CN-121982035-A - Method, device and equipment for identifying potential safety hazards of fuel gas

CN121982035ACN 121982035 ACN121982035 ACN 121982035ACN-121982035-A

Abstract

The application provides a method, a device and equipment for identifying potential safety hazards of fuel gas. The method comprises the steps of carrying out fine adjustment on a multi-modal large language model by adopting a reinforcement learning framework, carrying out iterative training on the fine-adjusted multi-modal large language model to obtain a potential safety hazard identification model capable of outputting position information corresponding to potential safety hazards and carrying out interpretation analysis, carrying out image preprocessing on an original image to obtain an image to be identified, the size of the image to be identified is matched with that of the potential safety hazard identification model, establishing a reference coordinate system for the position coordinates of the potential safety hazards to be output subsequently, inputting the image to be identified into the potential safety hazard identification model to obtain position information corresponding to the indicated potential safety hazards and an interpretation analysis identification result output by the potential safety hazard identification model, and overcoming the defects of interpretation deficiency, insufficient positioning precision and the like in the existing gas potential safety hazard identification technology, and improving the reliability and the accuracy of the potential safety hazard identification result.

Inventors

CHEN YAOJIA
ZHENG CHAO
QIN DONGMEI
ZHANG MENG
WANG WANG
CHEN CONGCONG
ZHANG JIAJI
WANG SHUANG

Assignees

金卡智能集团股份有限公司
易联云计算(杭州)有限责任公司

Dates

Publication Date: 20260505
Application Date: 20260409

Claims (10)

1. The method for identifying the potential safety hazard of the fuel gas is characterized by comprising the following steps of: fine tuning the multi-modal large language model by adopting a reinforcement learning framework; performing iterative training on the trimmed multi-modal large language model through a training data set to obtain a potential safety hazard identification model; Acquiring an original image, and performing image preprocessing on the original image to obtain an image to be identified, wherein the size of the image to be identified is matched with a potential safety hazard identification model; Inputting the image to be identified into the potential safety hazard identification model to obtain an identification result output by the potential safety hazard identification model, wherein the identification result is used for indicating position information corresponding to the potential safety hazard and carrying out interpretable analysis.
2. The method according to claim 1, wherein the iterative training of the trimmed multimodal large language model by the training data set to obtain the safety hazard identification model comprises: The method comprises the steps of obtaining a training data set, wherein the training data set comprises a training image set, a real hidden danger type, real position information and a real judgment basis corresponding to each training image; Inputting the training image set into the finely tuned multi-modal large language model to obtain a prediction recognition result output by the multi-modal large language model, wherein the prediction recognition result comprises a prediction hidden danger type, prediction position information and a prediction judgment basis corresponding to the training image; Optimizing the multi-modal large language model by adopting a multi-modal rewarding function based on the real hidden danger type, the real position information, the real judging basis, the predicted hidden danger type, the predicted position information and the predicted judging basis until the training times of the fine-tuned multi-modal large language model reach the preset times or the loss value corresponding to the predicted recognition result output by the fine-tuned multi-modal large language model is smaller than the preset loss value, so as to obtain the hidden danger recognition model.
3. The method of claim 2, wherein optimizing the multi-modal large language model using a multi-modal reward function based on real hidden danger types and predicted hidden danger types comprises: determining an overlapping bonus point of the real hidden danger type and the predicted hidden danger type; Optimizing the multimodal large language model based on the overlapping bonus points.
4. The method of claim 3, wherein optimizing the multimodal large language model using a multimodal rewards function based on the real location information and the predicted location information comprises: Determining euclidean distance between the real location information and the predicted location information; Determining a distance rewards score based on the Euclidean distance and fault tolerance radius; Optimizing the multimodal large language model based on the distance reward points.
5. The method of claim 4, wherein optimizing the multi-modal large language model using a multi-modal reward function based on the true decision basis and the predictive decision basis comprises: Determining semantic similarity rewarding points of the real judging basis and the predicting judging basis; Determining a keyword matching score corresponding to the prediction judgment basis; Optimizing the multimodal large language model based on the semantic similarity reward score and the keyword matching score.
6. The method according to claim 1, wherein the performing image preprocessing on the original image to obtain an image to be identified includes: Determining a corresponding scaling factor based on a maximum allowed number of pixels and an original number of pixels of the original image; According to the scaling factor, scaling the original image to obtain an intermediate image; and adjusting the intermediate image by adopting standard plaque size to obtain the image to be identified.
7. The method of claim 6, wherein the determining the corresponding scaling factor based on the maximum allowed number of pixels and the original number of pixels of the original image comprises: determining that the scaling factor is 1 in the case that the original pixel number is not greater than the maximum allowable pixel number; In the case that the original pixel number is greater than the maximum allowable pixel number, the scaling factor is determined based on the maximum allowable pixel number and the original pixel number of the original image.
8. The method according to claim 1, wherein the method further comprises: Performing image marking processing on the original image based on hidden danger types, position information and interpretability analysis corresponding to at least one potential safety hazard; And carrying out visualization processing on the original image after the marking processing.
9. A gas safety hazard identification device, the device comprising: The training module is used for fine-tuning the multi-modal large language model by adopting the reinforcement learning framework, and carrying out iterative training on the fine-tuned multi-modal large language model through the training data set to obtain a potential safety hazard identification model; the acquisition module is used for acquiring an original image; The image processing module is used for carrying out image preprocessing on the original image to obtain an image to be identified, and the size of the image to be identified is matched with the potential safety hazard identification model; The identification module is used for inputting the image to be identified into the potential safety hazard identification model to obtain an identification result output by the potential safety hazard identification model, wherein the identification result is used for indicating the position information corresponding to the potential safety hazard and the identification result capable of being interpreted and analyzed.
10. The gas safety hidden trouble identification device is characterized by comprising a memory and a processor; Wherein the memory stores computer-executable instructions; The processor executes computer-executable instructions stored in the memory to implement the method of any one of claims 1-8.

