A pre-assessment method for earthquake-induced building damage.

Abstract

This application discloses a pre-assessment method for earthquake-damaged buildings, relating to the fields of artificial intelligence and disaster assessment. By synergistically employing three strategies—historical and random transitions, multi-information social learning, and collision tracking—it can more effectively search for the globally optimal solution in a complex hyperparameter space. Compared to traditional grid search, random search, or single intelligent optimization algorithms, it can find a set of hyperparameters with superior performance. Applying this optimal set of hyperparameters to the pre-assessment model for earthquake-damaged buildings significantly improves the accuracy and generalization ability of identifying building damage levels. Deploying the trained pre-assessment model on a drone allows for rapid access to disaster areas that are difficult or dangerous for humans to reach, enabling image acquisition and identification, and outputting the pre-assessment results for earthquake-damaged buildings, thus improving assessment efficiency.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and disaster assessment technology, and more specifically, to a method for pre-assessing the degree of damage to buildings during earthquakes. Background Technology

[0002] Earthquakes are highly destructive natural disasters that pose a serious threat to people's lives and property. After an earthquake, a rapid and accurate assessment of the damage to buildings in the disaster area is crucial for developing rescue plans, allocating rescue resources, and evaluating the extent of the damage.

[0003] Traditional building damage assessment mainly relies on manual on-site inspections, which has many drawbacks: First, it is inefficient and difficult to complete large-scale assessments within the golden rescue time; second, it is highly dangerous, as aftershocks and secondary disasters pose a threat to the lives of assessment personnel; third, it is highly subjective, and the assessment results are easily influenced by the assessment personnel's experience, emotions, and other factors, lacking objectivity and consistency.

[0004] In recent years, with the development of drone technology and computer vision technology, disaster assessment using drone aerial images has become a research hotspot. Analyzing aerial images using deep learning models (such as semantic segmentation networks) can automatically identify the damaged areas and extent of buildings. However, the performance of deep learning models is highly dependent on the settings of their hyperparameters, such as learning rate, batch size, and number of network layers. Traditional hyperparameter optimization methods, such as grid search and random search, suffer from high computational cost, long processing time, and difficulty in finding the global optimum. Even some advanced optimization algorithms (such as genetic algorithms and particle swarm optimization) may still face challenges such as slow convergence speed and susceptibility to local optima when dealing with high-dimensional, nonlinear hyperparameter optimization problems. Summary of the Invention

[0005] This application aims to provide a pre-assessment method for the degree of earthquake-induced building damage, which addresses the technical problems of low efficiency, high risk, and strong subjectivity in existing building damage assessments, as well as the difficulty in optimizing hyperparameters of deep learning models, resulting in low assessment accuracy.

[0006] This application provides a method for pre-assessing the degree of earthquake damage to buildings, including:

[0007] Acquire sample image data pre-stored by staff or obtained through human-computer interaction, as well as the data label corresponding to each sample image data;

[0008] A pre-defined assessment model for earthquake-induced building damage is constructed using a deep learning algorithm, and a hyperparameter population is generated based on the hyperparameters of the pre-defined assessment model for earthquake-induced building damage.

[0009] Based on the sample image data and the data label corresponding to each sample image data, a multi-level tracking algorithm is used to train each hyperparameter individual in the hyperparameter population to obtain the target hyperparameter individual;

[0010] Based on the target hyperparameter individuals, a pre-set assessment model for earthquake-induced building damage is obtained after training;

[0011] The trained earthquake-damage assessment model was deployed on a drone, and the drone was dispatched to conduct a pre-assessment of earthquake-damage to buildings in the earthquake-stricken area.

[0012] In one possible implementation, a pre-defined assessment model of earthquake-induced building damage is constructed using a deep learning algorithm. Based on the hyperparameters of this pre-defined assessment model, a hyperparameter population is generated, including:

[0013] A pre-defined assessment model for earthquake-induced building damage was constructed using the DeepLabV3+ algorithm.

[0014] For the hyperparameters to be trained in the pre-set assessment model of earthquake-induced building damage, the hyperparameters are randomly initialized between the upper and lower limits, and the randomly initialized hyperparameters are encoded into vectors to obtain individual hyperparameters.

[0015] By repeatedly generating multiple different hyperparameter individuals, a hyperparameter population can be obtained.

[0016] In one possible implementation, based on the sample image data and the data label corresponding to each sample image data, a multi-level tracking algorithm is used to train each hyperparameter individual in the hyperparameter population to obtain the target hyperparameter individual, including:

[0017] For each hyperparameter individual in the hyperparameter population, the fitness value corresponding to the hyperparameter individual is obtained based on the sample image data and the data label corresponding to each sample image data.

