Intelligent detection method and system for preventing abnormal operation behavior of seal point leakage detection

Abstract

The present application relates to the technical field of industrial seal point leakage detection and supervision, in particular to an intelligent detection method and system for preventing abnormal operation behavior of seal point leakage detection. The method comprises: while carrying out physical leakage detection operation on a target seal point, obtaining an on-site operation optical image reflecting the spatial relationship between the detection instrument and the seal point through a sampling terminal; the system calls a standard filing image representing the physical structure features and the surrounding texture, and performs consistency comparison based on the feature structure with the on-site image to analyze the matching condition of the instrument operation position and the target point; based on the similarity index generated by the comparison, it is determined whether the detection actually occurs at the specified point; if it is determined to be true, the concentration data and the similarity are associated and stored; if it is determined to be abnormal, the validity of the detection data is marked as doubtful. The present application ensures that each piece of leakage concentration data strictly corresponds to the real physical seal point, thereby ensuring the authenticity and reliability of the detection data from the source.

Description

Technical Field

[0001] This invention relates to the field of industrial sealing point leakage detection and monitoring technology, specifically to an intelligent detection method and system for preventing abnormal operation behavior in sealing point leakage detection. Background Technology

[0002] With the continuous improvement of industrial safety and environmental protection management standards, the scale of leak detection and repair projects is expanding, and the number of sealing points to be tested is increasing significantly. This large-scale testing operation has brought huge challenges to quality supervision and data authenticity verification.

[0003] Currently, the supervision of leak detection at sealing points mainly relies on satellite positioning technology or manual spot checks. Inspectors use handheld testing instruments to physically inspect each sealing point and record gas concentration values. However, traditional supervision methods mainly rely on the subjective awareness of operators and coarse-grained location data. Although these methods can record the detection trajectory, they are difficult to effectively identify and prevent abnormal operations such as the detection probe not contacting the sealing point, detecting air, or mislabeling the object being tested at the micro-operation level. In addition, relying on manual spot checks has the defects of low coverage and poor timeliness, making it difficult to verify the authenticity of a large number of on-site detection activities in real time, resulting in blind spots in supervision.

[0004] Therefore, how to use intelligent means to verify the accuracy of the physical location of leak detection operations in real time, so as to effectively prevent abnormal detection operations, has become a technical problem that urgently needs to be solved in this field.

[0005] The information disclosed in the background section above is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention discloses an intelligent detection method and system for preventing abnormal operation behavior in sealing point leak detection. Specifically, the technical solution of this invention is: an intelligent detection method for preventing abnormal operation behavior in sealing point leak detection, comprising: simultaneously performing physical leak detection on a target sealing point, acquiring on-site optical images containing the spatial relationship between the detection instrument and the tested sealing point through a sampling terminal; recalling a pre-stored standard archival image of the target sealing point, wherein the standard archival image characterizes the physical structural features of the target sealing point and the texture of the surrounding environment; performing a consistency comparison between the on-site optical image and the standard archival image based on feature structure to analyze whether the spatial position of the detection instrument during operation matches the physical position of the target sealing point; determining whether the current leak detection operation actually occurred at the specified physical sealing point based on a similarity index generated by the consistency comparison; if the determination result is true, associating and storing the concentration data collected during the leak detection operation with the similarity index; if the determination result is abnormal, marking the validity of the leak detection data as questionable.

[0007] Preferably, the step of acquiring on-site operational optical images containing the spatial relationship between the detection instrument and the tested sealing point includes: synchronously recording a continuous on-site detection video stream during the sampling period when the leak detection probe is activated to draw in a gas sample; parsing a time-series frame sequence containing the contact action of the detection probe from the on-site detection video stream; performing environmental noise filtering and illumination normalization processing on the time-series frame sequence to eliminate the interference of changes in on-site light on the identification of the physical texture of the sealing point surface, and generating a set of test operation frames for consistency verification.

[0008] Preferably, the step of performing a consistency comparison between the on-site operational optical image and the standard archival image based on feature structure includes: using a deep learning model containing a visual feature extraction layer to extract static structural feature vectors from the standard archival image and dynamic operational feature vectors from the on-site operational optical image; the deep learning model is configured to identify the physical contour features of industrial components such as flanges, valves, and welds at sealing points; by calculating the projection distance between the static structural feature vector and the dynamic operational feature vector in the feature space, the degree of overlap between the two in physical texture and geometric structure is quantified, and the similarity index is output.

