A seismic rescue zoning and classification method based on a deep learning hybrid model

Abstract

The present invention discloses a seismic rescue zoning and classification method based on a deep learning hybrid model, including the following processes: division and determination of basic research units; generation of time series data of mobile phone positioning amounts for research units; analysis of the law of change of positioning amounts over time to determine the types of seismic rescue areas for research units; standardization and normalization of data and creation of a data set; construction, training and parameter tuning of a hybrid model; and use of the hybrid model to predict the types of key search and rescue areas for research units. The present invention constructs a hybrid model based on CNN-BiLSTM, extracts high-dimensional features through the feature abstraction ability of CNN, and performs time series prediction on the high-dimensional features of the mobile phone positioning amount sequence through the BiLSTM layer, fully integrating the advantages of the two network models of CNN and RNN. Therefore, the hybrid model of the present invention has high prediction accuracy and strong generalization ability, providing a reliable classification method for seismic rescue area division.

Description

Technical Field

[0001] The present invention relates to a method for classifying earthquake rescue areas, and particularly to a method for classifying earthquake rescue areas based on a deep learning hybrid model. Background Art

[0002] With the full popularization of smart phones, through precise positioning by LBS technology, the state of population flow can be comprehensively and immediately grasped. Mobile phone positioning data has the characteristics of high coverage rate, diverse positioning accuracies, and easy acquisition. Therefore, using mobile phone positioning population data can assist in judging the disaster situation, determining the severely affected areas, and effectively improving the emergency disaster relief ability.

[0003] After a major earthquake or a particularly large earthquake disaster occurs, the scope that needs to be rescued may involve multiple districts, counties, towns, etc. Geographical zoning of the affected areas can ensure the effective coordination of search and rescue forces. The division of rescue areas and the determination of rescue priorities are the basis for carrying out rescue operations, and also the premise for improving the efficiency of rescue work and ensuring the timeliness and coordination of rescue work. The United Nations put forward the concept of "International Search and Rescue Operation Zoning" in long-term international rescue operations. According to the zoning scope, combined with the refined disaster situation assessment within the zone, the rescue priority is divided, and the rescue forces are reasonably dispatched and coordinated. This method is conducive to formulating a better search and rescue action plan and improving the efficiency of multiple teams carrying out search and rescue operations simultaneously. Currently, international measures for rescue area zoning and rescue priority division are manual, which is time-consuming and laborious, and is easily affected by human factors. The characteristics of population spatio-temporal variation distribution are not considered, and it is difficult to achieve the division of rescue priorities in combination with the real-time dynamics of the population distribution in the earthquake disaster situation. Summary of the Invention

[0004] In order to solve the deficiencies of the above-mentioned technologies, the present invention provides a method for classifying earthquake rescue areas based on a deep learning hybrid model.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is: A method for classifying earthquake rescue areas based on a deep learning hybrid model, including the following processes:

[0006] S1. Division and determination of basic research units;

[0007] S2. Generating time series data of mobile phone positioning amounts of research units;

[0008] S3. Analyzing the law of the change of the positioning amount over time, and determining the earthquake rescue area type of the research unit;

[0009] S4. Standardizing and normalizing the data and creating a data set;

[0010] S5. Building, training, and tuning the parameters of the hybrid model;

[0011] S6. Use a hybrid model to predict the types of key search and rescue areas for research units.

[0012] Further, in step S1, first, assume a certain urban block as the earthquake rescue task area, process the road network data of the urban blocks to obtain a road network line layer, and then use the GIS method to convert the road network line layer into a surface layer. The areas in the surface layer are divided according to the following rules: 1) Areas with an area greater than 0.15 square kilometers are used as basic research units, that is, the smallest rescue areas; 2) Areas with an area less than 0.15 square kilometers are merged nearby; 3) Areas without building distributions are deleted.

[0013] Further, in step S2, taking the basic research unit as the scale, the mobile phone positioning volume data at the same time intervals within 24 hours of 22 days for each research unit is statistically analyzed, thereby constructing a time series data of mobile phone positioning volume with a data dimension of 24 * 22.

[0014] Further, in step S3, analyze the law of the change of the mobile phone positioning volume of each basic research unit over time, obtain the corresponding mobile phone positioning volume curve, and divide the research unit into the corresponding earthquake rescue area types according to the curve.

