Fire identification model construction method and device

Abstract

The invention provides a fire identification model construction method and device. The method comprises the following steps: acquiring multi-source fire monitoring data from different data platforms; aiming at each identification model group, taking all the environment monitoring variables as the input of the identification model group, taking the fire characteristic variables corresponding to the identification model group as the output of the identification model group, and training each fire identification sub-model in the identification model group; the recognition result output by each fire recognition sub-model is compared with a real fire record, and the initial weight of each fire recognition sub-model is determined according to the comparison result; and performing weighted integration on the plurality of fire identification sub-models in the identification model group based on the initial weight, and constructing a fire identification model corresponding to the identification model group so as to construct a comprehensive fire identification model based on different types of fire identification models. By adopting the fire identification model construction method and device, the accuracy and stability of fire identification are improved.

Description

Technical Field

[0001] The present application relates to the field of fire monitoring technology, and in particular to a method and device for constructing a fire identification model. Background Art

[0002] Forest fires, a global ecological disaster, not only cause irreversible damage to forest ecosystems but also pose a serious threat to human life and property. Achieving accurate early detection and rapid response to fires has become a core challenge in forest fire prevention. In recent years, satellite remote sensing technology, with its wide-area coverage and periodic observations, has demonstrated tremendous potential in forest fire identification, becoming a crucial tool in modern forest fire prevention.

[0003] However, traditional satellite-based fire identification technology has obvious shortcomings: first, fire identification usually relies on a single characteristic indicator (such as brightness temperature) and a preset threshold, which is easily interfered with in complex environments, resulting in high misjudgment and missed detection rates; second, satellites are constrained by objective factors such as cloud cover and detection time, which can easily lead to gaps in monitoring data. Identifying fires based only on data from a few satellites will bring greater uncertainty to the identification results. Summary of the Invention

[0004] In view of this, the purpose of this application is to provide a fire identification model construction method and device to overcome at least one of the above-mentioned defects.

[0005] In a first aspect, an embodiment of the present application provides a method for constructing a fire identification model, comprising:

[0006] Acquire multi-source fire monitoring data from different data monitoring platforms. The multi-source fire monitoring data includes multiple environmental monitoring variables and multiple fire characteristic variables. Each fire characteristic variable corresponds to a recognition model group, and the recognition model group includes multiple fire recognition sub-models.

[0007] For each recognition model group, all environmental monitoring variables are used as the input data of the recognition model group, and the fire characteristic variables corresponding to the recognition model group are used as the output of the recognition model group. Each fire recognition sub-model in the recognition model group is trained using multi-source fire monitoring data;

[0008] Comparing the recognition results output by each fire recognition sub-model in the recognition model group with the actual fire records, and determining the initial weight of each fire recognition sub-model in the recognition model group according to the comparison results;

[0009] Based on the initial weights, a weighted integration is performed on multiple fire recognition sub-models under the recognition model group to construct a fire recognition model corresponding to the recognition model group, so as to construct a comprehensive fire recognition model based on fire recognition models under different types.

[0010] In an optional embodiment, the comparison result includes recognition accuracy, and the initial weight of each fire recognition sub-model in the recognition model group is determined in the following manner: for each fire recognition sub-model, the initial weight of the fire recognition sub-model in the recognition model group is determined based on the ratio of the recognition accuracy of the fire recognition sub-model to the sum of the recognition accuracy of all fire recognition model sub-models in the model group.

[0011] In an optional embodiment, the multiple fire characteristic variables include a first fire characteristic variable characterizing the location of the fire, a second fire characteristic variable characterizing the range of the fire, and a third fire characteristic variable characterizing the intensity of the fire; the recognition accuracy of each fire recognition sub-model in the first recognition model group is determined based on the accuracy of the recognition result output by the fire recognition sub-model, and the first recognition model group is the recognition model group corresponding to the first fire characteristic variable; the recognition accuracy of each fire recognition sub-model in the second recognition model group and the third recognition model group is determined based on the root mean square error of the recognition result output by the fire recognition sub-model, the second recognition model group is the recognition model group corresponding to the second fire characteristic variable, and the third recognition model group is the recognition model group corresponding to the third fire characteristic variable.

[0012] In an optional embodiment, each fire recognition sub-model in the recognition model group is trained in the following manner: when the recognition model group is the first recognition model group, binary cross entropy is used as the objective function, and ten-fold cross validation is used to adjust the hyperparameters of each fire recognition sub-model in the recognition model group; when the recognition model group is the second recognition model group or the third recognition model group, mean square error is used as the objective function, and ten-fold cross validation is used to adjust the hyperparameters of each fire recognition sub-model in the recognition model group.

