Detailed Description
In order to make the objects, technical solutions and advantages of the embodiments of the present invention more clear, the technical solutions of 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.
It will be apparent that the described embodiments are some, but not all, embodiments of the invention.
The components of the embodiments of the present invention generally described and illustrated in the figures herein may be arranged and designed in a wide variety of different configurations.
The following describes specific embodiments of the present invention in detail with reference to the drawings.
In order to solve the foregoing technical problems in the background art, fig. 1 is a schematic flow chart of a fire spread simulation model training method based on big data according to an embodiment of the present disclosure, and the following describes the fire spread simulation model training method based on big data in detail.
Step S201, a historical fire spread data set and a real-time fire spread data set are obtained, wherein the historical fire spread data set comprises historical fire spread data, and the real-time fire spread data set comprises real-time fire spread data;
Step S202, a pre-training model is built according to at least two fire feature learning networks, the at least two fire feature learning networks are trained by the historical fire spread data, the pre-training model comprises an independent learning network and a joint learning network, and the independent learning network comprises a first independent learning network and a second independent learning network which are consistent in architecture and operate simultaneously; wherein the first individual learning network is used for performing feature coding on the historical fire spread data, and the second individual learning network is used for performing feature coding on the real-time fire spread data; the joint learning network is used for carrying out integrated analysis on the characteristic coding results of the independent learning network;
Step S203, performing training operation on the pre-training model through the historical fire spread data and the real-time fire spread data to obtain a fire spread simulation model, where the fire spread simulation model is used to refer to a fire spread simulation stage of the real-time fire spread data after optimizing learning strategies of at least two fire feature learning networks corresponding to the historical fire spread data set.
In the present embodiment, it is assumed, by way of example, that fire departments in one city collect a large amount of historical fire spread data through fire reports recorded over the years, data reported by firefighters, and sensor networks installed in buildings and public areas.
The data includes information such as the start time, place, spreading speed, spreading route, etc. of the fire.
Meanwhile, various data of the fire disaster which is currently happening, such as current fire intensity, wind direction, air temperature and the like, are acquired through the real-time sensor network.
To construct the pre-training model, two fire feature learning networks may be designed.
The first learning network is specially used for carrying out feature coding and learning on historical fire spread data, for example, extracting key features such as fire intensity, combustion object type, combustion temperature and the like.
The second learning network is focused on feature coding of real-time fire spread data, such as current intensity of fire, wind direction, air temperature, etc.
The two learning networks have the same architecture and operate simultaneously to extract and learn key fire features.
These data are transmitted into a pre-training model for training operations using both the historical fire spread data set and the real-time fire spread data set.
The model predicts the spread of a fire based on features in the historical data, such as learning the spread rate of a certain combustion substance under certain weather conditions based on past fire data.
The model performance is optimized by continuously adjusting the weight and parameters of the model, so that the model can simulate the spreading process of fire more accurately.
Through training operation, a fire spreading simulation model is obtained.
The model comprehensively utilizes the learning strategy of the fire feature learning network in the historical fire spread data set, and optimizes by combining the real-time fire spread data.
In a real-time fire spread simulation stage, the model can predict the spread condition of the fire according to the current real-time data.
For example, based on the feature weights learned in the past, the model can predict the path and speed of fire spread by combining the current wind direction, air temperature and other real-time data, and provide accurate early warning information and decision support for firefighters.
In the embodiment of the present invention, the foregoing step S202 may be implemented by the following example execution.
(1) According to the sequence relation of the at least two fire feature learning networks, taking the fire feature learning network at the last position as a starting position, and sequentially selecting at least one fire feature learning network to obtain the joint learning network;
(2) Determining fire feature learning networks other than the joint learning network of the at least two fire feature learning networks as pending individual learning networks;
(3) Determining the undetermined individual learning network as the first individual learning network or the second individual learning network, and establishing the individual learning network.
In the embodiment of the invention, three fire feature learning networks, A, B and C respectively, are assumed as an example.
And selecting the last C as a starting position according to the sequence relation, and sequentially selecting one or more fire feature learning networks to obtain a joint learning network.
Thus, C may be chosen as the joint learning network.
The fire disaster feature learning network is a network model for coding and learning features in the historical fire disaster spread data set through a machine learning technology.
These characteristics may include intensity of fire, type of combustion, rate of spread, etc.
In the former scenario, C is taken as the joint learning network.
Then a and B other than C will be determined to be pending individual learning networks.
The undetermined individual learning network refers to other fire feature learning networks in the pre-training model than the joint learning network.
These networks are further determined in subsequent steps and used to build individual learning networks.
Let a be determined as a first individual learning network and B be determined as a second individual learning network.
Thus, two separate learning networks, a and B, respectively, are established.
The individual learning network refers to a network model that performs feature coding and learning on the historical fire spread data set and the real-time fire spread data set, respectively.
The first individual learning network is used for feature encoding historical fire spread data, and the second individual learning network is used for feature encoding real-time fire spread data.
In the embodiment of the invention, when the undetermined individual learning network is constructed as the first individual learning network, the second individual learning network is acquired by cloning by the first individual learning network; or when the pending individual learning network is constructed as the second individual learning network, a first individual learning network of the individual learning networks is acquired by cloning by the second individual learning network.
In the embodiment of the present invention, it is assumed, by way of example, that the first individual learning network a and the pending individual learning network B have been constructed.
According to this step, the second individual learning network B can be acquired by cloning the first individual learning network a such that both individual learning networks have the same structure and parameters.
Cloning refers to replicating an existing network model such that the new network has the same structure and parameters as the original network.
In this case, the second individual learning network B is obtained from the first individual learning network a by cloning.
It is assumed that the second individual learning network B and the pending individual learning network a have been constructed.
According to this step, the first individual learning network a of the individual learning networks can be acquired by cloning the second individual learning network B so that both individual learning networks have the same structure and parameters.
In this case, the first individual learning network a is obtained from the second individual learning network B in a clonal manner such that both have the same feature learning capability.
In the embodiment of the present invention, the step of sequentially selecting at least one fire feature learning network to obtain the joint learning network by using the fire feature learning network located at the last position as the starting position may be implemented by the following example.
