A method and system for identifying power cable multi-source fault data of a deep neural network

Abstract

The application relates to the technical field of artificial intelligence, and discloses a power cable multi-source fault data recognition method and system of a deep neural network, the method comprising the following steps: collecting standardized data; based on a pre-trained deep neural network, performing heterogeneous feature separation on current-voltage signals and partial discharge signals to obtain armature transient characteristics and insulation deterioration characteristics; performing cross-domain adaptive fusion on the armature transient characteristics and the insulation deterioration characteristics to obtain unified fault representation; mapping the unified fault representation to topology data of the power cable, performing graph attention propagation on the topology data, and obtaining graph diffusion characteristics; based on the graph diffusion characteristics, performing dynamic risk assessment to obtain risk quantification indexes; based on preset fault risk association rules, performing multi-level collaborative decision on the risk quantification indexes to obtain fault type results and risk levels of the power cable; and the application can improve the accuracy and efficiency of the deep neural network for recognizing power cable multi-source fault data.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method and system for identifying multi-source fault data of power cables using deep neural networks. Background Technology

[0002] In the current field of power cable fault diagnosis, traditional fault identification methods mainly rely on the analysis of single-type data signals, such as processing only current and voltage signals or partial discharge signals independently. This method struggles to comprehensively capture the complex fault characteristics of power cables during operation, resulting in limited fault identification accuracy and a high risk of missed detections. Furthermore, existing technologies lack effective feature fusion mechanisms when processing multi-source heterogeneous data, hindering the collaborative analysis of cross-domain information and further limiting the accuracy and timeliness of fault diagnosis.

[0003] Existing fault assessment methods are mostly based on static models and fail to incorporate dynamic risk propagation analysis based on the actual cable topology, resulting in a lack of systematicness and foresight in fault location and risk assessment results. Due to the failure to fully utilize advanced artificial intelligence technologies such as deep learning and graph computing, existing systems perform poorly when dealing with multi-source, high-dimensional, and nonlinear cable fault data, exhibiting low overall identification efficiency and failing to meet the real-time requirements of smart grids for fault early warning and operation and maintenance decision-making. Summary of the Invention

[0004] This invention provides a method and system for identifying multi-source fault data in power cables using deep neural networks, in order to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides a method for identifying multi-source fault data in power cables using deep neural networks, comprising:

[0006] S1. Collect standardized data of power cables;

[0007] S2. Based on a pre-trained deep neural network, heterogeneous feature separation is performed on the current and voltage signals and partial discharge signals in the standardized data to obtain the armature transient features and insulation degradation features of the standardized data.

[0008] S3. Perform cross-domain adaptive fusion of the armature transient characteristics and the insulation degradation characteristics to obtain a unified fault representation of the standardized data;

[0009] S4. Map the unified fault representation to the topology data of the power cable, and perform graph attention propagation on the topology data to obtain the graph diffusion characteristics of the power cable;

[0010] S5. Based on the graph diffusion characteristics, perform dynamic risk assessment on the power cable to obtain the risk quantification index of the power cable;

[0011] S6. Based on preset fault risk association rules, perform multi-level collaborative decision-making on the risk quantification indicators to obtain the fault type results and risk level of the power cable.

[0012] In a preferred embodiment, the collection of standardized data from power cables includes:

[0013] Acquire the raw current and voltage signals and raw partial discharge signals of the power cable;

[0014] The original current and voltage signals and the original partial discharge signal are time-aligned to obtain the synchronization signal data of the power cable;

[0015] The synchronization signal data is normalized to obtain the standardized data of the power cable.

[0016] In a preferred embodiment, the pre-trained deep neural network performs heterogeneous feature separation on the current and voltage signals and partial discharge signals in the standardized data to obtain the armature transient features and insulation degradation features of the standardized data, including:

[0017] The current and voltage signals and partial discharge signals in the standardized data are respectively input into the convolutional paths set in parallel in the pre-trained deep neural network;

[0018] Based on the convolution path, multi-scale spatiotemporal feature extraction is performed on the current-voltage signal and the partial discharge signal to obtain the current-voltage depth feature and the partial discharge depth feature of the standardized data.

[0019] The current-voltage depth features and the partial discharge depth features are cross-integrated to obtain a joint feature representation of the standardized data;

[0020] The joint feature representation is decoupled and separated to obtain the initial armature transient features and initial insulation degradation features of the standardized data;

[0021] The initial armature transient characteristics and the initial insulation degradation characteristics are subjected to feature enhancement optimization to obtain the armature transient characteristics and insulation degradation characteristics of the standardized data.

[0022] In a preferred embodiment, the step of decoupling and separating the joint feature representation to obtain the initial armature transient features and initial insulation degradation features of the standardized data includes:

[0023] The joint feature representation is subjected to an orthogonal projection transformation to obtain the feature component matrix of the joint feature representation. The calculation formula for the orthogonal projection transformation is as follows:

[0024] ;

[0025] In the formula, The feature component matrix, The matrix representing the joint features. The parameters are the preset granularity of the decoupling process. Let be the orthogonal projection matrix of the joint feature representation. Let be the whitening transformation matrix of the joint feature representation. The adaptive scaling factor for the joint feature representation. The adaptive scaling diagonal matrix is ​​the adaptive scaling factor;

[0026] The feature component matrices are sorted by importance to obtain a priority score sequence for the feature component matrices;

[0027] Based on the priority scoring sequence, a bidirectional feature subspace projection is performed on the feature component matrix to obtain the armature-related feature components and insulation-related feature components of the feature component matrix;

[0028] Principal component reconstruction is performed on the armature-related feature components and the insulation-related feature components to obtain the initial armature transient features and initial insulation degradation features of the standardized data.

[0029] In a preferred embodiment, the cross-domain adaptive fusion of the armature transient characteristics and the insulation degradation characteristics to obtain a unified fault representation of the standardized data includes:

[0030] The armature transient characteristics are transformed in the time-frequency domain to obtain the frequency-domain armature characteristics of the armature transient characteristics;

[0031] The insulation degradation features are mapped in a feature space to obtain the shared space insulation features of the insulation degradation features;

[0032] An adaptive weight allocation is performed on the frequency domain armature features and the shared space insulation features to obtain the feature weight coefficients of the frequency domain armature features and the shared space insulation features;

[0033] Based on the feature weighting coefficients, the frequency domain armature features and the shared space insulation features are weighted and combined to obtain a preliminary fault feature representation of the standardized data.

[0034] The preliminary fault feature representation is compressed to obtain a unified fault representation.

[0035] In a preferred embodiment, mapping the unified fault representation to the topology data of the power cable includes:

[0036] The independent cable segments of the power cable are abstracted as edges, and the connection endpoints of the power cable are abstracted as nodes, in order to construct the topology graph of the power cable;

[0037] The unified fault representation is mapped to the topology graph to obtain the edge data of the topology graph;

[0038] Construct an adjacency matrix based on the node connection relationships in the topology graph;

[0039] The edge data and the adjacency matrix together constitute the topology data of the power cable.