[0013] In an optional embodiment, the method further includes: for the target fire identification sub-model, comparing the real fire record with the identification result output by the target fire identification sub-model, and determining whether the identification accuracy of the target fire identification sub-model meets the accuracy requirement based on the comparison result; if the accuracy requirement is not met, determining the dynamic weight for model updating based on the identification accuracy; and updating the weight of the target fire identification sub-model based on the dynamic weight and the initial weight of the target fire identification sub-model.

[0014] In an optional embodiment, the dynamic weight for model updating is determined in the following manner, including: when the recognition model group is the first recognition model group, the dynamic weight is determined based on the mean of the accuracy of the recognition results output by the target fire recognition sub-model; when the recognition model group is the second recognition model group or the third recognition model group, the dynamic weight is determined based on the mean of the root mean square error of the recognition results output by the target fire recognition sub-model.

[0015] In an optional embodiment, whether the accuracy requirement is met is determined by comparing the recognition accuracy of the most recent preset number of times with the historical recognition accuracy to determine the frequency of decline in recognition accuracy; if the frequency of decline in recognition accuracy is greater than or equal to a preset threshold, it is determined that the accuracy requirement is not met.

[0016] In an optional embodiment, the fire recognition model is updated in the following manner: the weight of each fire recognition sub-model after update is determined by summing the initial weight and the dynamic weight.

[0017] In an optional embodiment, the fire identification result output by the fire identification model is obtained by: collecting current multi-source fire monitoring data, extracting current environmental monitoring data from the current multi-source fire monitoring data; and inputting the current multi-source fire monitoring data into the comprehensive fire identification model to determine the current fire identification result.

[0018] In a second aspect, an embodiment of the present application further provides a fire identification model construction device, the device comprising:

[0019] A data acquisition module is used to obtain multi-source fire monitoring data from different data monitoring platforms. The multi-source fire monitoring data includes multiple environmental monitoring variables and multiple fire characteristic variables. Each fire characteristic variable corresponds to an identification model group, and the identification model group includes multiple fire identification sub-models;

[0020] A model training module is used to train each fire recognition sub-model in each recognition model group using all environmental monitoring data as the input of the recognition model group and the fire characteristic variables corresponding to the recognition model group as the output of the recognition model group, using multi-source fire monitoring data;

[0021] A weight calculation module is used to compare the recognition results output by each fire recognition sub-model in the recognition model group with the real fire records, and determine the initial weight of each fire recognition sub-model in the recognition model group according to the comparison results;

[0022] The model construction module is used to perform weighted integration of multiple fire recognition sub-models under the recognition model group based on initial weights, construct a fire recognition model corresponding to the recognition model group, and construct a comprehensive fire recognition model based on fire recognition models of different types.

[0023] The embodiments of the present application bring the following beneficial effects:

[0024] The embodiments of the present application provide a fire identification model construction method and apparatus that utilizes a constructed comprehensive fire identification model for fire identification. The comprehensive fire identification model can identify the values ​​of different fire characteristic variables, and the value of each fire characteristic variable is jointly determined by multiple fire identification sub-models, thus avoiding the problem of poor identification accuracy caused by relying solely on a single characteristic indicator to identify fires. Furthermore, the method can obtain multi-source fire monitoring data from different data monitoring platforms, avoiding the problem of low stability of identification results produced by identifying fires based on a small amount of satellite data. Compared with the fire identification model construction methods in the prior art, this method solves the problems of poor accuracy and low stability of fire identification.

[0025] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following is a brief introduction to the drawings required for use in the embodiments. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can be obtained based on these drawings without creative work.

[0027] Figure 1 A flow chart showing a method for constructing a fire identification model provided in an embodiment of the present application is shown;

[0028] Figure 2 A flowchart showing the training steps of the fire identification sub-model provided in an embodiment of the present application is shown;

[0029] Figure 3 A flowchart showing the steps for determining the initial weight provided in an embodiment of the present application is shown;

[0030] Figure 4 A flowchart showing the steps of forest fire identification provided by an embodiment of the present application is shown;

[0031] Figure 5 A flowchart showing the model weight updating steps provided in an embodiment of the present application is shown;

[0032] Figure 6 A schematic diagram of the structure of a fire identification model building device provided in an embodiment of the present application is shown;

[0033] Figure 7 A schematic structural diagram of an electronic device provided in an embodiment of the present application is shown. DETAILED DESCRIPTION

[0034] In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all of the embodiments. The components of the embodiments of the present application generally described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the application for which protection is claimed, but merely represents the selected embodiments of the present application. Based on the embodiments of the present application, each other embodiment obtained by those skilled in the art without making creative work falls within the scope of protection of the present application.

