Data processing method and device for electric power tower

Abstract

The invention discloses a data processing method and device for an electric power tower, and belongs to the technical field of electric power towers. The method comprises the following steps: firstly, acquiring multi-source data of a power tower, extracting associated data according to components, generating multi-modal fusion data, solving the problems of dispersion of the multi-source data and low association degree of the components, and realizing data aggregation taking the components as the core; preprocessing the fused data to obtain standardized data, eliminating format and unit differences, and improving consistency; then, a multi-dimensional index system is generated, type checking and warm restoration are conducted, rapid retrieval is supported, and it is guaranteed that data are legal and accurate; and finally, storing the index and the repaired data in a structural health database for an AI model training platform and the like to efficiently call by borrowing the index, so that the multi-platform data requirement is met, the data utilization value is improved, and the digital management and intelligent application of the power tower are promoted.

Description

Technical Field

[0001] This application relates to the field of power pole technology, and in particular to a data processing method and apparatus for power poles. Background Technology

[0002] With the expansion of power transmission line coverage and the improvement of intelligent levels, power poles, as critical infrastructure of power transmission lines, increasingly rely on digital means for structural modeling, operation monitoring, and maintenance management. However, current power pole data processing has significant shortcomings: existing pole modeling largely relies on static drawings or simplified models, making it difficult to acquire and integrate multi-source data covering categories such as 3D spatial coordinates and sensor monitoring data. Furthermore, it lacks dedicated data processing logic for each component of the power pole, failing to extract component-related data information from multi-source data and generate corresponding multimodal fusion data, resulting in fragmented data dimensions. Simultaneously, existing processing methods do not effectively preprocess the multimodal fusion data corresponding to components to obtain standardized data, nor do they generate a multidimensional index system for components based on standardized data, nor perform type checking and Windsor-style repair, resulting in insufficient data quality and structure. In addition, there is a lack of a mechanism to store the multidimensional index system of components and the repaired standardized data in a structural health database, hindering the development of AI. Model training platforms, analysis systems, and visualization engines struggle to efficiently access data, failing to meet the demands for high-quality, reusable data in scenarios such as pole structure identification, condition assessment, and digital twin modeling. This severely restricts the digital and intelligent development of power poles. Summary of the Invention

[0003] This application provides a data processing method and apparatus for power poles, aiming to solve the problem that the inability to meet the demand for high-quality, reusable data in scenarios such as pole structure identification, condition assessment, and digital twin modeling has seriously restricted the digital and intelligent development of power poles.

[0004] In a first aspect, this application provides a data processing method for power poles, the method comprising: Acquire multi-source data on power poles; For each component of the power pole, data information associated with the component is extracted from the multi-source data, and multimodal fusion data corresponding to the component is generated based on the data information associated with the component. For each component of the power pole, the standardized data after repair is obtained based on the multimodal fusion data corresponding to the component. The multidimensional index system and the repaired standardized data corresponding to each component of the power pole are stored in a preset structural health database so that the AI ​​model training platform, analysis system and visualization engine can use the multidimensional index system corresponding to the component to call the repaired standardized data of the component.

[0005] Secondly, this application provides a data processing device for power poles, the device comprising: The first unit is used to acquire multi-source data on power poles; The second unit is used to extract data information associated with each component of the power pole from the multi-source data, and to generate multimodal fusion data corresponding to the component based on the data information associated with the component. The third unit is used to preprocess the multimodal fusion data corresponding to each component of the power pole to obtain the standardized data corresponding to the component. The fourth unit is used to generate a multi-dimensional index system corresponding to each component of the power pole based on the standardized data corresponding to the component, and to perform type checking and Windsorization repair on the standardized data to obtain the repaired standardized data. The fifth unit is used to store the multi-dimensional index system and the repaired standardized data corresponding to each component of the power pole into a preset structural health database, so that the AI ​​model training platform, analysis system and visualization engine can use the multi-dimensional index system corresponding to the component to call the repaired standardized data of the component.

[0006] Thirdly, this application provides a readable medium including executable instructions, which, when executed by a processor of an electronic device, cause the electronic device to perform any of the methods described in the first aspect.

[0007] Fourthly, this application provides an electronic device including a processor and a memory storing execution instructions, wherein when the processor executes the execution instructions stored in the memory, the processor performs the method as described in any of the first aspects.

[0008] As can be seen from the above technical solution, the data processing method for power poles provided in this application obtains multi-source data of power poles, first extracts related data information for each component and generates multimodal fusion data, effectively solving the problems of scattered multi-source data and low correlation with components in traditional processing, realizing multi-source data aggregation with components as the core, laying the foundation for subsequent accurate processing; then, based on the multimodal fusion data corresponding to the component, the repaired standardized data corresponding to the component is obtained, which not only provides index support for rapid data retrieval, but also ensures the legality and accuracy of standardized data; finally, the multidimensional index system and the repaired standardized data are stored in the structural health database, enabling the AI ​​model training platform, analysis system and visualization engine to efficiently call the required component data with the help of the multidimensional index system, which not only meets the needs of AI model training for high-quality data, but also provides reliable data support for the structural analysis of the analysis system and the visualization display of the visualization engine, greatly improving the utilization value of power pole data and promoting the development of digital management and intelligent application of power poles.

[0009] The further effects of the aforementioned non-conventional preferred method will be explained below in conjunction with specific embodiments. Attached Figure Description

[0010] To more clearly illustrate the embodiments of this application or the existing technical solutions, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 A flowchart illustrating a data processing method for power poles provided in this application; Figure 2 A flowchart illustrating a data processing method for power poles provided in this application; Figure 3 A flowchart illustrating a data processing method for power poles provided in this application; Figure 4 A flowchart illustrating a data processing method for power poles provided in this application; Figure 5 A flowchart illustrating a data processing method for power poles provided in this application; Figure 6 A flowchart illustrating a data processing method for power poles provided in this application; Figure 7 A schematic diagram of the structure of a data processing device for power poles provided in this application; Figure 8 This is a schematic diagram of the structure of an electronic device provided in this application. Detailed Implementation

[0012] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0013] The various non-limiting embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0014] The inventors' research revealed that with the expansion of transmission line coverage and the improvement of intelligence levels, the reliance on digital methods for structural modeling, operation monitoring, and maintenance management of power poles, as key infrastructure of transmission lines, is becoming increasingly urgent. However, current power pole data processing has significant shortcomings: existing pole modeling largely relies on static drawings or simplified models, making it difficult to acquire and integrate multi-source data covering categories such as 3D spatial coordinates and sensor monitoring data. Furthermore, it lacks dedicated data processing logic for each component of the power pole, failing to extract component-related data information from multi-source data and generate corresponding multimodal fusion data, resulting in fragmented data dimensions. Simultaneously, existing processing methods do not effectively preprocess the multimodal fusion data corresponding to components to obtain standardized data, nor do they generate a multidimensional index system for components based on standardized data, nor perform type checking and Windsor-style repair, resulting in insufficient data quality and structure. In addition, there is a lack of a mechanism to store the multidimensional index system of components and the repaired standardized data in a structural health database, hindering the development of AI. Model training platforms, analysis systems, and visualization engines struggle to efficiently access data, failing to meet the demands for high-quality, reusable data in scenarios such as pole structure identification, condition assessment, and digital twin modeling. This severely restricts the digital and intelligent development of power poles.

[0015] See Figure 1 This application illustrates a data processing method for power poles in an embodiment of the present application. The method includes the following steps: S101: Acquire multi-source data of power poles.