[0009] Preferably, the step of quantifying the degree of overlap between the two in terms of physical texture and geometric structure specifically includes: setting the static structural feature vector as the reference physical fingerprint and the dynamic operation feature vector as the fingerprint of the environment to be tested; using a cosine similarity algorithm to calculate the cosine value of the angle between the reference physical fingerprint and the fingerprint of the environment to be tested, so as to characterize the physical consistency between the detection background and the archived background; performing statistical analysis on the cosine values ​​corresponding to multiple samplings, and extracting the maximum matching value as the final similarity value to confirm the validity of this detection operation.

[0010] Preferably, the step of determining whether the current leak detection operation actually occurred at the specified physical sealing point based on the similarity index generated by the consistency comparison is configured as follows: retrieving the validity threshold range established based on the historical normal detection records of the target sealing point; comparing the final similarity value with the validity threshold range; when the final similarity value falls within the validity threshold range, it is determined that the detection probe is in the correct sampling position, and the generated anti-abnormal operation behavior judgment report concludes that the detection environment is consistent; when the final similarity value is lower than the lower limit of the validity threshold, it is determined that the detection probe is not in the predetermined sampling position or there is a substitute shooting behavior, and the generated anti-abnormal operation behavior judgment report concludes that the detection environment is abnormal.

[0011] Preferably, the method further includes the step of generating a comprehensive test report: obtaining sensor readings from the leak detection instrument, including gas concentration values, ambient temperature, and detection timestamps; fusing the sensor readings with the abnormal operation behavior judgment report; extracting the key frame with the highest clarity from the on-site operation optical image as a location verification image; and generating a comprehensive test report containing physical detection data, location verification images, and authenticity judgment conclusions to prove the compliance of the sealing point leak analysis process.

[0012] Preferred methods also include: receiving verification requests from the management platform via an industrial data interface; outputting the comprehensive test report in a structured data format to prove to the management platform that the gas concentration value originates from a real physical sealing point; and recording test operation records deemed abnormal into a risk database for subsequent targeted review of specific test paths or operators.

[0013] An intelligent detection system for preventing abnormal operation behavior during leak detection at sealing points includes: a sampling and monitoring module configured to simultaneously acquire a live video stream reflecting the contact state between the probe and the sealing point while the leak detection probe collects gas data; an environmental consistency analysis module configured to use a feature extraction algorithm to compare the physical texture of the equipment in the live video stream with pre-stored archived images and calculate the spatial location matching degree; a detection validity verification module configured to verify the authenticity of the gas concentration value source based on the spatial location matching degree and mark the authenticity of the detection results; and a data traceability module configured to establish an associated archive containing leak concentration data and on-site optical evidence for traceability of detection quality.

[0014] Compared with the prior art, the present invention has the following beneficial effects:

[0015] 1. This invention solves the problems of abnormal operations such as void detection and substitution detection in traditional detection by simultaneously acquiring on-site images containing the spatial relationship between the instrument and the sealing point and comparing them with standard archived images based on feature structure consistency. The system does not rely solely on GPS or QR codes, but deeply analyzes the physical structural features and environmental texture of the sealing point. Only when the spatial location during detection highly matches the archived physical location is the data considered valid. This mechanism ensures that every leakage concentration data strictly corresponds to the actual physical sealing point, guaranteeing the authenticity and reliability of the detection data from the source.

[0016] 2. This invention employs a deep learning model that includes a visual feature extraction layer, specifically designed to identify the physical contours of industrial components such as flanges, valves, and welds. By calculating the projection distance and cosine of the included angle between static structural features and dynamic operational features in the feature space, the degree of overlap of physical textures is quantified. Combined with environmental noise filtering and illumination normalization processing, this method can effectively eliminate the interference of changes in ambient light and differences in shooting angles on the identification. This means that even in complex industrial sites with strong outdoor light or insufficient light, the system can accurately lock the physical fingerprint of the target equipment, ensuring the accuracy of consistency judgment.

[0017] 3. This invention achieves microsecond-level correlation between gas concentration values ​​and operational behavior images by synchronously recording video streams during the sampling period and analyzing time-series frames containing probe contact actions. The system performs deep data fusion with sensor readings, abnormal operation behavior judgment reports, and high-definition location verification images to generate a comprehensive detection report. This design breaks the previous disconnect between detection data and on-site conditions, providing an unalterable digital evidence package for each detection, improving the compliance of the sealing point leakage analysis process, and providing a solid basis for subsequent quality audits.