[0018] Based on the fitness value corresponding to the hyperparameter individual, determine the optimal hyperparameter individual and the historical optimal value corresponding to each hyperparameter individual from all hyperparameter individuals;

[0019] Based on the historical optimal value corresponding to the hyperparameter individual, a dual-strategy tracking strategy of historical and random transition is adopted to track the hyperparameter individual and obtain the hyperparameter individual after dual-strategy tracking.

[0020] Based on the optimal hyperparameter individual and the historical optimal value corresponding to each hyperparameter individual, a multi-information social tracking strategy is adopted to track the population information of the hyperparameter individuals after the dual-strategy tracking, and to obtain the hyperparameter individuals after the population information tracking.

[0021] Based on the optimal hyperparameter individual, a collision tracking strategy is used to perform collision tracking on the hyperparameter individuals after the population information tracking, to obtain the hyperparameter individuals after collision tracking;

[0022] Obtain the current training count and determine the training stage based on the current training count; wherein, the training stage includes a training completion stage or a training incomplete stage;

[0023] If the training phase is incomplete, then based on the hyperparameter individuals after collision tracking, return to the step of obtaining the fitness value corresponding to the hyperparameter individuals, and proceed to the next training step.

[0024] When the training phase is completed, the target hyperparameter individual is obtained based on the hyperparameter individuals after collision tracking.

[0025] In one possible implementation, for each hyperparameter individual in the hyperparameter population, the fitness value corresponding to the hyperparameter individual is obtained based on the sample image data and the data label corresponding to each sample image data, including:

[0026] For each hyperparameter individual in the hyperparameter population, the hyperparameters contained in the hyperparameter individual are applied to the pre-set assessment model of earthquake building damage, resulting in the pre-set assessment model of earthquake building damage after applying the hyperparameters.

[0027] The sample image data is used as the actual input to the pre-set assessment model of earthquake building damage after applying hyperparameters, and the data label corresponding to the sample image data is used as the expected output of the pre-set assessment model of earthquake building damage after applying hyperparameters to obtain the cross-entropy loss function value.

[0028] The fitness value corresponding to the hyperparameter individual is obtained by adding the cross-entropy loss function value to the non-zero constant term and taking the reciprocal.

[0029] In one possible implementation, based on the historical optimal value corresponding to the hyperparameter individual, a dual-strategy tracking strategy of historical and random transition tracking is used to track the hyperparameter individual to obtain the hyperparameter individual after dual-strategy tracking, including:

[0030] Obtain the current training iteration count, and based on the current training iteration count, obtain the dual-strategy decision factor as follows:

[0031]

[0032] in, Indicates the dual-strategy decision factor. This represents the maximum value of the dual-strategy decision factor. This represents the minimum value of the dual-strategy decision factor. Represents the sine function. Represents pi (π). Indicates the current number of training iterations. Indicates the maximum number of training iterations;

[0033] Based on the dual-strategy decision factors and the historical optimal values ​​corresponding to the hyperparameter individuals, dual-strategy tracking is performed on the hyperparameter individuals to obtain the hyperparameter individuals after dual-strategy tracking:

[0034]

[0035]

[0036] in, Indicates the first t During the training process, the first k The first hyperparameter of the individual d dimensional hyperparameters, Indicates the first k The first hyperparameter individual after the dual-strategy tracking d dimensional hyperparameters, k =1,2,…,K, where K represents the total number of hyperparameter individuals. d =1,2,…,Dim, where Dim represents the total dimension of hyperparameters in an individual hyperparameter. Indicates the first t During the training process, the first k The historical best value corresponding to the individual hyperparameters d dimensional hyperparameters, Represents the cosine function. Indicates the first tracking coefficient. Indicates the second tracking coefficient. Denotes the first hyperparameter of an individual with random hyperparameters. d dimensional hyperparameters, This represents the random transition tracking control coefficient. Indicates the order of a random transition. This represents the total number of random transitions.

