Substation detection report automatic generation method based on artificial intelligence
Through an artificial intelligence-based approach, combined with explainable AI anomaly detection and automated document generation technology, efficient collection of substation multimodal detection data and automatic generation of intelligent reports are achieved, solving the problems of low efficiency and insufficient standardization in substation detection data analysis and report generation, achieving a closed-loop and standardized process for detection report generation, and improving the intelligence level of substation detection management.
Patent Information
- Application Number
- CN202511349900.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-22
- Publication Date
- 2025-10-28
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The existing substation inspection data analysis and report generation process relies on manual or semi-automatic methods, and has problems such as low data processing efficiency, reliance on manual experience for anomaly identification, insufficient report standardization, and difficulty in archiving and tracing. It cannot meet the needs of large-scale, real-time, and intelligent operation and maintenance management.
By adopting an artificial intelligence-based approach, combined with explainable AI anomaly detection and automated document generation technology, we can achieve efficient collection of multimodal detection data, accurate anomaly identification, and automatic generation of intelligent reports. Through adaptive template matching, structured summary suggestion output, and multi-role hierarchical collaborative review, we can achieve a closed-loop and standardized process for detection report generation.
It significantly improves the automation level and efficiency of the inspection process, reduces the risk of human omissions, improves the standardization and traceability of inspection results, realizes the closed-loop automation and standardization of the entire process of inspection report generation, and enhances the intelligent decision-making capabilities of substation inspection management.
Smart Images

Figure CN120850986A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of artificial intelligence and intelligent document processing technology, and in particular to an automatic generation method for substation inspection reports based on artificial intelligence. Background Technology
[0002] With the continuous improvement of the intelligence and automation level of power systems, substation inspection work increasingly relies on the rapid acquisition and automatic processing of multimodal data from the field. However, existing substation inspection data analysis and report generation processes mostly adopt manual or semi-automatic methods, which have problems such as low data processing efficiency, reliance on human experience for anomaly identification, insufficient report standardization, and difficulty in archiving and tracing, and cannot meet the needs of large-scale, real-time, and intelligent operation and maintenance management.
[0003] In recent years, although some technical solutions have introduced anomaly detection algorithms and templated reporting tools based on deep learning, they generally suffer from shortcomings such as poor interpretability, insufficient adaptability to data types, weak automatic summarization capabilities, and incomplete collaborative management loops. How to achieve efficient and interpretable anomaly analysis of multimodal substation detection data, as well as the automatic generation, review, archiving, and distribution of detection reports, has become a key technical problem that urgently needs to be solved in the field of intelligent detection in the power industry.
[0004] Therefore, how to provide an AI-based method for automatically generating substation inspection reports is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention
[0005] One objective of this invention is to propose an AI-based method for automatically generating substation inspection reports. This invention deeply integrates interpretable AI anomaly detection with automated document generation technology, achieving efficient acquisition of multimodal substation inspection data, accurate anomaly identification, and intelligent automatic report generation. This not only significantly improves the automation level and efficiency of the inspection process and reduces the risk of human error, but also achieves a closed-loop and standardized process for inspection report generation through mechanisms such as adaptive template matching, structured summary suggestion output, multi-role hierarchical collaborative review, electronic signature, and archiving distribution. This enhances the standardization, traceability, and intelligent decision-making capabilities of substation inspection management, demonstrating outstanding engineering practical value and broad application prospects.
[0006] An automatic substation inspection report generation method based on artificial intelligence according to an embodiment of the present invention includes the following steps: By acquiring multimodal data, raw test data from the substation testing site is obtained, and the raw test data is preprocessed to obtain primary test data. The rule engine is used to perform preliminary verification of the primary detection data according to industry standards, and trend analysis and anomaly detection are performed on the verified primary detection data to generate structured detection data. The structured detection data is input into the automatic summary and suggestion generation module based on the BART model to generate an analysis summary and professional suggestions for the detection results; Fill the standardized test report template with structured test data, analysis summaries, and professional recommendations. The template is matched according to the type of test item. An automated typesetting method is used to generate a standardized test report. The test report is then pushed to a multi-role collaboration module, where relevant reviewers perform electronic signatures and hierarchical review to obtain a confirmed draft. After review and confirmation, the manuscript is archived and distributed to the relevant management system or designated terminal according to the set rules.
[0007] Furthermore, the multimodal data acquisition includes: The type of inspection item is obtained from the inspection task order or work order system; The equipment category and equipment number of the testing equipment can be obtained through equipment information database, RFID scanning or image recognition. The system collects physical parameters of substation monitoring equipment, including electrical quantities, ambient temperature and humidity, and vibration signals, using multiple types of sensors. Image information from equipment nameplates, pointer instrument readings, or test forms is acquired through an image acquisition device, and text information is extracted from the images using optical character recognition technology. All information is tagged with a unified timestamp and normalized and structured in the data acquisition module to generate complete primary test data including test item type, equipment category, and test parameters.
