An AI-based gas security inspection intelligent assistance method

Abstract

The application discloses an AI-based gas security inspection intelligent auxiliary method and belongs to the technical field of gas safety detection. The method comprises the following steps: obtaining a structured hidden danger questionnaire matched with a security inspection work order and a security inspection image collected on site; inputting the image into a target detection model to output hidden danger labels and basic recognition confidence; dynamically adjusting and generating a judgment threshold according to prior physical environment parameters such as the equipment age related to the work order; mapping the hidden danger labels to corresponding items from a rule base, and automatically preselecting an abnormality when the confidence reaches the dynamic threshold; receiving manual review to confirm the final abnormality item; calculating the overall risk score according to the abnormality item and the operation type to generate a structured security inspection data package. The application reduces artificial subjective differences, realizes accurate closed-loop mapping from unstructured images to structured forms, and improves the objective uniformity of hidden danger detection rate and risk rating.

Description

Technical Field

[0001] This invention belongs to the field of gas safety detection technology, specifically an AI-based intelligent auxiliary method for gas safety inspection. Background Technology

[0002] Gas safety inspections are a crucial link in ensuring the operation of urban gas supply systems, and their quality directly affects user gas safety and public safety. In recent years, with the continuous expansion of urban gas coverage and the increasing complexity of user gas usage scenarios, the traditional in-home safety inspection model is facing increasing pressure in terms of efficiency and accuracy.

[0003] Currently, gas safety inspections at homes are mainly conducted by inspectors who follow a pre-set checklist, recording the results of hazard identification through mobile devices or paper forms. This manual inspection method adequately met basic safety inspection needs in the early stages of the gas industry's development. However, as application scenarios have deepened and the industry's requirements for inspection quality have continued to rise, its inherent limitations have become increasingly apparent.

[0004] From the perspective of security inspection execution, different security personnel often have inconsistent standards for judging the same potential hazard. This subjective difference mainly stems from the varying levels of experience among inspectors and differing understandings of hazard definitions, leading to significant individual variations in inspection results and affecting the uniformity and comparability of hazard identification. From the perspective of information recording, on-site photos are primarily used for post-inspection documentation and do not provide real-time assistance in hazard identification during the inspection. Security personnel need to make secondary judgments based on memory or photos after completing the on-site inspection, thus limiting the timeliness and accuracy of information transmission.

[0005] From a business process perspective, when the on-site environment is complex, there are many potential hazards, or the security inspectors lack experience, problems such as missed inspections, incorrect filling, and disconnect between evidence and questions are likely to occur. The rectification methods, rectification deadlines, signature information, and security inspection results are usually scattered across multiple interfaces, lacking an effective correlation and integration mechanism, which is not conducive to subsequent review, accountability, and in-depth data value mining.

[0006] With the rapid iteration and development of computer vision technology and the significant improvement of the computing power of mobile smart terminals, the use of artificial intelligence models to automatically detect potential hazards in security inspection images has reached a mature engineering application level at the algorithm level.

[0007] However, objectively speaking, most existing technical solutions only focus on the simple display of recognition results, failing to effectively integrate image recognition results with security questionnaire items, human professional judgment, rectification parameter configuration, and business submission processes. This prevents artificial intelligence recognition capabilities from being truly embedded into the complete closed loop of security inspection operations. This disconnect between recognition results and business logic not only limits the maximization of the effectiveness of artificial intelligence technology in actual security inspection scenarios but also makes it difficult for security personnel to conveniently utilize recognition results to optimize their inspection decision-making processes.

[0008] In summary, how to construct an intelligent auxiliary solution that can deeply integrate artificial intelligence recognition capabilities with security inspection business processes, so that image recognition results can be automatically associated with corresponding security inspection questionnaire items, and generate complete structured security inspection data packages based on recognition results, questionnaire answers, risk scores, and rectification parameters, has become a key technical challenge and an urgent technical problem to be solved by those skilled in the art. Summary of the Invention

[0009] The purpose of this invention is to provide an AI-based intelligent auxiliary method for gas safety inspection, in order to solve the problems in existing gas safety inspection technologies mentioned in the background art, such as large subjective differences in inspection results, lack of correlation between image recognition and questionnaire items, scattered hidden danger information and rectification parameters, and failure of artificial intelligence recognition capabilities to be deeply embedded in the business closed loop.

[0010] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: An AI-based intelligent assistance method for gas safety inspection includes the following steps: Step S1: Obtain a structured hazard questionnaire adapted to the current security inspection work order. The structured hazard questionnaire contains multiple hazard items with different risk levels. Step S2: Obtain images of the security inspection target objects collected on-site; Step S3: Input the image into the pre-trained target detection model to detect potential hazards, and output hazard labels and basic recognition confidence scores; Step S4: Based on the prior physical environment parameters of the on-site equipment related to the security inspection work order, dynamically adjust and generate the corresponding dynamic judgment threshold. Step S5: In the pre-set hazard label and item rule base, determine the target hazard item that is mapped to the hazard label. When the basic identification confidence is greater than or equal to the dynamic judgment threshold, automatically switch the status of the target hazard item in the structured hazard questionnaire to the abnormal pre-selection status. Step S6: Receive manual review instructions for the abnormal pre-selected status to confirm the potential issues that are ultimately in an abnormal status; Step S7: Based on the hazard items that are ultimately in an abnormal state and the operation type of the manual review instruction, calculate the overall risk score value, and generate a structured security inspection data packet based on the risk score value.

[0011] Based on the above technical solution, the steps for obtaining a structured hazard questionnaire adapted to the current security inspection work order specifically include: A hierarchical rule matching algorithm is adopted. First, a basic question set is established based on the user type basic classification rules. Then, the mandatory questions in the industry security inspection mandatory inspection rule base are incorporated into the basic question set and deduplicated according to the question identifier. Finally, non-mandatory questions are filtered according to the security inspection scenario adaptation rules. When there are conflicts among the rules, the industry security inspection mandatory inspection rule base has the highest priority, followed by the user type basic classification rules, and the security inspection scenario adaptation rules have the lowest priority. A structured hazard questionnaire containing hazard category grouping identifiers, question type configurations, and question item metadata lists is generated.

[0012] Based on the above technical solution, the steps for outputting hazard labels and basic identification confidence levels specifically include: The target detection model outputs multiple candidate boxes and corresponding classification score vectors for each hazard category. The non-maximum suppression algorithm is used to remove overlapping and redundant boxes, and the coordinates of the effective region are retained as the location boxes of potential hazards. Normalize the classification score vector corresponding to the effective region coordinates to obtain the probability distribution of each category, and take the maximum probability as the basic recognition confidence. The hazard category with the highest probability is matched with the built-in standard gas hazard label library to generate hazard labels; if the probability difference between the top two hazard categories is less than a preset threshold, a set of multiple candidate labels is output.

[0013] According to the above technical solution, the steps of dynamically adjusting and generating corresponding dynamic judgment thresholds based on the prior physical environment parameters of the on-site equipment related to the security inspection work order specifically include: Determine the baseline threshold that matches the risk level of the hazard label; Obtain the scenario adjustment coefficient set based on the user type in the security check work order; Obtain the equipment age adjustment coefficient set based on the service life of the on-site equipment. The longer the service life of the equipment, the smaller the corresponding equipment age adjustment coefficient. The dynamic judgment threshold for the current work order and issue is calculated by multiplying the baseline threshold, the scenario adjustment coefficient, and the equipment age adjustment coefficient.