[0018] 4. This invention can not only determine the effectiveness of a single operation in real time, but also link with the management platform through the industrial data interface to automatically add abnormal detection records to the risk database. This mechanism allows the management focus to shift from massive amounts of routine data to specific high-risk detection paths or operators. Through targeted review of questionable data and statistical analysis of abnormal behavior, enterprises can promptly identify loopholes and human negligence in the detection process, achieving refined and intelligent management of LDAR projects and reducing safety and environmental risks. Attached Figure Description

[0019] The present invention will be further explained below with reference to the accompanying drawings and embodiments:

[0020] Figure 1 This is a flowchart of the method of the present invention.

[0021] Figure 2 This is a system structure diagram of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0023] Example 1:

[0024] Please see Figure 1 Intelligent detection methods and systems for preventing leaks at sealing points and detecting abnormal operational behaviors include:

[0025] While performing physical leak detection on the target sealing point, the sampling terminal acquires on-site optical images showing the spatial relationship between the detection instrument and the tested sealing point. Pre-stored standard archival images of the target sealing point are then retrieved; these images characterize the physical structure of the target sealing point and the texture of its surrounding environment. A consistency comparison based on feature structure is performed between the on-site optical images and the standard archival images to analyze whether the spatial position of the detection instrument during operation matches the physical position of the target sealing point. Based on the similarity index generated from the consistency comparison, it is determined whether the current leak detection operation actually occurred at the specified physical sealing point. If the determination result is positive, the concentration data collected during the leak detection operation and the similarity index are associated and stored. If the determination result is negative, the validity of the leak detection data is marked as questionable.

[0026] This embodiment provides a method for preventing abnormal operation behavior during sealing point leakage detection. This method relies on computer vision and deep learning technologies to automatically verify the authenticity of the detection operation by comparing the consistency of features between the archived data and the on-site operation video. The method specifically includes the following steps: Step S1: Data acquisition. The system acquires archived images of the sealing points. These images are high-quality, standardized equipment photos taken in advance during the equipment ledger management stage, serving as the true value benchmark for subsequent comparison. Simultaneously, the system acquires on-site sealing point detection video generated by a sampling terminal with image acquisition capabilities at the work site. In this embodiment, the sampling terminal refers to a handheld LDAR detector or a matching visualization auxiliary device integrating an explosion-proof camera and an intelligent processing unit. When the operator performs the detection, the terminal automatically records the entire detection process, ensuring synchronization between the detection behavior and video recording.

[0027] Step S2: Video frame extraction and preprocessing. The system extracts video frames and preprocesses images from the on-site inspection video of the sealing point to construct a set of frames to be detected. Due to the large amount of video data and the existence of redundancy, directly processing the entire video is inefficient. This step aims to extract the most representative key frame images from the continuous video stream and remove environmental noise interference to provide high-quality input data for subsequent feature extraction.

[0028] Step S3: Feature Extraction. The system uses a pre-defined multimodal neural network model to extract image features from the archival images and the set of frames to be detected, generating archival image feature vectors and detection frame feature vectors. In this step, the image is no longer treated as a pixel matrix, but is mapped to a high-dimensional semantic space; the archival image is mapped to an archival image feature vector. Each frame in the set of frames to be detected is mapped to a series of detection frame feature vectors. This characteristic indicates rotational invariance and illumination robustness, which can overcome interference caused by on-site shooting angle deviations and lighting changes.

[0029] Step S4: Similarity calculation. The system uses the cosine similarity algorithm to calculate the spatial distance between the feature vector of the archived image and the feature vector of the detection frame, and calculates the detection similarity value. The cosine similarity measures the consistency of the direction by calculating the cosine value of the angle between the two vectors. The closer the value is to 1, the more similar the two are.

[0030] Step S5: Authenticity determination. The system compares the detected similarity value with the preset authenticity determination threshold, determines the authenticity category of the detected operation based on the comparison result, and generates an anti-abnormal operation behavior determination report. This step discretizes the continuous similarity value into a binary determination result, realizing automated compliance review.

[0031] Step S6: Historical data update. The system stores the abnormal operation behavior judgment report and the corresponding detection similarity value into the historical record database to update the system's historical detection data. This is not only used for archiving this task, but also provides data support for subsequent optimization of judgment thresholds and audit trails.

