to assess the insulation health of equipment and develop preventative maintenance strategies.

[0003] Currently, image-based analysis techniques based on Phase Resolved Partial Discharge (PRPD) and Phase Resolved Partial Discharge Signal (PRPS) have become mainstream methods. PRPD / PRPS, by mapping partial discharge signals onto power frequency phase maps and characterizing discharge density or amplitude distribution using pixel brightness or color, can intuitively reflect discharge characteristics. However, in practical applications, PRPD / PRPS images are often affected by strong noise, equipment differences, and sensor link drift, leading to unstable image quality. Furthermore, the diverse types of discharges (such as internal discharge, surface discharge, and corona discharge) and the potential overlap of discharge signals with other interference sources (such as switching actions and wireless communication) further increase the difficulty of identification. Existing technologies struggle to balance long-range phase dependency modeling and local fine-grained feature extraction in complex noise and cross-device scenarios, resulting in insufficient classification accuracy and robustness, failing to meet the real-time, reliability, and engineering deployment requirements of online monitoring of power equipment. Summary of the Invention

[0004] This application provides a partial discharge identification method, apparatus, device, medium, and program product to improve the classification accuracy and robustness of partial discharge identification, thereby meeting the requirements of online monitoring of power equipment for real-time performance, reliability, and engineering deployment.

[0005] In a first aspect, embodiments of this application provide a partial discharge identification method, including:

[0006] Input the phase-resolved partial discharge map or phase-resolved partial discharge signal map into the preset dual-branch identification model;

[0007] The adaptive long short-term memory module in the preset dual-branch identification model extracts features from the input phase-resolved partial discharge map or phase-resolved partial discharge signal map to obtain the first output representation.

[0008] The second output representation is obtained by extracting features from the input phase-resolved partial discharge map or phase-resolved partial discharge signal map through the adaptive linear state-space module in the preset dual-branch identification model.

[0009] The first output representation and the second output representation are fused together to obtain the partial discharge identification result.

[0010] Secondly, embodiments of this application provide a partial discharge identification device, comprising:

[0011] The input module is used to input phase-resolved partial discharge diagrams or phase-resolved partial discharge signal diagrams into a preset dual-branch identification model;

[0012] The first extraction module is used to extract features from the input phase-resolved partial discharge map or phase-resolved partial discharge signal map through the adaptive long short-term memory module in the preset dual-branch identification model to obtain a first output representation;

[0013] The second extraction module is used to extract features from the input phase-resolved partial discharge map or phase-resolved partial discharge signal map through the adaptive linear state space module in the preset dual-branch identification model to obtain a second output representation.

[0014] The fusion module is used to fuse the first output representation and the second output representation to obtain the partial discharge identification result.

[0015] Thirdly, embodiments of this application provide a partial discharge identification device, including: a memory and a processor;

[0016] The memory stores computer-executed instructions;

[0017] The processor executes computer execution instructions stored in the memory, causing the processor to perform the first aspect and / or various possible implementations of the first aspect as described above.

[0018] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the first aspect and / or various possible implementations of the first aspect.

[0019] Fifthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the first aspect and / or various possible implementations of the first aspect.

[0020] The partial discharge identification method, apparatus, device, medium, and program products provided in this application embodiment input phase-resolved partial discharge map or phase-resolved partial discharge signal map into a preset dual-branch identification model. An adaptive long short-term memory module is used to extract local fine-grained features to capture the dynamic changes and local details of the discharge. Simultaneously, an adaptive linear state-space module is used to extract global features to model long-range phase dependencies and integrate multi-directional information. Finally, the output representations of the two modules are fused to obtain more comprehensive features, thereby achieving high-precision and robust partial discharge identification. This effectively solves the problems of insufficient classification accuracy and poor robustness in existing technologies under complex noise and cross-device scenarios, meeting the requirements of online monitoring of power equipment for real-time performance, reliability, and engineering deployment. Attached Figure Description

[0021] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0022] Figure 1 Flowchart of the partial discharge identification method provided in this application Figure 1 ;

[0023] Figure 2 Flowchart of the partial discharge identification method provided in this application Figure 2 ;

[0024] Figure 3 This is a schematic diagram illustrating the process of training the dual-branch recognition model in this application;

[0025] Figure 4 A schematic diagram illustrating the processing steps of the partial discharge identification method provided in this application;

[0026] Figure 5 A schematic diagram of the partial discharge identification device provided in this application;

[0027] Figure 6 A schematic diagram of the partial discharge identification device provided in this application.

[0028] The accompanying drawings have illustrated specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to specific embodiments. Detailed Implementation

[0029] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0030] In power systems, especially in critical scenarios such as densely populated urban areas, aerospace power links, and forest power transmission and distribution networks, the insulation health of high-voltage equipment directly affects the reliability of continuous power supply. Oil-immersed transformers and shunt reactors, among other equipment, are subjected to high voltage and thermal stress over long periods, making their insulation defects susceptible to gradual evolution into breakdown accidents via partial discharge (PD). Partial discharge refers to the microscopic discharge phenomenon occurring in localized areas within or on the surface of insulating materials, typically accompanied by charge migration and electromagnetic, acoustic, optical, and thermal effects. If not identified and intervened in a timely manner, partial discharge can lead to erosion of the insulating dielectric structure, localized temperature rise, carbonization, and even fire, ultimately causing power outages and equipment damage. Therefore, achieving online, rapid, and accurate identification of partial discharge types and assessment of their development is a core aspect of power equipment condition assessment and preventative maintenance.