[0015] Further, in step S3, there are five types of earthquake rescue area types divided as follows:

[0016] 1) Work area: When there are significant differences in the mobile phone positioning volume curve between weekdays and weekends, with a dense population during weekdays and a sparse population on weekends, it is defined as a work area;

[0017] 2) Residential area: When there are obvious peaks and valleys in the mobile phone positioning volume curve, and the largest peaks within a day are at 12 - 13 o'clock noon and from 19 o'clock to 7 o'clock the next day, it is defined as a residential area;

[0018] 3) Integrated office and living area: When the change in the positioning volume within a day is relatively small, and the rising and falling intervals of the mobile phone positioning volume curves on weekdays and weekends are the same, it is defined as an integrated office and living area;

[0019] 4) Commercial area: When there are no obvious peaks and valleys in the mobile phone positioning volume curve, and the high - value area appears from 10 o'clock in the morning to 21 o'clock, it is defined as a commercial area;

[0020] 5) Entertainment area: When there are obvious peaks in the mobile phone positioning volume curve at around 8 o'clock in the morning and around 19 o'clock in the evening, and there is a small fluctuation between 8 o'clock and 19 o'clock with no obvious valleys, it is defined as an entertainment area.

[0021] Further, in step S4, first, the Z-score standardization method is used to standardize the mobile phone positioning volume data of each research unit, and then the Z-score normalization method is used to normalize the 48-hour mobile phone positioning volume of each research unit for 24 hours on weekdays and weekends:

[0022]

[0023] In the formula, r represents the research unit, and Z r score(t) represents the normalized data of research unit r, represents the standard deviation of the positioning volume data on the 22nd in research unit r, represents the average value of the 48-hour mobile phone positioning volume of research unit r, is the mobile phone positioning volume at a certain moment of research unit r. The normalized data is shuffled, and the shuffled data is grouped according to the ratio of 6:2:2, which are used as the training data set, validation data set, and test data set respectively.

[0024] Further, in step S5, first, a CNN model with three branches is built. Each CNN branch is sequentially equipped with three CNN1D convolutional layers and a global pooling layer. The three CNN branches are integrated into the double-layer BiLSTM layer through the joint operation, and then connected to the double-layer fully connected layer to complete the construction of the hybrid model; the training data set with a data dimension of 24*22 is input into the hybrid model for model training, and the accuracy of the model is verified through the validation data set; the parameters of the hybrid model are adjusted through data translation and overfitting sample balancing to improve the accuracy of the model.

[0025] Further, in step S5, the number of convolutional kernels of each CNN1D convolutional layer is 128. The convolutional kernel sizes of the three CNN1D convolutional layers of CNN branch 1 are 21*1, 10*1, and 5*1 respectively. The convolutional kernel sizes of the three CNN1D convolutional layers of CNN branch 2 are 16*1, 8*1, and 3*1 respectively. The convolutional kernel sizes of the three CNN1D convolutional layers of CNN branch 3 are 11*1, 7*1, and 2*1 respectively.

[0026] Further, in step S6, the test data set is input into the hybrid model to predict the type of earthquake rescue area corresponding to the research unit. Different types of earthquake rescue areas correspond to different types of key search and rescue areas: if the research unit is a work area, it corresponds to the key search and rescue area from 9:00 to 17:00; if the research unit is a residential area, it corresponds to the key search and rescue areas from 12:00 to 13:00 and from 19:00 to 7:00 the next day; if the research unit is an integrated office and living area, it corresponds to the key search and rescue area throughout the day; if the research unit is a commercial area, it corresponds to the key search and rescue area from 10:00 to 21:00; if the research unit is an entertainment area, it corresponds to the key search and rescue area from 8:00 to 19:00.

[0027] The present invention constructs a hybrid model based on CNN-BiLSTM. The high-dimensional features are extracted through the feature abstraction ability of CNN, and the high-dimensional features of the mobile phone positioning quantity sequence are predicted in time series through the BiLSTM layer, fully integrating the advantages of the two network models of CNN and RNN. Therefore, the hybrid model of the present invention has high prediction accuracy and strong generalization ability, providing a reliable classification method for earthquake rescue area division. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 is a schematic diagram of the overall process of the present invention.

[0029] Figure 2 is a distribution map of the basic research units in the embodiment of the present invention.

[0030] Figure 3 is a curve graph showing the change of the mobile phone positioning quantity over time in the embodiment of the present invention.

[0031] Figure 4 is a structural diagram of the hybrid model in the embodiment of the present invention.