[0008] Furthermore, the preliminary verification of the initial detection data using a rule engine according to industry standards includes: For each piece of preliminary test data, the system automatically matches the corresponding numerical range threshold, data unit specifications, and logical relationships between parameters according to the industry standards to which the test item belongs. A set of procedural rules is used to perform range verification, unit verification, and cross-parameter consistency verification on the primary detection data in sequence, and data that does not conform to the rules is marked and classified as anomaly. The trend analysis and anomaly detection include: For the preprocessed multimodal primary detection data, the dominant time series component of each verified primary detection data is extracted. Based on the physical operating conditions of the substation equipment, the dominant time series component is normalized with the same type of data under historical health conditions. The Isomap manifold learning algorithm is used to embed each detection parameter component into a high-dimensional feature space in a manifold space. Combined with the distribution of historical health condition samples in the manifold space, a health status feature cluster is dynamically constructed. The projection distance and local density distribution of the current monitoring data in the manifold space are jointly analyzed. If the manifold embedding distance of the current data exceeds the set threshold of the health status feature cluster center and the local density is significantly reduced, accompanied by a sudden change in the energy ratio of the physical parameter, it is judged as abnormal. Based on the topological location of anomalies in the manifold space, their historical evolution trajectory, and typical failure modes in the domain knowledge base, anomalies are attributed and classified, and structured detection data including manifold anomaly degree, anomaly attribution type, risk level, and analysis suggestions are automatically output.
[0009] Furthermore, the automatic summary and suggestion generation module includes: The manifold anomaly degree, anomaly attribution type, risk level and analysis recommendations in the structured detection data are converted into the serialized text form required for the BART model input according to the set field mapping rules, including the concatenation of field names and corresponding values or categories, and arranged into an input sequence according to a predefined order; The input sequence is paired with manually annotated historical detection conclusions and suggested texts to form training sample pairs. The BART model is then trained in a supervised manner. During the training process, an improved SeqGAN loss function is used to optimize the similarity between the generated text and the reference text. In the automatic summarization stage, the newly collected structured detection data is processed in the same way and input into the trained BART model to obtain the natural language text sequence output by the model; The natural language text sequence is divided into analysis summary segments and professional advice segments according to preset delimiters, and the segmented text is sent to subsequent modules in the report generation process.
[0010] Furthermore, the improved SeqGAN loss function specifically includes: The structured detection data is processed by fields and then input into the BART generator to obtain the generated text sequence; The discriminator distinguishes between the generated text sequence and the manually annotated text sequence, and outputs a probability score that the text is manually annotated. Based on the discriminator's output score, the generator parameters are updated according to the policy gradient method to make the generated text sequence obtain a higher probability score; The discriminator parameters are optimized based on the discrimination error between the real text and the generated text; The discriminator includes multiple discriminant sub-units for structural differences and is equipped with a dynamic gating mechanism. Based on the parameter type, anomaly category, or risk level of the input structured detection data, it automatically selects or weights the outputs of the discriminant sub-units and uses them as the final discriminant score to participate in the loss function calculation and model parameter optimization.
[0011] Furthermore, the step of filling the standardized testing report template with structured testing data, analysis summaries, and professional recommendations includes: The system automatically retrieves and matches the corresponding standardized report template based on the type of test item, equipment category, and test parameters. The template includes a structured field area comprising a basic information area, a test result area, an anomaly analysis area, and an operation and maintenance suggestion area. Based on the mapping between the parameter labels in the structured detection data and the template fields, the detection parameter values, anomaly criteria, trend components, risk levels, and attribution labels are filled into the corresponding positions in the template; For analysis summaries and professional recommendations, the BART model is used to automatically determine the chapters and display order of their embedded templates. If there are abnormal types or risk events in the detection report, the content fields of the abnormal analysis area and the recommendation area are automatically expanded. After the data is populated, verify the integrity of the template, the consistency of the fields, and the logical order.
[0012] Furthermore, the step of generating a test report with standardized layout using automated typesetting specifically includes: The layout of the completed test report template is optimized using an automated typesetting engine, automatically adjusting the title, section, chart, paragraph layout and font style; The automatically generated test report is pushed to a collaborative system with multi-role hierarchical management in editable electronic document or PDF format, supporting project leaders, technical reviewers, and management reviewers to participate in the review process in parallel or sequentially. In the collaborative system, relevant reviewers review, annotate, and process the content of the test report through their respective authorized nodes, and use integrated electronic signature technology to perform hierarchical electronic signatures on the review opinions and confirmation drafts. Once the review is completed, the system automatically generates a confirmation draft of the test report with a complete signature chain and review log for archiving and distribution.
