A wind reconnaissance report automatic generation method and system

Abstract

The application discloses a kind of wind reconnaissance report automatic generation method and system, by the feature matching of effective image to scene, and based on the pre-set engineering risk checklist, the confidence score of each risk item is determined, qualitative judgment of traditional dependence on engineer personal experience is changed into quantitative evaluation based on standardization checklist, and this mechanism eliminates the judgment standard of artificial factor and ensures that the identification result of the same risk item in different scenes has high consistency and objectivity, significantly improves the standardization level of risk assessment;Second, according to the engineering report template, the risk items included in the engineering report are automatically generated structured report content, and the original scene survey image is associated, the automatic conversion from data to report is realized, at the same time, by automatically associating the original survey image, it is ensured that the report content is strongly bound with the scene evidence, and the generation speed of the report and the data traceability are significantly improved.

Description

Technical Field

[0001] This invention relates to the field of wind survey report generation technology, and in particular to a method and system for automatically generating wind survey reports. Background Technology

[0002] In the field of wind survey (wind load survey) and related engineering testing, on-site survey is the core link to obtain engineering structural safety data. The traditional wind survey report generation process is highly dependent on manual operation, and there are the following pain points: (1) Risk identification is highly subjective and has a low degree of standardization: Traditional risk identification mainly relies on the personal experience of engineers. The judgment standards for the same risk item are different, resulting in poor consistency and objectivity of the report. There is a lack of automated matching methods based on preset checklists and quantitative confidence scores, making it difficult to accurately locate risks; (2) Report writing is cumbersome and data association is difficult: The report writing process involves a lot of repetitive text input and chart organization work, and the association between the original image and the report conclusion is often loose, making it difficult to achieve accurate traceability of text and image correspondence, which increases the difficulty of later review. Summary of the Invention

[0003] In view of this, the present invention proposes an automatic risk survey report generation method and system, which can effectively solve the defects of existing technologies, such as strong subjectivity in risk identification, low standardization, cumbersome report writing, and difficulty in data correlation.

[0004] The technical solution of this invention is implemented as follows:

[0005] A method for automatically generating wind survey reports, specifically including:

[0006] Acquire on-site survey image data and perform image quality screening on the on-site survey image data to obtain valid on-site images;

[0007] Feature matching is performed on the valid images at the site, and the confidence score of each risk item is determined based on the preset engineering risk checklist;

[0008] The risk items to be included in the project report are obtained by comparing the confidence score with the preset threshold.

[0009] Based on the engineering report template, the system automatically generates structured report content by combining the risk items included in the engineering report, and links it to the original site survey images.

[0010] As a further optional solution to the aforementioned method for automatically generating wind survey reports, the step of performing image quality screening on the field survey image data to obtain valid field images specifically includes:

[0011] The acquired field survey image data is denoised and its color space is converted to obtain a grayscale image of the field survey.

[0012] The frequency domain energy of the grayscale image of the field survey was calculated using the Laplace variance method, and the frequency domain energy variance value was obtained.

[0013] By comparing the frequency domain energy variance value with the dynamic sharpness threshold, clear on-site survey images are obtained.

[0014] Based on clear on-site survey images, obtain the perspective matrix formed by the vanishing points of the images, and call the preset standard survey view vector;

[0015] Calculate the cosine similarity between the perspective matrix formed by the vanishing points of the image and the standard viewpoint vector of the image, and select the on-site survey images with the correct shooting angle based on the cosine similarity.

[0016] A lightweight CNN classification network is used to classify the scene of the on-site survey images with the correct shooting angle, so as to obtain the on-site survey images belonging to the scene related to wind survey.

[0017] As a further optional solution to the aforementioned automatic wind survey report generation method, the setting of the dynamic clarity threshold specifically includes:

[0018] Extract the global illumination intensity value from the grayscale image of the on-site survey;

[0019] A base sharpness threshold is set as the initial value, which is the Laplace variance determination threshold under standard lighting conditions;

[0020] The light attenuation coefficient is calculated based on the ratio of the global light intensity value to the standard light intensity value.

[0021] Multiply the base sharpness threshold by the illumination attenuation coefficient to obtain the dynamic sharpness threshold corresponding to the current scene.