[0040] In a preferred embodiment, performing graph attention propagation on the topology data to obtain the graph diffusion characteristics of the power cable includes:

[0041] Extract the node connection relationships and edge data from the topology data;

[0042] Based on the node connection relationships, the attention weights between the nodes and their neighboring nodes in the topology graph are calculated to obtain the attention weights between the nodes and their neighboring nodes. The formula for calculating the attention weights is as follows:

[0043] ;

[0044] In the formula, The attention weights, For the nodes in the topology data Features For the nodes in the topology data Features The feature weight matrix of the node is... The feature weight matrix of the neighboring nodes. The edge data of the topology data. For feature splicing operations, For activation function, For the node The neighborhood group, This is the transpose of the learnable attention vector in the deep neural network. For summation operations, it is an exponential function;

[0045] Based on the attention weights, the edge data of the neighboring nodes are weighted and aggregated to generate the neighborhood aggregation features of the nodes.

[0046] The neighborhood aggregation features and the edge data of the nodes are combined to obtain the node features of the topological data;

[0047] The node features of the topology data are subjected to multiple rounds of iterative propagation processing to obtain the graph diffusion features of the power cable.

[0048] In a preferred embodiment, the step of performing a dynamic risk assessment on the power cable based on the graph diffusion characteristics to obtain a risk quantification index for the power cable includes:

[0049] Based on the graph diffusion characteristics, fault path analysis is performed on the power cable to obtain the fault propagation range parameters of the power cable.

[0050] Based on the graph diffusion characteristics, the state change assessment of the power cable is performed to obtain the fault intensity development parameters of the power cable;

[0051] Based on the fault propagation range parameter and the fault intensity development parameter, a risk quantification index for the power cable is generated.

[0052] In a preferred embodiment, the step of performing multi-level collaborative decision-making on the risk quantification indicators based on preset fault risk association rules to obtain the fault type and risk level of the power cable includes:

[0053] Based on preset fault risk association rules, the risk quantification indicators are matched to obtain preliminary fault type results for the power cable.

[0054] The confidence level of the preliminary fault type results is assessed to obtain a confidence score for the preliminary fault type results;

[0055] The preliminary fault type results and the confidence score are integrated to obtain the fault type results of the power cable.

[0056] The risk level of the power cable is obtained by mapping the fault type results and the risk quantification indicators.

[0057] To address the aforementioned problems, the present invention also provides a deep neural network-based multi-source fault data identification system for power cables, the system comprising:

[0058] The data acquisition module is used to collect standardized data of power cables;

[0059] The feature separation module is used to perform heterogeneous feature separation on the current and voltage signals and partial discharge signals in the standardized data based on a pre-trained deep neural network, so as to obtain the armature transient features and insulation degradation features of the standardized data.

[0060] The feature fusion module is used to perform cross-domain adaptive fusion of the armature transient features and the insulation degradation features to obtain a unified fault characterization of the standardized data.

[0061] The topology analysis module is used to map the unified fault characterization to the topology data of the power cable, and to perform graph attention propagation on the topology data to obtain the graph diffusion characteristics of the power cable.

[0062] The risk assessment module is used to perform dynamic risk assessment on the power cable based on the graph diffusion characteristics, and obtain the risk quantification index of the power cable.

[0063] The decision output module is used to perform multi-level collaborative decision-making on the risk quantification indicators based on preset fault risk association rules, so as to obtain the fault type result and risk level of the power cable.

[0064] Compared with the prior art, the present invention has the following beneficial effects:

[0065] 1. This invention achieves efficient and accurate identification of multi-source fault data in power cables by combining deep neural networks with graph attention mechanisms. This method can perform deep heterogeneous feature separation of current, voltage, and partial discharge signals, effectively extracting armature transient features and insulation degradation features characterizing different fault mechanisms. Through cross-domain adaptive fusion technology, features from different sources and of different properties are unified into a comprehensive fault representation, significantly improving the completeness and robustness of fault feature representation.

[0066] 2. This invention maps a unified fault representation to the cable topology and performs graph attention propagation, enabling the system to accurately capture the potential propagation path and impact range of faults in the cable network. Based on this, the dynamic risk assessment and multi-level collaborative decision-making not only accurately output specific fault types but also provide quantified risk level indicators, thereby comprehensively improving the depth of fault diagnosis and the reliability of decision support, providing an efficient technical means for the intelligent operation and maintenance of power cables. Attached Figure Description

[0067] Figure 1 This is a flowchart illustrating a method for identifying multi-source fault data in power cables using a deep neural network, as provided in an embodiment of the present invention.

[0068] Figure 2 A functional block diagram of a deep neural network-based power cable multi-source fault data identification system provided in an embodiment of the present invention;

[0069] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0070] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0071] This application provides a method for identifying multi-source fault data in power cables using a deep neural network. The execution entity of this method includes, but is not limited to, at least one of the following electronic devices that can be configured to execute the method provided in this application: a server, a terminal, etc. In other words, the method can be executed by software or hardware installed on a terminal device or a server device. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cluster of cloud servers. The server can be an independent server or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms.

[0072] Reference Figure 1 The diagram shown is a flowchart illustrating a method for identifying multi-source fault data in power cables using a deep neural network, according to an embodiment of the present invention. In this embodiment, the method includes:

[0073] S1. Collect standardized data of power cables;

[0074] In this embodiment of the invention, the collection of standardized data from power cables includes:

[0075] Acquire the raw current and voltage signals and raw partial discharge signals of the power cable;

[0076] The original current and voltage signals and the original partial discharge signal are time-aligned to obtain the synchronization signal data of the power cable;

[0077] The synchronization signal data is normalized to obtain the standardized data of the power cable.

[0078] Specifically, the raw current and voltage signals of the power cable are acquired by installing Rogowski coil current sensors and capacitive voltage divider sensors at the cable joints. The Rogowski coil converts the changing current in the conductor into a measurable voltage signal through electromagnetic induction, while the capacitive voltage divider proportionally reduces the high voltage to a measurable range through a series capacitor network. The analog signals output by these sensors are transmitted via shielded cables to a 16-bit precision data acquisition device, where they are digitized at a frequency of 100,000 samples per second, forming complete raw current and voltage waveform data.

[0079] Specifically, the acquisition of the raw partial discharge signal of the power cable is accomplished using an ultra-high frequency electromagnetic wave detection method. A specially designed ultra-high frequency sensor is installed at the cable terminal, with its operating frequency range set between 300 MHz and 1500 MHz, which can effectively capture the electromagnetic wave signal radiated when partial discharge occurs inside the cable insulation layer. The sensor output signal is amplified by 40 dB by a preamplifier and then acquired by a high-speed acquisition card at a sampling rate of 2 gigabits per second, ensuring complete recording of nanosecond-level pulse waveforms to form the raw partial discharge signal containing the pulse sequence.

[0080] Furthermore, time alignment of the raw current and voltage signals and the raw partial discharge signal is achieved through a distributed sampling clock synchronization system. All data acquisition units are connected to the master clock source via a fiber optic network, which uses a rubidium atomic clock to provide a time reference accurate to the nanosecond level. The timestamp of each sampling point is directly written into the data frame header by the FPGA hardware circuit, ensuring that the absolute time error between different acquisition devices is less than 100 nanoseconds, thus achieving synchronized signal acquisition at the hardware level.

[0081] Furthermore, when there are slight differences in the sampling times of the signal, a resampling technique based on Lagrange interpolation is used for time alignment correction. This method approximates the original sampling points by constructing an nth-degree polynomial function and calculates new sampling values ​​at the target time point. For each asynchronous sampling point, four points—two known sampling points before and after it, totaling four points—are taken as interpolation nodes to calculate the precise amplitude at the target time, ensuring that all signals are perfectly aligned on the same time axis, forming strictly synchronized power cable synchronization signal data.