[0035] It is worth noting that before this application was filed, forest fires, as a global ecological disaster, not only caused irreversible damage to forest ecosystems, but also seriously threatened human life and property. Realizing early and accurate detection and rapid response to fires has become the core challenge of current forest fire prevention. In recent years, satellite remote sensing technology, with its advantages of wide-area coverage and periodic observations, has shown great potential in the field of forest fire identification and has become an important means of modern forest fire prevention. However, traditional satellite-based fire identification technology has obvious shortcomings: First, fire identification usually relies on a single characteristic indicator (such as brightness temperature) and a preset threshold, which is easily interfered with in complex environments, resulting in high misjudgment and missed judgment rates; second, satellites are restricted by objective factors such as cloud cover and detection time, which can easily lead to gaps in monitoring data. Identifying fires based on only a few satellite data will bring greater uncertainty to the identification results.

[0036] Based on this, an embodiment of the present application provides a method for constructing a fire identification model to improve the accuracy and stability of fire identification.

[0037] To facilitate understanding of this embodiment, each of the above exemplary steps provided in the embodiment of the present application is described below.

[0038] See also Figure 1 , Figure 1 This is a flow chart of a method for constructing a fire identification model provided in an embodiment of the present application. Figure 1 As shown, the fire identification model construction method provided in the embodiment of the present application includes:

[0039] Step S101, acquiring multi-source fire monitoring data from different data monitoring platforms;

[0040] Step S102: For each recognition model group, all environmental monitoring variables are used as inputs of the recognition model group, and the fire characteristic variables corresponding to the recognition model group are used as outputs of the recognition model group. Each fire recognition sub-model in the recognition model group is trained using multi-source fire monitoring data.

[0041] Step S103, comparing the recognition results output by each fire recognition sub-model in the recognition model group with the actual fire records, and determining the initial weight of each fire recognition sub-model in the recognition model group according to the comparison results;

[0042] Step S104 , weighted integration is performed on multiple fire recognition sub-models under the recognition model group based on the initial weights to construct a fire recognition model corresponding to the recognition model group, so as to construct a comprehensive fire recognition model based on the fire recognition models under different types.

[0043] The multi-source fire monitoring data includes multiple environmental monitoring variables and multiple fire characteristic variables. Each fire characteristic variable corresponds to an identification model group, and the identification model group includes multiple fire identification sub-models.

[0044] The fire identification model construction method provided in the embodiments of the present application can utilize the constructed comprehensive fire identification model to perform fire identification. The comprehensive fire identification model can identify the values ​​of different fire characteristic variables, and the value of each fire characteristic variable is jointly determined by multiple fire identification sub-models, avoiding the problem of poor identification accuracy caused by relying solely on a single characteristic indicator to identify fires. At the same time, it can obtain multi-source fire monitoring data from different data monitoring platforms, avoiding the problem of low stability of identification results produced by identifying fires based on a small number of satellite data, and solving the problems of poor accuracy and low stability of fire identification.

[0045] To facilitate understanding of this embodiment, the fire identification model construction method provided in the embodiment of the present application is applied to a terminal device as an example, and each of the above exemplary steps provided in the embodiment of the present application is described separately.

[0046] In step S101 , multi-source fire monitoring data is acquired from different data monitoring platforms.

[0047] In this step, different data monitoring platforms include but are not limited to: satellite monitoring platform, meteorological data platform, fire database, and terrain mapping platform.

[0048] Among them, satellite monitoring platforms may refer to MODIS, VIIRS / NPP, Sentinel-2, Landsat-9, Himawari-8 and other satellite monitoring platforms; meteorological data platforms may refer to ERA5 (ECMWF Atmospheric Reanalysis Dataset); fire databases may refer to the Global Fire Emissions Database (GEFD); and topographic mapping platforms may refer to SRTM (Shuttle Radar Topography Mission).

[0049] Multi-source fire monitoring data refers to fire monitoring data obtained from multiple channels. Multi-source fire monitoring data covers historical time periods. Multi-source fire monitoring data includes environmental monitoring data corresponding to environmental monitoring variables and fire characteristic data corresponding to fire characteristic variables. Environmental monitoring data includes smoke pattern data, thermal anomaly data, meteorological data, and topographic data; fire characteristic data includes fire location, fire extent, and fire intensity.

[0050] Smoke morphology data includes smoke texture and shape; meteorological data includes fire risk index, drought index, temperature, humidity, wind speed, and wind direction; and topographic data includes slope and aspect. Fire location refers to the location of the fire point; fire extent refers to the burned area; and fire intensity refers to the radiant power of the fire.