(1) Setting model parameters of a later preset number of fire feature learning networks in the at least two fire feature learning networks as constant values according to the sequence relation of the at least two fire feature learning networks;
(2) Based on the real-time fire spread data in the real-time fire spread data set, training the fire feature learning networks except for the preset number of fire feature learning networks after being excluded from the at least two fire feature learning networks to obtain a training fire spread result;
(3) And determining the combined learning network constructed by a preset number of fire feature learning networks in response to the error value between the training fire spread result and the marked real-time fire spread expected result corresponding to the real-time fire spread data meeting a preset condition.
In an embodiment of the present invention, five fire feature learning networks A, B, C, D and E are assumed for exemplary purposes.
And selecting E at the last position as a starting position according to the sequence relation, and setting the preset number to be 2.
According to this step, the model parameters of C and D can be set to constant values, i.e. no training is performed.
Model parameters refer to adjustable parameters in the fire signature learning network, also referred to as weights.
The learning capacity of the model parameter of the part of fire characteristic learning network can be fixed by setting the model parameter as a fixed value, so that the follow-up operation is convenient.
In the previous scenario, the two fire feature learning networks, C and D, were excluded.
Now, training operations are performed on A, B and E using data in the real-time fire spread data set to obtain training fire spread results.
The training operation refers to the use of the marked data set to adjust the model parameters of the fire feature learning network so that the real-time fire spread result can be predicted better.
Through such training operations, the learning ability of the fire feature learning network can be improved.
During the training process, error values between the training fire spread results and actual fire spread expected results marked in the real-time fire spread data set are compared.
If the error value meets a preset condition, i.e., the expected accuracy requirement is met, then a joint learning network constructed from a later preset number of fire feature learning networks is determined.
The error value refers to a measure of the difference between the training fire spread result and the actual fire spread expected result.
In this step, if the error value satisfies a preset condition, i.e., the accuracy requirement is met, a joint learning network constructed from a later preset number of fire feature learning networks is confirmed.
In an embodiment of the present invention, the pre-training model further includes a category identifier for identifying a spread trend type of the historical fire spread data and the real-time fire spread data; the aforementioned step S203 may be implemented by the following example execution.
(1) Determining historical fire spread data cost parameters corresponding to the historical fire spread data and real-time fire spread data cost parameters corresponding to the real-time fire spread data through the pre-training model by the historical fire spread data and the real-time fire spread data;
(2) Based on a class identifier in the pre-training model, judging the type of the spread trend of the historical fire spread data and the real-time fire spread data, and determining a class identification error value;
(3) And executing training operation on the pre-training model according to the historical fire spread data cost parameter, the real-time fire spread data cost parameter and the class identification error value to obtain the fire spread simulation model.
In an exemplary embodiment of the present invention, a category identifier in the pre-trained model may be used to identify the type of spread trend for both historical fire spread data and real-time fire spread data.
For example, by analyzing past fire data and fire data acquired in real time, the category identifier may determine different types of fire spread trend, such as rapid spread, slow spread, etc.
The category identifier refers to adding a component or a sub-network to the pre-training model to classify the input data and determine the type of the spread trend to which the input data belongs.
This class identifier may help to further understand and predict the trend of fire spread.
In this step, a training operation is performed using the historical fire spread data and the real-time fire spread data, thereby obtaining a fire spread simulation model.
By inputting the data into the pre-training model, cost parameters corresponding to the historical fire spread data and the real-time fire spread data can be determined, and spread trend type judgment and error calculation can be performed based on the category identifier.
The fire spread simulation model refers to a model which is obtained through training operation and can simulate the fire spread process.
By training using historical data and real-time data, the model can predict and simulate the spreading trend of the fire under different conditions and generate corresponding cost parameters.
In the step, training operation is carried out according to cost parameters corresponding to the historical fire spread data and the real-time fire spread data and the class identification error value, so that a final fire spread simulation model is obtained.
By adjusting parameters and weights of the pre-training model, the performance of the model can be optimized, so that the fire spread situation can be predicted better.
In this step, training operation is performed on the pre-training model according to cost parameters and class identification error values of the historical data and the real-time data, so as to optimize the performance of the model and enable the model to more accurately simulate the fire spreading process.
In the embodiment of the present invention, the step of performing training operation on the pre-training model according to the historical fire spread data cost parameter, the real-time fire spread data cost parameter and the class identification error value to obtain the fire spread simulation model may be implemented by the following example.
(1) Multiplying the class identification error value with a preset negative parameter to obtain a return cost parameter for carrying out error return;
(2) According to the return cost parameter, performing a first optimization operation on the parameter corresponding to the pre-training model, and determining a parameter updating value;
(3) And executing a second optimization operation on the pre-training model according to the historical fire spread data cost parameter, the real-time fire spread data cost parameter and the parameter updating value to obtain the fire spread simulation model.
In the embodiments of the present invention, it is assumed, by way of example, that a fire spread simulation model is used to predict the development of a fire.
By simulating actual fire data, a set of prediction results are generated and class identification error values are calculated.
In order to further optimize the model performance, the class identification error value is multiplied by a preset negative parameter, so that a return cost parameter for carrying out error return is obtained.
The return cost parameter is obtained by multiplying the class identification error value with a preset negative parameter.
This return cost parameter is used to measure the error of the current model in the subsequent optimization operations and to guide the updating of the parameters.
In this step, the first optimization operation of the model is performed using the backhauled cost parameter.
The parameter updating value can be determined by analyzing the returned cost parameter and the current parameter of the model, so that the performance of the model is improved.
This optimization operation may involve using gradient descent or the like algorithms to adjust parameters of the model.
The first optimization operation is to update parameters in the pre-training model according to the return cost parameters so that the model is better adapted to historical fire spread data and real-time fire spread data.
By calculating the parameter update values, the model can be guided to adjust to a more accurate direction and errors can be reduced.
In this step, the historical fire spread data cost parameter, the real-time fire spread data cost parameter and the previously calculated parameter update value are considered.
By inputting these parameters into the pre-trained model, a second optimization operation is performed, further improving model performance.
