Method and device for managing partial discharge data of high-voltage cable

Abstract

The invention belongs to the technical field of power equipment state monitoring, and discloses a partial discharge data management method and device for a high-voltage cable. The data management method comprises the following steps: acquiring a multi-modal original defect signal for feature extraction to obtain an initial defect feature; inputting the initial defect feature into a pre-constructed multi-scale noise adaptive separation network for processing to obtain a target noise probability and a target noise feature; when the target noise probability is greater than a preset noise probability threshold, storing the target noise feature into a noise library; otherwise, judging whether the similarity between the discharge defect feature and each preset discharge type cluster in the discharge type library meets a preset similarity requirement, and calculating a target discharge type probability of the discharge defect feature; or a new discharge type cluster is established, so that the intelligent management of the partial discharge data is realized, and the identification accuracy of the partial discharge defect is also improved.

Description

Technical Field

[0001] This invention relates to the field of power equipment condition monitoring technology, and in particular to a method and apparatus for managing partial discharge data of high-voltage cables. Background Technology

[0002] Partial discharge (PD) signals are an early characteristic of insulation defects in high-voltage cables. However, with the development of smart grids, partial discharge data is experiencing a massive increase, and traditional partial discharge data management models exhibit significant limitations when facing this new data format.

[0003] At the data characteristic level: partial discharge signals are easily interfered with by a mixture of various types of noise, such as complex electromagnetic environment and background sound waves. The time and frequency domain characteristics of different noises are significantly different, and they are often highly coupled with the effective signal in the time and frequency domains, making it difficult for traditional filtering methods to achieve accurate separation. At the same time, the spectra of the same type of defect are different under different working conditions, which increases the complexity of sample classification.

[0004] At the data storage and management level: there are a large number of similar patterns in massive data (such as the same defect being monitored repeatedly under the same working conditions). Traditional databases lack targeted redundancy processing mechanisms, resulting in inefficient use of storage resources. As data accumulates, this redundancy will continue to expand, consuming a large amount of storage resources and computing costs.

[0005] At the level of intelligent analysis models: existing defect identification models rely on training with historical samples, but in reality, cable discharge types may exhibit new characteristics as operating time and environment change. Traditional models need to be updated manually on a regular basis, which makes it difficult to adapt to the dynamic monitoring needs, resulting in a gradual decline in identification accuracy.

[0006] Therefore, those skilled in the art are dedicated to developing a high-voltage cable partial discharge data management method and index management method that features adaptive separation of multiple types of noise, flexible storage of massive redundant data, and online iterative evolution of the discharge type identification model. Summary of the Invention

[0007] In view of the above-mentioned deficiencies of the prior art, this application provides a method and apparatus for managing partial discharge data of high-voltage cables, so as to comprehensively improve the quality of partial discharge data, storage efficiency and intelligent analysis accuracy, and provide a solid data foundation and technical support for the lean operation and maintenance of power grid equipment.

[0008] The first aspect of this application provides a method for managing partial discharge data of high-voltage cables, comprising: acquiring multimodal raw defect signals; extracting features from the raw defect signals to obtain initial defect features; inputting the initial defect features into a pre-constructed multi-scale noise adaptive separation network for processing to obtain target noise features; calculating the target noise probability of the target noise features; storing the target noise features in a noise database when the target noise probability is greater than a preset noise probability threshold; otherwise, purifying the initial defect features according to a corresponding filtering strategy to obtain discharge defect features; determining whether the similarity between the discharge defect features and each preset discharge type cluster in the discharge type database meets a preset similarity requirement; when the similarity meets the preset similarity requirement, calculating the target discharge type probability of the discharge defect features; determining the target discharge type cluster of the discharge defect features based on the target discharge type probability and outputting the corresponding target discharge type; otherwise, storing the discharge defect features in a verification buffer; and establishing a new discharge type cluster based on the discharge defect features when the feature data in the verification buffer meets the preset new cluster creation conditions; calculating an index priority score based on the access information of the discharge type database and managing the index hierarchically based on the priority score.

[0009] A second aspect of this application provides a partial discharge data management device for high-voltage cables, characterized in that it includes: an acquisition module configured to acquire multimodal raw defect signals, extract features from the raw defect signals to obtain initial defect features; a noise probability calculation module configured to input the initial defect features into a pre-constructed multi-scale noise adaptive separation network for processing to obtain target noise features, and calculate the target noise probability of the target noise features; and a noise judgment module configured to store the target noise features in a noise database when the target noise probability is greater than a preset noise probability threshold; otherwise, to purify the initial defect features according to a corresponding filtering strategy to obtain discharge defect features; and defect data. The processing module is configured to determine whether the similarity between the discharge defect feature and each preset discharge type cluster in the discharge type library meets the preset similarity requirements. When the similarity meets the preset similarity requirements, the module calculates the target discharge type probability of the discharge defect feature, determines the target discharge type cluster of the discharge defect feature based on the target discharge type probability, and outputs the corresponding target discharge type. Otherwise, the module stores the discharge defect feature in the verification buffer. When the feature data in the verification buffer meets the preset new cluster creation conditions, a new discharge type cluster is established based on the discharge defect feature. The index management module is configured to calculate the index priority score based on the access information of the discharge type library and perform hierarchical management of the index based on the priority score.

[0010] The beneficial effects of this invention are as follows: By constructing a closed-loop data management system covering the entire process, integrating multi-scale adaptive noise separation for accurate noise identification and separation, and accurately identifying the discharge defect type using the effective signal after noise separation, this invention fundamentally solves the problem of severe noise interference leading to errors in identification results in traditional partial discharge data processing. Furthermore, by analyzing the access information of different indexes in real time and adjusting priorities for scheduling and management of different indexes, this invention improves the storage management of different indexes and enhances retrieval efficiency. Attached Figure Description

[0011] Figure 1 This is a schematic diagram of a partial discharge data management process for a high-voltage cable provided in an embodiment of this application;

[0012] Figure 2 This is a schematic diagram of a ResNet-34 network structure provided in an embodiment of this application;

[0013] Figure 3 This is a schematic diagram of a full-process intelligent monitoring system for a data management system provided in an embodiment of this application;

[0014] Figure 4 This is a schematic diagram of the structure of a partial discharge data management device for a high-voltage cable provided in an embodiment of this application. Detailed Implementation

[0015] The present invention will be further described below with reference to the accompanying drawings and embodiments. It should be noted that in the description of the present invention, the terms "first", "second", "third", etc. are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0016] This application provides an intelligent management method for partial discharge data of high-voltage cables. Before managing the partial discharge data, those skilled in the art have also designed a partial discharge data management system. The overall architecture of this system is based on a distributed database and microservice modular design. The system hardware adopts a distributed cluster based on MongoDB, with an initial configuration including 3 primary data nodes and 2 backup nodes. Node configuration can be adjusted according to data volume, concurrent access volume, fault tolerance level, and geographical distribution requirements. For example, regarding data volume, the node size can be dynamically and elastically scaled down based on the data volume: the effective storage capacity of a single node is set to 10TB (including 20% ​​redundancy). When the incremental monitoring data in a single region reaches 5 million records, the system can automatically trigger expansion, adding one primary node and one backup node; no restrictions are imposed here.