Description

Method, device and equipment for identifying potential safety hazards of fuel gas Technical Field The application relates to the technical field of gas safety hazard identification, in particular to a gas safety hazard identification method, device and equipment. Background The fuel gas is used as an important clean energy source and has wide application in household life and industrial production. However, the potential safety hazards such as gas leakage, pipe network corrosion and unexpected equipment faults are highlighted, serious explosion, fire and other malignant accidents can be possibly caused, and the potential safety hazards to public safety and lives and properties of people are greatly threatened. At present, gas safety hidden trouble identification mainly relies on manual inspection, namely security inspection personnel periodically enter a household to inspect a user side gas pipeline, a valve, a meter and gas equipment, and data are acquired through portable detection equipment (such as a combustible gas detector). After the inspection data is obtained, the inspection result is usually recorded by a manual input or a mobile terminal and is matched with random spot checks of management staff to carry out quality control, or the inspection data is intelligently inspected by utilizing the image understanding capability of a multi-mode large language model, so that the inspection result is obtained. However, the accuracy and reliability of the inspection result in the manual quality inspection mode are affected by the professional experience and operation standardization of inspection personnel, the existing spot check proportion is generally low, missing inspection or misjudgment is difficult to effectively avoid, the identification result of the identification scheme using the multi-mode large language model generally only gives out whether hidden danger exists and the type of hidden danger, the judgment basis (i.e. lack of interpretability) cannot be clearly clarified, and the positioning accuracy of the hidden danger is insufficient, so that secondary manual re-inspection is easy to be initiated. Disclosure of Invention The application provides a method, a device and equipment for identifying potential safety hazards of fuel gas, which are used for solving the defects that in the prior art, under the scene of identifying the potential safety hazards of fuel gas, only whether the potential hazards exist and the types of the potential hazards are given, the judgment basis (namely lack of interpretability) cannot be clearly clarified, the positioning accuracy of the potential hazards is insufficient, and secondary manual rechecking is easy to cause. In a first aspect, the present application provides a method for identifying a gas safety hazard, the method comprising: fine tuning the multi-modal large language model by adopting a reinforcement learning framework; performing iterative training on the trimmed multi-modal large language model through a training data set to obtain a potential safety hazard identification model; Acquiring an original image, and performing image preprocessing on the original image to obtain an image to be identified, wherein the size of the image to be identified is matched with a potential safety hazard identification model; inputting the image to be identified into the potential safety hazard identification model to obtain an identification result output by the potential safety hazard identification model, wherein the identification result is used for indicating position information corresponding to the potential safety hazard and an identification result capable of being interpreted and analyzed. In one possible implementation manner, the performing iterative training on the trimmed multi-modal large language model through the training data set to obtain a potential safety hazard identification model includes: The method comprises the steps of obtaining a training data set, wherein the training data set comprises a training image set, a real hidden danger type, real position information and a real judgment basis corresponding to each training image; Inputting the training image set into the finely tuned multi-modal large language model to obtain a prediction recognition result output by the multi-modal large language model, wherein the prediction recognition result comprises a prediction hidden danger type, prediction position information and a prediction judgment basis corresponding to the training image; Optimizing the multi-modal large language model by adopting a multi-modal rewarding function based on the real hidden danger type, the real position information, the real judging basis, the predicted hidden danger type, the predicted position information and the predicted judging basis until the training times of the fine-tuned multi-modal large language model reach the preset times or the loss value corresponding to the predicted recognition result output by th