High-voltage cable equipment state prediction method and device based on deep learning, electronic equipment and storage medium

Abstract

The invention relates to the technical field of power equipment state monitoring, and provides a high-voltage cable equipment state prediction method and device based on deep learning, electronic equipment and a storage medium. According to the implementation scheme, historical monitoring data of the high-voltage cable equipment are obtained, and the historical monitoring data are reconstructed to obtain a training data set; performing fusion processing on each feature in the training data set to obtain a first fusion feature; performing enhancement processing on the first fusion feature based on the local correlation and the distribution discrete degree among the features to obtain a second fusion feature; training an equipment state prediction model based on the second fusion feature to obtain a target prediction model; and taking real-time monitoring data of the high-voltage cable equipment as input of the target prediction model to obtain a state prediction result output by the target prediction model. According to the embodiment of the invention, the accuracy and reliability of the state prediction result of the high-voltage cable equipment can be improved.

Description

Technical Field

[0001] This invention relates to the field of power equipment condition monitoring technology, and in particular to a method, device, electronic device and storage medium for predicting the condition of high-voltage cable equipment based on deep learning. Background Technology

[0002] Currently, the condition monitoring and maintenance of high-voltage cables mainly rely on two methods: one is regular inspection and preventive testing. This method requires power interruption and has a fixed testing cycle, making it difficult to detect hidden defects that are dynamically changing during cable operation in real time, resulting in unreliable inspection results.

[0003] Secondly, while intelligent monitoring methods based on the Internet of Things (IoT) and big data technologies are being developed, applying these methods to the condition prediction of high-voltage cables in old urban areas presents the following challenges: cable faults are low-probability events, and the unique multi-factor coupling faults of cables in old urban areas, such as acidic soil corrosion, high-temperature operation, and short-term overload, are complex in their occurrence conditions and difficult to monitor. This results in a small number of effective fault samples available for model training, which may lead to a serious class imbalance problem in the training dataset, thus causing inaccurate model prediction results. Summary of the Invention

[0004] This invention provides a method, apparatus, electronic device, and storage medium for predicting the status of high-voltage cable equipment based on deep learning, which can solve at least one of the above-mentioned technical problems.

[0005] In a first aspect, embodiments of the present invention provide a method for predicting the state of high-voltage cable equipment based on deep learning, comprising: Historical monitoring data of high-voltage cable equipment is acquired and reconstructed to obtain a training dataset; The features in the training dataset are fused to obtain the first fused feature; Based on the local correlation and distribution dispersion among the various features, the first fused feature is enhanced to obtain the second fused feature; The device status prediction model is trained based on the second fusion feature to obtain the target prediction model; The real-time monitoring data of the high-voltage cable equipment is used as the input to the target prediction model to obtain the state prediction result output by the target prediction model.

[0006] Secondly, embodiments of the present invention provide a high-voltage cable equipment state prediction device based on deep learning, comprising: The data reconstruction module is used to acquire historical monitoring data of high-voltage cable equipment and reconstruct the historical monitoring data to obtain a training dataset. The fusion processing module is used to fuse the various features in the training dataset to obtain the first fused feature; An enhancement processing module is used to enhance the first fused feature based on the local correlation and distribution dispersion between the various features to obtain a second fused feature. The model training module is used to train the device state prediction model based on the second fusion feature to obtain the target prediction model; The state prediction module is used to take the real-time monitoring data of the high-voltage cable equipment as input to the target prediction model and obtain the state prediction result output by the target prediction model.

[0007] Thirdly, embodiments of the present invention also provide an electronic device, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the method described in any one of the embodiments of the present invention.

[0008] Fourthly, embodiments of the present invention also provide a non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to cause a computer to perform the method described in any one of the embodiments of the present invention.

[0009] By employing the technical solution of this invention, a high-quality training dataset is constructed through the systematic reconstruction and fusion of historical monitoring data of high-voltage cable equipment. Firstly, features are initially fused to form a first fused feature, enabling multi-dimensional operational information to be comprehensively expressed in a unified feature space and enhancing the correlation between different features. Subsequently, based on the local correlation and distribution dispersion of each feature, the first fused feature is selectively strengthened and reorganized to generate a second fused feature. This gives key feature dimensions higher weights during training, while redundant or low-correlation features are suppressed, further improving the effectiveness and discriminative power of feature expression. Based on this, the equipment status prediction model is trained using the second fused feature. The model can fully capture the complex mapping relationship between equipment operating status and key fault features, making the trained target prediction model more sensitive and accurate to the input real-time monitoring data. Thus, when real-time monitoring data of high-voltage cable equipment is input into the model, the target prediction model can quickly and accurately output equipment status prediction results based on the strengthened feature expression, achieving a reliable assessment of equipment health status and improving the accuracy and reliability of status prediction.

[0010] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of the present invention, nor is it intended to limit the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0011] The accompanying drawings are provided for a better understanding of this solution and do not constitute a limitation of the invention. Wherein: Figure 1 This is a flowchart of a deep learning-based high-voltage cable equipment state prediction method according to an embodiment of the present invention; Figure 2 This is a structural block diagram of a high-voltage cable equipment state prediction device based on deep learning according to an embodiment of the present invention. Figure 3 This is a schematic block diagram of an electronic device used to implement the methods of embodiments of the present invention. Detailed Implementation

[0012] The following description, in conjunction with the accompanying drawings, illustrates exemplary embodiments of the present invention, including various details to aid understanding. These details should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope of the invention. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0013] This application provides a method, apparatus, electronic device, and storage medium for predicting the condition of high-voltage cable equipment based on deep learning. This deep learning-based high-voltage cable equipment condition prediction method is applied to the condition management of high-voltage cable equipment in enterprises. The execution entity of this method can be the deep learning-based high-voltage cable equipment condition prediction apparatus provided in this application, or a computer device integrating the deep learning-based high-voltage cable equipment condition prediction apparatus. The deep learning-based high-voltage cable equipment condition prediction apparatus can be implemented in hardware or software, and the computer device can be a terminal or a server.

[0014] Figure 1 This is a flowchart of a deep learning-based high-voltage cable equipment status prediction method according to an embodiment of the present invention.

[0015] like Figure 1 As shown, this deep learning-based high-voltage cable equipment condition prediction method may include: S110: Obtain historical monitoring data of high-voltage cable equipment and reconstruct the historical monitoring data to obtain a training dataset; S120, perform fusion processing on the various features in the training dataset to obtain the first fused feature; S130, Based on the local correlation and distribution dispersion of each feature, the first fusion feature is enhanced to obtain the second fusion feature; S140, The device status prediction model is trained based on the second fusion feature to obtain the target prediction model; S150 uses real-time monitoring data of high-voltage cable equipment as input to the target prediction model to obtain the state prediction results output by the target prediction model.

[0016] For example, historical monitoring data refers to the multi-dimensional original operating parameters and status records collected by sensors and monitoring systems during the operation of high-voltage cable equipment, including information such as cable partial discharge signals, surface temperature, grounding current, soil acidity, and load current.

[0017] For example, the partial discharge intensity curve and corresponding ambient temperature and load current values ​​collected every minute from 0:00 to 24:00 on January 1st for a certain section of high-voltage cable.

[0018] For example, the first fusion feature refers to the comprehensive feature vector obtained by initially fusing the features of multidimensional historical monitoring data on the basis of the training dataset, which is used to reflect the basic correlation between different features.