[0022] As a further optional solution to the aforementioned method for automatically generating risk survey reports, the step of performing feature matching on valid on-site images and determining the confidence score for each risk item based on a preset engineering risk checklist specifically includes:

[0023] The effective images from the site are input into a pre-trained visual encoder to extract image feature vectors. The descriptive text of each risk item in the pre-set engineering risk checklist is input into a text encoder to generate text feature vectors corresponding to each risk item.

[0024] Calculate the cosine similarity between the image feature vector and the text feature vector corresponding to each risk item to obtain the preliminary matching score for each risk item;

[0025] By incorporating weighting coefficients and contextual factors from the industry knowledge base, and combining them with the preliminary matching score, the final confidence score for each risk item is calculated.

[0026] As a further optional solution to the aforementioned method for automatically generating wind survey reports, the step of automatically generating structured report content based on an engineering report template, incorporating risk items included in the engineering report, and associating it with original site survey images specifically includes:

[0027] Based on the JSON or XML schema structure of the engineering report template, obtain the mapping relationship between report chapters and risk item IDs;

[0028] Based on the mapping relationship, the risk items included in the engineering report are populated into the corresponding nodes of the engineering report template, and natural language descriptions are generated based on slot filling technology combined with the risk items;

[0029] Extract the EXIF ​​information of the original site survey images as metadata and embed it into the report attachment index. Calculate the SHA-256 hash value of the original site survey images and write it into the report's digital signature area.

[0030] As a further optional solution to the aforementioned method for automatically generating risk survey reports, the generation of natural language descriptions based on slot filling technology combined with risk items specifically includes:

[0031] The system calls upon a pre-set expert experience script library and, based on slot filling technology, matches the corresponding natural language description template from the script library according to the risk name, risk level, confidence score, and location information of each risk item.

[0032] Replace the slot variables in the natural language description template with the risk name, risk level, confidence score and location information of each risk item to generate natural language description text.

[0033] As a further optional solution to the aforementioned method for automatically generating wind survey reports, the method further includes performing integrity verification on the generated structured report, specifically including:

[0034] Iterate through all risk item IDs in the preset engineering risk check table, and check one by one whether there is a conclusion field in the report data structure corresponding to each risk item ID;

[0035] If the conclusion field corresponding to all risk item IDs is not empty, the report integrity check is considered to have passed.

[0036] If the conclusion field corresponding to any risk item ID is empty, the report integrity check is deemed to have failed, triggering an anomaly alert.

[0037] An automatic wind survey report generation system, comprising:

[0038] The image quality filtering module is used to acquire on-site survey image data and perform image quality filtering on the on-site survey image data to obtain valid on-site images;

[0039] The feature matching and confidence calculation module is used to perform feature matching on valid on-site images and determine the confidence score for each risk item based on a preset engineering risk checklist.

[0040] The risk item screening module is used to compare the confidence score with a preset threshold to obtain the risk items to be included in the project report;

[0041] The report generation module is used to automatically generate structured report content based on the engineering report template, combined with the risk items included in the engineering report, and to associate it with the original site survey images.

[0042] A computing device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of any of the above-described methods for automatically generating a wind survey report.

[0043] A computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of any of the above-described methods for automatically generating a wind survey report.

[0044] The beneficial effects of this invention are as follows: By performing feature matching on effective on-site images and determining the confidence score of each risk item based on a preset engineering risk checklist, the traditional qualitative judgment relying on the engineer's personal experience is transformed into a quantitative assessment based on a standardized checklist. This mechanism eliminates inconsistent judgment standards caused by human factors, ensuring a high degree of consistency and objectivity in the identification results of the same risk item in different scenarios, and significantly improving the standardization level of risk assessment. Secondly, according to the engineering report template, structured report content is automatically generated by combining the risk items included in the engineering report and linking them with the original on-site survey images, realizing the automated conversion from data to report. At the same time, by automatically linking the original survey images, a strong binding between the report content and on-site evidence is ensured, solving the problems of loose correspondence between text and images and difficulty in tracing the source in traditional reports, and significantly improving the report generation speed and data traceability. Attached Figure Description

[0045] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, 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 of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 This is a flowchart illustrating an automatic wind survey report generation method according to the present invention.