[0032] Figure 5 is an effect graph of the hybrid model training in the embodiment of the present invention.

[0033] Figure 6 is a distribution map of the key earthquake search and rescue areas in the embodiment of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0034] The present invention will be further described in detail below with reference to the drawings and specific embodiments.

[0035] The present invention discloses a method for classifying earthquake rescue areas based on a deep learning hybrid model, as Figure 1 shown, including the following processes:

[0036] S1. Division and determination of basic research units:

[0037] First, take the planar area enclosed by the core main roads in Lanzhou City as the earthquake rescue task area. Process the road network data of this area to obtain the road network line layer, and then use the GIS method to convert the road network line layer into a polygon layer. Divide the areas in the polygon layer according to the following rules: 1) Areas with an area greater than 0.15 square kilometers are used as basic research units, that is, the smallest rescue areas; 2) Areas with an area less than 0.15 square kilometers are merged nearby; 3) Areas without building distribution are deleted. After the division, 320 basic research units as shown in Figure 2 are obtained.

[0038] S2. Generate the time series data of mobile phone location quantities for the research units:

[0039] Taking the basic research units as the scale, the mobile phone location quantity data with an hourly time interval within 24 hours of 22 days for each basic research unit is statistically analyzed, thereby constructing the time series data of mobile phone location quantities with a data dimension of 24 * 22. The statistical data of a certain basic research unit is shown in Table 1-3 below.

[0040] Table 1

[0041]

[0042] Table 2

[0043]

[0044]

[0045] Table 3

[0046]

[0047] S3. Analyze the law of the change of the location quantity over time and determine the types of earthquake rescue areas for the research units:

[0048] Select 83 sample units from the 320 basic research units. According to the time series data of mobile phone location quantities statistically analyzed for the sample units, analyze the law of the change of the mobile phone location quantity over time for each sample unit, that is, the time series characteristics of the mobile phone location quantity, as shown in Figure 3 , and obtain the corresponding mobile phone location quantity curves. Divide the sample units into the following five types of earthquake rescue areas according to the curves:

[0049] 1) Work area: When there are significant differences in the mobile phone location quantity curve between weekdays and weekends, with a dense population during weekdays and a sparse population on weekends, it is defined as a work area;

[0050] 2) Residential area: When there are obvious peaks and valleys in the mobile phone location quantity curve, and the largest peaks within a day are at 12 - 13:00 noon and from 19:00 to 7:00 the next day, it is defined as a residential area;

[0051] 3) Integrated office and living area: When the change in the positioning volume within a day is relatively small and the rising and falling intervals of the mobile phone positioning volume curves on weekdays and weekends are the same, it is defined as an integrated office and living area;

[0052] 4) Commercial area: When there are no obvious peaks and valleys in the mobile phone positioning volume curve and the high-value area appears from 10:00 to 21:00, it is defined as a commercial area;

[0053] 5) Entertainment area: When there are obvious peaks in the mobile phone positioning volume curve around 8:00 am and around 19:00 pm, and there are small fluctuations and no obvious valleys between 8:00 and 19:00, it is defined as an entertainment area.

[0054] S4. Standardize and normalize the data and create a dataset:

[0055] First, use the Z-score standardization method to standardize the mobile phone positioning volume data of each sample unit, and then use the Z-score normalization method to normalize the 48-hour mobile phone positioning volume of each sample unit for 24 hours on weekdays and weekends:

[0056]

[0057] In the formula, r represents the sample unit, Z r score(t) represents the normalized data of the sample unit r, represents the standard deviation of the positioning volume data on the 22nd within the sample unit r, represents the average value of the 48-hour mobile phone positioning volume of the sample unit r, The mobile phone positioning volume at a certain moment of the sample unit r, the minimum value after normalization is -2, and the maximum value is 2. Randomize the normalized data and group the randomized data according to the ratio of 6:2:2, which are used as the training dataset, validation dataset, and test dataset respectively.

