Engineering entity test detection auxiliary analysis method based on deep learning

Abstract

The invention belongs to the technical field of engineering detection, and particularly discloses an engineering entity test detection auxiliary analysis method based on deep learning, which comprises the following steps: inputting collected detection data into an intelligent model, screening and identifying the type of the detection data, and calling detection processing software to process the detection data according to the type to generate a to-be-checked document; manually judging whether the to-be-checked document is qualified or not; if the to-be-checked document is unqualified, generating a first calibration instruction, sending the first calibration instruction to the intelligent model to regenerate the to-be-checked document until the to-be-checked document is qualified, inputting the qualified to-be-checked document into the intelligent model, generating an evaluation report according to preset evaluation parameters, and manually judging whether the currently generated evaluation report is qualified or not; if the judgment result is unqualified, generating a second calibration instruction and sending the second calibration instruction to the intelligent model to regenerate the evaluation report until the judgment result is qualified; generating a detection data processing log file according to the qualified to-be-checked document, the evaluation report and the corresponding detection data; the method has the advantages that an intelligent closed loop is constructed, and the detection efficiency and accuracy are improved.

Description

Technical Field

[0001] This invention relates to the field of engineering testing technology, and more specifically, to a deep learning-based auxiliary analysis method for engineering entity testing. Background Technology

[0002] The current standard procedure for engineering entity testing involves on-site testing personnel using various testing equipment to collect large volumes of data on bridges, tunnels, and building structures within construction or operation windows. The raw data is then brought back to the project site for data analysis and report compilation, or transmitted via network to the company's back office for centralized processing by senior engineers. With the acceleration of infrastructure construction, the testing market size and workload have increased dramatically. While the number of testing organizations has grown rapidly, the growth rate of qualified and experienced high-level testing personnel has lagged far behind demand, resulting in uneven technical skill levels among frontline teams. Simultaneously, the coexistence of high-intensity fieldwork in short periods and the processing of massive amounts of multi-dimensional data has pushed the existing model to its limits in terms of data processing capacity, report timeliness, and the reliability of conclusions.

[0003] Against this backdrop, the core problems faced by existing technologies include: the imbalance between supply and demand of personnel leading to insufficient on-site analysis capabilities; frontline personnel being prone to judgment bias, omissions, and textual errors when processing big data under fatigue; the time delay introduced by large file network transmission and decentralized processing significantly weakens the timeliness of detection results; and the risk of misjudgment arising from different personnel's varying abilities to understand complex data.

[0004] To address these issues, a deep learning-based auxiliary analysis method for engineering entity testing and detection is proposed. Summary of the Invention

[0005] The present invention aims to provide a deep learning-based auxiliary analysis method for engineering entity testing, in order to solve or improve the problem that existing engineering testing modes are unable to balance the accuracy and timeliness of testing results under conditions of limited personnel, high-intensity operations and massive data.

[0006] In view of this, the first aspect of the present invention is to provide a deep learning-based auxiliary analysis method for engineering entity test detection.

[0007] A second aspect of the present invention is to provide a system.

[0008] A third aspect of the present invention is to provide an electronic device.

[0009] A fourth aspect of the present invention is to provide a computer-readable storage medium.

[0010] The first aspect of this invention provides a deep learning-based auxiliary analysis method for engineering entity testing, comprising the following steps: inputting collected testing data into an intelligent model; filtering and identifying the type of the testing data; calling corresponding testing processing software to process the testing data according to the type to generate a test document; manually judging whether the currently generated test document is qualified; if the judgment result is unqualified, generating a first calibration command and sending it to the intelligent model to regenerate the test document until the judgment result is qualified; inputting the qualified test document into the intelligent model and generating an evaluation report according to preset evaluation parameters; manually judging whether the currently generated evaluation report is qualified; if the judgment result is unqualified, generating a second calibration command and sending it to the intelligent model to regenerate the evaluation report until the judgment result is qualified; generating a testing data processing log file based on the qualified test document, the evaluation report, and the corresponding testing data.

[0011] In any of the above technical solutions, the step of screening and identifying the type of the detection data includes: identifying the type of the detection data through the deep learning neural network of the intelligent model, and removing invalid data, erroneous data, and data that are outside the processing range from the detection data according to the type.

[0012] In any of the above technical solutions, the step of calling the corresponding detection processing software according to the type to process the detection data and generate the document to be verified includes: calling the detection processing software written in Python language according to the type to complete the data standardization processing to generate an image or report; inputting the image or report into the convolutional neural network model of the intelligent model to perform data recognition and annotation, and using the annotated image or report to form the document to be verified.

[0013] In any of the above technical solutions, the step of generating a first calibration instruction and sending it to the intelligent model to regenerate the document to be verified includes: generating feedback data for identification defects, annotation errors, annotation omissions, and unclear images in the document to be verified, sending it to the intelligent model to perform feedback learning, and reprocessing the data to generate the document to be verified.

[0014] In any of the above technical solutions, the feedback learning step includes: the intelligent model uses a backpropagation mechanism to reduce the loss gradient, performs supervised training and fine-tuning, and optimizes the model through a manual feedback reinforcement mechanism to improve the generalization ability of the intelligent model in continuous operation.

[0015] In any of the above technical solutions, the preset evaluation parameters include: engineering entity testing standards, national or industry standard parameters, or project-specific evaluation parameters set by the testing organization itself.

[0016] In any of the above technical solutions, the detection data processing log file includes: The types of detection data, qualified documents to be verified and evaluation reports, invalid data, erroneous data, and data that are out of processing range, as well as the first calibration instruction and the second calibration instruction.

[0017] A second aspect of the present invention provides a system comprising: a data input module for receiving and uploading collected test data; an intelligent analysis module for identifying data types, filtering and preprocessing the test data, calling corresponding test processing software to generate the document to be verified, and reprocessing the document to be verified or the evaluation report according to the received first calibration instruction or second calibration instruction; a human interaction module for manually reviewing the document to be verified and the evaluation report generated by the intelligent analysis module, and generating and sending the first calibration instruction or the second calibration instruction to the intelligent analysis module when the review is unsatisfactory; and a log generation module for packaging the approved document to be verified, the evaluation report and the corresponding test data to generate a test data processing log file, and storing and managing its traceability.

[0018] A third aspect of the present invention provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the steps of the method described above.

[0019] A fourth aspect of the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described method.