[0032] Through the above methods, this invention constructs a closed loop for preventing abnormal work behavior throughout the entire process from data collection to intelligent judgment. Compared with traditional methods such as manual sampling or relying solely on GPS positioning, this method directly performs deep comparison based on visual content, which can effectively identify abnormal work behavior such as mislabeling or false detection. By utilizing the strong generalization ability of deep learning models, the efficiency and accuracy of supervision are greatly improved, and the falsification phenomenon in manual inspection is eliminated.

[0033] The steps for acquiring on-site operational optical images containing the spatial relationship between the detection instrument and the tested sealing point include: synchronously recording a continuous on-site detection video stream during the sampling period when the leak detection probe draws in a gas sample; parsing the time-series frame sequence containing the contact action of the detection probe from the on-site detection video stream; performing environmental noise filtering and illumination normalization on the time-series frame sequence to eliminate the interference of changes in on-site lighting on the identification of the physical texture of the sealing point surface, and generating a set of test operation frames for consistency verification.

[0034] This embodiment provides a detailed explanation of the method for constructing the set of frames to be detected in step S2 above. To ensure the data quality of the input model, the process includes the following sub-steps: Original frame sequence acquisition: The system analyzes the on-site inspection video of the sealing point frame by frame in chronological order to obtain the original frame sequence. This step adopts a dual strategy of timed sampling + quality filtering: Candidate frames are extracted at preset time intervals, the Laplacian gradient value of the candidate frames is calculated, images with gradient values ​​lower than a preset blur threshold are removed, and the top N frames with the highest gradient values ​​among the remaining images are used as the set of frames to be detected to balance computational load and information integrity.

[0035] Noise Removal Processing: The system performs noise removal processing on each frame of the original frame sequence to obtain denoised frame data. In this embodiment, Gaussian filtering or median filtering algorithms are used to smooth the image to remove high-frequency noise and retain edge features, in order to remove dust and water mist interference that may exist in the chemical site.

[0036] Size Adjustment and Color Normalization: The system adjusts the denoised frame data to a preset standard input size and performs color normalization to generate normalized frame data. The preset standard input size strictly corresponds to the input layer requirements of the subsequent AI model. Color normalization usually involves mapping the pixel values ​​of the RGB channels from [0,255] to the range of [0,1] or [-1,1], and subtracting the mean of the dataset and dividing by the standard deviation. The dataset here specifically refers to the ImageNet dataset used to pre-train this multimodal neural network model.

[0037] The specific normalization calculation formula is as follows: ;in, Represents the color channel index of the image; This represents the pixel value of the original image in channel c; This represents the pixel value after normalization. and These represent the mean and standard deviation of the pre-training dataset on channel c, respectively. In this embodiment, the parameter values ​​correspond to the ImageNet dataset statistics, i.e., the corresponding mean vector. =[0.485,0.456,0.406], standard deviation vector =[0.229,0.224,0.225]; This means that when c=R, =0.485, =0.229.

[0038] Set Construction: The system labels all normalized frame data as frames to be detected, and the set of frames to be detected is composed of all frames to be detected. The preset standard input size is consistent with the dimensional parameters required by the input layer of the multimodal neural network model. This constraint is a necessary condition to ensure that tensor operations can be performed smoothly. Through a standardized preprocessing pipeline, this embodiment eliminates the impact of inconsistent video source quality on the recognition results. In particular, denoising and normalization processing significantly improve the signal-to-noise ratio of feature extraction. The uniform adjustment of the size enables the system to adapt to acquisition terminals with different resolutions, enhancing the hardware compatibility of the system.

[0039] The steps for performing a consistency comparison between on-site operational optical images and standard archival images based on feature structure include: using a deep learning model containing a visual feature extraction layer to extract static structural feature vectors from the standard archival images and dynamic operational feature vectors from the on-site operational optical images; configuring the deep learning model to identify the physical contour features of industrial components such as flanges, valves, and welds at sealing points; and quantifying the degree of overlap between the two in terms of physical texture and geometric structure by calculating the projection distance between the static structural feature vectors and the dynamic operational feature vectors in the feature space, and outputting a similarity index.

[0040] This embodiment provides a detailed explanation of the method for generating feature vectors in step S3, focusing on the application mechanism of the AI ​​model; Model loading and selection: The system loads a multimodal neural network model, which includes at least one of the CLIP model and the DINOv3 model; The CLIP model is a model trained on large-scale image-text pairs, which has strong zero-shot transfer capability and is suitable for handling semantic matching; The DINOv3 model is a model based on visual self-supervised learning, which can capture extremely subtle local texture features.