[0031] The development of partial discharge monitoring technology began with the current pulse method, which acquires pulse waveforms through coupling capacitors or current sensors and identifies them based on parameter criteria. However, the robustness of this method is limited in environments with strong noise and complex grounding. Subsequently, the ultra-high frequency (UHF) method emerged. By acquiring electromagnetic radiation signals on the order of hundreds of MHz, the UHF method can naturally suppress power frequency and some external interference, facilitating non-intrusive deployment. Nevertheless, the UHF signal in the field may still be mixed with interference from switching actions, wireless communication, etc., making direct classification based on the original waveform susceptible to influence. To improve reliability, adaptive decomposition and denoising techniques were introduced, reconstructing the one-dimensional discharge signal in the time and frequency domain before extracting features for classification. Although this method improves accuracy, the computational cost of decomposing and reconstructing each signal is high, making it difficult to meet the real-time requirements of large-scale online monitoring.

[0032] To balance efficiency and accuracy, the industry maps time-domain signals to phase-resolved maps, including Phase-Resolved Partial Discharge Maps (PRPDs) and Phase-Resolved Partial Discharge Signal Maps (PRPSs). PRPDs describe the phase location of discharge occurrences using power frequency phase and characterize discharge density or amplitude distribution using pixel brightness or color, revealing information such as the location of discharge, the density of discharge activity, and the intensity of the discharge. PRPSs, on the other hand, focus more on describing the signal characteristics of partial discharges. By mapping one-dimensional partial discharge signals (such as UHF signals) onto a power frequency phase map, they demonstrate the relationship between the phase and amplitude of the discharge signal, effectively distinguishing partial discharges from other noise and interference, helping engineers more accurately assess the insulation health of equipment in complex power grid environments.

[0033] Under the premise of using PRPD / PRPS as a unified representation, the dominant paradigm of partial discharge identification has shifted from the traditional "parameter-feature engineering" to a deep learning-based computer vision paradigm. This new paradigm is image-centric, using end-to-end trained models to automatically learn discriminative representations directly from PRPD / PRPS, replacing tedious manual rules and signal-by-signal reconstruction. This method aims to minimize front-end modifications and preprocessing. Through unified image normalization and data augmentation, it drives a multi-scale feature backbone to preserve discharge texture and details at different spatial scales, and utilizes attention / saliency mechanisms to highlight key "phase-amplitude" regions and suppress invalid background. When periodic relationships need to be represented, a phase-aware sequence modeling module supplements long-range dependencies. The discrimination end combines uncertainty and confidence assessment to achieve stable output of discharge type and evolution degree. In terms of training and deployment, it combines transfer / incremental learning, style normalization, and lightweight inference to reduce reliance on front-end cleaning and manual features, balancing cross-device / cross-scenario robustness and online real-time performance, thus better meeting the engineering needs of condition assessment and preventative maintenance.

[0034] This application proposes an end-to-end recognition framework for PRPD / PRPS images. It employs a parallel collaborative approach using a "temporal channel with multi-scale convolution and long short-term memory modules" and an "adaptive channel based on the Mamba model," systematically addressing key shortcomings of existing technologies in complex field and cross-device applications. First, addressing the dynamic range and statistical distribution drift caused by strong noise, differences in acquisition links, and variations in operating conditions, this application significantly improves robustness across devices and scenarios through adaptive state-space modeling and multi-scale local statistical alignment, eliminating the need for tedious data cleaning and frequent parameter tuning. Second, addressing the core contradiction of "difficulty in reconciling global phase dependence with local details," this application explicitly captures long-range temporal dependencies of half-cycles / full cycles on the phase axis using a long short-term memory network. Simultaneously, it preserves fine-grained information such as speckles and narrowband textures through pyramid-style cross-scale fusion in space, and establishes long-distance inter-block connections using selective scanning of the Mamba model, achieving a balance between global consistency and local fidelity. Furthermore, addressing the issues of insufficient multi-scale representation and detail dilution caused by downsampling, this application simultaneously performs multi-scale feature extraction and cross-layer fusion within dual channels. This ensures that shallow textures and deep semantics are aligned at a unified scale, preventing weak discharges and small targets from being masked during resampling. Further, to meet the constraints of real-time performance and cost for engineering deployment, this application replaces heavy self-attention with near-linear state-space computation, coupled with lightweight convolution and parameter-sharing design, effectively reducing inference latency and computational cost, adapting to the online deployment requirements of station-level and edge devices. Finally, regarding output reliability, this application simultaneously provides discharge type and severity estimates in the discriminant head, along with calibrable confidence levels, providing strong support for alarm classification and preventative maintenance decisions, overcoming the shortcomings of evaluating solely based on accuracy.

[0035] Specifically, addressing the issue of insufficient attention to local image regions in Transformer-based methods (caused by patching, windowed self-attention, and general positional encoding), this application introduces an adaptive module based on the Mamba model, providing stronger local inductive bias and continuous selective scanning capabilities. On one hand, without relying on large windows and global attention, it enhances the response to fine-grained structures such as sparse specks, narrow stripes, and weak discharges; on the other hand, while maintaining near-linear complexity, it preserves global context modeling, mechanistically alleviating the Transformer's weakening of local details and the fragmentation of cross-window phase rhythms, achieving the coordinated expression of "global-local-phase" information.

