Petrochemical fire type identification method
The flame shape in petrochemical fire scenes is obtained through infrared temperature measurement imagers and visible light cameras, a flame material library is built and a deep object detection model is trained, and combined with infrared thermal imaging and video recognition technology, the insufficient accuracy of fire recognition in complex scenarios of petrochemical enterprises is solved. The problem of this problem has achieved high accuracy and real-time fire identification and early warning.
Patent Information
- Application Number
- CN202310531863.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2023-05-11
- Publication Date
- 2024-11-12
AI Technical Summary
Existing fire identification technology is difficult to accurately identify multiple fires in complex scenarios of petrochemical enterprises, and there are many disturbing factors, so the recognition accuracy is difficult to ensure.
An infrared temperature measurement imager and visible light camera are used to simultaneously obtain various flame forms in petrochemical fire scenes, build a flame material library, train a deep object detection model, and combine infrared thermal imaging and video recognition technology to achieve flame recognition and alarm.
It effectively improves the accuracy and real-time identification of multi-form fires in complex environments of petrochemical enterprises, lays the foundation for accident positioning and rapid analysis, and improves the intelligence level of fire monitoring.
Smart Images

Figure CN118941836A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the field of flame identification, and in particular to a petrochemical fire type identification method. Background Art
[0002] There are many dangerous factors in the production, storage and transportation of petrochemical enterprises. In particular, raw materials, intermediate products and final products are mostly flammable and explosive hazardous chemicals. Once they leak and form explosive mixtures, they will explode when encountering open flames, causing fire accidents.
[0003] In recent years, with the rapid development of big data and artificial intelligence, especially the widespread application of deep learning algorithms, fire identification methods based on computer software have become possible. With the rapid development of computer vision technology, the image-based fire detection system based on video image recognition of fireworks has the advantages of long control distance, large protection area, and fast response speed. It has now developed into the mainstream technology for flame identification in various fields.
[0004] The fire detection algorithm based on video recognition mainly analyzes the image features of flames and smoke, detects and recognizes the features, and then realizes fire alarm. At present, it is divided into traditional recognition detection algorithms based on digital image processing and target detection algorithms based on deep learning. The former algorithm mainly uses conventional feature extraction, and the algorithm is relatively classic, but the recognition accuracy is low and is only applicable to simple scenarios. The latter relies on big data for training and has a high recognition accuracy.
[0005] With the rapid development of computer hardware performance, deep learning-based target detection algorithms have begun to be applied to all walks of life and have become the mainstream technology for image recognition today. However, the petrochemical plant area is not only extremely complex, but also has different flame forms and types. Relying solely on visible light video detection technology has many problems such as single identification means and multiple interference factors, and the accuracy is difficult to guarantee.
[0006] In addition, the flame and fire recognition algorithms currently developed by relevant institutions are mainly based on model training based on fire materials such as forest fires, building fires, and simple open flames in experiments. They have better flame recognition effects in corresponding forest fire and building fire scenarios, but it is difficult to achieve satisfactory recognition effects for specific fire scenarios and flame forms that may occur in complex equipment areas of petrochemical enterprises.
[0007] Therefore, it is necessary to provide a new technical solution to address the shortcomings of the above-mentioned technologies and lay the foundation for the rapid identification and early warning response of flames in complex petrochemical scenes. Summary of the invention
[0008] The purpose of the embodiment of the present invention is to provide a technical solution that can effectively improve the accuracy and real-time performance of identifying multi-form fires in complex environments of petrochemical enterprises, laying the foundation for the next step of accident location and rapid analysis.
[0009] In order to achieve the above object, an embodiment of the present invention provides a petrochemical fire type identification method, the method comprising:
[0010] Step 1: using an infrared temperature imaging camera to obtain various flame forms in various petrochemical fire scenes, and constructing a first flame material library according to different types of flame forms; using a visible light camera to obtain various flame forms in various petrochemical fire scenes, and constructing a second flame material library according to different types of flame forms;
[0011] Step 2, randomly selecting different images from the first flame material library as a first training set and a first verification set; randomly selecting different images from the second flame material library as a second training set and a second verification set;
[0012] Step 3, using the first training set to train the selected flame recognition model to obtain a first flame recognition model, and using the first verification set to verify the recognition effect of the first flame recognition model; using the second training set to train the selected flame recognition model to obtain a second flame recognition model, and using the second verification set to verify the recognition effect of the second flame model;
[0013] Step 4, using an infrared temperature imaging device to obtain a first image of the scene to be measured, and using a visible light camera to obtain a second image of the scene to be measured; inputting the first image into the first flame recognition model to obtain a first recognition result, and inputting the second image into the second flame recognition model to obtain a second recognition result;
[0014] Step 5: Sending an alarm signal based on the first recognition result and the second recognition result.