[0014] According to the above technical solution, the process of determining the target hazard item that maps to the hazard label also includes: The scenario keyword matching rules are used for secondary verification. User type and equipment scenario in the security inspection work order are extracted as keywords. Keywords are used as filtering conditions to exclude mapping results that are irrelevant to the hazard label and question rule base.

[0015] According to the above technical solution, the manual review instruction includes confirmation, rejection, and manual data entry; the steps for receiving a manual review instruction for an abnormal pre-selected status specifically include: If a rejection operation is received, the status of the corresponding potential hazard item will be restored to normal and it will not be included in the calculation set of subsequent risk score values. If a manual supplementary entry is received, the system will respond to the security personnel's manual selection of potential hazards not identified by the target detection model, and directly change the status of the corresponding hazard to abnormal.

[0016] According to the above technical solution, the step of calculating the overall risk score is implemented using a multi-factor weighted summation algorithm, specifically including: For each hazard item that is ultimately in an abnormal state, obtain its corresponding hazard level weight, confidence coefficient, manual confirmation factor, and historical rule factor, and multiply the above four factors to obtain the individual risk product of each hazard item; The individual risk products of all the potential problems in abnormal states are summed to obtain the conventional summation result; The result of the conventional summation calculation is multiplied by a preset scoring scaling factor, and the result of the multiplication is compared with the set upper limit of the full score range. The minimum value between the two is taken as the final output risk score value.

[0017] According to the above technical solution, the assignment of the confidence coefficient specifically includes: If a potential hazard item has a matching basic identification confidence level, then the original value of the basic identification confidence level is used as the confidence coefficient of that potential hazard item. If the abnormal state of a potential problem item is triggered by a manual data entry operation, then the confidence coefficient of that potential problem item will be forcibly assigned to the preset maximum full confidence value.

[0018] According to the above technical solution, the AI-based intelligent auxiliary method for gas safety inspection incorporates the highest-level mandatory coverage business rules; the process of calculating the overall risk score also includes: When there are serious hidden danger items with a level 1 hazard among all the hidden danger items that are ultimately in an abnormal state, and which have been confirmed by confirmation operation or manual supplementation operation, determine whether the score after multiplying the result of the regular summation calculation by the scoring scaling factor is less than the preset level 1 risk lower limit. If so, skip the usual calculation step of taking the minimum value and directly force the risk score value to be the lower limit of the first-level risk; when it is determined to be a serious leak, directly force the risk score value to be the upper limit of the full score range.

[0019] According to the above technical solution, the steps for generating structured security inspection data packets based on risk score values ​​specifically include: Based on the preset scoring range in which the risk score value falls, the overall hazard level corresponding to the current security check record is determined; Based on the overall level of hazard, the corresponding rectification method and rectification period parameters are automatically matched; The final status of the structured hazard questionnaire, the results of hazard identification and manual confirmation, image evidence, signature information, and the matching rectification methods and rectification deadlines are integrated into a single data package in JSON format to generate a structured security inspection data package. The generated structured security data packets are hashed using a digital signature algorithm and then digitally signed to achieve tamper-proof protection.

[0020] Compared with the prior art, the present invention has the following beneficial effects: This invention constructs a dynamic judgment threshold adaptive adjustment mechanism, which applies prior physical environment parameters such as user type and equipment age to the threshold in advance. While maintaining the objectivity of the underlying probability inference of the model, it effectively solves the problem of unstable triggering of universal confidence in specific scenarios, significantly improves the accuracy of detecting hidden dangers of high-risk or aging equipment and reduces human subjective differences.

[0021] This invention constructs a precise mapping channel from unstructured image data to structured security inspection forms. By utilizing a dual verification and matching system of hazard labels-question rule base and scene keywords, it achieves automated pre-selection and visual highlighting of abnormal questions, reducing the time for converting recognition results into business actions to the second level, and significantly improving the efficiency and accuracy of on-site form filling.

[0022] This invention integrates information such as the final state of potential hazards, human-computer interaction process, image evidence, and electronic signatures, and combines digital signature and hash verification mechanisms to generate structured security inspection data packets, forming a data closed loop with strong binding of evidence, operation, and conclusion. This fundamentally prevents result tampering and fake security inspections, and fully ensures the standardization and end-to-end traceability of security inspection results. Attached Figure Description

[0023] Figure 1 This is a flowchart of the intelligent auxiliary method for gas safety inspection of the present invention; Figure 2 This is a diagram illustrating the overall technical architecture of the intelligent auxiliary method for gas safety inspection of the present invention. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. 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.

[0025] Example 1 like Figure 1 and Figure 2 As shown, this invention provides an AI-based intelligent auxiliary method for gas safety inspections. Its overall technical architecture consists of six core functional modules: a questionnaire loading module, an image acquisition module, an AI recognition module, a question mapping module, a manual review module, and a work order submission module. These modules exchange information through standardized data interfaces, forming a complete business loop from work order access to safety inspection data submission.

[0026] Specifically, the questionnaire loading module, as the starting point of the entire security check process, is responsible for receiving security check work orders and generating a structured hazard questionnaire adapted to the current work order. The core technology of this module relies on two major technical components: a hierarchical rule matching algorithm and a dynamic form rendering model.

[0027] like Figure 1 As shown, in the specific implementation of the hierarchical rule matching algorithm, the algorithm logic is executed sequentially in three levels. The first level is the basic classification rule for user types. This level performs initial classification based on the user attribute information bound to the work order. When the user type field value is "residential" (i.e., residential user), a basic hazard question bank specifically for residential users is loaded. This question bank contains six primary hazard categories: pipeline inspection, stove inspection, alarm inspection, meter inspection, hose inspection, and gas usage environment inspection, totaling 45 sub-items of hazard. When the user type field value is "non-residential" (i.e., industrial and commercial user), a basic hazard question bank specifically for industrial and commercial users is loaded. This question bank adds three special groups to the residential user question bank: pipeline explosion-proof inspection, gas metering equipment special inspection, and large gas-using equipment inspection, totaling 68 sub-items of hazard. The differences between the question banks for residential users and industrial and commercial users are mainly reflected in two aspects: the adaptability of gas usage scale and equipment configuration, and daily gas usage risk points. Residential users focus on the inspection of household gas appliances and indoor facilities, while industrial and commercial users focus on the safety and metering compliance inspection of industrial gas-using equipment.

