Safety production hidden danger intelligent diagnosis system and method

Abstract

The invention discloses a safety production hidden danger intelligent diagnosis system and method, and belongs to the technical field of safety production management, and the method specifically comprises the steps: collecting and preprocessing a field image through a sensor, and inputting the image into a multi-target recognition model to recognize each target element in the image; determining a production scene containing all the target elements by using a preset production scene knowledge graph, calling a corresponding specification file, and outputting specification requirements through the large model; dividing the image into a large-scale information area and a small-scale information area according to specification requirements, splitting the field image into a large-scale sub-image and a small-scale sub-image according to the large-scale information area and the small-scale information area, and respectively inputting the sub-images into a first deep learning model and a second deep learning model of a corresponding production scene for hidden danger identification; identifying and recording potential safety hazards by combining output results of the first deep learning model and the second deep learning model; according to the invention, the accuracy and efficiency of production field potential safety hazard identification are improved.

Description

Technical Field

[0001] The invention belongs to the technical field of production safety management, and in particular to an intelligent diagnosis system and method for production safety hidden dangers. Background Art

[0002] In the field of production safety, timely and accurate identification of hidden dangers is crucial to preventing accidents. Traditional production safety management relies on manual inspections, which have problems such as low efficiency, easy omissions, and slow response. With the development of image recognition and artificial intelligence technology, it has become possible to use image recognition technology for intelligent identification of production safety hazards. However, existing technologies are often limited to single target recognition or simple scene analysis, and it is difficult to comprehensively and accurately identify safety hazards in complex production environments.

[0003] For example, the Chinese patent application with the publication number CN109345148A discloses an intelligent diagnostic method for identifying safety hazards of natural gas pipelines and station process equipment, including: data collection: collecting data of natural gas pipelines and station equipment, and establishing a ledger; establishing a database: inputting data into the server, optimizing the selection, establishing a database, and performing data statistics and cloud computing analysis; identifying hidden dangers: using phased array detection technology and injection and production string vibration fatigue life prediction technology, targeting intrinsic safety issues, checking and detecting the integrity of key components of equipment, pipelines, pipe fittings and valve bodies, online intelligent remote monitoring technology and on-site surveys, collecting hidden danger data, and performing test calculations. This technical solution conducts preliminary work to understand the risk sources and digital construction status of each gas transmission pipeline, conducts technology selection, formulates implementation plans, and conducts project risk analysis to ensure the safe operation of natural gas production pipelines.

[0004] The above existing technologies all have the following problems: weak generalization ability and adaptability; poor real-time and efficiency; high technical complexity and specialization; they are optimized for specific application scenarios of natural gas pipelines and site process equipment, but may be difficult to adapt to other types of safety hazard identification tasks. Summary of the invention

[0005] In view of the shortcomings of the prior art, the present invention proposes an intelligent diagnosis system and method for production safety hazards, which collects on-site images through sensors and preprocesses them, and inputs the images into a multi-target recognition model to identify each target element in the image; uses a preset production scene knowledge graph to determine the production scene containing all target elements, and calls the corresponding specification file to output the specification requirements through a large model; according to the specification requirements, the image is divided into a large-scale information area and a small-scale information area, and the on-site image is split into large-scale sub-images and small-scale sub-images accordingly, and these sub-images are respectively input into a first deep learning model and a second deep learning model of the corresponding production scene for hazard identification; combines the output results of the first deep learning model and the second deep learning model to identify and record safety hazards; the present invention improves the accuracy and efficiency of identifying safety hazards at production sites.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] An intelligent diagnosis method for production safety hazards, comprising:

[0008] Acquire a scene image of a sensor, wherein the scene image includes a plurality of target elements;

[0009] Determine a production scenario including all target elements according to the target elements and a preset production scenario knowledge graph;

[0010] Dividing the on-site image into a large-scale information area and a small-scale information area according to the production scene and at least one specification file corresponding to the production scene;

[0011] According to the large-scale information area and the small-scale information area, safety production hazards are identified on the on-site image.

[0012] Specifically, the process of determining the production scenario including all target elements includes:

[0013] The on-site image is input into the multi-target recognition model, and the multi-target recognition model outputs each target element in the image;

[0014] All identified target elements are input into the preset production scenario knowledge graph, and the production scenario including all target elements is determined based on the association relationships in the production scenario knowledge graph.

[0015] Specifically, the process of dividing the large-scale information region and the small-scale information region includes:

[0016] According to the production scenario, at least one specification file corresponding to the production scenario is called, and the big model is called to perform specification semantic output on all specification files to obtain multiple specification requirements;

[0017] Based on the specification requirements, the image is divided into a large-scale information area and a small-scale information area.

[0018] Specifically, the process of identifying production safety hazards on the on-site image according to the large-scale information area and the small-scale information area includes:

[0019] According to the divided large-scale information area and small-scale information area, the preprocessed scene image is split into at least one large-scale sub-image and at least one small-scale sub-image; the large-scale sub-image may overlap with the small-scale sub-image area;

[0020] Inputting the large-scale sub-image and the large-scale information into a first deep learning model corresponding to the production scene, and inputting the small-scale sub-image into a second deep learning model corresponding to the production scene;

[0021] Combine the output results of the first deep learning model and the second deep learning model to identify safety hazards at the production site, and record the results of each hazard identification.