[0015] Preferably, the obtaining of various flame forms in various petrochemical fire scenes includes:
[0016] Collect flame videos of various petrochemical fire scenes;
[0017] Extract frames from the flame video to obtain a flame sample image;
[0018] The flame video covers the scene conditions of various equipment areas of different petrochemical enterprises, and there are no open areas in the background area and complex structures with partial pipeline obstruction.
[0019] Preferably, the various flame forms include flowing fire, liquid pool fire, jet fire, tank fire, tank fire, hydrogen combustion fire, electrical fire, and pipe gallery obstruction fire; there are no fewer than 5,000 flame sample images of each form, covering no fewer than 10 scene conditions, and no more than 1,000 sample images extracted under each scene condition.
[0020] Preferably, the obtaining of various flame forms in various petrochemical fire scenes also includes: performing standardized preprocessing on the flame sample images to construct the first flame material library and the second flame material library; the standardized preprocessing includes image denoising and distortion correction to make the image ratio, field of view and target size the same or similar, and using a rectangular annotation box to select the coordinates and types of various flame targets in the image, ensuring that the annotation box completely covers the flame target and the area of each flame target in the annotation box accounts for no less than 1:3.
[0021] Preferably, the resolution of the infrared thermometer is not less than 800*600 pixels, the resolution of the visible light camera is not less than 1024*768 pixels, and the target scale of the multiple flame forms is greater than 20*20 pixels.
[0022] Preferably, the ratio of the training set to the validation set is not less than 4:1; and the flame recognition model is a Yolov5 deep target detection network model.
[0023] Preferably, the step 5 comprises: when the first recognition result and the second recognition result simultaneously indicate abnormal flame, sending an alarm signal externally.
[0024] Preferably, the step 5 further comprises: when one of the first recognition result and the second recognition result indicates an abnormal flame, sending the first image frame and the second image frame to a manual review platform for manual analysis and judgment by the manual review platform.
[0025] Preferably, before step 2, the method further includes: performing data set preprocessing on the first flame material library and the second flame material library respectively, wherein the data set preprocessing includes:
[0026] The original images in the first flame library and the original images in the second flame library are respectively subjected to random addition of Gaussian noise, CutMix data enhancement and mosaic data enhancement operations, and the images after mosaic data enhancement are respectively placed in the first flame library and the second flame library.
[0027] Preferably, before step 4, the method further includes: packaging the first flame recognition model and the second flame recognition model.
[0028] The technical solution provided by the embodiment of the present invention has the following beneficial effects:
[0029] The embodiment of the present invention simultaneously adopts an infrared temperature imaging device and a visible light camera to respectively obtain various flame forms in petrochemical fire scenes, effectively integrating the two monitoring methods of infrared thermal imaging and video recognition, and using various flame forms in petrochemical fire scenes as data samples of the flame material library to train and verify the flame recognition model, thereby performing flame recognition, which can effectively improve the recognition accuracy and real-time performance of multi-form fires in complex environments of petrochemical enterprises, and lay the foundation for the next step of accident location and rapid analysis.
[0030] Other features and advantages of the embodiments of the present invention will be described in detail in the subsequent detailed description. BRIEF DESCRIPTION OF THE DRAWINGS
[0031] The accompanying drawings are used to provide a further understanding of the embodiments of the present invention and constitute a part of the specification. Together with the following specific implementations, they are used to explain the embodiments of the present invention, but do not constitute a limitation on the embodiments of the present invention. In the accompanying drawings:
[0032] Figure 1 The figure is a flow chart of a petrochemical fire type identification method provided by an embodiment of the present invention.
[0033] Figure 2 This is a flow chart of constructing a flame material library provided by an embodiment of the present invention. DETAILED DESCRIPTION
[0034] The specific implementation of the embodiment of the present invention is described in detail below in conjunction with the accompanying drawings. It should be understood that the specific implementation described here is only used to illustrate and explain the embodiment of the present invention, and is not used to limit the embodiment of the present invention.