[0019] For example, features such as partial discharge peak value, average soil acidity, and operating temperature fluctuation can be concatenated or weighted to form the feature vector of each sample.

[0020] For example, the second fusion feature refers to the optimized feature obtained by selectively strengthening and reorganizing the features based on the first fusion feature by utilizing the local correlation, distribution discreteness and interaction relationship between the features, so as to improve the model's sensitivity to key features.

[0021] For example, based on the first fusion feature, the combined feature of "high temperature operation-partial discharge" can be enhanced by attention mechanism or local-global enhancement method to form a more discriminative vector representation for each sample.

[0022] For example, the target prediction model refers to a neural network or other machine learning model trained using the second fusion feature, used to predict the operating status or potential faults of high-voltage cable equipment. For instance, a trained deep neural network takes the second fusion feature of each sample as input and outputs a probability distribution of the equipment's health status, such as "normal operation 0.92, minor abnormality 0.05, fault 0.03".

[0023] For example, the state prediction result refers to the equipment state assessment result calculated by the target prediction model based on the input real-time monitoring data, which is used to guide operation and maintenance or alarm decisions. For instance, for real-time monitoring data collected on a certain section of cable, the model outputs a state prediction result of "slight anomaly probability 0.85, normal operation probability 0.10, and failure probability 0.05", which is used to alert operation and maintenance personnel that there may be a risk of local overload.

[0024] According to the above implementation method, a high-quality training dataset is obtained by acquiring and reconstructing historical monitoring data of high-voltage cable equipment. Based on this dataset, various features are fused to form a first fused feature, which unifies the expression of multi-dimensional operational information and preserves the correlation between key features. Subsequently, based on the local correlation and distribution dispersion of features, the first fused feature is selectively enhanced to obtain a second fused feature, thereby strengthening key feature information, suppressing redundant features, and improving the effectiveness and discriminative ability of model training. The second fused feature is used to train the equipment status prediction model, forming a target prediction model that can accurately capture the complex mapping relationship between equipment operating status and fault features. Finally, real-time monitoring data of the high-voltage cable equipment is input into the target prediction model, which can quickly and accurately output equipment status prediction results, achieving a reliable assessment of equipment health status. This significantly improves the accuracy and reliability of status prediction while reducing interference introduced by redundant or weakly correlated features, thereby enhancing the intelligent monitoring and early warning capabilities for the operating status of high-voltage cables.

[0025] In one implementation, historical monitoring data of high-voltage cable equipment is acquired and reconstructed to obtain a training dataset. This includes: classifying the historical monitoring data to obtain historical fault samples and historical normal samples; extracting multiple key feature dimensions corresponding to each historical fault sample based on the corresponding operating condition records; using each key feature dimension as a sample point and filtering each sample point based on the Euclidean distance between them to obtain multiple neighborhood sample points; randomly interpolating the connections between each sample point and its neighborhood sample points to generate synthetic fault samples; using the synthetic fault samples and historical fault samples as input to a pre-defined conditional generative adversarial network and reconstructing the synthetic fault samples under multi-factor coupling constraints to obtain enhanced fault samples; and mixing the enhanced fault samples with each historical normal sample at a pre-defined ratio and adjusting the distribution of the mixed samples to achieve equalization, thus obtaining a training dataset.

[0026] For example, historical monitoring data is sorted out, which includes multi-dimensional parameters such as partial discharge signals, surface temperature, grounding current, soil acidity, and load current collected during cable operation, as well as corresponding sample status identifiers. Based on the sample status identifiers, all historical fault samples marked as fault states and historical normal samples marked as normal operation states are selected. Then, the data integrity of the two types of samples is verified, and invalid samples with missing parameters or excessive abnormal fluctuations are removed. Finally, a set of historical fault samples and a set of historical normal samples that can be used for subsequent processing are obtained.

[0027] For example, the operating condition records corresponding to each historical fault sample were retrieved, including soil environmental parameters, ambient temperature data, and load change curves at the time of the fault. The correlation frequency between different operating conditions and fault occurrences was statistically analyzed, revealing that the probability of fault occurrence was significantly higher in acidic soil environments, high-temperature environments, and during short-term high-load operation than in other conditions. Correlation analysis was used to calculate the correlation coefficients between the parameters of each operating condition and the characteristic parameters of the fault samples, verifying that the soil acidity parameter corresponding to acidic soil, the ambient temperature parameter corresponding to high-temperature operation, and the load current parameter corresponding to short-term overload had the strongest correlation with the fault sample characteristics. Finally, acidic soil corrosion, high-temperature operation, and short-term overload were identified as key feature dimensions for the generation of historical fault samples, and the corresponding parameter data for each historical fault sample under these dimensions were extracted.

[0028] For example, historical fault samples under key feature dimensions are transformed into three-dimensional feature vectors to construct a feature space. Each sample point in the feature space is traversed, and the Euclidean distance between that sample point and other sample points is calculated, selecting the nearest neighboring sample points. Then, for each sample point, a nearest neighbor sample point is selected from its neighboring sample points, and the line connecting the two points is determined as an interpolation path. At least one interpolation point is randomly selected along this path at preset intervals. Based on the position of the interpolation point on the path, the parameter values ​​of each key feature dimension are calculated, and combined with the mean parameter values ​​of other non-key feature dimensions in the historical fault samples, a synthetic fault sample with complete parameters is generated. Simultaneously, the generated synthetic fault sample undergoes a rationality check, and abnormal synthetic samples exceeding the actual operating range of the high-voltage cable are removed.

[0029] For example, a conditional generative adversarial network (GAN) is constructed, consisting of a generator and a discriminator. The generator employs a combination of multi-layer fully connected layers and convolutional layers to reconstruct sample features, while the discriminator uses a combination of multi-layer convolutional layers and pooling layers to determine the authenticity of samples. Synthetic fault samples and their corresponding key feature dimensions (acidic soil corrosion, high-temperature operation, short-term overload) are fed into the generator as joint input data to generate candidate enhanced fault samples. These candidate enhanced fault samples, along with historical fault samples based on the key feature dimensions, are then input into the discriminator, which outputs the authenticity judgment result. The loss values ​​of the generator and discriminator are calculated based on the judgment result. The network parameters are iteratively updated using a backpropagation algorithm to continuously optimize the authenticity of the samples generated by the generator until the discriminator can no longer effectively distinguish between candidate enhanced fault samples and historical fault samples. At this point, network training is stopped, and samples that meet the multi-factor coupled fault characteristics are selected from the generator's output samples to form an enhanced fault sample set.

[0030] For example, a preset mixing ratio of enhanced fault samples to historical normal samples is determined based on the training requirements of the equipment status prediction model, typically 1:1 or 1:2. The number of historical normal samples is counted, the required number of enhanced fault samples is calculated, and corresponding samples are selected from the enhanced fault sample set and mixed with the historical normal samples to obtain an initial mixed dataset. The proportion of fault samples to normal samples in the initial mixed dataset is calculated; if it deviates from the preset ratio, adjustments are made by increasing or decreasing the number of enhanced fault samples. Data distribution analysis methods are used to check the distribution of the two types of samples across various feature dimensions. Feature dimensions with large distribution differences are smoothed to reduce the impact of feature imbalance on model training. Finally, sample distribution verification ensures that the historical fault samples and historical normal samples in the mixed dataset achieve a balance in both quantity and feature distribution, forming a training dataset suitable for model training.