[0036] In summary, this application aims to achieve robust, fine-grained, and low-latency online detection and classification of PRPD / PRPS images without increasing heavy preprocessing and on-site modifications; effectively resist noise and cross-device distribution drift; uniformly model long-range phase dependencies and local details; and provide usable outputs of type, severity, and confidence with high accuracy to meet the engineering needs of practical operation and maintenance scenarios. These objectives are achieved through a dual-channel structure (multi-scale long short-term memory channels and an adaptive channel based on the Mamba model) and its cross-scale fusion and end-to-end training mechanisms.

[0037] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0038] Figure 1 Flowchart of the partial discharge identification method provided in this application Figure 1 ,like Figure 1 As shown, the method includes:

[0039] S101. Input the phase-resolved partial discharge diagram or phase-resolved partial discharge signal diagram into the preset dual-branch identification model.

[0040] In this step, the preset dual-branch recognition model is a deep learning model for partial discharge recognition, which includes two parallel feature extraction modules: an adaptive long short-term memory module and an adaptive linear state space module. It can extract different types of features from the input PRPD / PRPS image, providing a foundation for subsequent fusion processing and recognition.

[0041] S102. The first output representation is obtained by extracting features from the input phase-resolved partial discharge map or phase-resolved partial discharge signal map through the adaptive long short-term memory module in the preset dual-branch recognition model.

[0042] Adaptive Long Short-Term Memory Module (ALSTM) is an improved Long Short-Term Memory (LSTM) network specifically designed for processing partial discharge image data. It extracts local features from the input image through multi-scale convolutional operations, and then uses long short-term memory mechanisms such as gated recurrent units (GRUs) to perform temporal modeling of these features along the phase axis. This effectively captures the long-range phase dependence and dynamic changes of partial discharges, thus obtaining a time-series-related feature representation.

[0043] The first output representation is the feature representation extracted by the adaptive long short-term memory module. It contains the temporal features and local details of partial discharge in the input image, providing a temporal description of the partial discharge features for subsequent fusion processing.

[0044] S103. The input phase-resolved partial discharge map or phase-resolved partial discharge signal map is feature extracted by the adaptive linear state-space module in the preset dual-branch identification model to obtain the second output representation.

[0045] The Adaptive Linear State-Space Module (ALSS), based on State-Space Models (SSMs) and a hybrid scan space module, is capable of global feature extraction from the input image, modeling global phase dependencies and connections between distant blocks. Furthermore, this module employs a selective scan strategy, enhancing the capture of local details while preserving global contextual information.

[0046] The second output representation is the feature representation extracted by the adaptive linear state space module. It contains the global features and spatial dependency information of partial discharge in the input image, providing a spatial description of the partial discharge features for subsequent fusion processing.

[0047] S104. The first output representation and the second output representation are fused to obtain the partial discharge identification result.

[0048] The feature representations extracted by the adaptive long short-term memory module and the adaptive linear state-space module are fused to integrate local and global features. The fusion method can be a simple concatenation followed by linear mapping, or a more complex learnable weighted fusion strategy. Through fusion processing, a more comprehensive and accurate feature representation can be obtained, which includes both the temporal features of partial discharge and the global spatial features, thereby improving the expressive power and recognition performance of partial discharge features.

[0049] Based on the fused feature representation, classification is performed using discriminative models such as a multilayer perceptron (MLP) classifier, outputting the type of partial discharge (e.g., internal discharge, surface discharge, corona discharge, etc.) and its confidence level. Furthermore, it can provide information such as discharge severity estimation, offering strong support for the condition assessment and preventative maintenance of power equipment.

[0050] The partial discharge identification method provided in this application involves inputting a phase-resolved partial discharge map or a phase-resolved partial discharge signal map into a preset dual-branch identification model. An adaptive long short-term memory module is used to extract local fine-grained features to capture the dynamic changes and local details of the discharge. Simultaneously, an adaptive linear state-space module is used to extract global features to model long-range phase dependencies and integrate multi-directional information. Finally, the output representations of the two modules are fused to obtain more comprehensive features, thereby achieving high-precision and robust partial discharge identification. This effectively solves the problems of insufficient classification accuracy and poor robustness in existing technologies under complex noise and cross-device scenarios, meeting the requirements of online monitoring of power equipment for real-time performance, reliability, and engineering deployment.

[0051] Figure 2 Flowchart of the partial discharge identification method provided in this application Figure 2 ,like Figure 2 As shown, in this embodiment... Figure 1 Based on the embodiments, the partial discharge identification method is described in detail, which includes:

[0052] S201. Input the phase-resolved partial discharge diagram or phase-resolved partial discharge signal diagram into the preset dual-branch identification model.

[0053] Reference Figure 3 The diagram illustrates the process of training the dual-branch recognition model in this application. First, sample images of power equipment PRPD / PRPS are collected, reflecting partial discharge activity during equipment operation. The acquired dataset is divided into different subsets: a training set, a validation set, and a test set. This division facilitates model training, parameter tuning, validation, and final testing, ensuring the model's generalization ability. The training set data is preprocessed, and then used to train the dual-branch recognition model. The model learns features and patterns from the data, progressively optimizing its parameters to improve the accuracy of partial discharge recognition. During model training, validation set data is used to evaluate the model's performance. This step helps monitor the model's learning progress, prevents overfitting, and provides a basis for parameter tuning. Based on the model's performance on the validation set, the model structure or hyperparameters are adjusted to further improve model performance. Once the model training is complete and performs well on the validation set, preprocessed test set data is input into the model for final performance evaluation. This process allows for the systematic training and evaluation of the dual-branch recognition model, ensuring its effectiveness and reliability in practical applications.