[0082] Furthermore, the normalization of the time-aligned synchronization signal data is implemented using the extreme value normalization method. This method first scans all sampling points of each signal channel to identify the global maximum and global minimum values ​​of the data in that channel. Then, it subtracts the minimum value of that channel from the value of each sampling point, and divides by the difference between the maximum and minimum values, compressing the amplitude range of all signals into a closed interval between zero and one, thus eliminating the data imbalance caused by differences in dimensions and amplitudes between different signals.

[0083] Furthermore, a signal quality verification step was performed after normalization. The mean, variance, and signal-to-noise ratio (SNR) of each signal channel were calculated and compared with preset quality thresholds. For signal channels with an SNR below 30 dB, a re-acquisition procedure was initiated to ensure that the standardized data input to subsequent processing steps maintained a consistent high-quality standard, providing a reliable data foundation for subsequent feature extraction.

[0084] In summary, through the complete signal acquisition, time alignment, and normalization process described above, the final standardized data possesses a unified time reference, a standard amplitude range, and verified signal quality. This data retains all the key characteristics of the original signal while eliminating data inconsistencies caused by differences in acquisition equipment, environmental interference, and other factors.

[0085] In summary, during the generation of standardized data, the time alignment process ensures strict synchronization of different signals in the time domain. This enables accurate establishment of the correspondence between current and voltage signals and partial discharge signals during subsequent feature extraction. Signal normalization unifies the dimensions and amplitude range of each signal, allowing deep neural networks to process different types of signals with equal importance weights, thus avoiding feature extraction bias caused by differences in signal amplitude.

[0086] In summary, this complete preprocessing workflow ensures the repeatability and verifiability of the conversion process from raw signals to standardized data. Each processing step employs deterministic mathematical methods, without relying on random parameters or empirical adjustments, guaranteeing that data collected from different batches and cable segments have completely consistent data specifications and quality standards after the same processing, providing solid data support for the accurate identification of subsequent multi-source fault data.

[0087] S2. Based on a pre-trained deep neural network, heterogeneous feature separation is performed on the current and voltage signals and partial discharge signals in the standardized data to obtain the armature transient features and insulation degradation features of the standardized data.

[0088] In this embodiment of the invention, the pre-trained deep neural network performs heterogeneous feature separation on the current and voltage signals and partial discharge signals in the standardized data to obtain the armature transient features and insulation degradation features of the standardized data, including:

[0089] The current and voltage signals and partial discharge signals in the standardized data are respectively input into the convolutional paths set in parallel in the pre-trained deep neural network;

[0090] Based on the convolution path, multi-scale spatiotemporal feature extraction is performed on the current-voltage signal and the partial discharge signal to obtain the current-voltage depth feature and the partial discharge depth feature of the standardized data.

[0091] The current-voltage depth features and the partial discharge depth features are cross-integrated to obtain a joint feature representation of the standardized data;

[0092] The joint feature representation is decoupled and separated to obtain the initial armature transient features and initial insulation degradation features of the standardized data;

[0093] The initial armature transient characteristics and the initial insulation degradation characteristics are subjected to feature enhancement optimization to obtain the armature transient characteristics and insulation degradation characteristics of the standardized data.

[0094] The process of decoupling and separating the joint feature representation to obtain the initial armature transient features and initial insulation degradation features of the standardized data includes:

[0095] The joint feature representation is subjected to an orthogonal projection transformation to obtain the feature component matrix of the joint feature representation. The calculation formula for the orthogonal projection transformation is as follows:

[0096] ;

[0097] In the formula, The feature component matrix, The matrix representing the joint features. The parameters are the preset granularity of the decoupling process. Let be the orthogonal projection matrix of the joint feature representation. Let be the whitening transformation matrix of the joint feature representation. The adaptive scaling factor for the joint feature representation. The adaptive scaling diagonal matrix is ​​the adaptive scaling factor;

[0098] The feature component matrices are sorted by importance to obtain a priority score sequence for the feature component matrices;

[0099] Based on the priority scoring sequence, a bidirectional feature subspace projection is performed on the feature component matrix to obtain the armature-related feature components and insulation-related feature components of the feature component matrix;

[0100] Principal component reconstruction is performed on the armature-related feature components and the insulation-related feature components to obtain the initial armature transient features and initial insulation degradation features of the standardized data.

[0101] Specifically, the current and voltage signals and partial discharge signals from the standardized data are input into parallel convolutional pathways within a pre-trained deep neural network. These two convolutional pathways employ the exact same network architecture but have independent parameter weights, each dedicated to processing signal data with different physical characteristics. The current and voltage signal pathway specifically analyzes power transmission characteristics, while the partial discharge signal pathway specifically monitors changes in insulation state. Each convolutional pathway contains three cascaded convolutional layers: the first layer uses a narrow convolutional kernel with a width of 5 sampling points to focus on local signal details; the second layer uses a medium-width convolutional kernel with a width of 15 sampling points to capture periodic signal features; and the third layer uses a wide convolutional kernel with a width of 30 sampling points to extract the overall signal trend, forming a multi-level feature extraction system from micro to macro levels. Each convolutional layer is followed by a batch normalization layer and a ReLU activation function to ensure the stability and non-linear expressive power of the feature extraction.

[0102] Specifically, multi-scale spatiotemporal feature extraction is performed on current-voltage signals and partial discharge signals based on convolutional pathways. In each convolutional layer, the input signal is computed using a sliding window with the convolutional kernel, and a feature response map is generated through an activation function. The first layer uses 32 convolutional kernels, and the output retains the high-frequency components and transient features of the signal; the second layer uses 64 convolutional kernels, and the output highlights the mid-frequency features and periodic fluctuations of the signal; the third layer uses 128 convolutional kernels, and the output emphasizes the low-frequency components and long-term trends of the signal. After dimensionality reduction through max pooling with a stride of 2, the outputs of each layer are concatenated according to the channel dimension to form 1024-dimensional current-voltage depth features and 1024-dimensional partial discharge depth features. These depth features not only retain the key information of the original signal but also have a higher level of abstract representation, laying the foundation for subsequent feature fusion.

[0103] Furthermore, feature cross-integration is performed on the current and voltage depth features and the partial discharge depth features. A bidirectional attention mechanism is used to calculate the correlation weight between the two feature sets. First, the two feature matrices are transposed and multiplied to obtain the correlation matrix. Then, the attention distribution is calculated using the softmax function. Features are selectively enhanced based on the correlation strength, with important features receiving higher weights and secondary features receiving lower weights. Through three steps—feature concatenation, weight allocation, and weighted summation—the two originally independent deep features are merged into a 2048-dimensional joint feature representation. This representation simultaneously contains complementary information on armature operating status and insulation condition, comprehensively reflecting the operating status of the cable system.

[0104] Furthermore, feature decoupling and separation are performed on the joint feature representation. First, singular value decomposition is performed on the joint feature representation matrix to obtain the eigenvector matrix and eigenvalue matrix. Then, a whitening transformation is used to eliminate the linear correlation between features, and the eigenvector matrix is ​​multiplied by the inverse square root of the eigenvalue matrix. Next, adaptive scaling is performed according to the variance of each feature dimension, with feature dimensions with larger variances receiving higher scaling factors. Finally, a feature component matrix is ​​obtained, where each feature component represents an independent physical mode, and the components maintain orthogonality.