[0031] According to the above implementation method, historical monitoring data is classified to distinguish between historical fault samples and historical normal samples. Multiple key feature dimensions are extracted based on the operating condition records corresponding to the fault samples. Each fault sample's key feature dimension is used as a sample point, and neighborhood sample points are obtained by filtering them based on Euclidean distance. Preliminary synthetic fault samples are then generated through random interpolation by connecting these points. Subsequently, the synthetic fault samples and historical fault samples are input into a conditional generative adversarial network (GAN), and the synthetic samples are reconstructed under multi-factor coupling constraints to obtain enhanced fault samples. This expands the number of fault samples while retaining key operating condition-related features. Finally, the enhanced fault samples and historical normal samples are mixed in a preset ratio, and the distribution of the mixed samples is balanced to obtain the training dataset. This not only significantly increases the diversity and coverage of fault samples, improving the model's ability to identify abnormal operating conditions, but also avoids the interference of class imbalance on training results through equalization processing, thereby improving the accuracy and reliability of the equipment condition prediction model.

[0032] In one implementation, the various features in the training dataset are fused to obtain a first fused feature, including: using samples in the training dataset as input to a preset neural network algorithm, calculating multi-scale deep features of the samples layer by layer through forward propagation to obtain a multi-dimensional first feature vector; calculating the importance score of each feature channel on the multi-dimensional first feature vector channel by channel attention mechanism, wherein the channel attention mechanism includes at least two fully connected layers and a nonlinear activation function located between the two fully connected layers, the multi-dimensional first feature vector is mapped by the two fully connected layers to obtain the original importance score corresponding to each feature channel, and the original importance score is normalized by the nonlinear activation function to obtain an attention weight vector; multiplying the attention weight vector with the multi-dimensional first feature vector channel by channel to obtain a weighted feature vector; integrating the weighted feature vector by channel dimension, and normalizing and fusing the integrated features by the Softmax function (normalization exponential function) to obtain the first fused feature.

[0033] For example, the neural network algorithm includes multiple convolutional layers and pooling layers.

[0034] In this example, firstly, the samples in the training dataset are format adapted, and the high-voltage cable partial discharge signal and time-domain environmental parameters (such as soil acidity, operating temperature, load current, etc.) contained in each sample are uniformly converted into a matrix format that the neural network can accept. At the same time, the state labels corresponding to the samples are retained to assist in the subsequent verification of the effectiveness of feature extraction.

[0035] Subsequently, a neural network based on a convolutional structure was constructed. This network contains 3-5 convolutional layers and 2-3 pooling layers. The convolutional layers use 3×3 or 5×5 convolutional kernels to extract feature information from different local ranges, and the pooling layers use max pooling to retain key features and reduce data dimensionality.

[0036] Next, the formatted training dataset is input into the neural network, and forward propagation computation is initiated. The first convolutional layer extracts the pulse features of the partial discharge signal and the trend features of the time-domain environmental parameters. Subsequent convolutional layers perform in-depth mining on the output features of the previous layer, extracting more abstract multi-scale deep features. After each convolutional layer, a pooling layer is used to filter key features and reduce redundancy. Finally, the features processed by all convolutional and pooling layers are integrated into a multi-dimensional first feature vector, with each feature channel corresponding to a feature that is meaningful for representing the device state.

[0037] For example, firstly, a multidimensional first feature vector is obtained, and the number of feature channels it contains and the feature dimension of each channel are determined. Then, a channel attention mechanism module is constructed, which includes at least two fully connected layers and a nonlinear activation function (e.g., a Corrected Linear Unit (ReLU)) located between the two fully connected layers. The first fully connected layer maps each feature channel to a low-dimensional feature vector to reduce computational complexity. The nonlinear activation function enhances the nonlinear expressiveness of the features and prevents gradient vanishing. The second fully connected layer maps the low-dimensional feature vector back to the original feature channel number dimension, outputting the original importance score of each channel. Then, the original importance score of each channel is input into a Sigmoid activation function (S-shaped function) for normalization, resulting in an attention weight vector in the 0-1 interval. This vector intuitively reflects the contribution of each feature channel to the device state prediction; high-contribution channels receive larger weights, and low-contribution channels receive smaller weights.

[0038] It should be noted that in the deep learning-based high-voltage cable equipment state prediction method of this invention, the high contribution threshold in the 0-1 range output by the channel attention mechanism through the Sigmoid activation function is not an absolutely fixed value, but an engineering-optimal threshold range determined by combining the scene characteristics of high-voltage cable state prediction, the actual representational value of the feature channels, and the design goal of the attention mechanism to highlight strongly correlated key features and suppress redundant / interference features, specifically 0.7~1.0.

[0039] The determination of this threshold range is strongly related to the technical solution of this invention, and is based on the following: Priority of feature channel representation: The core fault representation features of high-voltage cables in this invention are partial discharge signals, load current mutation features corresponding to short-term overload, and temperature rise features corresponding to high-temperature operation. These features are directly related to the identification and prediction of equipment fault states and are the core basis for model training. They need to be given significantly higher weights than other features by the attention mechanism. The range of 0.7 to 1.0 can reflect the high contribution of these features and form a clear distinction from medium and low contribution features.

[0040] The purpose of introducing the channel attention mechanism in this invention is to automatically highlight key features that are strongly correlated with the device state and suppress redundant / interference features. The lower threshold of 0.7 is an effective discrimination threshold verified by model training. When the feature channel weight is ≥0.7, its influence on the device state prediction result will be significantly improved, ensuring that the feature occupies a dominant position in the subsequent weighted fusion. However, a weight below 0.7 cannot reflect the core feature attributes that are strongly correlated.

[0041] In the normalized output of the Sigmoid activation function, the range of 0.7 to 1.0 is considered a high-confidence contribution range. In deep learning algorithms for device status monitoring and fault identification, this range is a common high-contribution threshold set by the industry for core feature channels, which is highly compatible with the model training requirements of this invention (improving the sensitivity of fault status identification and accurately characterizing the feature-state mapping relationship).

[0042] Furthermore, for the feature channel types in this invention, precise matching can be performed by combining this threshold range: High contribution channels (weight 0.7~1.0): partial discharge characteristic channel, load current characteristic channel, and operating temperature characteristic channel. These channels directly reflect the core operating status of high-voltage cables and are the key basis for fault prediction. Medium contribution channel (weight 0.3~0.7): Soil acidity characteristic channel. This characteristic belongs to the slowly changing fault inducing characteristics, not the direct fault characterization characteristics. Its impact on equipment status is indirect and gradual, so it is assigned a medium contribution weight. Low contribution channels (weight 0~0.3): Redundant feature channels such as environmental noise and unrelated background parameters. These channels have no practical value for equipment status prediction and may even cause interference, so they are given low contribution weights.