[0017] Regarding concurrent access, the system is designed to support a peak concurrent query rate of 500 times / second and a data write rate of 200 records / second. Requests are distributed through a load balancing algorithm to avoid single-point overload. The system employs an N+1 fault-tolerance mechanism, with each master node configured with two data replicas, keeping replica synchronization latency within one second to ensure uninterrupted service in the event of a master node failure. For cross-regional power grid applications, regional node clusters can be deployed according to district / county administrative regions. Data is synchronized between different district / county administrative regions periodically (hourly) via dedicated lines, enabling local queries to prioritize access to regional nodes, effectively reducing network latency. No specific limitations are specified here.

[0018] Furthermore, the database is logically divided into two core sub-databases: a partial discharge (PD) signal database and a noise database.

[0019] The partial discharge signal library is used to store partial discharge defect data of high-voltage cables. The initial design of the database is based on the typical defects of partial discharge in high-voltage cables (i.e., tip discharge, surface discharge, internal air gap discharge, and floating potential discharge), and the data model design includes three main categories of fields: 1. Raw data field: stores 512×512 pixel raw discharge maps in PNG or CSV format; 2. Feature field: stores a 136-dimensional multimodal feature vector, which is composed of 128-dimensional basic features automatically extracted based on deep convolutional networks and 8-dimensional supplementary features designed based on domain knowledge (such as discharge statistical features, phase distribution features, etc.); 3. Metadata field: contains information such as timestamps accurate to milliseconds, latitude and longitude of equipment (monitoring point), cable equipment model, operating voltage level, and ambient temperature and humidity.

[0020] The noise database is designed to be pre-partitioned according to electromagnetic interference, mechanical vibration, ambient sound waves, and other interference types. It stores noise data similar to partial discharge data (noise samples and their corresponding 32-dimensional feature vectors (including 8-dimensional time-domain features, 12-dimensional frequency-domain features, and 12-dimensional time-frequency-domain features), noise intensity levels, and acquisition scene labels). Time-domain features can include: peak value, kurtosis, skewness, pulse duration, rising slope, falling slope, mean, and variance. Frequency-domain features can include: peak power spectral density, center frequency, bandwidth, spectral entropy, harmonic order, fundamental amplitude ratio, and 12 other items. Time-frequency-domain features can include: wavelet decomposition (db4 wavelet, 5 layers) energy entropy at each scale, and STFT spectral texture features (contrast, correlation), totaling 12 items. No further limitations or specific explanations are provided here.

[0021] To support efficient management of massive amounts of data, the data management system employs a dynamic partitioned storage architecture for defect clusters. It uses a three-dimensional composite key—"defect cluster ID + acquisition time + region code"—as the sharding key for the discharge type library. During system initialization, the system is divided into 24 basic shards based on four typical discharge defects and six geographical regions. The system also features automatic management capabilities: when a new defect pattern (discharge defect features that do not match those stored in the system) is detected and the conditions for creating a new shard are met (e.g., 30 consecutive samples with a cluster similarity greater than 85% and confirmed by manual verification), the data management system's intelligent management engine automatically creates the corresponding new shard, allocates it to low-load storage nodes, and updates the sharding routing table.

[0022] The data management system also introduces a hot partition identification mechanism, which marks the data partitions with the highest access frequency in recent times (e.g., those with ≥20% access volume in the last 7 days) as hot partitions. Hot partition data is automatically migrated to high-performance nodes with stronger configurations and larger memory, and high-capacity memory (e.g., 100GB L1 cache) is allocated to hot partition data to ensure sub-second response times for frequently accessed data. Regarding data security, the data management system can perform incremental or full backups periodically. Backup data is stored in an off-site disaster recovery center more than 100 kilometers away from the main cluster, and backup reliability is ensured through MD5 verification and 5% data integrity sampling checks (no restrictions are placed here).

[0023] Before the data management system is operational, historical discharge defect data needs to be imported and preprocessed in batches. Historical partial discharge data from the past 5 years can be imported, ensuring that there are no fewer than 5000 labeled samples for each type of defect (no specific limit), covering different operating conditions. Simultaneously, imported labeled noise samples of various types, with no fewer than 3000 samples for each type (no limit).

[0024] Data preprocessing may include: 1. Map standardization: scaling all original maps to 512×512 pixels using bilinear interpolation, converting to PNG (or CSV) format, and applying adaptive threshold segmentation for background noise reduction; 2. Feature standardization: performing Min-Max normalization to the [0,1] interval for numerical features, one-hot encoding for categorical features, and Z-Score standardization (mean 0, variance 1) for temporal features; 3. Outlier handling: identifying outlier feature values ​​using the 3σ criterion and replacing them with a threshold of mean ± 3σ, while recording outlier logs for model optimization. The preprocessed 136-dimensional multimodal feature vectors also need to undergo batch normalization to accelerate the training and convergence of subsequent deep learning models.

[0025] For noisy data samples, time-domain stretching (stretching coefficient 0.8-1.2), frequency-domain shifting (±10% center frequency), and additive white Gaussian noise superposition (signal-to-noise ratio 10-20dB) can be performed to expand each class of original samples to 8000 samples, thereby enhancing the noise samples and improving the identification of subsequent noise data.

[0026] After building a data management system and performing data preprocessing, the system can be used to filter noise and identify discharge types in newly acquired raw defect signals, enabling self-verification, self-reduction, and self-growth of data within the system.

[0027] Figure 1 This is a schematic diagram of a partial discharge data management process for a high-voltage cable provided in an embodiment of this application, as shown below. Figure 1 As shown, the method includes:

[0028] S1, acquire the original defect signal of the multimodal mode, extract features from the original defect signal to obtain the initial defect features.

[0029] Specifically, a standardized data acquisition interface can be used to receive raw defect signals in real time from various sensors (pulse current, ultrasonic, electromagnetic wave sensors, etc.) deployed at the field monitoring points. Different sensors acquire raw defect signals corresponding to different modes; for example, pulse current sensors correspond to electrical modes, ultrasonic sensors to acoustic modes, and electromagnetic wave sensors to electromagnetic modes. These raw defect signals can be electrical signals suspected of being partial discharges. Of course, in practical applications, all electrical signals, including noise, discharge, and normal data, can be acquired. A preliminary filtering mechanism can be designed to process the defect signals after filtering out normal signals; no restrictions are placed here.

[0030] S2, the initial defect features are input into a pre-constructed multi-scale noise adaptive separation network (Multi-Scale Attention Network, MS-ANet) for processing to obtain target noise features, and the target noise probability of the target noise features is calculated.

