CN111723720A - An intelligent visualization real-time online monitoring system for organic gas leakage - Google Patents

Abstract

The invention relates to a real-time detection system for organic gas leakage based on Faster RCNN, comprising a storage tank area, an infrared camera, a communication module and a central control room. There will be organic gas leakage in the storage tank area. The infrared camera monitors the storage tank area and transmits the monitoring video to the server in the central control room through the communication cable. The server stores the backup monitoring video and performs leak detection on the real-time uploaded monitoring video. The display will pass the leak detection. The video shows an early warning. The working steps of leak detection are to call the surveillance video uploaded in real time, identify the leakage of the video through the fast regional convolutional neural network identification model, determine the source of leakage and the type of leakage, and output the identified video. Compared with sensors and common infrared cameras, the present invention can avoid the influence of environment and human factors, ensure higher leakage identification performance, and at the same time low cost.

Description

An intelligent visualization real-time online monitoring system for organic gas leakage

technical field

The invention relates to an intelligent visual real-time online monitoring system for organic gas leakage, which belongs to the technical field of organic gas leakage monitoring.

Background technique

In process industries such as ethylene cracking, organic gas leakage is a dangerous source of fire, explosion and poisoning accidents. Therefore, the key to avoid such accidents is the timely early warning, control and disposal of organic gases. Industrial refineries use sensors or infrared cameras for leak detection. The sensor monitoring method is easily affected by wind speed, wind direction and plant environment, and it is difficult to accurately locate the leakage source and distinguish the leakage type. At the same time, the general monitoring area requires multiple sensors to form a monitoring network, which is costly. Compared with sensor methods, infrared thermal imagers can monitor organic gas leaks in areas several meters away, reducing maintenance time and costs. However, the monitoring video of infrared thermal imager requires manual observation of organic gas leakage, which causes waste of personnel, and the unreliability of people also has a huge negative impact on the real-time monitoring effect.

In video or image anomaly recognition, traditional data-driven methods are limited in the accuracy of anomaly pattern recognition. On the contrary, deep learning neural networks show high accuracy in the field of image and sound recognition. Among them, the Faster Regions with Convolutional Neural Network (Faster RCNN) proposed in 2015 uses the Region Proposal Networks (RPN) to replace the Selective Search in the previous RCNN and Fast RCNN. Search) method, which integrates Feature Extraction, Proposal Extraction, Bounding Box Regression, and Classification into one network, which improves the detection accuracy and speed of CNN. The end-to-end target detection framework is truly realized, and the detection speed is particularly improved. Therefore, Faster RCNN can meet the task of anomaly identification in real-time monitoring video. However, since the training of Faster RCNN requires the support of big data, and the data sets collected in the industrial field are not enough to train high-performance recognition models, it is necessary to directly transfer the parameters of the pre-trained models on other data sets. In the pre-training model, the data set type, feature extractor type and the number of candidate regions all affect the performance of the training model, so sensitivity analysis should be performed to determine the parameter selection principle of the pre-training model.

In view of the hazards of organic gas leakage and the shortcomings of traditional monitoring methods, it is urgent to invent an intelligent visualization real-time online monitoring system for organic gas leakage, which can realize intelligent monitoring of organic gas leakage and provide security for the process industry.

SUMMARY OF THE INVENTION

The invention provides an intelligent visualization real-time online monitoring system for organic gas leakage, which is used to solve the defects in the prior art.

The present invention is achieved through the following technical solutions:

An intelligent visualization real-time online monitoring system for organic gas leakage includes a storage tank area, an infrared camera, a communication module, a server, a leakage identification module and a display. The storage tank area stores materials, which will cause organic gas leakage; the infrared camera is installed in the storage tank area, and the installation height is set according to the monitoring range; the communication module includes a wired communication link and a backup wireless communication module, responsible for transmitting infrared monitoring Video data; the server is responsible for storing backup and calling video data; the leak identification module calls the video data through the secondary development of the infrared camera SDK, is responsible for marking the leak label of the video data, and determines the leak source and leak type; the display is responsible for displaying Video data for leak identification; the working steps of the leak identification module are to call the video data uploaded in real time, identify the leak of the video data through the Faster RCNN identification model, determine the source and type of leakage, and output the identified video. The construction of the leak identification module includes the following steps:

Step S1, data collection: collecting image data;

Step S2, data labeling: label the image with a leak label;

Step S3, model training: model training based on transfer learning;

Step S4, model evaluation: perform model evaluation based on the model accuracy index and the model speed index;

Step S5, the optimal model: determine the optimal model based on the game idea.