This operation may involve calculating a loss function and adjusting model parameters using a suitable optimization algorithm to ultimately arrive at a simulated model of the fire spread.
The second optimization operation is to further adjust the pre-training model by using the historical fire spread data cost parameters, the real-time fire spread data cost parameters and the calculated parameter update values.
Therefore, the model can be better fit with historical data and real-time data, and the accuracy of fire spread simulation is improved.
In the embodiment of the present invention, the step of performing the second optimization operation on the pre-training model according to the historical fire spread data cost parameter, the real-time fire spread data cost parameter and the parameter update value to obtain the fire spread simulation model may be performed by the following example.
(1) Optimizing the pre-training model according to the historical fire spread data cost parameter, the real-time fire spread data cost parameter and the parameter updating value to obtain a to-be-determined fire spread simulation model;
(2) And executing a second optimization operation on the model weight of the undetermined fire spread simulation model based on the real-time fire spread data in the real-time fire spread data set to obtain the fire spread simulation model.
In an exemplary embodiment of the present invention, consider a fire spread simulation system, where a second optimization operation has been performed on the pre-trained model by performing the foregoing steps.
The model is now further optimized using historical fire spread data cost parameters, real-time fire spread data cost parameters, and parameter update values.
And obtaining a pending fire spreading simulation model by adjusting parameters and weights of the model.
The undetermined fire spread simulation model is obtained after the second optimization operation, is a model which is further optimized, and uses historical fire spread data cost parameters, real-time fire spread data cost parameters and parameter updating values.
There is now a data set containing real-time fire spread data.
These data are input into a model of the simulation of the spread of the fire to be determined and used to further optimize the model.
And obtaining a fire spreading simulation model after the second optimization operation by adjusting the model weight.
The purpose of this step is to perform a second optimization operation on the model weights of the pending fire spread simulation model using the real-time fire spread data.
By adjusting the model weight by using real-time data, the model can be better adapted to the current fire situation, thereby obtaining the final fire spread simulation model.
In an embodiment of the present invention, the foregoing step of performing a second optimization operation on the model weights of the pending fire spread simulation model based on the real-time fire spread data in the real-time fire spread data set to obtain the fire spread simulation model may be performed by the following example.
(1) Setting a network weight corresponding to a joint learning network in the to-be-determined fire spread simulation model as a fixed value;
(2) Optimizing network weights corresponding to the independent learning networks in the undetermined fire spread simulation model based on the real-time fire spread data in the real-time fire spread data set, and determining the optimized independent learning networks;
(3) And forming the optimized independent learning network and the combined learning network into the fire spread simulation model.
In the embodiment of the invention, a fire spread simulation model based on a neural network is assumed as an example, wherein a joint learning network and a separate learning network are included.
In this scenario, the network weights corresponding to the joint learning network in the pending fire spread simulation model are set to initial values or fixed values that have been optimized.
The aim of this step is to set the network weights of the joint learning networks in the model of the fire spread simulation to a certain value, so that only the optimization of the individual learning networks is concerned in the following optimization process.
In the above scenario, real-time data in the real-time fire spread data set is used as input and only a separate learning network in the pending fire spread simulation model is of interest.
By adjusting and optimizing the network weights corresponding to the individual learning networks, the network can be improved for the current real-time fire situation.
According to an optimization algorithm (such as gradient descent), the weights of the individual learning networks are updated to better adapt to the fire situation in the real-time dataset.
The goal of this step is to optimize the weights of the individual learning networks by using real-time fire spread data.
By adjusting and optimizing the network for fire information in the real-time data set, accuracy and adaptability of the model under the current fire condition can be improved.
In the previous step, the individual learning networks in the simulated model of fire spread to be determined have been optimized.
The optimized individual learning network and the joint learning network are combined to form a final fire spread simulation model.
The model fully utilizes the advantages of the independent learning network and the combined learning network to provide accurate fire spread prediction and simulation results.
The method aims at combining the optimized independent learning network with the original joint learning network to form a final fire spread simulation model.
By integrating the results of the two networks, the model can more fully take into account aspects of fire spread and provide more accurate simulation results.
In the embodiment of the invention, the historical fire spread data is correspondingly marked with a historical fire spread expected result, and the real-time fire spread data is correspondingly marked with a real-time fire spread expected result; the step of determining the historical fire spread data cost parameter corresponding to the historical fire spread data and the real-time fire spread data cost parameter corresponding to the real-time fire spread data by using the pre-training model according to the historical fire spread data and the real-time fire spread data cost parameter corresponding to the real-time fire spread data can be implemented by the following example.
(1) Inputting the historical fire spread data and the real-time fire spread data into the pre-training model, and determining a historical fire spread output result corresponding to the historical fire spread data and a real-time fire spread output result corresponding to the real-time fire spread data;
(2) Determining historical fire spread data cost parameters corresponding to the historical fire spread data based on a deviation state between the historical fire spread output result and the historical fire spread expected result;
(3) And determining real-time fire spread data cost parameters corresponding to the real-time fire spread data based on the deviation state between the real-time fire spread output result and the real-time fire spread expected result.
In the present embodiment, it is assumed, by way of example, that there is a fire spread data set that includes both historical and real-time fire spread conditions.
In this scenario, each sample in the historical dataset is annotated with its corresponding historical fire spread expected result, while each sample in the real-time dataset is annotated with its corresponding real-time fire spread expected result.
The purpose of this step is to provide a prediction outcome annotation for each sample in the historical and real-time fire spread data sets, respectively.
The expected results of the historical and real-time data are annotated to measure the difference between the model output and the true value.
In the previous step, a pre-trained model has been performed.
Now, historical and real-time fire spread data is entered into the pre-trained model.
And determining the historical fire spread output result corresponding to the historical fire spread data and the real-time fire spread output result corresponding to the real-time fire spread data according to the output result of the model.
The goal of this step is to predict historical and real-time fire spread data using a pre-trained model.
The deviation between the history and real-time data and the corresponding expected results can be calculated by inputting the data and obtaining the output results through the model, so that the history fire spread data cost parameters corresponding to the history fire spread data and the real-time fire spread data cost parameters corresponding to the real-time fire spread data are determined.