[0016] In this embodiment, the multi-source data of the power pole includes: three-dimensional spatial coordinates, sensor monitoring data, material properties, spatial location information, environmental load and historical operation and maintenance records; the multi-source data is extracted from structured or semi-structured files, which include at least one of the following formats: PDF, CSV, TXT, JSON format, tables obtained by parsing images or PDFs, and parametric component information.

[0017] Specifically, this embodiment extracts multi-source data of power poles from the following structured or semi-structured files: Three-dimensional spatial coordinates: obtained from CAD design drawings (PDF format), including the three-dimensional coordinates (unit: m) of components such as tower legs and crossarms, such as the coordinates of the center point of the tower leg foundation (X=5002.34, Y=3201.56, Z=100.20). Sensor monitoring data: extracted from the monitoring system log in CSV format, including tower strain values ​​(unit: με) collected by strain sensors and crossarm deflection angles (unit: °) collected by tilt sensors, with a sampling frequency of 10Hz; Material properties: Extracted from material manuals in TXT format, such as yield strength (unit: MPa) and elastic modulus (unit: GPa) of angle steel. Spatial location information: obtained from GIS system data in JSON format, including the latitude and longitude (unit: °) and altitude (unit: m) of the tower. Environmental loads: extracted from the structured database of the weather station, including wind speed (unit: m / s) and icing thickness (unit: mm). Historical maintenance records: These are parsed from PDF maintenance reports and contain text information such as component repair time and replacement records.

[0018] The above data can be extracted in batches using automated parsing tools (such as PDF table extractors and JSON parsers) to ensure the integrity of the original data.

[0019] The specific implementation may include the following sub-processes: S1 Data Source Determination and Format Routing Distribution The system first identifies the type of the input file, determining whether it is text or image data, and then initiates the corresponding parsing process accordingly. This process relies on both the file header signature and file extension for verification, ensuring that the correct parsing path is selected.

[0020] S2 Media File Processing (Images and PDFs) The system employs a method based on projection histograms and connected component analysis to identify page structure in PDFs or images, dividing them into table areas and text areas. For documents with dense structures, the Docstrum algorithm or lightweight graph neural networks such as the LayoutLMv2 model can be used to enhance layout understanding.

[0021] Within the segmented table regions, an algorithm based on morphological operations and Hough line transform is used to identify grid lines, extract cell boundaries, and construct a two-dimensional cell coordinate matrix. For frameless tables, a YOLO table detector (such as TableNet) combined with an affine alignment strategy can be used for segmentation.

[0022] In the obtained text area, the identified text area is sent to the OCR engine for text recognition. Based on the recognition results, the structural component features of the single-line diagram in the two-dimensional drawing are analyzed: through geometric pattern matching and depth image recognition algorithms, the bar intersection type, cross section type and tower head type are parametrically identified and encoded into standardized fields to form a structural feature dataset that can be used for subsequent index construction and topological association analysis.

[0023] S3 is used for text file processing (CSV / TXT). Field parsing: The system parses text using regular expressions and an adaptive delimiter recognition strategy, prioritizing common delimiters (such as commas, tabs, and spaces) and automatically selecting the optimal delimiter mode based on the criterion of minimum field variation.

[0024] Structural reconstruction: Field localization and structural restoration are achieved through heuristic algorithms based on row and column alignment or semantic header recognition (such as using BiLSTM+Attention network to extract header fields).

[0025] S4 Semantic Awareness and Table Content Unification All information from image or text paths is ultimately mapped into a preliminary table and sent to the downstream semantic analysis module. If there are mixed units or inconsistent field names in the original fields, the system will perform normalization processing based on the custom mapping table and rule base (e.g., replacing mm and cm with m).

[0026] S102: For each component of the power pole, extract data information associated with the component from the multi-source data, and generate multimodal fusion data corresponding to the component based on the data information associated with the component.

[0027] As an example, data information associated with the component can be extracted from the multi-source data based on the component's unique identifier or spatial anchor information. Then, semantic unification of the fields in the component-associated data information can be performed to eliminate naming differences between fields from different sources. Next, the attribution relationship between the data information and the component can be determined through spatial anchoring or explicit anchoring. Spatial anchoring is achieved through spatial neighborhood queries, setting a spatial threshold; if the distance between the data point coordinates and the component's center point coordinates is less than the spatial threshold, the data information is determined to belong to the component. Explicit anchoring involves directly matching the explicit component number contained in the data information with the component's map. Finally, using the component's unique identifier as the primary key, all data information belonging to the component can be aggregated to form multimodal fusion data corresponding to the component.

[0028] For example, for each component of a power pole (such as tower leg A, tower body segment B, crossarm C, etc.), the associated data can be extracted and multimodal fusion data can be generated in the following way.

[0029] 1. Data Association Extraction: Based on the unique identifier of the component (such as "Tower Leg A-2023-001") or spatial anchor point information (such as the coordinates of the tower leg center point), filter the associated information from the acquired multi-source data. For example, for "Crossarm C", extract its three-dimensional coordinates, strain monitoring data, material properties and historical maintenance records.

[0030] 2. Unified field semantics: Eliminate naming differences between data from different sources. For example, unify the naming of "Strain value", "Strain", and "ε" as "Strain value (με)"; unify the naming of "Inspection date" and "Maintenance time" as "Maintenance time (YYYY-MM-DD)".

[0031] 3. Determination of ownership: Spatial Anchoring: A spatial threshold of 0.5m is set. If the distance between the installation coordinates of a strain sensor (X=5003.10, Y=3202.05, Z=15.30) and the center point coordinates of crossarm C (X=5003.00, Y=3202.00, Z=15.00) is calculated to be 0.32m (less than 0.5m), then the sensor data is determined to belong to crossarm C. The distance calculation formula is:

[0032] In the formula: (x1, y1, z1) are the coordinates of the data points, (x2, y2, z2) are the coordinates of the center point of the component, and d is the straight-line distance between the two points.

[0033] Explicit anchoring: If the environmental load data contains an explicit number for the active component, "Crossbearing C", then it is directly matched with the component diagram to determine the attribution relationship.

[0034] 4. Data Aggregation: Using the unique identifier of the component "Crossarm C-2023-002" as the primary key, all information such as its three-dimensional coordinates, strain data, and material properties are aggregated to form multimodal fusion data containing multiple types of data such as numerical values, text, and coordinates.

[0035] Specifically, firstly, standardization mapping can be performed by customizing the mapping table and rule base to ensure that data from different sources maintains semantic consistency.

[0036] After semantic unification is completed, the system anchors the component affiliation of data records.

[0037] In one embodiment, the system performs a spatial neighborhood query (e.g., using a KD-tree spatial index) between the point and a known set of component coordinates to determine which component it belongs to. A spatial threshold ε can be set, and a point is considered to belong to component i if the following condition is met:

[0038] Where p is the coordinate of the data point, c i Let be the coordinates of the center point of component i.

[0039] If an explicit component number (such as "JGT-T04-H03") exists in the data field, direct matching can be achieved through the component map.

[0040] Once anchoring is complete, the system will aggregate the multimodal data under the component using the component ID as the primary key. This includes: structural response data (such as strain and displacement), material and geometric properties (such as material, cross section, and connection type), environmental information (such as temperature, humidity, and wind load), and historical operation and maintenance tags (such as defect type and maintenance time), thus constructing a unified data structure.

[0041] This structure can be converted to JSON format for output. Optionally, it can be saved as a table or graph database, and provides a unified semantic interface for subsequent label generation, standardized field construction, and model training.

[0042] like Figure 2 As shown, the implementation process of step 102 is as follows: Figure 3 In the S4 semantic perception and table content unification, all fields from image or text paths are ultimately mapped into a preliminary table and sent to the downstream semantic analysis module.