[0013] Furthermore, the step of archiving and distributing the confirmed test report after review and confirmation includes: The confirmed test report, which has completed all hierarchical review and electronic signature, is archived synchronously to the local database and the remote cloud archiving system in an encrypted archiving format. The archived information includes the report's unique number, archiving time, review signature chain, and metadata index. According to the preset distribution rules, the confirmation draft of the test report will be automatically distributed in the form of an electronic document to relevant management systems, project file systems, supervision platforms or designated terminals, including enterprise management information systems, equipment operation and maintenance platforms or the email address of the operation and maintenance manager. The distribution process supports automatic notifications, hierarchical access permissions, and distribution logs. All archiving and distribution operations can be retrieved and traced by testing item, device number, and time field.
[0014] The beneficial effects of this invention are: This invention deeply integrates an interpretable AI anomaly detection algorithm with an automated document generation process, enabling efficient acquisition of multimodal substation inspection data, accurate anomaly identification, and automatic intelligent report generation. Compared to traditional manual analysis and template-based reporting methods, this invention significantly improves the automation level of data processing and report generation, substantially shortens the inspection process cycle, reduces the risk of human error, and enhances the standardization and traceability of inspection results.
[0015] This invention also achieves full-process closed-loop automation and standardization of test report generation through mechanisms such as adaptive template matching, structured summary suggestion output, multi-role hierarchical collaborative review and electronic signature, archiving and intelligent distribution. It improves the efficiency, transparency and intelligent decision support capabilities of substation test management, and has high engineering practical value and broad application prospects. Attached Figure Description
[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:
[0017] Figure 1 This is a flowchart of an artificial intelligence-based automatic substation inspection report generation method proposed in this invention; Figure 2 This is a closed-loop structure diagram of the intelligent report generation method for automatically generating substation inspection reports based on artificial intelligence, as proposed in this invention. Detailed Implementation
[0018] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.
[0019] refer to Figure 1-2 An AI-based method for automatically generating substation inspection reports includes the following steps: Through multimodal data acquisition, raw detection data from the substation inspection site is obtained. This raw data is preprocessed to obtain primary detection data. Multiple sensors and acquisition devices, including those for current, voltage, temperature, vibration, partial discharge, and images, are used to synchronously acquire data from the target equipment in the substation. All types of raw data are then uniformly aggregated through edge acquisition terminals and undergo standardized preprocessing, including missing value completion, noise filtering, normalization, and time-series alignment, to obtain a structurally consistent primary detection dataset.
[0020] The system utilizes a rule engine to perform preliminary verification of initial testing data according to industry standards. Following verification, it performs trend analysis and anomaly detection on the initial testing data, generating structured testing data. The rule engine is then invoked to automatically load industry standards or enterprise specifications based on the testing task sheet or equipment category, performing range, unit, and consistency checks on each piece of initial testing data. Data that passes verification proceeds to the trend analysis and anomaly detection process. Signal decomposition and manifold learning methods can be used to identify and attribute anomalies, ultimately outputting structured testing data with multi-field annotations.
[0021] The structured detection data is input into the automatic summary and suggestion generation module based on the BART model, which generates an analysis summary and professional suggestions for the detection results. The structured detection data is mapped to the standard input format of the BART model. The finely tuned BART model automatically generates the detection conclusion summary and professional suggestion text, and segments them according to the conclusion and suggestion fields. Structured test data, analysis summaries, and professional recommendations are populated into the standardized test report template. The template is matched according to the test project type and automatically selects the matching standardized test report template (including basic information area, test result area, anomaly analysis area, and recommendation area) based on the test work order or equipment asset information, and populates the data into the corresponding modules according to the field rules. An automated typesetting method is used to generate standardized test reports. These reports are then pushed to a multi-role collaboration module where relevant reviewers perform electronic signatures and hierarchical review. Once a confirmed draft is obtained, the typesetting engine automatically adjusts the report layout, table of contents, pagination, fonts, and chart formats, exporting a standard electronic report. The report circulates through the collaboration system to multiple roles, including testers, reviewers, and managers. Reviewers at each level annotate, confirm, and complete electronic signatures online.
[0022] After review and confirmation, the manuscript is archived and distributed to relevant management systems or designated terminals according to the set rules. The final confirmation report is archived in local and cloud storage systems by project number, time, device and other tags, and is automatically pushed to the management platform, operation and maintenance platform or relevant responsible terminals to achieve traceable archiving and automatic distribution.