[0047] Figure 2This is a schematic diagram of the components of an automatic wind survey report generation system according to the present invention;

[0048] Figure 3 This is a schematic diagram of the composition of a computing device according to the present invention. Detailed Implementation

[0049] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0050] refer to Figures 1 to 3 A method for automatically generating wind survey reports, specifically including:

[0051] Acquire on-site survey image data and perform image quality screening on the on-site survey image data to obtain valid on-site images; in some embodiments, the step of performing image quality screening on the on-site survey image data to obtain valid on-site images specifically includes:

[0052] The acquired field survey image data is denoised and its color space is converted to obtain a grayscale image of the field survey.

[0053] The frequency domain energy of the grayscale image of the field survey was calculated using the Laplace variance method, and the frequency domain energy variance value was obtained.

[0054] By comparing the frequency domain energy variance value with the dynamic sharpness threshold, clear on-site survey images are obtained.

[0055] Based on clear on-site survey images, obtain the perspective matrix formed by the vanishing points of the images, and call the preset standard survey view vector;

[0056] Calculate the cosine similarity between the perspective matrix formed by the vanishing points of the image and the standard viewpoint vector of the image, and select the on-site survey images with the correct shooting angle based on the cosine similarity.

[0057] A lightweight CNN classification network is used to classify the scene of the on-site survey images with the correct shooting angle, so as to obtain the on-site survey images belonging to the scene related to wind survey.

[0058] Specifically, the process involves acquiring on-site survey image data, performing denoising and color space conversion on the acquired on-site survey image data to obtain a grayscale image of the on-site survey. In the specific operation, Gaussian filtering is used to remove high-frequency noise in the image, and then the RGB image is converted into a grayscale image for subsequent frequency domain energy analysis.

[0059] The frequency domain energy of the grayscale image of the field survey is calculated using the Laplacian variance method to obtain the frequency domain energy variance value. Then, based on a preset dynamic sharpness threshold (which can be adaptively adjusted according to the scene illumination intensity), the frequency domain energy variance value is compared with the dynamic sharpness threshold. If the variance value is lower than the dynamic sharpness threshold, the image is determined to be a blurry image and is removed; if the variance value is higher than or equal to the dynamic sharpness threshold, it is determined to be a clear field survey image and proceeds to the next step of screening.

[0060] Based on clear on-site survey images, a perspective matrix composed of the vanishing points of the images is obtained. At the same time, a preset standard survey perspective vector is invoked (this vector represents the standard shooting angle parameters in wind survey operations).

[0061] Calculate the cosine similarity between the perspective matrix formed by the vanishing points of the image and the standard view vector of the image. Set a similarity threshold (e.g., 0.95). If the cosine similarity is higher than the threshold, the shooting angle is determined to be correct, and the on-site survey images with correct shooting angles are selected. If it is lower than the threshold, the shooting angle deviation is determined to be too large, and it is rejected.

[0062] A lightweight CNN classification network is used to classify on-site survey images taken from the correct angle. The network is trained on wind survey-related scenes (such as wind turbine foundations, towers, blades, etc.) and can identify whether an image belongs to a wind survey-related scene. If the classification result is a wind survey-related scene, it is finally confirmed as a valid on-site image and used for subsequent risk identification and report generation.

[0063] Thus, by combining the Laplacian variance method with a dynamic sharpness threshold, low-quality images caused by inaccurate focus, motion blur, or insufficient lighting can be accurately eliminated. This avoids sending blurry images to the subsequent feature matching module, thereby preventing false or missed detections due to missing image information and ensuring that all data entering the risk assessment process are high-quality, valid images. Secondly, by using the cosine similarity between the perspective matrix and the standard viewpoint vector for screening, the on-site images are required to conform to the standard survey viewpoint. This solves the problem of structural dimension measurement errors caused by arbitrary angles in traditional manual shooting, ensuring that subsequent image-based risk item judgments (such as spacing, occlusion, etc.) have geometric accuracy and comparability. In addition, by using a lightweight CNN classification network for scene classification, images of non-situational scenes (such as sky, vegetation, irrelevant buildings, etc.) are effectively eliminated. This reduces the computational load of the system in processing invalid data, improves the operating efficiency of the feature matching module, and also avoids erroneous risk warnings caused by irrelevant background interference.

[0064] In some embodiments, setting the dynamic sharpness threshold specifically includes:

[0065] Extract the global illumination intensity value from the grayscale image of the on-site survey;

[0066] A base sharpness threshold is set as the initial value, which is the Laplace variance determination threshold under standard lighting conditions;

[0067] The light attenuation coefficient is calculated based on the ratio of the global light intensity value to the standard light intensity value.