In the previous step, the historical fire output results and the historical fire expected results have been obtained by a pre-training model.
From these data, historical fire spread data cost parameters corresponding to the historical fire data can be calculated.
This cost parameter may reflect the degree of difference between the historical fire data and its expected outcome.
The method comprises the steps of determining historical fire spread data cost parameters corresponding to historical fire spread data according to deviation states between historical fire output results and historical fire expected results.
By comparing the predicted result and the actual expected result of the historical fire data, the cost parameter of the historical fire data can be calculated.
This cost parameter may be used to measure the magnitude of the error or deviation between the historical fire data and its expected outcome.
In the above scenario, real-time fire output results and real-time fire expected results have been obtained from a pre-trained model.
By comparing the differences between the two, the real-time fire spread data cost parameter corresponding to the real-time fire data can be calculated.
This cost parameter reflects the degree of error between the real-time fire data and its expected outcome.
The goal of this step is to determine a cost parameter corresponding to the real-time fire spread data based on the deviation status between the real-time fire output result and the real-time fire expected result.
By calculating the difference between the real-time data and the expected result, the accuracy and reliability of the real-time fire data can be assessed.
In an embodiment of the present invention, after the foregoing step S201, the method further includes the following examples.
(1) Inputting the historical fire spread data in the historical fire spread data set into at least two undetermined fire feature learning networks which are arranged continuously, and determining a historical fire spread data processing result corresponding to the historical fire spread data;
(2) And based on the error between the historical fire spread data processing result and the historical fire spread expected result corresponding to the historical fire spread data, training the undetermined fire feature learning network to obtain the at least two fire feature learning networks which are arranged continuously.
In the present embodiment, it is assumed that there is a historical fire spread data set that contains fire conditions over the past few years.
These historical fire spread data are passed as input to a model consisting of at least two pending fire signature learning networks.
Each network is responsible for extracting different fire spread characteristics, such as wind direction, wind speed, topography, etc.
The output generated by these models can represent the result of processing the historical fire spread data through the calculation and processing of the models.
This step utilizes a pending fire signature learning network to process historical fire spread data.
Each network learns and extracts characteristic information related to the spread of fire, thereby generating a processing result of historical fire spread data.
A network of successive permutations may capture more complex feature relationships through different levels of abstraction.
In the preceding step, the processing result of the historical fire data has been acquired through the pending fire characteristics learning network.
These treatment results are now compared with the corresponding historical fire spread expectations and the error between them is calculated.
For example, a Mean Square Error (MSE) or other suitable indicator may be used to measure the difference between the actual output result and the expected result.
The goal of this step is to reduce the error between the historical fire data processing results and the historical fire spread expected results through training operations such as back propagation.
The parameters of the model are optimized by adjusting the weight and the bias of the undetermined fire feature learning network, so that the historical fire spread situation can be predicted more accurately, and the performance of the model is improved.
In the embodiment of the invention, the fire spread simulation model is used for judging the spread trend type of the image; after obtaining the fire spread simulation model, the embodiments of the present invention also provide the following examples.
(1) Acquiring key fire image data from the real-time fire spread data set;
(2) Inputting the key fire image data into the fire spread simulation model, and judging the spread trend type of the key fire image data by the fire spread simulation model to obtain the fire spread trend type corresponding to the key fire image data;
(3) And determining the type of the fire spreading trend as a current fire spreading simulation result corresponding to the key fire image data.
In the embodiment of the present invention, it is assumed, by way of example, that there is a fire spread simulation model which is trained to determine the trend type of the fire spread from the inputted image data.
For example, the model may distinguish between conditions of spread, fade, or leveling of a fire in the image and divide it into different types of spread trends, such as fast spread, slow spread, or steady state, etc.
The goal of this step is to make a determination of the type of spread trend on the image using a fire spread simulation model.
The model has the capability of analyzing and classifying the image data through learning and training, so that the type of the fire spread trend displayed by the image can be judged.
It is assumed that a real-time fire spread data set is provided, which contains fire image data at various moments.
These images may come from a surveillance camera, satellite image, or other fire monitoring system.
The key fire image refers to an image having an important meaning in the current analysis task, for example, an image representing a significant change in fire or an image of a specific area.
The goal of this step is to select and obtain key fire image data from the real-time fire spread data set that is relevant to the current analysis task.
These image data will be used for subsequent fire spread trend type decisions.
And transmitting the selected key fire image data to a fire spread simulation model for analysis.
The model identifies the corresponding type of the fire spreading trend according to the fire behavior in the image.
For example, the model may determine that certain key fire image data represents a fast-propagating fire, a slow-propagating fire, or a steady-state fire.
The goal of this step is to analyze and determine the images from the model by inputting the critical fire image data into a simulated model of the fire spread.
The model can infer the type of the fire spreading trend represented by the image according to the characteristics and the fire behavior in the image.
Through the previous steps, the type of trend of fire spread (such as rapid spread) corresponding to the key fire image data has been obtained.
This type of spread trend is now determined as the current fire spread simulation result corresponding to the key fire image data.
This means that, based on the determination of the model, it can be inferred that the type of trend of the spread of the fire represented by the image matches the actual spread of the fire.
The goal of this step is to determine the type of the fire spread trend determined by the model as the current fire spread simulation result corresponding to the key fire image data.
Through the step, the image can be analyzed by using the model, and the result is compared with the actual fire spread condition, so that the current fire spread simulation result is obtained.
In the embodiment of the present invention, the following examples are also provided.
(1) Acquiring a current fire spread simulation result of a target area and current fire control scheduling information aiming at the current fire spread simulation result;
(2) Determining fire-fighting related features of the target area based on past fire-fighting information of the target area, wherein the past fire-fighting information comprises past fire spread simulation results of the target area and past fire-fighting scheduling information aiming at the past fire spread simulation results;
(3) And determining a scheduling decision score based on the fire control related characteristics of the target area, the current fire spread simulation result and the current fire control scheduling information, wherein the scheduling decision score is used for representing the adaptation condition between the current fire control scheduling information and the target area.
In the embodiment of the invention, a fire spread simulation system is assumed, and the system can simulate and generate the current fire spread simulation result in the target area in real time.