[0051] Here, multiple fire characteristic variables include a first fire characteristic variable characterizing the fire location, a second fire characteristic variable characterizing the fire burning area, and a third fire characteristic variable characterizing the radiation power; environmental detection variables include smoke texture, smoke shape, fire risk index, drought index, temperature, humidity, wind speed, wind direction, slope and slope direction.

[0052] In an embodiment of the present application, smoke texture and shape are extracted from data monitored by visible and near-infrared products of a satellite monitoring platform based on a gray-level co-occurrence matrix, thermal anomaly data are extracted from data monitored by thermal infrared and microwave products of a satellite monitoring platform based on an absolute threshold method, meteorological data are obtained based on the ERA5 meteorological data platform, and terrain data such as slope and aspect are calculated based on the SRTM platform.

[0053] In step S102, for each recognition model group, all environmental monitoring variables are used as the input of the recognition model group, and the fire characteristic variables corresponding to the recognition model group are used as the output of the recognition model group. Each fire recognition sub-model in the recognition model group is trained using multi-source fire monitoring data.

[0054] In this step, each fire characteristic variable corresponds to a recognition model group, and the recognition model group includes multiple fire recognition sub-models.

[0055] For example, the fire location corresponds to the first recognition model group, the fire burning area corresponds to the second recognition model group, and the radiation power corresponds to the third recognition model group. Each recognition model group includes multiple fire recognition sub-models of different types.

[0056] The types of multiple fire recognition sub-models in different recognition model groups can be the same. For example, each recognition model group includes a support vector machine, a random forest, a convolutional neural network, a long short-term neural network, and a recurrent neural network. Then, the three types of fire feature data correspond to a total of fifteen fire recognition sub-models.

[0057] In one example, the training process of each fire sub-model in different recognition model groups is the same, but the output numbers of different recognition model groups are different, and the output data of different recognition model groups are determined by the fire characteristic variables.

[0058] For example, the output of the first recognition model group is the fire location, the output of the second recognition model group is the fire burning area, and the output of the third recognition model group is the radiation power.

[0059] Refer to the following Figure 2 Let’s introduce the training process of the fire recognition sub-model in detail.

[0060] Figure 2 A flowchart showing the training steps of the fire identification sub-model provided in the embodiment of the present application is shown in FIG. Figure 2 As shown in Figure 2, the training steps of the fire recognition sub-model include:

[0061] Step S1021 , when the recognition model group is the first recognition model group, binary cross entropy is used as the objective function, and ten-fold cross validation is used to adjust the hyperparameters of each fire recognition sub-model in the recognition model group.

[0062] The first recognition model group is used to identify the fire location and obtain discrete category labels. It is suitable for qualitative analysis and belongs to the classification task.

[0063] Taking the training of the random forest model in the first recognition model group using multi-source fire monitoring data as an example, 70% of the multi-source fire monitoring data is used as the training set, and 30% of the data is used as the validation set. After training the random forest model with the training set, a trained random forest model is obtained. After all fire recognition sub-models in the first recognition model group are trained, the model matrix corresponding to the first recognition model group is obtained. The model matrix corresponding to the first recognition model group includes the environmental monitoring data, the five fire recognition sub-models in the first recognition model group, and the fire locations corresponding to each fire recognition sub-model.

[0064] At this point, for each candidate hyperparameter combination, a ten-fold cross-validation method was used to adjust the hyperparameters of the random forest model. Environmental monitoring data was input into the random forest model to obtain the model's output prediction matrix. This prediction matrix was used as the predicted values, and the fire locations in the training set were used as the true values. The predicted and true values ​​were input into the binary cross-entropy to calculate the binary cross-entropy loss. The binary cross-entropy losses corresponding to 10 validation sets were obtained, and the average of these 10 binary cross-entropy losses was taken as the final score for the candidate hyperparameter combination. From the multiple candidate hyperparameter combinations, the candidate hyperparameter combination with the lowest average validation loss was selected as the optimal solution for the random forest model.

[0065] With reference to the above process, the optimal solution of the parameters of each fire identification sub-model in the first identification model group can be determined.

[0066] Step S1022 , when the recognition model group is the second recognition model group or the third recognition model group, the mean square error is used as the objective function, and ten-fold cross validation is used to adjust the hyperparameters of each fire recognition sub-model in the recognition model group.

[0067] The second identification model group or the third identification model group is used to identify the fire burning area and radiation power to obtain specific values, which is suitable for quantitative analysis and belongs to a regression task.