[0020] The beneficial effects of this invention compared to the prior art are as follows: Through a closed-loop mechanism of intelligent model, detection processing software, manual review, and calibration instructions, the system can perform multi-angle recognition and standardized processing of multi-source detection data such as images, waveforms, and numerical reports. It uniformly applies national standards, industry specifications, or project-defined parameters to complete defect judgment and report compilation. Driven by each round of calibration instructions, the intelligent model instantly backpropagates, learns, or undergoes manual parameter tuning, continuously optimizing recognition boundaries and judgment thresholds. This significantly reduces missed judgments, misjudgments, and textual errors caused by fatigued work. Simultaneously, the processing of large volumes of raw data is moved to the field or near-end nodes; only the documents to be reviewed and the evaluation report need to be uploaded to complete the review. This significantly reduces the data transmission and report generation cycle compared to the traditional remote centralized processing mode, alleviating the queuing and delay risks caused by a shortage of highly skilled inspection personnel.

[0021] Equipped with comprehensive data, image, logical deduction, text output, and deep learning capabilities, the system continuously receives feedback data from various engineering projects, accumulates challenging examples, and automatically iterates its generalization ability. Testing organizations can upgrade the system simply by entering new specifications or thresholds through a graphical interface, requiring no programming experience. The testing data processing log file records the entire process of data processing, manual review, calibration instructions, and version information, ensuring compliant traceability of testing results and high-quality accumulation of model training data. As a result, the system not only significantly reduces errors and biases caused by manual processing of massive amounts of data but also lowers the technical barrier for non-programming professionals in the subsequent maintenance and functional expansion of the intelligent model, driving the engineering testing industry towards standardization, intelligence, and high efficiency.

[0022] Additional aspects and advantages of embodiments of the invention will become apparent in the following description or may be learned by practice of embodiments of the invention. Attached Figure Description

[0023] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which: Figure 1 This is a flowchart of the method steps of the present invention; Figure 2 This is a flowchart illustrating a specific implementation method of the present invention; Figure 3 This is a schematic diagram of the specific support device of the present invention; Figure 4 This is a system logic block diagram of the present invention; Figure 5 This is a schematic diagram of the structure of an electronic device according to the present invention. Detailed Implementation

[0024] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0025] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0026] Please see Figures 1-5 The following describes a deep learning-based auxiliary analysis method for engineering entity testing and detection, based on some embodiments of the present invention.

[0027] An embodiment of the first aspect of the present invention proposes a deep learning-based auxiliary analysis method for engineering entity test detection. In some embodiments of the present invention, such as... Figure 1 As shown, the method includes the following steps: S101 inputs the collected detection data into the intelligent model, filters and identifies the type of detection data, and calls the corresponding detection processing software to process the detection data and generate a document to be verified based on the type.

[0028] Here, raw test data acquired manually or by unmanned testing equipment during the physical testing of construction projects will be input into the intelligent model. This test data may include, but is not limited to, images, waveforms, curves, digital signals, or combined data from concrete rebound hammers, ultrasonic testing instruments, structural strain sensors, thermal imaging equipment, or visual camera units. This data is characterized by diverse sources, complex formats, large quantities, and significant noise interference; therefore, it must first undergo preliminary screening and classification by the intelligent model.

[0029] The intelligent model, through its internally built deep learning neural network architecture, automatically performs structured analysis on the input detection data by utilizing the distribution characteristics of a large amount of engineering inspection sample data learned during the training phase. Specifically, the model determines the detection type of the data based on feature parameters such as dimensionality, signal pattern, image texture, and frequency domain response; specifically, whether it is rebar scan data, concrete strength data, crack image data, or structural strain time history data, etc. Through this identification process, not only can the type be classified, but some obviously invalid data can also be eliminated during the classification process, such as null data collected during equipment initialization, distorted data caused by severe environmental noise interference, or duplicate and redundant data caused by improper operation.

[0030] After completing type recognition, the intelligent model calls the matching detection and processing software from a pre-set software library based on the specific data type identified. Specifically, the detection and processing software can be an image enhancement and edge extraction program for structural image recognition, a filtering and demodulation algorithm tool based on signal analysis, or an engineering evaluation process engine encapsulated in Python scripts. The calling process may include operations such as software path identification, automatic parameter configuration, and data format adaptation and conversion.

[0031] The processing software performs formatting and index extraction on the filtered data. Specifically, it converts ultrasonic test data into attenuation waveforms and extracts effective amplitude values, binarizes crack images and identifies crack direction and length information, and generates corresponding parameter units in a standardized report template from the rebound value array. Through these processing steps, the raw data can be effectively transformed into intermediate structured documents with engineering significance.

[0032] As described above, the key responsibilities of intelligent screening, automatic classification, and preliminary standardization of data collected from the engineering site are to ensure that subsequent model identification, standard judgment, and report generation operations are based on reliable data, laying a solid foundation for the automated analysis process of the entire engineering entity testing and inspection. This step, through the collaborative efforts of deep learning models and specialized software, significantly improves the efficiency and accuracy of the raw data preprocessing stage, reducing the risk of human judgment bias and repetitive work.

[0033] Specifically, the steps for screening and identifying the types of detection data include: The intelligent model uses a deep learning neural network to identify the type of detection data, and then removes invalid, erroneous, and out-of-process data based on the type.

[0034] As described above, the raw detection data uploaded by on-site inspectors or unmanned inspection equipment is input into the intelligent model. The deep learning neural network integrated within the intelligent model performs parallel analysis of the input data's dimensional features, texture patterns, waveform morphology, and sampling distribution through convolutional feature extraction, temporal feature fusion, and frequency domain spectral analysis. Subsequently, based on the feature dictionary constructed during the training phase, the detection data is quickly categorized into preset types such as crack image data, concrete strength rebound data, ultrasonic attenuation data, and rebar positioning GPR data. Specifically, for high-resolution crack photographs, the neural network captures pixel texture differences in low-level convolutional layers, focuses on linear crack morphology in high-level feature maps, and outputs crack image data labels. For ultrasonic amplitude sequences uploaded in CSV format, the network uses one-dimensional convolution combined with Fast Fourier Transform to extract waveform spectral features and labels them as ultrasonic attenuation data. For rebound value matrices, the network confirms they are concrete strength rebound data through statistical distribution matching. After completing type identification, the intelligent model immediately calls the detection processing software corresponding to each type, laying a precise matching foundation for subsequent standardized processing.

