Power data analysis method based on multi-mode data processing and fusion

Abstract

The application discloses a power data analysis method based on multi-mode data processing and fusion, comprising the following steps: S1, constructing a knowledge graph; S2, inputting power data into the knowledge graph and analyzing and fusing the power data; wherein S2 comprises the following steps: S21, segmenting, matching and analyzing multi-modal and cross-media semantics; S22, segmenting and semantically analyzing power point cloud data; and S23, fusing multi-modal and cross-level power data. The application has the beneficial effect that the fusion of multi-modal and cross-level power data is realized by segmenting, matching and analyzing multi-modal and cross-media semantics and segmenting and semantically analyzing power point cloud data, so that the knowledge graph can be used.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of power systems, and in particular to a power data analysis method based on multi-modal data processing and fusion. BACKGROUND

[0002] With the continuous expansion of the power grid scale and the continuous addition of new energy and new equipment, the power source structure and the power grid pattern have undergone profound changes, the power system has become increasingly complex, and the safe and stable operation of the power grid faces great challenges. At the same time, under the background of digital transformation in power grid production, enterprise operation and customer service, it is urgent to use digital technology to comprehensively improve the holographic perception, efficient processing and sharing service capabilities of the power grid.

[0003] However, the current power grid business applications are evolving towards online, mobile, collaborative and intelligent, and therefore, it is urgent to overcome the difficulties of low efficiency in sharing, collaboration, interaction and decision-making between traditional modes and professional platforms, and to promote the transformation of government and enterprise decision-making to a digital-driven decision-making mode. In practice, there are four problems: data multi-source heterogeneity, data time granularity dispersion, large geographical and business space span, privacy protection and data security.

[0004] With the continuous development of artificial intelligence technology, intelligent retrieval and analysis methods based on knowledge graphs are gradually applied in the field of smart grids. On the basis of the knowledge graph, further construction of a power grid big data intelligent service platform can break through the interaction channel of data flow and business flow between operation and maintenance, marketing, dispatching and power supply service professional platforms, and fully realize the fusion of business flow and data flow, thereby realizing the full "perception" of power grid data and promoting and developing innovative businesses.

[0005] The Chinese patent "A semantic matching method and device for power transformer knowledge question answering", publication number CN113919366A, publication date January 11, 2022, discloses the application of a knowledge graph in the power industry. However, power business data has characteristics such as multi-source, heterogeneity and multi-modal, and how to fuse multi-source heterogeneous data is the key to knowledge question answering and is also the focus and difficulty of current research on subject knowledge service and intelligent knowledge discovery in the field of power engineering. SUMMARY

[0006] In order to solve the shortcomings and deficiencies in the prior art, the application provides a power data analysis method based on multi-modal data processing and fusion, which fuses multi-source heterogeneous data in the power system and realizes the full "perception" of power grid data.

[0007] To achieve the above technical purposes, the power data analysis method based on multi-modal data processing and fusion provided by the present application comprises: S1: constructing a knowledge graph; S2: inputting power data into the knowledge graph and analyzing and fusing the power data; wherein S2 comprises: S21: segmenting, matching and analyzing the semantics of multi-modal and cross-media; S22: segmenting and semantically analyzing power point cloud data; S23: fusing multi-modal and cross-level power data.

[0008] Optionally, the knowledge graph comprises: a basic data layer containing various types of structured and unstructured data involved in fault handling; a graph construction layer extracting relevant knowledge from the corpus to form a structured knowledge network; an information analysis layer analyzing and structuring real-time signals received during power grid dispatching and matching, retrieving or extracting relevant data and knowledge from the graph database of the knowledge graph; an inference decision layer using knowledge inference methods to query, analyze and process the structured fault information formed based on various dispatching experiences and rules accumulated over a long period of time; and a human-computer interaction layer pushing friendly and understandable structured knowledge based on the knowledge graph to display and remind key and implicit information, check dispatching operations and review historical experiences.

[0009] Optionally, S21 comprises: a lightweight image semantic segmentation network based on multi-level features in parallel, which is used as a benchmark network structure; a hollow residual enhancement module and a deep hollow spatial pyramid module, which are used to enhance the feature information of different levels.

[0010] Optionally, S21 further comprises: an end-to-end fully convolutional neural network model for fusing several scale features to restore image details; an efficient lossless shape coding algorithm based on contour and chain code representation for contour extraction, segmentation, coding and compression; a training feature mapping network and a modal discrimination network, which are constructed using an adversarial learning method framework to analyze the correlation under several modes.