Meanwhile, the current fire control dispatching information formulated for the current fire spread simulation results, such as dispatching fire-fighting teams, indicating evacuation routes and the like, can be obtained.
The goal of this step is to obtain the current fire spread simulation result of the target area and the current fire control schedule information corresponding thereto.
By obtaining this information, further analysis and decision making may be performed.
In the past fire event, past fire information of the target area is accumulated, including past fire spread simulation results and corresponding past fire schedule information.
By analyzing such information, fire-related characteristics of the target area, such as a fire spread rate, a fire spread path, and fire measures effective in different situations, etc., can be determined.
The goal of this step is to utilize past fire information for the target area to determine relevant fire related features.
These features may provide historical data and experience regarding fire spread and fire dispatch at the target area, facilitating subsequent dispatch decisions.
The fire control associated characteristics of the target area, the current fire spread simulation result and the current fire control scheduling information are combined, so that a scheduling decision score can be calculated.
This score is used to represent the fit between the current fire schedule information and the target area.
For example, if the current fire spread simulation results show rapid spread of fire and the current fire dispatch information contains emergency measures such as rapid dispatch of personnel and indication of evacuation routes, the dispatch decision score may be high.
The goal of this step is to determine a dispatch decision score based on the fire-related characteristics of the target area, the current fire spread simulation results, and the current fire dispatch information.
This score may help evaluate the degree of match between the current fire dispatch measure and the fire spread situation in the target area.
In another implementation of the present embodiment, it is assumed that an emergency situation of a large factory fire is being handled.
Fire spread simulation software is used to simulate and predict fire spread in different areas of a factory in real time.
At the same time, current fire dispatch information corresponding to these simulation results is also obtained, such as dispatching fire teams to specific areas, instructing employees to evacuate along evacuation routes, etc.
The goal of this step is to obtain the current fire spread simulation result of the target area and the current fire control schedule information corresponding thereto.
By obtaining this information, the current fire situation and related countermeasures can be evaluated in a subsequent analysis and decision process.
Consider a region in a city where historically excessive fire events have occurred.
By analyzing these past fire information, fire-related features associated with the target area can be identified.
For example, based on past fire spread simulation results and corresponding fire control schedule information, the rate of fire spread in a particular area, the location of a building susceptible to impact, past effective fire control measures, etc. may be determined.
The goal of this step is to determine fire-related features associated therewith by analyzing past fire information for the target area.
These features may provide historical data and experience regarding fire spread and fire dispatch in the target area, thereby providing a basis for subsequent dispatch decisions.
The scheduling decision score can be calculated by combining fire control related features of the target area, the current fire spread simulation result and the current fire control scheduling information.
This score is used to represent the fit between the current fire schedule information and the target area.
For example, if the current fire spread simulation results show rapid spread of fire and the current fire dispatch information contains emergency measures such as rapid dispatch of personnel and indication of evacuation routes, the dispatch decision score may be high.
The goal of this step is to determine a dispatch decision score using the fire-related characteristics of the target area, the current fire spread simulation results, and the current fire dispatch information.
By calculating this score, the degree of match between the current fire dispatch measure and the fire spread situation in the target area can be evaluated, thereby providing guidance for decision making.
The relationship of each step of this embodiment can be further explained by integrating the above scenarios:
Acquiring a current fire spread simulation result of a target area and current fire control scheduling information aiming at the current fire spread simulation result:
The fire spread simulation result in the target area is obtained through the real-time fire spread simulation system, and the current fire control scheduling information is combined, including dispatching fire control teams, indicating evacuation routes and the like.
This step provides the current fire situation and corresponding countermeasures.
Based on past fire control information of the target area, determining fire control associated features of the target area:
fire-related features of the target area, such as fire spread rate, vulnerable building location, and effective fire protection measures, are identified by analyzing past fire spread simulation results and related fire dispatch information.
These features provide historical data and experience for subsequent scheduling decisions.
Determining a scheduling decision score based on fire control correlation characteristics of the target area, a current fire spread simulation result and current fire control scheduling information: and calculating a scheduling decision score by combining fire control related characteristics of the target area, a current fire spread simulation result and current fire control scheduling information, wherein the scheduling decision score is used for measuring the adaptation condition between the current fire control scheduling information and the target area.
This score may help evaluate the effectiveness of the current scheduling measure and provide guidance for decision making.
Through the association of the three steps, the current fire spread simulation result, fire control scheduling information and past fire control experience can be comprehensively considered, so that the accuracy and the effectiveness of fire control scheduling are ensured.
Such an approach may help better address fire emergencies and provide reasonable scheduling decisions.
In the embodiment of the present invention, the step of determining the fire-fighting related characteristics of the target area based on the past fire-fighting information of the target area may be implemented by the following example.
(1) Determining a plurality of corresponding target fire spread simulation nodes of the target area in a fire spread state diagram structure based on past fire spread simulation results of the target area, wherein the fire spread state diagram structure is used for representing correlation among the plurality of fire spread simulation nodes;
(2) Determining a plurality of target fire control scheduling schemes corresponding to the target area in a fire control scheduling strategy diagram structure based on past fire control scheduling information of the target area, wherein the fire control scheduling strategy diagram structure is used for representing correlation among the plurality of fire control scheduling schemes;
(3) And determining fire-fighting related characteristics of the target area based on the target fire spread simulation nodes and the target fire-fighting scheduling schemes.
In the present embodiment, it is assumed, by way of example, that a residential fire emergency in a city is being handled.
By analyzing past fire spread simulation results of the area, the fire spread condition at a plurality of different time points can be identified.
These different time points form a fire spread status diagram structure, wherein the nodes represent the fire spread conditions at different time points.
Based on past fire spread simulation results of the target area, a plurality of corresponding target fire spread simulation nodes of the target area in the fire spread state diagram structure can be determined.
These nodes represent the spread of fire at different moments, helping to understand the trend and pattern of fire.
Consider a fire emergency situation in a business center.
By analyzing past fire dispatch information for the area, a number of different fire dispatch schemes can be identified, such as dispatching a different number and type of fire fighters, using different fire extinguishing equipment, and the like.