[0043] S103: For each component of the power pole, obtain the repaired standardized data corresponding to the component based on the multimodal fusion data corresponding to the component.

[0044] Specifically, for each component of the power pole, the multimodal fusion data corresponding to the component can be preprocessed to obtain the standardized data corresponding to the component.

[0045] In this embodiment, the preprocessing method may include at least one of the following: Convert data of different formats in the multimodal fusion data corresponding to the components into a preset standard format; The physical quantity units of the multimodal fusion data corresponding to the components are uniformly converted into preset standard units; A multi-scale segmented time matching method is adopted to enable time synchronization of multi-modal fusion data corresponding to components by sensor data with different sampling frequencies; A time-series prediction model is used to predict and fill in missing values ​​in the multimodal fusion data corresponding to the components based on historical data patterns. A multimodal anomaly identification algorithm is used to detect outliers in the multimodal fusion data corresponding to the components, and then the outliers are repaired by combining historical data.

[0046] It is understood that in this embodiment, the multimodal fusion data of each component is preprocessed to obtain standardized data. The specific processing methods include: 1. Consistent format and units: Convert all data to JSON format, and unify the units of physical quantities to the International System of Units (SI). For example, convert the wind speed unit from "km / h" to "m / s" (conversion formula: 1km / h=0.2778m / s), and convert the temperature unit from "℉" to "℃" (conversion formula: ℃=(℉-32) / 1.8).

[0047] 2. Time synchronization: A multi-scale segmented time matching method is adopted to synchronize strain data with a sampling frequency of 10Hz and wind speed data of 1Hz: the high-frequency strain data is averaged in 1-minute segments to align the timestamps of the two (e.g., to be unified to the 0th second of each minute).

[0048] 3. Missing value imputation: For the missing strain data of crossarm C during a certain period, the LSTM time series prediction model is used to fill in the missing data based on the historical data patterns of the previous 72 hours (such as the strain pattern caused by the diurnal temperature difference) to ensure data continuity.

[0049] 4. Outlier repair: The isolated forest algorithm (a type of multimodal anomaly identification algorithm) is used to detect outliers (such as a sudden jump in strain value to 10,000 με, which is far beyond the normal range). Combined with the historical data of the component in the same period (normal strain value is -500 to 500 με), the outliers are restored to predicted values ​​based on historical trends.

[0050] Thus, after the above processing, standardized data with uniform format, time alignment, and no anomalies is obtained, which is the standardized data corresponding to the component.

[0051] As an example, this step involves timestamp alignment, missing value imputation, and outlier detection and labeling.

[0052] In addition, to facilitate subsequent AI training and other uses, the raw monitoring data can be transformed into structured labels with high semantic density through multimodal data fusion, nonlinear feature extraction, and probabilistic graphical reasoning, so that intelligent operation and maintenance, fault prediction, and knowledge reasoning models can directly call them.

[0053] Timestamp alignment is used to eliminate the impact of differences in sampling frequencies or time formats of different sensors on data alignment. Different methods are used to map data from different types of sensors.

[0054] In one exemplary embodiment, such as Figure 4 As shown, to eliminate the differences in sampling frequency, time format, trigger delay, and clock drift among multiple sensor sources, this system adopts a time alignment method based on Dynamic Time Warping (DTW) and combines it with a multi-scale segmentation strategy to achieve high-precision time mapping. First, establish a unified standard reference timeline:

[0055] In the formula: t i This represents the time value at the i-th moment. Indicates the time step.

[0056] For the original sequence With reference sequence Calculate the cost matrix:

[0057] In the formula: X represents the original sampled data, Y represents the reference sampled data, and D is the cost matrix of both.

[0058] Euclidean distance, Manhattan distance, or Mahalanobis distance can be selected to accommodate sensor data of different dimensions.

[0059] Then, dynamic programming is used to solve for the optimal matching path:

[0060] The optimal path set obtained The mapping relationship from the original time point to the reference time point is given to achieve global nonlinear alignment. Here, DTW represents the dynamic time warping operator, a mathematical symbol without actual physical meaning; D... i,j denoted as the value at a specific position in the cost matrix; P represents the set point of the optimal path in the cost matrix.

[0061] To address short-term, high-frequency fluctuations from different sensors, this invention introduces a multi-scale segmentation strategy within the DTW framework: the original sequence is divided into multiple time segments, DTW is performed independently on each segment, and the results are aligned through weighted fusion.

[0062] in, The segment weights are adaptively adjusted based on the signal-to-noise ratio within each segment. Let S be the original sequence segment S; S represents the number of segments. This is a series of aligned results after weighted fusion.

[0063] Of course, these processing methods will inevitably deviate from the information obtained from actual measurements. Therefore, in actual measurements, it is necessary to ensure that the sampling frequency and time base are consistent as much as possible. Missing value imputation and outlier identification require the system to automatically identify missing fields or null value markers and perform repair on missing segments. Identification and repair strategies include manifold learning and GRU time series prediction imputation.

[0064] Traditional single methods (such as threshold filtering and interpolation) struggle to simultaneously address multimodal anomaly detection and time-series data missing data completion. This embodiment proposes an anomaly detection and repair method that combines manifold learning with gated recurrent unit (GRU) prediction. This method can identify latent anomalies and perform high-precision completion using historical patterns, thereby improving the quality and consistency of artificial intelligence training data.

[0065] This method consists of two stages: Phase 1 uses manifold learning to discover latent outliers after multimodal fusion and outputs outlier labels.

[0066] Phase 2 uses GRU time series prediction to predict and complete missing information and time series parts marked as anomalous, and outputs the repaired data using historical data patterns.

[0067] In this embodiment, Phase 1 involves manifold learning and detection on the time-series processed multimodal data. First, the k-nearest neighbor method is used to calculate weights to construct the adjacency matrix and weight matrix:

[0068] In the formula: x i y j This represents two sets of time series to be processed; Represents the mean squared error of the sequence; This represents the weight matrix.

[0069] Anomalies are then detected using manifold learning mapping and local density calculation.

[0070]

[0071] In the formula: Represented as a manifold learning mapping function; It is represented as a local density value.

[0072] if It is then identified as an outlier and an outlier label is output. This is the outlier threshold, typically 1.8.

[0073] In Phase Two of this embodiment, for outliers and missing values ​​(null markers or non-physical outliers) in the aligned sensor matrix, the system uses a deep temporal prediction model based on gated recurrent units (GRUs) to complete the data. Input time series vectors in sliding window form Where p is the window length, Z t It can contain multi-sensor, multi-variable values ​​at the same time, where t represents the number of the t-th sequence.

[0074] The gating mechanism of GRU is as follows:

[0075] In the formula: z t Enter information at the current time h t-1 The hidden state from the previous time step. The hidden state acts as the neural network's memory, containing information about the data seen by previous nodes. h t The hidden state passed to the next moment. Candidate hidden state r t Reset the door u t Update Gate The sigmoid function can be used to convert data into a value in the range of 0-1.

[0076] tanh: The tanh function can transform data into values ​​in the range [-1, 1]. The reset gate is used to control the influence of the previous hidden state ht on the current candidate state, and the update gate is used to control the fusion ratio between the current candidate state and the historical state.

[0077] In the prediction phase, the formula for predicting missing values ​​is:

[0078] In the formula: For missing predicted values; is the GRU gate control function; L is the reconstruction error.

[0079] During the inference phase, if the reconstruction error Lt of a sample exceeds the threshold δ at a certain moment, it is determined to be an anomaly:

[0080] The formula is the same as the latter half above, representing only the reconstruction error at a specific time t. It should be noted that for outliers discovered through manifold learning, users can choose whether to use the predicted values ​​obtained from GRU time series prediction as the output.

