Environmental fire identification method based on edge algorithm

Abstract

The invention relates to the technical field of image recognition, in particular to an environmental fire recognition method and system based on an edge algorithm, and the method comprises the steps: obtaining a target state model in advance; acquiring a target real-time image by using an external device; loading the target state model, and generating an identification data result; transmitting an identification data result; and circularly acquiring a target real-time image. The system comprises a high-altitude spherical monitor, an image acquisition device, an edge calculation center unit and a data transmission unit. An edge algorithm is introduced to be combined with a target state model, so that real-time analysis of a fire scene picture can be carried out in time, meanwhile, the loss of a boiler unit can be effectively reduced, fire extinguishing and rescue operation can be realized in time, and meanwhile, the information transmission and analysis process of a monitoring end and a ground operation end is optimized.

Description

Technical Field

[0001] The present invention relates to the technical field of image recognition, and in particular to an environmental fire recognition method based on an edge algorithm, and also to a system. Background Art

[0002] At present, the understanding of the fire scene of boiler units in the special working environment of thermal power plants is mainly based on ground reconnaissance by personnel. The technology of high-altitude reconnaissance is still relatively backward and remains at the stage of simple video recording. When a fire occurs, it is impossible to quickly obtain real-time images of the fire location, and thus the fire point and degree cannot be controlled, delaying the rescue and fire-fighting time, resulting in the scrapping of the boiler unit, direct losses of up to millions of yuan, and even endangering lives in serious cases. Summary of the invention

[0003] Through research, the inventors found that the boiler units of thermal power plants have very high requirements for ambient temperature and are also more likely to explode. Therefore, operators must strictly abide by rules and regulations during actual operations. However, in actual operations, the prescribed requirements are often not met, which results in the boiler units being more prone to explosion. When an explosion occurs, the relevant operators are unable to evacuate the scene in time, and the rescue and fire-fighting personnel are unable to quickly determine the real-time situation of the explosion center, resulting in the inability to carry out rescue and fire-fighting operations in a timely manner, which in turn causes subsequent major economic losses and personal safety risks.

[0004] The purpose of the present invention is to provide an environmental fire identification method and system based on edge algorithm. By establishing a target state model and connecting an external high-altitude spherical monitor, and combining the edge algorithm to perform real-time image processing of the fire scene, the technical problem of the prior art that rescue personnel are unable to immediately determine the real-time situation of the fire location after an explosion occurs in a boiler unit in a thermal power plant is solved.

[0005] According to one aspect of the present invention, an environmental fire identification method based on an edge algorithm is provided, S1. a target state model is acquired in advance; S2. a real-time target image is acquired using an external device, wherein the external device is a high-altitude spherical monitor; S3. the target state model is loaded, and the real-time target image acquired by the external device is identified based on the edge algorithm to generate identification data results; S4. the identification data results are transmitted to a rear monitoring terminal for fire extinguishing strategy deployment; S5. the real-time target image is acquired cyclically until the fire extinguishing strategy deployment work is completed.

[0006] In some embodiments, the target state model is a past deflagration scenario model of a boiler unit in a thermal power plant.

[0007] In some embodiments, the target state model pre-acquisition process is: S11. Acquire a target image data set, wherein the image data set includes visible light image features of a boiler unit in a thermal power plant after deflagration and temperature numerical features detected by infrared detection; S12. Train the acquired target image data group to obtain a target state model.

[0008] In some embodiments, the target image data set training process is: S121. Using a lightweight model to train visible light image features and temperature numerical features of infrared detection in the target image data set; S122. Using a heavyweight model to train visible light image features and infrared temperature numerical features in the target image data set; S123. Define the model after training the lightweight model as the training model, and define the model after training the heavyweight model as the training model. Defined as a standardized model; S124. Performing pre-pruning processing on the training model to obtain a pre-pruning model; performing post-pruning processing on the training model to obtain a post-pruning model; S125. Perform model distillation on the pre-pruning model and the post-pruning model together with reference to the standardized model to obtain an optimized model, wherein the optimized model is divided into a first optimized model and a second optimized model, and after comparing the accuracy, an optimized model is obtained; S126. Perform model quantization on the optimization model to obtain a target state model.

[0009] In some embodiments, the comparison accuracy comparison target is the mean square error loss size.

[0010] In some embodiments, the process of training the visible light image features and infrared detected temperature numerical features in the target image data group using a lightweight model is as follows: defining the parameter space of the visible light image features and infrared temperature features of the explosion of the boiler unit in the thermal power plant; determining the parameter network structure from the parameter space based on the basic acquisition strategy; evaluating the parameter network structure, and if the evaluation result meets the preset conditions, determining the candidate parameter network structure that meets the preset conditions as the training model.