[0068] Multiply the base sharpness threshold by the illumination attenuation coefficient to obtain the dynamic sharpness threshold corresponding to the current scene.

[0069] Specifically, after denoising and grayscale processing of the on-site survey images, the global illumination intensity value of the grayscale image is extracted simultaneously. This value represents the overall brightness level of the current image and reflects the actual lighting environment at the site.

[0070] A baseline sharpness threshold is preset. This threshold is the Laplacian variance determination value obtained through a large number of samples under standard lighting conditions (such as indoor standard light source or sunny noon), which serves as the benchmark for judging whether an image is sharp or not. Then, the ratio of the global illumination value of the current image to the standard illumination value is calculated. If the current illumination is weak (ratio less than 1), an illumination attenuation coefficient greater than 1 is calculated. If the illumination is strong, the coefficient is adjusted accordingly. This step aims to quantify the impact of insufficient illumination on the perception of image sharpness.

[0071] The calculated light attenuation coefficient is multiplied by the base sharpness threshold to generate a dynamic sharpness threshold that adapts to the current scene. For example, in a dark twilight environment, the standard Laplacian variance may be low, but by multiplying it by the light attenuation coefficient, the judgment standard will be automatically reduced, thereby avoiding the misjudgment of "false blurry" images caused by insufficient light as invalid images, and ensuring that valid field survey images can still be screened in low light environments.

[0072] Thus, traditional fixed thresholds are prone to misclassifying valid but dark images as blurry images and rejecting them when there is insufficient light. This solution, by dynamically adjusting the threshold, can intelligently identify the decrease in clarity caused by insufficient light and exempt it. This greatly expands the applicable scenarios of the system, allowing valuable field survey data to be retained even under complex lighting conditions such as dusk, cloudy days, or backlighting, avoiding data waste. Secondly, by introducing a light attenuation coefficient, the system no longer relies on a single absolute value for judgment, but calibrates the judgment standard in real time according to the ambient light. This mechanism enables the image quality screening module to automatically adapt to various environmental changes from strong light to weak light, maintaining high-precision screening results without manual intervention.

[0073] Feature matching is performed on valid on-site images, and a confidence score for each risk item is determined based on a pre-set engineering risk checklist, specifically including:

[0074] The effective images from the site are input into a pre-trained visual encoder to extract image feature vectors. The descriptive text of each risk item in the pre-set engineering risk checklist is input into a text encoder to generate text feature vectors corresponding to each risk item.

[0075] Calculate the cosine similarity between the image feature vector and the text feature vector corresponding to each risk item to obtain the preliminary matching score for each risk item;

[0076] By incorporating weighting coefficients and contextual factors from the industry knowledge base, and combining them with the preliminary matching score, the final confidence score for each risk item is calculated.

[0077] Specifically, the qualified images from the site that have passed quality screening are input into a pre-trained visual encoder (such as ViT or ResNet) to extract high-dimensional image feature vectors. These vectors contain key information about texture, edges, and object structure in the images. At the same time, a pre-set engineering risk checklist is read, and the standardized description text of each risk item (such as "loose foundation bolts" or "rust on the tower surface") is input into a text encoder to generate corresponding text feature vectors.

[0078] The step of calculating the cosine similarity between the image feature vector and the text feature vector of each risk item aims to measure the semantic similarity between the content of the on-site image and the description of a specific risk. For example, if obvious rust texture is detected in the image, its feature vector will have a high cosine similarity with the vector of the text description "rust on the surface of the tower", thus obtaining a high preliminary matching score.

[0079] To enhance the professionalism of the assessment, an industry-specific knowledge base was introduced. Based on the severity level of risk items (e.g., general risk, major risk), corresponding weight coefficients were retrieved from the knowledge base. Simultaneously, contextual factors were incorporated (e.g., if the visual features of "bolt" and "loose" appear simultaneously in the image, the contextual factors would enhance the score for the "loose foundation bolt" item) to weight and correct the initial matching score. Finally, a final confidence score was calculated for each risk item, which not only reflects the matching degree between the image and the text but also incorporates the experience weights of industry experts.