[0037] In one possible implementation, based on the optimal hyperparameter individual and the historical best value corresponding to each hyperparameter individual, a multi-information social tracking strategy is used to track the population information of the hyperparameter individuals after the dual-strategy tracking, obtaining the hyperparameter individuals after the population information tracking, including:

[0038] Based on the fitness value of the optimal hyperparameter individual, the adaptive inertia weight is obtained as follows:

[0039]

[0040] in, Indicates adaptive inertia weights, This represents the maximum value of the adaptive inertia weight. This represents the minimum value of the adaptive inertia weight. Indicates the first i The fitness values ​​of hyperparameter individuals after dual-strategy tracking. This represents the fitness value of the individual with the best hyperparameters. This represents the fitness value of the worst-performing hyperparameter individual. This represents the mean fitness value of all hyperparameter individuals after dual-strategy tracking;

[0041] The first and second social tracking factors were obtained as follows:

[0042]

[0043]

[0044] in, Indicates the first i The first social tracking factor corresponding to the hyperparameters of individuals after dual-strategy tracking. Indicates the first i The second social tracking factor corresponding to the hyperparameters of individuals after dual-strategy tracking. Represents the social tracking angle coefficient, and , A This represents the first constant term between [5, 10]. Represents the first random number between (0,1). Represents the social tracking coefficient, and , This represents the tracking period coefficient between [0.5, 2]. Represents the hyperparameters of individuals after all dual-policy tracking. The maximum value in, Represents the hyperparameters of individuals after all dual-policy tracking. The maximum value in;

[0045] Based on the optimal hyperparameter individual, the historical best value corresponding to each hyperparameter individual, the adaptive inertia weight, the first social tracking factor, and the second social tracking factor, population information tracking is performed on the hyperparameter individuals after dual-strategy tracking to obtain the hyperparameter individuals after population information tracking:

[0046]

[0047]

[0048] in, Indicates the first i The first hyperparameter individual after the dual-strategy tracking d dimensional hyperparameters, Indicates the first i The first hyperparameter of an individual after tracking population information. d dimensional hyperparameters, The first hyperparameter of the individual represents the optimal hyperparameter. d dimensional hyperparameters, The first individual represents the mean. d dimensional hyperparameters, Represents the convergence coefficient. Indicates the first learning factor. Indicates the second learning factor. This represents the maximum value of the convergence coefficient. This represents the minimum value of the convergence coefficient. Let s represent the natural constant, and let s represent the trend adjustment coefficient of the convergence coefficient. This represents the adjustment factor for the magnitude of the convergence coefficient.

[0049] In one possible implementation, based on the optimal hyperparameter individual, a collision tracking strategy is used to perform collision tracking on the hyperparameter individuals after the population information tracking, to obtain the hyperparameter individuals after collision tracking, including:

[0050] The collision tracking coefficients are obtained as follows:

[0051]

[0052] in, Indicates the collision tracking coefficient;

[0053] Based on the collision tracking coefficient, the base collision tracking quantity corresponding to the hyperparameter individual after the population information tracking is determined as follows:

[0054]

[0055] in, Indicates the first m The first hyperparameter of an individual after tracking population information. d dimensional hyperparameters, express The corresponding base collision tracking data;

[0056] Based on the aforementioned basic collision tracking data, the real-time collision tracking data is obtained as follows:

[0057]

[0058] in, express The corresponding real-time collision tracking data, Indicates the first m The fitness values ​​of hyperparameter individuals after tracking population information. Indicates the first m The Euclidean distance between hyperparameter individuals and the optimal hyperparameter individual after tracking population information;

[0059] Based on the collision tracking coefficient, the basic collision tracking amount, and the real-time collision tracking amount, collision tracking is performed on the hyperparameter individuals after the population information tracking, resulting in the hyperparameter individuals after collision tracking:

[0060]

[0061] in, Indicates the first m The first hyperparameter individual after the first collision tracking d dimensional hyperparameters, Let represent the second random number between (0,1), and exp represent the exponential function with the natural constant e as the base.

[0062] In one possible implementation, obtaining the current training iteration count and obtaining the training phase based on the current training iteration count includes:

[0063] Get the current number of training iterations;

[0064] Determine whether the current number of training iterations is greater than or equal to the preset maximum number of training iterations. If so, determine that the current training stage is the training completion stage; otherwise, determine that the current training stage is the training incomplete stage.

[0065] In one possible implementation, based on the target hyperparameter individual, a pre-set assessment model for earthquake-induced building damage is obtained after training, including: using the hyperparameters in the target hyperparameter individual as the final hyperparameters of the pre-set assessment model for earthquake-induced building damage, to obtain the pre-set assessment model for earthquake-induced building damage after training.

[0066] In one possible implementation, the trained earthquake-damage building pre-assessment model is deployed on a drone, and the drone is scheduled to conduct an earthquake-damage building pre-assessment at the earthquake-stricken area, including:

[0067] The trained earthquake-damage assessment model was deployed on a drone.

[0068] The drones are deployed to collect building image data in the earthquake-stricken area, and the collected building image data is identified by a pre-set assessment model of the degree of earthquake damage to buildings to obtain the pre-set assessment results of the degree of earthquake damage to buildings.