[0054] To improve the model's discriminative ability, focus loss is used for optimization, and the objective function is expressed as follows:

[0055] ;

[0056] The loss function consists of two parts: classification loss and segmentation loss. Specifically, S represents the probability that the model predicts the image belongs to each category; c is the true category label of the input image.

[0057] S202. The first output representation is obtained by extracting features from the input phase-resolved partial discharge map or phase-resolved partial discharge signal map through the adaptive long short-term memory module in the preset dual-branch recognition model.

[0058] In one possible implementation, the first output representation is obtained by extracting features from the input phase-resolved partial discharge map or phase-resolved partial discharge signal map through the adaptive long short-term memory module in the preset dual-branch identification model. Specifically, this may include the following steps:

[0059] In the adaptive long short-term memory module, feature extraction is performed on the input phase-resolved partial discharge map or phase-resolved partial discharge signal map to obtain basic features;

[0060] Multi-scale feature extraction is performed on the extracted basic features to obtain the feature tensor;

[0061] The feature tensor is expanded into a sequence column-wise, and the sequence is processed sequentially through two layers of gated loop units;

[0062] The sequence processed by two layers of gated recurrent units is mapped through a linear layer to obtain the first output representation.

[0063] In this embodiment, the adaptive long short-term memory module first extracts features from the input PRPD or PRPS image to obtain basic features. Multi-scale feature extraction is then performed on the extracted basic features to obtain a feature tensor. Multi-scale feature extraction helps capture features of the image at different scales, which is crucial for understanding complex structures and patterns in the image. The feature tensor is expanded column-wise into a sequence, and the sequence is processed sequentially through two layers of gated recurrent units (GRUs). The purpose of this step is to utilize the gated recurrent units to process temporal data and capture the temporal dependencies between features, which is critical for understanding the dynamic changes in partial discharge activity. The sequence processed by the two layers of gated recurrent units is mapped through a linear layer to obtain the first output representation.

[0064] In practical implementation, the input PRPD or PRPS image can be represented as Here, H, W, and C represent the image height, width, and number of channels, respectively. First, the input image is convolved using a 3×3 convolutional block. Next, a Gaussian Error Linear (GELU) activation function is applied to the output of the convolutional processing for a non-linear transformation. Then, another 3×3 convolutional block is used to convolve the output of the non-linear transformation. After that, the Gaussian Error Linear (GELU) activation function is applied again to the output of the convolutional processing for a non-linear transformation. Finally, the output of this non-linear transformation is pooled to obtain the basic features. .

[0065] Next, basic features The data is fed into three parallel convolutional branches of different sizes (3×3, 5×5, and 7×7), each branch corresponding to a different output channel (respectively). , , These convolutional blocks of different sizes can capture features of different sizes, thus achieving multi-scale feature extraction. The extracted features are concatenated along the channel dimension to form a feature tensor. .

[0066] To align the channel dimensions and compress the feature tensor, a 1×1 convolutional block is used to convolve the feature tensor to obtain the aligned feature tensor. .

[0067] To characterize the difference between column discharge intensity and structure, the aligned feature tensors are... Expand into a sequence by column. each , This represents an average or weighted pooling operation performed along the height direction. This step converts the two-dimensional features into a one-dimensional sequence, preparing for subsequent processing of timing information by the gated recurrent unit. The sequence is processed sequentially through two layers of gated recurrent units to achieve channel timing recalibration. The second layer of gated recurrent units uses the hidden states of the first layer as conditions to assign greater gating weights to high-response channels, thereby better capturing the timing features of discharge activity. Finally, the sequence processed by the two layers of gated recurrent units is dimension-mapped through a linear layer to obtain the first output representation. .

[0068] S203. The second output representation is obtained by extracting features from the input phase-resolved partial discharge map or phase-resolved partial discharge signal map through the adaptive linear state-space module in the preset dual-branch identification model.

[0069] In one possible implementation, the second output representation is obtained by extracting features from the input phase-resolved partial discharge map or phase-resolved partial discharge signal map using an adaptive linear state-space module in a pre-defined dual-branch identification model. Specifically, this may include the following steps:

[0070] In the adaptive linear state-space module

[0071] Feature extraction is performed on the input phase-resolved partial discharge map or phase-resolved partial discharge signal map to obtain preliminary features;

[0072] The extracted preliminary features are input into the hybrid scan space module in the adaptive linear state space module to extract global information features, and the extracted preliminary features are input into the convolution module in the adaptive linear state space module to extract local information features.

[0073] The extracted global and local information features are fused to obtain fused features.

[0074] Perform convolution processing on the fused features;

[0075] The fused features after convolution are mapped through a linear layer to obtain a second output representation.

[0076] In this embodiment, in the adaptive linear state space module, feature extraction is first performed on the input PRPD or PRPS image to obtain preliminary features. The extracted preliminary features are then input into the hybrid scan space module and the convolution module to extract global and local information features. The extracted global and local information features are fused to obtain fused features. The fused features are then convolved. Finally, the convolved fused features are mapped through a linear layer to obtain a second output representation.

[0077] Specifically, two convolutional layers and activation functions can be applied to the input image to obtain preliminary feature representations, and then these preliminary features can be processed... The data is fed into a hybrid scanning space module, and after being stacked through n consecutive hybrid scanning space modules, enhanced global information features are obtained. :

[0078] .

[0079] On the other hand, preliminary features can be fed into a convolutional module, and processed through 1×1 convolutional blocks. Convolution processing is performed to obtain local information features. :

[0080] .