[0023] In this embodiment, the multimodal data acquisition includes: The inspection task or work order system automatically obtains the inspection item type corresponding to the current inspection task and writes it as a data tag into the primary inspection data. The equipment category and equipment number of the testing equipment are automatically obtained through equipment information database retrieval, RFID code scanning identification or equipment nameplate image recognition, and then linked and bound to the testing data; By using multiple types of sensors, including current, voltage, insulation resistance, temperature, humidity, and vibration sensors, the on-site operating conditions of the target equipment are collected synchronously to obtain the raw data stream of various physical parameters. Image information including equipment nameplates, pointer meter readings, and test forms is acquired through image acquisition devices (such as handheld terminals and cameras), and optical character recognition (OCR) technology is used to automatically extract equipment names, numerical readings, and handwritten text information from the images. All collected data are time-stamped with a unified timestamp, and preprocessing operations, including data normalization, missing value imputation, and field structuring, are uniformly completed in the data acquisition module. Finally, primary detection data containing detection item type, equipment category, equipment number, and detection parameters are generated, providing a data foundation for subsequent rule verification, analysis, and automatic matching of report templates.
[0024] In this embodiment, the preliminary verification of the initial detection data using a rule engine according to industry standards includes: For each piece of preliminary test data, the system automatically matches the corresponding numerical range threshold, data unit specifications, and logical relationships between parameters according to the industry standards to which the test item belongs. For each piece of primary test data, first read the test item type, equipment category, and test parameter fields associated with that data; The built-in rule database automatically retrieves the corresponding industry standard clauses by using the detection project type, equipment category and parameter name as composite search conditions. This includes the standard numerical range of the parameter, unit specifications and logical relationships with other parameters (such as the correlation curve between the main transformer winding temperature and the load current). The retrieved standard information is stored in the rule database in the form of tables or key-value pairs, as shown in Table 1: Table 1. Example table of rule database
[0025] During the verification process, the programmed rule set sequentially calls the aforementioned standard data to process the input primary detection data: Range verification (e.g., insulation resistance ≥10MΩ); Unit verification (e.g., if the data collection unit is kΩ, it needs to be converted to MΩ and compared with the standard); Cross-parameter logic relationship verification (such as matching the sum of multiphase currents with zero-sequence current, or judging the rationality of temperature and load). If data is found to be outside the allowed range, with inconsistent units, or with abnormal logical relationships during the verification process, the data will be automatically marked as abnormal and classified into abnormal types (such as "out of limit", "unit abnormality", "logical contradiction"). Abnormal data and their types will be written into structured detection data for subsequent processing and report generation.
[0026] The trend analysis and anomaly detection include: For the preprocessed multimodal primary detection data, the CEEMDAN (Complete Ensemble Empirical Mode Decomposition) method is used to decompose the signal of each calibrated primary detection parameter (such as current, temperature, vibration, and insulation resistance). The dominant intrinsic mode component (IMF) is selected from the decomposition results as the representative time-series component reflecting the dynamic characteristics of the equipment's current operating condition. Subsequently, combined with the equipment's real-time operating parameters (such as load, grid frequency, and ambient temperature and humidity), the aforementioned dominant time-series component is Z-score normalized with the same type of dominant component under the same operating conditions and historical health status of the same equipment.
[0027] ; in, The dominant component after normalization The feature vector of the current dominant component. and These represent the mean and standard deviation of the corresponding features for the health history samples.
[0028] The normalized dominant components of all detection parameters are concatenated according to the detection time points to form observation samples, thus constructing a multidimensional input feature matrix. The Isomap (isometric mapping) manifold learning algorithm is used to embed this feature matrix from high to low dimensions. Specifically, this involves: first, selecting an appropriate number of K-nearest neighbors, calculating the distance between samples based on Euclidean distance, and constructing a K-nearest neighbor graph; then, calculating the shortest path distance between all sample pairs, and performing multidimensional scaling (MDS) analysis on the distance matrix, ultimately obtaining the embedding coordinates of each sample in a low-dimensional manifold space (such as two-dimensional or three-dimensional space).
[0029] For historical healthy operating condition samples, the above processing procedure is repeated, and their embedding coordinates in the manifold space are used as the set of healthy reference points. DBSCAN (density-based spatial clustering algorithm) is used to perform cluster analysis on these healthy samples in the manifold space, dynamically forming clusters of healthy status features. The system automatically calculates the center coordinates, radius, and average neighborhood point density of the healthy clusters, which serve as the baseline for subsequent anomaly detection.
[0030] In the actual detection process, the newly acquired current monitoring data is processed using the same feature extraction, normalization, and Isomap embedding to obtain the manifold space coordinates of the current detection sample. Subsequently, the Euclidean distance from the embedding point to the center of the healthy state feature cluster is calculated, and the number of healthy samples within a neighborhood of this point with a set radius (such as the radius of the healthy cluster) is counted to obtain its local point density.