These schemes form a fire dispatch strategy graph structure in which nodes represent different fire dispatch schemes.
Based on past fire control scheduling information of the target area, a plurality of target fire control scheduling schemes corresponding to the target area in the fire control scheduling strategy diagram structure can be determined.
These schemes represent different scheduling measures taken in the past for fire emergency situations, helping to understand the effectiveness and feasibility of fire scheduling.
Assume that fire spread simulation nodes at different time points and different fire control scheduling schemes are comprehensively considered.
It is found that at some point in time, the fire spreads at the fastest rate, and the corresponding fire dispatch scheme is to dispatch a large number of firefighters and use large fire extinguishing equipment.
Such a combination may be considered a fire-fighting-related feature of the target area.
Based on the plurality of target fire spread simulation nodes and the plurality of target fire dispatch schemes, fire-fighting correlation characteristics of the target area may be determined.
These features represent correlations between the spread of fire at different points in time and the corresponding fire schedule, which helps to understand the impact of different factors on fire spread and countermeasures.
Thus, based on the past fire spread simulation result and the fire control scheduling information, a fire spread state diagram structure and a fire control scheduling strategy diagram structure can be established, and fire control related characteristics of a target area can be determined according to the structures.
These features can be used to analyze the trend of fire spread, evaluate the effectiveness of fire dispatch measures, and provide guidance and basis for emergency decisions.
In the embodiment of the present invention, the step of determining the fire-related characteristics of the target area based on the plurality of target fire spread simulation nodes and the plurality of target fire-fighting scheduling schemes may be implemented by the following example execution.
(1) And executing integration operation on the fire disaster simulation feature vectors and the fire fighting scheduling feature vectors to obtain fire fighting correlation features of the target area, wherein the fire disaster simulation feature vectors are features corresponding to the target fire disaster spread simulation nodes, and the fire fighting scheduling feature vectors are features corresponding to the target fire fighting scheduling scheme.
In the embodiment of the present invention, it is assumed, by way of example, that a fire emergency in a city is being handled, and in the process, respective eigenvectors have been obtained by analyzing a plurality of target fire spread simulation nodes and a plurality of target fire control schedule schemes.
For example, for each target fire spread simulation node, there are characteristics of fire spread speed, fire source position, and the like; for each target fire-fighting scheduling scheme, the characteristics of dispatch of the number of fire-fighting teams, the type of equipment used and the like are provided.
By integrating the feature vectors, the fire-fighting related features of the target area can be obtained.
Performing an integration operation on the plurality of fire simulation feature vectors and the plurality of fire dispatch feature vectors means that the relevant features in the different feature vectors are combined to obtain a more comprehensive and comprehensive fire-fighting correlation feature.
These features reflect the correlation between the fire spread simulation node and the fire dispatch scheme, helping to better understand the fire situation and corresponding fire measures.
In the embodiment of the present invention, before the above-mentioned integrating operation is performed on the plurality of fire simulation feature vectors and the plurality of fire control dispatch feature vectors to obtain the fire control related features of the target area, the following implementation manner is further provided.
(1) Performing a graph structure convolution operation on the fire spread state graph structure to obtain characteristics of a plurality of fire spread simulation nodes in the fire spread state graph structure;
(2) And executing the graph structure convolution operation on the fire control scheduling strategy graph structure to obtain the characteristics of a plurality of fire control scheduling schemes in the fire control scheduling strategy graph structure.
In an embodiment of the present invention, consider, by way of example, a fire situation within a large industrial area.
Prior to analysis of the structure of the fire spread state diagram, fire spread simulation nodes have been established, representing the spread of fire at different points in time.
Now, in order to obtain more characteristic features, a graph structure convolution operation is performed on the fire spread state graph structure.
In this way, the fire spread simulation nodes can be processed through the graph convolution algorithm to obtain the aggregation characteristics of the nodes, such as the importance degree of the nodes, the fire spread condition of adjacent nodes and the like.
By performing the graph structure convolution operation on the fire spread state graph structure, the characteristics of the fire spread simulation node can be extracted.
These features are obtained by convolving a node with its neighbors.
The graph structure convolution operation can capture the relevance and the context information among the nodes, and is helpful for more comprehensively understanding the fire spread condition.
Assuming a fire emergency situation in a city is being handled, a fire dispatch strategy graph structure has been established that includes nodes of different fire dispatch schemes.
To obtain more representative features, a graph structure convolution operation is performed on the fire scheduling policy graph structure.
In this way, the fire control scheduling scheme node can be processed through a graph volume integration algorithm, so that comprehensive characteristics of the node, such as importance of the node in the whole scheduling process, association with other nodes and the like, are obtained.
By performing a graph structure convolution operation on the fire dispatch strategy graph structure, features of the fire dispatch protocol may be extracted.
The graph structure convolution operation allows connections and relationships between nodes to be considered, thereby obtaining more comprehensive information about the fire dispatch scheme.
These features help to evaluate the merits of different scheduling schemes, understand the effectiveness of scheduling strategies, and provide support for emergency decisions.
In the embodiment of the present invention, the step of determining the scheduling decision score based on the fire control related characteristics of the target area, the current fire spread simulation result and the current fire control scheduling information may be implemented by the following example execution.
(1) Acquiring a first weight between the fire control related characteristic of the target area and the current fire spread simulation result;
(2) Acquiring a second weight between the fire control related characteristic of the target area and the current fire control scheduling information;
(3) Based on the first weight and the second weight, performing an integration operation on the current fire spread simulation result and the current fire control scheduling information to obtain a fusion feature vector of the target area;
(4) And determining the scheduling decision score based on the fusion feature vector of the fire control associated feature of the target area and the target area.
In an embodiment of the present invention, it is assumed, by way of example, that a fire situation in a building is being treated.
By analyzing fire-related features (such as building structures, evacuation channels and the like) of the target area and current fire spread simulation results (such as fire spread speed, hot spot distribution and the like), the first weight between the two can be calculated.
This weight may reflect the extent to which the fire spreads to affect the target area.
Acquiring a first weight between the fire-fighting associated feature of the target area and the current fire spread simulation result means determining the importance of the fire-fighting associated feature of the target area in the scheduling decision.