[0081] Global information features and local information features The features are concatenated along the channel dimension to obtain the fused features. These fused features are then convolved using 3×3 convolutional blocks. Finally, the convolutionally processed fused features are mapped through a linear layer to obtain the second output representation. :

[0082] .

[0083] In one possible implementation, the extracted preliminary features are input into the hybrid scanning space module to extract global information features, which may specifically include the following steps:

[0084] The extracted preliminary features are input into multiple hybrid scanning space modules. Each hybrid scanning space module includes a layer normalization layer, a linear layer, a convolutional layer, a hybrid scanning encoder, a state space model, and a hybrid scanning decoder connected in sequence.

[0085] In the first hybrid scanning space module, the initial features are processed by layer normalization layer, the normalized features are transformed by linear layer, the transformed features are extracted by convolutional layer, the extracted features are encoded by hybrid scanning encoder, the encoded features are modeled by state space model, and the modeled features are decoded by hybrid scanning decoder to obtain enhanced feature representation.

[0086] The enhanced feature representation is used as input and fed into the next hybrid scan space module;

[0087] After being processed by stacking multiple hybrid scanning spatial modules, global information features are extracted.

[0088] In this embodiment, the preliminary features extracted from the input image are input into multiple hybrid scanning space modules. Each hybrid scanning space module includes a layer normalization layer, a linear layer, a convolutional layer, a hybrid scanning encoder, a state space model, and a hybrid scanning decoder. After being processed by multiple modules in a step-by-step stacking manner, global information features are extracted. This step-by-step stacking processing method enables the model to gradually extract increasingly rich global information from the preliminary features, ultimately obtaining a feature representation that can comprehensively represent the global information of the input image.

[0089] The Hybrid Scan Space Module, drawing inspiration from the Mamba model, is specifically designed for processing image data to execute hybrid strategies and multi-directional scanning and fusion. The module aims to enhance the model's ability to capture local and global features in images, particularly when processing Phase-Resolved Partial Discharge Maps (PRPDs) or Phase-Resolved Partial Discharge Signal Maps (PRPSs). It comprises sequentially connected layer normalization layers, linear layers, convolutional layers, a hybrid scan encoder, a state-space model, and a hybrid scan decoder, which can be represented by the following formula:

[0090] ;

[0091] ;

[0092] The first formula represents the encoding process, and the second formula represents the decoding process. Using the input features, an enhanced feature representation is obtained through processing by a layer normalization layer, a linear layer, an activation function, a hybrid scanning encoder, a state-space model, and a hybrid scanning decoder. Further processing using linear layers and activation functions. and with the original input By combining these features, we obtain the updated feature representation. .

[0093] A state-space model is a mathematical model derived from control theory. It is used to describe the time-varying behavior of dynamic systems and can be represented as:

[0094] , ;

[0095] The derivative of the state represents the rate of change of the state over time; The state transition matrix describes how the internal state of the system evolves over time. The input matrix describes the input... How to affect the state ; The output matrix describes the state. How to map to the output .

[0096] With a sampling step size of The parameters can be discretized, and the discretized input matrix can be represented as:

[0097] ;

[0098] For matrix exponents, It is the identity matrix, used to map the parameters of a continuous model to a discrete model.

[0099] The discretized state-space model can be represented as:

[0100] , ;

[0101] The state at time t; This refers to the state at the previous time step. It is the output at time t.

[0102] The discrete model can be further represented in convolutional form:

[0103] , ;

[0104] It is a structured convolutional kernel, where L is the length of the input sequence x; This represents the convolution operation, where y is the result of the convolution of the input sequence x with the convolution kernel K, i.e., the output of the model.

[0105] In one possible implementation, the extracted features are encoded using a hybrid scan encoder, which may specifically include the following steps:

[0106] In the hybrid scan encoder, the extracted features are encoded according to the selected Hilbert scan strategy.

[0107] Hilbert scan is a scanning method based on Hilbert curves. Hilbert curves map a multidimensional space to a one-dimensional space through a series of continuous polygonal lines (or curves) while preserving proximity relationships in the space. This makes it stable in encoding local and global relationships, and it is particularly suitable for mitigating efficiency and information loss problems in long-range dependency modeling.

[0108] Hilbert curves can be recursively generated from an n-order Hilbert matrix:

[0109]

[0110] in, ; It is an n-order matrix of all 1s; Represents an n-order Hilbert matrix transpose; Represents an n-order Hilbert matrix Flip left and right; Represents an n-order Hilbert matrix Flip it up and down.

[0111] In this embodiment, the selected Hilbert scanning strategy may include: row forward scanning, row reverse scanning, column forward scanning, and column reverse scanning. The selection of these four scanning directions aims to comprehensively cover all regions of the image, ensuring that no important information is missed during feature extraction.

[0112] In one possible implementation, the features after state modeling are decoded by a hybrid scan decoder to obtain an enhanced feature representation, which may specifically include the following steps:

[0113] In the hybrid scan decoder, the features after state modeling are decoded and restored;

[0114] The enhanced feature representation is obtained by summing the feature components in different directions after decoding and restoration.

[0115] In this embodiment, the hybrid scan encoder selects the Hilbert scan strategy to encode the input features, which are then fed into the spatial state model to enhance the global representation capability. The hybrid scan decoder restores the output features of the spatial state model to the same direction as the original input, and sums the feature components in each direction to obtain the final output.

[0116] S204. Merge the first output representation and the second output representation.