[0011] In some embodiments, the process of using a heavyweight model to train visible light image features and infrared detected temperature numerical features in a target image data set is as follows: defining a parameter space of visible light image features and infrared temperature features of a past boiler unit explosion in a thermal power plant; determining a parameter network structure from the parameter space based on a basic acquisition strategy; and evaluating the parameter network structure. If the evaluation result meets a preset condition, the candidate parameter network structure that meets the preset condition is determined as a standardized model.

[0012] In some embodiments, the edge algorithms include: Ppyolo target detection algorithm, Ssd target detection algorithm and Mobilenet-ssd detection algorithm.

[0013] In some embodiments, the acquisition strategy is random search acquisition.

[0014] According to one aspect of the present invention, there is provided a system comprising: High altitude spherical monitor; An image acquisition device is fixedly installed inside the high-altitude spherical monitor to acquire a real-time image of the fire of the boiler unit of the thermal power plant; The edge computing center unit is embedded in the high-altitude spherical monitor, and is used to load the target state model, and to identify the target in the real-time image of the fire of the boiler unit of the thermal power plant in real time according to the target state model and the edge algorithm, and to generate a real-time identification result; The data transmission unit is connected to the edge computing center unit for timely transmission of real-time recognition results.

[0015] Compared with the prior art, the present invention has the following advantages: the present invention introduces edge algorithms and target state models, and can simultaneously perform high-altitude monitoring operations on real-time fire scenes; Based on the above advantages, it has the following beneficial effects: the introduction of edge algorithms combined with target state models can timely conduct real-time analysis of fire scene images, effectively reduce the losses of boiler units, and realize fire fighting and rescue operations in time, while optimizing the information transmission and analysis process between the monitoring end and the ground operation end. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0017] Figure 1 is a flow chart of an environmental fire identification method according to some embodiments of the present invention; Figure 2 is a flowchart of a target state model pre-acquisition process of some embodiments of the present invention; Figure 3 is a flowchart of a target image dataset training process according to some embodiments of the present invention. DETAILED DESCRIPTION

[0018] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. The following description of at least one exemplary embodiment is actually only illustrative and is by no means intended to limit the present invention and its application or use. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0019] This embodiment provides an environmental fire identification method and system based on edge algorithm. The environmental fire identification method and system based on edge algorithm are already in the testing stage and have been put into use in a small-scale factory. Figure 1-3 Give a description.

[0020] Figure 1 This is a flow chart of an environmental fire identification method according to some embodiments of the present invention. The method is divided into five steps in total.

[0021] First, in step S1, the target state model is obtained in advance. Since the method of the present invention is a targeted solution to the problem of thermal power plant boiler units, the target state model is a past deflagration scenario model of thermal power plant boiler units. The model acquisition process is divided into two steps, see the attached Figure 2 , respectively: In step S11, it is necessary to obtain a target image data set, which includes visible light image features after the explosion of the boiler unit in the thermal power plant and temperature numerical features detected by infrared, etc. After obtaining the above information, the image data set is processed by data augmentation and the like. The processing methods include: one or more of image flipping, random translation, random color perturbation, Gaussian noise, and Gaussian blur to increase data diversity to overcome the problem of unbalanced data categories. At the same time, in actual operation, the inventor also uses the smoothlabel method to smooth the labels to improve the generalization ability of the model in order to solve the problem of inaccurate labeling of the data set.

[0022] In step S12, the acquired target image data set is trained to obtain a target state model. In this embodiment, the specific process of training is divided into five steps. Figure 3 In step S121, a lightweight model is used to train the visible light image features in the target image data set and the temperature numerical features detected by infrared. At the same time, in step S122, a heavyweight model is used to train the visible light image features in the target image data set and the temperature numerical features detected by infrared. The lightweight model has a short training time and a poor training effect, while the heavyweight model has a long training time and a good training effect. Different models are obtained through different training. In step S123, the model obtained by training the lightweight model is defined as a training model, and the model obtained by training the heavyweight model is defined as a standardized model. The standardized model is also a teacher model known to those skilled in the art.

[0023] Among them, it should be separately explained that the process of using a lightweight model to train the visible light image features in the target image data group and the temperature numerical features of infrared detection is as follows: define the parameter space of the visible light image features and infrared temperature features of the explosion of the boiler unit in the thermal power plant; based on the basic acquisition strategy, determine the parameter network structure from the parameter space; evaluate the parameter network structure, and if the evaluation result meets the preset conditions, determine the candidate parameter network structure that meets the preset conditions as the training model.