[0011] Optionally, S22 comprises: a point cloud filtering and processing method, including a point cloud filtering algorithm combining double tensor voting and multi-scale normal vector estimation to filter power point cloud data.

[0012] Optionally, S22 further comprises: a filtering method based on a segmented energy function optimization, which gridded the point cloud data and designed a segmented function to obtain the corresponding ground elevation; a mountain point cloud filtering method to eliminate environmental interference through the echo characteristics of vegetation point cloud data and PTD algorithm.

[0013] Optionally, S22 further comprises: a fast three-dimensional point cloud semantic segmentation method based on clustering to construct a two-stage point cloud semantic segmentation framework; a point cloud scene re-identification based on semantic graph representation to describe the scene at the semantic level.

[0014] Optionally, S22 further comprises: a point cloud encoding method based on the compression sensing theory, to compress and encode the power point cloud data; a point cloud data sparse representation model and an encoding and reconstruction model based on an over-complete dictionary, to normalize the power point cloud data.

[0015] Optionally, S23 comprises: an incomplete multi-modal data fusion algorithm based on deep semantic matching, to perform deep correlation fusion of the incomplete multi-modal data.

[0016] Optionally, S23 further comprises: a fusion method for incremental multi-modal data clustering, to define a multi-modal data similarity measurement standard and perform parameter-free incremental clustering fusion of the multi-modal data; a fusion method for heterogeneous modal data migration, comprising: a heterogeneous modal data migration fusion algorithm based on multi-layer semantic matching, to complete the transfer learning from the source modal knowledge to the target modal task; a fusion method for multi-modal data low-dimensional sharing, to obtain robust cross-modal shared fusion features in a low-dimensional latent subspace.

[0017] Advantages of the present application:

[0018] (1) The multi-source, heterogeneous, multi-modal data intelligent processing and fusion technology for the new power system is proposed.

[0019] (2) The new method of cross-media hierarchical semantic learning based on semantic concepts is proposed, and the multi-modal, cross-level power data intelligent fusion technology based on deep learning and semantic driving is realized, which effectively improves the data fusion efficiency and performance. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 The knowledge graph framework in the case of an embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0021] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be further described in detail below in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only the best embodiment of the present application, which are used to explain the present application and do not limit the protection scope of the present application. All other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0022] Knowledge graphs, as a knowledge organization and construction method based on artificial intelligence technology, express information in a way that more closely resembles human cognition of the world. They can represent complex relationships at the semantic level, providing a better ability to manage and understand massive amounts of information. Knowledge graphs in the power sector aim to fully utilize the data information carried by the power Internet of Things (IoT) to structurally depict concepts, entities, events, and their relationships within the power system. This provides the power industry chain with a more effective cross-media big data organization, management, and cognitive capability. Constructing knowledge graphs in the power sector can fully explore the value of diverse and heterogeneous data from the power grid, and to some extent solve problems such as low accuracy and poor timeliness in business processing caused by differences and deficiencies in the knowledge reserves of various professionals in the power sector. It is an effective way to improve knowledge sharing, collaborative processing capabilities among professionals, and efficiency.

[0023] Specifically, the power data analysis method based on multi-mode data processing and fusion proposed in this application includes: S1: constructing a knowledge graph;

[0024] S2: Input power data into the knowledge graph to analyze and integrate the power data.

[0025] Among them, such as Figure 1 As shown, the knowledge graph comprises: a basic data layer, a graph construction layer, an information parsing layer, a reasoning and decision-making layer, and a human-computer interaction layer. The basic data layer mainly contains various structured and unstructured (semi-structured) data involved in fault handling, serving as the foundational corpus for graph construction. The graph construction layer extracts relevant knowledge from the corpus to form a structured knowledge network. The information parsing layer parses and structures the real-time signals received during power grid dispatching, and matches, retrieves, or extracts relevant data and knowledge from the knowledge graph's graph database. The reasoning and decision-making layer uses knowledge reasoning methods to query, analyze, and process the structured fault information based on long-term accumulated dispatching experience and rules. The human-computer interaction layer pushes user-friendly and understandable structured knowledge based on the knowledge graph, thereby displaying and reminding key and implicit information, verifying dispatching operations, and reviewing historical experience.