[0028] Furthermore, the second level is the industry-wide mandatory inspection rule base, which is bound to national gas safety regulations and internal enterprise safety inspection standards to form a fixed rule set. The mandatory inspection rule base has a built-in hazard classification and labeling system, which divides hazards into three levels of danger: Level 1 hazards, Level 2 hazards, and Level 3 hazards. Level 1 hazards include gas leak-related hazards (such as loose interfaces, obviously abnormal condensation or oil stains, corrosion and perforation, etc.), hose detachment, alarm failure, and other hazard types that directly endanger personal safety; Level 2 hazards include gas meter malfunctions, stoves without flameout protection, improper water heater installation, and damage to explosion-proof equipment components, and other hazard types that pose a greater safety risk; Level 3 hazards include non-standard concealed pipe installation (the visual judgment criteria include at least one of the following observable characteristics: pipes entering enclosed walls without visible maintenance port markings, pipes being completely covered by fixed ceilings without reserved inspection ports, pipes not having standard sleeves installed at wall penetrations, etc.) and minor pipe corrosion, etc. All items in the mandatory check rule base are marked as mandatory selection items and cannot be deleted when the questionnaire is generated. The question type configuration includes three types: single choice, multiple choice, and text note questions.

[0029] Furthermore, the third level consists of security inspection scenario adaptation rules. This level fine-tunes the questionnaire based on the security inspection scenario type corresponding to the work order. Security inspection scenario types include four categories: routine inspection, fault-related inspection, re-inspection, and special inspection. In routine inspection scenarios, a complete set of standard questions is loaded, with all primary hazard category groups and secondary hazard groups displayed. In fault-related inspection scenarios, fault-related hazard groups are prioritized; for example, when the fault type is gas leak, only the pipeline inspection group and hose inspection group are loaded, while other groups are hidden by default. In re-inspection scenarios, groups related to hazard rectification in the previous inspection record are prioritized to verify rectification implementation. In special inspection scenarios, corresponding groups are prioritized according to the special inspection requirements; for example, in alarm-specific inspection scenarios, the alarm inspection group is prioritized. When these three levels of rules are superimposed, they follow a priority judgment logic. When multiple rules conflict, the second-level industry-mandated security inspection rule library has the highest priority, followed by the first-level user type basic classification rules, with the third-level security inspection scenario adaptation rules serving as a fine-tuning supplement. The algorithm filters and assembles data layer by layer according to priority, ultimately generating a customized, interactive, structured hazard questionnaire tailored to the current work order. After the rule matching operation is completed, the output questionnaire structure data includes a list of group identifiers, a list of question type configurations, and a list of question metadata. This output is then transmitted to the dynamic form rendering model for page rendering.

[0030] To avoid conflicts between scenario adaptation rules and mandatory inspection rules, a basic set of questions is first established based on user type during questionnaire assembly. Then, mandatory questions from the industry's mandatory inspection rule library are incorporated into this set and deduplicated according to question identifiers. Finally, scenario filtering is performed only on non-mandatory questions. Therefore, even in scenarios involving fault-related security checks, re-inspections, or special security checks, mandatory questions remain visible and cannot be skipped; scenario rules only control the display order and activation status of the remaining questions.

[0031] In a preferred embodiment, the dynamic form rendering model uses a hazard classification hierarchical data model as its underlying data structure. This data model categorizes hazards into six primary hazard categories: pipelines, stoves, alarms, meters, hoses, and gas usage environment. Each primary hazard category includes several sub-hazard items. For example, the primary hazard category of stoves includes five sub-items: flameout protection device inspection, stove installation location inspection, stove connection pipe inspection, stove service life inspection, and stove brand and model verification. The form metadata model pre-stores three types of question templates, each bound to a specific question type attribute. The single-choice template is bound to an option set attribute, a required field identifier attribute, a hazard level attribute, an associated AI hazard tag attribute, and an input length limit attribute. The multiple-choice template adds an option quantity limit attribute to the single-choice template attributes. The text comment question template is bound to an image upload interface attribute and an input format validation rule attribute.

[0032] Specifically, after receiving the group identifiers, question type configurations, and option data output by the algorithm, the form rendering engine automatically generates a column-based interface displaying major hazard categories. Within each group, three question types—single-choice, multiple-choice, and text-annotated questions—are loaded as needed. Text-annotated questions are linked to an image upload interface during interface display, allowing security personnel to take photos of the scene and attach explanatory text. Single-choice and multiple-choice questions have two preset standard options, "normal" and "abnormal," during interface rendering. The status of single-choice questions will automatically change based on the judgment logic of the subsequent question mapping module, and will be visually highlighted to alert security personnel.

[0033] Regarding specific parameter configurations, in regular inspection scenarios, all six primary hazard categories and all sub-categories are loaded, with a total of 45 to 68 items, determined based on user type and inspection scenario. In fault-related inspection scenarios, only one or two fault-related hazard groups are activated, with 8 to 15 items. In re-inspection scenarios, only groups involving rectification and verification are loaded, with 3 to 8 items. The industrial and commercial user group adds explosion-proof inspection multiple-choice questions, gas metering multiple-choice questions, and large gas-consuming equipment inspection multiple-choice questions to the residential user group. The explosion-proof inspection multiple-choice questions contain 8 items, the gas metering multiple-choice questions contain 6 items, and the large gas-consuming equipment inspection multiple-choice questions contain 9 items; these 23 items, together with the 45 items in the residential user question bank, constitute the 68 items in the industrial and commercial user question bank. All grouping and question type configuration parameters are persistently stored in the configuration table of the database. The rules parameters can be modified through the backend management system. The modified rules will take effect immediately when the next work order is processed, without the need to change the rendering engine code.

[0034] Specifically, the image acquisition module is responsible for acquiring images of target objects related to gas safety inspections on-site, providing raw data input for subsequent AI recognition. The module's technical implementation includes three core functions: image acquisition range definition, image acquisition process control, and image metadata management.

[0035] The specific definition of the image acquisition range includes six types of target object images: gas environment images, gas meter images, gas appliance or water heater images, alarm images, connecting hose images, and safety inspection report images. Gas environment images capture environmental information such as the overall kitchen layout, ventilation conditions, and gas pipeline laying status to determine if there are potential hazards in enclosed spaces or improperly concealed pipeline installations. Gas meter images capture information such as the gas meter's appearance, reading, and model identification to verify meter information and detect any abnormalities. Gas appliance or water heater images capture information such as the stove's brand and model, flameout protection device installation status, water heater installation location, and exhaust pipe condition to determine if there are any safety hazards related to the stove or water heater installation. Alarm images capture information such as the alarm's appearance, display status, and installation location to determine if the alarm is functioning properly. Connecting hose images capture information such as hose material, connection status, and service life markings to determine if there are any potential hazards such as hose aging or detachment. The security checkpoint image captures handwritten information filled in by security personnel, which is used to verify the information against system records and extract signature information.

[0036] Furthermore, the image acquisition process control adopts a standardized operation guidance mechanism. The system automatically generates an image acquisition list based on the user type and security inspection scenario bound to the work order, specifying the types of images to be acquired and the recommended acquisition order. Security personnel acquire images item by item according to the list order, and the system performs real-time verification during the acquisition process. When a certain type of image is missing, the system displays a prompt message reminding security personnel to acquire it. When the image quality does not meet the requirements, such as blurry images, insufficient lighting, or the target area not being fully captured, the system provides immediate feedback and requests a retake. Multiple resolution options are supported during image acquisition; the default resolution is 1920×1080 pixels, which can be switched to 1280×720 pixels in scenarios with limited storage space.