[0068] Taking the training of the recurrent neural network model in the second recognition model group using multi-source fire monitoring data as an example, 70% of the multi-source fire monitoring data is used as the training set, and 30% of the data is used as the validation set. After training the recurrent neural network model with the training set, a trained recurrent neural network model is obtained. After all fire recognition sub-models in the second recognition model group are trained, the model matrix corresponding to the second recognition model group is obtained. The model matrix corresponding to the second recognition model group includes the environmental monitoring data, the five fire recognition sub-models in the second recognition model group, and the fire range corresponding to each fire recognition sub-model.

[0069] At this point, for each candidate hyperparameter combination, a ten-fold cross-validation method was used to adjust the hyperparameters of the recurrent neural network model. The environmental monitoring data was input into the recurrent neural network model to obtain the model's output prediction matrix. The prediction matrix was used as the predicted value, and the fire range in the training set was used as the true value. The mean squared error (MSE) was calculated based on the predicted and true values ​​to calculate the mean squared error loss. A total of 10 MSE losses corresponding to the validation set were obtained, and the average of these 10 MSE losses was taken as the final score for the candidate hyperparameter combination. From the multiple candidate hyperparameter combinations, the candidate hyperparameter combination with the lowest average validation loss was selected as the optimal solution for the recurrent neural network model.

[0070] With reference to the above process, the optimal solution of the parameters of each fire recognition sub-model in the second recognition model group and the third recognition model group can be determined.

[0071] In step S103, the recognition result output by each fire recognition sub-model in the recognition model group is compared with the real fire record, and the initial weight of each fire recognition sub-model in the recognition model group is determined according to the comparison result.

[0072] In this step, the recognition result is the output of the fire recognition sub-model. Taking the first recognition model group as an example, the recognition result is the fire location; taking the second recognition model group as an example, the recognition result is the fire burning area; taking the third recognition model group as an example, the recognition result is the radiation power.

[0073] The real fire record may refer to a data record corresponding to a real fire. For example, the real fire record may be a fire location, a fire burning area, and a radiation power in a verification set.

[0074] Refer to the following Figure 3 Let's introduce the process of determining the initial weights in detail.

[0075] Figure 3 A flowchart showing the steps for determining the initial weight provided in the embodiment of the present application is shown in FIG. Figure 3 As shown, the steps for determining the initial weight include:

[0076] Step S1031 : for each fire recognition sub-model, determine the recognition accuracy of the fire recognition sub-model.

[0077] In one case, for each fire recognition sub-model in the first recognition model group, multiple recognition results of the fire recognition sub-model are obtained, and based on the multiple recognition results output by the fire recognition sub-model and the actual fire records, the accuracy of the fire recognition sub-model is calculated, and then the recognition accuracy of the fire recognition sub-model is determined based on the accuracy.

[0078] In another case, for each fire recognition sub-model in the second recognition model group and the third recognition model group, multiple recognition results of the fire recognition sub-model are obtained, and based on the recognition results output by the fire recognition sub-model and the actual fire records, the root mean square error of the fire recognition sub-model is determined, and then the recognition accuracy of the fire recognition sub-model is determined based on the root mean square error.

[0079] Step S1032 : For each fire recognition sub-model, determine the initial weight of the fire recognition sub-model in the recognition model group based on the ratio of the recognition accuracy of the fire recognition sub-model to the sum of the recognition accuracy of all fire recognition sub-models in the recognition model group.

[0080] For example, the recognition accuracy of the nth fire recognition sub-model in the first recognition model group is recorded as: A n , the initial weight is recorded as: W n , then the initial weight of the nth fire recognition sub-model in the first recognition model group is:

[0081]

[0082] In step S104, a plurality of fire recognition sub-models under the recognition model group are weightedly integrated based on the initial weights to construct a fire recognition model corresponding to the recognition model group, so as to construct a comprehensive fire recognition model based on the fire recognition models under different types.

[0083] In this step, for each recognition model group, the initial weights of the five fire recognition sub-models in the recognition model group are determined. Then, the weights of these five fire recognition sub-models are integrated according to the initial weights to construct a recognition model group. Then, these three recognition model groups together constitute the comprehensive fire recognition model.

[0084] In this way, the constructed comprehensive fire identification model can be used to identify the fire that may occur at present.

[0085] Refer to the following Figure 4 To introduce the application process of the comprehensive fire identification model.

[0086] Figure 4 A flowchart of the forest fire identification steps provided in the embodiment of the present application is shown as follows: Figure 4 As shown in Figure 2, the steps for forest fire identification include:

[0087] Step S1041 : collecting current multi-source fire monitoring data, and extracting current environment monitoring data from the current multi-source fire monitoring data.

[0088] And by using a variety of methods such as gray level co-occurrence matrix and absolute threshold method, the current environmental monitoring data can refer to the current data of environmental monitoring variables.