[0035] While completing the type classification, the intelligent model uses the confidence score and anomaly detection submodule output by the deep learning neural network to screen the detection data for quality and automatically remove three types of data that do not meet the processing conditions: (1) invalid data - such as crack images with no texture and overexposure, ultrasonic signals with zero, and current noise collected under no-load conditions; (2) erroneous data - such as repeated rebound values ​​caused by device buffer overflow, missing GPS coordinates in GPR trajectory files, and disordered ultrasonic recording timestamps; (3) data that exceeds the processing range - such as crack image resolution below the system threshold, ultrasonic sampling rate less than 10kHz, and rebound values ​​exceeding the upper or lower limit of the device range. The removal operation is performed in memory in real time, and the removal results and reasons are recorded in the auxiliary log to provide abnormal samples for subsequent model training. The filtered valid detection data is labeled with type, quality and unique identifiers and transmitted through the data pipeline to the corresponding detection processing software (such as image enhancement and edge extraction module, ultrasonic attenuation curve analysis script, rebound value conversion tool, etc.) for the next step of standardization processing.

[0036] As described above, through the efficient multimodal feature recognition and anomaly removal mechanism, this step establishes a strict quality gate at the data flow entry point. This ensures that subsequent analysis stages only process high-confidence data, and continuously records invalid and erroneous data samples to feed back into the deep learning neural network, achieving online adaptive performance iteration. This not only significantly improves the efficiency of raw data quality control and reduces the workload of manual screening, but also provides a reliable and standardized consistent data foundation for subsequent document generation, evaluation report output, and archiving of detection data processing log files.

[0037] Specifically, the steps of calling the corresponding detection processing software to process the detection data and generate the document to be verified according to the type include: Based on the data type, a detection and processing software written in Python is invoked to perform data standardization processing in order to generate images or reports.

[0038] Images or reports are input into the convolutional neural network model of the intelligent model to perform data recognition and annotation, and the annotated images or reports are used to form a document to be verified.

[0039] Regarding the specific description above, after completing the identification and quality screening of the detection data types, the intelligent model, based on the assigned type labels, calls the corresponding detection processing software, encapsulated in Python, from the system's pre-built software repository. These processing software programs have built-in standardized processes for different engineering detection scenarios: for example, the image enhancement script for crack image data automatically performs brightness equalization, edge sharpening, and noise suppression, then generates a grayscale image conforming to the "Technical Guidelines for Images of Defects in Concrete Structures"; the signal analysis script for ultrasonic attenuation data first performs bandpass filtering and zero-phase shaping on the original waveform, then extracts the arrival time of the first wave and the attenuation curve of the main peak, finally outputting an attenuation spectrum containing the coordinates of key feature points; and for concrete strength rebound data, the processing script automatically performs temperature correction, angle correction, and carbonation depth correction according to the "Rebound Hammer Testing Procedure," generating a comparison strength conversion table. After all scripts are executed, they will uniformly output images or reports, with image files saved in PNG or TIFF format and report files saved in both CSV and PDF formats for flexible use in subsequent processes. Through this standardized process, the raw data is quickly transformed into intermediate results that are meaningful for engineering interpretation, have standardized formats, and are easy to visualize.

[0040] Subsequently, the intelligent model directly inputs the aforementioned images or reports into its internal convolutional neural network model for further identification and annotation. The convolutional neural network model first extracts multi-scale features from the input content and establishes semantic segmentation or object detection channels: taking a crack grayscale image as an example, the network automatically locates the crack trunk, outputs crack length, width, and orientation information, and overlays red polysegment lines on the image; taking an ultrasonic attenuation spectrum as an example, the network identifies the first wave arrival point, the peak value of the reflected wave, and the energy attenuation inflection point, annotates the coordinates of key peaks on the image, and outputs the corresponding values; for rebound strength reports, the network automatically reads the rebound value and correction parameters based on fixed column headers, calculates and annotates the estimated concrete strength range. After identification and annotation are completed, the convolutional neural network model also generates a confidence score for each defect or data anomaly and records the corresponding label in the file metadata, ensuring that the credibility of the annotation can be intuitively viewed during subsequent manual review.

[0041] As described above, the images or reports identified and labeled by the convolutional neural network model will be uniformly packaged into a document awaiting verification. The packaging process includes metadata such as document number, detection batch information, device number, detection location coordinates, and recognition timestamp. Simultaneously, the system generates a JSON index file to record the coordinates, confidence level, and associated original data path of each labeled object, achieving a complete mapping of the data link. This step achieves two key benefits: firstly, it ensures seamless integration between the processing software and intelligent recognition, guaranteeing that intermediate results meet engineering testing standards and visual interpretation requirements; secondly, it provides a well-structured, information-rich document for subsequent manual review, enabling rapid problem identification, thereby significantly improving review efficiency, reducing the risk of omissions, and laying a solid data foundation for the accurate generation of the final evaluation report.

[0042] S102, the current generated document to be verified is judged manually to determine whether it is qualified; if the judgment result is unqualified, the first calibration instruction is generated and sent to the intelligent model to regenerate the document to be verified, until the judgment result is qualified.

[0043] Here, the documents to be reviewed are submitted to qualified testing engineers for manual judgment. The documents to be reviewed are generated by the intelligent model in combination with the data type of the test and the corresponding testing processing software; specifically, they may include information such as image annotation results, parameter calculation reports, and structural recognition results. The manual judgment mainly focuses on the following aspects: (1) Whether the test images in the documents to be reviewed are clear and complete, and whether there are problems such as image jitter, occlusion, or insufficient exposure; (2) Whether the data annotation is accurate, whether the defect area in the image is completely selected, and whether the crack direction and length are correctly identified; (3) Whether the report information meets the standard format requirements of the current test type, whether the data units are consistent, and whether the fields are complete; (4) Whether there is a problem of mismatch between the test data and the analysis conclusion, such as the test data being qualified but marked as defective.

[0044] If the inspection engineer determines through manual judgment that the document to be inspected contains non-compliance with standards or technical errors, the manual judgment opinion will be transformed into a first calibration instruction in a structured form. Specifically, the first calibration instruction may include corrections to the location of identified defects in the image, suggestions for correcting erroneous annotations, suggestions for adjusting non-standard data formats, image clarity requirements, and explanations for missing structural information. This calibration instruction is fed back to the intelligent model through an interactive system.