[0037] In one optional image quality verification implementation, the system first scales the image to a fixed detection size, calculates the Laplacian variance as the sharpness index Vblur and the grayscale mean as the exposure index Lmean, and calculates the proportion Rtarget of the target candidate region to the image area. When Vblur is lower than a preset sharpness threshold, Lmean exceeds a preset exposure range, or Rtarget is lower than a preset integrity threshold, qualityPass is set to false and the corresponding retake reason is returned; otherwise, qualityPass is set to true and the AI ​​recognition process begins. Each threshold is determined by a calibration image set from different mobile terminal models and stored in the device parameter table, making the quality judgment parameters configurable and traceable.

[0038] Furthermore, image metadata management enables comprehensive recording of acquired information. Each time an image is acquired, the system synchronously records the following metadata: acquisition time accurate to milliseconds, acquisition location obtained via the mobile device's GPS module (latitude and longitude coordinates), device identifier (unique serial number of the acquisition device), acquisition personnel identifier (currently logged-in security personnel account), and work order identifier (currently processed security work order number). This metadata information is stored along with the image file and associated with the corresponding work order in subsequent processes. The metadata is stored in JSON format and includes eight fields: captureTime, captureLatitude, captureLongitude, deviceId, operatorId, workOrderId, imageType, and imageUuid.

[0039] Specifically, the AI ​​recognition module is the core technical component of this invention, responsible for automatically detecting and labeling potential hazards in the acquired on-site images. The technical implementation of this module includes four processing stages: image preprocessing, target detection, confidence extraction, and label generation. Each stage is executed sequentially to form a complete recognition processing pipeline.

[0040] In the image preprocessing stage, the system receives and standardizes various types of on-site images. This stage receives five types of raw input: gas environment images, gas meter images, gas appliance or water heater images, alarm images, and connecting hose images. Size standardization is performed to uniformly adjust the input images to 600×600 pixels. Gaussian filtering is used for noise reduction to eliminate random noise during image acquisition, and adaptive histogram equalization is used for illumination correction to eliminate the impact of different lighting conditions on recognition results. After preprocessing, a standardized image tensor is generated with dimensions of 3×600×600, and the three channels correspond to the RGB color space. Simultaneously, a unique identifier, `originalImageId`, is generated for each raw image. This identifier is a 32-bit UUID string used for subsequent image source tracing and evidence association. The `originalImageId` value is consistent with the `imageUuid` value in the image metadata and is output as the `imageId` field in the recognition result JSON. All three point to the same unique identifier of the original image. After the labeled image is generated, it is also assigned a unique identifier, annotatedImageId, which is also in 32-bit UUID format.

[0041] Furthermore, the target detection stage employs an AI-based gas hazard-specific target detection model for hazard identification. This model is built upon a deep learning target detection framework, using ResNet50 as the backbone network to extract multi-scale image features, a feature pyramid network to achieve multi-scale feature fusion, a region proposal network to generate candidate detection boxes, and a classification regression head to output the final detection results. The model training sample library contains over 12,000 standard image samples of various gas safety hazards, covering visual features related to aging hoses, hose detachment, alarm failure, stoves without flameout protection, improper water heater installation, and gas leaks. The model outputs multiple sets of candidate boxes, each carrying four pixel coordinate parameters, representing the coordinates of the top-left corner of the detection box. , and the coordinates of the bottom right corner , The algorithm includes the predicted score vectors for all hazard categories. In the target detection stage, the Non-Maximum Suppression (NMS) algorithm is used to remove overlapping and redundant boxes. The IoU threshold in the NMS algorithm is set to 0.75; for detection boxes with an overlap exceeding the threshold, only the one with the highest confidence is retained. After the algorithm is executed, the coordinates of the effective area are stored as the hazard location box. This location box is defined by... Formatted output.

[0042] During the model training phase, training samples are randomly divided into training, validation, and test sets based on work order identifiers rather than individual images to avoid data leakage caused by similar images from the same site crossing sets. Labeled data includes at least the hazard category and bounding box coordinates. During training, images are subjected to random scaling, brightness perturbation, horizontal flipping, and local occlusion enhancement, with the weighted sum of classification loss, bounding box regression loss, and region proposal loss used as the optimization objective. After each training round, precision, recall, and mean precision for each category are calculated on the validation set. Training stops when there is no improvement after a preset number of rounds, and the model parameters with the best validation metrics are saved. Finally, the online model is evaluated using only an independent test set.

[0043] In a preferred embodiment, the confidence extraction step normalizes the classification score vector of each valid detection box to obtain the probability distribution of each category, and takes the maximum probability as the basic recognition confidence. To maintain the objectivity of the model's probability inference, this module does not directly apply prior information to modify the confidence scores. Instead, it uses... The original value is output to the subsequent question mapping module, which then dynamically adapts it at the threshold.

[0044] The tag generation process incorporates a standard gas hazard tag library, which maps hazard category identifiers to standardized descriptive text. The library covers 28 standard hazard categories, including aging hoses, hose detachment, gas leak-related hazards (including loose connections, obvious abnormal condensation or oil stains, corrosion perforation, etc.), alarm malfunction, alarm not powered, stove without flameout protection, improper water heater installation, poor water heater exhaust, abnormal gas meter, pipe corrosion, and non-standard concealed pipe installation (judgment criteria include at least one of the following observable characteristics: pipes entering enclosed walls without visible inspection port markings, pipes being completely covered by fixed ceilings without reserved inspection ports, pipes not having standardized sleeves at wall penetrations), and enclosed spaces. During tag generation, the hazard category with the highest probability in a single frame is matched with the text identifier in the tag library to generate the hazard tag text. If the probability difference between the top two hazard categories in the detection frame is less than a preset threshold of 0.15, a set of multiple candidate tags is output for manual review. The label generation process ultimately integrates the original image identifier (originalImageId), the annotated image identifier (annotatedImageId), the hazard label, the pixel coordinates of the hazard location bounding box, and the final basic recognition confidence level. The complete output is sent to the question mapping module. The output data format of the recognition results is a JSON object, which includes fields such as imageId, annotatedImageId, labels, boundingBox, confidence, timestamp, modelVersion, and paramVersion. Among them, modelVersion records the model version number, and paramVersion records the dynamic threshold parameter version number, which is used to restore the judgment conditions of the current time during review.

[0045] Furthermore, the AI ​​recognition module supports both online and offline recognition modes to adapt to different network environments and device conditions. In online recognition mode, the mobile device transmits the captured images to the server via HTTPS protocol. The server is equipped with an NVIDIA Tesla V100 GPU for model inference, with a single image processing time of approximately 0.3 seconds. In offline recognition mode, a lightweight object detection model is deployed in the mobile app. The model file size is approximately 45MB, and model quantization technology is used to convert floating-point parameters to INT8 parameters to reduce computing resource requirements. The mobile inference time is approximately 0.8 seconds. The recognition results data format remains consistent in both modes, and the question mapping module and the manual review module do not need to distinguish the data source.

[0046] Specifically, the item mapping module is responsible for automatically associating and mapping the hazard detection results output by the AI ​​recognition module with specific items in the structured hazard questionnaire, thereby achieving accurate conversion of recognition results into business items. The technical implementation of this module includes three core functions: building a mapping relationship library, executing mapping rules, and dynamically adjusting thresholds.