[0031] Among them, the multi-scale noise adaptive separation network can be used to calculate the target noise probability of the target noise feature, so as to determine whether the initial defect feature is noise data based on the target noise probability.

[0032] S3. When the target noise probability is greater than the preset noise probability threshold, the target noise feature is stored in the noise library; otherwise, the initial defect feature is purified according to the corresponding filtering strategy to obtain the discharge defect feature.

[0033] Specifically, the system presets a noise probability threshold (e.g., 0.85). When the target noise probability is greater than 0.85, the target noise feature can be stored in the corresponding partition of the preset noise type cluster in the noise library, and the noise type of the target noise feature can be further determined. Otherwise, it indicates that the initial defect features are not primarily noise defects and cannot be simply categorized as noise data. It is highly likely that they are discharge defects containing noise data or discharge defect features with suspected noise data characteristics. There may be some noise data mixed in with the original defect signal (of course, if the noise distribution result is 0, it means the original defect signal does not contain noise data, but the probability of this occurring is extremely low). Therefore, the initial defect features can be purified to obtain the discharge defect features. The signal purification steps can include: based on the noise probability distribution output by the MS-ANet network, using specific filtering algorithms for the main noise components: for example, for power frequency electromagnetic interference, using an IIR notch filter with adaptive center frequency; for mechanical vibration noise, using an FIR low-pass filter with adjustable cutoff frequency; for mixed noise that is difficult to separate, using the FastICA blind source separation algorithm for initial separation, and then combining it with a U-Net structure signal repair network to complete and reconstruct the damaged effective signal waveform after separation, ensuring a high retention rate of the effective signal in the initial defect features, thus obtaining the discharge defect features.

[0034] S4. Determine whether the similarity between the discharge defect feature and each preset discharge type cluster in the discharge type library meets the preset similarity requirements. If the similarity meets the preset similarity requirements, calculate the target discharge type probability of the discharge defect feature, determine the target discharge type cluster of the discharge defect feature based on the target discharge type probability, and output the corresponding target discharge type. Alternatively, if the similarity does not meet the preset similarity requirements, store the discharge defect feature in the verification buffer. In response to the feature data in the verification buffer meeting the preset new cluster creation conditions, establish a new discharge type cluster based on the discharge defect feature.

[0035] Specifically, the preset conditions for creating new clusters can be adaptively set according to changes in discharge defects. For example, when the data in a new cluster reaches a certain number (e.g., 30), a new discharge type cluster will be created automatically. There are no restrictions here.

[0036] This enables self-verification (data storage) and self-growth (new cluster establishment) of discharge defect types, outputting accurate target discharge types.

[0037] S5. Calculate the index priority score based on the access information of the discharge type library, and perform hierarchical management of the index according to the priority score.

[0038] Specifically, the index can be used to improve the efficiency of discharge type identification after the system receives data. The index in this application can be constructed based on defect cluster ID + acquisition time + region code, and there are no restrictions here. Access information can represent the access frequency ratio, average hit rate, average query time ratio, etc. of the index, and there are no restrictions here.

[0039] Hierarchical management can refer to the adaptive management of indexes with different priority scores, such as adjusting the storage location and retrieval method accordingly.

[0040] In this way, by establishing a full-link processing framework from multimodal signal acquisition, adaptive noise identification, signal purification to intelligent defect feature classification and dynamic knowledge base updates, closed-loop management of massive, high-noise, multi-source partial discharge data is achieved. This effectively overcomes the problems of severe noise interference, rigid defect identification, and the inability of the knowledge base to evolve autonomously in traditional methods. Combining discharge type matching with self-verification and self-growth mechanisms and a knowledge base update strategy, the system can not only accurately identify known defects but also proactively discover and summarize new defect patterns, significantly improving the intelligence level and long-term adaptability of the data management system, and providing more reliable and timely data support for the condition assessment of high-voltage cables.

[0041] In some embodiments, feature extraction is performed on the original defect signal to obtain initial defect features, including: performing wavelet threshold denoising on the original defect signal to obtain a denoised original defect signal; performing cross-modal feature extraction on the denoised original defect signal through a cross-modal feature extraction module to obtain a multimodal feature vector; and performing batch normalization and regularization processing on the multimodal feature vector through a feature constraint module to obtain the initial defect features.

[0042] Specifically, the original defect signal first enters the preprocessing network layer. The first step in this layer is wavelet threshold filtering, which uses the db4 wavelet basis to decompose the original defect signal into five layers. The Stein unbiased risk estimation method is used to adaptively determine the threshold of each layer, effectively filtering out high-frequency random noise. Subsequently, the denoised original defect signal is input into a lightweight cross-modal feature extraction module. This module contains a feature extraction network, which includes three convolutional layers (3×3 kernel, stride 1, padding = Same, activation function ReLU) and two max pooling layers, outputting 64-dimensional convolutional features. The output layer of the cross-modal feature extraction module concatenates these 64-dimensional convolutional features with 72-dimensional handcrafted features (such as statistical moments, spectral features, etc.) to form a 136-dimensional multimodal feature vector. This integrates signal information from different sensors, and the resulting multimodal feature vector can more comprehensively characterize the physical nature of the discharge event, compensating for the deficiencies of single-modal information.

[0043] Furthermore, a Dropout layer and a batch normalization layer are used to perform regularization and batch normalization on the multimodal feature vectors with a dropout rate of 0.3 to obtain initial defect features. In actual execution, batch normalization can be performed on the multimodal feature vectors first, and then regularization can be applied to the result of batch normalization. This can effectively avoid interference from features with different dimensions and distributions on subsequent model training, accelerate the convergence speed of the deep learning model, improve the overall stability and generalization ability of the model, and thus ensure the robustness of the entire data processing chain from the source. No specific restrictions are imposed here.

[0044] Based on the specific implementation method provided in this embodiment, a high-quality and highly comparable initial defect feature can be constructed through a standardized preprocessing workflow of "wavelet threshold denoising—cross-modal feature extraction—regularization and batch normalization," laying a clean data foundation for subsequent analysis.

[0045] In some embodiments, step S2 includes: S21, extracting features from the initial defect features according to a preset noise feature extraction dimension to obtain an initial noise feature vector; S22, processing the initial noise feature vector according to a multi-scale feature extraction module to obtain multi-scale noise features; S23, sequentially performing spatial attention and channel attention processing on the multi-scale noise features according to an attention feature extraction module to obtain key noise features; S24, processing the key noise features according to a temporal feature extraction module to obtain deep noise features; S25, processing the deep noise features through a classification and regression module to obtain target noise features and a noise probability distribution, wherein the noise probability distribution is used to characterize the noise probability of the target noise features corresponding to different noise types, and the maximum noise probability in the noise probability distribution is determined as the target noise probability; wherein, the multi-scale noise adaptive separation network includes a multi-scale feature extraction module, an attention feature extraction module, a temporal feature extraction module, and a classification and regression module.