[0117] The first output refers to the phase-temporal correlation feature vector obtained by the adaptive long short-term memory module after performing hierarchical feature extraction on the input phase-resolved partial discharge map or phase-resolved partial discharge signal map: the input image is subjected to double convolutional encoding to extract the edge gradient features, texture distribution features and local aggregation pattern features of the discharge spots, and the basic features are formed after pooling compression; multi-scale feature extraction is performed on the basic features to obtain the feature tensor, and then it is expanded into a pseudo-temporal sequence along the phase dimension (column direction). The sequence is temporally modeled through two layers of gated recurrent units to capture the short-range evolution dependency of discharge intensity in adjacent phase intervals, and mapped into a fixed-dimensional feature vector through a linear layer.

[0118] The second output represents the global-local fusion feature vector obtained by the adaptive linear state-space module after multi-scale scanning feature extraction of the input phase-resolved partial discharge map or phase-resolved partial discharge signal map: Initial convolutional encoding is performed on the input image to extract primary features of the discharge distribution. These primary features are then input in parallel into the hybrid scanning space module and the convolution module. The hybrid scanning space module uses a four-way scanning strategy of Hilbert row scanning (along the phase dimension) and column scanning (along the amplitude dimension) to map the two-dimensional features into a one-dimensional sequence. After modeling with the state-space model, it captures the long-range statistical correlation (global information features) across phase intervals. The convolution module extracts the channel correlation and point detail features (local information features) of the primary features using a 1×1 convolution kernel. The global and local information features are concatenated and fused along the channel dimension, and after convolutional compression and linear mapping, a unified feature vector is formed. The first output provides temporal resolution capability for identifying the instantaneous change patterns and periodic fluctuation characteristics of discharge activity, while the second output provides spatial resolution capability for identifying the spatial distribution morphology and cross-period stability characteristics of discharge activity.

[0119] In this step, the "first output representation" obtained from the adaptive long short-term memory module and the "second output representation" obtained from the adaptive linear state space module are fused. This fusion can be achieved in various ways, such as simple concatenation, weighted summation, or other more complex fusion strategies.

[0120] S205. Perform layer normalization on the fused features.

[0121] S206. The features after layer normalization are linearly mapped to obtain the probability distribution of various discharge types, and the probability distribution is used as the result of partial discharge identification.

[0122] The fused features are subjected to layer normalization, and the normalized features are linearly mapped to obtain the probability distribution of various discharge types. The probability distribution is then used as the result of partial discharge identification.

[0123] In practice, the fused features can be fed into a multilayer perceptron classifier. In the multilayer perceptron classifier, the fused features are first normalized through layer normalization. The normalized features are then transformed linearly through a linear layer. After the linear transformation, the features are nonlinearly mapped through a Gaussian error activation function. After the nonlinear mapping, the features are output as log probabilities for each category through a linear layer. The log probabilities are then converted into a probability distribution of the discharge category through a normalized exponential function.

[0124] Reference Figure 4The diagram shown illustrates the processing flow of the partial discharge identification method provided in this application. This application proposes a dual-branch identification model for PRPD / PRPS images, consisting of an adaptive long short-term memory module and an adaptive linear state-space module, both of which are essential components. The adaptive long short-term memory branch extracts basic features by sequentially processing the input image through 3×3 convolution → Gaussian error linear activation function → 3×3 convolution → Gaussian error linear function → max pooling. These basic features are then fed into three parallel convolution branches of different sizes (3×3, 5×5, 7×7) for multi-scale feature extraction. Subsequently, channel alignment is performed through 1×1 convolution, and channel temporal recalibration is achieved through two layers of gated recurrent units before outputting the representation. The adaptive linear state-space branch employs stackable hybrid scan space modules for feature extraction from the input. Within these modules, feature extraction is achieved through layer normalization → linear layers → 3×3 depthwise separable convolutions → hybrid scan encoder → state-space model → hybrid scan decoder → linear layers → layer normalization. This is combined with internal residuals and 1×1 convolutional branches for joint modeling, resulting in the output representation. The two branches represent the fusion at the decision-making end, and the input to the multilayer perceptron classifier generates the final category prediction; the dashed boxes in the figure are used to supplement the detailed internal structure of the two branches.

[0125] This application proposes an innovative dual-channel network structure specifically optimized for PRPD and PRPS images. The core feature of this network structure lies in the parallel integration of two efficient feature extraction channels: a "multi-scale convolution + long short-term memory module" channel and a "Mamba-based adaptive vision" channel. The features extracted by these two channels are integrated through a specific cross-scale and cross-channel fusion strategy.

[0126] In addition, this application introduces a mechanism for explicit sequence modeling on the phase axis, which is combined with a multi-scale convolutional pyramid to simultaneously maintain high-resolution capture of local texture details and effectively capture long-range dependencies of half-cycles or whole cycles.

[0127] To enhance the response to sparse speckles and narrow textures in PRPD / PRPS images, this application designs an adaptive vision module based on Mamba. This module utilizes selective scanning and state-space equations to model long-distance dependencies on a two-dimensional feature map, and further improves the recognition ability of these subtle features through local enhancement units.

[0128] Regarding the fusion strategy and loss design, this application proposes a comprehensive approach, including weighted or gated fusion at the feature layer, a joint task head (combining type classification and severity regression) at the decision layer, and a combination of confidence calibration and robust regularization. This design not only improves the model's accuracy in identifying discharge types but also enhances its ability to estimate discharge severity and provides more reliable decision support through confidence calibration.