[0022] Specifically, the first deep learning model and the second deep learning model include:

[0023] The first deep learning model is established based on edge recognition. The first deep learning model first performs edge recognition on the large-scale sub-image to obtain boundary contour information, and determines whether the large-scale information is satisfied by comparing the similarity between the boundary contour information and the large-scale information;

[0024] The second deep learning model is used for material identification and realizes the accurate identification and analysis of subtle features in small-scale sub-images.

[0025] Specifically, when the large-scale sub-image is used for boundary contour information recognition, the small-scale sub-image is further separated from the original large-scale sub-image, which is a dynamic verification process. The specific steps of the dynamic verification include:

[0026] Obtain a large-scale sub-image, and calculate the texture complexity and grayscale change rate index for each local area of ​​the large-scale sub-image. When the texture complexity or grayscale change rate index of any local area exceeds a preset threshold, it is considered that there are hidden dangers in the details of the local area that need further detailed analysis;

[0027] According to the local area with hidden detail problems determined, the corresponding small-scale sub-image is separated from the large-scale sub-image;

[0028] According to the detailed hidden danger features in different production scenarios, image area selection rules associated with the detailed hidden danger features are established.

[0029] Specifically, the specific steps of the dynamic verification also include:

[0030] After the small-scale sub-image is stripped from the large-scale sub-image, a real-time feedback mechanism is established between the first deep learning model and the second deep learning model;

[0031] Performing a preliminary analysis on the stripped area using the first deep learning model, and passing the location, size, and preliminary analysis results of the stripped area to the second deep learning model;

[0032] After receiving the information transmitted by the first deep learning model, the second deep learning model performs a detailed analysis of the small-scale sub-image to detect and identify specific detailed hidden danger information, where the detailed hidden danger information includes the specific location and shape of the hidden danger, and its correlation with the large-scale information.

[0033] Specifically, the specific steps of the dynamic verification also include:

[0034] Feeding back the detected detailed hidden danger information to the first deep learning model;

[0035] During the multiple image stripping and model analysis processes, the verification standards and model parameters are adjusted based on the comparison between each verification result and the actual hidden danger situation;

[0036] The final analyzed hidden danger information, location, and severity will be output to the staff, and the process and results of each analysis will be recorded.

[0037] Specifically, the specific process of stripping the corresponding small-scale sub-image from the large-scale sub-image includes:

[0038] Determine a local area with hidden detail risks in the large-scale sub-image, where the local area with hidden detail risks is a bounding box formed by a plurality of pixel coordinates;

[0039] According to the determined local area of ​​detail hidden dangers, the parameters required for stripping out the small-scale sub-image are calculated, including the starting coordinates, width and height of the small-scale sub-image;

[0040] Using the calculated peeling parameters, peeling the corresponding small-scale sub-image from the large-scale sub-image through image cropping operation;

[0041] Check whether the stripped small-scale sub-image contains the complete potential risk area and does not contain unnecessary additional areas.

[0042] Specifically, when combining the output results of the first deep learning model and the second deep learning model, a weighted average method is used to assign different weights to the first deep learning model and the second deep learning model according to their accuracy. and .

[0043] An intelligent diagnosis system for production safety hazards, comprising: an image processing module, a production scene determination module, a sub-image generation module, a deep learning model analysis module, and a hazard identification and recording module;

[0044] The image processing module collects on-site images through sensors and performs preprocessing;

[0045] The production scene determination module is used to use a multi-target recognition model to identify each target element from the preprocessed image, and determine the current production scene based on the identified target elements in combination with a preset production scene knowledge graph;

[0046] The sub-image generation module is used to divide the image into regions according to the determined specification requirements, and split the pre-processed scene image into large-scale sub-images and small-scale sub-images according to the divided large-scale information regions and small-scale information regions;

[0047] The deep learning model analysis module is used to input the large-scale sub-image and large-scale information into a first deep learning model corresponding to the production scene, and input the small-scale sub-image into a second deep learning model corresponding to the production scene, to perform a preliminary analysis of safety hazards;

[0048] The hidden danger identification and recording module is used to combine the output results of the two deep learning models, identify safety hazards at the production site, and record the results of each hidden danger identification.

[0049] Specifically, the sub-image generation module includes: a specification file calling unit, a specification semantic output unit, a region division unit, and a sub-image cutting unit;

[0050] The specification file calling unit is used to call the corresponding specification file according to the determined production scenario;

[0051] The specification semantic output unit is used to perform semantic analysis on the specification document through the large model and output multiple specification requirements;

[0052] The region division unit is used to divide the image into a large-scale information region and a small-scale information region based on specification requirements;

[0053] The sub-image cropping unit is used to split the pre-processed scene image into a large-scale sub-image and a small-scale sub-image according to the divided large-scale information area and small-scale information area.