[0035] See also Figure 1-Figure 2 The embodiment of the present invention provides a petrochemical fire type identification method, the method comprising:
[0036] Step 1: Use an infrared temperature imaging camera to obtain various flame forms in various petrochemical fire scenes, and build a first flame material library based on different types of flame forms; use a visible light camera to obtain various flame forms in the various petrochemical fire scenes, and build a second flame material library based on different types of flame forms.
[0037] The embodiment of the present invention combines an infrared temperature measurement imager and a visible light camera to obtain a large amount of flame videos in complex scenes of petrochemical enterprises, wherein the video images captured by the visible light camera are standard videos or high-definition video files, and the resolution is preferably not less than 1024*768 pixels; the infrared temperature measurement imager outputs infrared temperature measurement imaging results, and the resolution is preferably not less than 800*600 pixels. The scale of various flame targets in the visible light video is preferably greater than 20*20 pixels to ensure that the model recognition accuracy meets the requirements.
[0038] In order to make the recognition results more accurate and reliable, preferably, the flame video covers various flame forms of various fire accidents in the petrochemical industry, including flowing fire, liquid pool fire, jet fire, tank fire, tank fire, hydrogen combustion fire, electrical fire, pipe gallery obstruction fire and other flames, etc. The specific form classification can be further refined and expanded on the basis of the above classification. The flame video collected by infrared temperature imaging cameras and visible light cameras should cover the scene conditions of various device areas of different petrochemical enterprises as much as possible, with no open areas in the background area, complex structures and partial pipeline obstruction, so that the first flame material library and the second flame material library are closer to the real scene to be tested, and more intelligent flame precision monitoring and recognition can be achieved.
[0039] In order to establish the first flame material library and the second flame material library based on the acquired massive flame videos, it is also necessary to perform video frame extraction and sample labeling on the acquired flame videos.
[0040] Video frame extraction refers to the use of video frame extraction tool software to extract frames from the acquired files. During the frame extraction process, the time interval setting requirement for the video file is not less than 500ms / frame. The samples extracted from the visible light video are classified in detail according to the various forms of flames, and are packaged into multiple folders such as flowing fire, liquid pool fire, jet fire, tank fire, tank fire, hydrogen combustion fire, electrical fire, pipe gallery shielding fire and other flames. The number of flame samples for each classification form is preferably no less than 5,000, the number of covered scene conditions is preferably no less than 10, and the number of sample images extracted under each scene condition is preferably no more than 1,000.
[0041] Sample labeling refers to the standardized preprocessing of sample images. Standardized preprocessing can use image denoising, distortion correction and other methods to make the image scale, field of view and target size the same or as close as possible. Sample images are labeled with flame targets using the image labeling tool LabelImg or other image labeling tools. In the labeling step, first set the labeled data set format type to YOLO. The sample file format after labeling is .txt. The classes corresponding to each flame target in the sample class file are as follows: flowing fire 0, liquid pool fire 1, jet fire 2, tank fire 3, tank fire 4, hydrogen combustion fire 5, electrical fire 6, pipe gallery shielding fire 7... and so on. New categories can be added according to the needs of the enterprise. When infrared thermal imaging recognition detection, only one category corresponds: thermal imaging recognition fire 0. When labeling samples, use a rectangular labeling box to select the coordinates and types of various flame targets in the image, ensuring that the labeling box completely covers the flame target, and the area of each flame target in the labeling box accounts for no less than 1:3.
[0042] Step 2: randomly select different images from the first flame material library as a first training set and a first verification set; randomly select different images from the second flame material library as a second training set and a second verification set.
[0043] The generated data is divided into independent and non-repetitive training sets and validation sets in a certain ratio by random sampling. The ratio of training set to validation set is generally not less than 4:1. The training set is used to provide the flame recognition model for training. The validation set is used to test the model effect in order to verify whether the model performance meets the application requirements.
[0044] Step 3: Use the first training set to train the selected flame recognition model to obtain a first flame recognition model, and use the first verification set to verify the recognition effect of the first flame recognition model; use the second training set to train the selected flame recognition model to obtain a second flame recognition model, and use the second verification set to verify the recognition effect of the second flame model.