[0031] Anomaly detection criteria include the following three items: If the distance from the current embedding point to the center of the healthy feature cluster is greater than a set multiple (e.g., 1.5 times) of the radius of the healthy cluster. If the number of healthy samples in the neighborhood of the current point is significantly lower than the average neighborhood density of healthy samples (e.g., less than 50%). The dominant component energy ratio of the current detection parameter is dynamically monitored. If the energy ratio shows a significant change in the historical healthy distribution (e.g., greater than the historical mean ± 2 standard deviations), it is judged as abnormal.
[0032] For all detected samples deemed abnormal, the system further analyzes their topological location in the manifold space, including the distribution characteristics of abnormal points on the periphery of healthy clusters, their distance relationship with healthy-fault samples, and the evolution trajectory of historical detection data. Subsequently, the system retrieves typical fault mode samples labeled in the domain knowledge base, compares the abnormal points with the manifold coordinate characteristics of these known modes, and automatically assigns anomaly types (e.g., insulation aging, mechanical loosening, short-circuit discharge) using minimum distance criteria or density distribution overlap criteria. Risk level labels (e.g., general, severe, extremely high) are assigned based on the aforementioned distance, density, and energy ratio indicators, while simultaneously generating analysis suggestion fields corresponding to the attribution type. Finally, all results are output in structured data format, with fields including manifold anomaly degree, anomaly attribution type, risk level, and analysis suggestions, providing a basis for subsequent automatic generation of detection reports and intelligent operation and maintenance decisions.
[0033] In this embodiment, the automatic summary and suggestion generation module includes: The manifold anomaly degree, anomaly attribution type, risk level and analysis recommendations in the structured detection data are converted into the serialized text form required for the BART model input according to the set field mapping rules, including the concatenation of field names and corresponding values or categories, and arranged into an input sequence according to a predefined order; The input sequence is combined with manually annotated historical detection conclusions and suggested texts to form training sample pairs. The BART model is then subjected to supervised fine-tuning training. During the training process, an improved SeqGAN loss function is used to jointly optimize the distribution difference between generated text and real text using the policy gradient method, thereby improving the model's expressive ability and generalization ability in complex detection scenarios. In the automatic summarization stage, the newly collected structured detection data is processed in the same way and input into the trained BART model to obtain the natural language text sequence output by the model; The natural language text sequence is divided into analysis summary segments and professional suggestion segments according to preset delimiters. The segmented text is sent to subsequent modules of the report generation process to realize the automatic filling and display of analysis conclusions and suggestions.
[0034] In this embodiment, the improved SeqGAN loss function specifically includes: The structured detection data, after field mapping and serialization, is concatenated into an input sequence according to a preset field order and then input into the BART text generator. The BART generator, based on the Transformer architecture, outputs a natural language text sequence corresponding to the input data.
[0035] Subsequently, the text sequence output by the BART generator and the manually annotated historical detection conclusions and suggested texts are input into the discriminator. The discriminator contains three structurally differentiated sub-discriminators: 1. A text discriminator based on a one-dimensional convolutional neural network (1D-CNN) to capture local feature patterns in short texts of detection reports; 2. A text discriminator based on a bidirectional long short-term memory network (BiLSTM) to extract temporal dependencies in long texts; and 3. A Transformer discriminator based on a multi-head self-attention mechanism to identify the association structure of key industry terms.
[0036] During each training iteration, the system dynamically calculates the weighting coefficients of the three discriminant sub-units in real time based on the parameter type (e.g., insulation parameters, environmental parameters, load parameters), anomaly category (e.g., insulation degradation, temperature anomaly, excessive load), and risk level (e.g., general, high, extremely high) in the input structured detection data. For example, when the input is an insulation parameter and the risk level is extremely high, the gating unit will increase the weights of the BiLSTM discriminator to enhance its ability to discriminate long text descriptions and anomaly attributions. When the input involves multiple parameters and complex anomalies, the gating mechanism allocates weights to each sub-discriminant unit, proportionally weights the three discriminant scores according to the input features, and uses the weighted score as the final dynamic discriminant score.
[0037] During the adversarial training phase, the REINFORCE policy gradient method is employed, using dynamic discrimination scores as reward signals to update the BART generator parameters. Specifically, for each round of sampled text, the probability of generating high-scoring text is increased based on the discrimination score, while the probability of generating low-scoring text is decreased. The discriminator parameters are then iteratively optimized using the cross-entropy loss function to determine the probability of classifying real text as true and generated text as false.
[0038] The entire model training process is iterative, and the final output BART generator can automatically generate highly credible and professional detection report summaries and suggestion texts. It can also adapt to various detection parameters and scenarios to meet the intelligent generation needs of multi-scenario detection reports in substations.