[0069] Beneficial effects:

[0070] This application provides a pre-assessment method for earthquake-damaged buildings. Through the synergy of three strategies—historical and random transitions, multi-information social learning, and collision tracking—it can more effectively search for the globally optimal solution in a complex hyperparameter space. Compared to traditional grid search, random search, or single intelligent optimization algorithms, it finds a set of hyperparameters with superior performance. Applying this optimal set of hyperparameters to the pre-assessment model for earthquake-damaged buildings significantly improves the accuracy and generalization ability of identifying building damage levels. Deploying the trained pre-assessment model on a drone allows for rapid image acquisition and identification in disaster areas that are difficult or dangerous for humans to access, and outputs the pre-assessment results for earthquake-damaged buildings, thus improving assessment efficiency. Attached Figure Description

[0071] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments of this application will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0072] Figure 1 This is a flowchart of a pre-assessment method for the degree of earthquake-induced building damage proposed in one embodiment of this application;

[0073] Figure 2 This is a flowchart of obtaining a target hyperparameter individual according to an embodiment of this application. Detailed Implementation

[0074] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0075] like Figure 1 As shown in the figure, this application provides a method for pre-assessing the degree of earthquake damage to buildings, including:

[0076] S101. Obtain sample image data pre-stored by staff or obtained through human-computer interaction, as well as the data label corresponding to each sample image data.

[0077] Specifically, aerial images of different earthquake-stricken areas from history are collected and labeled by experts. The segmented areas are categorized into types such as debris, water bodies, undamaged buildings, moderately damaged buildings, mainly damaged buildings, completely damaged buildings, vehicles, roads, trees, ponds, and sand, forming a sample image dataset with corresponding data labels. This not only enables the classification of different areas but also allows for pre-assessment of the degree of building damage based on the classification results.

[0078] S102. A pre-set assessment model for the degree of earthquake-induced building damage is constructed using a deep learning algorithm, and a hyperparameter population is generated based on the hyperparameters of the pre-set assessment model for the degree of earthquake-induced building damage.

[0079] The DeepLabV3+ network can be used as the base model. Determine the hyperparameters to be optimized, such as the initial learning rate, weight decay coefficient, batch size, and dilation rate of dilated convolutions. Each hyperparameter has a corresponding range of values. Randomly generate hyperparameters within this range and encode them into a vector, forming a single hyperparameter. Repeat this process to generate K (e.g., 50) hyperparameter individuals, forming the initial hyperparameter population.

[0080] S103. Based on the sample image data and the data label corresponding to each sample image data, a multi-level tracking algorithm is used to train each hyperparameter individual in the hyperparameter population to obtain the target hyperparameter individual.

[0081] Compared to traditional grid search, random search, or single intelligent optimization algorithms, the multi-level tracking algorithm provided in this application can find a set of hyperparameters with better performance. Applying this set of optimal hyperparameters to deep learning models such as DeepLabV3+ enables the model to learn fully on the training set and exhibit higher classification accuracy and more precise segmentation boundaries on the test set.

[0082] S104. Based on the target hyperparameter individual, obtain the pre-set assessment model of earthquake building damage degree after training.

[0083] In one possible implementation, based on the target hyperparameter individual, a pre-set assessment model for earthquake-induced building damage is obtained after training, including: using the hyperparameters in the target hyperparameter individual as the final hyperparameters of the pre-set assessment model for earthquake-induced building damage, to obtain the pre-set assessment model for earthquake-induced building damage after training.

[0084] S105. Deploy the trained earthquake building damage pre-assessment model on a drone and schedule the drone to conduct earthquake building damage pre-assessment at the earthquake disaster site.

[0085] This application deploys an optimized high-precision model on a drone, forming a complete automated assessment process. The drone can quickly enter disaster areas that are difficult or dangerous for humans to access, conducting large-scale, high-resolution image acquisition. The onboard model processes the images in real time and immediately outputs damage assessment results. This completely changes the inefficient and high-risk model that relies on manual labor, achieving automation and efficiency in the assessment process. The assessment results are objective and consistent, providing timely and reliable data support for the rescue command center, facilitating scientific decision-making, precise deployment of rescue forces, and minimizing casualties and property damage.

[0086] In one possible implementation, a pre-defined assessment model of earthquake-induced building damage is constructed using a deep learning algorithm. Based on the hyperparameters of this pre-defined assessment model, a hyperparameter population is generated, including:

[0087] A pre-assessment model for earthquake-induced building damage was constructed using the DeepLabV3+ algorithm. However, it's worth noting that this pre-assessment model can also be configured as other segmentation models for multi-classification, enabling segmentation and classification of multiple regions and ultimately achieving a pre-assessment of earthquake-induced building damage.