[0089] Step S1042: Input the current multi-source fire monitoring data into the comprehensive fire identification model to determine the current fire identification result.

[0090] For example: the current multi-source fire monitoring data is input into the trained comprehensive fire recognition model. The comprehensive fire recognition model will fuse the output of each fire recognition sub-model according to the initial weight of each fire recognition sub-model in the first recognition model group to obtain the final fire location; according to the initial weight of each fire recognition sub-model in the second recognition model group, the output of each fire recognition sub-model is fused and calculated to obtain the final fire burning area; according to the initial weight of each fire recognition sub-model in the third recognition model group, the output of each fire recognition sub-model is fused and calculated to obtain the final radiation power.

[0091] Step S1043: Determine the fire range and fire level based on the fire burning area and the spatial distribution of the radiation power.

[0092] For example, the radiation power can be divided into multiple power intervals, each corresponding to a fire severity. The current fire severity is determined based on the power interval within which the radiation power falls within the current time period. The fire severity is then used to characterize the size of the fire. Furthermore, the fire's extent can be determined based on the burning area and fire boundaries.

[0093] Step S1044: determining a change trend of the fire locations based on multiple fire locations within a short period of time, and determining a fire spreading direction based on the change trend.

[0094] For example, multiple fire locations within a short period of time can be connected to form a location curve. This location curve can be used to determine the changing trend of the fire location and, in turn, the direction of fire spread.

[0095] In order to ensure the accuracy of the identification results of the comprehensive fire identification model, it is necessary to establish a model evaluation mechanism to dynamically adjust the weights of each fire identification sub-model.

[0096] Refer to the following Figure 5 Let’s introduce the model weight update process in detail.

[0097] Figure 5 A flow chart showing the model weight updating steps provided in the embodiment of the present application is shown in FIG. Figure 5 As shown, the model weight update steps include:

[0098] Step S1045 : For the target fire recognition sub-model, the real fire record is compared with the recognition result output by the target fire recognition sub-model, and whether the recognition accuracy of the target fire recognition sub-model meets the accuracy requirement is determined based on the comparison result.

[0099] Here, the weight of each fire identification sub-model in the comprehensive fire identification model needs to be updated.

[0100] In one case, for the target fire recognition sub-model in the first recognition model group, the accuracy is selected as the evaluation indicator of the recognition accuracy, and the recognition accuracy corresponding to the most recent k recognition results of the target fire recognition sub-model is calculated based on the accuracy formula, and each recognition result corresponds to a recognition accuracy.

[0101] In another case, for the target fire recognition sub-model in the second recognition model group or the third recognition model group, the root mean square error (RMSE) is selected as the recognition accuracy evaluation indicator. The recognition accuracy corresponding to the target fire recognition sub-model's most recent k recognition results is calculated based on the RMS error. Each recognition result is assigned a recognition accuracy, where K is the preset number of times.

[0102] The recognition accuracy of the most recent preset number of times is then compared with the historical recognition accuracy to determine the frequency of decline in recognition accuracy. For example, the recognition accuracy of the fire recognition sub-model when the initial weights were determined is used as the historical recognition accuracy. Each recent recognition accuracy is compared with the historical recognition accuracy. The number of the most recent k times that have a lower recognition accuracy than the historical recognition accuracy is determined. The ratio of this number to the number k is used as the decline frequency. If 10 of the 20 most recent recognition accuracies are lower than the historical recognition accuracy, the decline frequency is 50%.

[0103] If the frequency of decrease in recognition accuracy is greater than or equal to the preset threshold, it is determined that the accuracy requirement is not met, and step S1046 is executed. If the frequency of decrease in recognition accuracy is less than the preset threshold, it is determined that the accuracy requirement is met, and there is no need to update the model weights. For example, if the preset threshold is 25%, using the above example, since the frequency of decrease of 50% is greater than 25%, it is determined that the accuracy requirement is not met.

[0104] Step S1046: If the accuracy requirement is not met, a dynamic weight for model update is determined based on the recognition accuracy.

[0105] In one case, when the recognition model group is the first recognition model group, the dynamic weight is determined based on the average of the accuracy rates of the recognition results output by the target fire recognition sub-models.

[0106] For example, taking the mean accuracy as the mean recognition accuracy, the mean accuracy of the target fire recognition sub-model for the last k times is recorded as: B0, and the mean accuracy of the nth fire recognition sub-model for the last k times is recorded as: B n , the dynamic weight of the fire identification sub-model is recorded as: The calculation formula for dynamic weight is:

[0107]

[0108] In another case, when the recognition model group is the second recognition model group or the third recognition model group, the dynamic weight is determined based on the mean of the root mean square errors of the recognition results output by the target fire recognition sub-model.