[0041] In response to a model selection command or an automatic configuration strategy, the system determines the current execution model from the CLIP model and the DINOv3 model; specifically, the system reads the average luminance value of the frame to be detected. With Laplace variance As an environmental assessment indicator; if the following Boolean logic is satisfied: ;in, To preset the brightness threshold, This is a preset texture complexity threshold.

[0042] The formulas for calculating the parameters upon which the above logical judgments rely are as follows:

[0043]

[0044] ;

[0045] in, This represents the average brightness value of the image; Representing coordinates The grayscale pixel value at the location, where H and W are the height and width of the image, respectively; Specifically, it refers to the variance value of the image after processing by the Laplacian operator, used to quantify texture sharpness; N is the total number of pixels in the image. Representing an image The response value of the i-th pixel after convolution with the Laplacian kernel. For all response values The arithmetic mean of the values. This metric is used to quantify the sharpness of an image; conversely, if... If the scene is deemed complex, it will automatically switch to DINOv3.

[0046] Among them, the preset brightness threshold The method for obtaining the data is as follows: Collect a set of unqualified identification samples under low-light conditions, calculate their average grayscale value, and take the upper limit of this value as the threshold; preset the texture complexity threshold. The method for obtaining the data is as follows: calculate the Laplacian variance of a set of samples containing high-density textures or complex backgrounds, and take the lower limit of its variance distribution as the threshold; for example, CLIP is selected first in scenes with sufficient lighting and simple backgrounds; and DINOv3 is automatically switched in scenes with dense devices and high texture similarity.

[0047] Feature extraction of archival images: The system inputs the archival image of the sealing point into the current execution model, extracts the corresponding multidimensional feature data, and performs vector normalization on the multidimensional feature data to generate the archival image feature vector.

[0048] Detection frame feature extraction: The system sequentially inputs each frame image in the set of frames to be detected into the current execution model, extracts the corresponding multi-dimensional feature data, and performs vector normalization on the multi-dimensional feature data to generate the detection frame feature vector.

[0049] Vector normalization is defined as follows: Vector normalization is configured to map the magnitude of an eigenvector to a unit length; specifically, for any eigenvector v, its normalized vector... .

[0050] This embodiment introduces the most advanced large-scale visual model as the core feature extractor. Compared with traditional handcrafted features such as SIFT / SURF, deep learning-based feature vectors can understand the deep semantics of images and have extremely high robustness to viewpoint changes, partial occlusion, and illumination changes. The strategy of supporting multiple model switching allows the system to flexibly adjust the strategy according to actual working conditions, balancing detection speed and accuracy. Vector normalization processing lays the mathematical foundation for subsequent fast calculation based on cosine similarity.

[0051] The steps for quantifying the degree of overlap between the two in terms of physical texture and geometric structure specifically include: setting the static structural feature vector as the baseline physical fingerprint and the dynamic operation feature vector as the fingerprint of the environment to be tested; using the cosine similarity algorithm to calculate the cosine value of the angle between the baseline physical fingerprint and the fingerprint of the environment to be tested, so as to characterize the physical consistency between the detection background and the archived background; performing statistical analysis on the cosine values ​​corresponding to multiple samplings, and extracting the maximum matching value as the final similarity value to confirm the validity of this detection operation.

[0052] This embodiment details the method for calculating the detection similarity value; the method is based on the vector space model in linear algebra.

[0053] Vector labeling: The system acquires the feature vector of the archived image and labels it as the baseline vector A; it acquires the feature vector of the detection frame and labels it as the target vector B.

[0054] Vector operations: The system calculates the dot product of the reference vector and the target vector, and obtains the vector dot product value (A). B); The system calculates the product of the magnitude of the reference vector and the magnitude of the target vector, and obtains the product value of the magnitudes ( The system divides the vector dot product by the product of their moduli to calculate the cosine similarity value; the specific calculation formula is shown below: ;in, and These represent the numerical components of the baseline vector and the target vector in the i-th feature dimension, respectively; i is the summation index variable, ranging from 1 to n; n is the total number of dimensions of the feature vectors output by the multimodal neural network model, for example, the CLIP model is usually configured with n=512 or n=768.

[0055] Since normalization has already been performed in the previous steps, that is In practice, the calculation can be simplified to a vector dot product, which reduces the computational complexity.