[0043] For example, the i-th weight value in the attention weight vector is multiplied element-wise with the i-th feature channel of the multi-dimensional first feature vector to enhance the feature information of high-contribution channels and weaken the interference information of low-contribution channels. Then, the features multiplied channel by channel are integrated into a unified feature matrix and input into the Softmax function for normalization and fusion. This ensures that each feature exhibits a probability distribution within the overall feature set, guaranteeing that the values ​​are within a reasonable range of 0-1 and eliminating feature suppression caused by differences in the numerical ranges of different channels. Finally, the normalized result output by the Softmax function is the first fused feature, which comprehensively reflects the key feature information of the samples in the training dataset and can be used for subsequent feature enhancement and device status prediction.

[0044] It should be noted that, for ease of explanation, the above example only uses one sample as an example. For each sample in the training dataset, the steps shown in the above example can be followed.

[0045] According to the above implementation method, by inputting samples from the training dataset into a preset neural network algorithm, multi-scale deep features are extracted layer by layer using forward propagation to obtain a multi-dimensional first feature vector, which fully expresses the potential features of the samples at different levels. Subsequently, a channel attention mechanism is used to evaluate the importance of each channel of the multi-dimensional first feature vector. This mechanism generates an attention weight vector through at least two fully connected layers and nonlinear activation functions in between, and performs normalization processing, thereby adaptively highlighting feature channels that are more critical to the prediction task. The attention weight vector is multiplied with the multi-dimensional first feature vector channel by channel to obtain a weighted feature vector. Then, the weighted feature vector is integrated according to the channel dimension and fused using the Softmax function to form the first fused feature. In this way, not only is the model's responsiveness to key features enhanced, and the discriminative and expressive power of feature representation improved, but also the interference of redundant or weakly correlated features on model training can be effectively suppressed, thereby improving the accuracy and stability of the device status prediction model.

[0046] In one implementation, a second fusion feature is obtained by enhancing the first fusion feature based on the local correlation and distribution dispersion of the various features. This includes: combining the feature vectors in the first fusion feature to obtain multiple feature vector groups; mapping the three feature vectors in each feature vector group to coordinate points in two-dimensional space, and determining the local correlation strength of the feature vector group based on the area of ​​the triangle formed by the three coordinate points; recombining the feature vectors based on the local correlation strength of each feature vector group to obtain locally enhanced features; and generating the second fusion feature based on the locally enhanced features and the first fusion feature.

[0047] For example, firstly, the type of feature vectors included in the first fusion feature is defined. These feature vectors cover key information related to multi-factor coupled faults of high-voltage cables, such as partial discharge signal feature vectors, operating temperature feature vectors, soil acidity feature vectors, and load current feature vectors.

[0048] Subsequently, based on the electro-thermal-chemical-mechanical multi-stress synergistic logic, feature vectors are selected from the first fusion feature for combination. For example, the partial discharge intensity vector, soil acidity change vector, and operating temperature fluctuation vector are combined into one group, or the load current peak vector, partial discharge frequency vector, and soil acidity mean vector are combined into another group.

[0049] Finally, ensure that each feature vector group contains three potentially related feature vectors, and ensure that all key feature vectors are included in at least one vector group, forming multiple feature vector groups with a unified structure.

[0050] For example, firstly, the three feature vectors in the feature vector group are mapped to coordinate points in a two-dimensional space. For instance, the value of the first feature vector is used as the x-coordinate, the value of the second feature vector as the y-coordinate, and the value of the third feature vector is converted into a weight parameter of the two-dimensional space point using a scaling factor. Then, the area of ​​the triangle formed by the three points is calculated based on the mapped coordinate values. This area value serves as an indicator of the strength of the local correlation of the corresponding feature vector group. The smaller the area value, the more similar the trend of the feature vectors within the group and the stronger the local correlation; the larger the area value, the weaker the local correlation.

[0051] For example, firstly, the area values ​​of the triangles in all feature vector groups are calculated, and the local correlation strength intervals are divided according to the area values, such as high correlation group, medium correlation group, and low correlation group. Then, the feature vectors in the high correlation group are weighted to enhance them, while redundant feature vectors in the low correlation group can have their weights reduced or be removed. Finally, the enhanced high correlation feature vectors are recombined according to multi-factor coupling logic; for example, enhanced feature vectors related to acidic soil corrosion and partial discharge are grouped into one category, forming locally enhanced features that highlight local correlation information. Subsequently, based on the locally enhanced features and the original first fusion features, a second fusion feature can be further generated, providing more representative input features for subsequent equipment status prediction.

[0052] According to the above implementation method, multiple feature vector groups are formed by combining the feature vectors in the first fusion feature, and the three feature vectors in each group are mapped to two-dimensional spatial coordinate points. The local correlation strength of the feature vector group is quantified by calculating the area of ​​the triangle, thereby objectively reflecting the degree of interrelation between different features. Subsequently, based on the local correlation strength of each feature vector group, the feature vectors are selectively reorganized and strengthened to generate locally enhanced features, highlighting highly correlated key features and suppressing weakly correlated or redundant features. Finally, the locally enhanced features are combined with the original first fusion feature to form the second fusion feature. In this way, not only can the collaborative information between key features be strengthened, improving the effectiveness and discriminative power of feature expression, but also the interference of noise and redundant information on model training can be reduced, thereby improving the accuracy and robustness of the equipment status prediction model.

[0053] In one implementation, generating a second fusion feature based on local enhancement features and a first fusion feature includes: mapping each feature vector in the local enhancement features to feature points in a multi-dimensional space to obtain a set of feature points; performing concave point removal on the feature point set using the Graham scan method to generate a feature distribution concave hull; generating a global distribution description feature based on the area value of the feature distribution concave hull and the density distribution information of each feature point in the feature point set; and updating the first fusion feature based on the global distribution description feature and the local enhancement features to obtain the second fusion feature.

[0054] For example, firstly, the number of dimensions of the multidimensional space is determined based on the key feature types contained in the local enhancement features. For instance, a four-dimensional feature space is constructed with partial discharge features, operating temperature features, soil acidity features, and load current features as the core, with each dimension corresponding to a key feature.

[0055] Subsequently, each feature vector in the local enhancement features is decomposed into parameter values ​​of the corresponding dimension. For example, if a feature vector contains a partial discharge peak of 0.8, a temperature of 35 degrees Celsius (°C), a soil acidity of 6.2, and a load current of 120 amperes (A), it is mapped to feature points in four-dimensional spatial coordinates (0.8, 35, 6.2, 120).

[0056] Finally, all local enhanced feature vectors are traversed to complete the mapping of each vector to a feature point in multidimensional space, and all feature points are summarized to form a feature point set to ensure that the set covers the feature point distribution under different operating conditions.

[0057] For example, a convex hull calculation is performed on the set of feature points to determine the smallest convex polygon that can enclose all feature points, and the coordinates and arrangement order of the feature points on the boundary are recorded. Then, three adjacent points are selected sequentially from the starting point of the convex hull boundary, and the cross product sign of the preceding and following vectors is calculated. If the cross product sign is negative, the concave point in the middle is removed. This iterative process is repeated until the cross product sign of all three adjacent points on the boundary is positive or zero, thereby obtaining a concave boundary that tightly encloses the set of feature points, forming the feature distribution concave hull.

[0058] For example, the concave hull is divided into multiple triangular sub-regions. The area of ​​each sub-region is calculated and summed to obtain the total area of ​​the concave hull. This total area serves as a global indicator of the distribution range and dispersion of feature points. The multidimensional space is divided into equal-volume grids, and the number of feature points within each grid is counted to obtain the density distribution information of the feature points. Parameters such as the total area, average density, and the proportion of locally high-density regions are integrated into a multidimensional feature vector, forming a global distribution description feature that comprehensively describes the distribution morphology and dispersion characteristics of the feature space, providing a basis for subsequent feature enhancement and the generation of the second fusion feature.