The weight may be determined by evaluating the extent to which the fire spread simulation results affect the target area.
Fire emergency situations are handled taking into account fire-related features of the target area (e.g., building structure, safety exits, etc.) and current fire dispatch information (e.g., dispatch fire teams, rescue actions taken, etc.).
By analyzing the relationship between the two, a second weight can be calculated.
This weight reflects the importance of the fire schedule to the target area.
Acquiring a second weight between the fire-fighting correlation characteristic of the target area and the current fire-fighting scheduling information represents the weight of the fire-fighting schedule in the scheduling decision.
The weight may be determined according to the degree of influence of fire dispatch information on the target area.
When coping with a fire emergency, the first weight and the second weight are applied to the current fire spread simulation result and the current fire control schedule information.
By means of the integration operation, a fusion feature vector can be generated, wherein the fusion feature vector contains comprehensive information of the two aspects.
This fused feature vector may represent the composite feature of the target region.
By performing an integration operation on the current fire spread simulation result and the current fire control scheduling information based on the first weight and the second weight, the two aspects of information are combined together, and a more comprehensive target area feature vector is obtained.
The fusion feature vector reflects the relevance between fire spread simulation and fire control dispatch and is helpful for evaluating the dispatch decision score.
By comparing and analyzing the fire-fighting associated features of the target area with the fused feature vectors of the target area, a scheduling decision score may be determined.
This score reflects the degree of correlation between the current fire spread, fire dispatch information, and target zone characteristics.
Based on the fused feature vector of the fire fighting associated feature of the target area and the target area, a scheduling decision score may be calculated.
The score reflects the overall correlation between fire-fighting correlation characteristics, current fire spread simulation results, and current fire-fighting schedule information.
It can be used to evaluate the merits of different scheduling decision schemes and help make optimal scheduling decisions.
In an embodiment of the present invention, the step of obtaining the first weight between the fire protection related feature of the target area and the current fire spread simulation result may be performed by the following example.
(1) Determining at least one corresponding comparison fire spread simulation node of the target area in a fire spread state diagram structure based on the current fire spread simulation result;
(2) And acquiring the first weight based on a dot product between the fire control related feature of the target area and a comparison fire disaster simulation feature vector, wherein the comparison fire disaster simulation feature vector is a feature corresponding to the comparison fire disaster spread simulation node.
In the present embodiment, it is assumed, by way of example, that a fire in a large building is being treated.
By analyzing the current fire spread simulation result, the position of the target area in the fire spread state diagram can be determined.
This state diagram shows the structure and path of the fire spread.
At least one comparison fire spread simulation node can be found as a reference according to the location of the target area.
Based on the current fire spread simulation results, determining at least one corresponding comparative fire spread simulation node for the target area in the fire spread state diagram structure means associating the target area with the fire spread state diagram.
This step helps determine the location of the target area on the fire propagation path for subsequent weight calculation.
In analyzing fire-related features (e.g., building structures, fire protection facilities, etc.) of a target area, a dot product between the feature and a comparative fire simulation feature vector may be calculated.
The contrast fire simulation feature vector is a feature corresponding to the contrast fire spread simulation node.
The first weight may be obtained by a dot product operation.
This weight reflects the correlation between the fire-fighting correlation characteristics of the target area and the contrast fire spread simulation node characteristics.
By calculating the dot product, the similarity degree between the target area and the comparison node can be evaluated, and then the first weight is determined.
An overall implementation flow of an embodiment of the present invention is provided below.
Given a large forest, many fires have occurred in the past, and a large amount of historical fire spread data has been collected.
These data contain information about how the fire has spread from the initial point and how quickly the fire has spread under different weather and environmental conditions.
Forest managers now want to use this historical data to predict how fire may spread if a fire reoccurs in the future.
Firstly, two independent fire disaster feature learning networks are trained through the historical data, wherein one network is used for performing feature coding on the historical fire disaster spread data, and the other network is used for performing feature coding on the real-time fire disaster spread data.
And then, carrying out integrated analysis on the feature coding results of the two independent learning networks to form a joint learning network.
Thus, a pre-trained model is obtained.
In addition, a category identifier is designed for identifying the trend type of the fire spread.
Then, the historical fire spread data and the real-time fire spread data are input into a pre-training model, and data cost parameters are determined according to the deviation between the output result and the expected result.
Then, based on the category identifier, the type of the spread trend is determined for the historical fire spread data and the real-time fire spread data, and the category identification error value is determined.
After the optimization operation is performed for a plurality of times, a fire spread simulation model is obtained.
The model can be used for optimizing the learning strategies of at least two fire feature learning networks corresponding to the historical fire spread data set and then introducing the learning strategies into a fire spread simulation stage of the real-time fire spread data.
When a new fire occurs, a forest manager can acquire key fire image data, then input the key fire image data into a fire spread simulation model, and the model judges the spread trend type of the key fire image data so as to obtain the possible spread direction and speed of the current fire.
In addition, the method can also be combined with fire control dispatching information, and the optimal fire control dispatching scheme is obtained by analyzing the past fire spread simulation result and the fire control dispatching strategy so as to control the spread of fire as soon as possible.
Forest managers may also obtain current fire spread simulation results for the target area and fire dispatch information for such results.
By using the past fire spread simulation results, a plurality of target fire spread simulation nodes in the fire spread state diagram structure are determined, and the association between the nodes can trace the possible spread paths of the fire.
Meanwhile, a plurality of target fire control scheduling schemes in the fire control scheduling policy map structure are determined according to the past fire control scheduling information.
And then integrating the fire simulation feature vectors and the fire control dispatching feature vectors to obtain fire control related features of the target area.
And then, according to the current fire spread simulation result, determining at least one corresponding comparison fire spread simulation node of the target area in the fire spread state diagram structure.
Based on the dot product between this fire-related feature and the comparative fire simulation feature vector, a first weight may be obtained.
Next, a second weight between the fire-fighting correlation characteristic of the target area and the current fire-fighting scheduling information is obtained.
Based on the two weights, the integration operation can be performed on the current fire spread simulation result and the fire control dispatching information, and the fusion feature vector of the target area is obtained.