[0105] Furthermore, the feature component matrices are ranked by importance. A priority score is obtained by calculating the matching degree between each feature component and a typical fault mode. The matching degree is calculated based on the cosine similarity between the feature component and the fault template in the feature space; the cosine of the angle between the two vectors is calculated, and the higher the similarity, the higher the score. All feature components are arranged in descending order of score to form a priority score sequence, ensuring that key features are retained first. A scoring threshold is also set; feature components below the threshold are filtered out to ensure feature quality.

[0106] Furthermore, a bidirectional feature subspace projection is performed on the feature component matrix based on the priority scoring sequence. The top 40% of high-scoring feature components are projected onto the armature feature subspace, which is spanned by the principal component vectors of historical armature fault samples. The middle 40% of feature components are projected onto the insulation feature subspace, which is spanned by the principal component vectors of historical insulation fault samples. The remaining 20% ​​of low-scoring feature components are discarded as noise. The Gramm-Schmidt orthogonalization method is used during the projection process to ensure the independence of the projected features, resulting in completely separated armature-related and insulation-related feature components.

[0107] Furthermore, principal component analysis (PCA) is used to reconstruct the armature-related and insulation-related feature components. PCA is employed to reduce the dimensionality of both types of feature components, calculate their covariance matrices, and perform eigenvalue decomposition. The top 80% of principal components with the largest variance contribution are retained, while minor components with smaller variance contributions are discarded. The main feature directions are selected by eigenvalue sorting, and the original features are projected onto the low-dimensional space spanned by these principal components, achieving feature refinement and noise reduction. The reconstructed features retain the main features of the original information while eliminating redundant noise, resulting in 256-dimensional initial armature transient features and 256-dimensional initial insulation degradation features.

[0108] Furthermore, feature enhancement optimizations were performed on the initial armature transient features and initial insulation degradation features. Dilated convolutions were used to expand the receptive field of the initial armature transient features, with an expansion rate set to 2, enabling each convolutional kernel to capture a wider range of temporal dependencies. A channel attention mechanism was applied to the initial insulation degradation features. Global average pooling was used to obtain the statistical features of each channel, and then a fully connected layer and a sigmoid function were used to generate channel weights, highlighting feature channels strongly correlated with insulation degradation. Finally, layer normalization was used to standardize the numerical distribution of the two types of features, eliminating internal covariate bias, resulting in the final armature transient features and insulation degradation features. These features maintain clear physical meaning and possess good numerical properties, facilitating subsequent processing.

[0109] In summary, through multi-scale feature extraction via parallel convolutional pathways, the system is able to capture rich temporal and frequency features from the original signal. The design of convolutional kernels at different scales ensures comprehensive feature coverage from details to the whole, while independent parameter weights guarantee specialized processing for different signal characteristics. This design fully utilizes the feature learning capabilities of deep neural networks while also taking into account the physical characteristics of power cable fault features.

[0110] In summary, feature cross-integration establishes the intrinsic connections between different signal sources, and intelligent feature fusion is achieved through an attention mechanism. Feature decoupling and separation ensure the independence and interpretability of various fault features, avoiding mutual interference between different types of features. The entire processing flow, through rigorous mathematical transformations and intelligent algorithm design, achieves precise localization and extraction of specific physical features from multi-source mixed signals.

[0111] In summary, feature importance ranking and bidirectional projection mechanisms effectively distinguish between armature-related and insulation-related features, while principal component reconstruction further refines the essential information of the features. The final feature enhancement optimization improves the representational capability of the features, ensuring that the final armature transient features and insulation degradation features retain the physical meaning of the original signals while possessing mathematical properties suitable for fault diagnosis. This complete heterogeneous feature separation process, through progressive processing steps, provides accurate and reliable input data for subsequent fault diagnosis, laying the technical foundation for multi-source fault data identification.

[0112] S3. Perform cross-domain adaptive fusion of the armature transient characteristics and the insulation degradation characteristics to obtain a unified fault representation of the standardized data;

[0113] In this embodiment of the invention, the step of performing cross-domain adaptive fusion of the armature transient characteristics and the insulation degradation characteristics to obtain a unified fault representation of the standardized data includes:

[0114] The armature transient characteristics are transformed in the time-frequency domain to obtain the frequency-domain armature characteristics of the armature transient characteristics;

[0115] The insulation degradation features are mapped in a feature space to obtain the shared space insulation features of the insulation degradation features;

[0116] An adaptive weight allocation is performed on the frequency domain armature features and the shared space insulation features to obtain the feature weight coefficients of the frequency domain armature features and the shared space insulation features;

[0117] Based on the feature weighting coefficients, the frequency domain armature features and the shared space insulation features are weighted and combined to obtain a preliminary fault feature representation of the standardized data.

[0118] The preliminary fault feature representation is compressed to obtain a unified fault representation.

[0119] Specifically, the time-frequency domain transformation of the armature transient characteristics is achieved using the short-time Fourier transform method. The armature transient characteristic sequence is divided into multiple time segments, and a Hanning window function is applied to each segment for windowing. Then, a fast Fourier transform is performed on the windowed signal to convert the time-domain characteristics into a frequency-domain representation. This process is performed across the entire signal segment with a fixed step size, ultimately yielding frequency-domain armature characteristics containing two-dimensional time-frequency information. These characteristics simultaneously preserve the temporal evolution and spectral distribution characteristics of the armature transient process.

[0120] Furthermore, feature space mapping of insulation degradation features is accomplished through kernel function transformation. Radial basis functions are used as the kernel function to calculate the similarity between insulation degradation features and a set of predefined feature basis vectors. By projecting the original insulation degradation features onto a high-dimensional feature space spanned by these basis vectors, the originally nonlinearly separable insulation degradation patterns become linearly separable in the new feature space, thus obtaining shared space insulation features.

[0121] Furthermore, adaptive weight allocation for frequency domain armature features and shared spatial insulation features is achieved through an attention mechanism. First, the two types of features are input into a fully connected neural network, and an attention score is calculated for each feature dimension. Then, a softmax function is used to convert the attention scores into normalized weight coefficients. These weight coefficients are automatically and dynamically adjusted according to the importance of each feature to fault diagnosis, ultimately yielding feature weight coefficients that accurately reflect the contribution of each feature.

[0122] Furthermore, the weighted combination of frequency-domain armature features and shared-space insulation features based on feature weight coefficients is accomplished through element-wise weighted summation. Each feature weight coefficient is multiplied by its corresponding feature dimension to enhance important features and suppress unimportant features. Then, the weighted frequency-domain armature features and shared-space insulation features are concatenated dimensionally, and a linear transformation is used to project the concatenated features onto a unified dimensional space, forming a preliminary fault feature representation.

[0123] Furthermore, information compression of the preliminary fault feature representation is achieved through an autoencoder network. The autoencoder consists of an encoder and a decoder. The encoder progressively reduces the feature dimensionality through three fully connected layers, extracting the most essential fault feature information. The decoder reconstructs the original features through a symmetrical network structure. During training, it minimizes the reconstruction error to ensure that key fault information is not lost during the compression process. Finally, a highly compressed unified fault representation is obtained from the encoder output.

[0124] In summary, through time-frequency domain transformation and feature space mapping, the system transforms features of different properties into a comparable representation space. The adaptive weight allocation mechanism ensures that key features dominate the fusion process, while information compression extracts the most essential fault feature information. The resulting unified fault representation provides a comprehensive and accurate feature foundation for subsequent fault diagnosis.