[0058] S5. Build, train, and tune the parameters of the hybrid model:

[0059] Such as Figure 4As shown in the figure, first, a CNN (Convolutional Neural Network) model with three branches is built. The CNN has three CNN1D convolutional layers and one global pooling layer for each CNN branch in sequence. The number of convolutional kernels for each CNN1D convolutional layer is 128. The sizes of the convolutional kernels for the three CNN1D convolutional layers of CNN branch 1 are 21*1, 10*1, and 5*1 respectively. The sizes of the convolutional kernels for the three CNN1D convolutional layers of CNN branch 2 are 16*1, 8*1, and 3*1 respectively. The sizes of the convolutional kernels for the three CNN1D convolutional layers of CNN branch 3 are 11*1, 7*1, and 2*1 respectively. The CNN branches can comprehensively extract the spatial features of the data under multiple scales and multiple receptive fields, that is, extract high-dimensional features through their feature abstraction ability. The three CNN branches are integrated into the double-layer BiLSTM (Bidirectional Long Short-Term Memory Neural Network) layer through the operation of Concatenate, and then connected to the double-layer fully connected layer (Dense) to complete the construction of the hybrid model. Among them, BiLSTM is developed from RNN (Recurrent Neural Network) and consists of a forward LSTM and a backward LSTM. The BiLSTM layer can fully consider the bidirectional correlation between the predecessors and successors of the sample data points, enhance the model's expression ability by increasing the depth of the bidirectional long short-term memory neural network, and complete model classification using the iterative update strategy.

[0060] The training data set with a data dimension of 24*22 is input into the hybrid model for model training, and the accuracy of the model is verified through the validation data set. Due to problems such as sample imbalance and small number of training samples in the training samples, the initial training accuracy of the hybrid model is 75% as shown in the upper half. Subsequently, the parameters of the hybrid model are adjusted through data translation and overfitting sample balancing. After multiple parameter adjustment comparisons, the Adam optimizer is used during training, the learning rate is 0.0001, the batch size (Batch_Size) is 32, the number of training epochs (Epoch) is 30, and the ACC is used as the loss function to evaluate the error between the predicted value and the true value, and the accuracy of the model is improved to 82% as shown in the lower half. Figure 5 Figure 5

[0061] S6. Use the hybrid model to predict the type of key search and rescue areas of the research unit:

[0062] Input the test data set into the hybrid model to predict the corresponding earthquake rescue area type for the research unit. Different earthquake rescue area types correspond to different key search and rescue area types: if the research unit is a work area, it corresponds to the key search and rescue area from 9:00 to 17:00; if the research unit is a residential area, it corresponds to the key search and rescue areas from 12:00 to 13:00 and from 19:00 to 7:00 the next day; if the research unit is an integrated office and living area, it corresponds to the key search and rescue area all day long; if the research unit is a commercial area, it corresponds to the key search and rescue area from 10:00 to 21:00; if the research unit is an entertainment area, it corresponds to the key search and rescue area from 8:00 to 19:00.

[0063] Input other basic research units into the model for prediction. According to the prediction results, a distribution map of the key search and rescue areas for 320 basic research units as shown in Figure 6 is drawn.

[0064] The present invention discloses a method for classifying earthquake rescue zones based on a deep learning hybrid model. On the basis of dividing basic research units, mobile phone positioning data is used to construct sequence samples. High-dimensional features are extracted through the feature abstraction ability of CNN, and the high-dimensional features of the mobile phone positioning sequence are predicted in time series through the BiLSTM layer. Finally, a hybrid model for classifying earthquake rescue zones based on CNN-BiLSTM is built, fully integrating the advantages of the two network models of CNN and RNN. Therefore, the hybrid model of the present invention has high prediction accuracy and strong generalization ability, providing a reliable classification method for earthquake rescue zone division.

[0065] The above embodiments are not limitations to the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by those skilled in the art within the scope of the technical solution of the present invention also fall within the protection scope of the present invention.