[0046] Specifically, the initial defect features are first extracted into a 32-dimensional initial noise feature vector according to the multi-scale noise feature extraction dimensions (e.g., time-domain feature dimension, frequency-domain feature dimension, and time-frequency domain feature dimension). The initial noise feature vector is then processed by three parallel convolutional layers in the multi-scale feature extraction module to output a 64-dimensional multi-scale noise feature. The convolutional kernels of the three parallel convolutional layers are 3×3, 5×5, and 7×7, respectively.

[0047] Furthermore, the multi-scale noise features are processed by the attention feature extraction module. The module sequentially passes through the channel attention submodule (learning the importance weights of different feature channels) and the spatial attention submodule (locating key regions of noise features in the spatial dimension), filtering out irrelevant information, strengthening the weights of key noise features, and outputting the key noise features.

[0048] Furthermore, considering the correlation of noise signals over time, key noise features are fed into a temporal feature extraction module of a two-layer Long Short-Term Memory (LSTM) network. This LSTM has 256 hidden units, uses the tanh activation function, and sets a forget gate threshold of 0.8 to capture the dynamic temporal patterns of noise features and output deep noise features.

[0049] Finally, the classification and regression module outputs the target noise features and the noise probability distribution of the 16 possible noise classes corresponding to the target noise features (the specific number of noise classes output can be adjusted according to noise variations, and is not limited here). Each dimension of this noise probability distribution corresponds to the probability of a preset noise type. The maximum noise probability is recorded as the target noise probability, and the noise type corresponding to the target noise probability is recorded as the category label of the target noise feature. The classification and regression module can sequentially include a fully connected layer with a structure of 1024-256-16 and an activation function layer (the activation function can be Softmax).

[0050] In this way, by organically combining multi-scale convolution, attention mechanisms, and temporal modeling, high-precision, fine-grained identification of complex noise is achieved, breaking through the limitations of traditional single-scale noise processing. It enables adaptive identification and accurate separation of various types and intensities of noise in high-voltage cable partial discharge signals. This progressive processing of "multi-scale perception – attention focusing – temporal understanding" collectively ensures the accuracy of noise identification and the ability to detect rare noises, providing a solid guarantee for the reliable separation of effective signals.

[0051] In some embodiments, whenever the cumulative number of newly added target noise features in the noise database reaches a preset threshold, incremental training of the multi-scale noise adaptive separation network is triggered; the incremental training process includes: fine-tuning the preset learning rate of the classification and regression module.

[0052] Specifically, whenever the cumulative number of newly added target noise features in the noise database reaches a preset threshold (e.g., 1000), incremental training of the MS-ANet network is automatically triggered. Incremental training can employ a transfer learning strategy, freezing the network parameters of the feature extraction part of the MS-ANet network and only fine-tuning the subsequent classification and regression modules. A low learning rate (e.g., 0.0005) is used to train for 50 epochs, enabling the model to continuously adapt to new changes in the ambient noise features.

[0053] In this way, by monitoring the amount of new data added to the noise database, it is possible to achieve accurate separation of different types of noise, while supporting dynamic updates of the noise database and automatic triggering of incremental training. By adopting a strategy of fine-tuning the classification and regression modules with a low learning rate, new knowledge can be efficiently absorbed while retaining the general noise feature representations already learned by the network.

[0054] In some embodiments, determining whether the similarity between the discharge defect feature and each preset discharge type cluster in the discharge type library meets the preset similarity requirement includes: calculating the cosine similarity between the discharge defect feature and the center of each preset discharge type cluster; if the cosine similarity is within a first range, then the similarity is determined to meet the preset similarity requirement; if the cosine similarity is within a second range, then the inter-cluster similarity between the discharge defect feature and all preset discharge type clusters is calculated; if the inter-cluster similarity is within a third range, then the similarity is determined to meet the preset similarity requirement; if the cosine similarity is within a fourth range, or the inter-cluster similarity is within a fourth range, then the similarity is determined to not meet the preset similarity requirement, and the values ​​of the first range, the second range, the third range, and the fourth range decrease sequentially.

[0055] Specifically, the cosine similarity A between this feature and the center point of each preset discharge type cluster in the discharge type library is calculated. The preset discharge type clusters can be divided into four typical defects: tip discharge, surface discharge, internal air gap discharge, and floating potential discharge. As new clusters are established, the preset discharge type clusters are also updated accordingly.

[0056] If the cosine similarity is within the first range (e.g., A>90%), then the discharge defect characteristics can be considered to be highly similar to the corresponding preset discharge type cluster, directly marked as "verified", and stored in that cluster.

[0057] If the cosine similarity is within the second range (e.g., 75% ≤ A ≤ 90%), it indicates insufficient confidence that the discharge defect feature is the same as or similar to the preset discharge type cluster. Therefore, the discharge defect feature cannot be directly classified and requires further refined verification. This can be achieved by calculating the Mahalanobis distance between the discharge defect feature and all preset discharge type clusters (calculated based on the covariance matrix within each cluster) and converting this Mahalanobis distance into an inter-cluster similarity B. If the inter-cluster similarity is within the third range (e.g., B > 75%), then a predefined defect rule base (e.g., typical feature thresholds for various defects, i.e., determining whether the matching degree between the discharge defect feature and the standard feature of a typical discharge defect exceeds the feature threshold; for example, exceeding 0.8 indicates a match with the standard feature of that typical discharge defect) is used for logical rule verification. If successful, the feature is classified; otherwise, it is temporarily stored in a "waiting-to-be-verified buffer."

[0058] If A < 75%, or B ≤ 75%, then the discharge defect characteristic is considered to be inconsistent with all existing preset discharge type clusters. It is then temporarily stored in a "verification buffer," triggering a system alarm and notifying experts for manual review. Once manual review confirms it is not a preset discharge type cluster, a new discharge type cluster can be created in the system, and the clustering model and dynamic partitioning architecture are updated synchronously. When a new discharge type cluster meets the requirement of "30 consecutive newly added feature data cluster similarity > 85% + manual verification confirmation," a new shard is automatically created and assigned to a low-load node, and the shard routing table and index structure are updated synchronously. The above does not impose specific limitations on the range.

[0059] This two-tiered verification mechanism of "coarse screening + fine judgment" significantly reduces false positives and false negatives, improving verification accuracy. At the same time, clear judgment boundaries (such as the first to fourth range) make the entire verification process clear and interpretable, and reduce the amount of data requiring manual review, greatly alleviating the burden on operations and maintenance personnel.

[0060] It should be noted that the data in the above-mentioned "buffer to be verified" can be manually reviewed to determine its specific discharge type, or a new discharge type cluster can be established for it. Of course, the intelligent management engine in the data management system can also directly monitor and re-verify the data in the "buffer to be verified" and / or generate a new discharge type cluster in real time. No specific restrictions or elaborations are made here.