[0109] For example, the mean value of the root mean square error is used as the mean value of recognition accuracy. The mean value of the root mean square error of the target fire recognition sub-model for the last k times is recorded as: C0, and the mean value of the root mean square error of the nth fire recognition sub-model for the last k times is recorded as: C n , the dynamic weight of the fire identification sub-model is recorded as: The calculation formula for dynamic weight is:

[0110]

[0111] Step S1047 : updating the weight of the target fire identification sub-model based on the dynamic weight and the initial weight of the target fire identification sub-model.

[0112] For example, the sum of the initial weight and the dynamic weight is used to determine the weight of each fire identification sub-model after the update. At this time, the initial weight is recorded as: The updated weight is recorded as: W n ,but Among them, β represents the weight coefficient of the initial weight.

[0113] Based on the same inventive concept, the embodiments of the present application also provide a fire identification model construction device corresponding to the fire identification model construction method. Since the principle of solving the problem by the device in the embodiments of the present application is similar to the above-mentioned fire identification model construction method in the embodiments of the present application, the implementation of the device can refer to the implementation of the method, and the repeated parts will not be repeated.

[0114] See also Figure 6 , Figure 6 This is a structural diagram of a fire identification model building device provided in an embodiment of the present application. Figure 6 As shown in , the fire identification model building device 200 includes:

[0115] The data acquisition module 201 is used to acquire multi-source fire monitoring data from different data monitoring platforms. The multi-source fire monitoring data includes multiple environmental monitoring variables and multiple fire characteristic variables. Each fire characteristic variable corresponds to a recognition model group, and the recognition model group includes multiple fire recognition sub-models.

[0116] The model training module 202 is used to train each fire recognition sub-model in each recognition model group using multi-source fire monitoring data, taking all environmental monitoring variables as the input of the recognition model group and the fire characteristic variables corresponding to the recognition model group as the output of the recognition model group;

[0117] The weight calculation module 203 is used to compare the recognition results output by each fire recognition sub-model in the recognition model group with the real fire records, and determine the initial weight of each fire recognition sub-model in the recognition model group according to the comparison results;

[0118] The model building module 204 is used to perform weighted integration on multiple fire recognition sub-models under the recognition model group based on initial weights to build a fire recognition model corresponding to the recognition model group, so as to build a comprehensive fire recognition model based on fire recognition models of different types.

[0119] See also Figure 7 , Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application. Figure 7 As shown in FIG, the electronic device 300 includes a processor 310 , a memory 320 and a bus 330 .

[0120] The memory 320 stores machine-readable instructions executable by the processor 310. When the electronic device 300 is running, the processor 310 communicates with the memory 320 via the bus 330. When the machine-readable instructions are executed by the processor 310, the above-mentioned Figure 1 The steps of the fire identification model construction method in the illustrated method embodiment and the specific implementation thereof can be found in the method embodiment, which will not be described in detail here.

[0121] The embodiment of the present application also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the computer program can execute the above-mentioned Figure 1 The steps of the fire identification model construction method in the illustrated method embodiment and the specific implementation thereof can be found in the method embodiment, which will not be described in detail here.

[0122] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0123] In the several embodiments provided in this application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. The device embodiments described above are merely schematic. For example, the division of the units is merely a logical function division. There may be other division methods in actual implementation. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed may be through some communication interface, indirect coupling or communication connection of devices or units, which may be electrical, mechanical or other forms.

[0124] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0125] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0126] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a non-volatile computer-readable storage medium that is executable by a processor. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0127] Finally, it should be noted that the above-described embodiments are only specific implementation methods of the present application, which are used to illustrate the technical solutions of the present application, rather than to limit them. The scope of protection of the present application is not limited thereto. Although the present application has been described in detail with reference to the above-mentioned embodiments, those skilled in the art should understand that any person skilled in the art can modify or easily conceive of changes to the technical solutions described in the above-mentioned embodiments within the technical scope disclosed in the present application, or perform equivalent replacements for some of the technical features thereof. These modifications, changes, or replacements do not deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A method for constructing a fire identification model, characterized in that: include: Acquire multi-source fire monitoring data from different data monitoring platforms, wherein the multi-source fire monitoring data includes multiple environmental monitoring variables and multiple fire characteristic variables, each fire characteristic variable corresponds to a recognition model group, and the recognition model group includes multiple fire recognition sub-models; For each recognition model group, all environmental monitoring variables are used as inputs of the recognition model group, and the fire characteristic variables corresponding to the recognition model group are used as outputs of the recognition model group, and each fire recognition sub-model in the recognition model group is trained using the multi-source fire monitoring data; Comparing the recognition results output by each fire recognition sub-model in the recognition model group with the actual fire records, and determining the initial weight of each fire recognition sub-model in the recognition model group according to the comparison results; Based on the initial weights, a plurality of fire recognition sub-models under the recognition model group are weightedly integrated to construct a fire recognition model corresponding to the recognition model group, so as to construct a comprehensive fire recognition model based on fire recognition models under different types.