[0129] This application significantly improves the robustness and transferability of the model by employing multi-scale feature alignment and adaptive modeling techniques. This method ensures that weak discharge signals and sparse details are stably preserved even during multiple downsampling and resampling processes, thereby greatly reducing performance fluctuations caused by differences in equipment, sensor links, and sampling conditions. Simultaneously, the end-to-end learning strategy of this application is performed directly in the image domain, avoiding reliance on large-scale one-dimensional signal decomposition and denoising processes, reducing sensitivity to parameter adjustments and dependence on human experience, and effectively lowering engineering maintenance costs.

[0130] In terms of computational efficiency and real-time performance, this application employs a near-linear state-space computation method to replace the computationally intensive self-attention mechanism. Combined with lightweight convolution and parameter-sharing design, this enables the model to achieve low-latency inference on edge GPUs or terminal devices, meeting the high throughput and power consumption constraints of multi-point concurrent monitoring. Compared to Transformer-based solutions, the Mamba adaptive module in this invention provides the model with stronger local inductive bias and continuous selective scanning capabilities. This significantly alleviates the problems of insufficient attention to local regions and dilution of detail information caused by patching or windowing processing without sacrificing global modeling capabilities.

[0131] Regarding the usability of the results, this application simultaneously outputs the discharge type and severity in the classification header and provides calibrable confidence / uncertainty indices. These indices can be used for alarm classification and maintenance prioritization, thereby compensating for the shortcomings of evaluating solely based on overall accuracy and improving the reliability and interpretability of on-site decisions. In summary, this application, through a mechanism of "dual-channel collaborative operation + multi-scale feature fusion + adaptive state-space modeling," balances long-range phase dependence, local detail fidelity, cross-domain robustness, and online processing efficiency, making it more suitable for the actual deployment needs of industrial power equipment.

[0132] Figure 5 A schematic diagram of the partial discharge identification device provided in this application is shown below. Figure 5 As shown, the partial discharge identification device 50 provided in this embodiment includes:

[0133] Input module 501 is used to input phase-resolved partial discharge diagram or phase-resolved partial discharge signal diagram into a preset dual-branch identification model;

[0134] The first extraction module 502 is used to extract features from the input phase-resolved partial discharge map or phase-resolved partial discharge signal map through the adaptive long short-term memory module in the preset dual-branch identification model to obtain the first output representation;

[0135] The second extraction module 503 is used to extract features from the input phase-resolved partial discharge map or phase-resolved partial discharge signal map through the adaptive linear state space module in the preset dual-branch identification model to obtain the second output representation.

[0136] The fusion module 504 is used to fuse the first output representation and the second output representation to obtain the partial discharge identification result.

[0137] In one possible implementation, the first extraction module is specifically used for:

[0138] In the adaptive long short-term memory module, feature extraction is performed on the input phase-resolved partial discharge map or phase-resolved partial discharge signal map to obtain basic features;

[0139] Multi-scale feature extraction is performed on the extracted basic features to obtain the feature tensor;

[0140] The feature tensor is expanded into a sequence column-wise, and the sequence is processed sequentially through two layers of gated loop units;

[0141] The sequence processed by two layers of gated recurrent units is mapped through a linear layer to obtain the first output representation.

[0142] In one possible implementation, the second extraction module is specifically used for:

[0143] In the adaptive linear state-space module

[0144] Feature extraction is performed on the input phase-resolved partial discharge map or phase-resolved partial discharge signal map to obtain preliminary features;

[0145] The extracted preliminary features are input into the hybrid scan space module in the adaptive linear state space module to extract global information features, and the extracted preliminary features are input into the convolution module in the adaptive linear state space module to extract local information features.

[0146] The extracted global and local information features are fused to obtain fused features.

[0147] Perform convolution processing on the fused features;

[0148] The fused features after convolution are mapped through a linear layer to obtain a second output representation.

[0149] In one possible implementation, the second extraction module is specifically used for:

[0150] The extracted preliminary features are input into multiple hybrid scanning space modules. Each hybrid scanning space module includes a layer normalization layer, a linear layer, a convolutional layer, a hybrid scanning encoder, a state space model, and a hybrid scanning decoder connected in sequence.

[0151] In the first hybrid scanning space module, the initial features are processed by layer normalization layer, the normalized features are transformed by linear layer, the transformed features are extracted by convolutional layer, the extracted features are encoded by hybrid scanning encoder, the encoded features are modeled by state space model, and the modeled features are decoded by hybrid scanning decoder to obtain enhanced feature representation.

[0152] The enhanced feature representation is used as input and fed into the next hybrid scan space module;

[0153] After being processed by stacking multiple hybrid scanning spatial modules, global information features are extracted.

[0154] In one possible implementation, the second extraction module is specifically used for:

[0155] In the hybrid scan encoder, the extracted features are encoded according to the selected Hilbert scan strategy.

[0156] In one possible implementation, the second extraction module is specifically used for:

[0157] In the hybrid scan decoder, the features after state modeling are decoded and restored;

[0158] The enhanced feature representation is obtained by summing the feature components in different directions after decoding and restoration.

[0159] In one possible implementation, the fusion module is specifically used for:

[0160] The first output representation and the second output representation are merged;

[0161] The features obtained by fusion are then subjected to layer normalization.

[0162] The features after layer normalization are linearly mapped to obtain the probability distributions of various discharge types, and the probability distributions are used as the results of partial discharge identification.

[0163] The partial discharge identification device provided in this embodiment can execute the method provided in the above method embodiment. Its implementation principle and technical effect are similar, and will not be described in detail here.