[0024] Similarly, the process of using the heavyweight model to train the visible light image features and infrared temperature numerical features in the target image data set is as follows: define the parameter space of the visible light image features and infrared temperature features of the past boiler unit explosion in the thermal power plant; based on the basic acquisition strategy, determine the parameter network structure from the parameter space; evaluate the parameter network structure, and if the evaluation result meets the preset conditions, determine the candidate parameter network structure that meets the preset conditions as the standardized model.

[0025] When using lightweight and heavyweight models for training, first, the parameters of the parameter space include: one or more of the number of training categories, anchor points, maximum number of iterations, and network input size. The number of training categories is the number of target categories that need to be identified, the anchor point is the size of the training pre-selected box, the maximum number of iterations is the maximum number of iterations in the training phase, and the training is stopped when it is reached. The network input size is the network input size obtained after processing the original image during the training process, which is related to the model training effect. Secondly, the basic acquisition strategy acquisition strategy in this implementation is random search acquisition, of course, it can also be an evolutionary algorithm, a gradient-based algorithm, etc.

[0026] In step S124, the training model is pre-pruned to obtain a pre-pruned model, and the training model is post-pruned to obtain a post-pruned model. Pruning can reduce the training time while ensuring recognition accuracy, and solve the problems of too large models, inability to recognize in real time, and long operation time. It should be noted that the purpose of pruning is to avoid overfitting of the model. In this embodiment, the training model is first pre-pruned to reduce the risk of overfitting, and significantly reduce the training time and test time. Then, the same training model is post-pruned to obtain a model with a lower risk of overfitting and better generalization performance. In this way, two models can be obtained through two pruning processes.

[0027] Then, in step S125, the pre-pruning model and the post-pruning model are distilled together with reference to the standardized model to obtain the optimized model. The optimized model is divided into the first optimization model and the second optimization model. After comparing the accuracy, the optimized model is obtained. Among them, the comparison accuracy comparison target is the mean square error loss size. The smaller the loss, the more optimized the model. At the same time, the above process overcomes the problem of single direction of existing model processing. In the acquisition of the explosion model of the boiler unit of the thermal power plant, if only a single model processing in a certain direction is performed without comparison processing, there will be problems of model selection omission or lack of pertinence. For example, when the power of the boiler unit of a thermal power plant is small, if the post-pruning processing model is used, it will be a fuss and waste of energy, because the pre-pruning operation can solve the problem more quickly, although there is a risk of overfitting, but it takes a short time. On the contrary, if it is a high-power boiler unit of a large thermal power plant, it is necessary to accurately obtain the model, and the post-pruning processing plays an important role at this time. After the pruning process is completed, the actual model distillation operation is performed according to the situation to obtain a suitable optimization model. Model distillation is a technology that compresses complex neural networks into small and fast networks. It can maintain high accuracy while reducing the time of computing resources. This embodiment adopts an adaptive temperature scaling method. The temperature is controlled by an adaptive parameter, and the parameter is adjusted with the training status. This is of great significance for the image recognition of the fire environment of the present invention.

[0028] In step S126, model quantization is performed on the optimization model to obtain a target state model.

[0029] Secondly, in step S2, in some embodiments, an external device is used to obtain a real-time image of the target, wherein the external device is a high-altitude spherical monitor. Of course, in other thermal power plants, the external device can be reasonably selected according to the actual plant structure and installation location, but the selection of high-altitude external equipment is required to be used in conjunction with the method of the present invention.

[0030] Next, in step S3, the target state model is loaded, and the real-time image of the target obtained by the external device is recognized based on the edge algorithm to generate recognition data results. In some embodiments, the edge algorithm includes: Ppyolo target detection algorithm, Ssd target detection algorithm and Mobilenet-ssd detection algorithm, and any one of the above can be selected in this implementation.

[0031] Further, in step S4, the identification data results are transmitted to the rear monitoring end for fire fighting strategy deployment. The transmission process is different from the existing transmission process, and does not require complicated manual analysis, machine recognition, conclusion combination, and transmission analysis, and can directly achieve the effect of rapid identification.

[0032] Finally, in step S5, real-time images of the target are acquired in a loop until the deployment of the fire-fighting strategy is completed.