[0054] Compared with the prior art, the present invention has the following beneficial effects:

[0055] 1. The present invention proposes an intelligent diagnosis system for production safety hazards, and optimizes and improves the architecture, operation steps and processes. The system has the advantages of simple process, low investment and operation costs, and low production work costs.

[0056] 2. The present invention proposes an intelligent diagnosis method for production safety hazards. By integrating multi-target recognition, knowledge graph application and deep learning model analysis, it realizes the intelligent and automated identification of safety hazards in production sites, improves recognition efficiency and accuracy, reduces manual intervention, and improves the level of production safety management. At the same time, the method shows high efficiency and flexibility in detail processing. By dividing large-scale information areas and small-scale information areas, and applying targeted deep learning models for analysis respectively, it not only improves the accuracy of hazard identification, but also can effectively cope with complex and changeable production scenarios, ensuring that safety hazards are identified and recorded without omissions. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 A schematic diagram of an intelligent diagnosis method for production safety hazards of the present invention;

[0058] Figure 2 This is a principle flow chart of an intelligent diagnosis method for production safety hazards of the present invention;

[0059] Figure 3 A schematic diagram of a lamp room of an intelligent diagnosis method for production safety hazards of the present invention;

[0060] Figure 4 A schematic diagram of a general control room of an intelligent diagnosis method for production safety hazards of the present invention;

[0061] Figure 5 A schematic diagram of a 10KV capacitor room ventilation window grille of an intelligent diagnosis method for production safety hazards of the present invention;

[0062] Figure 6 This is an architecture diagram of an intelligent diagnosis system for production safety hazards of the present invention. DETAILED DESCRIPTION

[0063] Example 1

[0064] See also Figures 1 to 5 , an embodiment of the present invention provides: a method for intelligent diagnosis of production safety hazards, the method comprising steps S101 to S104, wherein:

[0065] S101: Acquire a scene image of a sensor, where the scene image contains a plurality of target elements;

[0066] Furthermore, the specific steps of S101 include:

[0067] (1) Obtaining the scene image collected by the sensor (such as CMOS image sensor). The pixel array on the CMOS sensor is composed of many micro photodiodes. These photodiodes generate charges when exposed to light, completing the photoelectric conversion process. The generated charges are accumulated in the storage area of ​​each pixel and integrated;

[0068] (2) After the integration is completed, the charge signal of each pixel is read through the row and column scanning registers. The analog charge signals read out are amplified one by one and then converted into digital signals through the analog-to-digital converter;

[0069] (3) The digital signal is then sent to the back-end digital signal processor for processing, including noise removal, white balance adjustment, and color interpolation;

[0070] (4) Fusion of images with different resolutions or spectra to improve image quality and information content, and obtain preprocessed on-site images.

[0071] S102: The target elements and the preset production scenario knowledge graph determine the production scenarios including all the target elements;

[0072] The process of determining a production scenario that includes all target elements includes:

[0073] A1: Input the on-site image into the multi-target recognition model, and the multi-target recognition model outputs each target element in the image, where the target elements include but are not limited to equipment, personnel, and materials;

[0074] A2: Input all identified target elements into the preset production scenario knowledge graph, and determine the production scenario including all target elements based on the association relationships in the production scenario knowledge graph.

[0075] Among them, the multi-target recognition model refers to a deep learning model that can recognize multiple target objects in an image, which needs to be loaded in advance. The deep learning model is the prior art content in this field and is not the creative solution of this application, so it will not be elaborated here.

[0076] Furthermore, the specific steps of A2 include:

[0077] (1) Extract all target elements identified in the image from the output of the multi-target recognition model. The specific target elements depend on the production scenario and the design of the recognition model.

[0078] (2) Match the extracted target elements with the preset production scenario knowledge graph, where the knowledge graph is a graph structure database containing entities, attributes, and relationships. Entities represent objects in the production scenario, and relationships represent the interactions or connections between these objects.

[0079] The specific process of matching the extracted target elements with the preset production scenario knowledge graph includes:

[0080] a. Ensure that the preset production scenario knowledge graph is complete and accurate, including all relevant entities, attributes, relationships, and the hierarchical and associated information between them. The knowledge graph usually represents knowledge in the form of triples: entity-relationship-entity;

[0081] b. Using natural language processing methods to extract target elements from the data source to be matched, these elements may be instances of entities, attributes or relationships, wherein the natural language processing method is the prior art content in this field, is not the inventive solution of this application, and is not described in detail here;

[0082] c. Select a graph-based matching algorithm based on the characteristics of the target element and the knowledge graph;

[0083] d. Compare the extracted target element with the entities, attributes or relationships in the knowledge graph one by one, and calculate the similarity between the target element and each element in the knowledge graph according to the selected graph-based matching algorithm, wherein the similarity calculation formula is the prior art content in this field and is not the inventive solution of this application, and is not described in detail here;

[0084] e. According to the preset threshold, select the element with the highest matching degree as the matching result, and verify the matching result to ensure the accuracy and reliability of the matching result;

[0085] f. Adjust and optimize the matching algorithm based on the verification results to improve matching performance and accuracy.