[0059] According to the above implementation method, by mapping each feature vector in the local enhancement features to feature points in a multi-dimensional space, a feature point set is constructed, enabling the features to obtain an intuitive geometric representation in the multi-dimensional space. Subsequently, the Graham scan method is used to remove concave points from the feature point set, generating a feature distribution hull that tightly surrounds the feature point distribution, thereby accurately characterizing the spatial boundary shape of the features. Furthermore, by combining the area value of the hull and the density distribution information of the feature points, a global distribution description feature is generated, quantifying the overall distribution range, dispersion, and local density characteristics of the feature space. Finally, the first fused feature is updated based on the global distribution description feature and the local enhancement features to obtain the second fused feature. In this way, not only is comprehensive modeling of the global and local structure of the feature space achieved, enhancing the representation capability of key features, but also the interference of sparsely distributed or abnormal features is effectively suppressed, thereby improving the accuracy, stability, and sensitivity to abnormal operating conditions of the equipment condition prediction model.

[0060] In one implementation, the first fused feature is updated based on the global distribution description feature and the local enhancement feature to obtain the second fused feature, including: concatenating the local enhancement feature with the global distribution description feature to obtain a combined feature vector; standardizing the combined feature vector and selecting multiple feature vector pairs from the standardized combined feature vector; obtaining each cross product vector based on the cross product between the relative features; combining the cross product vectors in sequence to obtain an interaction matrix; performing matrix multiplication on the normalized interaction matrix and the standardized combined feature vector to obtain a weighted adjusted feature; performing a nonlinear transformation on the weighted adjusted feature to obtain a nonlinear transformed feature; and performing a residual connection between the nonlinear transformed feature and the first fused feature to obtain the second fused feature.

[0061] For example, firstly, the number of dimensions of the local enhancement features and the global distribution description features is determined, such as a 20-dimensional vector for the local enhancement features and a 5-dimensional vector for the global distribution description features. Then, the features are concatenated using a feature dimension stacking method, with the local enhancement features forming the first 20 dimensions of the combined feature vector and the global distribution description features forming the last 5 dimensions, while maintaining the parameter values ​​of both types of features unchanged during the concatenation process. Finally, the dimensions of the concatenated combined feature vector are verified to confirm that the dimension of the combined feature vector is equal to the sum of the dimensions of the local enhancement features and the global distribution description features, ensuring that both local correlation information and global distribution information are fully preserved, thus obtaining a combined feature vector that simultaneously contains local details and global patterns.

[0062] For example, the combined feature vector is standardized, and the values ​​of each dimension in the combined feature vector are converted into numerical forms that conform to a standard normal distribution by calculating the mean and standard deviation of each feature dimension, so as to eliminate the influence of differences in units between different feature dimensions. Then, according to the principle of multi-factor cross-correlation, multiple feature vector pairs are selected from the standardized combined feature vector to cover the main combination relationships between different features.

[0063] For example, for each pair of feature vectors, the cross product is calculated according to the calculation rules of the cross product and combined with the standardized vector values ​​to obtain the cross product vector that represents the orthogonal relationship and interaction strength between the corresponding pair of feature vectors.

[0064] For example, the selected feature vector pairs are numbered to determine their order. The cross product vectors corresponding to each feature vector pair are then arranged sequentially according to their numbers; for instance, the cross product vector numbered 1 is placed in the first row of the matrix, and the cross product vector numbered 2 is placed in the second row. Based on the number of feature vector pairs and the dimension of the cross product vectors, an interaction matrix is ​​constructed such that the number of rows equals the number of feature vector pairs, and the number of columns equals the dimension of the cross product vectors. This yields an interaction matrix used to comprehensively characterize the interactions between the features.

[0065] For example, the interaction matrix is ​​normalized to transform the matrix elements into a preset range to maintain the relative size relationship of each element in the matrix; then, the matching relationship between the normalized interaction matrix and the standardized combined feature vector in terms of dimension is confirmed so that the number of columns of the interaction matrix is ​​consistent with the dimension of the combined feature vector; then, according to the matrix multiplication rules, the normalized interaction matrix and the standardized combined feature vector are operated to obtain the weighted adjusted features that reflect the interaction weights of the features.

[0066] For example, a nonlinear activation function is selected for feature processing, and each element of the weighted feature is input into the nonlinear activation function for operation, so that the feature values ​​undergo nonlinear mapping, thereby enhancing the ability to represent the complex coupling relationship between features and obtaining the nonlinearly transformed features.

[0067] For example, it is confirmed that the features after nonlinear transformation and the first fused features maintain the same dimensionality. The nonlinearly transformed features are then added element-wise to the first fused features to introduce the nonlinearly transformed feature information while preserving the original information of the first fused features. This results in a second fused feature that combines the original information with complex interaction features, which is used for subsequent model processing.

[0068] According to the above implementation method, by concatenating local enhancement features with global distribution description features to form a combined feature vector, the features are uniformly expressed in terms of local correlation and global distribution information. Then, the combined feature vector is standardized, and multiple feature vector pairs are selected. The cross product vector is calculated based on the cross product between the feature vectors, and these are sequentially combined to form an interaction matrix, thus characterizing the higher-order interaction relationships between features. Next, the normalized interaction matrix is ​​multiplied with the standardized combined feature vector to obtain a weighted adjusted feature, and its expressive power is enhanced through nonlinear transformation. Finally, the nonlinearly transformed feature is residually concatenated with the first fused feature to form a second fused feature. This approach not only fully explores the potential interaction relationships between features and improves the discrimination ability of key features, but also suppresses the interference of noisy features while retaining the original feature information, achieving a richer and more robust feature representation. This significantly improves the accuracy, reliability, and adaptability to complex operating conditions of the equipment condition prediction model.

[0069] In one implementation, training a device state prediction model based on a second fusion feature to obtain a target prediction model includes: performing forward propagation calculation on the second fusion feature using the device state prediction model to obtain a preliminary state prediction result; calculating the prediction error between the preliminary state prediction result and the true label using a cross-entropy loss function; calculating the gradient of the prediction error with respect to each network parameter in the device state prediction model using a backpropagation algorithm, and updating the weight parameters of the device state prediction model using an adaptive moment estimation algorithm; and repeating the forward propagation, loss calculation, backpropagation, and weight parameter update process until the prediction error converges or reaches a preset number of training rounds to obtain the target prediction result.

[0070] For example, firstly, a device state prediction model is constructed, comprising multiple fully connected layers and several Dropout layers (random deactivation layers). The number of neurons in the fully connected layers gradually decreases from the input layer to the output layer. Dropout layers are placed between adjacent fully connected layers to randomly discard some neuron outputs during training to reduce the risk of overfitting. Secondly, the second fused feature is used as the input to the device state prediction model. The model performs linear transformations and nonlinear activation operations layer by layer to gradually extract high-level abstract features. Thirdly, the output layer uses the Softmax function to normalize each category, obtaining the probability distribution corresponding to each device state category. This probability distribution serves as the preliminary state prediction result.