[0039] In this embodiment, the step of filling the standardized test report template with structured test data, analysis summary, and professional recommendations includes: The system automatically retrieves and matches the corresponding standardized report template based on the type of test item, equipment category, and test parameters. The template includes a structured field area comprising a basic information area, a test result area, an anomaly analysis area, and an operation and maintenance suggestion area. Based on the mapping between the parameter labels in the structured detection data and the template fields, the detection parameter values, anomaly criteria, trend components, risk levels, and attribution labels are filled into the corresponding positions in the template; For analysis summaries and professional recommendations, the BART model is used to automatically determine the chapters and display order of their embedded templates. If there are abnormal types or risk events in the detection report, the content fields of the abnormal analysis area and the recommendation area are automatically expanded. After the data is populated, verify the integrity of the template, the consistency of the fields, and the logical order.
[0040] Specifically, it includes: During the test report generation phase, based on the current test item type, equipment category, and test parameters, the data acquisition module automatically reads relevant field information and uses this as search criteria to automatically retrieve and match the corresponding standardized report template in the template management library. This template is pre-divided into structured field areas such as basic information, test results, anomaly analysis, and maintenance suggestions. Each block has a fixed field name and order, and some blocks can be dynamically expanded based on the test content.
[0041] Subsequently, based on the mapping between the parameter labels in the structured inspection data and each field of the template, the collected inspection parameter values, anomaly criteria, trend components, risk levels, and attribution labels are automatically filled into the corresponding fields according to the template requirements. For example, parameters such as insulation resistance, current, and temperature are filled into the inspection results area, anomaly criteria and trend components are filled into the anomaly analysis area, and risk levels and attribution labels can be written into the corresponding positions in both the inspection results area and the anomaly analysis area. During the field mapping process, the system automatically verifies the consistency of parameter units, numerical ranges, and names to ensure the accuracy of the filling and the template specifications.
[0042] For the analysis summary and expert recommendations sections, the natural language text output by the finely tuned BART model is parsed into paragraphs. During text generation, the BART model uses special markers (such as "[Analysis Summary Paragraph]" and "[Expert Recommendation Paragraph]") or paragraph separators to segment the output text according to chapter content. Based on the paragraph tags in the BART output, the report generation engine automatically inserts the analysis summary paragraphs into the template's detection results area or anomaly analysis area, and the expert recommendations paragraphs into the operation and maintenance recommendations area. If the detection report contains multiple anomaly types, risk events, or requires multiple recommendations, the template extension interface can be automatically invoked to dynamically add corresponding field blocks to the anomaly analysis area and operation and maintenance recommendations area, enabling a multi-column display of multiple anomalies and recommendations.
[0043] Once all data and text content are populated, the template integrity is automatically checked. This includes verifying that all required fields contain data, that parameter units match the template, and that there are no inconsistencies between logical fields (e.g., anomalies exist but no suggestions are provided, or suggestions do not match the risk level). Simultaneously, the order of all fields is checked for consistency to ensure the generated test report has a standardized structure and no missing content.
[0044] Finally, the completed report will go through subsequent processes such as automated typesetting, review and signing to ensure the standardization, normalization and traceability of the report content.
[0045] In this embodiment, the step of generating a layout specification test report using automated typesetting specifically includes: The layout of the completed test report template is optimized using an automated typesetting engine, automatically adjusting the title, section, chart, paragraph layout and font style; The automatically generated test report is pushed to a collaborative system with multi-role hierarchical management in editable electronic document or PDF format, supporting project leaders, technical reviewers, and management reviewers to participate in the review process in parallel or sequentially. In the collaborative system, relevant reviewers review, annotate, and process the content of the test report through their respective authorized nodes, and use integrated electronic signature technology to perform hierarchical electronic signatures on the review opinions and confirmation drafts. Once the review is completed, a confirmation draft of the test report with a complete signature chain and review log is automatically generated for archiving and distribution.
[0046] In this embodiment, the step of archiving and distributing the confirmed test report after review and confirmation includes: The confirmed test report, which has completed all hierarchical review and electronic signature, is archived synchronously to the local database and the remote cloud archiving system in an encrypted archiving format. The archived information includes the report's unique number, archiving time, review signature chain, and metadata index. According to the preset distribution rules, the confirmation draft of the test report will be automatically distributed in the form of an electronic document to relevant management systems, project file systems, supervision platforms or designated terminals, including enterprise management information, equipment operation and maintenance platforms or the email address of the operation and maintenance manager. The distribution process supports automatic notifications, hierarchical access permissions, and distribution logs. All archiving and distribution operations can be retrieved and traced by testing item, device number, and time field.
[0047] Example 1: To verify the feasibility of this invention in practice, it was applied to the insulation condition detection scenario of the main transformer of a 110kV substation. Five consecutive sets of monitoring data from a certain day in May 2024 were selected, and the core parameters are shown in Table 2.