In the above-mentioned intelligent visualization real-time online monitoring system for organic gas leakage, the data collection in step S1 is to collect the video of organic gas leakage in the storage tank area, and then frame the collected video at fixed time intervals to obtain image data.

In the above-mentioned intelligent visualization real-time online monitoring system for organic gas leakage, the data labeling in step S2 uses Labelling software to label the image data for leakage, and only labels clearly visible gases near the leak source.

In the above-mentioned intelligent visualization real-time online monitoring system for organic gas leakage, the model training in step S3 uses transfer learning to optimize and improve the prediction accuracy of the constructed model with the help of parameters such as weights obtained from other general quantity sets.

The above-mentioned intelligent visualization real-time online monitoring system for organic gas leakage, the model evaluation described in step S4, the model evaluation indicators are mean average accuracy (mAP) and detection inference time (that is, response speed), wherein mAP is a plurality of categories The mean of the mean precision (AP).

The average precision is used to test whether the gas leak has been successfully identified. The AP value is the given coincidence threshold, and the average value of the corresponding precision under the 11 recall values (Recall) is:

In the formula, TP (True positive) is the number of frames for correct detection of gas leakage, and FP (False positive) is the number of false positives, that is, the number of frames for false detection. FN (False negative) is the number of false negatives, representing the number of undetected frames.

The coincidence degree of the truth ground box and the predicted label box (Bounding box) is:

In the formula, A is the true label box in the data annotation, and B is the label box predicted by the developed model. If the degree of coincidence between the two is greater than the given default degree of coincidence, the prediction result can be considered as TP, otherwise it is FP.

A kind of above-mentioned organic gas leakage intelligent visualization real-time online monitoring system, the optimal model described in step S5, adopts the game idea to judge, by constantly changing the network topology of the model, between the model accuracy and the model inference speed. Therefore, it can not only achieve higher model accuracy, but also greatly reduce the time required for model detection.

The advantages of the present invention are as follows: the present invention adopts an infrared camera to monitor the storage tank area of the refinery. Compared with the sensor monitoring method, it is not affected by factors such as terrain and environment, and has high leakage identification accuracy; the present invention adopts a migration learning training model, The training efficiency is improved, the high performance of the model is ensured, the game idea is used to determine the optimal model, and the accuracy of the model is improved under the premise of lower inference time; the invention uses the Faster RCNN target recognition model for the detection of organic gases in the video. The automatic identification can determine the leakage source and classify the leakage type; the invention saves labor cost, does not need to manually judge the leakage of the leakage video, avoids artificial unreliability, and realizes the automatic leakage identification.

Description of drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the accompanying drawings that need to be used in the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description These are some embodiments of the present invention, and for those of ordinary skill in the art, other drawings can also be obtained from these drawings without any creative effort.

Figure 1 is a schematic structural diagram of an intelligent real-time monitoring system for organic gas leakage; Figure 2 is a schematic diagram of the construction of a leakage identification module.

Reference numerals: 1 Tank farm, 2 Infrared cameras, 3 Communication cables, 4 Leak identification modules, 5 Servers, 6 Displays.

Detailed ways

In order to make the purposes, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments These are some embodiments of the present invention, but not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

An intelligent visualization real-time online monitoring system for organic gas leakage, as shown in FIG. For an industrial ethylene cracker, the infrared camera performs real-time monitoring, and the monitoring video data is transmitted to the server through the cable. The pre-trained leak identification module uses the infrared camera SDK secondary development to call the video data to identify and label the video data leakage. The monitor shows leaked surveillance video.

Leak identification module construction, the specific steps are:

Step S1, data collection: FLIR GF320 infrared thermal imager photographed organic gas leaked from the ethylene cracking unit, and the decomposed frames were 3205 images. The training set contains 2000 frames, and the validation set and test set respectively contain 605 frames and 600 frames. , the image pixels are 320×240.

Step S2, data labeling: Label the image data for leaks using Labelling software, and label only the gas that is clearly visible near the leak source.