[0045] The first flame recognition model and the second flame recognition model can both be YOLOv5 deep target detection network models, which are implemented based on Python program code and can be opened through common Python language editors such as pycharm, and training parameters are set in the model train file parse_opt (known = False) function. The core code parameters of the training file are set as follows:
[0046]
[0047] Among them, weights represents the model weight, which can be set to one of YOLOv5s.pt, YOLOv5m.pt, YOLOv5l.pt and YOLOv5x.pt; data represents the training data set, and the corresponding training set can be named fire.yaml; epochs represents the round of training, and the setting value is not less than 200; resume indicates whether to retrain or continue the previous training results during training. It is set to False for the first training, and is changed to True when the training process is not interrupted. Other parameters are set by default options. Training starts after the training settings are completed.
[0048] During the flame recognition model parameter training process, model indicator monitoring is also required. Specifically, the core indicator result values of the YOLOv5 deep target detection network model are dynamically saved in the results.csv file as the training progresses. During the training process, the metrics / mAP_0.5 in the results.csv file is continuously and dynamically monitored. When the training is completed, the metrics / mAP_0.5 result value should be no less than 0.99, otherwise the indicator can be improved by increasing the training rounds or expanding the training set. The core indicator monitoring table of the training process is as follows:
[0049]
[0050] After the model training is completed, the model needs to be verified. Specifically, the verification parameters are set in the parse_opt() of the model detect file. The model recognition effect is adjusted by adjusting the setting values of conf-thres and iou-thres. conf-thres represents the recognition confidence threshold, which is set to 0.25 by default, and iou-thres represents the intersection-over-union ratio threshold, which is set to 0.45 by default. The core code parameter settings of the verification file are as follows:
[0051]
[0052] Step 4: Use an infrared temperature imaging device to obtain a first image of the scene to be tested, and use a visible light camera to obtain a second image of the scene to be tested; input the first image into the first flame recognition model to obtain a first recognition result, and input the second image into the second flame recognition model to obtain a second recognition result.
[0053] The flame recognition model established using the relevant data of the infrared thermal imager is used as the first flame recognition model, and the flame recognition model established using the relevant data of the visible light camera is used as the second flame model. Before recognizing the flame, the first flame recognition model and the second flame recognition model are encapsulated. Specifically, the first flame model and the second flame model are respectively placed in the monitoring host computer, or in the video processing middle-end system of the monitoring center server, and the system binds the corresponding model service and loads the model.
[0054] After loading the model, the system enters the actual measurement phase, and the acquisition results of the infrared temperature imager and the visible light camera are respectively input into the first flame recognition model and the second flame recognition model, so as to obtain the recognition response results of the first flame recognition model and the recognition response results of the second flame recognition model.
[0055] Step 5: Sending an alarm signal based on the first recognition result and the second recognition result.
[0056] The alarm strategy for abnormal events mainly includes two main methods: single abnormal alarm and multiple abnormal alarm. From the perspective of alarm timeliness, it is divided into immediate alarm and delayed alarm. The system adopts a dual alarm mode in which visible light camera and infrared thermal imager respond in parallel. When the visible light video recognition result and the infrared thermal imager recognition result simultaneously find flame abnormality within a time difference of 1s, the alarm is immediately activated and the intelligent emergency response process is entered. If only one of the two alarm modes gives an abnormal flame recognition result, and the other mode does not respond within a 1s time interval, the emergency process will not be entered immediately after the alarm information is given. It must first be submitted to the manual auxiliary analysis module, and the monitoring personnel will determine whether it is a system misjudgment, and then decide whether to handle the alarm by initiating the emergency procedure.
[0057] Specifically, if the first recognition result and the second recognition result simultaneously indicate abnormal flames, an alarm signal is sent out to push the petrochemical fire type, for example, "Oil pump A01 may have a leak causing a jet fire, please confirm immediately!".
[0058] If one of the first recognition result and the second recognition result indicates an abnormal flame, that is, the two recognition models fail to give warning information at the same time, the first image screen and the second image screen are sent to the manual review platform, which performs manual analysis to determine whether it is a misjudgment, and determines whether to activate the alarm based on the analysis results.