[0026] As a more specific solution, S2 includes:

[0027] S21: Segment, match, and analyze the semantics of multimodal and cross-media communication;

[0028] S22: Segmentation and semantic analysis of 3D point cloud data of power systems;

[0029] S23: Fusion of multimodal, cross-level power data.

[0030] In the S21 step, when the network parameter quantity and the calculation quantity are too large, the timeliness of system feedback processing is affected, a multi-level feature parallel lightweight image semantic segmentation network, a hollow residual enhancement module, a deep hollow spatial pyramid module, an end-to-end full convolutional neural network model, and an efficient lossless shape coding algorithm based on contour and chain code representation are used to segment, match and analyze the multi-modal and cross-media semantics.

[0031] The multi-level feature parallel lightweight image semantic segmentation network selects a lightweight feature extraction network as a benchmark network structure, thereby improving the problem of large parameter quantity and large calculation quantity caused by the current semantic segmentation algorithm which generally uses a network with deep network depth and wide network width as a feature extraction benchmark network. At the same time, considering the segmentation accuracy of the network, the hollow residual enhancement module and the deep hollow spatial pyramid module are used in cooperation. The hollow residual enhancement module is composed of two residual branches, which are used to strengthen the edge contour information of different scale targets of shallow features. The deep hollow spatial pyramid module is composed of deepened hollow convolution with different receptive field sizes, which is used to strengthen the semantic information of different scale targets of deep features. Under the mutual influence of the multi-level feature parallel lightweight image semantic segmentation network and the hollow residual enhancement module and the deep hollow spatial pyramid module, the network parameter quantity and the calculation quantity can be reduced, the segmentation accuracy can be improved through deep and shallow hierarchical processing, the semantic segmentation accuracy and real-time performance are considered, thereby adapting to the real-time and accurate demand of large data quantity calculation in the power environment.

[0032] At the same time, the end-to-end full convolutional neural network model is used, the symmetric coding-decoding structure is used, the maximum pooling layer is used, the encoding part is used for down-sampling to extract multi-scale features, the decoding part uses a nonlinear up-sampling method, layer-by-layer up-sampling is used to fuse different scale features to restore image details, and a weighted loss function is used to train the model to compensate for the imbalance between the target and the background, improve the algorithm efficiency, and reduce the environmental interference factors in complex scenes, so as to adapt to the current complex power environment.

[0033] Considering the need for efficient and accurate semantic object encoding and representation, lossless shape compression is required. An efficient lossless shape coding algorithm based on contour and chain code representation is adopted, which is divided into contour extraction and segmentation and coding and compression. The algorithm first extracts the object contour and thins it to single-pixel width. Then it is converted into chain code representation and the target contour curve is segmented into several sub-segments based on the direction correlation of chain code, so that each sub-segment contains only two basic direction codes. Next, linear detection is performed on each contour sub-segment, and it is divided into ordinary sub-segment and straight line segment sub-segment. Finally, the length, type and direction code sequence of each sub-segment are encoded respectively, and the run length encoding is performed on the straight line segment part. Since there are only two basic direction codes in each sub-segment, each direction code only needs one bit to represent, so high coding efficiency can be achieved. For a given original binary shape image, the contours of all objects are first extracted. Since the extracted object contours are often not single-pixel wide, it is not conducive to subsequent efficient coding and compression. The extracted contour is further thinned to single-pixel width based on 8-adjacency. That is, for each edge point on the object contour, there are only 2 adjacent contour edge points in its 8-neighborhood except for the starting point and contour intersection point. After the target contour is extracted and thinned, it is further converted into chain code representation using the classic 8-direction Freeman chain code. Since the 8-direction chain code exactly matches the actual adjacency of the pixel points in the image, it can accurately describe the information of the center pixel point and its adjacent points. The contour intersection points are selected as feature points, and the entire contour curve is divided into several segments based on the feature points, so that each segment of the contour edge has zero or two feature points. If a segment of the object contour is a closed curve, it has no feature points, otherwise it contains two end-point feature points, thereby realizing efficient representation and coding of the shape. Further, based on the type of direction code, each edge contour is divided into several sub-segments, so that there are at most two types of direction codes in each sub-segment. Thus, any direction code in the sub-segment only needs one bit to encode and represent, compared with the original chain code which requires 3 bits for each direction code, which can effectively improve the compression efficiency. For each sub-segment in the segmented contour segment, if it is a straight line segment, run length encoding is performed, and its starting point and length are encoded respectively. If it is a non-straight line segment, the type, length and direction code of each link in the sub-segment need to be encoded respectively. Through the above scheme, the semantic object encoding and representation have the characteristics of high efficiency, lossless and accuracy at the same time.