[0086] (3) In the knowledge graph, based on the associations between target elements, such as spatial location, functional relationship, process flow, etc., analyze and determine the production scenarios to which they belong. This involves traversing and searching the nodes and edges in the knowledge graph to find the production scenarios associated with the target elements.

[0087] (4) Based on the results of the association analysis, determine the production scenario that includes all target elements. This step involves evaluating and comparing multiple candidate scenarios to select the scenario that best matches the association relationship of the target elements.

[0088] S103: Dividing the on-site image into a large-scale information area and a small-scale information area according to the production scene and at least one specification file corresponding to the production scene;

[0089] It should be noted that in production scenarios, different information has different requirements for decision-making and monitoring. Large-scale information may focus more on the overall structure and macro-environment, while small-scale information may focus more on detail features and minor changes. By dividing the image into information areas of different scales, key information can be captured and analyzed more accurately, and the precision and accuracy of the analysis can be improved. Different processing methods and algorithms can be used for information areas of different scales, thereby optimizing the entire processing process and improving efficiency and effectiveness. On the other hand, by dividing information areas of different scales, changes and potential risks can be captured more keenly.

[0090] The division process of large-scale information area and small-scale information area includes:

[0091] B1: According to the production scenario, call at least one specification file corresponding to the production scenario, and call the big model to perform specification semantic output on all specification files to obtain multiple specification requirements;

[0092] Among them, specification documents refer to enterprise specifications or operation manuals; the big model adopts a natural language processing model and uses the output function of the natural language processing model to convert the text content in the specification documents into structured specification requirements, which include operating procedures, quality standards, and safety specifications.

[0093] For example, let the production scenario be S, the specification document library be F, and the large model be M;

[0094] For a given production scenario S, find the corresponding specification file set in the specification file library F ,in, Indicates the nth specification file;

[0095] For each specification file , call the large model M for semantic analysis and obtain the set of specification requirements ,in, represents the mth specification requirement of the i-th specification file;

[0096] Merge the specification requirements of all specification documents to obtain the final specification requirements set ,in, Indicates the Nth specification requirement in the final specification requirement set;

[0097] The final set of specification requirements is collated and verified to obtain the final usable specification requirements.

[0098] B2: Based on the specification requirements, the image is divided into large-scale information areas and small-scale information areas.

[0099] S104: Identify production safety hazards on the on-site image according to the large-scale information area and the small-scale information area.

[0100] The process of identifying production safety hazards on the on-site image according to the large-scale information area and the small-scale information area includes:

[0101] C1: splitting the preprocessed scene image into at least one large-scale sub-image and at least one small-scale sub-image according to the divided large-scale information area and small-scale information area; the large-scale sub-image may overlap with the small-scale sub-image area, that is, the relationship between the large-scale sub-image and the small-scale sub-image area includes: intersection, inclusion and irrelevance;

[0102] In the present invention, the scene image is split into large-scale sub-images and small-scale sub-images in order to better analyze and utilize the multi-scale information in the image. Moreover, since the information in the image often does not exist in isolation, but is interrelated and influences each other, the information connection between different scales in the image can be better captured and retained through the overlapping area. Among them, the large-scale sub-image usually contains the overall structure and trend information, while the small-scale sub-image contains more details and texture information.

[0103] C2: Input the large-scale sub-image and large-scale information into the first deep learning model corresponding to the production scene, and input the small-scale sub-image into the second deep learning model corresponding to the production scene;

[0104] Furthermore, the specific steps of C2 include:

[0105] (1) Arranging the preprocessed large-scale sub-images and large-scale information into a data format suitable for input into a deep learning model. Similarly, arranging the small-scale sub-images into a data format suitable for input into a second deep learning model.

[0106] (2) Loading a pre-built first deep learning model and a pre-built second deep learning model according to the requirements of the production scenario and the image features, wherein the first deep learning model is used to process large-scale sub-images and large-scale information, and the second deep learning model is used to process small-scale sub-images;

[0107] Among them, the first deep learning model and the second deep learning model are constructed based on the convolutional neural network model architecture, and the convolutional neural network model is the prior art content in this field, and is not the creative solution of this application, and will not be elaborated here.

[0108] (3) Using the training data set to train the first deep learning model and the second deep learning model respectively, so that they can accurately identify and process large-scale sub-images, large-scale information and small-scale sub-images, wherein the training set needs to collect historical large-scale sub-images, large-scale information and small-scale sub-images in advance to obtain training set data;

[0109] (4) Deploy the trained and qualified first deep learning model and the second deep learning model into the production environment to process large-scale sub-images, large-scale information, and small-scale sub-images;

[0110] (5) In a production environment, the newly acquired large-scale sub-image and large-scale information are input into the first deep learning model for prediction, and the newly acquired small-scale sub-image is input into the second deep learning model for prediction.