[0061] In some embodiments, calculating the target discharge type probability of the discharge defect feature includes: generating a time-frequency spectrum corresponding to the discharge defect feature; extracting features from the time-frequency spectrum to obtain initial spectrum features; performing residual processing on the initial spectrum features through a residual module with an embedded feature-enhanced residual attention mechanism to obtain enhanced spectrum features; sequentially performing global average pooling, fully connected, and normalization processing on the enhanced spectrum features to obtain a discharge type probability distribution, which is used to characterize the discharge type probability of the discharge defect feature corresponding to each preset discharge type, and determining the maximum discharge type probability in the discharge type probability distribution as the target discharge type probability.

[0062] Specifically, the system calls the ResNet-34 network with a fusion feature-enhanced residual attention mechanism to perform accurate classification and output the probability of the specific target discharge type. Figure 2 This is a schematic diagram of a ResNet-34 network structure provided in an embodiment of this application, as shown below. Figure 2The specific data processing flow of the network shown is as follows: the discharge defect features are reconstructed into their corresponding time-frequency spectra (such as PRPD spectra), which are then input into the network as a 512×512×3 RGB image. Features are extracted from this time-frequency spectra through an initial convolutional layer and a max-pooling layer to obtain initial spectra features. The initial convolutional layer can include 64 7×7 convolutional kernels with a stride of 2 and padding=Same, activated by ReLU, outputting 64×256×256 features. The pooling layer can be a 3×3 max-pooling layer with a stride of 2, processing the features output by the initial convolutional layer to obtain 64×128×128 initial spectra features.

[0063] The initial spectral features obtained after initial convolution and pooling are processed through a residual module to obtain enhanced spectral features. This residual module can contain four sets of residual blocks, as follows:

[0064] Group 1 (2 residual blocks): 3×3 convolution kernels, 64 in number, stride 1, 1×1 convolution (dimension matching) shortcut, ReLU activation, output 64×128×128;

[0065] Group 2 (2 residual blocks): 3×3 convolution kernels, 128 in number, stride 2, shortcut dimension adjustment, ReLU activation, output 128×64×64;

[0066] Group 3 (2 residual blocks): 3×3 convolution kernels, 256 in number, stride 2, shortcut dimension adjustment, ReLU activation, output 256×32×32;

[0067] Group 4 (2 residual blocks): Embedded multi-scale noise adaptive separation network, convolution kernel 3×3, number 512, stride 2, output 512×16×16.

[0068] In the fourth set of residual blocks, a feature-enhanced residual attention mechanism is embedded. This mechanism executes in parallel: 1) Channel attention: dynamically calibrating the weights of each feature channel through a squeeze-excitation module; 2) Spatial attention: focusing on key discharge regions in the map through a convolutional attention module; 3) Feature enhancement branch: strengthening weak defect features through an additional 1×1 convolution + ReLU + 3×3 convolution path. After the outputs of the three are fused, they are then processed through global average pooling and a fully connected layer (including Dropout and L2 regularization) to finally output a 4-dimensional vector. After Softmax normalization, the discharge type probability distribution is obtained. The maximum value is taken as the discharge type probability. And according to the system's pre-set parameters, for example: if the target discharge type probability is ≥0.7, the corresponding type is directly output; if the target discharge type probability is between 0.5 and 0.7, the clustering similarity result is combined for comprehensive judgment (if the clustering similarity is >85%, the clustering result is adopted); if the target discharge type probability is <0.5, the discharge defect feature is sent back to the verification buffer.

[0069] For new defect patterns confirmed by manual review, the system will activate the knowledge base self-growth mechanism. In the verification buffer, when the amount of data belonging to the same new pattern accumulates to a certain level (e.g., 30 records) and the internal cluster similarity is >85%, the system will automatically create a new discharge type cluster based on these data, assign it a unique ID, and create a corresponding new shard in the dynamic partitioned storage architecture.

[0070] Simultaneously, the system will perform online iterative optimization of the defect classification model based on the newly added data. Optimization is typically performed during daily off-peak hours (e.g., 2-4 AM). All validated new samples (features of discharge defects already classified) from the past 24 hours are extracted to form an incremental training set. Class-balanced sampling is used to divide the validated new samples into a 70% training set, a 20% validation set, and a 10% test set. The model is incrementally trained with a learning rate of 0.0005, 100 epochs, and a batch size of 32, by freezing the feature extraction layer and fine-tuning the parameters of the high-level network (i.e., embedding a feature-enhanced residual attention mechanism). If the trained classification model shows improved performance on the validation set, it is updated to the online service model, thereby achieving a continuous annual improvement in model recognition accuracy. Otherwise, the update is discarded.

[0071] Thus, by introducing an image classification network based on a feature-enhanced residual attention mechanism, one-dimensional features are reconstructed into a two-dimensional map and processed using an advanced deep residual network, fully mining the spatial structure and morphological information contained in the partial discharge map. Channel attention optimizes the weight allocation of feature channels, spatial attention focuses on key discharge regions in the map, and feature enhancement branches strengthen weak but important defect features. These three elements work synergistically to significantly improve the model's ability to identify and distinguish defects in complex conditions and with indistinct features, resulting in a continuous annualized improvement in the accuracy of discharge type identification and a faster response speed to new defect features.

[0072] In some embodiments, after step S4, the method further includes: triggering a redundancy cleanup mechanism whenever the number of newly added discharge defect features in the same discharge type cluster reaches a preset value, or at a set time. The discharge type cluster includes a preset discharge type cluster and a new discharge type cluster. The specific redundancy cleanup mechanism includes the following steps: using kernel density estimation to analyze the distribution density of defect data within the discharge type cluster, and based on the peak value of the data distribution density, dividing the discharge type cluster into different density regions, namely, dense regions, transition regions, and sparse regions with decreasing density; setting differentiated redundancy reduction thresholds for different density regions, and calculating the weighted Euclidean distance of the defect data; clustering the defect data in the same discharge type cluster according to the weighted Euclidean distance, and performing redundancy removal based on the clustering results of different density regions in each discharge type cluster and the corresponding redundancy reduction threshold.

[0073] Specifically, to address the storage redundancy issue, the system designs an adaptive intra-cluster redundancy reduction mechanism for each discharge type cluster. This mechanism can be triggered in two situations: first, when the number of newly added features in a discharge type cluster reaches a preset value (e.g., 100); second, it is executed at a set time (e.g., 3 AM daily). Of course, no restrictions are imposed here.

[0074] First, kernel density estimation can be used to analyze the distribution density of all feature data within the discharge type cluster in the feature space. Based on the peak value of the density distribution, the discharge type cluster is divided into three regions: dense region (density ≥ 0.8 × peak density), transition region (0.2 × peak density < density < 0.8 × peak density), and sparse region (density ≤ 0.2 × peak density).