[0056] Final value determination: The system counts the cosine similarity values ​​corresponding to the feature vectors of all detected frames in the set of frames to be detected, and determines the largest value or the average value as the detection similarity value; in this embodiment, the strategy of the largest value is preferred; because as long as there is only one frame in the video that can highly match the archived image, it can be proven that the operator has indeed reached the point and taken a picture.

[0057] Cosine similarity is used instead of Euclidean distance because in high-dimensional feature space, the directional difference of vectors reflects the similarity of semantic content better than the distance difference. Through the aggregation strategy, this method tolerates possible brief jitter, blur or misalignment moments in the video. As long as the video segment contains a valid aligned frame, it can be judged as passing, which not only ensures the rigor of the judgment, but also conforms to the actual situation of on-site operation.

[0058] The steps for determining whether the current leak detection operation actually occurred at the specified physical sealing point based on the similarity index generated by consistency comparison are configured as follows: retrieve the validity threshold range established based on the historical normal detection records of the target sealing point; compare the final similarity value with the validity threshold range; when the final similarity value falls within the validity threshold range, it is determined that the detection probe is in the correct sampling position, and the generated anti-abnormal operation behavior judgment report concludes that the detection environment is consistent; when the final similarity value is lower than the lower limit of the validity threshold, it is determined that the detection probe is not in the predetermined sampling position or there is a substitute shooting behavior, and the generated anti-abnormal operation behavior judgment report concludes that the detection environment is abnormal.

[0059] This embodiment details a method for determining the authenticity category of a testing operation based on comparison results.

[0060] Threshold acquisition: The system acquires a preset authenticity judgment threshold, which is a value derived from the similarity distribution statistics of historical normal detection data; for example, the system analyzes the manual review detection data that has been confirmed as genuine in the past 1000 times, calculates its similarity distribution curve, and selects the lower bound value covering the 95% confidence interval as the threshold.

[0061] Normal judgment logic: If the detection similarity value is greater than or equal to the preset authenticity judgment threshold, the authenticity category of the detection operation is judged as normal detection, and the detection result is marked as real in the anti-abnormal operation behavior judgment report; this means that the target contained in the on-site video is highly consistent with the file target.

[0062] Anomaly detection logic: If the similarity value is less than the preset authenticity threshold, the authenticity of the detection operation is classified as a suspected abnormal operation, and the detection result is marked as abnormal in the anomaly detection report, while generating anomaly alarm data. This indicates that the on-site video failed to provide sufficient evidence to prove that it was filmed at the designated sealing point, and there may be abnormal operation behaviors such as air detection, detection of incorrect equipment, or use of substitute pictures. By setting a dynamic threshold based on statistical laws, this method avoids the subjectivity of manually setting thresholds. The binary classification judgment logic is simple and clear, and can directly trigger the anomaly alarm mechanism, enabling managers to pay attention to suspected problem points as soon as possible, realizing the transformation from post-event accountability to in-event intervention.

[0063] It also includes the steps of generating a comprehensive test report: obtaining sensor readings from the leak detection instrument, including gas concentration values, ambient temperature, and detection timestamps; fusing the sensor readings with the abnormal operation behavior judgment report; extracting the key frame with the highest clarity from the on-site operation optical image as a location verification image; and generating a comprehensive test report containing physical detection data, location verification images, and authenticity judgment conclusions to prove the compliance of the sealing point leak analysis process.

[0064] This embodiment details the method for generating a report on abnormal work behavior.

[0065] Metadata Acquisition: The system acquires the metadata of the detection operation, which includes at least the detection time, detection location, equipment number, and operator information; these data are sourced from the sensors of the sampling terminal and user login information.

[0066] Content Construction and Evidence Association: The system integrates metadata, detection similarity scores, and the determined authenticity category to construct the main content of the report. To enhance the persuasiveness of the report, the system extracts key frames from the sealing point filing images and on-site inspection videos of the sealing points as evidence attachments and associates them with the main content of the report. The key frame is usually the one with the highest similarity score.

[0067] Layout and rendering: The system lays out and renders the main content of the report and supporting evidence according to the preset report template format, generating an abnormal operation behavior judgment report; the report format can be PDF, HTML or image format, which is easy to read and archive; the report generated in this step is not just a simple result, but a complete chain of evidence; by integrating objective data, subjective data, spatiotemporal data and visual evidence, the report has extremely high legal effect and management value, and can be directly used for internal enterprise assessment or compliance review by environmental protection departments.