[0045] Upon receiving the first calibration instruction, the intelligent model initiates a feedback-driven document regeneration process. On one hand, it corrects parameters or optimizes paths in the previous processing flow based on the specific modifications required in the calibration instruction; on the other hand, it utilizes accumulated misjudged samples for rapid training and correction. Specifically, it optimizes the loss function through backpropagation or performs a few iterations to adjust the weight distribution, improving the model's adaptability to this type of problem. The newly generated documents awaiting verification are then submitted for manual review. If they still fail, the feedback and regeneration process is repeated.

[0046] This cyclical process will continue until the document to be reviewed is deemed acceptable by manual review. The criteria for acceptance are based on national standards, industry specifications, or technical agreements set by the client in the region or field where the project is located, and can be invoked from built-in templates in the system or by manually inputting parameters.

[0047] As described above, a data quality control mechanism based on human-machine collaboration has been implemented. Through an embedded calibration mechanism, the model achieves adjustability and self-correction capabilities while maintaining efficiency, and also lays a highly reliable data foundation for subsequent evaluation report generation. This step not only improves the accuracy of detection data analysis and document compliance, but also supports continuous learning and optimization of the model in actual operations, significantly enhancing the professional level and engineering adaptability of the engineering entity testing process.

[0048] Specifically, the steps of generating the first calibration command and sending it to the intelligent model to generate the calibration document again include: Feedback data is generated to address identification defects, annotation errors, annotation omissions, and unclear images in the documents to be verified. This feedback data is then sent to the intelligent model for feedback learning, and the model is further processed based on the feedback data to generate the documents to be verified.

[0049] Regarding the specific description above, when inspection engineers discover issues such as identification defects, annotation errors, omissions, or unclear images in the document to be verified during manual review, the system will refine these issues into structured feedback data and aggregate them into the first calibration instruction. Specifically, the system first locates the problematic elements in the document to be verified using coordinates or row / column indexes based on the manually marked positions or items; then, it records the problem type, location coordinates, text description, and suggested correction methods in a JSON array. For example, if the main crack in a crack image is not fully identified, the system will record the start and end coordinates of the gap and the missing length value; if the first wave arrival point in an ultrasonic attenuation spectrum is offset, the system will record the offset distance and correct coordinates; if a character is misidentified in a rebound report, the system will record the cell coordinates and expected value. After all issues are recorded, the system automatically encapsulates multiple sub-instructions such as defect annotation correction, image clarity enhancement, and annotation confidence below a threshold, along with timestamps, reviewer IDs, and the version number of the document to be verified, ultimately forming a machine-readable first calibration instruction and pushing it to the intelligent model.

[0050] Upon receiving the first calibration command, the intelligent model immediately enters a feedback-based reprocessing flow. On one hand, the intelligent model uses the coordinate information marked in the command to perform local backpropagation on the feature map of the convolutional neural network model; by dynamically adjusting the loss weights and learning rate, the network quickly converges in the defect-marked area, improving recognition accuracy. On the other hand, for sub-commands indicating unclear images, the intelligent model automatically calls image enhancement scripts such as adaptive histogram equalization or deep denoising networks to process the original image and replace the old image in the document to be verified. For text errors in reports, the system triggers the OCR error correction module to re-identify and automatically replace the errors after comparing with the standard dictionary; for missing defects, it enhances low-confidence candidate regions in the neural network's candidate boxes and re-executes non-maximum suppression. After completing all corrections, the intelligent model regenerates a new version of the document to be verified, automatically updating the document version number, iteration count, and correction summary, and pushes it back to the inspection engineer for review. This closed-loop feedback-learning-regeneration mechanism is employed.

[0051] As can be seen from the above, this not only ensures that the quality of the documents to be verified continues to approach the optimal level, but also accumulates difficult samples in real-world scenarios for the intelligent model, achieving online adaptive performance improvement, thereby significantly reducing the probability of similar errors occurring in the future, and continuously improving the accuracy and stability of detection data processing and annotation.

[0052] Specifically, the steps of feedback learning include: The intelligent model uses backpropagation to reduce the loss gradient, performs supervised training and fine-tuning, and optimizes the model through human feedback reinforcement to improve the generalization ability of the intelligent model in continuous operation.

[0053] Regarding the specific description above, the structured data fed back by the inspection engineer through the first or second calibration command is stored in the online training cache according to the triplet format of problem type-location information-expected output, and associated with the corresponding input samples to form a batch of fine-tuning data available in real time. When a new batch reaches a set threshold, such as 16 crack images or 64 ultrasonic waveform records, the intelligent model starts the incremental training process, performing a backpropagation mechanism on the convolutional neural network model or one-dimensional signal network model: first, the cross-entropy or smoothed L1 loss between the actual annotation and the target annotation is calculated at the output layer, and then the error signal is backpropagated layer by layer; to avoid overfitting, the system adopts a strategy of freezing the low-level convolutional kernels and only fine-tuning the high-level weights, while enabling adaptive learning rate decay, so that the loss gradient gradually decreases and converges stably near the local optimum. Taking the crack recognition scenario as an example, if the end of the crack is manually pointed out as a missed detection, the intelligent model will assign higher gradient weights to the high-level semantic feature channels, such as the fine crack detection filter, during the error backpropagation process, until the convolutional kernel can sensitively respond to the edge of the fine crack, thereby completing the supervised training fine-tuning.

[0054] As described above, after each closed-loop review is completed, the system dynamically calculates positive rewards or negative penalties based on the number of iterations required for new documents awaiting review, and writes these signals into the experience replay pool: a pass on the first attempt is recorded as +1, a pass on the second attempt as 0, and more than two passes are recorded as -n. Subsequently, the intelligent model uses this score as a delayed reward to update the weight distribution or adjust the confidence threshold through policy gradients; for example, if a crack identification fails to pass the review after three consecutive iterations, the system automatically lowers the crack confidence threshold and increases the attention weight of the edge detail channel. At the same time, the system adds high-value difficult examples marked by manual feedback to the priority sampling queue, increasing their display frequency in subsequent training batches, prompting the intelligent model to quickly learn from similar difficult examples and reduce future misjudgments. Driven by both the backpropagation mechanism to reduce the loss gradient and the manual feedback reinforcement mechanism to optimize the model, the intelligent model can continuously accumulate diverse data features from the engineering site during continuous operation, effectively improving its adaptability and generalization ability to input data under different detection equipment, different material structures, and different shooting lighting conditions, ultimately achieving a dynamic evolution effect of becoming more accurate with use.