[0111] In the present invention, because the information in the image often has multi-scale characteristics, that is, information of different scales plays different roles in image understanding and analysis, by processing large-scale and small-scale information separately, the complementarity of this information can be fully utilized to improve the comprehensiveness and accuracy of processing. Therefore, the large-scale sub-image and large-scale information are input into the first deep learning model corresponding to the production scene, and the small-scale sub-image is input into the second deep learning model corresponding to the production scene. This is based on the complementarity of multi-scale information, model adaptability considerations, and rational use of computing resources. Considerations such as the rational use of computing resources can more effectively capture the details and global features in the image, thereby improving the precision and accuracy of processing.

[0112] C3: Combine the output results of the first deep learning model and the second deep learning model to identify safety hazards at the production site and record the results of each hazard identification.

[0113] When combining the output results of the first deep learning model and the second deep learning model, a weighted average method is used to assign different weights to the first deep learning model and the second deep learning model according to their accuracy. and .

[0114] The first deep learning model and the second deep learning model include:

[0115] The first deep learning model is established based on edge recognition. The first deep learning model first performs edge recognition on the large-scale sub-image to obtain boundary contour information, and determines whether the large-scale information is satisfied by comparing the similarity between the boundary contour information and the large-scale information. The edge recognition adopts an edge detection algorithm. The edge detection algorithm is a prior art content in this field and is not an inventive solution of the present application, and is not described in detail here.

[0116] In the present invention, the large-scale information includes the identified dosing room, the total reduction control room, and the window grille information, while the small-scale information includes the corrosion area information of the lamps and wires, curtains, and window grilles in the dosing room.

[0117] The second deep learning model is used for material identification and realizes the accurate identification and analysis of subtle features in small-scale sub-images.

[0118] For example, Figure 3 As shown, in Figure 3 In the example, after the first deep learning model recognizes the large-scale sub-image, it is determined that the large-scale sub-image is the dosing room. Then, after the large-scale sub-image is recognized, detail recognition is performed to obtain small-scale sub-images, including but not limited to light bulbs, wires, and load-bearing walls. Figure 3 A in the figure is the wiring diagram of the lamps in the dosing room, which is a small-scale sub-image. When identifying the wiring connection details of the lamps in the dosing room, it can focus on the connection points between the wires and the lamps, the slight damage to the insulation layer of the wires, the tightness of the screws, and other features that can only be clearly presented at a small scale; by learning from a large number of such small-scale hidden danger samples, the model can determine whether these details meet the small-scale information requirements in the safety regulations, such as whether the wire connections are tight and whether the insulation layer is intact.

[0119] like Figure 4 As shown in FIG, after being identified by the first deep learning model, the large-scale sub-image is determined to be the total control room, and then a small-scale analysis is performed in the total control room, as shown in FIG. Figure 4 B in the figure is a curtain. For details related to the combustion performance of the curtain, the second deep learning model can analyze its microscopic characteristics such as fiber structure and weaving method, and compare them with the small-scale information in the fire protection regulations, such as the combustion characteristic indicators of specific fiber materials, to determine whether there is a fire hazard.

[0120] like Figure 5 As shown in FIG, after being identified by the first deep learning model, the large-scale sub-image is determined to be a window grille. Then, a small-scale analysis is performed on the window grille to identify the detail information, such as Figure 5 C in the figure is the window grille corrosion area. For the corrosion of the ventilation window grille in the 10kV capacitor room, the model can identify details such as tiny corrosion pits and metal thinning areas on the surface of the window grille, and determine whether it meets the small-scale specification requirements for strength and protection performance, thereby effectively detecting hidden dangers that may lead to the entry of small animals or other safety issues.

[0121] When the large-scale sub-image is used for boundary contour information recognition, the small-scale sub-image is further separated from the original large-scale sub-image. This is a dynamic verification process. The specific steps of dynamic verification include:

[0122] Obtain a large-scale sub-image, and calculate the texture complexity and grayscale change rate index for each local area of ​​the large-scale sub-image. When the texture complexity or grayscale change rate index of any local area exceeds a preset threshold, it is considered that there are hidden dangers in the details of the local area that need further detailed analysis;

[0123] It is important to know that the calculation process of texture complexity and grayscale change rate indicators includes:

[0124] (1) Cut out or extract a large-scale sub-image from the preprocessed image;

[0125] (2) Divide the large-scale sub-image into multiple overlapping or non-overlapping local regions. The size and step size of the division can be adjusted according to actual needs;

[0126] (3) For each local area, the texture complexity is calculated using the gray level co-occurrence matrix method;

[0127] Among them, the gray-level co-occurrence matrix is ​​a matrix obtained by calculating the frequency of occurrence of pixel pairs at a specific distance and direction in the image. Then, multiple texture features such as contrast, energy, homogeneity, and correlation can be extracted from the matrix. These features can be used to measure the texture complexity of the image. Further, when extracting multiple texture features from the matrix, the existing image processing library scikit-image is used to calculate these features, and scikit-image is the prior art content in this field, not the creative solution of this application, and will not be repeated here.