[0081] Meanwhile, the system allows loading user-defined or system-preset physical rules, common rules include but are not limited to: maximum allowable strain, maximum deflection angle, and minimum axial bearing capacity.

[0082] In this embodiment, the obtained data will be preliminarily processed. Semantics and tags will be generated through physical constraints (such as maximum allowable strain, maximum deflection angle, and minimum bearing capacity) built into the system or user-defined. The core is to transform the raw monitoring data into structured tags with high semantic density for subsequent high-level semantic analysis.

[0083] The goal of this step is to transform multi-source sensor-structure-environment data into standardized semantic health markers for use in AI training, inspection and early warning, and digital twin calibration.

[0084] The specific triggering rules are as follows: Perform real-time or batch calculations on the standardized component multimodal data, and determine whether the trigger logic is met one by one according to the trigger conditions of the rule base.

[0085] The triggering condition formula can be a simple inequality, and supports condition combinations and priority control.

[0086] After the rule is triggered, a semantic health flag containing the following fields is generated: In this embodiment, five levels can be used and associated with colors and action suggestions:

[0087] When the same component triggers multiple rules at the same time with different levels, the final level is determined according to the preset priority matrix (such as mechanical failure > trend anomaly > data quality).

[0088] Spatial consistency aggregation is applied between component-level, region-level, and tower-level labels: if a tower has ≥k adjacent components at or above S3 level, then the S3 label for that tower is automatically generated.

[0089] The generated semantic health flag is bound to the component ID, timestamp, and spatial location. The system will record a structural health semantic tag for the component. The tag is appended to the standard structure in the form of fields. The tag content includes: tag name, trigger rule number, over-limit value, and abnormal time.

[0090] All data points identified as anomalies will be marked and can be removed if desired. The anomaly type, detection method, and processing result will also be recorded.

[0091] Specifically, for each component of the power pole, a multidimensional index system corresponding to the component can be generated based on the standardized data corresponding to the component, and the standardized data can be type-checked and Windsorized to obtain the repaired standardized data.

[0092] In this embodiment, the multidimensional index system includes at least the following index fields: a component unique identifier index, a time index, a spatial index, and a semantic tag index; wherein, the component unique identifier index is used to uniquely identify the component; the time index supports range retrieval based on time periods; the spatial index is constructed based on the spatial coordinate information of the component and is used for fast lookup based on spatial neighborhood; the semantic tag index includes health status semantic tags, implicit pattern semantic tags, and predicted failure mode tags generated based on the standardized data corresponding to the component, and the semantic tag index is used to filter and call data according to the component status level.

[0093] It should be noted that the health status semantic label is generated based on the physical constraint rules of the component. The physical constraint rules include at least one of the component's maximum allowable strain, maximum deflection angle, and minimum bearing capacity. If the monitored value in the standardized data of the component exceeds the threshold corresponding to the physical constraint rule, different health status levels are divided according to the degree of exceedance, and the component is labeled with the corresponding health status semantic label.

[0094] The implicit pattern semantic labels are generated by combining manifold learning and graph neural networks. By performing dimensionality reduction and pattern clustering on the standardized data of components, the operating modes of components are identified and labeled.

[0095] The predicted failure mode labels are generated by combining a hybrid deep sequence model with state transition probability modeling, and are generated by predicting the failure modes that will occur in the component within a preset time period and labeling them.

[0096] In one implementation, the process of performing type checking and Windsorization repair on the standardized data is as follows: verify whether the data type of each field in the standardized data is consistent with the preset field type; if they are inconsistent, they are marked as data with abnormal type; for the data with abnormal type, convert it into data that conforms to the preset field type according to the preset repair rules; for abnormal data that cannot be directly converted, mark it as an item that requires manual confirmation and push it to the manual review platform to prompt staff to perform repair confirmation.

[0097] This can be understood as follows: for each component, a multi-dimensional indexing system is built, and standardized data is repaired as follows: (1) Generation of a multidimensional index system: An index system can include the following fields: Unique component identifier index: such as "crossarm C-2023-002", to ensure the unique identification of the component; Time index: The format is "YYYY-MM-DD HH:MM:SS", and it supports searching by time period (e.g., "2023-06-01 to 2023-06-30"). Spatial Index: An R-tree index is built based on the three-dimensional coordinates of the components, supporting spatial neighborhood queries (such as finding all components within a 5m range of tower leg A). Semantic tag index: includes: Semantic label for health status: generated according to physical constraint rules, such as the maximum allowable strain threshold of crossarm C is ±800με. If the monitored value is 950με (exceeding the threshold) during a certain period, it is labeled "Health status: mildly abnormal". Latent pattern semantic labeling: The label "daytime high temperature strain pattern" was identified by t-SNE manifold learning dimensionality reduction and GNN graph neural network clustering. Failure mode prediction labels: Using an LSTM-Transformer hybrid deep sequence model and combining state transition probabilities, the model predicts that "bolt loosening" may occur within the next 30 days and labels it accordingly.

[0098] (2) Type checking and Windsorization repair of standardized data The type check specifically verifies whether the data types of each field conform to preset rules (e.g., "Strain Value" should be a floating-point type, and "Operation and Maintenance Time" should be a date type). If the string "Abnormal" is found in the "Strain Value" field, it is marked as data with an abnormal type.

[0099] The Windsor restoration specifically involves: For convertible abnormal data (such as converting the string "850με" to the floating-point number 850), it is automatically repaired to data that conforms to the correct type; For data that cannot be converted (such as "Strain value: Severely exceeded"), mark it as a manually confirmed item and push it to the operation and maintenance platform to prompt staff to verify and repair it.

[0100] Understandably, this step is the final systematic verification before performing high semantic analysis on the standardized multimodal data. The verification objects include the integrity and consistency of field content, data type, and semantic labels, in order to avoid erroneous data from interfering with subsequent AI model training and structural diagnosis.

[0101] In one exemplary embodiment, the system performs the following validation on each field in the standardized fields: Type validity: Ensure that the field type is consistent with the type defined in the template (timestamp is ISO8601 timestamp, Force is float type).

[0102] Unit validity: For example, the force unit is "kN", the temperature unit is "°C", etc., and the system uses a built-in unit whitelist for judgment; Time sequence validity: The timestamp field must be monotonically increasing, and the interval must meet the sensor frequency definition.

[0103] For any illegal field-level entries detected, the system will trigger an error logging mechanism and select a repair strategy based on the field type.

[0104] The system checks the physical consistency between components in the standardized fields: The system also performs physical logic checks across fields and multiple components. Common rules include: Non-zero length check: If the start and end coordinates of a component are the same, it is considered an illegal component; Orientation Consistency: If the orientation deviation of a component is significantly inconsistent with the structural design (e.g., a horizontal component is perpendicular to the orientation), a warning will be triggered. Duplicate component detection: Removes duplicate ID components or redundant data records through hash verification. In one exemplary embodiment, an example of constructing a hash input string for a field is shown below: Convert the standardized output structure table into a string: Z Then perform a hash calculation, for example, using the SHA-256 algorithm:

[0105] The system records the hash values ​​of all components in an index table. If the hash value of new data already exists in the existing set, it is considered a redundant component and can be removed or prompted to be merged.

[0106] All legality verification anomalies and repair actions will be written to the system log, which includes: component number, anomaly type, repair method, timestamp, and executing module. If the system cannot determine the repair strategy, it will be marked as requiring manual confirmation and pushed to the manual review platform through the front-end interface to ensure that critical data repair does not lose its original meaning.