[0055] S103: Input the qualified document to be verified into the intelligent model and generate an evaluation report based on the preset evaluation parameters. Manually judge whether the currently generated evaluation report is qualified. If the judgment result is unqualified, generate a second calibration command and send it to the intelligent model to generate an evaluation report again until the judgment result is qualified.

[0056] Specifically, the preset evaluation parameters may include: structural safety limits, defect judgment thresholds, crack width classification, concrete strength grade requirements, upper limits of ultrasonic signal attenuation, and apparent defect control standards specified in national standards, industry specifications, and technical agreements designated for engineering projects. These preset evaluation parameters can be entered by professionals during the model deployment phase or dynamically loaded by manual selection of applicable standards during the testing process. During the analysis, the intelligent model compares the data extracted from the document to be verified with the preset evaluation parameters item by item. Based on mechanisms such as rule matching, upper and lower limit judgment, and grade assessment, it performs judgment and outputs a structured evaluation result, which is then filled into a standardized report template. The model automatically completes tasks such as formatting, numbering, and item listing to generate a complete preliminary evaluation report.

[0057] After the preliminary report is generated, the evaluation report will be submitted to a qualified testing engineer for manual review. The core of the manual review is to check the judgment conclusions, analysis paths, item content, logical relationships, report format and current project requirements in the evaluation report item by item. Specifically, this includes but is not limited to: (1) whether the evaluation conclusions are consistent with the information in the aforementioned documents to be reviewed; (2) whether the specifications called in the report are applicable to this testing scenario, and whether there are any issues with calling incorrect parameter templates; (3) whether the defect level classification is consistent with the actual image annotations; and (4) whether the report text, charts, numbering, wording and other aspects meet the standard expression requirements of engineering documents.

[0058] If the testing engineer determines that the evaluation report contains non-compliance issues, they will organize the review comments into a structured second calibration instruction. Specifically, the second calibration instruction may include: a prompt indicating incorrect specification citations, an explanation of logical deviations in the judgment, suggestions for correcting conclusion items, and annotations of formatting errors. This second calibration instruction will be fed back to the intelligent model through the human interaction module.

[0059] Upon receiving the second calibration instruction, the intelligent model automatically calibrates its aforementioned evaluation logic, report template calling logic, and parameter judgment path. On one hand, the intelligent model adjusts the parameter weights, processing order, and data mapping relationships in each step according to the calibration content. On the other hand, it can utilize the model's deep learning training mechanism to perform error backpropagation and weight fine-tuning for parts with misjudgments or rule deviations, thereby iteratively improving the model's processing capabilities in subsequent detection tasks. After correction according to the second calibration instruction, the intelligent model regenerates a new evaluation report and resubmits it for manual review. If it still fails, the feedback and correction process is repeated until the evaluation report is manually judged as qualified.

[0060] As mentioned above, after completing the intelligent analysis, a manual review and closed-loop correction mechanism is introduced to ensure that the evaluation report not only meets the testing standards and specifications, but also possesses the format compliance and on-site applicability of engineering documents. This step combines the automated processing of the intelligent model with the experience judgment of engineering technicians, enhancing accuracy and practicality while ensuring efficiency.

[0061] Specifically, the preset evaluation parameters include: Engineering entity testing and inspection standards, national or industry standard parameters, or project-specific evaluation parameters set by the testing organization itself.

[0062] Regarding the specific description above, the preset evaluation parameters play a crucial role in establishing a correlation between test results and engineering specifications, industry standards, and project requirements. The system is configured with a multi-level parameter library in the background, divided into three layers: national standard parameter sets, industry standard parameter sets, and project-defined parameter sets. The national standard parameter sets record mandatory clauses such as the "Code for Acceptance of Construction Quality of Concrete Structures" and the "Code for Maintenance of Highway Bridges and Culverts." The industry standard parameter sets store commonly used recommended specifications such as the "Technical Specification for Testing and Evaluation of Urban Bridges" and the "Technical Specification for Testing Defects in Railway Tunnels." The project-defined parameter sets are entered by the testing agency at the start of the project according to the client's requirements; for example, a highway maintenance project might strictly adjust the maximum crack width limit from 0.3mm to 0.2mm. When generating an evaluation report, the system first dynamically searches for project-defined parameters based on the project number field in the document to be verified. If no corresponding entry is found, it falls back to the industry standard parameters. If still no match is found, the national standard parameters are used as the final basis, forming a top-down priority application chain to ensure that a reference threshold is available for any testing scenario.

[0063] Once a qualified document is input into the intelligent model, the logic reasoning component automatically loads the corresponding preset evaluation parameter entries while parsing the detection type label. For example, if the detection type is crack image data and the project is an inspection of an urban elevated bridge structure, the system will call the grading indicators for the width, length, and density of cracks in existing highway bridge decks from the "Technical Specification for Inspection and Evaluation of Urban Bridges." If the detection type is ultrasonic attenuation data and the project's custom parameter set has specific requirements for the lower limit of concrete wave velocity, the system will replace the default threshold of 3.2 km / s in the industry standard with this custom threshold, such as 3.5 km / s. Inside the evaluation engine, defect features identified by the convolutional neural network model, such as a crack width of 0.28 mm and a direction angle of 32°, are mapped to the corresponding thresholds. The logic reasoning component gives a judgment of minor defect, moderate defect, or non-compliance based on the parameter entries, and simultaneously packages the relevant clause index, including the standard number, clause number, and publication date, into the evaluation report, making it easy for reviewers to quickly trace the source of the basis.

[0064] As described above, version control and encrypted signature management are implemented for the preset evaluation parameters. When national standards are revised or industry specifications are updated, the system administrator uploads the new version parameter XML file through the backend interface. The system automatically generates a hash value and records the update log. The old version is disabled by default in new projects but is retained for traceability. If special evaluation indicators are added during project execution (such as temporary limits after bridge reinforcement), the testing agency can add entries to the project's custom parameter set and mark the transitional parameters. When generating the evaluation report, the intelligent model will automatically output a judgment prompt based on the transitional parameters, ensuring that the report conclusions have sufficient documented support. Through the above-mentioned multi-level parameter library, dynamic priority call and version management mechanism, the preset evaluation parameters not only ensure the compliance and timeliness of the evaluation conclusions, but also provide flexible and refined judgment criteria for different engineering scenarios, further improving the adaptability and credibility of the method of this invention in cross-project applications.