[0048] Table 2 Monitoring Data Table
[0049] Taking test number 4 as an example, the original collected parameters are as follows: insulation resistance 11.7MΩ; winding temperature 80.9℃; load current 450A; ambient temperature 26.9℃; partial discharge quantity 42.5pC; The historical healthy sample mean and standard deviation were 25.0 MΩ and 2.0 MΩ, respectively. The normalized insulation resistance was: Similarly, normalize all parameters. Feed the normalized parameters into the Isomap algorithm (K=5, reduced to two dimensions) to obtain the sample's embedding coordinates (0.72, 0.95). The historical healthy cluster center is (0.30, 0.40), with a radius of 0.20. Calculate the distance to the healthy cluster center: Within a radius of 0.20, there is only one healthy sample in the neighborhood of this point. The average number of healthy samples is 5, and the density is 0.2, which is far below the average. The dominant component energy ratio is 0.37, while the mean of healthy samples is 0.13 and the standard deviation is 0.02. Therefore, the energy ratio of this sample is more than twice the standard deviation of the mean, and it is judged as an energy ratio mutation. The distance of this sample from the healthy center is 0.69 > 0.20, and the density is 0.2 < 0.5, indicating an energy ratio mutation. The anomaly criteria are met, and the sample is judged as abnormal.
[0050] Based on the manifold spatial distribution and historical fault knowledge base, this anomaly is attributed to "insulation deterioration + high temperature + overload," with a risk level of "extremely high." The system automatically generates the following recommendation: "Immediate shutdown and maintenance are recommended, along with a comprehensive diagnosis and maintenance of the main transformer's insulation system."
[0051] Other samples, serial numbers 1, 2, and 5, were processed in the same way and all fell within the radius of the healthy cluster. They showed high local density and no energy ratio mutations, so the system judged them to be normal and no special recommendations were needed.
[0052] This embodiment fully demonstrates how to directly determine the health / abnormal status of actual samples through data normalization, manifold embedding, and joint criterion calculation, and output attribution and suggestions. All calculation steps are supported by real data and conclusions, and can be directly reproduced in engineering.
[0053] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. A method for automatically generating substation inspection reports based on artificial intelligence, characterized in that, The steps include: By acquiring multimodal data, raw test data from the substation testing site is obtained, and the raw test data is preprocessed to obtain primary test data. The rule engine is used to perform preliminary verification of the primary detection data according to industry standards, and trend analysis and anomaly detection are performed on the verified primary detection data to generate structured detection data. The structured detection data is input into the automatic summary and suggestion generation module based on the BART model to generate an analysis summary and professional suggestions for the detection results; Fill the standardized test report template with structured test data, analysis summaries, and professional recommendations. The template is matched according to the type of test item. An automated typesetting method is used to generate a standardized test report. The test report is then pushed to a multi-role collaboration module, where relevant reviewers perform electronic signatures and hierarchical review to obtain a confirmed draft. After review and confirmation, the manuscript is archived and distributed to the relevant management system or designated terminal according to the set rules.
2. The method for automatically generating substation inspection reports based on artificial intelligence according to claim 1, characterized in that, The multimodal data acquisition includes: The type of inspection item is obtained from the inspection task order or work order system; The equipment category and equipment number of the testing equipment can be obtained through equipment information database, RFID scanning or image recognition. The system collects physical parameters of substation monitoring equipment, including electrical quantities, ambient temperature and humidity, and vibration signals, using multiple types of sensors. Image information from equipment nameplates, pointer instrument readings, or test forms is acquired through an image acquisition device, and text information is extracted from the images using optical character recognition technology. All information is tagged with a unified timestamp and normalized and structured in the data acquisition module to generate complete primary test data including test item type, equipment category, and test parameters.
3. The method for automatically generating substation inspection reports based on artificial intelligence according to claim 1, characterized in that, The preliminary verification of the initial detection data using a rule engine according to industry standards includes: For each piece of preliminary test data, the system automatically matches the corresponding numerical range threshold, data unit specifications, and logical relationships between parameters according to the industry standards to which the test item belongs. A set of procedural rules is used to sequentially perform range verification, unit verification, and cross-parameter consistency verification on the initial detection data, and data that does not conform to the rules is marked and classified as anomaly. The trend analysis and anomaly detection include: For the preprocessed multimodal primary detection data, the dominant time series component of each verified primary detection data is extracted. Based on the physical operating conditions of the substation equipment, the dominant time series component is normalized with the same type of data under historical health conditions. The Isomap manifold learning algorithm is used to embed each detection parameter component into a high-dimensional feature space in a manifold space. Combined with the distribution of historical health condition samples in the manifold space, a health status feature cluster is dynamically constructed. The projection distance and local density distribution of the current monitoring data in the manifold space are jointly analyzed. If the manifold embedding distance of the current data exceeds the set threshold of the health status feature cluster center and the local density is significantly reduced, accompanied by a sudden change in the energy ratio of the physical parameter, it is judged as abnormal. Based on the topological location of anomalies in the manifold space, their historical evolution trajectory, and typical failure modes in the domain knowledge base, anomalies are attributed and classified, and structured detection data including manifold anomaly degree, anomaly attribution type, risk level, and analysis suggestions are automatically output.