[0080] Thus, by calculating the cosine similarity between the image and the risk description text, ambiguous visual information is transformed into a calculable numerical score. Combined with industry weighting coefficients, the risk assessment is based on data and standard calculation results. This significantly reduces missed or false detections caused by differences in human experience, ensuring the consistency and objectivity of risk assessment. Secondly, by introducing contextual correlation factors, the logical relationships between multiple visual elements in the image can be understood. For example, simply detecting "tower" may be just a normal scene, but if "tower" and "crack" are detected simultaneously, the system will significantly increase the confidence of the high-risk item "structural crack" through contextual correlation. This mechanism effectively solves the defect of single feature matching being easily interfered with by background, and improves the ability to accurately capture risks in complex engineering sites. It should be noted that the final confidence score calculation formula is ScoreI=Si×Wi×Rcontext; where Si is the initial matching score, Wi is the weighting coefficient in the industry knowledge base, and Rcontext is the contextual correlation factor.

[0081] The risk items to be included in the project report are obtained by comparing the confidence score with the preset threshold.

[0082] Specifically, the final confidence score of each risk item is compared with a preset inclusion threshold, which is usually set according to the engineering safety level. For example, for major risk items, the threshold may be set to 0.85 or higher to ensure that only high-confidence risks are included in the report. If the confidence score of a risk item is higher than or equal to the preset threshold, the risk item is determined to be a valid risk and is marked as a risk item to be included in the engineering report.

[0083] Iterate through all risk items, and summarize the IDs of all risk items that meet the threshold conditions, along with their corresponding confidence scores, risk descriptions, and other information, to generate the final list of risk items to be included in the engineering report. For example, if the confidence score of tower corrosion detected in the field image is 0.92, which is higher than the threshold of 0.85, then the tower corrosion will be officially included in the report. If a certain item has a score of only 0.60, it will be judged as a low-confidence risk and will not be included in this report to avoid interfering with the engineer's judgment.

[0084] Based on the engineering report template, the system automatically generates structured report content by incorporating the risk items included in the engineering report, and links it to the original site survey images. Specifically, this includes:

[0085] Based on the JSON or XML schema structure of the engineering report template, obtain the mapping relationship between report chapters and risk item IDs;

[0086] Based on the mapping relationship, the risk items included in the engineering report are populated into the corresponding nodes of the engineering report template, and natural language descriptions are generated based on slot filling technology combined with the risk items;

[0087] Extract the EXIF ​​information of the original site survey images as metadata and embed it into the report attachment index. Calculate the SHA-256 hash value of the original site survey images and write it into the report's digital signature area.

[0088] Specifically, the system reads a pre-defined engineering report template (such as JSON or XML Schema format), parses the mapping relationship between the report chapter structure and risk item IDs, for example, mapping risk item ID R001 (tower corrosion) to the "Main Structure Inspection" chapter in the report; then, based on this mapping relationship, the system fills the selected risk item data (including risk ID, name, level, confidence score, etc.) into the corresponding node positions in the template. At the same time, it calls a pre-defined expert language library and uses slot filling technology to fill the specific values ​​of the risk items (such as corrosion area, severity) into the natural language description template, generating structured text such as "Upon inspection, the tower surface has moderate corrosion, with a corrosion area of ​​approximately 2 square meters".

[0089] Extract the EXIF ​​information (including shooting time, GPS coordinates, camera parameters, etc.) from the original on-site survey images and embed it as metadata in the report's attachment index. At the same time, calculate the SHA-256 hash value of the original image file and write the hash value into the digital signature area of ​​the report. This step ensures that the images cited in the report correspond uniquely to the original on-site data, prevents the images from being tampered with, and achieves a strong binding between the report content and the on-site evidence.

[0090] In this way, by parsing the template structure and automatically filling in risk data, engineers are freed from tedious text entry and formatting adjustments, and the generated report format is standardized, completely eliminating human typesetting errors. Secondly, by embedding EXIF ​​information and calculating SHA-256 hash values, a strong correlation is established between the report content and the original on-site images. This not only facilitates later review and traceability (the original high-definition images can be viewed directly by clicking on the image index in the report), but also ensures data integrity through digital signature technology, preventing the report content from being maliciously tampered with.