[0107] It should also be noted that in this step, the repaired data will be further processed to form high-level semantics, so as to form semantic and label dimensions (i.e., multi-dimensional index system) that AI can directly use. Its core is to transform the original monitoring data into structured labels with high semantic density through multimodal data fusion, nonlinear feature extraction and probabilistic graph reasoning, so that intelligent operation and maintenance, fault prediction and knowledge reasoning models can directly call them.

[0108] The semantic and tag dimensions include at least the following: 1. Semantic tags for health status Based on multimodal input X (including geometric, dynamic, environmental, and material parameters), an end-to-end multi-task deep neural network is constructed. Joint Embedded Nonlinear Feature Extractor Combined with the Bayesian uncertainty estimation module, it outputs the health index H and its confidence interval. .

[0109] The calculation process can be represented as follows:

[0110] in, This represents a multimodal concatenation operation; To predict variance, it is used to characterize the uncertainty of health status, facilitating risk assessment in operation and maintenance decisions; These are represented as geometric data, dynamic data, and environmental data, respectively. For feature functions of deep neural networks; This is expressed as a health index value.

[0111] The raw, multi-source sensor data (geometric morphology, dynamic response, environmental parameters, etc.) is compressed into a unified health index Hs (with confidence intervals). This transforms scattered raw monitoring data with different physical units into health scores that can be directly compared and ranked.

[0112] 2. Implicit pattern semantic tags This approach combines manifold learning with graph neural networks (GNNs). First, it encodes the spatial topology and physical connectivity into a graph structure G=(V,E) using a graph constructor G(.). Then, it generates low-dimensional embeddings Z using spectral clustering and high-dimensional nonlinear mapping, and performs probabilistic soft clustering to generate state pattern labels C. K and probability distribution vector p k .

[0113] The mathematical expression is:

[0114] Where A is the adjacency matrix, Let k be the center vector of the pattern class. Within-class variance; Construct a graph constructor to handle geometric input; This is a graph neural network function used to implement learning and training. k Let be the probability distribution vector. and analogy , No further details will be provided; X and X geo All of these are represented as modal input parameters.

[0115] This label can perform dimensionality reduction and pattern clustering on high-dimensional, multimodal structural monitoring data, identifying different "state clusters" or "operational modes." Label C... k Indicates the current structural mode category (e.g., normal operation, minor anomaly, early stage of structural deformation, etc.), p k This represents the probability distribution of the judgment. It helps AI identify the operating patterns of different towers under similar working conditions, facilitating knowledge transfer across structures.

[0116] 3. Predicted Failure Mode Labels In the time series analysis phase, a hybrid deep sequence model (such as the Transformer-TemporalConvolutional Network hybrid architecture) is used to analyze the historical time window X. t T:t Multi-scale convolution and self-attention encoding are performed, and a Markov Jump Process is used to model the failure state transition probability, outputting the future T. f Failure modes that may occur within a time period F m and its probability of occurrence P m .

[0117] The calculation process is as follows:

[0118] in, Let F be the initial failure probability vector, and T be the failure mode state transition matrix. m For failure mode, P m X represents the probability of occurrence. t-T:t For historical time windows; Transformer is a deep sequence model; TCN is a hybrid deep sequence model; H is a failure feature.

[0119] 4. Semantic Indexing Matrixing and High-Dimensional Tensorization Representation The above three types of labels are uniformly encapsulated into a fourth-order tensor. Where n is the number of indexed objects, d1 is the number of tag types (health index, pattern tag, failure probability, etc.), d2 is the feature dimension of each tag type, and t is the time step. Mathematically, S is a fourth-order tensor; a fourth-order tensor has four numerical dimensions, meaning there are four symbols in the upper right corner of R.

[0120] This means:

[0121] This quantized index can not only be directly used as input for AI models, but also extract potential correlation patterns through tensor decomposition, enabling predictive maintenance optimization and full lifecycle management decision support. Here, H... s p k p m These represent computational functions for different sequences and have no specific practical physical meaning. S i This represents a certain fourth-order tensor mentioned above.

[0122] S104: Store the multi-dimensional index system and the repaired standardized data corresponding to each component of the power pole to a preset structural health database so that the AI ​​model training platform, analysis system and visualization engine can use the multi-dimensional index system corresponding to the component to call the repaired standardized data of the component.

[0123] In this embodiment, the preset structural health database includes at least one of a relational database, a document database, a graph database, and a time-series database. The relational database stores component data with a clear field hierarchy and frequent queries. The document database stores component data with key-value pair data structures or heterogeneous fields. The graph database stores topological structure data between components to adapt to the graph neural network construction requirements. The time-series database stores continuous monitoring data of components to adapt to high-frequency data acquisition and analysis requirements.

[0124] In other words, the multidimensional index system corresponding to each component of the power pole and the repaired standardized data can be stored in the following structural health database: Relational database (MySQL): Store data with clear fields such as material properties and basic component parameters; Document database (MongoDB): Store heterogeneous field data such as multi-modal fusion data; Graph database (Neo4j): Store the topological relationships between components (such as "Tower Leg A - Connection - Tower Section B"); Time-series database (InfluxDB): Store high-frequency sensor monitoring data (such as strain data at 10Hz).

[0125] As an example, the specific implementation method for facilitating the AI model training platform, analysis system, and visualization engine to utilize the repaired standardized data of components through the multi-dimensional index system corresponding to the components can be as shown in the following introduction content. Provide API interfaces to the AI model training platform, the analysis system, and the visualization engine; among them, the API interfaces support at least one protocol among RESTful, GraphQL, and WebSocket. The AI model training platform, the analysis system, and the visualization engine input query conditions based on the multi-dimensional index system corresponding to the components through the API interfaces, where the query conditions include at least one of component type, time range, spatial region, and health status level. The structural health database matches the standardized data of the corresponding components according to the query conditions, and feeds back the standardized data of the components to the caller. At the same time, it supports exporting the standardized data of the components into a dataset format adapted for AI model training, and the dataset format includes at least one of CSV, JSON, TFRecord, and NPY.

[0126] It can be understood that this embodiment can provide data calling capabilities for the AI model training platform, analysis system, and visualization engine. Specifically, it can include the following capabilities: Providing API interfaces: Deploy API interfaces that support RESTful and WebSocket protocols, such as / api / query?componentId=Crossarm C-2023-002&timeRange=2023-06.

[0127] Multi-condition query: (1) The AI model training platform inputs query conditions (such as "Component type = Crossarm, Health status = Abnormal, Time range = 2023") through the interface to obtain tagged training data; (2) The visualization engine inputs spatial region conditions (such as "Longitude 116° - 117°, Latitude **********") to call the spatial and status data of the corresponding tower components.

[0128] Data Export: Supports exporting query results to formats such as CSV (for traditional machine learning), TFRecord (for TensorFlow training), and NPY (for PyTorch training), directly adapting to AI model training needs.

[0129] In one implementation of this embodiment, before the step of storing the multidimensional index system corresponding to each component of the power pole and the repaired standardized data into a preset structural health database, the method further includes: A hash algorithm is used to generate a unique hash value for the standardized data of each component after repair. The hash value is compared with the hash value of the standardized data corresponding to the component stored in the structural health database. If there is a duplicate hash value and the time interval between the corresponding data is less than a preset time threshold, it is determined to be duplicate data. The latest standardized data corresponding to the component is retained and the duplicate standardized data corresponding to the component is removed.

[0130] This can be understood as requiring deduplication before storage. For example, the SHA-256 hash algorithm can be used to generate a unique hash value for the standardized data after repair of each component. If a piece of data has the same hash value as data already stored in the database, and the time interval (e.g., 2 minutes) is less than a preset threshold (5 minutes), it is determined to be duplicate data, and only the latest data is retained.