[0071] For example, the ground truth labels corresponding to the training samples are obtained. These labels are annotated based on the historical operation records and test results of the high-voltage cable equipment and represented using one-hot encoding. The probability distribution of the preliminary state prediction results is compared with the ground truth labels, and the difference between the two is measured using the cross-entropy loss function to obtain the loss value for each sample. The loss values ​​of all training samples are summarized and averaged to obtain the prediction error for the current training round. The prediction error reflects the degree of deviation between the preliminary state prediction results and the ground truth labels.

[0072] For example, starting from the output layer of the device state prediction model, the gradient values ​​of the prediction error with respect to each network parameter are calculated layer by layer based on the backpropagation algorithm. The network parameters include the weight parameters and bias parameters in each fully connected layer.

[0073] For example, the Adaptive Moment Estimation (Adam) algorithm is used to estimate the first and second moments of the gradient, and the update step size of each network parameter is adaptively adjusted in conjunction with a preset learning rate. The weight parameters of the device state prediction model are corrected according to the update step size, so that the prediction error gradually decreases during the iteration process.

[0074] For example, a preset number of training rounds and a prediction error convergence threshold are set. Then, in each training round, forward propagation calculation of the second fused feature is performed sequentially to obtain a preliminary state prediction result, loss function calculation is performed to obtain the prediction error, and backpropagation and weight parameter update processes are performed to train the device state prediction model. When the change in prediction error is less than the convergence threshold in several consecutive training rounds, or when the preset number of training rounds is reached, training is stopped, and the prediction result corresponding to the device state prediction model at this time is output as the target prediction result.

[0075] According to the above implementation method, by inputting the second fused feature into the equipment state prediction model, a preliminary state prediction result is obtained using forward propagation, thereby achieving a preliminary judgment of the equipment operating state. Subsequently, the prediction error between the preliminary prediction result and the true label is calculated using the cross-entropy loss function, and the gradient of the error with respect to each network parameter of the model is obtained through the backpropagation algorithm. The model weights are then updated using the adaptive moment estimation algorithm to achieve iterative optimization of the model. The forward propagation, loss calculation, backpropagation, and weight update are repeated until the prediction error converges or reaches the preset training rounds, finally obtaining the target prediction result. In this way, not only can the model be trained efficiently to capture the complex mapping relationship between equipment state and key features, but also the network parameters can be adaptively adjusted to improve the prediction accuracy and robustness of the model under different operating conditions, thereby ensuring reliable monitoring and intelligent early warning capabilities for the operating state of high-voltage cable equipment.

[0076] In one implementation, real-time monitoring data of the high-voltage cable equipment is used as input to the target prediction model to obtain the state prediction result output by the target prediction model.

[0077] For example, firstly, the real-time acquired monitoring data is preprocessed to obtain preprocessed real-time data. The real-time monitoring data includes partial discharge signals, operating temperature data, soil acidity parameters, load current data, and environmental noise data. The real-time monitoring data is then cleaned by setting reasonable numerical thresholds to remove outliers and missing values. Next, the cleaned data undergoes denoising to reduce the impact of environmental electromagnetic interference and external temperature fluctuations on the monitoring data. Subsequently, the denoised data is standardized to map data of different dimensions to a preset numerical range, resulting in preprocessed real-time data.

[0078] Then, the preprocessed real-time data is input into the target prediction model. The feature extraction network in the target prediction model performs forward propagation processing on the preprocessed real-time data to extract the pulse features of the partial discharge signal, the temporal variation features of the operating temperature, the variation trend features of the soil acidity, and the peak features of the load current, forming a multidimensional initial feature vector containing multiple feature channels.

[0079] Next, based on the multidimensional initial feature vector, the weights of each feature channel are calculated using the attention mechanism in the target prediction model, and the feature channels are then weighted and fused to obtain the primary fused features. Among them, feature channels with a high correlation to the device state are assigned larger weights, while noise feature channels with a low correlation to the device state are assigned smaller weights, thereby highlighting key feature information and suppressing interfering feature information.

[0080] Furthermore, local correlation enhancement and global distribution evaluation operations are sequentially performed on the primary fusion features to construct an interaction matrix. Based on this interaction matrix, the features are further enhanced to obtain the final enhanced fusion features. Specifically, according to the multi-stress cross-correlation principle, the feature vectors in the primary fusion features are grouped. The local correlation strength is determined based on the geometric relationship between each group of feature vectors, and the highly correlated feature vectors are enhanced. Simultaneously, the enhanced feature vectors are mapped to feature points in multi-dimensional space. A feature distribution hull is generated based on the feature point set, and the global distribution description features are obtained by combining the hull area and feature point density information. The locally enhanced features and the global distribution description features are then concatenated to construct a combined feature vector. An interaction matrix is ​​generated based on the cross product relationship between the combined feature vectors, and the combined feature vectors are weighted and nonlinearly transformed using the interaction matrix to obtain the final enhanced fusion features.

[0081] Finally, the enhanced fusion features are input into the prediction network of the target prediction model for forward propagation calculation, yielding the state prediction results output by the target prediction model. The state prediction results represent the probability distribution or corresponding classification labels of the high-voltage cable equipment in different state categories. These probability distributions or classification labels are then output as the final state prediction results, characterizing the real-time operating status of the high-voltage cable equipment.

[0082] According to the above implementation method, the real-time collected monitoring data (real-time monitoring data) is cleaned, denoised, and standardized to remove outliers, environmental interference noise, and dimensional differences, ensuring the accuracy and consistency of the preprocessed real-time data. The preprocessed real-time monitoring data is input into the target prediction model, and a multi-dimensional initial feature vector is extracted through a feature extraction network. This allows for the comprehensive capture of the electrical, thermal, chemical, and mechanical fundamental information of the high-voltage cable from the real-time data, providing reliable initial data support for subsequent feature fusion and enhancement. Based on the multi-dimensional initial feature vector, the weights of each feature channel are calculated using an attention mechanism, automatically identifying the more critical features for equipment status prediction in the real-time data and quantifying the contribution of different features. Weighted fusion of features yields primary fused features, highlighting the role of key features and suppressing redundant or interfering features, allowing the primary fused features to focus more on the core characterization information of the equipment status. Local correlation enhancement is then applied to the primary fused features sequentially to capture the dynamic correlations between features in the real-time data (such as the short-term linkage changes between partial discharge intensity and temperature), supplementing the detailed interaction information of the features. Global distribution assessment allows us to grasp the distribution patterns of features in real-time space. Simultaneously, constructing an interaction matrix and using it for feature enhancement further characterizes the real-time coupling effects between multiple factors, ensuring that the final enhanced fused features possess real-time performance, local correlation, and global regularity, accurately reflecting the real-time operating status of the equipment. Inputting the final enhanced fused features into a prediction network for forward propagation calculations allows for rapid mapping of the relationship between real-time features and equipment status using trained network parameters, efficiently obtaining preliminary prediction results and meeting the real-time requirements of high-voltage cable condition monitoring. Analyzing the preliminary prediction results yields probability values ​​or classification labels representing the equipment's health status, providing a clear visual representation of the equipment's current state. This facilitates rapid understanding of the equipment's condition by maintenance personnel and provides a clear basis for subsequent maintenance decisions (such as whether to initiate maintenance).

[0083] Figure 2 This is a structural block diagram of a high-voltage cable equipment state prediction device based on deep learning according to an embodiment of the present invention.