[0128] (4) For each local area, calculate its grayscale change rate , to measure the speed of grayscale change, where Represents the horizontal gradient of the image gray value. Represents the gradient of the image gray value in the vertical direction.

[0129] When making judgments, it is important to understand that by comparing the texture complexity and grayscale change rate of each local area with a preset threshold, if any indicator exceeds the threshold, it is considered that the local area has detailed hidden dangers that require further detailed analysis.

[0130] For example, when analyzing the large-scale sub-image of a 10kV capacitor room, if the grayscale changes in the area near the ventilation window are drastic and the texture is complex, it may indicate corrosion or foreign matter adhesion on the window grille. In this case, the system automatically performs small-scale sub-image peeling in this area so that the second deep learning model can perform more detailed detection.

[0131] According to the local area with hidden detail problems determined, the corresponding small-scale sub-image is separated from the large-scale sub-image;

[0132] According to the detailed hidden danger features in different production scenarios, image area selection rules associated with the detailed hidden danger features are established;

[0133] Furthermore, the process of establishing an image region selection rule associated with detail hidden danger features includes:

[0134] (1) Analyze various production scenarios and understand the production process, equipment layout, and operating specifications;

[0135] (2) Identify and record the detailed characteristics of potential hazards that may occur in each scenario, such as equipment aging, abnormal sounds, vibrations, leakage, and improper operation;

[0136] (3) Based on the identified detailed hidden danger features, formulate corresponding image area selection criteria, which may include area size, shape, location, color, texture and other features;

[0137] (4) Based on the specific conditions of the production scenario and the image area selection criteria, establish detailed image area selection rules. The rules should clearly define which areas need to be selected for further analysis, as well as the specific conditions and methods for selecting these areas;

[0138] Among them, image region selection rules refer to a set of rules or standards that determine which image regions need to be selected or focused on according to specific goals or requirements during image processing and analysis. These rules are usually based on certain features or attributes in the image, such as color, texture, shape, size, etc., as well as the association between these features and specific goals or requirements.

[0139] (5) Apply image region selection rules in actual production environments and collect feedback data;

[0140] (6) Adjust and optimize the rules based on feedback data to ensure their accuracy and effectiveness.

[0141] For example, in the large-scale sub-image analysis of the dosing room, if abnormal shadows or blurred areas are detected around the lamp and are associated with the hidden danger characteristics of electrical faults, the system will prioritize the area for small-scale sub-image stripping, focusing on checking details such as lamp wiring and bulb status. This active image stripping based on hidden danger characteristics can improve the targetedness and efficiency of detection.

[0142] After the small-scale sub-image is stripped from the large-scale sub-image, a real-time feedback mechanism is established between the first deep learning model and the second deep learning model;

[0143] Performing a preliminary analysis on the stripped area using the first deep learning model, and passing the location, size, and preliminary analysis results of the stripped area to the second deep learning model;

[0144] After receiving the information transmitted by the first deep learning model, the second deep learning model performs a fine analysis on the small-scale sub-image to detect and identify specific detailed hidden danger information, where the detailed hidden danger information includes the specific location and shape of the hidden danger, and the correlation with the large-scale information;

[0145] Feeding back the detected detailed hidden danger information to the first deep learning model;

[0146] For example, in the large-scale sub-image analysis of the indoor curtains under total control, if the first deep learning model finds that the overall area of ​​the curtains is large but the color or texture of some local areas are abnormal, after stripping off the small-scale sub-images, the second deep learning model detects that there are loose fibers in the area and other details that may affect the combustion performance, and feeds this information back to the first deep learning model, so that the first deep learning model can comprehensively judge the overall safety status of the curtains and realize collaborative work and dynamic verification between models.

[0147] During the multiple image stripping and model analysis processes, the verification standards and model parameters are adjusted based on the comparison between each verification result and the actual hidden danger situation;

[0148] For example, after processing a series of images of electrical equipment rooms, if it is found that certain types of hidden dangers are missed or misjudged under specific conditions, the system automatically adjusts the relevant parameters to improve the accuracy of subsequent detections, so that the entire dynamic verification process can continuously adapt to different production scenarios and changes in hidden dangers.

[0149] The final analyzed hidden danger information, location, and severity will be output to the staff, and the process and results of each analysis will be recorded.

[0150] The specific process of stripping the corresponding small-scale sub-image from the large-scale sub-image includes:

[0151] Determine a local area with hidden detail risks in the large-scale sub-image, where the local area with hidden detail risks is a bounding box formed by a plurality of pixel coordinates;

[0152] According to the determined local area of ​​detail hidden dangers, the parameters required for stripping out the small-scale sub-image are calculated, including the starting coordinates, width and height of the small-scale sub-image;

[0153] Among them, when calculating the parameters required for stripping the small-scale sub-image, the calculation is performed by using the functions provided by the image processing library OpenCV. OpenCV is the prior art content in this field and is not the creative solution of this application, so it will not be described here.