[0131] It is important to emphasize that this step aims to push the multimodal data, after field standardization, semantic tag generation, and legality verification, to the structural health database system, and to build a data indexing system that supports multi-condition queries and efficient invocation. The system is compatible with multiple database types, supports structured retrieval and AI service invocation, and meets the requirements of downstream structural recognition, state assessment, and training tasks for high-quality datasets.

[0132] Data can be stored in any of the following database systems: relational databases (such as PostgreSQL), suitable for tasks with clear field hierarchical structures and frequent queries; document databases (such as MongoDB), suitable for key-value pair data structures or heterogeneous field storage; graph databases (such as Neo4j), suitable for topological modeling between components and graph neural network construction; and time-series databases (such as InfluxDB), suitable for high-frequency acquisition and analysis of continuous monitoring data.

[0133] Secondly, based on the data entered into the database, the system constructs a multi-dimensional composite index system, which includes the following key index fields: Component ID Index: Used to uniquely identify components; Time index: Supports range retrieval based on time period; Spatial index (component center point coordinates): used for fast lookup based on spatial neighborhood; Tag index: Used to filter calls by status level; Health Status Semantic Label Index: Based on multi-source monitoring data, the health index and confidence interval are calculated by a deep learning model to reflect the overall health of components or the entire tower and support trend analysis. Latent pattern semantic label index: State pattern categories and probability distributions extracted through manifold learning and pattern clustering algorithms, used to identify structural operation modes and potential abnormal features; Predicted Failure Mode Tag Index: Based on time-series prediction models and state transition probability matrices, this index calculates future failure types and their probabilities of occurrence, and is used for predictive maintenance and risk classification management.

[0134] In an exemplary embodiment, the above index structure can be represented as the following quadruple:

[0135] Wherein, ID represents the component number; t For time sequence number; l The length of the component; The geometric center point, determined by the start and end coordinates of the component, is calculated using the following formula:

[0136] In the formula: Represented as the coordinates of the starting point; Represented as the coordinates of the termination point.

[0137] The above indexing system supports the execution of multi-dimensional joint search statements, including: querying the status changes of a specified component within a certain time period; querying all components in a warning state within a certain spatial region; and querying the set of all components labeled "fail" within a continuous time window.

[0138] Finally, the system provides the aforementioned data to external platforms or model training tasks, enabling structured access through standardized API interfaces. These interfaces support protocols such as RESTful, GraphQL, and WebSocket, and possess capabilities for field filtering, time aggregation, status indexing, and export, meeting the data input needs of AI platforms, BIM / GIS system integration, and user visualization.

[0139] The system also supports formatted export of training data, such as .csv, .json, .tfrecord, .npy, etc., and supports constructing graph structure data with components as nodes and spatial relationships as edges for input modeling of graph neural networks.

[0140] It is important to note that the process of parametrically identifying and classifying single-line diagrams of towers, such as... Figure 5 As shown, during the parsing of a single-line diagram in a two-dimensional drawing, after completing page structure recognition and OCR text extraction, the system further enters the parametric recognition sub-process of structural components. This sub-process realizes the automatic recognition and encoding of the cross-type, cross-section type and tower head type of the members based on the component topology hierarchy of the overall tower single-line diagram.

[0141] Specifically, the following steps are included: Tower head type identification: The system matches the geometric features of the tower head and crossarm area in the drawings (located at the top of the overall tower, with a specific horizontal extension ratio and member arrangement pattern) to the tower head feature template library (Tower Head Library ⑥). Through pattern matching and feature vector similarity calculation, the system determines the tower head structure type and generates tower head type parameters. The parameter values ​​are derived from the encoding results of the tower head library.

[0142] Rod intersection type identification: For the portion of the tower body without a base, the system first extracts the centerline of the members, calculates the angle between adjacent members, and compares it with a preset threshold range for intersection structures. When a symmetrical quadrilateral is formed, it is determined to be an "X-type structure" and encoded into "X-type library ②"; When a member has an acute angle with an upward opening, it is identified as a "Λ-type structure" and coded into "Λ-type library ③"; When a member has an acute angle with a downward opening, it is identified as a "V-shaped structure" and coded into "V-shaped library ④".

[0143] The system stores the recognition results as cross-type parameters and associates them with the location coordinates and structural node numbers.

[0144] Section type identification: In the pedestal, tower legs, and tower body components, the system uses two-dimensional section symbol features (such as drawing symbols for channel steel, angle steel, and round pipe) to match with the section template library, and combines the section geometry parameters extracted by Hough transform to generate section type parameters. For components that require additional crossbars, the system records the section type and the number of crossbars after matching and associates them with the section library⑤.

[0145] Parameterized storage and association: The system stores the three types of parameters in a multi-dimensional index database as standardized fields, and associates them with the corresponding component ID, spatial coordinates, and structural hierarchy path to achieve a fully parameterized expression of the tower head, tower body, and tower legs.

[0146] Using the above method, this application can achieve automated identification, encoding and parameterized recording of component types in two-dimensional single-line diagram analysis, providing high-precision and reusable basic data support for subsequent multi-dimensional index establishment, structural topology analysis and intelligent retrieval.

[0147] As can be seen from the above technical solution, the data processing method for power poles provided in this application obtains multi-source data of power poles, first extracts related data information for each component and generates multimodal fusion data, effectively solving the problems of scattered multi-source data and low correlation with components in traditional processing, realizing multi-source data aggregation with components as the core, laying the foundation for subsequent accurate processing; then, the multimodal fusion data is preprocessed to obtain standardized data, eliminating differences in data format, units, etc., and improving data consistency; next, a multi-dimensional index system corresponding to the component is generated and type checking and Windsor-style repair are performed, which not only provides index support for rapid data retrieval, but also ensures the legality and accuracy of standardized data; finally, the multi-dimensional index system and the repaired standardized data are stored in the structural health database, enabling the AI ​​model training platform, analysis system and visualization engine to efficiently call the required component data with the help of the multi-dimensional index system, which not only meets the needs of AI model training for high-quality data, but also provides reliable data support for the structural analysis of the analysis system and the visualization display of the visualization engine, greatly improving the utilization value of power pole data and promoting the development of digital management and intelligent application of power poles.

[0148] In other words, this embodiment achieves systematic processing of multi-source data from power poles through the above steps: raw data extracted from multi-format files is correlated, fused, and standardized to form high-quality repaired data; the construction of a multi-dimensional indexing system improves data retrieval efficiency by more than 80%; collaborative storage of multiple database types meets the access needs of different scenarios; and the open API interface supports cross-platform calls, providing reliable data support for structural health assessment, AI prediction model training, and visual monitoring of power poles. The entire process strictly follows the technical features of claims 1-10, ensuring the feasibility and practicality of the method.

[0149] like Figure 7 The image shows a specific embodiment of a data processing device for power poles provided in this application. The device described in this embodiment is the physical device used to execute the method described in the above embodiments. Its technical solution is essentially the same as that of the above embodiments, and the corresponding descriptions in the above embodiments are also applicable to this embodiment. The device in this embodiment includes: Unit 701 is used to acquire multi-source data of power poles; The second unit 702 is used to extract data information associated with each component of the power pole from the multi-source data, and to generate multimodal fusion data corresponding to the component based on the data information associated with the component. The third unit 703 is used to obtain the repaired standardized data corresponding to each component of the power pole based on the multimodal fusion data corresponding to the component. The fourth unit 704 is used to store the multi-dimensional index system and the repaired standardized data corresponding to each component of the power pole to a preset structural health database, so that the AI ​​model training platform, analysis system and visualization engine can use the multi-dimensional index system corresponding to the component to call the repaired standardized data of the component.