[0075] Different redundancy thresholds are set for different regions: the threshold δ1 for dense regions is the strictest (e.g., 0.12), δ2 for sparse regions is the most lenient (e.g., 0.18), and the threshold δ3 for transition regions uses linear interpolation, for example, δ3 = 0.12 + 0.06 × (density - 0.2 × peak density) / (0.6 × peak density). Subsequently, the system uses a pre-trained random forest model to analyze and derive the importance weight of each dimension in the 136-dimensional features. Based on this weight, the weighted Euclidean distance between any two defective data points within a cluster is calculated.

[0076] Finally, hierarchical clustering is performed on the defect data within each cluster based on this weighted Euclidean distance. Within each sub-cluster, only the most representative defect data closest to the cluster center is retained, while the remaining data is safely deleted as redundant. After each reduction operation, the system evaluates the dispersion of features within the cluster. If the dispersion decreases by more than 20% of the original level, the system automatically calls back the threshold and re-executes the operation, ensuring that while storage is reduced by more than 30%, the diversity retention rate of features within the cluster is no less than 85%.

[0077] In this way, clustering and reduction are performed using Euclidean distance weighted by feature importance, ensuring that the retained data best represents the core features of the defective cluster. Through this reduction mechanism, the system achieves self-reduction, reducing storage requirements while maintaining feature coverage within clusters, thus maintaining the database's "health" and efficiency in the long term.

[0078] In some embodiments, step S5 further includes: analyzing access information of different indexes in real time, and calculating the priority score based on the weight coefficient of each access information, wherein the access information includes index access frequency, query time and hit rate; and performing scheduling management on different indexes based on the priority score, wherein the scheduling management includes performing corresponding storage management on different indexes based on the priority score.

[0079] In other embodiments, storage management is performed on different indexes based on priority scores, including: sorting different indexes in descending order of priority scores; marking the first N indexes in the sorting results as high-priority indexes and loading the high-priority indexes into the memory cache; and storing the N+1th and subsequent indexes in the sorting results on the disk.

[0080] Specifically, to achieve millisecond-level retrieval of massive amounts of data, this application proposes a dynamic index priority scheduling method. The system constructs a multi-layered composite index system:

[0081] The main index is a composite B+ tree index based on "defect cluster ID + acquisition time + region code", which optimizes the query by discharge type and time range.

[0082] Auxiliary indexes include inverted indexes, similarity indexes, geospatial indexes, and full-text indexes.

[0083] Inverted Index: For a 136-dimensional multimodal feature vector, each feature dimension is divided into 15 intervals (based on feature distribution statistics), and an inverted index is built in the format of "feature dimension - interval number → data ID list", which supports multi-feature range queries.

[0084] Similarity Index: Based on the cluster centers of each discharge type, a mapping of "cluster center → {data ID, similarity}" is established to optimize similarity queries.

[0085] GeoHash Index: Converts the latitude and longitude of the collected points into GeoHash encoding (6-bit precision), supporting geographic range queries.

[0086] Full-text indexing: An inverted index based on the BM25 algorithm is built for text fields such as device model and defect description, supporting fuzzy keyword search. No specific restrictions are imposed.

[0087] The dynamic index priority scheduling architecture dynamically loads indexes when a query request hits, with intelligent management at its core. The system continuously monitors query logs and analyzes access information in real time. Based on the weighting coefficients of the access information, it calculates a dynamic priority score for each index. For example: Priority Score = 0.6 × Access Frequency Percentage + 0.3 × Average Hit Rate - 0.1 × Average Query Time Percentage, where 0.6 is the weighting coefficient for access frequency percentage. Other factors are calculated similarly, without imposing any restrictions on the weighting of each access information. Based on this score, the system categorizes indexes into high, medium, and low priorities. High-priority indexes (e.g., Top 20%, where N indexes are in the top 20%) are permanently loaded into a high-speed memory cache (e.g., Redis, 50GB capacity); medium and low-priority indexes are stored on disk (e.g., SSDs) and dynamically loaded on demand during queries. Index updates also employ a differentiated strategy: primary indexes and geohash indexes are updated in real time; inverted indexes and similarity indexes are batch-merged and updated hourly to reduce I / O overhead. The system also has the ability to self-monitor index health. Once it detects that a query is performing a full table scan due to index failure, it will immediately trigger an alarm and automatically analyze the cause within 10 minutes, providing optimization suggestions or performing index rebuilding.

[0088] In this way, intelligent index management technology based on query behavior analysis can achieve efficient collaboration of multiple types of indexes by dynamically evaluating index priority and optimizing index storage and access strategies, thereby improving retrieval response speed and concurrent processing capabilities and effectively solving the bottleneck of retrieval efficiency for massive multimodal data.

[0089] Based on the above embodiments, the data management system of this application achieves self-verification through a multimodal cross-cluster verification mechanism; self-reduction through an adaptive intra-cluster redundancy reduction mechanism; and self-growth through a feature-enhanced residual attention mechanism. For these three self-mechanisms and the system's full-process collaborative monitoring, embodiments of this application also provide a system collaboration and monitoring method, specifically including the implementation of intelligent management engine operation and retrieval functions.

[0090] 1. Intelligent management engine in operation

[0091] Core functions:

[0092] Module scheduling: Based on the task queue (RabbitMQ), the system manages the tasks of the "three self" modules and innovative technology modules, and schedules them according to priority (data writing > model iteration > index optimization > redundancy cleanup) to ensure that core tasks are executed first;

[0093] Status monitoring: Collects module running status (CPU, memory, progress, error code), database status (storage capacity, read / write rate, node load, index hit rate), and query performance (response time, success rate) once per second.

[0094] Anomaly Handling: When anomalies occur (e.g., verification accuracy < 95%, storage reduction < 30%, retrieval delay > 1 second), multi-level alarms are triggered (SMS + email + platform pop-up), and emergency plans are automatically executed (starting backup nodes, pausing non-core tasks, and adjusting index cache).

[0095] Resource scheduling: Computing resources are dynamically allocated based on node load (CPU utilization > 80% is considered high load). Model training tasks are preferentially allocated to GPU nodes, while query tasks are allocated to CPU nodes.

[0096] 2. Implementation of search function (optimized based on dynamic index priority scheduling architecture)

[0097] Search process:

[0098] Users can initiate multi-condition combined queries (supporting discharge type, time, feature range, region, etc.).

[0099] The engine parses the query conditions and matches the optimal index (e.g., region queries match the GeoHash index, feature queries match the inverted index).

[0100] Index query: Multiple index results are processed using a bitmap merging algorithm (efficiently handling intersections / unions) to locate the data ID list;

[0101] Results sorting: Sort by similarity / collection time, prioritize returning data from hot partitions, and return the Top 2000 results;

[0102] Results returned: Output in JSON format (including graphs, features, and metadata), supporting visualization and CSV / PDF export;

[0103] Performance Guarantee:

[0104] Response time ≤ 1 second (2000+ samples), cache hit rate ≥ 85%;

[0105] It supports 1000+ concurrent queries and ensures stability through rate limiting (maximum queue length 5000) and load balancing.