[0034] In the fusion, because the data characteristics of different modalities cannot be directly compared when the cross-modal data is fused, an adversarial learning method framework is adopted to construct a feature mapping network and a modal discrimination network, so that different modal data under the same semantics has a small distance in the space, and the same modal data under different semantics has a large distance. The semantic distribution and similarity are used as the training basis for the feature mapping network, and the modal discrimination network is responsible for determining the modal of different data in the space. Based on the adversarial learning, the two networks are alternately trained, so that the data obtained by the feature mapping network is consistent with the original data in semantics, and the modal characteristics are eliminated. Finally, the similarity is used for correlation analysis in the same space.

[0035] Based on the above, the algorithm complexity is reduced, the system real-time performance and accuracy are improved, and the multi-modal and cross-media semantics are segmented, matched and analyzed.

[0036] In the S22 step, considering the problem of large noise of power point cloud data, an efficient point cloud filtering and processing method is adopted. First, a point cloud filtering algorithm combining double tensor voting and multi-scale normal vector estimation is used. The principal component analysis method is used to estimate the normal vector of each point with a larger scale, and the feature points are extracted by double tensor voting. The normal vector of the extracted feature points is estimated with a smaller scale, and a random sampling method is used to remove small-range noise planes. Curvature is used to filter the remaining noise to obtain the final point cloud data, which can effectively remove noise points and better preserve the sharp features of the three-dimensional model, laying a foundation for subsequent point cloud registration and three-dimensional reconstruction.

[0037] At the same time, in the S22 step, considering the complexity of the power environment, a filtering method based on piecewise energy function optimization is adopted to remove noise and grid the point cloud data. A piecewise energy function is designed, and the ground elevation corresponding to each grid is calculated by energy minimization. If the difference between the lowest point elevation in the grid and the elevation value is within the quantization error range, the point is set as a ground seed point. According to the seed point, a Delaunay triangular network is constructed, and the points within the threshold range are taken as ground points. This method has a simple algorithm and is easy to set parameters, and takes into account the accuracy and efficiency.

[0038] Further, since the power environment is also possible in high mountainous areas, the terrain of such a hydropower project area is large in relief, densely vegetated, and there are some buildings, and when extracting the terrain of these areas, the airborne LiDAR point cloud filtering method has the problem of low precision, in the S22 step, a mountain point cloud filtering method is also adopted which comprehensively considers the echo characteristics of point clouds, the progressive triangulated network encryption (PTD) algorithm, the improved surface fitting algorithm and the fine processing of ground points, which, on the basis of denoising of LiDAR point cloud data, first removes part of the vegetation points using the echo characteristics of the vegetation point cloud, then uses the PTD algorithm to perform two iterations to obtain a set of partial ground points, then uses the set of partial ground points obtained as seed points for the improved surface fitting algorithm to perform grid regionalization surface fitting to obtain the ground points in the original point cloud data, and finally removes the low vegetation points mixed in the ground points through fine processing of the point cloud, thereby obtaining the final set of ground points, reducing the influence of the regional terrain on point cloud denoising, and improving the overall system precision.

[0039] In the S22 step, a fast three-dimensional point cloud semantic segmentation method based on clustering and a point cloud scene re-identification based on semantic graph expression are also included for point cloud semantic segmentation. Through a two-stage point cloud semantic segmentation framework, the problems of traditional segmentation methods in obtaining semantics and point cloud networks in direct application to large scene point clouds are overcome, the advantages of the two are effectively combined, the system calculation amount is reduced, and the segmentation accuracy and efficiency are improved. At the same time, through topological graph expression, the point cloud semantic information and semantic relationship are effectively utilized to describe the scene at the semantic level, and the scene similarity is evaluated through a graph similarity measurement network. By introducing point cloud semantic information that is more robust to scene environment transformation, the precision, robustness and generalization performance of the algorithm are improved.