[0150] Optionally, the multi-source data of the power pole includes: three-dimensional spatial coordinates, sensor monitoring data, material properties, spatial location information, environmental load and historical operation and maintenance records; the multi-source data is extracted from structured or semi-structured files, which include at least one of the following formats: PDF, CSV, TXT, JSON format, tables obtained by parsing images or PDFs, and parametric component information.

[0151] Optionally, the step of extracting data information associated with the component from the multi-source data, and generating multimodal fusion data corresponding to the component based on the data information associated with the component, includes: Based on the unique identifier or spatial anchor point information of the component, extract data information associated with the component from the multi-source data; Unify the semantics of fields in data information associated with components to eliminate differences in field naming from different data sources; The attribution relationship between the data information and the component is determined by spatial anchoring or explicit anchoring. The spatial anchoring method is achieved through spatial neighborhood query. A spatial threshold is set. If the distance between the coordinates of the data point and the coordinates of the center point of the component is less than the spatial threshold, it is determined that the data information belongs to the component. The explicit anchoring method is to directly match the explicit component number contained in the data information with the component map. Using the unique identifier of the component as the primary key, all data information belonging to the component is aggregated to form the multimodal fusion data corresponding to the component.

[0152] Optionally, for each component of the power pole, obtaining the repaired standardized data corresponding to the component based on the multimodal fusion data corresponding to the component includes: For each component of the power pole, the multimodal fusion data corresponding to the component is preprocessed to obtain the standardized data corresponding to the component. Based on the standardized data corresponding to the component, a multidimensional index system corresponding to the component is generated, and the standardized data is subjected to type checking and Windsorization repair to obtain the repaired standardized data.

[0153] Optionally, the preprocessing method includes at least one of the following: Convert data of different formats in the multimodal fusion data corresponding to the components into a preset standard format; The physical quantity units of the multimodal fusion data corresponding to the components are uniformly converted into preset standard units; A multi-scale segmented time matching method is adopted to enable time synchronization of multi-modal fusion data corresponding to components by sensor data with different sampling frequencies; A time-series prediction model is used to predict and fill in missing values ​​in the multimodal fusion data corresponding to the components based on historical data patterns. A multimodal anomaly identification algorithm is used to detect outliers in the multimodal fusion data corresponding to the components, and then the outliers are repaired by combining historical data.

[0154] Optionally, the multidimensional index system includes at least the following index fields: a component unique identifier index, a time index, a spatial index, and a semantic tag index; wherein, the component unique identifier index is used to uniquely identify the component; the time index supports range retrieval based on time periods; the spatial index is constructed based on the spatial coordinate information of the component and is used for fast lookup based on spatial neighborhood; the semantic tag index includes health status semantic tags, implicit pattern semantic tags, and predicted failure mode tags generated based on the standardized data corresponding to the component, and the semantic tag index is used to filter and call data according to the component status level.

[0155] Optionally, the health status semantic label is generated based on the physical constraint rules of the component, which include at least one of the component's maximum allowable strain, maximum deflection angle, and minimum bearing capacity; if the monitored value in the standardized data of the component exceeds the threshold corresponding to the physical constraint rule, different health status levels are divided according to the degree of exceedance, and the component is labeled with the corresponding health status semantic label. The implicit pattern semantic labels are generated by a combination of manifold learning and graph neural networks. By performing dimensionality reduction and pattern clustering on the standardized data of components, the operating modes of components are identified and labeled. The predicted failure mode labels are generated by combining a hybrid deep sequence model with state transition probability modeling, and are generated by predicting the failure modes that will occur in the component within a preset time period and labeling them.

[0156] Optionally, the step of performing type checking and Windsorization repair on the standardized data includes: Verify whether the data type of each field in the standardized data is consistent with the preset field type; if they are inconsistent, mark them as data with abnormal type. For data with abnormal types, it is converted into data that conforms to the preset field types according to the preset repair rules. For abnormal data that cannot be directly converted, it is marked as an item that requires manual confirmation and pushed to the manual review platform to prompt staff to perform repair confirmation.

[0157] Optionally, before the step of storing the multidimensional index system corresponding to each component of the power pole and the repaired standardized data into a preset structural health database, the method further includes: A hash algorithm is used to generate a unique hash value for the standardized data of each component after repair. The hash value is compared with the hash value of the standardized data corresponding to the component stored in the structural health database. If there is a duplicate hash value and the time interval between the corresponding data is less than a preset time threshold, it is determined to be duplicate data. The latest standardized data corresponding to the component is retained and the duplicate standardized data corresponding to the component is removed.

[0158] Optionally, the preset structural health database includes at least one of relational database, document database, graph database, and time-series database; wherein, the relational database is used to store component data with a clear field hierarchy structure and frequent queries; the document database is used to store component data with key-value pair data structures or heterogeneous fields; the graph database is used to store topological structure data between components to adapt to the graph neural network construction requirements; and the time-series database is used to store continuous monitoring data of components to adapt to the high-frequency data acquisition and analysis requirements.

[0159] Optionally, the step of enabling AI model training platforms, analysis systems, and visualization engines to access the repaired standardized data of components using the multi-dimensional indexing system corresponding to the components includes: The system provides API interfaces to the AI ​​model training platform, the analysis system, and the visualization engine; wherein the API interfaces support at least one of the following protocols: RESTful, GraphQL, and WebSocket. The AI ​​model training platform, the analysis system, and the visualization engine input query conditions based on the multidimensional index system corresponding to the component through the API interface. The query conditions include at least one of component type, time range, spatial region, and health status level. The structural health database matches the standardized data of the corresponding components according to the query conditions and feeds back the standardized data of the components to the caller. At the same time, it supports exporting the standardized data of the components into a dataset format adapted for AI model training. The dataset format includes at least one of CSV, JSON, TFRecord, and NPY.

[0160] Figure 8 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. At the hardware level, the electronic device includes a processor, and optionally also includes an internal bus, a network interface, and a memory. The memory may include RAM, such as high-speed random-access memory (RAM), or non-volatile memory, such as at least one disk storage device. Of course, the electronic device may also include other hardware required for other services.

[0161] The processor, network interface, and memory can be interconnected via an internal bus, which can be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus, or an EISA (Extended Industry Standard Architecture) bus, etc. This bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 8 The symbol is represented by a single double-headed arrow, but this does not mean that there is only one bus or one type of bus.

[0162] Memory is used to store instructions for execution. Specifically, instructions for execution are computer programs that can be executed. Memory can include main memory and non-volatile memory, and it provides the processor with execution instructions and data.

[0163] In one possible implementation, the processor reads the corresponding execution instructions from non-volatile memory into main memory and then executes them. Alternatively, it may obtain the corresponding execution instructions from other devices to form a data processing device for power poles at the logical level. The processor executes the execution instructions stored in the memory to implement the data processing method for power poles provided in any embodiment of this application.

[0164] The above is as stated in this application. Figure 1The method executed by the data processing device for power poles provided in the illustrated embodiment can be applied to a processor or implemented by a processor. The processor may be an integrated circuit chip with signal processing capabilities. During implementation, each step of the above method can be completed by integrated logic circuits in the processor's hardware or by instructions in software form. The processor can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it can also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor can be a microprocessor or any conventional processor.

[0165] The steps of the method disclosed in the embodiments of this application can be directly manifested as being executed by a hardware decoding processor, or executed by a combination of hardware and software modules in the decoding processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. This storage medium is located in memory, and the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method.