[0047] The mapping relationship library uses a hazard label and question rule base as its core benchmark. This rule base is independent of the gas hazard standard label library used in the label generation process: the gas hazard standard label library is used to generate hazard label text from hazard categories, while the hazard label and question rule base is used to map questionnaire questions from the generated hazard labels. The hazard label and question rule base pre-defines the correspondence between hazard labels and safety inspection questionnaire questions. Each mapping rule includes four attributes: hazard label identifier, question identifier, matching mode, and priority weight. The hazard label identifier corresponds to the hazard label text output by the AI ​​recognition module, the question identifier corresponds to the specific question number in the structured hazard questionnaire, and the matching mode includes two types: exact matching and fuzzy matching. For example, the hazard label "hose aging" exactly matches the questionnaire question "Does the connecting hose show signs of aging or cracking?", and the hazard label "alarm malfunction" exactly matches the questionnaire question "Is the alarm working properly?".

[0048] During the execution of the mapping rules, a secondary validation is performed using scenario keyword matching rules. These rules extract information such as user type, equipment scenario, and gas consumption scale from the work order as keywords, and then use these keywords as filtering conditions to exclude irrelevant mapping results. For example, when the work order user type is residential, the mapping rules automatically filter out hazard tags and item mappings that only apply to industrial and commercial users.

[0049] Each mapping rule also records the ruleId, ruleVersion, effectiveTime, sceneScope, and enabled status. During mapping, an exact match is first performed based on the hazard label identifier; if the exact match result is empty, a fuzzy match is then performed on the standardized label text and question keywords, sorted by priority weight. If multiple questions have the highest priority or the score difference between multiple candidate results is less than a preset conflict threshold, the question status is not automatically switched; instead, a candidate set is generated for manual review. The system writes the matched ruleId and ruleVersion to the manualReviewResult, ensuring traceability of the mapping basis.

[0050] In a preferred embodiment, the threshold determination logic employs a dynamic threshold adaptive adjustment mechanism. To address the technical problem of unstable detection performance and potential for missed detections or false alarms in specific scenarios (such as new or old devices, different user environments) when using a fixed confidence threshold, the system uses the basic recognition confidence level output by the AI ​​recognition module. With dynamic judgment threshold A comparison is made, and this is used as the trigger condition for automatic pre-selection. This invention does not use the method of directly modifying the model's posterior probability, but dynamically determines the threshold. The adjustment formula is as follows:

[0051] In the formula: This represents the final judgment threshold dynamically generated for the current work order and question item; This represents the baseline threshold, determined by the hazard identification accuracy calibration (set for the identification category corresponding to Level 1 hazards). To improve detection sensitivity; to set up routine identification categories for Level 2 and Level 3 hazards. ); This represents the scenario adjustment coefficient, determined based on the work order user type. For residential user scenarios... Industrial and commercial user scenarios ; This represents the equipment age adjustment factor, which is determined based on the service life of the equipment on site. Equipment with less than one year of service life is included. =1.05, equipment with a service life of one year or more but not exceeding three years. =1.0, equipment that has been in use for more than three years =0.90.

[0052] benchmark threshold Calibration can be performed separately according to hazard category: Candidate thresholds are traversed on an independent calibration set, and precision, recall, and false alarm rate are calculated. For Level 1 hazards, the threshold with the lowest false alarm rate is selected from the candidate values ​​that meet the preset minimum recall rate constraint. For Level 2 and Level 3 hazards, the threshold with the highest F1 score is selected. Scenario adjustment coefficient. Equipment age adjustment factor Offline calibration is performed based on the missed detection rate grouped according to different user types and device age. The parameter version number and model version number are written into the recognition result so that the judgment conditions of the previous time can be restored during the review.

[0053] The specific execution logic is as follows: If the hidden danger label matches the question description of the single-choice question successfully, and satisfies the following conditions... If the condition is met, the system automatically switches the single-choice question to an abnormal pre-selection state, highlights it in yellow on the interface, and displays the confidence level value and a thumbnail of the highlighted image next to the question. If the hazard label corresponds to multiple candidate questions, a set of candidate results awaiting manual confirmation is generated. At this time, the system only generates a prompt message and displays it in the prompt area at the bottom of the interface, without directly triggering automatic pre-selection.

[0054] Specifically, the manual review module is responsible for receiving manual confirmation, correction, or supplementary notes from security personnel regarding the automatically pre-selected results, thus forming the final hazard assessment. This module provides four operation interfaces: confirmation, rejection, manual entry, and text note input, supporting the complete review process for security personnel. The confirmation, rejection, and manual entry operations are manual review instructions that change the status of the question items, while the text note input is used to supplement information and does not change the status of the question items.

[0055] The confirmation interface displays a list of pre-selected abnormal items automatically selected by the system. Each pre-selected item is associated with a corresponding marked image and a box indicating the location of the potential hazard. After the security inspector clicks the confirmation button, the status of that item is officially changed to abnormal, and the associated image evidence is simultaneously locked and saved.

[0056] The rejection interface allows security personnel to reject automatically pre-selected results. When security personnel determine that the pre-selected abnormal status is invalid, they can click the "Reject" button. The rejection operation requires security personnel to select a reason for rejection. Rejection reason options include: image misdetection, hazard has been rectified, not within the scope of inspection, and other reasons. When selecting other reasons, the system will display a text input box requiring security personnel to enter the specific reason for rejection. It is important to note that after a rejection operation, the status of that item will revert to "normal," and therefore it will no longer be included in the abnormal item set or participate in subsequent risk scoring calculations. Rejection operation records are used for subsequent statistical analysis of AI recognition accuracy and optimization of model parameters.

[0057] The text annotation input interface allows security personnel to add necessary textual explanations. The text annotation input box supports up to 500 Chinese characters, and the input content is stored in association with the question status. The interface also links text annotation questions to an image upload interface; when additional on-site photos are needed for explanation, the image acquisition module's shooting function can be directly invoked.

[0058] The manual entry interface allows security personnel to manually select potential hazards that were not automatically detected by the system. When security personnel discover hazards that the system failed to identify during on-site inspections, they can use the manual entry function to add the hazard information to the security inspection record. The status of manually entered items will also change to "abnormal" and be stored in association with the current work order.

[0059] The manual review module uses a question-and-answer interactive process to guide security personnel through the confirmation process. For Level 1 hazards, the system automatically highlights them with a red border and displays a clear prompt: "High-risk hazard detected, please review carefully." The marked image displays the hazard area in a highlighted manner, with the hazard area marked by a red rectangle and filled with diagonal lines, facilitating quick location and verification by security personnel.

[0060] Specifically, the work order submission module is responsible for integrating data from the entire security check process and generating structured security check data packages, which are then submitted to the security check business system. The technical implementation of this module includes three core functions: data fusion and encapsulation, risk scoring calculation, and rectification parameter matching.

[0061] The data fusion and encapsulation process integrates questionnaire responses, identification and confirmation results, photographic evidence, signature information, and rectification parameters. Photographic evidence includes the storage paths of the original and annotated images, using file service addresses within a distributed storage architecture. Signature information includes electronic signature image data of both the user's and security personnel's signatures, stored in Base64 encoded format.