Claims

1. A method for classifying earthquake rescue zones based on a deep learning hybrid model, characterized in that: It includes the following processes: S1. Division and determination of basic research units; S2. Generating time-series data of mobile phone positioning amounts for research units; S3. Analyzing the law of the change of the positioning amount over time and determining the earthquake rescue area types of research units; S4. Standardizing and normalizing the data and creating a data set; S5. Building, training and tuning a hybrid model; S6. Using the hybrid model to predict the key search and rescue area types of research units; In step S2, taking the basic research unit as the scale, the mobile phone positioning amount data at the same time interval within 24 hours of 22 days for each research unit was statistically analyzed, thereby constructing time-series data of mobile phone positioning amounts with a data dimension of 24 * 22; In step S3, the law of the change of the mobile phone positioning amount over time for each basic research unit was analyzed to obtain the corresponding mobile phone positioning amount curve, and the research units were divided into corresponding earthquake rescue area types according to the curve; In step S5, first, a CNN model with three branches was built. Each CNN branch was successively equipped with three CNN1D convolutional layers and a global pooling layer. The three CNN branches were integrated into a double-layer BiLSTM layer through the joint operation, and then connected to a double-layer fully connected layer to complete the construction of the hybrid model; the training data set with a data dimension of 24 * 22 was input into the hybrid model for model training, and the accuracy of the model was verified through the validation data set; the hybrid model was tuned by data translation and overfitting sample balancing to improve the accuracy of the model; In step S5, the number of convolutional kernels for each CNN1D convolutional layer was 128. The convolutional kernel sizes of the three CNN1D convolutional layers of CNN branch 1 were 21 * 1, 10 * 1, and 5 * 1 respectively. The convolutional kernel sizes of the three CNN1D convolutional layers of CNN branch 2 were 16 * 1, 8 * 1, and 3 * 1 respectively. The convolutional kernel sizes of the three CNN1D convolutional layers of CNN branch 3 were 11 * 1, 7 * 1, and 2 * 1 respectively.

2. The method for classifying earthquake rescue zones based on a deep learning hybrid model according to claim 1, wherein: In step S1, first, a certain urban block was assumed as the earthquake rescue task area, the road network data of the urban block was processed to obtain a road wire layer, and then the road wire layer was converted into a surface layer by using the GIS method. The areas in the surface layer were divided according to the following rules: 1) Areas with an area greater than 0.15 square kilometers were used as basic research units, that is, the smallest rescue areas; 2) Areas with an area less than 0.15 square kilometers were merged nearby; 3) Areas without building distributions were deleted.

3. The method for classifying earthquake rescue zones based on a deep learning hybrid model according to claim 2, wherein: In step S3, there are five types of divided earthquake rescue area types as follows: 1) Work area: When there are huge differences in the mobile phone positioning amount curve between weekdays and weekends, and the population is dense during weekdays and sparse on weekends, it is defined as a work area; 2) Residential area: When there are obvious peaks and valleys in the mobile phone positioning amount curve, and the largest peaks within a day are from 12:00 to 13:00 and from 19:00 to 7:00 the next day, it is defined as a residential area; 3) Integrated office and living area: When the change of the positioning amount within a day is relatively small and the rising and falling intervals of the mobile phone positioning amount curves on weekdays and weekends are the same, it is defined as an integrated office and living area; 4) Commercial area: When there are no obvious peaks and valleys in the mobile phone positioning volume curve, and the high-value area appears from 10:00 to 21:00 in the morning, it is defined as a commercial area; 5) Entertainment area: When there are obvious peaks in the mobile phone positioning volume curve around 8:00 in the morning and around 19:00 in the evening, and there is a small fluctuation between 8:00 and 19:00 without obvious valleys, it is defined as an entertainment area.

4. The method for classifying earthquake rescue zones based on a deep learning hybrid model according to claim 3, wherein: In step S4, first use the Z-score standardization method to standardize the mobile phone positioning volume data of each research unit, and then use the Z-score normalization method to normalize the 48-hour mobile phone positioning volume of each research unit for 24 hours on weekdays and weekends: ， where r represents the research unit, represents the normalized data of the research unit r, represents the standard deviation of the location data on the 22nd within the research unit r, represents the average value of the mobile phone location quantity in the research unit r for 48 hours, The mobile phone location quantity at a certain moment of the research unit r. The normalized data is shuffled, and the shuffled data is grouped according to the ratio of 6:2:2, which are used as the training data set, the validation data set, and the test data set respectively.

5. The method for classifying earthquake rescue zones based on a deep learning hybrid model according to claim 4, characterized in that: In step S6, input the test data set into the hybrid model to predict the corresponding earthquake rescue area type of the research unit. Different earthquake rescue area types correspond to different key search and rescue area types: If the research unit is a work area, it corresponds to the key search and rescue area from 9:00 to 17:00; if the research unit is a residential area, it corresponds to the key search and rescue areas from 12:00 to 13:00 and from 19:00 to 7:00 the next day; if the research unit is an integrated office and living area, it corresponds to the key search and rescue area throughout the day; if the research unit is a commercial area, it corresponds to the key search and rescue area from 10:00 to 21:00; if the research unit is an entertainment area, it corresponds to the key search and rescue area from 8:00 to 19:00.

Images