[0106] Figure 3 This is a schematic diagram of a full-process intelligent monitoring system for a data management system provided in an embodiment of this application, such as... Figure 3 As shown, multi-sensor signals are fused through an external access layer and uniformly scheduled by the core central hub—the intelligent management engine. Data preprocessing, adaptive noise separation, feature self-balancing, and model self-growth are sequentially completed at the core processing layer. Efficient data organization and retrieval are achieved through the intelligent index cluster in the storage layer. Finally, the service and interaction layer provides users with second-level retrieval, visualization, and intelligent alarm services, realizing closed-loop management and continuous optimization of the entire data chain from collection, processing, storage to application. Further details are omitted here.

[0107] Figure 4 This is a schematic diagram of a partial discharge data management device for high-voltage cables provided in an embodiment of this application. The device includes:

[0108] The acquisition module 401 is configured to acquire the original defect signal of the multimodal mode, extract features from the original defect signal, and obtain initial defect features.

[0109] The noise probability calculation module 402 is configured to input the initial defect features into a pre-constructed multi-scale noise adaptive separation network for processing to obtain target noise features and calculate the target noise probability of the target noise features.

[0110] The noise judgment module 403 is configured to store the target noise feature into a noise library when the target noise probability is greater than a preset noise probability threshold; otherwise, it will perform signal purification on the initial defect feature according to the corresponding filtering strategy to obtain the discharge defect feature.

[0111] The defect data processing module 404 is configured to determine whether the similarity between the discharge defect feature and each preset discharge type cluster in the discharge type library meets a preset similarity requirement. When the similarity meets the preset similarity requirement, the module calculates the target discharge type probability of the discharge defect feature, determines the target discharge type cluster of the discharge defect feature based on the target discharge type probability, and outputs the corresponding target discharge type. Otherwise, the module stores the discharge defect feature in a verification buffer. In response to the feature data in the verification buffer meeting the preset new cluster creation conditions, a new discharge type cluster is established based on the discharge defect feature.

[0112] The index management module 405 is configured to calculate an index priority score based on the access information of the discharge type library, and to perform hierarchical management of the index based on the priority score.

[0113] In some embodiments, the acquisition module 401 is specifically used to perform wavelet threshold denoising on the original defect signal to obtain the denoised original defect signal; to perform cross-modal feature extraction on the denoised original defect signal through the cross-modal feature extraction module to obtain multimodal feature vectors; and to perform batch normalization and regularization processing on the multimodal feature vectors through the feature constraint module to obtain initial defect features.

[0114] In some embodiments, the noise probability calculation module 402 is specifically used to extract features from the initial defect features according to a preset noise feature extraction dimension to obtain an initial noise feature vector; process the initial noise feature vector according to the multi-scale feature extraction module to obtain multi-scale noise features; sequentially perform spatial attention and channel attention processing on the multi-scale noise features according to the attention feature extraction module to obtain key noise features; process the key noise features according to the temporal feature extraction module to obtain deep noise features; process the deep noise features through the classification and regression module to obtain the target noise features and a noise probability distribution, wherein the noise probability distribution is used to characterize the noise probability of the target noise features corresponding to different noise types; and determine the maximum noise probability in the noise probability distribution as the target noise probability; wherein the multi-scale noise adaptive separation network includes the multi-scale feature extraction module, the attention feature extraction module, the temporal feature extraction module, and the classification and regression module.

[0115] In some embodiments, whenever the cumulative number of newly added target noise features in the noise database reaches a preset threshold, incremental training of the multi-scale noise adaptive separation network is triggered; the incremental training process includes: fine-tuning the preset learning rate of the classification and regression module.

[0116] In some embodiments, the defect data processing module 404 is specifically used to calculate the cosine similarity between the discharge defect feature and the center of each preset discharge type cluster. If the cosine similarity is within a first range, it is determined that the similarity meets the preset similarity requirement. If the cosine similarity is within a second range, the Mahalanobis distance between the discharge defect feature and all preset discharge type clusters is calculated, and the inter-cluster similarity of each preset discharge type cluster is calculated in combination with the Mahalanobis distance of each preset discharge type cluster. If the inter-cluster similarity is within a third range, it is determined that the similarity meets the preset similarity requirement. If the cosine similarity is within a fourth range, or the inter-cluster similarity is within a fourth range, it is determined that the similarity does not meet the preset similarity requirement, and the values ​​of the first range, the second range, the third range, and the fourth range decrease sequentially.

[0117] In some embodiments, the defect data processing module 404 is specifically used to generate a time-frequency spectrum corresponding to the discharge defect feature, extract features from the time-frequency spectrum to obtain initial spectrum features, perform residual processing on the initial spectrum features through a residual module with an embedded feature-enhanced residual attention mechanism to obtain enhanced spectrum features, and sequentially perform global average pooling, fully connected processing and normalization processing on the enhanced spectrum features to obtain a discharge type probability distribution. The discharge type probability distribution is used to characterize the discharge type probability of the discharge defect feature corresponding to different discharge types, and the maximum discharge type probability in the discharge type probability distribution is determined as the target discharge type probability.

[0118] In some embodiments, the defect data processing module 404 is specifically used to trigger a redundancy cleanup mechanism whenever the number of newly added discharge defect features in the same discharge type cluster reaches a preset value, or at regular intervals. The redundancy cleanup mechanism includes the following steps: using kernel density estimation to analyze the distribution density of defect data within the discharge type cluster, and dividing the discharge type cluster into dense regions, transition regions, and sparse regions based on the peak value of the distribution density; setting differentiated redundancy reduction thresholds for different density regions, and calculating the weighted Euclidean distance of the defect data; clustering the defect data in the same discharge type cluster according to the weighted Euclidean distance, and removing redundancy based on the clustering results and the corresponding redundancy reduction thresholds.

[0119] In some embodiments, the index management module 405 is specifically used to analyze the access information of different indexes in real time, and calculate the priority score according to the weight coefficient of each access information. The access information includes the index access frequency, query time and hit rate. The scheduling management is performed on different indexes according to the priority score. The scheduling management includes the corresponding storage management of different indexes based on the priority score.

[0120] In some embodiments, the index management module 405 is specifically used to sort different indexes in descending order of priority score; mark the first N indexes in the sorting result as high priority indexes and load the high priority indexes into the memory cache; and store the indexes from the N+1th to the end of the sorting result on the disk.