[0062] The structured security inspection data packet is organized in JSON format and includes the following fields: workOrderId, customerInfo, operatorInfo, questionerResult, aiRecognitionResult, manualReviewResult, imageEvidence, signatureData, riskScore, riskLevel, rectificationMethod, rectificationDeadline, and submitTime. After the data packet is generated, it is protected for integrity using digital signature technology. The signature algorithm is SHA256 with RSA, and the signature data is appended to the end of the data packet. The above thirteen fields from workOrderId to submitTime are business fields. After the digital signature is completed, four additional signature verification fields are added: signatureAlgorithm, keyId, signatureValue, and payloadHash.

[0063] During packet signing, the system normalizes and serializes thirteen business fields (excluding the four signature verification fields: signatureAlgorithm, keyId, signatureValue, and payloadHash) according to a preset field order. These fields are uniformly encoded in UTF-8 and a SHA-256 digest is calculated. The digest is then signed using the RSA private key corresponding to keyId. Finally, the signatureAlgorithm, keyId, signatureValue, and payloadHash fields are appended. The receiving end reconstructs the byte stream to be verified using the same normalization rules. It first compares the payloadHash and then verifies the signatureValue using the public key. If any verification fails, the data is rejected and an audit event is recorded.

[0064] In one specific implementation, the risk scoring calculation unit performs a fusion calculation based on a multi-factor weighted summation algorithm, outputting an objective and unified risk score value and the corresponding hazard level determination result. The calculation formula of this model is as follows:

[0065] In the formula: This represents the final output risk score value (within the range of values). ); This represents the total number of items in a single security check record that ultimately fall into an abnormal state. Indicates the first The hazard level weights for each abnormal item are set as follows: Level 1 hazard is set to 1.0, Level 2 hazard to 0.7, and Level 3 hazard to 0.4. Indicates the first The confidence coefficient of the abnormal item. If the abnormal item has an AI recognition result, then... The value is the basic recognition confidence score output by the AI ​​module. (scope If the question item is manually entered by security personnel during on-site inspection, then it will be forcibly assigned... The principle behind this is that manual data entry represents a higher degree of certainty than AI probability inference, and assigning it full confidence ensures that high-risk risks discovered manually are not diluted or down-weighted in the calculation. Indicates the first The manual confirmation factor for each abnormal item. A value of 1.0 is used when the issue is confirmed as a potential hazard or manually entered, indicating that manual confirmation enhances the reliability of the hazard; a value of 0.8 is used when the system pre-selection is performed without manual review. Indicates the first Historical rule factor for abnormal items. If the user has rectification records of the same type of hidden danger within the past year, the value is 1.2, indicating that the hidden danger is a recurring problem that requires special attention; if there is no historical record, the value is 1.0. This is the scoring scaling factor (percentage conversion factor), preferably configured to be 50. This value is set to ensure that a reasonable combination of hazards (such as two independently identified Level 1 hazards) can normally overflow into the 100-point range.

[0066] Preferably, to prevent a single serious hazard that directly endangers personal safety from being diluted by other parameters in the formula weighting, this model incorporates a top-level mandatory coverage business rule: when all [certain] hazard is in an abnormal state... Among the items, there are Level 1 hidden dangers that have been manually confirmed or manually added (i.e., simultaneously meeting the following conditions). and If the result of the conventional summation calculation of the above formula is S<90, the system will skip the conventional cumulative calculation and directly force a risk score of S=90 (or force a score of 100 when it is determined to be a serious leak).

[0067] The risk scoring process is as follows: First, calculate the weighted sum and multiply it by K, then compare it with 100 and take the smaller value; next, determine the mandatory coverage condition for Level 1 hazards. If this condition is met and the current score is less than 90, increase it to 90; if the severe leakage rule is met, increase it to 100; finally, output the integer score according to the rounding rule and match the risk level accordingly. All participants in the calculation... , , , The K, rule version, and reason for coverage are all written into the scoring details for easy review and auditing.

[0068] The overall hazard level is divided into three levels based on the scoring range: Level 1 is 90 points and above, indicating a major safety risk that requires immediate action; Level 2 is 70 points to 89 points, indicating a significant safety risk that requires action within a specified timeframe; Level 3 is 50 points to 69 points, indicating a general safety risk that requires a reminder for rectification; and below 50 points is considered normal or a minor issue.

[0069] The rectification parameter matching system automatically matches the corresponding rectification method and deadline based on the hazard level output by the risk score. Rectification method templates include five categories: immediate rectification, rectification within a specified timeframe, rectification with reminder, gas outage rectification, and no rectification required. Level 1 hazards are matched for immediate rectification, with a rectification deadline of the same day or the next day; for Level 1 hazards determined to be serious leaks, gas outage rectification is also matched. Level 2 hazards are matched for rectification within a specified timeframe, with a rectification deadline set to 3 to 7 days depending on the hazard type. Level 3 hazards are matched for rectification with reminder, with a rectification deadline set to 30 days. No rectification is matched if there is no hazard. When generating structured security inspection data packets, the work order submission module writes the matched rectification method and rectification deadline codes into the corresponding fields and pushes them to the security inspection business system through the business interface.

[0070] The technical effects of the present invention will be described in detail below with reference to specific implementation examples.

[0071] Case Study 1: Routine Gas Safety Inspection Scenario for Residential Users. A gas safety inspector conducts a routine gas safety inspection on a residential user in Chaoyang District, Beijing. Work order number WO-2024-001234. The user type is residential, and the inspection scenario is routine. The questionnaire loading module generates a structured hazard questionnaire using a hierarchical matching algorithm, including six major hazard categories: pipelines, stoves, alarms, meters, hoses, and gas usage environment, totaling 45 hazard items. The known parameters of the user's on-site equipment are: stove approximately 2 years old (corresponding to equipment age adjustment factor). The service life of the connecting hose is approximately 4 years (adjusted for the corresponding equipment lifespan). ).

[0072] Safety inspectors collected six images from the site according to the image collection list generated by the system. After image collection, the AI ​​recognition module performed hazard detection on the images. A hazard of missing flameout protection device was detected in the stove image, and the hazard label "Stove without flameout protection" (Level 2 hazard) was output, with a basic recognition confidence level of [not specified]. For this typical category of questions, the system's dynamic judgment threshold is as follows. The image of the connecting hose detected potential hose aging, and the hazard label "Hose Aging" (Level 2 Hazard) was output. The basic identification confidence level was... For this typical category of questions, combined with the parameters of the old hose, the system's dynamic judgment threshold is determined as follows. If no abnormal potential hazards reaching the threshold are detected in the alarm image (indicating that no visible abnormality is currently seen in the alarm), this result does not trigger abnormality pre-selection.

[0073] The question mapping module performs automatic pre-selection based on the mapping rules between hazard labels and questionnaire questions. The two hazards mentioned above... All are greater than the dynamic judgment threshold of the corresponding scenario. Therefore, the system automatically switched these two question items to an abnormal status. The manual review module displays the automatically pre-selected results and corresponding labeled images. Security personnel verify the accuracy and click the confirmation button to complete the review. ), and none of them have historical records ( ).