[0040] Considering the difficulty of reducing the huge point cloud data for system storage, transmission and processing, a point cloud coding method based on the compression sensing theory is provided, a three-dimensional point cloud data normalization method is established by utilizing the local spatial similarity of three-dimensional point cloud data, a point cloud data sparse representation model and coding and reconstruction model based on an over-complete dictionary are proposed. By studying the geometric space characteristics and local similarity characteristics of the three-dimensional point cloud model, a K-nearest neighbor-based point cloud data normalization method is proposed, which effectively utilizes the local spatial similarity of point cloud data, improves the similarity of three-dimensional point cloud data in coordinate values, and provides an important guarantee for sparse representation of three-dimensional point cloud data. At the same time, considering the self-similarity between the normalized point cloud data, a K-SVD-based dictionary training algorithm is proposed to obtain the sparse representation basis of the normalized point cloud data, so that the normalized point cloud data can be sparsely represented under the over-complete dictionary.

[0041] In practical applications, multi-modal data has the problems of incomplete modal, unbalanced modal, and difficult semantic expression. In step S23, a multi-modal and cross-level power data intelligent fusion method based on deep learning and semantic driving is included.

[0042] In the method, an incomplete multi-modal data fusion algorithm based on deep semantic matching is used. The high-level semantic abstraction characteristics of the deep learning network are utilized to design a unified deep model coupling modal private deep network and incomplete modal shared feature learning, to realize deep correlation fusion of incomplete multi-modal data and reduce the semantic deviation of modal shared features. Based on the geometric characteristics of the modal space, a local invariant graph regularization factor is designed to couple the multi-modal shared features in the subspace and the original modal features, further improving the accuracy of the fusion result.

[0043] Further, an incremental multi-modal data clustering-oriented fusion method is used, including a parameter-free multi-modal data incremental co-clustering fusion algorithm and an adaptive modal weight updating mechanism. By defining a new multi-modal data similarity measurement standard and designing three incremental clustering strategies, namely cluster creation, cluster merging and instance division, parameter-free incremental clustering fusion of multi-modal data is performed, improving the algorithm efficiency while ensuring the robustness of the fusion result. Through the adaptive modal weight updating mechanism, the modal weight is dynamically adjusted during the co-clustering fusion process, meeting the dynamic change demand of the influence of the modal on the fusion result and improving the scalability of the algorithm.

[0044] As a more specific scheme, a heterogeneous modal data migration-oriented fusion method is used, including a heterogeneous modal data migration fusion algorithm based on multi-level semantic matching, coupling modal deep network and modal semantic related model, and designing a unified deep network architecture based on multi-level semantic matching. Through the related matching of the cross-modal features of each layer, the semantic deviation between heterogeneous modalities is gradually reduced. The maximum correlation of the item layer output features is used to optimize and adjust the modal network as a whole, further improving the correlation of the modal deep semantics. A new objective function is defined to jointly optimize the heterogeneous modal deep matching network, obtaining a cross-modal high-level semantic fusion subspace, and completing the source modal knowledge to target modal task migration learning in the subspace.

[0045] And a multi-modal data low-dimensional sharing-oriented fusion method is used, a co-learning model of modal private (irrelevant or negatively correlated) features and cross-modal shared (correlated) features is designed, the accuracy of low-dimensional shared feature representation is improved by separating modal private features. The coupling of shared features is used to establish a joint optimization objective function of each modal, and the modal invariant graph regularization and projection matrix sparsification auxiliary model are used to optimize the process, further improving the accuracy of the fusion result. Through iterative related and unrelated feature co-training and updating, robust cross-modal shared fusion features in the low-dimensional latent subspace are obtained.

[0046] Considering the fusion of laser point cloud and visible light image can combine the advantages of both, a laser point cloud and image fusion algorithm based on feature matching of depth-echo intensity image and visible light image is adopted, the laser point cloud is projected on a two-dimensional image plane, and is converted into a depth-echo intensity image; then, feature point matching is performed with the visible light image to obtain the mapping relationship between the two, and then the fusion of laser point cloud and visible light image is realized; the kernel Fisher discriminant analysis is integrated into the SIFT algorithm, the process of generating the feature descriptor is improved, and the robustness of the algorithm is improved.

[0047] As Figure 1 shown, it is a knowledge graph framework in the present application. In entity extraction, the present application adopts a feature-based method and a neural network method. The former uses a pre-labeled corpus to train a model, so that the model learns the probability of a certain word or character as a part of an entity, and then calculates the probability value of a candidate field as an entity; the neural network model can automatically represent and capture effective features from the text, and then complete entity recognition.