[0091] In some embodiments, the generation of natural language descriptions based on slot filling technology combined with risk items specifically includes:

[0092] The system calls upon a pre-set expert experience script library and, based on slot filling technology, matches the corresponding natural language description template from the script library according to the risk name, risk level, confidence score, and location information of each risk item.

[0093] Replace the slot variables in the natural language description template with the risk name, risk level, confidence score and location information of each risk item to generate natural language description text.

[0094] Specifically, based on the attributes of the current risk item (e.g., "tower surface corrosion", "level: moderate", "confidence level: 0.92", "location: tower in area A"), the most matching natural language description template is retrieved from the preset expert experience script database. For example, for the "corrosion" risk, a general template such as "Upon on-site verification, {risk name} was found at {location}, the severity was assessed as {risk level}, and the confidence level was determined to be {confidence score}" will be matched.

[0095] The extracted specific risk item data is used as variables to replace the corresponding slots in the template (such as {location}, {risk name}, etc.); finally, specific natural language description text is generated, such as: "Upon on-site inspection, corrosion was found on the surface of the tower in Area A. The severity is assessed as moderate, and the confidence level is 0.92". This process is fully automated and requires no manual intervention.

[0096] In this way, by using slot filling technology, structured risk data (such as confidence scores and location information) can be accurately converted into fluent natural language. This not only eliminates the conversion errors between data and text, but also enables non-professionals to clearly understand the risk conclusions in the report, significantly improving the readability of the report and the efficiency of information transmission.

[0097] Perform integrity checks on the generated structured report, specifically including:

[0098] Iterate through all risk item IDs in the preset engineering risk check table, and check one by one whether there is a conclusion field in the report data structure corresponding to each risk item ID;

[0099] If the conclusion field corresponding to all risk item IDs is not empty, the report integrity check is considered to have passed.

[0100] If the conclusion field corresponding to any risk item ID is empty, the report integrity check is deemed to have failed, triggering an anomaly alert.

[0101] Specifically, the system reads the preset engineering risk checklist and obtains a list of all risk item IDs (e.g., R001 to R050). Then, it iterates through the ID list and checks whether there are conclusion fields corresponding to each risk item ID in the generated structured report data. For example, it checks whether the report contains a conclusion description for R001 (tower corrosion) or R002 (loose foundation bolts).

[0102] Determine if the conclusion field corresponding to all inspected risk item IDs is not empty; if all fields are filled with valid content, the report integrity check is passed and the report can be exported or published; if any conclusion field corresponding to any risk item ID (such as R015) is found to be empty (i.e. missing), the report integrity check is failed, an anomaly alarm is immediately triggered, and the engineer is prompted to supplement the missing data.

[0103] In this way, by fully traversing the checklist IDs, it is possible to ensure that every preset checkpoint is reflected in the report. This mechanism eliminates the risk of missed checks due to human negligence from a technical perspective, ensuring that the engineering report covers all safety hazards and significantly improving the completeness and rigor of the report.

[0104] An automatic wind survey report generation system, comprising:

[0105] The image quality filtering module is used to acquire on-site survey image data and perform image quality filtering on the on-site survey image data to obtain valid on-site images;

[0106] The feature matching and confidence calculation module is used to perform feature matching on valid on-site images and determine the confidence score for each risk item based on a preset engineering risk checklist.

[0107] The risk item screening module is used to compare the confidence score with a preset threshold to obtain the risk items to be included in the project report;

[0108] The report generation module is used to automatically generate structured report content based on the engineering report template, combined with the risk items included in the engineering report, and to associate it with the original site survey images.

[0109] A computing device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of any of the above-described methods for automatically generating a wind survey report.

[0110] A computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of any of the above-described methods for automatically generating a wind survey report.

[0111] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for automatically generating wind survey reports, characterized in that, Specifically, it includes: Acquire on-site survey image data and perform image quality screening on the on-site survey image data to obtain valid on-site images; Feature matching is performed on the valid images at the site, and the confidence score of each risk item is determined based on the preset engineering risk checklist; The risk items to be included in the project report are obtained by comparing the confidence score with the preset threshold. Based on the engineering report template, the system automatically generates structured report content by combining the risk items included in the engineering report, and links it to the original site survey images.