[0084] like Figure 2 As shown, the deep learning-based high-voltage cable equipment state prediction device may include: The data reconstruction module 510 is used to acquire historical monitoring data of high-voltage cable equipment and reconstruct the historical monitoring data to obtain a training dataset. The fusion processing module 520 is used to perform fusion processing on the various features in the training dataset to obtain the first fused feature; The enhancement processing module 530 is used to enhance the first fused feature based on the local correlation and distribution dispersion between the various features to obtain the second fused feature; The model training module 540 is used to train the device state prediction model based on the second fusion feature to obtain the target prediction model; The state prediction module 550 is used to take the real-time monitoring data of the high-voltage cable equipment as the input of the target prediction model and obtain the state prediction result output by the target prediction model.

[0085] In one embodiment, the data reconstruction module includes: A classification unit is used to classify the historical monitoring data to obtain various historical fault samples and various historical normal samples; The unit is used to extract multiple key feature dimensions corresponding to each of the historical fault samples based on the operating condition records corresponding to each of the historical fault samples. The filtering unit is used to take each of the historical fault samples as a sample point, and filter each sample point based on the Euclidean distance between each sample point to obtain multiple neighborhood sample points. A random interpolation unit is used to randomly interpolate the connections between each of the sample points and each of the neighboring sample points to generate a synthetic fault sample. The sample reconstruction unit is used to take the synthetic fault sample and each of the historical fault samples as input to a preset conditional generative adversarial network, and reconstruct the synthetic fault sample with multi-factor coupling conditions as constraints to obtain an enhanced fault sample. The mixing and adjustment unit is used to mix the enhanced fault samples with each of the historical normal samples according to a preset ratio, and to balance the distribution of the mixed samples to obtain the training dataset.

[0086] In one embodiment, the fusion processing module includes: The extraction unit is used to take the samples in the training dataset as input to a preset neural network algorithm, and extract the multi-scale deep features of the samples layer by layer through forward propagation to obtain a multi-dimensional first feature vector. The importance score calculation unit is used to calculate the importance score of each feature channel of the multidimensional first feature vector through a channel attention mechanism. The channel attention mechanism includes at least two fully connected layers and a nonlinear activation function between the two fully connected layers. The multidimensional first feature vector is mapped through the two fully connected layers to obtain the original importance score corresponding to each feature channel. The original importance scores are then normalized by the nonlinear activation function to obtain the attention weight vector. The channel-wise multiplication unit is used to multiply the attention weight vector with the multidimensional first feature vector channel-wise to obtain a weighted feature vector; The normalization fusion unit is used to integrate the weighted feature vectors by channel dimension and to normalize and fuse the integrated features by the Softmax function to obtain the first fused feature.

[0087] In one embodiment, the enhancement processing module includes: A combination unit is used to combine the feature vectors in the first fused feature to obtain multiple feature vector groups; The mapping unit is used to map the three feature vectors in each feature vector group to coordinate points in a two-dimensional space, and to determine the local correlation strength of the feature vector group based on the area of ​​the triangle formed by the three coordinate points. The recombination unit is used to recombine the feature vectors based on the local correlation strength of each feature vector group to obtain locally enhanced features; The second fusion feature generation unit is used to generate the second fusion feature based on the local enhancement feature and the first fusion feature.

[0088] In one embodiment, the second fusion feature generation unit includes: The feature mapping subunit is used to map each feature vector in the local enhancement feature to feature points in a multi-dimensional space to obtain a set of feature points. The concave point removal subunit is used to remove concave points from the feature point set based on the Graham scan method to generate a feature distribution concave hull. A global distribution description feature generation subunit is used to generate global distribution description features based on the area value of the feature distribution hull and the density distribution information of each feature point in the feature point set; An update subunit is used to update the first fused feature based on the global distribution description feature and the local enhancement feature to obtain the second fused feature.

[0089] In one implementation, the updating subunit is specifically used for: The local enhancement features are concatenated with the global distribution description features to obtain a combined feature vector; The combined feature vectors are standardized, and multiple feature vector pairs are selected from the standardized combined feature vectors. Based on the cross product of the vectors between each of the aforementioned features, each cross product vector is obtained; The cross product vectors are combined sequentially to obtain the interaction matrix; Perform matrix multiplication on the normalized interaction matrix and the standardized combined feature vector to obtain the weighted adjusted features; The weighted features are subjected to a nonlinear transformation to obtain the nonlinearly transformed features; The features after nonlinear transformation are residually connected with the first fused features to obtain the second fused features.

[0090] In one implementation, the model training module includes: The forward propagation calculation unit is used to perform forward propagation calculation on the second fused feature through the device state prediction model to obtain a preliminary state prediction result; The prediction error unit is used to calculate the prediction error between the preliminary state prediction result and the true label using the cross-entropy loss function. The backpropagation algorithm calculation unit is used to calculate the gradient of the prediction error with respect to each network parameter in the device state prediction model through the backpropagation algorithm, and to update the weight parameters of the device state prediction model using the adaptive moment estimation algorithm. The convergence unit is used to repeatedly execute the forward propagation, loss calculation, backpropagation and weight parameter update process until the prediction error converges or the preset training rounds are reached to obtain the target prediction result.

[0091] The specific functions and examples of each module and submodule of the system in this embodiment of the invention can be found in the relevant descriptions of the corresponding steps in the above method embodiments, and will not be repeated here.

[0092] The acquisition, storage, and application of user personal information involved in the technical solution of this invention all comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0093] This invention also provides an electronic device, comprising: At least one processor; and a memory communicatively connected to said at least one processor; The memory stores instructions that can be executed by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform the method described in any one of the embodiments of the present invention.

[0094] The beneficial effects of the electronic device in this embodiment of the invention are equivalent to the beneficial effects of the above-described deep learning-based high-voltage cable equipment state prediction method, and will not be repeated here.

[0095] This invention also provides a non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to cause a computer to perform the method described in any one of the embodiments of this invention.

[0096] The beneficial effects of the storage medium of the present invention are equivalent to the beneficial effects of the above-described deep learning-based high-voltage cable equipment state prediction method, and will not be repeated here.

[0097] Figure 3 A schematic block diagram of an example electronic device 800 that can be used to implement embodiments of the present invention is shown. Electronic device 800 is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. Electronic device 800 may also represent various forms of mobile devices, such as personal digital assistants, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0098] like Figure 3 As shown, the electronic device 800 includes a computing unit 801, which can perform various appropriate actions and processes based on a computer program stored in a read-only memory (ROM) 802 or a computer program loaded from a storage unit 808 into a random access memory (RAM) 803. The RAM 803 may also store various programs and data required for the operation of the electronic device 800. The computing unit 801, ROM 802, and RAM 803 are interconnected via a bus 804. An input / output (I / O) interface 805 is also connected to the bus 804.