[0121] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A method for managing partial discharge data of high-voltage cables, characterized in that, include: S1, acquire the original defect signal of the multimodal mode, extract features from the original defect signal to obtain the initial defect features; S2, the initial defect features are input into a pre-constructed multi-scale noise adaptive separation network for processing to obtain target noise features, and the target noise probability of the target noise features is calculated; S3, when the target noise probability is greater than a preset noise probability threshold, the target noise feature is stored in the noise database; Otherwise, the initial defect features are purified according to the corresponding filtering strategy to obtain the discharge defect features; S4, determine whether the similarity between the discharge defect feature and each preset discharge type cluster in the discharge type library meets the preset similarity requirements. When the similarity meets the preset similarity requirements, calculate the target discharge type probability of the discharge defect feature, determine the target discharge type cluster of the discharge defect feature based on the target discharge type probability, and output the corresponding target discharge type; otherwise, store the discharge defect feature in the verification buffer. In response to the feature data in the verification buffer meeting the preset new cluster creation conditions, establish a new discharge type cluster based on the discharge defect feature. S5. Calculate the index priority score based on the access information of the discharge type library, and perform hierarchical management of the index based on the priority score.

2. The method according to claim 1, characterized in that, The step of extracting features from the original defect signal to obtain initial defect features includes: The original defect signal is subjected to wavelet threshold denoising to obtain the denoised original defect signal. The original defect signal after denoising is subjected to cross-modal feature extraction by the cross-modal feature extraction module to obtain a multimodal feature vector; The initial defect features are obtained by performing batch normalization and regularization on the multimodal feature vectors through the feature constraint module.

3. The method according to claim 1, characterized in that, Step S2 includes: S21, extract features from the initial defect features according to the preset noise feature extraction dimension to obtain the initial noise feature vector; S22, The initial noise feature vector is processed by the multi-scale feature extraction module to obtain multi-scale noise features; S23, according to the attention feature extraction module, spatial attention and channel attention processing are sequentially performed on the multi-scale noise features to obtain key noise features; S24, The key noise features are processed by the temporal feature extraction module to obtain deep noise features; S25, the deep noise features are processed by the classification and regression module to obtain the target noise features and the noise probability distribution. The noise probability distribution is used to characterize the noise probability of the target noise features corresponding to different noise types. The maximum noise probability in the noise probability distribution is determined as the target noise probability. The multi-scale noise adaptive separation network includes a multi-scale feature extraction module, an attention feature extraction module, a temporal feature extraction module, and a classification and regression module.

4. The method according to claim 3, characterized in that, The method further includes: Whenever the cumulative number of newly added target noise features in the noise database reaches a preset threshold, incremental training of the multi-scale noise adaptive separation network is triggered; the incremental training process includes: fine-tuning the preset learning rate of the classification and regression module.

5. The method according to claim 1, characterized in that, The step of determining whether the similarity between the discharge defect features and each preset discharge type cluster in the discharge type library meets preset similarity requirements includes: Calculate the cosine similarity between the discharge defect feature and the center of each preset discharge type cluster. If the cosine similarity is within the first range, then determine that the similarity meets the preset similarity requirement. If the cosine similarity is within the second range, then the Mahalanobis distance between the discharge defect feature and all preset discharge type clusters is calculated. Combining the Mahalanobis distance of each preset discharge type cluster, the inter-cluster similarity of each preset discharge type cluster is calculated. If the inter-cluster similarity is within the third range, then the similarity is determined to meet the preset similarity requirement. If the cosine similarity is within the fourth range, or the inter-cluster similarity is within the fourth range, then it is determined that the similarity does not meet the preset similarity requirement, and the values ​​of the first range, the second range, the third range, and the fourth range decrease sequentially.

6. The method according to claim 5, characterized in that, The calculation of the target discharge type probability of the discharge defect characteristics includes: Generate a time-frequency spectrum corresponding to the discharge defect characteristics, and extract features from the time-frequency spectrum to obtain initial spectrum features; The initial spectral features are processed by a residual module that embeds a feature-enhanced residual attention mechanism to obtain enhanced spectral features; Global average pooling, full connection, and normalization are sequentially performed on the enhanced spectral features to obtain a discharge type probability distribution. The discharge type probability distribution is used to characterize the discharge type probability of the discharge defect feature corresponding to each preset discharge type. The maximum discharge type probability in the discharge type probability distribution is determined as the target discharge type probability.

7. The method according to claim 1, characterized in that, Following step S4, the following is also included: Whenever the number of newly added discharge defect features in the same discharge type cluster reaches a preset value, or at a set time, a redundancy cleanup mechanism is triggered. The discharge type cluster includes the preset discharge type cluster and the new discharge type cluster. The redundancy cleanup mechanism includes the following steps: The distribution density of defect data within the same discharge type cluster is analyzed using kernel density estimation. Based on the peak value of the distribution density, the same discharge type cluster is divided into different density regions, including dense regions, transition regions and sparse regions with density decreasing sequentially. Differential redundancy reduction thresholds are set for regions with different densities, and the weighted Euclidean distance of the defect data is calculated. The defect data in the same discharge type cluster are clustered according to the weighted Euclidean distance, and redundancy is removed according to the clustering results and the corresponding redundancy reduction threshold.

8. The method according to claim 1, characterized in that, Step S5 includes: The system analyzes access information for different indexes in real time and calculates the priority score based on the weight coefficient of each access information. The access information includes index access frequency, query time, and hit rate. The scheduling management of different indexes is based on the priority score, and the scheduling management includes corresponding storage management for different indexes based on the priority score.

9. The method according to claim 8, characterized in that, The storage management based on the priority score for different indexes includes: Arrange the different indexes in descending order according to the aforementioned priority scores; The top N indices in the sorting results are marked as high-priority indices, and these high-priority indices are loaded into the memory cache. Store the indexes from the N+1th index onwards in the sorted results to disk.

10. A partial discharge data management device for high-voltage cables, characterized in that, include: The acquisition module is configured to acquire the original defect signal of the multimodal mode, and perform feature extraction on the original defect signal to obtain initial defect features; The noise probability calculation module is configured to input the initial defect features into a pre-constructed multi-scale noise adaptive separation network for processing, obtain target noise features, and calculate the target noise probability of the target noise features. The noise judgment module is configured to store the target noise features into a noise database when the target noise probability is greater than a preset noise probability threshold. Otherwise, the initial defect features are purified according to the corresponding filtering strategy to obtain the discharge defect features; The defect data processing module is configured to determine whether the similarity between the discharge defect feature and each preset discharge type cluster in the discharge type library meets a preset similarity requirement. When the similarity meets the preset similarity requirement, the module calculates the target discharge type probability of the discharge defect feature, determines the target discharge type cluster of the discharge defect feature based on the target discharge type probability, and outputs the corresponding target discharge type. Otherwise, the module stores the discharge defect feature in a verification buffer. In response to the feature data in the verification buffer meeting the preset new cluster creation conditions, a new discharge type cluster is established based on the discharge defect feature, and the discharge type library is dynamically updated based on the data increment and feature clustering similarity within the new discharge type cluster. The index management module is configured to calculate an index priority score based on the access information of the discharge type library, and to perform hierarchical management of the index based on the priority score.

Images