2. The method for automatically generating wind survey reports according to claim 1, characterized in that, The process of filtering the on-site survey image data to obtain valid on-site images specifically includes: The acquired field survey image data is denoised and its color space is converted to obtain a grayscale image of the field survey. The frequency domain energy of the grayscale image of the field survey was calculated using the Laplace variance method, and the frequency domain energy variance value was obtained. By comparing the frequency domain energy variance value with the dynamic sharpness threshold, clear on-site survey images are obtained. Based on clear on-site survey images, obtain the perspective matrix formed by the vanishing points of the images, and call the preset standard survey view vector; Calculate the cosine similarity between the perspective matrix formed by the vanishing points of the image and the standard viewpoint vector of the image, and select the on-site survey images with the correct shooting angle based on the cosine similarity. A lightweight CNN classification network is used to classify the scene of the on-site survey images with the correct shooting angle, so as to obtain the on-site survey images belonging to the scene related to wind survey.

3. The method for automatically generating wind survey reports according to claim 2, characterized in that, The setting of the dynamic sharpness threshold specifically includes: Extract the global illumination intensity value from the grayscale image of the on-site survey; A base sharpness threshold is set as the initial value, which is the Laplace variance determination threshold under standard lighting conditions; The light attenuation coefficient is calculated based on the ratio of the global light intensity value to the standard light intensity value. Multiply the base sharpness threshold by the illumination attenuation coefficient to obtain the dynamic sharpness threshold corresponding to the current scene.

4. The method for automatically generating wind survey reports according to claim 1, characterized in that, The process of performing feature matching on valid on-site images and determining the confidence score for each risk item based on a pre-set engineering risk checklist specifically includes: The effective images from the site are input into a pre-trained visual encoder to extract image feature vectors. The descriptive text of each risk item in the pre-set engineering risk checklist is input into a text encoder to generate text feature vectors corresponding to each risk item. Calculate the cosine similarity between the image feature vector and the text feature vector corresponding to each risk item to obtain the preliminary matching score for each risk item; By incorporating weighting coefficients and contextual factors from the industry knowledge base, and combining them with the preliminary matching score, the final confidence score for each risk item is calculated.

5. The method for automatically generating wind survey reports according to claim 1, characterized in that, The process involves automatically generating a structured report based on the engineering report template, incorporating risk items included in the engineering report, and linking it to original site survey images. Specifically, this includes: Based on the JSON or XML schema structure of the engineering report template, obtain the mapping relationship between report chapters and risk item IDs; Based on the mapping relationship, the risk items included in the engineering report are populated into the corresponding nodes of the engineering report template, and natural language descriptions are generated based on slot filling technology combined with the risk items; Extract the EXIF ​​information of the original site survey images as metadata and embed it into the report attachment index. Calculate the SHA-256 hash value of the original site survey images and write it into the report's digital signature area.

6. The method for automatically generating wind survey reports according to claim 5, characterized in that, The generation of natural language descriptions based on slot filling technology combined with risk items specifically includes: The system calls upon a pre-set expert experience script library and, based on slot filling technology, matches the corresponding natural language description template from the script library according to the risk name, risk level, confidence score, and location information of each risk item. Replace the slot variables in the natural language description template with the risk name, risk level, confidence score and location information of each risk item to generate natural language description text.

7. The method for automatically generating wind survey reports according to claim 6, characterized in that, The method also includes performing integrity verification on the generated structured report, specifically including: Iterate through all risk item IDs in the preset engineering risk check table, and check one by one whether there is a conclusion field in the report data structure corresponding to each risk item ID; If the conclusion field corresponding to all risk item IDs is not empty, the report integrity check is considered to have passed. If the conclusion field corresponding to any risk item ID is empty, the report integrity check is deemed to have failed, triggering an anomaly alert.

8. An automatic wind survey report generation system, characterized in that, include: The image quality filtering module is used to acquire on-site survey image data and perform image quality filtering on the on-site survey image data to obtain valid on-site images; The feature matching and confidence calculation module is used to perform feature matching on valid on-site images and determine the confidence score for each risk item based on a preset engineering risk checklist. The risk item screening module is used to compare the confidence score with a preset threshold to obtain the risk items to be included in the project report; The report generation module is used to automatically generate structured report content based on the engineering report template, combined with the risk items included in the engineering report, and to associate it with the original site survey images.

9. A computing device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the automatic generation method for wind survey reports according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the steps of the automatic generation method for wind survey reports according to any one of claims 1-7.

Images