[0125] In summary, this cross-domain adaptive fusion process effectively addresses the domain differences between different types of features. Through scientific feature transformation and intelligent weight allocation, it achieves the organic integration of multi-source fault features. The unified fault representation retains the physical meaning of the original features while possessing mathematical properties suitable for deep neural network processing, laying a solid foundation for building an accurate fault diagnosis system.

[0126] S4. Map the unified fault representation to the topology data of the power cable, and perform graph attention propagation on the topology data to obtain the graph diffusion characteristics of the power cable;

[0127] In this embodiment of the invention, the step of performing graph attention propagation on the topology data to obtain the graph diffusion characteristics of the power cable includes:

[0128] Extract the node connection relationships and edge data from the topology data;

[0129] Based on the node connection relationships, the attention weights between the nodes and their neighboring nodes in the topology graph are calculated to obtain the attention weights between the nodes and their neighboring nodes. The formula for calculating the attention weights is as follows:

[0130] ;

[0131] In the formula, The attention weights, For the nodes in the topology data Features For the nodes in the topology data Features The feature weight matrix of the node is... The feature weight matrix of the neighboring nodes. The edge data of the topology data. For feature splicing operations, For activation function, For the node The neighborhood group, This is the transpose of the learnable attention vector in the deep neural network. For summation operations, It is an exponential function;

[0132] Based on the attention weights, the edge data of the neighboring nodes are weighted and aggregated to generate the neighborhood aggregation features of the nodes.

[0133] The neighborhood aggregation features and the edge data of the nodes are combined to obtain the node features of the topological data;

[0134] The node features of the topology data are subjected to multiple rounds of iterative propagation processing to obtain the graph diffusion features of the power cable.

[0135] The dynamic risk assessment of the power cable based on the graph diffusion characteristics yields a quantitative risk index for the power cable, including:

[0136] Based on the graph diffusion characteristics, fault path analysis is performed on the power cable to obtain the fault propagation range parameters of the power cable.

[0137] Based on the graph diffusion characteristics, the state change assessment of the power cable is performed to obtain the fault intensity development parameters of the power cable;

[0138] Based on the fault propagation range parameter and the fault intensity development parameter, a risk quantification index for the power cable is generated.

[0139] Specifically, mapping the unified fault representation to the topology data of power cables is achieved by abstracting each independent cable segment in the cable system as an edge in the topology graph and the cable connection points as nodes. The feature vector of each edge is filled with the unified fault representation of the corresponding cable segment, forming 128-dimensional edge feature data. Simultaneously, an adjacency matrix is ​​constructed based on the power wiring diagram. This matrix records the connectivity between nodes in binary form; when two nodes are directly connected, the corresponding position in the matrix is ​​1, otherwise it is 0. The edge data and the adjacency matrix together constitute the complete power cable topology data, containing n nodes and m edges, fully describing the connection structure and fault characteristic distribution of the cable network.

[0140] Furthermore, the node connectivity and edge data of the topology data are extracted by parsing the adjacency matrix and edge feature vectors. A depth-first search algorithm is used to traverse the adjacency matrix, establishing a list of neighboring nodes for each node and determining the connection topology between nodes. Simultaneously, a 128-dimensional fault feature vector is extracted from the edge data for each edge. These feature vectors contain key information such as the fault severity and fault type probability distribution of the corresponding cable segment. Node connectivity defines the structural features of the topology graph, while edge data carries the fault state information of each cable segment; together, they form the input basis for graph neural network processing.

[0141] Furthermore, the attention weights between nodes and their neighbors in the topological graph, based on node connectivity, are calculated using a multi-head graph attention mechanism. For each central node, its 128-dimensional feature vector is first projected onto eight different feature subspaces through a linear transformation, with each subspace corresponding to an attention head. In each attention head, the features of the central node and its neighbors are multiplied by a learnable weight matrix and then concatenated, incorporating edge feature information. The concatenated 256-dimensional features are then input into a single-layer neural network to calculate a similarity score. After processing the score using the LeakyReLU activation function (with a negative slope of 0.2), the score is normalized using the softmax function on the set of neighboring nodes to obtain the attention weights of each neighboring node on each attention head.

[0142] Furthermore, the weighted aggregation of neighboring node edge data based on attention weights is performed in parallel using a multi-head attention mechanism. In each attention head, the edge feature vector of a neighboring node is multiplied by its corresponding attention weight, and then all weighted feature vectors are summed to obtain the neighborhood aggregation feature for that attention head. The output feature vectors of the eight attention heads are concatenated dimensionally to form a 1024-dimensional intermediate feature representation. This feature is then projected onto 128 dimensions through a linear transformation to generate the neighborhood aggregation feature of the current node. This feature integrates information from neighboring nodes from multiple angles and scales, highlighting the contributions of important neighboring nodes.

[0143] Furthermore, the feature combination of neighborhood aggregation features and node edge data is achieved through a gating mechanism. First, the original edge data of the node and the neighborhood aggregation features are concatenated dimensionally to form a 256-dimensional concatenated feature. Then, an update gate and a reset gate are calculated through a gating recurrent unit to control the degree of integration of neighborhood information. The update gate determines how much of the original feature is retained, while the reset gate controls the utilization of the neighborhood feature. Finally, feature fusion is performed through a fully connected layer and a tanh activation function to obtain the updated 128-dimensional node feature. This process ensures that each node feature retains its own fault information while also including the propagation influence of neighboring nodes, achieving effective information transmission.

[0144] Furthermore, the multi-round iterative propagation processing of node features in the topology data is achieved through a layered graph attention network. Three graph attention layers are set up, with the output of each layer serving as the input to the next. In the first layer, node features primarily incorporate information from direct neighbors; in the second layer, node features capture the influence of two-hop neighbors; and in the third layer, node features contain fault propagation information within the three-hop neighborhood. Through this layered design, fault information achieves multi-hop propagation in the topology graph. Finally, after all node features converge and stabilize, these features are combined to form the graph diffusion feature of the power cable, which contains global fault state information of the cable network.

[0145] Specifically, fault path analysis of power cables based on graph diffusion characteristics is implemented using an improved Dijkstra algorithm. Starting from the source node with the highest fault probability, the optimal path for fault propagation is calculated based on the connectivity and feature similarity between nodes. A path cost function is defined, comprehensively considering the physical distance, feature similarity, and connection strength between nodes. By traversing all possible propagation paths, the primary and secondary channels of fault propagation are identified. The number, distribution density, and propagation depth of affected nodes are statistically analyzed. Combined with the fault intensity index of each node, a standardized fault propagation range parameter is calculated. This parameter ranges from 0 to 1, with a larger value indicating a wider fault impact range.

[0146] Furthermore, the assessment of power cable condition changes based on graph diffusion characteristics is accomplished through sliding time window analysis. A graph diffusion characteristic sequence of 24 consecutive time points is collected to construct a time trajectory of characteristic changes. For each node, the Mahalanobis distance between the current characteristic and the historical normal state baseline characteristic is calculated as an indicator of condition deviation. Simultaneously, linear regression analysis is used to analyze the characteristic change trend over the most recent six time points, calculating the change gradient and acceleration. Combining the importance weights of nodes, the mean and variance of the overall network fault intensity are weighted and statistically analyzed to assess the stability and abrupt change risk of fault development. Finally, a comprehensive fault intensity development parameter is obtained, reflecting the current severity and future development trend of the fault.