2. The method according to claim 1, characterized in that The comparison result includes recognition accuracy, which is determined by the initial weight of each fire recognition sub-model in the recognition model group in the following way: For each fire recognition sub-model, the initial weight of the fire recognition sub-model in the recognition model group is determined based on the ratio of the recognition accuracy of the fire recognition sub-model to the sum of the recognition accuracy of all fire recognition sub-models in the model group.

3. The method according to claim 1, characterized in that The plurality of fire characteristic variables include a first fire characteristic variable characterizing a fire location, a second fire characteristic variable characterizing a fire range, and a third fire characteristic variable characterizing a fire intensity; The recognition accuracy of each fire recognition sub-model in the first recognition model group is determined based on the accuracy of the recognition result output by the fire recognition sub-model, and the first recognition model group is the recognition model group corresponding to the first fire characteristic variable; The recognition accuracy of each fire recognition sub-model in the second recognition model group and the third recognition model group is determined based on the root mean square error of the recognition results output by the fire recognition sub-model. The second recognition model group is the recognition model group corresponding to the second fire characteristic variable, and the third recognition model group is the recognition model group corresponding to the third fire characteristic variable.

4. The method according to claim 3, characterized in that Each fire recognition sub-model in the recognition model group is trained in the following way: When the recognition model group is the first recognition model group, the binary cross entropy is used as the objective function, and ten-fold cross validation is used to adjust the hyperparameters of each fire recognition sub-model in the recognition model group; When the recognition model group is the second recognition model group or the third recognition model group, the mean square error is used as the objective function, and ten-fold cross validation is used to adjust the hyperparameters of each fire recognition sub-model in the recognition model group.

5. The method according to claim 1, wherein The method further comprises: For the target fire identification sub-model, comparing the real fire record with the identification result output by the target fire identification sub-model, and determining whether the identification accuracy of the target fire identification sub-model meets the accuracy requirement based on the comparison result; If the accuracy requirement is not met, determining a dynamic weight for model updating based on the recognition accuracy; The weight of the target fire identification sub-model is updated based on the dynamic weight and the initial weight of the target fire identification sub-model.

6. The method according to claim 5, characterized in that The dynamic weights used for model updates are determined by: When the recognition model group is the first recognition model group, determining the dynamic weight based on the average of the accuracy rates of the recognition results output by the target fire recognition sub-models; When the recognition model group is the second recognition model group or the third recognition model group, the dynamic weight is determined based on the mean of the root mean square errors of the recognition results output by the target fire recognition sub-models.

7. The method according to claim 5, characterized in that Whether the accuracy requirements are met is determined by: Comparing the recognition accuracy of the most recent preset number of times with the historical recognition accuracy to determine the frequency of decline in recognition accuracy; If the frequency of decrease in the recognition accuracy is greater than or equal to a preset threshold, it is determined that the accuracy requirement is not met.

8. The method according to claim 5, characterized in that The fire identification model is updated in the following way: The updated weight of each fire identification sub-model is determined by summing the initial weight and the dynamic weight.

9. The method according to claim 1, characterized in that The method further comprises: Collecting current multi-source fire monitoring data, and extracting current environment monitoring data from the current multi-source fire monitoring data; The current multi-source fire monitoring data is input into the comprehensive fire identification model to determine a current fire identification result.

10. A fire identification model construction device, characterized in that: include: A data acquisition module is used to acquire multi-source fire monitoring data from different data monitoring platforms. The multi-source fire monitoring data includes multiple environmental monitoring variables and multiple fire characteristic variables. Each fire characteristic variable corresponds to a recognition model group, and the recognition model group includes multiple fire recognition sub-models. A model training module is used to train each fire recognition sub-model in each recognition model group using all environmental monitoring data as the input of the recognition model group and the fire characteristic variables corresponding to the recognition model group as the output of the recognition model group, using multi-source fire monitoring data; A weight calculation module is used to compare the recognition results output by each fire recognition sub-model in the recognition model group with the real fire records, and determine the initial weight of each fire recognition sub-model in the recognition model group according to the comparison results; The model construction module is used to perform weighted integration on multiple fire recognition sub-models under the recognition model group based on the initial weights, to construct a fire recognition model corresponding to the recognition model group, so as to construct a comprehensive fire recognition model based on fire recognition models of different types.

Images