[0154] Using the calculated peeling parameters, peeling the corresponding small-scale sub-image from the large-scale sub-image through image cropping operation;

[0155] Check whether the stripped small-scale sub-image contains the complete potential risk area and does not contain unnecessary additional areas.

[0156] Example 2

[0157] See also Figure 6 Another embodiment provided by the present invention is a smart diagnosis system for production safety hazards, comprising:

[0158] Image processing module, production scene determination module, sub-image generation module, deep learning model analysis module, hidden danger identification and recording module;

[0159] The image processing module collects on-site images through sensors and performs preprocessing, including denoising, enhancement, and scaling, to ensure that the image quality meets the requirements of subsequent processing;

[0160] The production scene determination module is used to use the multi-target recognition model to identify each target element from the preprocessed image, and determine the current production scene based on the identified target elements and the preset production scene knowledge graph;

[0161] A sub-image generation module is used to divide the image into regions according to the determined specification requirements, and split the pre-processed scene image into large-scale sub-images and small-scale sub-images according to the divided large-scale information regions and small-scale information regions;

[0162] A deep learning model analysis module is used to input large-scale sub-images and large-scale information into a first deep learning model corresponding to the production scene, and input small-scale sub-images into a second deep learning model corresponding to the production scene, to conduct a preliminary analysis of safety hazards;

[0163] The hidden danger identification and recording module is used to combine the output results of the two deep learning models to identify safety hazards at the production site and record the results of each hidden danger identification.

[0164] The image processing module includes: an image acquisition unit and a preprocessing unit;

[0165] An image acquisition unit, used for acquiring on-site images through a sensor;

[0166] The preprocessing unit is used to perform preprocessing operations such as denoising, enhancement, and scaling on the collected on-site images.

[0167] The production scenario determination module includes: a target recognition unit and a knowledge graph matching unit;

[0168] A target recognition unit, used to recognize each target element from the preprocessed image using a multi-target recognition model;

[0169] The knowledge graph matching unit is used to determine the current production scenario based on the identified target elements and the preset production scenario knowledge graph.

[0170] The sub-image generation module includes: a specification file calling unit, a specification semantic output unit, a region division unit, and a sub-image cutting unit;

[0171] A specification file calling unit, used to call a corresponding specification file according to a determined production scenario;

[0172] The specification semantic output unit is used to perform semantic analysis on the specification document through a large model and output multiple specification requirements;

[0173] A region division unit, used for dividing the image into a large-scale information region and a small-scale information region based on specification requirements;

[0174] The sub-image cropping unit is used to split the pre-processed scene image into a large-scale sub-image and a small-scale sub-image according to the divided large-scale information area and the small-scale information area.

[0175] The hidden danger identification and recording module includes: hidden danger identification unit and result recording unit;

[0176] The hidden danger identification unit is used to integrate the output results of the two deep learning models to identify safety hazards at the production site;

[0177] The result recording unit is used to record the results of each hidden danger identification for subsequent analysis and processing.

[0178] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation modes, which are merely illustrative rather than restrictive. Under the guidance of the present invention, ordinary technicians in the field may also change, modify, replace and modify the above-mentioned embodiments without departing from the purpose and scope of protection of the present invention, and all of these are within the protection of the present invention.

Claims

1. An intelligent diagnosis method for production safety hazards, characterized in that: include: Acquire a scene image of a sensor, wherein the scene image includes a plurality of target elements; Determine a production scenario including all target elements according to the target elements and a preset production scenario knowledge graph; Dividing the on-site image into a large-scale information area and a small-scale information area according to the production scene and at least one specification file corresponding to the production scene; According to the large-scale information area and the small-scale information area, safety production hazards are identified on the on-site image.

2. The intelligent diagnosis method for production safety hazards according to claim 1, characterized in that: The process of determining a production scenario including all target elements includes: The on-site image is input into the multi-target recognition model, and the multi-target recognition model outputs each target element in the image; All identified target elements are input into the preset production scenario knowledge graph, and the production scenario including all target elements is determined based on the association relationships in the production scenario knowledge graph.

3. The intelligent diagnosis method for production safety hazards according to claim 1, characterized in that: The process of dividing the large-scale information area and the small-scale information area includes: According to the production scenario, at least one specification file corresponding to the production scenario is called, and the big model is called to perform specification semantic output on all specification files to obtain multiple specification requirements; Based on the specification requirements, the image is divided into a large-scale information area and a small-scale information area.

4. The intelligent diagnosis method for production safety hazards according to claim 1, characterized in that: The process of identifying production safety hazards on the on-site image according to the large-scale information area and the small-scale information area includes: Splitting the preprocessed scene image into at least one large-scale sub-image and at least one small-scale sub-image according to the divided large-scale information area and small-scale information area; Inputting the large-scale sub-image and the large-scale information into a first deep learning model corresponding to the production scene, and inputting the small-scale sub-image into a second deep learning model corresponding to the production scene; Combine the output results of the first deep learning model and the second deep learning model to identify safety hazards at the production site, and record the results of each hazard identification.