[0147] Furthermore, the generation of risk quantification indicators for power cables based on fault propagation range parameters and fault intensity development parameters is accomplished using a fuzzy logic system. A dual-input, single-output fuzzy inference system is established, using the propagation range and intensity development parameters as input variables, defining their respective fuzzy sets and membership functions. Nine fuzzy inference rules are designed, covering various parameter combinations. Through defuzzification, the fuzzy output is converted into precise risk quantification indicators. Simultaneously, a time decay factor is introduced to add risk scores for persistent faults and appropriately reduce risk scores for transient faults. The final output is a risk quantification indicator ranging from 0 to 100, providing intuitive numerical basis for operation and maintenance decisions.

[0148] In summary, by mapping topology data and using graph attention propagation, the system transforms point-like fault characteristics into network-based diffusion characteristics, accurately capturing the propagation patterns of faults within the cable system. The graph attention mechanism adaptively learns the influence weights between nodes, and multiple iterations ensure that fault information is fully diffused throughout the entire network. This topology-based feature propagation method effectively solves the technical challenge of handling fault propagation effects using traditional methods, providing a more comprehensive feature foundation for risk assessment.

[0149] In summary, dynamic risk assessment integrates both spatial propagation and temporal evolution dimensions, considering both the scope of fault spread within the cable network and analyzing the evolution trend of fault intensity. Through improved path analysis algorithms and refined time series analysis, the system can accurately quantify the current impact and potential risks of faults. The generated risk quantification indicators comprehensively reflect the overall safety status of the cable system, providing a scientific basis for the formulation of differentiated operation and maintenance strategies.

[0150] In summary, this process achieves a leap from local fault characteristics to global risk assessment by deeply fusing physical topology with fault data through graph neural network technology. This method can not only identify the current fault state but also predict fault development trends and assess the risk of fault propagation. This systematic risk assessment method provides accurate and reliable decision-making basis for the preventive maintenance of power cables, significantly improving the intelligence level of cable operation and maintenance.

[0151] In this embodiment of the invention, the step of performing a dynamic risk assessment on the power cable based on the graph diffusion characteristics to obtain a risk quantification index for the power cable includes:

[0152] Based on the graph diffusion characteristics, fault path analysis is performed on the power cable to obtain the fault propagation range parameters of the power cable.

[0153] Based on the graph diffusion characteristics, the state change assessment of the power cable is performed to obtain the fault intensity development parameters of the power cable;

[0154] Based on the fault propagation range parameter and the fault intensity development parameter, a risk quantification index for the power cable is generated.

[0155] Specifically, fault path analysis of power cables based on graph diffusion characteristics is implemented using an improved depth-first search algorithm. This algorithm starts with a source node whose initial fault probability exceeds 0.8 and traverses the topology graph at three depths along the edge connections. In each traversal, the algorithm records the path length, path weight, and feature values ​​of each node traversed from the current node to the source node. By calculating the feature cosine similarity between nodes and the weights of connecting edges, the primary path (similarity > 0.9) and secondary paths (similarity between 0.7 and 0.9) of fault propagation are determined. The number of all affected nodes, their spatial distribution density, and their average distance relative to the source node are statistically analyzed, ultimately calculating a fault propagation range parameter between 0 and 1, where 0 represents no propagation and 1 represents propagation throughout the entire network.

[0156] Furthermore, the assessment of power cable condition changes based on graph diffusion features is accomplished through a multi-dimensional feature dynamic comparison method. A feature benchmark library containing 30 days of historical normal conditions is established, and the graph diffusion features at the current moment are compared node by node with the benchmark library. For each node, the Euclidean distance and cosine similarity of its feature vectors are calculated to obtain a comprehensive condition deviation index. Simultaneously, the feature sequence of the most recent 24 hours is extracted, and the instantaneous rate of change is calculated using first-order differencing and the acceleration of change is calculated using second-order differencing. An exponentially weighted moving average method is used to smooth noise, and combined with the absolute value of the current condition deviation and the intensity of the change trend, a fault intensity development parameter between 0 and 100 is obtained, where a higher value indicates a more severe fault and a faster development.

[0157] Furthermore, the generation of risk quantification indicators for power cables based on fault propagation range parameters and fault intensity development parameters is achieved through a deep multilayer perceptron model. This model contains three hidden layers: the first layer has 64 neurons, using the ReLU activation function to standardize the input parameters and extract preliminary features; the second layer has 32 neurons, using the tanh activation function to learn the complex nonlinear relationship between the two parameters; and the third layer has 16 neurons, further abstracting the feature representation. The output layer uses the sigmoid activation function to map the results to a risk score of 0-100. The model is trained using 5000 sets of historical fault data, minimizing the mean squared error loss function through the Adam optimizer, with a training cycle of 1000 epochs. Early stopping is used to prevent overfitting, ensuring the accuracy and generalization ability of the risk quantification indicators.

[0158] Furthermore, the calculation of the fault propagation range parameter involves multi-level spatial dimension analysis. First, based on the activation degree of nodes in the graph diffusion characteristics, nodes are divided into core influence areas (activation degree > 0.8), secondary core influence areas (activation degree 0.5-0.8), and peripheral influence areas (activation degree < 0.5). The proportion of nodes in each region and the spatial distribution dispersion are calculated. Combined with the total number of nodes in the topological network and the average degree centrality, the basic propagation coefficient is obtained. Then, the characteristics of the fault propagation path are analyzed, including indicators such as the number of paths, average path length, and path redundancy. After dimensionality reduction through principal component analysis, the path complexity coefficient is obtained. Finally, the basic propagation coefficient and the path complexity coefficient are weighted and fused to obtain the standardized fault propagation range parameter.

[0159] Furthermore, the calculation of fault intensity development parameters employs multi-scale time dimension analysis. Sliding windows at three time scales—1 hour, 6 hours, and 24 hours—are set to extract short-term fluctuations, medium-term trends, and long-term changes in the graph diffusion characteristics, respectively. At each time scale, statistical measures such as the mean, variance, skewness, and kurtosis of the eigenvalues ​​are calculated. An autoregressive integral moving average model is used to predict the characteristic change trend for the next hour, obtaining the trend slope and confidence interval. Simultaneously, the Mahalanobis distance between the current eigenvalue and the historical normal interval is calculated to assess the degree of anomaly. Finally, the statistical characteristics from multiple time scales, the prediction results, and the degree of anomaly are fused using an attention mechanism to obtain comprehensive fault intensity development parameters.

[0160] Furthermore, environmental factor correction was incorporated into the risk quantification index generation process. A correction coefficient matrix was established, considering actual on-site factors such as cable load rate, ambient temperature, and service life. The initial risk index output by the neural network was multiplied by the correction coefficient matrix to eliminate the interference of environmental factors on the assessment results. Simultaneously, risk level thresholds were set: 0-20 for safe level, 21-40 for level of concern, 41-60 for warning level, 61-80 for level of danger, and 81-100 for level of emergency. Corresponding handling recommendations were developed for different levels, improving the practicality and operability of the risk assessment results.

[0161] In summary, the dynamic risk assessment process fully utilizes the spatial topological information and temporal evolution information of graph diffusion characteristics. An improved graph traversal algorithm accurately characterizes the fault propagation path, and multi-scale time series analysis is combined to deeply uncover patterns of state changes. This assessment system not only considers the current impact range of the fault but also predicts its development trend, achieving a leap from static assessment to dynamic early warning. The environmental adaptability of the deep learning model ensures accurate risk assessment results under different operating conditions.