And finally, determining a scheduling decision score according to the fire control associated features and the fusion feature vector of the target area.
This score is used to represent the degree of adaptation between the current fire schedule information and the target area.
A high score means a better fit, i.e. that the fire resources can be used more efficiently.
In general, the fire spread simulation model training method based on big data can help better understand and predict the spread trend of fire and how to most effectively deploy fire resources.
It is assumed that the forest manager has a plurality of pending fire feature learning networks arranged in series.
Historical fire spread data is input into the networks, and corresponding historical fire spread data processing results are determined.
And then, training the to-be-determined fire feature learning network based on the error between the processing result and the historical fire spreading expected result to obtain at least two fire feature learning networks.
Now, a new fire occurs.
Key fire image data is acquired from the real-time fire spread data set, and then these data are input into the fire spread simulation model.
The model can judge the type of the spreading trend of the key fire image data, and the possible spreading direction and speed of the fire can be obtained.
Such information can help more accurately predict the development of fire and formulate corresponding fire strategies.
In addition, the current fire spread simulation result of the target area and fire control schedule information for the result are also acquired.
The optimal fire control dispatching scheme is obtained according to the past fire spread simulation result and the fire control dispatching strategy by combining the fire control related characteristics of the target area, the current fire spread simulation result and the current fire control dispatching information.
The method also carries out the convolution operation of the graph structure, and obtains the characteristics of a plurality of fire spread simulation nodes in the fire spread state graph structure and the characteristics of a plurality of fire control scheduling schemes in the fire control scheduling strategy graph structure.
These feature vectors are then integrated to form fire-fighting related features.
Thus, an optimal fire dispatch scheme can be determined based on the characteristics.
By the fire spread simulation model training method based on big data, forest managers can not only predict possible spread trend of fire, but also make an optimal fire control scheduling scheme, so that fire is controlled more effectively, and loss caused by fire is reduced.
Next, this fire spread simulation model training method based on big data will be further studied.
When dealing with real-time fire conditions, the manager needs to acquire the current fire spread simulation result of the target area and the current fire control scheduling information for the result.
Then, a plurality of target fire spread simulation nodes associated with the fire spread state diagram structure are determined using the past fire spread simulation results for the target area.
Meanwhile, a plurality of target fire control scheduling schemes related to the fire control scheduling strategy diagram structure are determined according to the past fire control scheduling information of the target area.
And then, the manager integrates each fire simulation feature vector and the fire control dispatching feature vector to obtain the fire control related features of the target area.
And then, according to the current fire spread simulation result, determining at least one corresponding comparison fire spread simulation node of the target area in the fire spread state diagram structure.
The first weight may be obtained using a dot product between the fire-related feature and the comparative fire simulation feature vector.
Meanwhile, a second weight between the fire control related characteristic of the target area and the current fire control scheduling information is acquired.
And then, integrating the current fire spread simulation result with fire control scheduling information according to the two weights to obtain a fusion feature vector of the target area.
And finally, determining a scheduling decision score based on the fusion feature vector of the fire control associated feature of the target area and the target area.
This score can represent the degree of adaptation between the current fire dispatch information and the target area, thereby helping to make optimal fire dispatch decisions.
Overall, the method provides a comprehensive, systematic and scientific way to predict fire spread and schedule fire protection through big data and machine learning techniques to minimize the harm of fire to human and environment.
In the fire spread simulation model training method based on big data, the execution of the graph structure convolution operation on the fire spread state graph structure and the fire control scheduling policy graph structure is also an important link.
To implement this step, the manager needs to first acquire a fire spread status diagram structure and a fire schedule policy map structure.
Then, a graph structure convolution operation is performed on the two graph structures to obtain characteristics of the plurality of fire spread simulation nodes and characteristics of the plurality of fire control schedule schemes.
Then, the manager integrates the fire simulation feature vectors and the fire control dispatching feature vectors to obtain fire control related features of the target area.
The fire-fighting related features not only contain the information of fire spread, but also contain the information of fire-fighting scheduling, so that the fire condition and the coping strategy of the target area can be reflected more comprehensively.
With the fire-fighting correlation feature, when a new fire occurs, the manager can determine a scheduling decision score based on the fire-fighting correlation feature of the target area, the current fire spread simulation result and fire-fighting scheduling information.
This score can help evaluate whether the current fire dispatch strategy is appropriate for the fire conditions of the target area, thereby making a more scientific, accurate decision.
By the fire spread simulation model training method based on big data, spread of fire can be effectively predicted and controlled, a manager can be helped to better utilize fire resources, fire efficiency is improved, and life and property safety of people is protected to the greatest extent.
The embodiment of the invention provides a computer device 100, wherein the computer device 100 comprises a processor and a nonvolatile memory comprising computer instructions, and when the computer instructions are executed by the processor, the computer device 100 executes the fire spread simulation model training method based on big data.
As shown in fig. 2, fig. 2 is a block diagram of a computer device 100 according to an embodiment of the present invention.
The computer device 100 comprises a memory 111, a processor 112 and a communication unit 113.
For data transmission or interaction, the memory 111, the processor 112 and the communication unit 113 are electrically connected to each other directly or indirectly.
For example, the elements may be electrically connected to each other via one or more communication buses or signal lines.
The foregoing description, for purpose of explanation, has been presented with reference to particular embodiments.
The illustrative discussions above are not intended to be exhaustive or to limit the disclosure to the precise forms disclosed.
Many modifications and variations are possible in light of the above teaching.
The embodiments were chosen and described in order to best explain the principles of the disclosure and its practical application, to thereby enable others skilled in the art to best utilize the disclosure and various embodiments with various modifications as are suited to the particular use contemplated.
The foregoing description, for purpose of explanation, has been presented with reference to particular embodiments.
The illustrative discussions above are not intended to be exhaustive or to limit the disclosure to the precise forms disclosed.
Many modifications and variations are possible in light of the above teaching.
The embodiments were chosen and described in order to best explain the principles of the disclosure and its practical application, to thereby enable others skilled in the art to best utilize the disclosure and various embodiments with various modifications as are suited to the particular use contemplated.