[0059] The embodiment of the present invention simultaneously adopts an infrared temperature imaging device and a visible light camera to respectively obtain various flame forms in petrochemical fire scenes, effectively integrating the two monitoring methods of infrared thermal imaging and video recognition, and using various flame forms in petrochemical fire scenes as data samples of the flame material library to train and verify the flame recognition model, thereby performing flame recognition, which can effectively improve the recognition accuracy and real-time performance of multi-form fires in complex environments of petrochemical enterprises, and lay the foundation for the next step of accident location and rapid analysis.
[0060] In a preferred embodiment of the present invention, before step 2, the method further includes: performing data set preprocessing on the first flame material library and the second flame material library respectively, so that the images in the first flame material library and the second flame material library are closer to the actual test scene, thereby improving the accuracy and real-time performance of flame monitoring.
[0061] The data set preprocessing includes: randomly adding Gaussian noise, CutMix data enhancement and mosaic data enhancement operations to the original images in the first flame material library and the original images in the second flame material library, and placing the images after mosaic data enhancement into the first flame material library and the second flame material library respectively.
[0062] Among them, Gaussian noise refers to a type of noise whose probability density function obeys Gaussian distribution (i.e., normal distribution). Common Gaussian noise includes fluctuation noise, cosmic noise, thermal noise, shot noise, etc. CutMix data enhancement randomly cuts out a rectangular area from an image in the data set, and fills the corresponding areas of other images into the cut image to obtain a new data set sample. Mosaic data enhancement randomly flips, transforms the color gamut, scales, and other processes on four images in the data set, and then places the processed four images in four corners in turn to obtain a new image that is a mixture of four images of the same size as the original image.
[0063] The embodiments of the present invention can be widely used in the field of petrochemical enterprises, and the system can be deployed in the operation room or video monitoring command center of the device. By reading the monitoring video stream of the key risk points in the device area, the rapid identification of flame risks and the multi-form classification alarm of flames can be realized, and the emergency duty personnel can be effectively assisted in the fire accident monitoring and early warning in complex petrochemical scenes, thereby improving the intelligent monitoring level of fire accidents in complex scenes of petrochemical enterprises. The technical solution provided by the embodiment of the present invention is further explained below in combination with application examples.
[0064] Embodiment 1:
[0065] Hikvision explosion-proof camera high-definition full-color explosion-proof hemispherical camera is used as a visible light camera, and Dali D843NT infrared temperature measurement imaging module is used as an infrared temperature measurement imager. Both are installed at the high point of the perimeter of the abandoned liquid hydrocarbon spherical tank area used for training in a petrochemical enterprise as a dual-light video acquisition module, and the host computer is used to simultaneously collect visible light video and infrared imaging video. A Hikvision high-performance video analysis server that supports fast processing of 4 RTX 3090 graphics cards is placed in the inner operation room, and the target detection model based on the Yolov5 network and the dual-light detection recognition and early warning module provided by the present invention are loaded as the background service center of this system. Video data and instructions are transmitted through the petrochemical enterprise intranet, and the petrochemical fire type recognition method described above is used for on-site monitoring.
[0066] The simulated jet fire was ignited by the leakage of the liquid hydrocarbon spherical tank A01, and the actual measurement stage was entered. First, it was determined whether there was an abnormal flame state in the real-time video images of the two channels of the visible light camera and the infrared temperature imager. If an abnormal flame occurs, the recognition response results of the first recognition model and the second recognition model are checked: if the first recognition model and the second recognition model give abnormal flame results at the same time, the emergency procedure is intelligently started, an alarm message is sent to the safety management personnel, and the petrochemical fire type is pushed according to the flame classification results given by the recognition model, such as "Liquid hydrocarbon spherical tank A01 may have a jet fire caused by leakage, please confirm immediately!"; if the first recognition model and the second recognition model fail to give warning information at the same time, they will go to the manual analysis module to determine whether it is a misjudgment based on the results of manual analysis and observation, and determine whether to alarm based on the analysis results.
[0067] Embodiment 2:
[0068] A high-definition full-color explosion-proof hemispherical camera of a certain brand was used as a visible light camera, and a Dali D843NT infrared temperature measurement imaging module was used as an infrared temperature measurement imager. The two were installed in the pipeline corridor of the pump area of an abandoned catalytic cracking unit used for training in a petrochemical enterprise as a dual-light video acquisition module, and the host computer was used to simultaneously acquire visible light video and infrared imaging video. A Hikvision high-performance video analysis server that supports fast processing of 4 RTX 3090 graphics cards was placed in the inner operation room, loaded with a multi-modal flame target detection model based on the YOLOv5 network and a dual-light detection recognition and early warning module, as the background service center of this system, video data and instructions were transmitted through the petrochemical enterprise intranet, and the petrochemical fire type recognition method described above was used for on-site monitoring.