[0074] The work order submission module calculates risk scores based on the hazard results. Both hazards are classified as Level 2 hazards. The calculation process is as follows:

[0075]

[0076]

[0077] The system rounded the score to 62, which fell to... The overall hazard level for the selected area is assessed as Level 3. The system's chosen rectification method is a reminder, with a 30-day rectification period. The final generated structured security inspection data package includes the corresponding work order information, a risk score of 62, and an overall hazard level of Level 3, and is submitted to the security inspection business system via the business interface.

[0078] A comparison was made with traditional manual security checks. In this method, the same security personnel inspect the same user, relying solely on manual inspection and visual judgment. Due to differences in experience and judgment, the personnel failed to identify potential hazards such as missing flameout protection devices on stoves and aging hoses, simply selecting "normal" on the questionnaire. Comparative data shows that traditional manual security checks, which rely heavily on on-site observation and questionnaire completion by security personnel, are prone to overlooking subtle hazards due to differences in experience, missed observations, or incorrectly filled-out questions. With this invention, the AI ​​model can automatically associate image recognition results with corresponding questionnaire items, which are then verified by security personnel using the labeled images. This reduces omissions and errors, improving the consistency and traceability of security check results.

[0079] Case Study 2: Comprehensive Safety Inspection Scenario for Industrial and Commercial Users. An inspector conducted a comprehensive gas safety inspection on a catering business, work order number WO-2024-005678. The user type was industrial and commercial. Due to the comprehensive nature of the inspection, the questionnaire loading module generated 68 hazard items, including all hazard groups and specific hazard groups. The inspector collected twelve images from the site according to the checklist.

[0080] The AI ​​recognition module detected potential hazards in twelve images. In the alarm image, it detected a "no power to alarm" hazard (Level 1 hazard) with a basic recognition confidence level of 0.92. In the gas appliance or water heater image, it detected a "no flameout protection on stove" hazard (Level 2 hazard) with a confidence level of 0.89. In the gas environment image, it detected a "damaged explosion-proof equipment component (damaged explosion-proof junction box)" hazard (Level 2 hazard) with a confidence level of 0.78. The item mapping module automatically pre-selected these three locations as anomalies according to mapping rules. Safety inspectors confirmed this during manual review and also discovered an additional hazard not covered by the AI: "improper installation of connecting hose" (Level 3 hazard, corresponding weight). Security personnel manually added the entry as abnormal using the manual data entry function. The confidence coefficient for this data entry is then... This work order is for an industrial or commercial user; the scenario adjustment coefficient is [not specified]. =1.0; The service life of the on-site alarms, gas appliances, and explosion-proof equipment is all at least one year and no more than three years, with an equipment service life adjustment factor. All values ​​are set to 1.0; this user has no rectification records for the same type of hidden danger in the past year, and all historical rule factors are [not specified]. All values ​​are set to 1.0; after the review is completed, the system summarizes the final hazard results.

[0081] The work order submission module calculates risk scores. This safety inspection record includes one Level 1 hazard (alarm not powered), two Level 2 hazard items (stove lacks flameout protection, explosion-proof equipment components damaged), and one Level 3 hazard (irregular hose installation), all of which were manually confirmed or supplemented. The standard factor summation calculation process is as follows:

[0082]

[0083]

[0084] Because the weighted summation result has reached the upper limit of the full score range, the system's overall risk score is determined to be 100 points, and the overall hazard level is classified as Level 1. The system's recommended rectification method is immediate rectification, with a rectification deadline of the following day.

[0085] Comparative experimental data show that the AI-assisted security inspection method of this invention achieves a hazard detection rate of 98% in industrial and commercial user scenarios, significantly higher than the 65% of traditional manual security inspections. The automatic mapping and association between AI recognition and questionnaire items reduces the conversion time from recognition results to business actions from an average of 3 minutes to less than 15 seconds. The unified hazard level and rectification parameters generated by risk scoring calculation ensure the comparability and standardization of inspection results from different security personnel.

[0086] The data comparison table is as follows:

[0087]

[0088] Case Study 3: Network-Constrained Scenario. A security inspector is conducting a security check in a gas equipment room in a basement. The network signal in this area is weak, making online AI recognition impossible. The system automatically switches to offline recognition mode, utilizing a lightweight model deployed on a mobile device for recognition processing. The mobile lightweight model performs local inference processing on the acquired images. After recognition, the system temporarily stores the results in a local database, automatically uploading them to the server for further processing once the network is restored. After network recovery, the item mapping module and the manual review module perform dynamic threshold comparison and manual review according to standard procedures, ultimately generating a structured security inspection data packet. This embodiment verifies the usability of the invention under network-constrained conditions, ensuring the continuity of security inspection operations in various complex environments.

[0089] When uploading offline results, an idempotent key is constructed using workOrderId, imageUuid, and modelVersion. The server retains only one valid record for duplicate idempotent keys. If the mobile client's rule version differs from the server's current version, the server retains the original recognition results and re-executes the item mapping according to the latest rules, recording the original version, the new version, and the recalculated result. After successful upload, the server returns a confirmation flag. Upon receiving the confirmation flag, the mobile client marks its local record as synchronized, thus avoiding duplicate work orders or evidence loss due to retransmissions in weak network conditions.

[0090] Specifically, this invention employs multiple technical means to ensure the security and integrity of security inspection data. Image data is transmitted using a TLS encrypted channel, and stored using a distributed storage architecture with a multi-replica backup strategy. Three copies are stored, located in the main data center, disaster recovery center, and edge nodes. The structured security inspection data packets employ an anti-tampering mechanism; a hash value is calculated and a digital signature is attached when the data packet is generated, and any change to any field after submission will cause the hash verification to fail. All operations of the manual review module are logged, supporting post-audit queries. User signatures and security personnel signatures are stored as electronic signature images along with metadata such as the signer and signing time, and are bound to the work order content through the aforementioned structured data packet digital signature mechanism; the receiving end uses an asymmetric cryptographic algorithm to verify that the data packets have not been tampered with.

[0091] Specifically, this invention possesses excellent scalability and adaptability. The hazard label and issue rule base adopts a configurable data structure, supporting the addition, modification, and deletion of mapping rules through the backend management interface. The AI ​​recognition model employs a modular design; when new hazard categories emerge, recognition capabilities can be expanded through incremental model training, and training samples are configured according to category complexity when adding new hazard categories. The risk scoring calculation model supports custom configuration of weight parameters, with adjustments taking effect immediately. This invention also supports multilingual and localized deployment; prompts utilize an internationalized string resource management mechanism, and the system supports Simplified Chinese, Traditional Chinese, and English versions. The work order submission module provides standardized business interfaces, supporting both batch submission and single-item submission modes. The batch submission interface can process up to 100 work orders at a time, while the single-item submission interface is used for real-time processing scenarios.