[0162] In summary, this assessment mechanism establishes a complete three-dimensional assessment system encompassing space, time, and environment, transforming complex network conditions into intuitive and quantifiable risk assessment indicators. The spatial dimension analyzes the breadth and depth of fault propagation, the temporal dimension assesses the speed and trend of fault development, and the environmental dimension eliminates interference from external factors. This multi-dimensional integrated assessment method significantly improves the accuracy and practicality of risk assessment, providing a scientific basis for the refined operation and maintenance of cable systems.

[0163] In summary, the dynamic risk assessment based on graph diffusion characteristics innovatively combines graph neural networks with time series analysis, considering both the topological characteristics of cable networks and the dynamic patterns of fault development. Through system parameter design and intelligent algorithm integration, a comprehensive and accurate cable risk quantification system is constructed. This system can promptly identify potential risks and accurately assess risk levels, providing strong technical support for the safe operation of power cables.

[0164] S6. Based on preset fault risk association rules, perform multi-level collaborative decision-making on the risk quantification indicators to obtain the fault type results and risk level of the power cable.

[0165] In this embodiment of the invention, the step of performing multi-level collaborative decision-making on the risk quantification indicators based on preset fault risk association rules to obtain the fault type result and risk level of the power cable includes:

[0166] Based on preset fault risk association rules, the risk quantification indicators are matched to obtain preliminary fault type results for the power cable.

[0167] The confidence level of the preliminary fault type results is assessed to obtain a confidence score for the preliminary fault type results;

[0168] The preliminary fault type results and the confidence score are integrated to obtain the fault type results of the power cable.

[0169] The risk level of the power cable is obtained by mapping the fault type results and the risk quantification indicators.

[0170] Specifically, rule matching of risk quantification indicators based on preset fault risk association rules is achieved through a generative rule system. This system contains a set of rules in the form of "condition-conclusion," where the condition part defines different combinations of threshold ranges for the risk quantification indicators, and the conclusion part corresponds to a specific fault type. The actually acquired risk quantification indicators are matched one by one with the conditions in the rule base. When all conditions are met, the corresponding conclusion is triggered, thereby obtaining the preliminary fault type result for the power cable.

[0171] Furthermore, the confidence assessment of the preliminary failure type results is accomplished using a probabilistic inference model. This model analyzes the statistical correlation strength between risk quantification indicators and failure types, calculating the support and coverage of the current matching results. Support reflects the frequency of the failure type in historical data, while coverage reflects the ability of the risk quantification indicators to represent the failure type. These two indicators are weighted and fused to obtain a confidence score ranging from 0 to 1.

[0172] Furthermore, the preliminary fault type results and confidence scores are collaboratively integrated through a decision fusion mechanism. When multiple rules match for the same fault type, a weighted voting method is used to integrate these results, with the weights determined by the corresponding confidence scores. For conflicting fault type judgments, the result with the highest confidence score is selected as the final output. Through this collaborative integration process, accurate and reliable power cable fault type results are obtained.

[0173] Furthermore, the correlation mapping between fault type results and risk quantification indicators is accomplished through a risk matrix. The rows of this matrix correspond to different fault types, the columns correspond to the numerical ranges of the risk quantification indicators, and the matrix elements define the risk level corresponding to a specific combination of fault type and risk value. By querying this risk matrix, the fault type results are combined with the risk quantification indicators to determine the final risk level output.

[0174] Furthermore, a four-level risk classification system is adopted, from low to high: Normal, Caution, Warning, and Danger. The Normal level indicates that the cable is in good operating condition; the Caution level indicates a potential risk requiring enhanced monitoring; the Warning level indicates that obvious fault characteristics have appeared and maintenance needs to be arranged; and the Danger level means that the cable is in an emergency and requires immediate attention. This classification system provides a clear basis for operation and maintenance decisions.

[0175] In summary, through four steps—rule matching, confidence assessment, collaborative integration, and correlation mapping—the system transforms quantified risk indicators into specific fault type judgments and clear risk level classifications. This multi-level collaborative decision-making process ensures both the accuracy of fault identification and provides intuitive risk assessment results.

[0176] In summary, the multi-level decision-making mechanism based on preset rules effectively combines the advantages of expert experience and data analysis. It ensures the transparency and interpretability of the decision-making process, and improves the reliability of the decision results through confidence assessment and collaborative integration. Ultimately, it provides a complete fault diagnosis solution for the safe operation and maintenance of power cables.

[0177] like Figure 2 The diagram shown is a functional block diagram of a deep neural network-based power cable multi-source fault data identification system provided in an embodiment of the present invention.

[0178] The deep neural network-based power cable multi-source fault data identification system 100 of this invention can be installed in an electronic device. Depending on the functions implemented, the deep neural network-based power cable multi-source fault data identification system 100 may include a data acquisition module 101, a feature separation module 102, a feature fusion module 103, a topology analysis module 104, a risk assessment module 105, and a decision output module 106. The module described in this invention can also be called a unit, which refers to a series of computer program segments that can be executed by the processor of an electronic device and can perform a fixed function, and are stored in the memory of the electronic device.

[0179] In this embodiment, the functions of each module / unit are as follows:

[0180] The data acquisition module 101 is used to acquire standardized data of power cables;

[0181] The feature separation module 102 is used to perform heterogeneous feature separation on the current and voltage signals and partial discharge signals in the standardized data based on a pre-trained deep neural network, so as to obtain the armature transient features and insulation degradation features of the standardized data.

[0182] The feature fusion module 103 is used to perform cross-domain adaptive fusion of the armature transient features and the insulation degradation features to obtain a unified fault characterization of the standardized data.

[0183] The topology analysis module 104 is used to map the unified fault characterization to the topology data of the power cable, and to perform graph attention propagation on the topology data to obtain the graph diffusion characteristics of the power cable.

[0184] The risk assessment module 105 is used to perform dynamic risk assessment on the power cable based on the graph diffusion characteristics to obtain the risk quantification index of the power cable.

[0185] The decision output module 106 is used to perform multi-level collaborative decision-making on the risk quantification indicators based on preset fault risk association rules, so as to obtain the fault type result and risk level of the power cable.

[0186] In the several embodiments provided by this invention, it should be understood that the disclosed methods and systems can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and other division methods may be used in actual implementation.

[0187] The modules described as separate components may or may not be physically separate. The components shown as modules 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 modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0188] Furthermore, the functional modules 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. The integrated unit can be implemented in hardware or in the form of hardware plus software functional modules.

[0189] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0190] The embodiments of this application can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence is the theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0191] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for identifying multi-source fault data in power cables using deep neural networks, characterized in that, The method includes: S1. Collect standardized data of power cables; S2. Based on a pre-trained deep neural network, heterogeneous feature separation is performed on the current and voltage signals and partial discharge signals in the standardized data to obtain the armature transient features and insulation degradation features of the standardized data. S3. Perform cross-domain adaptive fusion of the armature transient characteristics and the insulation degradation characteristics to obtain a unified fault representation of the standardized data; S4. Map the unified fault representation to the topology data of the power cable, and perform graph attention propagation on the topology data to obtain the graph diffusion characteristics of the power cable; S5. Based on the graph diffusion characteristics, perform dynamic risk assessment on the power cable to obtain the risk quantification index of the power cable; S6. Based on preset fault risk association rules, perform multi-level collaborative decision-making on the risk quantification indicators to obtain the fault type results and risk level of the power cable.