5. The intelligent diagnosis method for production safety hazards according to claim 4, characterized in that: The first deep learning model and the second deep learning model include: The first deep learning model is established based on edge recognition. The first deep learning model first performs edge recognition on the large-scale sub-image to obtain boundary contour information, and determines whether the large-scale information is satisfied by comparing the similarity between the boundary contour information and the large-scale information; The second deep learning model is used for material identification and realizes the accurate identification and analysis of subtle features in small-scale sub-images.

6. The intelligent diagnosis method for production safety hazards according to claim 5, characterized in that: When the large-scale sub-image is used for boundary contour information recognition, the small-scale sub-image is further separated from the original large-scale sub-image, which is a dynamic verification process. The specific steps of the dynamic verification include: Obtain a large-scale sub-image, and calculate the texture complexity and grayscale change rate index for each local area of ​​the large-scale sub-image. When the texture complexity or grayscale change rate index of any local area exceeds a preset threshold, it is considered that there are hidden dangers in the details of the local area that need further detailed analysis; According to the local area with hidden detail problems determined, the corresponding small-scale sub-image is separated from the large-scale sub-image; According to the detailed hidden danger features in different production scenarios, image area selection rules associated with the detailed hidden danger features are established.

7. The intelligent diagnosis method for production safety hazards according to claim 6, characterized in that: The specific steps of the dynamic verification also include: After the small-scale sub-image is stripped from the large-scale sub-image, a real-time feedback mechanism is established between the first deep learning model and the second deep learning model; Performing a preliminary analysis on the stripped area using the first deep learning model, and passing the location, size, and preliminary analysis results of the stripped area to the second deep learning model; After receiving the information transmitted by the first deep learning model, the second deep learning model performs a detailed analysis of the small-scale sub-image to detect and identify specific detailed hidden danger information, where the detailed hidden danger information includes the specific location and shape of the hidden danger, and its correlation with the large-scale information.

8. The intelligent diagnosis method for production safety hazards according to claim 7, characterized in that: The specific steps of the dynamic verification also include: Feeding back the detected detailed hidden danger information to the first deep learning model; During the multiple image stripping and model analysis processes, the verification standards and model parameters are adjusted based on the comparison between each verification result and the actual hidden danger situation; The final analyzed hidden danger information, location, and severity will be output to the staff, and the process and results of each analysis will be recorded.

9. The intelligent diagnosis method for production safety hazards according to claim 6, characterized in that: The specific process of stripping the corresponding small-scale sub-image from the large-scale sub-image includes: Determine a local area with hidden detail risks in the large-scale sub-image, where the local area with hidden detail risks is a bounding box formed by a plurality of pixel coordinates; According to the determined local area of ​​detail hidden dangers, the parameters required for stripping out the small-scale sub-image are calculated, including the starting coordinates, width and height of the small-scale sub-image; Using the calculated peeling parameters, peeling the corresponding small-scale sub-image from the large-scale sub-image through image cropping operation; Check whether the stripped small-scale sub-image contains the complete potential risk area and does not contain unnecessary additional areas.

10. The intelligent diagnosis method for production safety hazards according to claim 4, characterized in that: When combining the output results of the first deep learning model and the second deep learning model, a weighted average method is used to assign different weights to the first deep learning model and the second deep learning model according to their accuracy. and .

11. An intelligent diagnosis system for production safety hazards, used to implement an intelligent diagnosis method for production safety hazards according to any one of claims 1 to 10, characterized in that: include: Image processing module, production scene determination module, sub-image generation module, deep learning model analysis module, hidden danger identification and recording module; The image processing module collects on-site images through sensors and performs preprocessing; The production scene determination module is used to use a multi-target recognition model to identify each target element from the preprocessed image, and determine the current production scene based on the identified target elements in combination with a preset production scene knowledge graph; The sub-image generation module is used to divide the image into regions according to the determined specification requirements, and split the pre-processed scene image into large-scale sub-images and small-scale sub-images according to the divided large-scale information regions and small-scale information regions; The deep learning model analysis module is used to input the large-scale sub-image and large-scale information into a first deep learning model corresponding to the production scene, and input the small-scale sub-image into a second deep learning model corresponding to the production scene, to perform a preliminary analysis of safety hazards; The hidden danger identification and recording module is used to combine the output results of the two deep learning models, identify safety hazards at the production site, and record the results of each hidden danger identification.

12. The intelligent diagnosis system for production safety hazards according to claim 11, characterized in that: The sub-image generation module includes: a specification file calling unit, a specification semantic output unit, a region division unit, and a sub-image cutting unit; The specification file calling unit is used to call the corresponding specification file according to the determined production scenario; The specification semantic output unit is used to perform semantic analysis on the specification document through the large model and output multiple specification requirements; The region division unit is used to divide the image into a large-scale information region and a small-scale information region based on specification requirements; The sub-image cropping unit is used to split the pre-processed scene image into a large-scale sub-image and a small-scale sub-image according to the divided large-scale information area and small-scale information area.

Images