[0068] The method also includes: receiving verification requests from the management platform through an industrial data interface; outputting a comprehensive test report in a structured data format to prove to the management platform that the gas concentration value comes from the actual physical sealing point; recording the test operation records that are determined to be abnormal into the risk database for subsequent targeted review of specific test paths or operators; this embodiment involves the system's data interaction and sharing functions.

[0069] External call response: The system receives call requests from external industrial management systems through the RESTful API interface; in response to the call request, it returns the abnormal operation behavior judgment report to the external industrial management system in JSON format to achieve data sharing; RESTful API is a web application interface design style based on the HTTP protocol, and JSON is a lightweight data exchange format; this design allows the system to be easily integrated into the enterprise's ERP, LDAR management platform or environmental monitoring system.

[0070] Historical Query: In response to historical record query commands, the system retrieves and outputs historical detection records that meet the query conditions from the historical record database; it supports multi-dimensional queries by time period, by personnel, by device area, etc.; through standardized interface design, this invention breaks down information silos, enabling the detection results of abnormal operation behavior to flow to the core management of the enterprise in real time; the universality of JSON format ensures the system's high compatibility; the historical record retrieval function provides a convenient tool for long-term trend analysis and compliance auditing.

[0071] Example 2:

[0072] Please see Figure 2The system includes a sampling and monitoring module, configured to simultaneously acquire a live video stream reflecting the contact status between the probe and the sealing point while the leak detection probe collects gas data; an environmental consistency analysis module, configured to use a feature extraction algorithm to compare the live video stream with the physical texture of the equipment in pre-stored archived images and calculate the spatial location matching degree; a detection validity verification module, configured to verify the authenticity of the gas concentration value based on the spatial location matching degree and mark the authenticity of the detection results; and a data traceability module, configured to establish a correlation archive containing leak concentration data and on-site optical evidence for traceability of detection quality.

[0073] This embodiment provides a system for detecting abnormal operating behaviors to prevent leaks at sealing points. The system implements the aforementioned method in hardware logic. The system mainly includes the following modules: a data acquisition module: configured to acquire images of sealing points and receive on-site inspection videos of sealing points uploaded by sampling terminals with image acquisition capabilities; this module is responsible for mapping from the physical world to the digital world; a preprocessing module: configured to extract and standardize frames from the on-site inspection videos of sealing points, constructing a set of frames to be detected; this module is responsible for cleaning the data; a feature extraction module: configured to integrate a multimodal neural network model to extract features from the images of the sealing points and the set of frames to be detected, generating feature vectors for the archived images and the detection frames; this module is the brain of the system, responsible for understanding the image content; a similarity calculation module: configured to calculate the cosine similarity between the feature vectors of the archived images and the feature vectors of the detection frames, outputting the detection similarity value; this module is responsible for quantifying the difference; and a judgment and reporting module: configured to compare the detection similarity value with a preset authenticity judgment threshold, determine the authenticity of the detection, generate an anti-abnormal operating behavior judgment report, and update the historical record database; this module is responsible for decision output.

[0074] The system adopts a modular design, with each functional unit decoupled, which facilitates independent maintenance and upgrades. For example, when a more advanced AI model emerges, only the feature extraction module needs to be upgraded, without having to reconstruct the entire system. This architectural design ensures the long-term vitality and scalability of the system, enabling it to adapt to the ever-changing needs of industrial sites.

[0075] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. An intelligent detection method for preventing leaks at sealing points and detecting abnormal operational behaviors, characterized in that, include: While performing physical leak detection on the target sealing point, the sampling terminal acquires on-site optical images containing the spatial relationship between the detection instrument and the sealing point being tested; The pre-stored standard archival image of the target sealing point is invoked, and the standard archival image represents the physical structural features of the target sealing point and the texture of the surrounding environment. A consistency comparison based on feature structure is performed between the on-site optical image and the standard archival image to analyze whether the spatial position of the detection instrument during operation matches the physical position of the target sealing point. Based on the similarity index generated by the consistency comparison, it is determined whether the current leak detection operation actually occurred at the specified physical sealing point. If the determination result is true, the concentration data collected during the leak detection operation is associated with and stored with the similarity index; if the determination result is abnormal, the validity of the leak detection data is marked as questionable.