Step S3, model training: using pre-trained models of 5 mainstream datasets (COCO, KITTI, Open Images, INaturalistSpecies, AVAv2.1) to perform parameter migration to train a new target recognition model.

Step S4, model evaluation: using the indicators mAP, mAP (large), mAP (medium), mAP (small) and inference time to determine the selection data set is coco, the number of candidate regions is 100, and the feature extractor is Resnet 101.

Step S5, the optimal model: determine the optimal model based on the game idea, and determine that the Faster RCNN model constructed under the Faster_rcnn_resnet101_coco pre-training model, the Resnet101 feature extractor, and the number of candidate regions is 100 as the optimal model, and the mAP of this model can reach 71 %, while the inference time is only 55ms.

Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, but not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that it can still be The technical solutions described in the foregoing embodiments are modified, or some technical features thereof are equivalently replaced; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims (6)

1. The utility model provides a visual real-time on-line monitoring system of organic gas leakage intelligence which characterized in that: the system comprises a tank storage area, an infrared camera, a communication module, a server, a leakage identification module and a display; the storage tank area stores materials, and organic gas leakage can be generated; the infrared camera is arranged in the tank storage area, and the installation height is set according to the monitoring range; the communication module comprises a wired communication link and a standby wireless communication module and is responsible for transmitting infrared monitoring video data; the server is responsible for storing backup and calling video data; the leakage identification module calls video data through secondary development of an infrared camera SDK, is responsible for marking a leakage label of the video data and determines a leakage source and a leakage type; the display is responsible for displaying the video data of the leakage identification; the working steps of the leakage identification module are to call video data uploaded in real time, carry out leakage identification on the video data through a fast RCNN identification model, determine a leakage source and a leakage type, and output an identified video, wherein the construction of the leakage identification module comprises the following steps:

step S1, data acquisition: collecting image data;

step S2, data annotation: labeling the leakage label on the image;

step S3, model training: performing model training based on transfer learning;

step S4, model evaluation: performing model evaluation based on the model precision index and the model speed index;

step S5, optimal model: and judging an optimal model based on the game idea.

2. The intelligent visual real-time online monitoring system for organic gas leakage according to claim 1, characterized in that: and S1, acquiring the data acquisition, acquiring the organic gas leakage video of the storage tank area, and framing the acquired video according to a fixed time interval to acquire image data.

3. The intelligent visual real-time online monitoring system for organic gas leakage according to claim 1, characterized in that: and (S2) data labeling, namely labeling the leakage of the image data by using Labelling software, and labeling only the gas which is clearly visible near the leakage source.

4. The intelligent visual real-time online monitoring system for organic gas leakage according to claim 1, characterized in that: and step S3, training the model, and optimizing and improving the prediction accuracy of the constructed model by using transfer learning and by means of the parameters such as the weight and the like acquired from other universal quantity sets.

5. The intelligent visual real-time online monitoring system for organic gas leakage according to claim 1, characterized in that: evaluating the model in step S4, wherein the evaluation indexes of the model are mean average accuracy (mAP) and detection inference time (i.e. response speed), wherein the mAP is a mean of average Accuracies (APs) of a plurality of categories;

average accuracy is used to test whether a gas leak was successfully identified, with the AP value being the threshold for a given overlap, and the average of the corresponding accuracies (precisions) at 11 Recall values (recalls) being:

wherein TP (true positive) is the number of frames for correctly detecting gas leakage; FP (false positive) is the number of false reports, namely the number of false detection frames; FN (false negative) is a missing report, which represents the number of undetected frames;

the coincidence ratio of the real label box (Truth ground box) and the predicted label box (Bounding box) is as follows:

in the formula, A is a real label frame marked on data, and B is a label frame predicted by a developed model; if the contact ratio of the two is larger than the given default contact ratio, the predicted result can be regarded as TP, otherwise, the result is FP.

6. The intelligent visual real-time online monitoring system for organic gas leakage according to claim 1, characterized in that: and step S5, the optimal model is judged by adopting a game idea, and a game is played between the model accuracy and the model reasoning speed by continuously changing the network topology structure of the model, so that the purposes of achieving higher model accuracy and greatly reducing the time required by model detection are achieved.

Images