[0065] S104. Generate a test data processing log file based on the qualified documents to be tested, the evaluation report, and the corresponding test data.

[0066] Here, after the document to be verified is confirmed as qualified by manual review, and the evaluation report is confirmed as qualified by manual review, the system will collect and process the two final confirmed outputs together with the corresponding detection data from their sources. Qualified documents to be verified include standardized documents such as the detection image annotation results, parameter calculation reports, and structural information extraction content output by the intelligent model and confirmed by manual review; the evaluation report includes a structured conclusion document generated based on preset evaluation parameters and approved by manual review, covering the status assessment of the detected object, defect level assessment, and qualification determination results; the corresponding detection data is the dataset collected during the original detection process, including raw input data in the form of images, signals, text, or tables.

[0067] The three types of data mentioned above are associated using their unique identifiers and then integrated into the log building module. In this module, the system will generate a standard-format detection data processing log file. The log file contains, but is not limited to, the following information: The type of detection data, the time of collection, the serial number of the collection device, and the serial number of the detection point; The generated version of the document to be verified, an overview of the document content, coordinate information of the image annotation area, and the processing software call records; The evaluation report includes the judgment conclusions, the types of preset evaluation parameters used, the source of the standards, and the model judgment process path. The complete content of the first and second calibration instructions, their generation time, the operator's identification, and the number of model responses; The intelligent model processes each round of data using the following parameters, identification paths, and iterative adjustment logs, including information such as changes in the loss function and learning rate adjustments. The data processing steps involving manual review, the text of review comments, the information of the final confirmer, and the review timestamp, etc. The system includes status indicators such as the overall task number, project ownership information, report output number, electronic signature status of the report, and whether printing has been authorized.

[0068] The test data processing log files can be constructed using structured JSON or XML format and simultaneously stored as PDF and traceable electronic archive documents, facilitating regulatory auditing, test archiving, and report management. The system can upload log files to the project database, regulatory system interface, or test institution document management platform according to project requirements, and supports generating data summary hash values ​​to achieve data authenticity verification and tamper prevention.

[0069] As mentioned above, this log file serves not only as a recording unit for the entire testing process but also as a crucial link in ensuring data traceability, transparency in the processing, and the authority of the testing results. Through the systematic construction of this log file, every key node in the execution of the engineering entity testing and auxiliary analysis method can be clearly reconstructed, strengthening quality management capabilities, reducing the risk of ambiguous responsibility definitions, and providing a solid, high-quality data feedback foundation for subsequent optimization of artificial intelligence models, thus forming a closed-loop intelligent auxiliary testing process and data governance system.

[0070] Specifically, the detection data processing log files include: The types of test data, qualified documents to be verified and evaluation reports, invalid data, erroneous data, and data that are out of processing range, as well as the first calibration instruction and the second calibration instruction.

[0071] Regarding the specific description above, the detection data processing log file serves as the complete record and traceability carrier for the entire auxiliary analysis process. It is generated using a hierarchical JSON+PDF dual-format: the main JSON file stores structured fields, embeds file hash values ​​and version numbers, and is encrypted and signed; the synchronously exported PDF is presented in the format required for engineering archives, facilitating project archiving and regulatory review. The log file's main structure records six major fields in chronological order: the type of detection data, qualified documents awaiting verification and evaluation reports, invalid data removed, erroneous data, and data exceeding the processing scope, as well as the first and second calibration instructions. All records for the same detection task are linked together using a unique data link ID. When generating the log, the system automatically extracts metadata such as project name, batch number, equipment number, and personnel number, writing it to a common information segment. Subsequently, the following archiving strategy is executed for each data stream: The types of detection data are stored in a list format, such as crack image data, concrete strength rebound data, and ultrasonic attenuation data. Each type is associated with the original file name, acquisition timestamp, and device positioning coordinates. Qualified documents awaiting review and evaluation reports are stored using document ID + version number index, and are accompanied by SHA-256 hash value, generation time, and manual review approval time; The invalid data, erroneous data, and data that are out of processing range that are removed are listed in three subsections, including the enumeration code of the reason for removal (0=invalid, 1=error, 2=out of range), the original file path, the removal time, and the review engineer number; The first and second calibration instructions are recorded with instruction sequence number, instruction content summary, corresponding document version number to be verified, issuance time and number of intelligent model responses. An array of problem entries is embedded in the instruction body. Each problem entry includes problem type, coordinate index, expected modification value and final resolution status.

[0072] A specific example is as follows: In a special inspection of cracks on the K25+600 bridge of the Shanghai-Chongqing Expressway, inspectors uploaded 120 crack images, 8 sets of ultrasonic attenuation CSV files, and 200 rebound value matrices at once. The system identified the data as crack image data, ultrasonic attenuation data, and concrete strength rebound data. Three crack images were overexposed, and two rebound value matrices contained consecutive zero values, thus being marked as invalid data. One set of CSV files was marked as data exceeding the processing range due to insufficient sampling rate; all were added to the removal list in the log file. The remaining valid data was standardized using a Python script and then labeled by a convolutional neural network model, generating a verification document v1.0. The inspection engineer found that one crack image missed the end crack and two images had offset annotation boxes. Based on this, the system issued the first calibration instruction #20250623-001, recording three crack annotation correction issues. After the intelligent model completed local backpropagation and image enhancement, it output a verification document v1.1, which was then reviewed and approved before entering the evaluation process. The subsequently generated evaluation report v1.0 was found to be non-compliant due to using an outdated industry standard version. The system then generated a second calibration instruction #20250623-002, which recorded the standard version replacement issue within its body. After updating the parameters, the intelligent model output evaluation report v1.1, which passed the review. Finally, the log file stores the hash values ​​of the pending document v1.1 and the evaluation report v1.1 in the qualified document segment, and records the complete content of the two calibration instructions and the intelligent model's response summary in chronological order in the calibration instruction segment. This log file, along with the encrypted hash, is stored in the project database, providing authoritative evidence for subsequent model iterations and quality accountability.