2. The intelligent detection method for preventing leaks at sealing points and detecting abnormal operating behavior according to claim 1, characterized in that, The steps for acquiring on-site optical images containing the spatial relationship between the testing instrument and the sealed point under test include: During the sampling period when the leak detection probe draws in a gas sample, a continuous on-site detection video stream is recorded simultaneously. The time frame sequence containing the contact action of the detection probe is parsed from the on-site detection video stream; The time-series frame sequence is subjected to environmental noise filtering and illumination normalization to eliminate the interference of on-site light changes on the identification of physical texture on the sealing point surface, thereby generating a set of test operation frames for consistency verification.

3. The intelligent detection method for preventing leaks at sealing points and detecting abnormal operational behavior as described in claim 2, characterized in that, The steps for performing a consistency comparison between the on-site operational optical image and the standard archival image based on feature structure include: Using a deep learning model that includes a visual feature extraction layer, static structural feature vectors in the standard archival images and dynamic operation feature vectors in the on-site operation optical images are extracted respectively. The deep learning model is configured to identify the physical contour features of industrial components such as flanges, valves, and welds at sealing points; By calculating the projection distance between the static structural feature vector and the dynamic job feature vector in the feature space, the degree of overlap between the two in physical texture and geometric structure is quantified, and the similarity index is output.

4. The intelligent detection method for preventing leaks at sealing points and detecting abnormal operational behavior as described in claim 3, characterized in that, The step of quantifying the degree of overlap between the two in terms of physical texture and geometric structure specifically includes: The static structural feature vector is set as the baseline physical fingerprint, and the dynamic operation feature vector is set as the fingerprint of the environment to be tested. The cosine similarity algorithm is used to calculate the cosine value of the angle between the reference physical fingerprint and the fingerprint of the environment to be tested, so as to characterize the physical consistency between the detected background and the archived background. Statistical analysis was performed on the cosine values ​​corresponding to multiple samples, and the maximum matching value was extracted as the final similarity value to confirm the validity of this detection operation.

5. The intelligent detection method for preventing leaks at sealing points and detecting abnormal operating behavior according to claim 4, characterized in that, The step of determining whether the current leak detection operation actually occurred at the specified physical sealing point based on the similarity index generated by the consistency comparison is configured as follows: Retrieve the validity threshold range established based on the historical normal detection records of the target sealing point; The final similarity value is compared with the validity threshold range; When the final similarity value falls within the validity threshold range, it is determined that the detection probe is in the correct sampling position, and the generated anti-abnormal operation behavior judgment report concludes that the detection environment is consistent. When the final similarity value is lower than the lower limit of the validity threshold, it is determined that the detection probe is not in the predetermined sampling position or there is a substitute shooting behavior, and the generated anti-abnormal operation behavior judgment report concludes that the detection environment is abnormal.

6. The intelligent detection method for preventing leaks at sealing points and detecting abnormal operational behavior according to claim 5, characterized in that, It also includes the step of generating a comprehensive test report: Acquire sensor readings from the leak detection instrument, including gas concentration, ambient temperature, and detection timestamp; The sensor readings are fused with the abnormal operation behavior determination report; The keyframe with the highest clarity from the on-site optical image is selected as the location verification image; Generate a comprehensive inspection report that includes physical inspection data, location verification images, and authenticity determination conclusions to prove the compliance of the sealing point leakage analysis process.

7. The intelligent detection method for preventing leaks at sealing points and detecting abnormal operational behavior as described in claim 6, characterized in that, The method further includes: Receive verification requests from the management platform via the industrial data interface; The comprehensive test report is output in a structured data format to prove to the management platform that the gas concentration value originates from the actual physical sealing point; Records of inspection operations deemed abnormal will be entered into the risk database for subsequent targeted review of specific inspection paths or operators.

8. An intelligent detection system for preventing abnormal operation behavior during sealing point leakage detection, used to execute the intelligent detection method for preventing abnormal operation behavior during sealing point leakage detection as described in any one of claims 1-7, characterized in that, include: The sampling and monitoring module is configured to simultaneously acquire a live video stream reflecting the contact status between the probe and the sealing point while the leak detection probe acquires gas data. The environmental consistency analysis module is configured to use a feature extraction algorithm to compare the physical texture of the equipment in the on-site video stream with the pre-stored archived images and calculate the spatial location matching degree. The detection validity verification module is configured to verify the authenticity of the gas concentration value source based on the spatial location matching degree, and to mark the detection results as true or false. The data traceability module is configured to create a linked archive containing leakage concentration data and on-site optical evidence for the purpose of tracing the quality of the test.

Images