[0164] Figure 6 This is a schematic diagram of the partial discharge identification device provided in this application. Figure 6 As shown, the partial discharge identification device 60 provided in this embodiment includes at least one processor 601 and a memory 602. Optionally, the device 60 further includes a communication component 603. The processor 601, memory 602, and communication component 603 are connected via a bus.

[0165] In a specific implementation, at least one processor 601 executes computer execution instructions stored in memory 602, causing at least one processor 601 to perform the above-described method.

[0166] The specific implementation process of processor 601 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.

[0167] In the above embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.

[0168] The memory may include random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage device.

[0169] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0170] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0171] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the above-described method.

[0172] The aforementioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0173] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components in the device.

[0174] The division of units is merely a logical functional division; in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0175] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0176] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0177] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0178] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0179] Finally, it should be noted that other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein, and is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. A method for identifying partial discharge, characterized in that, include: Input the phase-resolved partial discharge map or phase-resolved partial discharge signal map into the preset dual-branch identification model; The adaptive long short-term memory module in the preset dual-branch identification model extracts features from the input phase-resolved partial discharge map or phase-resolved partial discharge signal map to obtain the first output representation. The second output representation is obtained by extracting features from the input phase-resolved partial discharge map or phase-resolved partial discharge signal map through the adaptive linear state-space module in the preset dual-branch identification model. The first output representation and the second output representation are fused together to obtain the partial discharge identification result.

2. The method according to claim 1, characterized in that, The step of extracting features from the input phase-resolved partial discharge map or phase-resolved partial discharge signal map using the adaptive long short-term memory module in the preset dual-branch identification model to obtain a first output representation includes: In the adaptive long short-term memory module, feature extraction is performed on the input phase-resolved partial discharge map or phase-resolved partial discharge signal map to obtain basic features; Multi-scale feature extraction is performed on the extracted basic features to obtain a feature tensor; The feature tensor is expanded into a sequence column-wise, and the sequence is processed sequentially through two layers of gated loop units; The sequence processed by two layers of gated recurrent units is mapped through a linear layer to obtain the first output representation.

3. The method according to claim 1, characterized in that, The step of extracting features from the input phase-resolved partial discharge map or phase-resolved partial discharge signal map using the adaptive linear state-space module in the preset dual-branch identification model to obtain a second output representation includes: In the adaptive linear state-space module, Feature extraction is performed on the input phase-resolved partial discharge map or phase-resolved partial discharge signal map to obtain preliminary features; The extracted preliminary features are input into the hybrid scan space module in the adaptive linear state space module to extract global information features, and the extracted preliminary features are input into the convolution module in the adaptive linear state space module to extract local information features. The extracted global information features and local information features are fused to obtain fused features; The fused features are then subjected to convolution processing; The fused features after convolution are mapped through a linear layer to obtain the second output representation.

4. The method according to claim 3, characterized in that, The step of inputting the extracted preliminary features into the hybrid scanning space module to extract global information features includes: The extracted preliminary features are input into multiple hybrid scanning space modules. Each hybrid scanning space module includes a layer normalization layer, a linear layer, a convolutional layer, a hybrid scanning encoder, a state space model, and a hybrid scanning decoder connected in sequence. In the first hybrid scanning space module, the preliminary features are processed by a layer normalization layer, the normalized features are then transformed by a linear layer, the transformed features are extracted by a convolutional layer, the extracted features are encoded by the hybrid scanning encoder, the encoded features are modeled by the state space model, and the modeled features are decoded by the hybrid scanning decoder to obtain an enhanced feature representation. The enhanced feature representation is then input into the next hybrid scan space module; After being processed by stacking multiple hybrid scanning spatial modules, the global information features are extracted.

5. The method according to claim 4, characterized in that, The features extracted are encoded by the hybrid scanning encoder, including: In the hybrid scan encoder, the extracted features are encoded according to the selected Hilbert scan strategy.

6. The method according to claim 5, characterized in that, The features after state modeling are decoded by the hybrid scan decoder to obtain an enhanced feature representation, including: In the hybrid scan decoder, the features after state modeling are decoded and restored; The enhanced feature representation is obtained by summing the feature components in different directions after decoding and restoration.

7. The method according to claim 1, characterized in that, The step of fusing the first output representation and the second output representation to obtain the partial discharge identification result includes: The first output representation and the second output representation are merged; The features obtained by fusion are then subjected to layer normalization. The features after layer normalization are linearly mapped to obtain the probability distributions of various discharge types, and the probability distributions are used as the partial discharge identification results.

8. A partial discharge identification device, characterized in that, include: The input module is used to input phase-resolved partial discharge diagrams or phase-resolved partial discharge signal diagrams into a preset dual-branch identification model; The first extraction module is used to extract features from the input phase-resolved partial discharge map or phase-resolved partial discharge signal map through the adaptive long short-term memory module in the preset dual-branch identification model to obtain a first output representation; The second extraction module is used to extract features from the input phase-resolved partial discharge map or phase-resolved partial discharge signal map through the adaptive linear state space module in the preset dual-branch identification model to obtain a second output representation. The fusion module is used to fuse the first output representation and the second output representation to obtain the partial discharge identification result.

9. A partial discharge identification device, characterized in that, include: Memory, processor; The memory stores instructions that the computer executes; The processor executes computer execution instructions stored in memory, causing the processor to perform the method as claimed in any one of claims 1-7.

10. A computer-readable storage medium or computer program product, characterized in that, A computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as claimed in any one of claims 1-7; or, The computer program product includes a computer program that, when executed by a processor, implements the method of any one of claims 1-7.

Images