[0069] By igniting the oil pump A01 to simulate the flowing fire, the actual measurement stage is entered. The method used in the actual measurement stage is similar to that in Example 1, except that the type of petrochemical fire pushed outward is different, such as "The oil pump A01 may have a leak that causes a flowing fire, please confirm immediately!".
[0070] Embodiment 3:
[0071] A high-definition full-color explosion-proof hemispherical camera of a certain brand was used as a visible light camera, and a Dali D843NT infrared temperature measurement imaging module was used as an infrared temperature measurement imager. The two were installed at the high point of the perimeter of the abandoned diesel tank area device used for training in a petrochemical enterprise as a dual-light video acquisition module, and the host computer was used to simultaneously collect visible light video and infrared imaging video. An Inspur high-performance server that supports fast processing of 4 RTX3070 graphics cards was placed in the inner operation room, loaded with a multi-modal flame target detection model based on the YOLOv5 network and a dual-light detection recognition and early warning module, as the background service center of this system, video data and instructions were transmitted through the petrochemical enterprise intranet, and the petrochemical fire type recognition method described above was used for on-site monitoring.
[0072] By igniting the oil tank D101 to simulate a tank fire, the actual measurement stage is entered. The method used in the actual measurement stage is similar to that in Example 1, except that the type of petrochemical fire pushed outward is different, such as "an oil tank fire may occur in the oil tank D101, please confirm immediately!".
[0073] Embodiment 4:
[0074] A certain brand of explosion-proof camera high-definition full-color explosion-proof hemispherical camera is used as a visible light camera, and a Dali D843NT infrared temperature measurement imaging module is used as an infrared temperature measurement imager. The two are installed on the intelligent inspection robot outside the abandoned electrical workshop substation of a dangerous chemical accident emergency rescue base training center as a dual-light video acquisition module to simultaneously collect visible light video and infrared imaging video. An Inspur high-performance server that supports fast processing of 4 RTX3070 graphics cards is placed in the inner operation room, loaded with a multi-modal flame target detection model based on the YOLOv5 network and a dual-light detection recognition and early warning module, as the background service center of this system, video data and instructions are transmitted through the petrochemical enterprise intranet, and the petrochemical fire type recognition method described above is used for on-site monitoring.
[0075] By igniting the simulated fire point in the substation, the actual measurement stage is entered. The method used in the actual measurement stage is similar to that in Example 1, except that the type of petrochemical fire pushed outward is different, such as "an electrical fire may occur in the substation, please confirm immediately!".
[0076] Embodiment 5:
[0077] A high-definition full-color explosion-proof hemispherical camera of a certain brand was used as a visible light camera, and a Dali D843NT infrared temperature measurement imaging module was used as an infrared temperature measurement imager. The two were installed at the high point of the perimeter of the abandoned diesel tank area device used for training in a petrochemical enterprise as a dual-light video acquisition module, and the host computer was used to simultaneously collect visible light video and infrared imaging video. An Inspur high-performance server that supports fast processing of 4 RTX3070 graphics cards was placed in the inner operation room, loaded with a multi-modal flame target detection model based on the YOLOv5 network and a dual-light detection recognition and early warning module, as the background service center of this system, video data and instructions were transmitted through the petrochemical enterprise intranet, and the petrochemical fire type recognition method described above was used for on-site monitoring.
[0078] The tank fire is simulated by igniting the oil tank filled with diesel next to the oil tank, and the actual measurement stage begins. The method used in the actual measurement stage is similar to that in Example 1, except that the type of petrochemical fire pushed outward is different, such as "An oil tank fire may occur in the oil tank in the diesel tank area, please confirm immediately!".
[0079] Through the specific application examples of the above embodiments 1 to 5, it is verified that the technical solution provided by the embodiments of the present invention can accurately and real-timely realize the identification and early warning of petrochemical fire types.
[0080] The above are only embodiments of the present application and are not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included within the scope of the claims of the present application.