[0073] This invention provides a deep learning-based auxiliary analysis method for engineering entity testing. The intelligent model can automatically complete the type identification, invalid data removal, defect labeling, and specification comparison of large batches of test data, replacing a large amount of repetitive manual work. This frees qualified testing engineers from intensive data processing, allowing them to focus on result review and decision-making. Even during peak periods or when testing institutions are short-staffed, it maintains stable processing capacity and service levels, reducing the risk of queuing and delays due to insufficient manpower.

[0074] Standardized deep learning neural networks and logical reasoning components maintain consistent judgment thresholds and rules across multi-dimensional data (images, waveforms, numerical reports), avoiding interpretation biases caused by individual experience differences. The human-machine collaborative first and second calibration instruction mechanism enables real-time backpropagation learning after human correction suggestions, achieving learning on the job and continuously improving the model's accuracy in identifying rare defects and complex working conditions.

[0075] Large volumes of raw data that previously required traveling to and from the project site or across networks can now be preprocessed and initially labeled directly on-site or at near-end edge computing nodes. Only documents awaiting verification and evaluation reports are uploaded for manual review, significantly reducing transmission time. Automated report templates and real-time formatting output functions enable approved reports to be finalized within minutes, meeting the needs of urgent testing tasks with high timeliness requirements.

[0076] Every data processing step, manual review, calibration instruction issuance, and model feedback is written to the testing data processing log file, forming a complete data chain and version record, providing authoritative evidence for quality responsibility traceability, regulatory audits, and subsequent project reviews. A multi-level preset evaluation parameter library (national standards, industry specifications, and project-defined parameters) and version control ensure that all judgments have clear and verifiable source clauses.

[0077] The intelligent model continuously collects challenging examples and feedback data from various engineering projects, constantly iterating its generalization capabilities to quickly adapt to different testing scenarios such as bridges, tunnels, and building structures. Testing organizations can upgrade functionality by importing new standards or parameter files without the need for deep programming, thereby reducing subsequent system maintenance and technical barriers and accelerating the standardization and intelligentization of the industry's overall testing level.

[0078] Another embodiment of the first aspect of the present invention proposes a specific implementation method for an engineering entity test detection auxiliary analysis method based on deep learning. In some embodiments of the present invention, such as... Figure 2 As shown, the specific implementation method includes the following steps: The engineering entity testing and inspection process begins with acquiring data from manual or unmanned equipment: on-site testing personnel or unmanned equipment upload raw images, waveforms, or numerical files to the system in real time. After uploading, the intelligent model filters and identifies the data type, distinguishing categories such as crack images, ultrasonic attenuation waveforms, and concrete rebound values ​​based on a deep learning feature dictionary. Subsequently, it calls Python to run professional testing and processing software, such as image enhancement scripts, waveform filtering scripts, and rebound value correction scripts, to complete the standardization process. The processing results are again assessed by the intelligent model to determine the data processing quality and generate a preliminary score. Testing engineers then manually verify the image processing quality of key samples. If the two assessments match, the data enters the usable branch; otherwise, a three-step self-learning mechanism—backpropagation to reduce gradients, supervised training for fine-tuning, and manual feedback for reinforcement—is triggered for parameter adjustment and reprocessing until the data is deemed usable. Qualified data is uniformly output as a test map and the test data is saved, completing the data screening and processing stage.

[0079] After entering the evaluation phase of the inspection results, the system hands over the inspection map to the intelligent model for recognition, extracting key features such as crack size and wave velocity threshold. Then, based on national standards, industry specifications, or project-defined thresholds, the intelligent model evaluates and derives the defect level and pass / fail conclusion. Subsequently, inspection engineers manually review the draft report. If deviations are found in the judgment logic, threshold calls, or text format, supervised training fine-tuning and human feedback reinforcement loops prompt the intelligent model to make immediate corrections until manual confirmation that the accuracy conditions are met. Finally, the system generates a dataset and outputs an evaluation report via a Python template script. Simultaneously, the inspection data types, qualified documents awaiting review and the evaluation report, details of rejected data, and all feedback instructions are encapsulated in an inspection data processing log file, achieving full-process data traceability, quality auditing, and continuous model iteration.

[0080] Another embodiment of the first aspect of the present invention proposes a specific supporting device for an engineering entity test and detection auxiliary analysis method based on deep learning. In some embodiments of the present invention, such as... Figure 3 As shown, the specific supporting equipment includes: The specific supporting equipment consists of handheld devices, unmanned data collection devices, testing equipment, data input terminals, 5G communication terminals, artificial intelligence model servers, mobile review terminals, and data review and printing terminals working together.

[0081] At the start of the inspection operation, the on-site operator uses the handheld device to collect crack photographs or rebound values; simultaneously, unmanned data acquisition equipment (such as a multi-rotor drone equipped with a high-resolution camera and ultrasonic sensors) acquires top-view images and attenuation waveforms in the arch area of ​​the bridge or tunnel; when large components require offline mechanical testing, the testing equipment outputs raw data such as load-displacement and strain-stress. The aforementioned heterogeneous data is first aggregated at the data input terminal, which has a built-in type recognition script and high-speed cache, allowing for initial local grouping, compression, and inclusion of acquisition location information.

[0082] The data input terminal establishes a low-latency link with the artificial intelligence model server via the 5G communication terminal. The server is equipped with a deep learning neural network and logical reasoning components of the intelligent model. During the upload process, the data packet carries a task number and a source identification tag. The server uses this information to call the corresponding Python detection and processing script to perform data standardization, image enhancement, or waveform filtering. Subsequently, the intelligent model performs defect identification, threshold comparison, and outputs the document to be reviewed and a preliminary evaluation report. The system instantly pushes the generated results to the mobile review terminal (the inspection engineer's smartphone or tablet). The engineer can manually review key information on-site or remotely. If an annotation error or inappropriate threshold call is found, the mobile review terminal generates a first calibration instruction or a second calibration instruction and returns it to the server in real time, triggering the intelligent model to perform backpropagation, supervised training fine-tuning, and manual feedback reinforcement to complete rapid correction and resubmit for review.