[0033] In order to better understand the present invention and more clearly explain the usage and usage conditions of the present invention, the present embodiment also provides a system to implement the above-mentioned environmental fire identification method, which specifically includes a high-altitude spherical monitor; an image acquisition device, fixedly installed inside the high-altitude spherical monitor, to obtain real-time images of fires in boiler units of thermal power plants; an edge computing center unit, embedded in the high-altitude spherical monitor, for loading a target state model, and according to the target state model and the edge algorithm, real-time identification of targets in the real-time images of fires in boiler units of thermal power plants, and generating real-time identification results; a data transmission unit, connected to the edge computing center unit, for timely transmission of real-time identification results.

[0034] In summary, the present invention realizes accurate identification of fire conditions of boiler units in thermal power plants by utilizing edge algorithms, learning algorithms and target state models, with rapid feedback, and can perform real-time dynamic monitoring of targets. At the same time, it combines visible light features and infrared temperature features to greatly improve the accuracy of target identification.

[0035] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0036] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for environmental fire identification based on edge algorithm, characterized in that: S1. Pre-acquire the target state model; S2. Using an external device to obtain a real-time image of the target, wherein the external device is a high-altitude spherical monitor; S3. Load the target state model, identify the real-time target image obtained by the external device based on the edge algorithm, and generate the identification data result; S4. Transmit the identification data results to the rear monitoring end for fire fighting strategy deployment; S5. Loop to obtain real-time images of the target until the fire-fighting strategy deployment is completed.

2. The environmental fire identification method according to claim 1, characterized in that: The target state model is a past deflagration scenario model of a boiler unit in a thermal power plant.

3. The environmental fire identification method according to claim 2, characterized in that: The target state model pre-acquisition process is: S11. Acquire a target image data set, wherein the image data set includes visible light image features of a boiler unit in a thermal power plant after deflagration and temperature numerical features detected by infrared detection; S12. Train the acquired target image data group to obtain a target state model.

4. The environmental fire identification method according to claim 3, characterized in that: The target image data set training process is: S121. Using a lightweight model to train visible light image features and temperature numerical features of infrared detection in the target image data set; S122. Using a heavyweight model to train visible light image features and infrared temperature numerical features in the target image data set; S123. Define the model after training the lightweight model as the training model, and define the heavyweight model as the training model. The trained model is defined as the standardized model; S124. Performing pre-pruning processing on the training model to obtain a pre-pruning model; performing post-pruning processing on the training model to obtain a post-pruning model; S125. Perform model distillation on the pre-pruning model and the post-pruning model together with reference to the standardized model to obtain an optimized model, wherein the optimized model is divided into a first optimized model and a second optimized model, and after comparing the accuracy, an optimized model is obtained; S126. Perform model quantization on the optimization model to obtain a target state model.

5. The environmental fire identification method according to claim 4, characterized in that: The comparison accuracy comparison target is the mean square error loss size.

6. The environmental fire identification method according to claim 4, characterized in that: The process of using the lightweight model to train the visible light image features and the temperature numerical features detected by infrared in the target image data set is as follows: defining the parameter space of the visible light image features and the infrared temperature features of the past boiler unit deflagration in the thermal power plant; Determining a parameter network structure from the parameter space based on a basic acquisition strategy; The parameter network structure is evaluated, and if the evaluation result meets the preset conditions, the candidate parameter network structure that meets the preset conditions is determined as the training model.

7. The environmental fire identification method according to claim 4, characterized in that: The process of using the heavyweight model to train the visible light image features and the temperature numerical features detected by infrared in the target image data set is as follows: defining the parameter space of the visible light image features and the infrared temperature features of the past boiler unit deflagration in the thermal power plant; Determining a parameter network structure from the parameter space based on a basic acquisition strategy; The parameter network structure is evaluated, and if the evaluation result meets the preset conditions, the candidate parameter network structure that meets the preset conditions is determined as the standardized model.

8. The environmental fire identification method according to claim 1, characterized in that: The edge The algorithms include: Ppyolo target detection algorithm, Ssd target detection algorithm and Mobilenet-ssd detection algorithm.

9. The environmental fire identification method according to any one of claims 6 or 7, characterized in that: The acquisition strategy is random search acquisition.

10. A system for implementing the environmental fire identification method according to any one of claims 1 to 9, characterized in that: include: High altitude spherical monitor; An image acquisition device is fixedly installed inside the high-altitude spherical monitor to acquire a real-time image of the fire of the boiler unit of the thermal power plant; The edge computing center unit is embedded in the high-altitude spherical monitor, and is used to load the target state model, and to identify the target in the real-time image of the fire of the boiler unit of the thermal power plant in real time according to the target state model and the edge algorithm, and to generate a real-time identification result; The data transmission unit is connected to the edge computing center unit for timely transmission of real-time recognition results.

Images