2. The method for identifying multi-source fault data in power cables using deep neural networks as described in claim 1, characterized in that, The standardized data collected from power cables includes: Acquire the raw current and voltage signals and raw partial discharge signals of the power cable; The original current and voltage signals and the original partial discharge signal are time-aligned to obtain the synchronization signal data of the power cable; The synchronization signal data is normalized to obtain the standardized data of the power cable.

3. The method for identifying multi-source fault data in power cables using deep neural networks as described in claim 1, characterized in that, The pre-trained deep neural network performs heterogeneous feature separation on the current and voltage signals and partial discharge signals in the standardized data to obtain the armature transient features and insulation degradation features of the standardized data, including: The current and voltage signals and partial discharge signals in the standardized data are respectively input into the convolutional paths set in parallel in the pre-trained deep neural network; Based on the convolution path, multi-scale spatiotemporal feature extraction is performed on the current-voltage signal and the partial discharge signal to obtain the current-voltage depth feature and the partial discharge depth feature of the standardized data. The current-voltage depth features and the partial discharge depth features are cross-integrated to obtain a joint feature representation of the standardized data; The joint feature representation is decoupled and separated to obtain the initial armature transient features and initial insulation degradation features of the standardized data; The initial armature transient characteristics and the initial insulation degradation characteristics are subjected to feature enhancement optimization to obtain the armature transient characteristics and insulation degradation characteristics of the standardized data.

4. The method for identifying multi-source fault data in power cables using deep neural networks as described in claim 3, characterized in that, The process of decoupling and separating the joint feature representation to obtain the initial armature transient features and initial insulation degradation features of the standardized data includes: The joint feature representation is subjected to an orthogonal projection transformation to obtain the feature component matrix of the joint feature representation. The calculation formula for the orthogonal projection transformation is as follows: ； In the formula, The feature component matrix, The matrix representing the joint features. The parameters are the preset granularity of the decoupling process. Let be the orthogonal projection matrix of the joint feature representation. Let be the whitening transformation matrix of the joint feature representation. The adaptive scaling factor for the joint feature representation. The adaptive scaling diagonal matrix is ​​the adaptive scaling factor; The feature component matrices are sorted by importance to obtain a priority score sequence for the feature component matrices; Based on the priority scoring sequence, a bidirectional feature subspace projection is performed on the feature component matrix to obtain the armature-related feature components and insulation-related feature components of the feature component matrix; Principal component reconstruction is performed on the armature-related feature components and the insulation-related feature components to obtain the initial armature transient features and initial insulation degradation features of the standardized data.

5. The method for identifying multi-source fault data in power cables using deep neural networks as described in claim 1, characterized in that, The cross-domain adaptive fusion of the armature transient characteristics and the insulation degradation characteristics to obtain a unified fault representation of the standardized data includes: The armature transient characteristics are transformed in the time-frequency domain to obtain the frequency-domain armature characteristics of the armature transient characteristics; The insulation degradation features are mapped in a feature space to obtain the shared space insulation features of the insulation degradation features; An adaptive weight allocation is performed on the frequency domain armature features and the shared space insulation features to obtain the feature weight coefficients of the frequency domain armature features and the shared space insulation features; Based on the feature weighting coefficients, the frequency domain armature features and the shared space insulation features are weighted and combined to obtain a preliminary fault feature representation of the standardized data. The preliminary fault feature representation is compressed to obtain a unified fault representation.

6. The method for identifying multi-source fault data in power cables using deep neural networks as described in claim 1, characterized in that, The process of mapping the unified fault representation to the topology data of the power cable includes: The independent cable segments of the power cable are abstracted as edges, and the connection endpoints of the power cable are abstracted as nodes, in order to construct the topology graph of the power cable; The unified fault representation is mapped to the topology graph to obtain the edge data of the topology graph; Construct an adjacency matrix based on the node connection relationships in the topology graph; The edge data and the adjacency matrix together constitute the topology data of the power cable.

7. The method for identifying multi-source fault data in power cables using deep neural networks as described in claim 1, characterized in that, The step of performing graph attention propagation on the topology data to obtain the graph diffusion characteristics of the power cable includes: Extract the node connection relationships and edge data from the topology data; Based on the node connection relationships, the attention weights between the nodes and their neighboring nodes in the topology graph are calculated to obtain the attention weights between the nodes and their neighboring nodes. The formula for calculating the attention weights is as follows: ； In the formula, The attention weights, For the nodes in the topology data Features For the nodes in the topology data Features The feature weight matrix of the node is... The feature weight matrix of the neighboring nodes. The edge data of the topology data. For feature splicing operations, For activation function, For the node The neighborhood group, This is the transpose of the learnable attention vector in the deep neural network. For summation operations, It is an exponential function; Based on the attention weights, the edge data of the neighboring nodes are weighted and aggregated to generate the neighborhood aggregation features of the nodes. The neighborhood aggregation features and the edge data of the nodes are combined to obtain the node features of the topological data; The node features of the topology data are subjected to multiple rounds of iterative propagation processing to obtain the graph diffusion features of the power cable.

8. The method for identifying multi-source fault data of power cables using deep neural networks as described in claim 1, characterized in that, The dynamic risk assessment of the power cable based on the graph diffusion characteristics yields a quantitative risk index for the power cable, including: Based on the graph diffusion characteristics, fault path analysis is performed on the power cable to obtain the fault propagation range parameters of the power cable. Based on the graph diffusion characteristics, the state change assessment of the power cable is performed to obtain the fault intensity development parameters of the power cable; Based on the fault propagation range parameter and the fault intensity development parameter, a risk quantification index for the power cable is generated.

9. The method for identifying multi-source fault data of power cables using deep neural networks as described in claim 1, characterized in that, The method, based on preset fault risk association rules, performs multi-level collaborative decision-making on the risk quantification indicators to obtain the fault type and risk level of the power cable, including: Based on preset fault risk association rules, the risk quantification indicators are matched to obtain preliminary fault type results for the power cable. The confidence level of the preliminary fault type results is assessed to obtain a confidence score for the preliminary fault type results; The preliminary fault type results and the confidence score are integrated to obtain the fault type results of the power cable. The risk level of the power cable is obtained by mapping the fault type results and the risk quantification indicators.

10. A deep neural network-based multi-source fault data identification system for power cables, characterized in that, The system includes: The data acquisition module is used to collect standardized data of power cables; The feature separation module is used to perform heterogeneous feature separation on the current and voltage signals and partial discharge signals in the standardized data based on a pre-trained deep neural network, so as to obtain the armature transient features and insulation degradation features of the standardized data. The feature fusion module is used to perform cross-domain adaptive fusion of the armature transient features and the insulation degradation features to obtain a unified fault characterization of the standardized data. The topology analysis module is used to map the unified fault characterization to the topology data of the power cable, and to perform graph attention propagation on the topology data to obtain the graph diffusion characteristics of the power cable. The risk assessment module is used to perform dynamic risk assessment on the power cable based on the graph diffusion characteristics, and obtain the risk quantification index of the power cable. The decision output module is used to perform multi-level collaborative decision-making on the risk quantification indicators based on preset fault risk association rules, so as to obtain the fault type result and risk level of the power cable.

Images