[0092] In summary, this invention achieves deep integration of AI recognition capabilities with the security inspection process by dynamically mapping AI image recognition results to a structured hazard questionnaire using threshold correlation, and by combining risk scoring weighted calculation with manual review mechanisms. This generates a complete structured security inspection data package containing hazard results, image evidence, rectification parameters, and signature information. Examples and comparative experimental data demonstrate that this invention significantly improves the hazard detection rate in gas security inspections, reduces subjective differences in manual inspections, enhances the uniformity and traceability of inspection results, and effectively solves several technical problems existing in current gas security inspection technologies.

[0093] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0094] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. 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. An AI-based intelligent auxiliary method for gas safety inspection, characterized in that: Includes the following steps: Step S1: Obtain a structured hazard questionnaire adapted to the current security inspection work order. The structured hazard questionnaire contains multiple hazard items with different risk levels. Step S2: Obtain images of the security inspection target objects collected on-site; Step S3: Input the image into the pre-trained target detection model to detect potential hazards, and output hazard labels and basic recognition confidence scores; Step S4: Based on the prior physical environment parameters of the on-site equipment related to the security inspection work order, dynamically adjust and generate the corresponding dynamic judgment threshold. Step S5: In the pre-set hazard label and item rule base, determine the target hazard item that is mapped to the hazard label. When the basic identification confidence is greater than or equal to the dynamic judgment threshold, automatically switch the status of the target hazard item in the structured hazard questionnaire to the abnormal pre-selection status. Step S6: Receive manual review instructions for the abnormal pre-selected status to confirm the potential issues that are ultimately in an abnormal status; Step S7: Based on the hazard items that are ultimately in an abnormal state and the operation type of the manual review instruction, calculate the overall risk score value, and generate a structured security inspection data packet based on the risk score value.

2. The AI-based intelligent auxiliary method for gas safety inspection according to claim 1, characterized in that: The steps to obtain a structured hazard questionnaire adapted to the current security inspection work order include: A hierarchical rule matching algorithm is adopted. First, a basic question set is established based on the user type basic classification rules. Then, the mandatory questions in the industry security inspection mandatory inspection rule base are incorporated into the basic question set and deduplicated according to the question identifier. Finally, non-mandatory questions are filtered according to the security inspection scenario adaptation rules. When there are conflicts among the rules, the industry security inspection mandatory inspection rule base has the highest priority, followed by the user type basic classification rules, and the security inspection scenario adaptation rules have the lowest priority. A structured hazard questionnaire containing hazard category grouping identifiers, question type configurations, and question item metadata lists is generated.

3. The AI-based intelligent auxiliary method for gas safety inspection according to claim 1, characterized in that: The steps for outputting hazard labels and basic identification confidence levels specifically include: The target detection model outputs multiple candidate boxes and corresponding classification score vectors for each hazard category. The non-maximum suppression algorithm is used to remove overlapping and redundant boxes, and the coordinates of the effective region are retained as the location boxes of potential hazards. Normalize the classification score vector corresponding to the effective region coordinates to obtain the probability distribution of each category, and take the maximum probability as the basic recognition confidence. The hazard category with the highest probability is matched with the built-in standard gas hazard label library to generate hazard labels; if the probability difference between the top two hazard categories is less than a preset threshold, a set of multiple candidate labels is output.

4. The AI-based intelligent auxiliary method for gas safety inspection according to claim 1, characterized in that: Based on the prior physical environment parameters of the on-site equipment related to the security inspection work order, the steps of dynamically adjusting and generating corresponding dynamic judgment thresholds specifically include: Determine the baseline threshold that matches the risk level of the hazard label; Obtain the scenario adjustment coefficient set based on the user type in the security check work order; Obtain the equipment age adjustment coefficient set based on the service life of the on-site equipment. The longer the service life of the equipment, the smaller the corresponding equipment age adjustment coefficient. The dynamic judgment threshold for the current work order and issue is calculated by multiplying the baseline threshold, the scenario adjustment coefficient, and the equipment age adjustment coefficient.

5. The AI-based intelligent auxiliary method for gas safety inspection according to claim 1, characterized in that: The process of determining the target hazard items that correspond to the hazard labels also includes: The execution scenario keyword matching rules are used for secondary verification. User type and equipment scenario in the security inspection work order are extracted as keywords. These keywords are used as filtering conditions to exclude mapping results that are irrelevant to the hazard label and question rule base.

6. The AI-based intelligent auxiliary method for gas safety inspection according to claim 1, characterized in that: Manual review instructions include confirmation, rejection, and manual data entry. The steps for receiving manual review instructions for abnormal pre-selected states specifically include: If a rejection operation is received, the status of the corresponding potential hazard item will be restored to normal and it will not be included in the calculation set of subsequent risk score values. If a manual supplementary entry is received, the system will respond to the security personnel's manual selection of potential hazards not identified by the target detection model, and directly change the status of the corresponding hazard to abnormal.

7. The AI-based intelligent auxiliary method for gas safety inspection according to claim 6, characterized in that: The steps for calculating the overall risk score are implemented using a multi-factor weighted summation algorithm, specifically including: For each hazard item that is ultimately in an abnormal state, obtain its corresponding hazard level weight, confidence coefficient, manual confirmation factor, and historical rule factor, and multiply the above four factors to obtain the individual risk product of each hazard item; The individual risk products of all the potential problems in an abnormal state are summed to obtain the conventional summation result; The result of the conventional summation calculation is multiplied by a preset scoring scaling factor, and the result of the multiplication is compared with the set upper limit of the full score range. The minimum value between the two is taken as the final output risk score value.

8. The AI-based intelligent auxiliary method for gas safety inspection according to claim 7, characterized in that: The assignment of confidence coefficients specifically includes: If a potential hazard item has a matching basic identification confidence level, then the original value of the basic identification confidence level is used as the confidence coefficient of that potential hazard item. If the abnormal state of a potential problem item is triggered by a manual data entry operation, then the confidence coefficient of that potential problem item will be forcibly assigned to the preset maximum full confidence value.

9. The AI-based intelligent auxiliary method for gas safety inspection according to claim 7, characterized in that: The AI-based intelligent assistance method for gas safety inspections incorporates the highest-level mandatory coverage business rules; the calculation of the overall risk score also includes: When there are serious hidden danger items with a level 1 hazard among all the hidden danger items that are ultimately in an abnormal state, and which have been confirmed by confirmation operation or manual supplementation operation, determine whether the score after multiplying the result of the regular summation calculation by the scoring scaling factor is less than the preset level 1 risk lower limit. If so, skip the usual calculation step of taking the minimum value and directly force the risk score value to be the lower limit of the first-level risk; when it is determined to be a serious leak, directly force the risk score value to be the upper limit of the full score range.

10. The AI-based intelligent auxiliary method for gas safety inspection according to claim 1, characterized in that: The steps for generating structured security inspection data packets based on risk score values ​​specifically include: Based on the preset scoring range in which the risk score value falls, the overall hazard level corresponding to the current security check record is determined; Based on the overall level of hazard, the corresponding rectification method and rectification period parameters are automatically matched; The final status of the structured hazard questionnaire, the results of hazard identification and manual confirmation, image evidence, signature information, and the matching rectification methods and rectification deadlines are integrated into a single data package in JSON format to generate a structured security inspection data package. The generated structured security data packets are hashed using a digital signature algorithm and then digitally signed to achieve tamper-proof protection.

Images