[0048] In view of the problems of non-standard natural language expression and different expressions of entity names in various business systems in the power field, a name extraction method based on conditional random field is adopted to extract entity names from text data. The actual obtained text can get the label of each single character through this trained model, that is, it is not necessary to manually set the label for each text, so that the device information involved in the dispatching text can be obtained through the model, and then the power grid text data can be conveniently searched and statistically analyzed according to the device name.

[0049] In the process of knowledge fusion and processing, the device entity of the power grid has a clear definition and naming in the database of the dispatching system, which can be used as the main standard for knowledge merging. Through automatic means, the existing system database is converted into a triple form of resource description framework (RDF), and other forms of extracted knowledge elements are fused with it. For quality evaluation, a method combining expert manual evaluation and automatic evaluation is adopted, quality evaluation functions are defined according to different business needs, and comprehensive evaluation is performed to determine the final quality score of the knowledge graph. Through the credibility quantitative evaluation process of the new knowledge before it is added to the knowledge base, the knowledge errors or conflicts are eliminated.

[0050] The above specific embodiments are the preferred embodiments of the power data analysis method based on multi-modal data processing and fusion of the present application, and do not limit the specific implementation range of the present application. The scope of the present application includes but is not limited to the specific embodiments, and equivalent changes made according to the shape and structure of the present application are within the protection scope of the present application.

Claims

1. A power data analysis method based on multi-mode data processing and fusion, characterized in that: include: S1: Construct a knowledge graph; S2: Input power data into the knowledge graph, and analyze and integrate the power data; S2 includes: S21: Segment, match, and analyze the semantics of multimodal and cross-media communication; S22: Segment and semantic analysis of power point cloud data; S23: Fusion of multimodal, cross-level power data; S21 includes: A lightweight image semantic segmentation network based on multi-level feature parallelism is used as the baseline network structure. The void residual enhancement module and the deep void spatial pyramid module are used to enhance feature information at different levels. An end-to-end fully convolutional neural network model is used to fuse features at several scales to recover image details; An efficient and lossless shape encoding algorithm based on contour and chain code representation is used for contour extraction, segmentation, encoding, and compression. The feature mapping network and the modality discrimination network are trained, and an adversarial learning framework is used to construct a correlation analysis under several modalities. S22 includes: A point cloud filtering and processing method, including a point cloud filtering algorithm combining dual tensor voting and multi-scale normal vector estimation, to filter the power point cloud data; A filtering method based on piecewise energy function optimization is used to grid point cloud data and design piecewise functions to obtain the corresponding ground elevation. A mountain point cloud filtering method is proposed to eliminate environmental interference by utilizing the echo characteristics of vegetation point cloud data and the PTD algorithm. A fast 3D point cloud semantic segmentation method based on clustering is proposed to construct a two-stage point cloud semantic segmentation framework. Point cloud scene re-identification based on semantic graph representation, to describe the scene at the semantic level; A point cloud encoding method based on compressed sensing theory is used to compress and encode the power point cloud data. The power point cloud data is normalized based on an overcomplete dictionary-based sparse representation model and encoding and reconstruction model. S23 includes: A deep semantic matching-based algorithm for incomplete multimodal data fusion is proposed to perform deep correlation fusion of incomplete multimodal data. A fusion method for incremental multimodal data clustering is proposed, which defines a similarity metric for multimodal data and performs parameter-free incremental clustering fusion on multimodal data. Fusion methods for heterogeneous modal data transfer include: a heterogeneous modal data transfer fusion algorithm based on multi-layer semantic matching, which completes the transfer learning from source modal knowledge to target modal task; A fusion method for low-dimensional sharing of multimodal data is proposed to obtain robust cross-modal sharing fusion features in low-dimensional latent subspaces.

2. The power data analysis method based on multi-mode data processing and fusion as described in claim 1, characterized in that: The knowledge graph includes: The basic data layer contains various structured and unstructured data involved in fault handling; The graph construction layer extracts relevant knowledge from the corpus to form a structured knowledge network; The information parsing layer parses and structures the real-time signals received during the power grid dispatching process, and matches, retrieves, or extracts relevant data and knowledge from the graph database of the knowledge graph. The reasoning and decision-making layer, based on various scheduling experiences and rules accumulated over a long period of time, uses knowledge reasoning methods to query, analyze, and process the structured fault information formed. The human-computer interaction layer pushes user-friendly and understandable structured knowledge based on knowledge graphs, thereby displaying and reminding key and implicit information, verifying scheduling operations, and reviewing historical experience.

Images