[0166] This application also proposes a readable medium that stores execution instructions. When the stored execution instructions are executed by the processor of an electronic device, the electronic device can execute the data processing method for power poles provided in any embodiment of this application, and specifically perform the above-mentioned evaluation method.

[0167] The electronic devices described in the foregoing embodiments may be computers.

[0168] Those skilled in the art will understand that the embodiments of this application can be provided as methods or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or a combination of software and hardware.

[0169] The various embodiments in this application are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the device embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0170] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0171] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A data processing method for power poles, characterized in that, The method includes: Acquire multi-source data on power poles; For each component of the power pole, data information associated with the component is extracted from the multi-source data, and multimodal fusion data corresponding to the component is generated based on the data information associated with the component. For each component of the power pole, the standardized data after repair is obtained based on the multimodal fusion data corresponding to the component. The multidimensional index system and the repaired standardized data corresponding to each component of the power pole are stored in a preset structural health database so that the AI ​​model training platform, analysis system and visualization engine can use the multidimensional index system corresponding to the component to call the repaired standardized data of the component.

2. The method according to claim 1, characterized in that, The multi-source data of power poles includes: three-dimensional spatial coordinates, sensor monitoring data, material properties, spatial location information, environmental loads and historical operation and maintenance records; the multi-source data is extracted from structured or semi-structured files, which include at least one of the following formats: PDF, CSV, TXT, JSON, tables obtained by parsing images or PDFs, and parametric component information.

3. The method according to claim 1 or 2, characterized in that, The step of extracting data information associated with the component from the multi-source data, and generating multimodal fusion data corresponding to the component based on the data information associated with the component, includes: Based on the unique identifier or spatial anchor point information of the component, extract data information associated with the component from the multi-source data; Unify the semantics of fields in data information associated with components to eliminate differences in field naming from different data sources; The attribution relationship between the data information and the component is determined by spatial anchoring or explicit anchoring. The spatial anchoring method is achieved through spatial neighborhood query. A spatial threshold is set. If the distance between the coordinates of the data point and the coordinates of the center point of the component is less than the spatial threshold, it is determined that the data information belongs to the component. The explicit anchoring method is to directly match the explicit component number contained in the data information with the component map. Using the unique identifier of the component as the primary key, all data information belonging to the component is aggregated to form the multimodal fusion data corresponding to the component.

4. The method according to claim 1 or 2, characterized in that, For each component of the power pole, the standardized data after repair for that component is obtained based on the multimodal fusion data corresponding to that component, including: For each component of the power pole, the multimodal fusion data corresponding to the component is preprocessed to obtain the standardized data corresponding to the component. Based on the standardized data corresponding to the component, a multidimensional index system corresponding to the component is generated, and the standardized data is subjected to type checking and Windsorization repair to obtain the repaired standardized data.

5. The method according to claim 4, characterized in that, The preprocessing method includes at least one of the following: Convert data of different formats in the multimodal fusion data corresponding to the components into a preset standard format; The physical quantity units of the multimodal fusion data corresponding to the components are uniformly converted into preset standard units; A multi-scale segmented time matching method is adopted to enable time synchronization of multi-modal fusion data corresponding to components by sensor data with different sampling frequencies; A time-series prediction model is used to predict and fill in missing values ​​in the multimodal fusion data corresponding to the components based on historical data patterns. A multimodal anomaly identification algorithm is used to detect outliers in the multimodal fusion data corresponding to the components, and then the outliers are repaired by combining historical data.

6. The method according to claim 1 or 2, characterized in that, The multidimensional index system includes at least the following index fields: component unique identifier index, time index, spatial index, and semantic tag index; wherein, the component unique identifier index is used to uniquely identify the component; the time index supports range retrieval based on time period; the spatial index is constructed based on the spatial coordinate information of the component and is used for fast search based on spatial neighborhood; the semantic tag index includes health status semantic tags, implicit pattern semantic tags, and predicted failure mode tags generated based on the standardized data corresponding to the component, and the semantic tag index is used to filter and call data according to the component status level.

7. The method according to claim 6, characterized in that, The health status semantic label is generated based on the physical constraint rules of the component. The physical constraint rules include at least one of the component's maximum allowable strain, maximum deflection angle, and minimum bearing capacity. If the monitored value in the standardized data of the component exceeds the threshold corresponding to the physical constraint rule, different health status levels are divided according to the degree of exceedance, and the component is labeled with the corresponding health status semantic label. The implicit pattern semantic labels are generated by a combination of manifold learning and graph neural networks. By performing dimensionality reduction and pattern clustering on the standardized data of components, the operating modes of components are identified and labeled. The predicted failure mode labels are generated by combining a hybrid deep sequence model with state transition probability modeling, and are generated by predicting the failure modes that will occur in the component within a preset time period and labeling them.

8. The method according to claim 4, characterized in that, The process of performing type checking and Windsorization repair on the standardized data includes: Verify whether the data type of each field in the standardized data is consistent with the preset field type; if they are inconsistent, mark them as data with abnormal type. For data with abnormal types, it is converted into data that conforms to the preset field types according to the preset repair rules. For abnormal data that cannot be directly converted, it is marked as an item that requires manual confirmation and pushed to the manual review platform to prompt staff to perform repair confirmation.

9. The method according to claim 1 or 2, characterized in that, Before the step of storing the multidimensional index system corresponding to each component of the power pole and the repaired standardized data into a preset structural health database, the method further includes: A hash algorithm is used to generate a unique hash value for the standardized data of each component after repair. The hash value is compared with the hash value of the standardized data corresponding to the component stored in the structural health database. If there is a duplicate hash value and the time interval between the corresponding data is less than a preset time threshold, it is determined to be duplicate data. The latest standardized data corresponding to the component is retained and the duplicate standardized data corresponding to the component is removed.

10. The method according to claim 1 or 2, characterized in that, The preset structural health database includes at least one of relational database, document database, graph database, and time-series database; wherein, the relational database is used to store component data with a clear field hierarchy structure and frequent queries; the document database is used to store component data with key-value pair data structures or heterogeneous fields; the graph database is used to store topological structure data between components to adapt to the graph neural network construction requirements; and the time-series database is used to store continuous monitoring data of components to adapt to the high-frequency data acquisition and analysis requirements.

11. The method according to claim 1 or 2, characterized in that, The method to enable AI model training platforms, analysis systems, and visualization engines to access the repaired, standardized data of components using the multi-dimensional indexing system corresponding to the components includes: API interfaces are provided to the AI ​​model training platform, the analysis system, and the visualization engine; wherein the API interfaces support at least one of the protocols RESTful, GraphQL, and WebSocket. The AI ​​model training platform, the analysis system, and the visualization engine input query conditions based on the multidimensional index system corresponding to the component through the API interface. The query conditions include at least one of component type, time range, spatial region, and health status level. The structural health database matches the standardized data of the corresponding components according to the query conditions and feeds back the standardized data of the components to the caller. At the same time, it supports exporting the standardized data of the components into a dataset format adapted for AI model training. The dataset format includes at least one of CSV, JSON, TFRecord, and NPY.

12. A data processing device for power poles, characterized in that, The device includes: The first unit is used to acquire multi-source data on power poles; The second unit is used to extract data information associated with each component of the power pole from the multi-source data, and to generate multimodal fusion data corresponding to the component based on the data information associated with the component. The third unit is used to obtain the repaired standardized data corresponding to each component of the power pole based on the multimodal fusion data corresponding to the component. The fourth unit is used to store the multi-dimensional index system and the repaired standardized data corresponding to each component of the power pole to a preset structural health database, so that the AI ​​model training platform, analysis system and visualization engine can use the multi-dimensional index system corresponding to the component to call the repaired standardized data of the component.

Images