[0099] Multiple components in electronic device 800 are connected to I / O interface 805, including: input unit 806, such as keyboard, mouse, etc.; output unit 807, such as various types of displays, speakers, etc.; storage unit 808, such as disk, optical disk, etc.; and communication unit 809, such as network card, modem, wireless transceiver, etc. Communication unit 809 allows electronic device 800 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0100] The computing unit 801 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 801 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 801 performs the various methods and processes described above, such as the deep learning-based high-voltage cable equipment state prediction method. For example, in some embodiments, the deep learning-based high-voltage cable equipment state prediction method can be implemented as a computer software program tangibly contained in a machine-readable medium, such as storage unit 808. In some embodiments, part or all of the computer program can be loaded and / or installed on the electronic device 800 via ROM 802 and / or communication unit 809. When the computer program is loaded into RAM 803 and executed by the computing unit 801, one or more steps of the deep learning-based high-voltage cable equipment state prediction method described above can be performed. Alternatively, in other embodiments, the computing unit 801 may be configured by any other suitable means (e.g., by means of firmware) to perform a deep learning-based high-voltage cable equipment state prediction method.

[0101] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0102] The program code used to implement the methods of the present invention can be written in any combination of one or more programming languages. This program code can be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code can be executed entirely on the machine, partially on the machine, as a standalone software package partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0103] In the context of this invention, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable media can include, but are not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0104] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0105] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0106] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact via communication networks. Client-server relationships are created by computer programs running on the respective computers and having a client-server relationship with each other. Servers can be cloud servers, servers in distributed systems, or servers incorporating blockchain technology.

[0107] It should be understood that the various forms of processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this invention can be achieved, and this is not limited herein.

[0108] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the principles of this invention should be included within the scope of protection of this invention.

Claims

1. A method for predicting the condition of high-voltage cable equipment based on deep learning, characterized in that, include: Historical monitoring data of high-voltage cable equipment is acquired and reconstructed to obtain a training dataset; The features in the training dataset are fused to obtain the first fused feature; Based on the local correlation and distribution dispersion among the various features, the first fused feature is enhanced to obtain the second fused feature; The device status prediction model is trained based on the second fusion feature to obtain the target prediction model; The real-time monitoring data of the high-voltage cable equipment is used as the input to the target prediction model to obtain the state prediction result output by the target prediction model.

2. The method according to claim 1, characterized in that, The process of acquiring historical monitoring data of high-voltage cable equipment and reconstructing the historical monitoring data to obtain a training dataset includes: The historical monitoring data is classified to obtain various historical fault samples and various historical normal samples. Based on the operating condition records corresponding to each of the historical fault samples, multiple key feature dimensions corresponding to each of the historical fault samples are extracted. Each of the historical fault samples is treated as a sample point, and the sample points are filtered based on the Euclidean distance between them to obtain multiple neighborhood sample points. Random interpolation is performed on the connections between each of the sample points and each of the neighboring sample points to generate synthetic fault samples; The synthesized fault samples and each of the historical fault samples are used as inputs to a preset conditional generative adversarial network, and the synthesized fault samples are reconstructed under the constraint of multi-factor coupling conditions to obtain enhanced fault samples. The enhanced fault samples are mixed with each of the historical normal samples according to a preset ratio, and the distribution of the mixed samples is adjusted to achieve equalization, thereby obtaining the training dataset.

3. The method according to claim 1, characterized in that, The process of fusing the features in the training dataset to obtain the first fused feature includes: The samples in the training dataset are used as input to a preset neural network algorithm. Multi-scale deep features of the samples are extracted layer by layer through forward propagation to obtain a multi-dimensional first feature vector. The importance score of each feature channel is calculated for the multidimensional first feature vector through a channel attention mechanism. The channel attention mechanism includes at least two fully connected layers and a nonlinear activation function between the two fully connected layers. The original importance score of each feature channel is obtained after the multidimensional first feature vector is mapped through the two fully connected layers. The original importance score is then normalized by the nonlinear activation function to obtain the attention weight vector. The attention weight vector is multiplied channel by channel with the multidimensional first feature vector to obtain the weighted feature vector; The weighted feature vectors are integrated along the channel dimension, and the integrated features are normalized and fused using the Softmax function to obtain the first fused feature.

4. The method according to claim 1, characterized in that, The enhancement process for the first fused feature, based on the local correlation and distribution dispersion among the various features, yields the second fused feature, including: The feature vectors in the first fused feature are combined to obtain multiple feature vector groups; For each of the feature vector groups, the three feature vectors in the feature vector group are mapped to coordinate points in two-dimensional space, and the local correlation strength of the feature vector group is determined based on the area of ​​the triangle formed by the three coordinate points. Based on the local correlation strength of each of the feature vector groups, the feature vectors are recombined to obtain locally enhanced features; The second fusion feature is generated based on the local enhancement feature and the first fusion feature.

5. The method according to claim 4, characterized in that, The step of generating the second fusion feature based on the local enhancement feature and the first fusion feature includes: Each feature vector in the local enhancement features is mapped to a feature point in a multi-dimensional space to obtain a set of feature points; The feature point set is subjected to concave point removal based on the Graham scan method to generate a feature distribution concave hull; Based on the area value of the feature distribution concave hull and the density distribution information of each feature point in the feature point set, a global distribution description feature is generated; Based on the global distribution description features and the local enhancement features, the first fusion feature is updated to obtain the second fusion feature.

6. The method according to claim 5, characterized in that, The step of updating the first fused feature based on the global distribution description feature and the local enhancement feature to obtain the second fused feature includes: The local enhancement features are concatenated with the global distribution description features to obtain a combined feature vector; The combined feature vectors are standardized, and multiple feature vector pairs are selected from the standardized combined feature vectors. Based on the cross product of the vectors between each of the aforementioned features, each cross product vector is obtained; The cross product vectors are combined sequentially to obtain the interaction matrix; Perform matrix multiplication on the normalized interaction matrix and the standardized combined feature vector to obtain the weighted adjusted features; The weighted features are subjected to a nonlinear transformation to obtain the nonlinearly transformed features; The features after nonlinear transformation are residually connected with the first fused features to obtain the second fused features.

7. The method according to claim 1, characterized in that, The step of training the device status prediction model based on the second fusion feature to obtain the target prediction model includes: The device state prediction model is used to perform forward propagation calculation on the second fused feature to obtain a preliminary state prediction result. The prediction error between the preliminary state prediction result and the true label is calculated using the cross-entropy loss function. The gradient of the prediction error with respect to each network parameter in the device state prediction model is calculated using the backpropagation algorithm, and the weight parameters of the device state prediction model are updated using the adaptive moment estimation algorithm. Repeat the forward propagation, loss calculation, backpropagation and weight parameter update process until the prediction error converges or the preset training rounds are reached to obtain the target prediction result.

8. A high-voltage cable equipment condition prediction device based on deep learning, characterized in that, include: The data reconstruction module is used to acquire historical monitoring data of high-voltage cable equipment and reconstruct the historical monitoring data to obtain a training dataset. The fusion processing module is used to fuse the various features in the training dataset to obtain the first fused feature; An enhancement processing module is used to enhance the first fused feature based on the local correlation and distribution dispersion between the various features to obtain a second fused feature. The model training module is used to train the device state prediction model based on the second fusion feature to obtain the target prediction model; The state prediction module is used to take the real-time monitoring data of the high-voltage cable equipment as input to the target prediction model and obtain the state prediction result output by the target prediction model.

9. An electronic device, characterized in that, include: At least one processor; and a memory communicatively connected to the at least one processor; The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-7.

10. A non-transitory computer-readable storage medium storing computer instructions, characterized in that, The computer instructions are used to cause the computer to perform the method according to any one of claims 1-7.

Images