4. The method for automatically generating substation inspection reports based on artificial intelligence according to claim 1, characterized in that, The automatic summary and suggestion generation module includes: The manifold anomaly degree, anomaly attribution type, risk level and analysis recommendations in the structured detection data are converted into the serialized text form required for the BART model input according to the set field mapping rules, including the concatenation of field names and corresponding values or categories, and arranged into an input sequence according to a predefined order; The input sequence is paired with manually annotated historical detection conclusions and suggested texts to form training sample pairs. The BART model is then trained in a supervised manner. During the training process, an improved SeqGAN loss function is used to optimize the similarity between the generated text and the reference text. In the automatic summarization stage, the newly collected structured detection data is processed in the same way and input into the trained BART model to obtain the natural language text sequence output by the model; The natural language text sequence is divided into analysis summary segments and professional advice segments according to preset delimiters, and the segmented text is sent to subsequent modules in the report generation process.
5. The method for automatically generating substation inspection reports based on artificial intelligence according to claim 4, characterized in that, The improved SeqGAN loss function specifically includes: The structured detection data is processed by fields and then input into the BART generator to obtain the generated text sequence; The discriminator distinguishes between the generated text sequence and the manually annotated text sequence, and outputs a probability score that the text is manually annotated. Based on the discriminator's output score, the generator parameters are updated according to the policy gradient method to make the generated text sequence obtain a higher probability score; The discriminator parameters are optimized based on the discrimination error between the real text and the generated text; The discriminator includes multiple discriminant sub-units for structural differences and is equipped with a dynamic gating mechanism. Based on the parameter type, anomaly category, or risk level of the input structured detection data, it automatically selects or weights the outputs of the discriminant sub-units and uses them as the final discriminant score to participate in the loss function calculation and model parameter optimization.
6. The method for automatically generating substation inspection reports based on artificial intelligence according to claim 1, characterized in that, The steps for filling structured test data, analysis summaries, and professional recommendations into the standardized test report template include: The system automatically retrieves and matches the corresponding standardized report template based on the type of test item, equipment category, and test parameters. The template includes a structured field area comprising a basic information area, a test result area, an anomaly analysis area, and an operation and maintenance suggestion area. Based on the mapping between the parameter labels in the structured detection data and the template fields, the detection parameter values, anomaly criteria, trend components, risk levels, and attribution labels are filled into the corresponding positions in the template; For analysis summaries and professional recommendations, the BART model is used to automatically determine the chapters and display order of the embedded templates. If there are abnormal types or risk events in the detection report, the content fields of the abnormal analysis area and the recommendation area are automatically expanded. After the data is populated, verify the integrity of the template, the consistency of the fields, and the logical order.
7. The method for automatically generating substation inspection reports based on artificial intelligence according to claim 1, characterized in that, The steps for generating a layout specification test report using automated typesetting methods specifically include: The layout of the completed test report template is optimized using an automated typesetting engine, automatically adjusting the title, section, chart, paragraph layout and font style; The automatically generated test report is pushed to a collaborative system with multi-role hierarchical management in editable electronic document or PDF format, supporting project leaders, technical reviewers, and management reviewers to participate in the review process in parallel or sequentially. In the collaborative system, relevant reviewers review, annotate, and process the content of the test report through their respective authorized nodes, and use integrated electronic signature technology to perform hierarchical electronic signatures on the review opinions and confirmation drafts. Once the review is completed, the system automatically generates a confirmation draft of the test report with a complete signature chain and review log for archiving and distribution.
8. The method for automatically generating substation inspection reports based on artificial intelligence according to claim 1, characterized in that, The steps for archiving and distributing the confirmed test report after review and confirmation include: The confirmed test report, which has completed all hierarchical review and electronic signature, is archived synchronously to the local database and the remote cloud archiving system in an encrypted archiving format. The archived information includes the report's unique number, archiving time, review signature chain, and metadata index. According to the preset distribution rules, the confirmation draft of the test report will be automatically distributed in the form of an electronic document to relevant management systems, project file systems, supervision platforms or designated terminals, including enterprise management information systems, equipment operation and maintenance platforms or the email address of the operation and maintenance manager. The distribution process supports automatic notifications, hierarchical access permissions, and distribution logs. All archiving and distribution operations can be retrieved and traced by testing item, device number, and time field.
Citation Information
Cited By
Intelligent auditing method and system for daily report data of power production and operation
CN121639122A
Intelligent auditing method and system for power production operation daily report data
CN121639122B
Engineering production line report automatic generation method and system based on artificial intelligence
CN121808469A
Structural health report intelligent generation method based on generative AI
CN121936556A