[0083] Once the mobile review terminal confirms that both the document to be reviewed and the evaluation report are qualified, the server writes the document hash and review record into the detection data processing log file, and sends it along with the detection results to the data review printing terminal. The printing terminal automatically formats the data according to the project template, affixes an electronic signature, and outputs a paper report. After on-site authorization and signature, the detection loop is completed. Throughout the process, the 5G communication terminal ensures millisecond-level round-trip communication between large volumes of data and feedback instructions; the continuous learning of the artificial intelligence model server enables the system to be transferred between multiple projects. Non-programming technicians only need to select or enter new evaluation parameters on the mobile review terminal to complete the model upgrade, thereby achieving highly efficient, low-error, and traceable auxiliary analysis for engineering entity testing.

[0084] A system is proposed in a second aspect of the present invention, such as Figure 4 As shown, the system includes: The data input module is used to receive and upload the collected detection data.

[0085] The intelligent analysis module is used to identify the data type, filter and preprocess the test data, call the corresponding test processing software to generate the document to be verified, and reprocess the document to be verified or the evaluation report according to the first or second calibration instruction received.

[0086] The human interaction module is used for human review of the documents to be verified and the evaluation reports generated by the intelligent analysis module, and to generate and send the first or second calibration instruction to the intelligent analysis module when the review is unsatisfactory.

[0087] The log generation module is used to package the approved documents to be verified, evaluation reports and corresponding test data into test data processing log files, and to store and manage them.

[0088] An embodiment of the third aspect of the present invention provides an electronic device. In some embodiments of the present invention, such as... Figure 5 As shown, an electronic device is provided, which may include: a desktop computer, a laptop, a handheld computer, and a cloud server, etc. The electronic device 3 may include, but is not limited to, a processor 301 and a memory 302. Those skilled in the art will understand that... Figure 5 This is merely an example of electronic device 3 and does not constitute a limitation on electronic device 3. It may include more or fewer components than shown, or different components.

[0089] The processor 301 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.

[0090] The memory 302 can be an internal storage unit of the electronic device 3, such as a hard disk or RAM of the electronic device 3. The memory 302 can also be an external storage device of the electronic device 3, such as a plug-in hard disk, Smart MediaCard (SMC), Secure Digital (SD) card, or Flash Card equipped on the electronic device 3. The memory 302 can also include both internal and external storage units of the electronic device 3. The memory 302 is used to store computer programs and other programs and data required by the electronic device.

[0091] An embodiment of the fourth aspect of the present invention provides a computer-readable storage medium. In some embodiments of the present invention, a computer-readable storage medium is provided that, when executed by processor 301, implements the steps of the above-described method. Therefore, the computer-readable storage medium provided in the fourth aspect of the present invention has all the technical effects of the above-described steps, which will not be repeated here.

[0092] If an integrated module / unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program may include computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. A computer-readable medium may include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in a computer-readable medium may be appropriately added to or subtracted according to the requirements of legislation and patent practice in a jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media may not include electrical carrier signals and telecommunication signals.

[0093] The above embodiments are only used to illustrate the technical solutions of this disclosure, and are not intended to limit it. Although this disclosure has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this disclosure, and should all be included within the protection scope of this disclosure.

Claims

1. A deep learning-based auxiliary analysis method for engineering entity testing, characterized in that, Includes the following steps: The collected detection data is input into the intelligent model, the type of the detection data is filtered and identified, and the corresponding detection processing software is called according to the type to process the detection data and generate a document to be verified. The system manually determines whether the currently generated document to be verified is qualified; if the result is unqualified, a first calibration instruction is generated and sent to the intelligent model to regenerate the document to be verified, until the result is qualified. The qualified document to be verified is input into the intelligent model, and an evaluation report is generated according to the preset evaluation parameters. The current evaluation report is judged by the human to be qualified. If the judgment result is unqualified, a second calibration command is generated and sent to the intelligent model to generate an evaluation report again, until the judgment result is qualified. Generate a test data processing log file based on the qualified documents to be verified, the evaluation report, and the corresponding test data.

2. The auxiliary analysis method for engineering entity testing according to claim 1, characterized in that, The step of filtering and identifying the type of the detection data includes: The intelligent model uses a deep learning neural network to identify the type of the detection data, and then removes invalid data, erroneous data, and data that is outside the processing range based on the type.

3. The auxiliary analysis method for engineering entity testing according to claim 2, characterized in that, The step of calling the corresponding detection processing software to process the detection data and generate the document to be verified according to the type includes: Based on the type, the detection and processing software written in Python is invoked to complete the data standardization process in order to generate images or reports; The images or reports are input into the convolutional neural network model of the intelligent model to perform data recognition and annotation, and the annotated images or reports are used to form the document to be verified.

4. The auxiliary analysis method for engineering entity testing according to claim 1, characterized in that, The step of generating the first calibration command and sending it to the intelligent model to generate the calibration document again includes: Feedback data is generated for identification defects, annotation errors, annotation omissions, and unclear images in the document to be verified. This feedback data is sent to the intelligent model to perform feedback learning, and the document to be verified is generated again based on the feedback data.

5. The auxiliary analysis method for engineering entity testing according to claim 4, characterized in that, The steps of the feedback learning include: The intelligent model uses backpropagation to reduce the loss gradient, performs supervised training and fine-tuning, and optimizes the model through human feedback reinforcement to improve the generalization ability of the intelligent model in continuous operation.

6. The auxiliary analysis method for engineering entity testing according to claim 1, characterized in that, The preset evaluation parameters include: Engineering entity testing and inspection standards, national or industry standard parameters, or project-specific evaluation parameters set by the testing organization itself.

7. The auxiliary analysis method for engineering entity testing according to claim 2, characterized in that, The detection data processing log file includes: The types of detection data, qualified documents to be verified and evaluation reports, invalid data, erroneous data, and data that are out of processing range, as well as the first calibration instruction and the second calibration instruction.

8. A system for implementing the auxiliary analysis method for testing and detecting engineering entities according to any one of claims 1-7, characterized in that, include: The data input module is used to receive and upload the collected detection data; The intelligent analysis module is used to identify the data type, filter and preprocess the detection data, call the corresponding detection processing software to generate the document to be verified, and reprocess the document to be verified or the evaluation report according to the first calibration instruction or the second calibration instruction received. The human interaction module is used to manually review the document to be verified and the evaluation report generated by the intelligent analysis module, and generate and send the first calibration instruction or the second calibration instruction to the intelligent analysis module when the review is unsatisfactory. The log generation module is used to package the approved documents to be verified, evaluation reports and corresponding test data into test data processing log files, and to store and manage them.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1 to 8.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 8.

Images