Claims
1. A method for identifying the type of petrochemical fire, characterized in that: The method comprises: Step 1: using an infrared temperature imaging camera to obtain various flame forms in various petrochemical fire scenes, and constructing a first flame material library according to different types of flame forms; using a visible light camera to obtain various flame forms in various petrochemical fire scenes, and constructing a second flame material library according to different types of flame forms; Step 2: randomly selecting different images from the first flame material library as a first training set and a first verification set; randomly selecting different images from the second flame material library as a second training set and a second verification set; Step 3, using the first training set to train the selected flame recognition model to obtain a first flame recognition model, and using the first verification set to verify the recognition effect of the first flame recognition model; using the second training set to train the selected flame recognition model to obtain a second flame recognition model, and using the second verification set to verify the recognition effect of the second flame model; Step 4, using an infrared temperature imaging device to obtain a first image of the scene to be measured, and using a visible light camera to obtain a second image of the scene to be measured; inputting the first image into the first flame recognition model to obtain a first recognition result, and inputting the second image into the second flame recognition model to obtain a second recognition result; Step 5: Sending an alarm signal based on the first recognition result and the second recognition result.
2. The petrochemical fire type identification method according to claim 1, characterized in that: The method of obtaining various flame forms in various petrochemical fire scenes includes: Collect flame videos of various petrochemical fire scenes; Extract frames from the flame video to obtain a flame sample image; The flame video covers the scene conditions of various equipment areas of different petrochemical enterprises, and there are no open areas in the background area and complex structures with partial pipeline obstruction.
3. The petrochemical fire type identification method according to claim 2, characterized in that: The various flame forms include flowing fire, liquid pool fire, jet fire, trough fire, tank fire, hydrogen combustion fire, electrical fire, and pipe gallery obstruction fire; there are no fewer than 5,000 flame sample images of each type of form, covering no fewer than 10 scene conditions, and no more than 1,000 sample images extracted under each scene condition.
4. The petrochemical fire type identification method according to claim 2, characterized in that: The acquiring of various flame forms in various petrochemical fire scenes further includes: performing standardized preprocessing on the flame sample images to construct the first flame material library and the second flame material library; The standardized preprocessing includes image denoising and distortion correction to make the image ratio, field of view and target size the same or similar, and to select the coordinates and types of various flame targets in the image with a rectangular annotation frame to ensure that the annotation frame completely covers the flame target and the area of each flame target accounts for no less than 1:3 in the annotation frame.
5. The petrochemical fire type identification method according to claim 1, characterized in that: The resolution of the infrared thermometer is not less than 800*600 pixels, the resolution of the visible light camera is not less than 1024*768 pixels, and the target scale of the multiple flame forms is greater than 20*20 pixels.
6. The petrochemical fire type identification method according to claim 1, characterized in that: The ratio of the training set to the validation set is not less than 4:1; the flame recognition model is a Yolov5 deep target detection network model.
7. The petrochemical fire type identification method according to claim 1, wherein step 5 comprises: When the first recognition result and the second recognition result simultaneously indicate abnormal flame, an alarm signal is sent out.
8. The petrochemical fire type identification method according to claim 7, wherein step 5 further comprises: When one of the first recognition result and the second recognition result indicates an abnormal flame, the first image frame and the second image frame are sent to a manual review platform for manual analysis and judgment.
9. The petrochemical fire type identification method according to any one of claims 1 to 8, characterized in that: Before step 2, the step further includes: performing data set preprocessing on the first flame material library and the second flame material library respectively, wherein the data set preprocessing includes: The original images in the first flame library and the original images in the second flame library are respectively subjected to random addition of Gaussian noise, CutMix data enhancement and mosaic data enhancement operations, and the images after mosaic data enhancement are respectively placed in the first flame library and the second flame library.
10. The petrochemical fire type identification method according to any one of claims 1 to 8, further comprising before step 4: The first flame recognition model and the second flame recognition model are encapsulated.
Citation Information
Patent Citations
Method for fire identification through infrared and visible-light video image fusion
CN105512667A
Forest fire identification method and system, program and storage medium
CN112257554A
Smoke and fire identification method and device, computer equipment and storage medium
CN112560657A
Two-channel flame detection algorithm for inspection robot
CN113947711A
Extremely early thermal imaging fire alarm system in chemical industry, petrochemical industry and pharmaceutical industry
CN114519918A