[0088] For the hyperparameters to be trained in the pre-set assessment model of earthquake-induced building damage, random initialization is performed between the upper and lower limits of the hyperparameters, and the randomly initialized hyperparameters are encoded into vectors to obtain individual hyperparameters.

[0089] By repeatedly generating multiple different hyperparameter individuals, a hyperparameter population can be obtained.

[0090] like Figure 2 As shown, based on the sample image data and the data labels corresponding to each sample image data, a multi-level tracking algorithm is used to train each hyperparameter individual in the hyperparameter population to obtain the target hyperparameter individual, including:

[0091] S201. For each hyperparameter individual in the hyperparameter population, obtain the fitness value corresponding to the hyperparameter individual based on the sample image data and the data label corresponding to each sample image data.

[0092] S202. Based on the fitness value corresponding to the hyperparameter individual, determine the optimal hyperparameter individual and the historical optimal value corresponding to each hyperparameter individual from all hyperparameter individuals.

[0093] S203. Based on the historical best value corresponding to the hyperparameter individual, a dual-strategy tracking strategy of historical and random transition is adopted to track the hyperparameter individual and obtain the hyperparameter individual after dual-strategy tracking.

[0094] S204. Based on the optimal hyperparameter individual and the historical optimal value corresponding to each hyperparameter individual, a multi-information social tracking strategy is adopted to track the population information of the hyperparameter individuals after the dual-strategy tracking, and to obtain the hyperparameter individuals after the population information tracking.

[0095] S205. Based on the optimal hyperparameter individual, a collision tracking strategy is used to perform collision tracking on the hyperparameter individuals after the population information tracking, to obtain the hyperparameter individuals after collision tracking.

[0096] S206. Obtain the current training count and determine the training stage based on the current training count. The training stage may include a training completion stage or a training incomplete stage.

[0097] S207. If the training phase is incomplete, then based on the hyperparameter individuals after collision tracking, return to the step of obtaining the fitness value corresponding to the hyperparameter individuals, and proceed to the next training step.

[0098] S208. When the training phase is completed, the target hyperparameter individual is obtained based on the hyperparameter individuals after collision tracking. For example, the optimal hyperparameter individual is re-obtained based on the hyperparameter individuals after collision tracking to obtain the target hyperparameter individual.

[0099] Traditional hyperparameter search methods require trying numerous combinations, resulting in extremely high computational costs. This application's multi-level tracking algorithm, through an intelligent search strategy, can converge to high-quality hyperparameter solutions with fewer iterations and in a shorter time. For example, the dynamic balancing mechanism of the history and random jump strategies avoids ineffective searches; the multi-information social tracking strategy utilizes collective intelligence to accelerate the overall convergence of the population. This means that the entire cycle from model building to final deployment is significantly shortened, enabling the faster preparation of high-precision evaluation models after a disaster, buying valuable time for emergency response. The collision tracking strategy in this application is specifically designed to address this problem. When population diversity decreases or individuals stagnate, the collision tracking strategy applies an adaptive, perturbative "collision force" based on the individual's distance from the optimal solution and its fitness, helping the individual escape local optima and continue exploring the global optimum. This greatly enhances the algorithm's robustness, ensuring that the final target hyperparameter individual is a truly high-quality solution.

[0100] This algorithm, through the synergy of three strategies—history and random transitions, multi-information social learning, and collision tracking—can more effectively search for the global optimum in a complex hyperparameter space. Compared to traditional grid search, random search, or single intelligent optimization algorithms, this method finds a set of hyperparameter combinations with superior performance. Applying this optimal set of hyperparameters to deep learning models such as DeepLabV3+ allows the model to learn fully on the training set and exhibit higher classification accuracy and more precise segmentation boundaries on the test set, thereby significantly improving the accuracy and generalization ability in identifying the degree of damage to buildings (e.g., undamaged, slightly damaged, severely damaged, completely collapsed).

[0101] In one possible implementation, for each hyperparameter individual in the hyperparameter population, the fitness value corresponding to the hyperparameter individual is obtained based on the sample image data and the data label corresponding to each sample image data, including:

[0102] For each hyperparameter individual in the hyperparameter population, the hyperparameters contained in the hyperparameter individual are applied to the pre-set assessment model of earthquake building damage, resulting in the pre-set assessment model of earthquake building damage after applying the hyperparameters.

[0103] The sample image data is used as the actual input to the pre-set assessment model of earthquake building damage after applying hyperparameters, and the data labels corresponding to the sample image data are used as the expected output of the pre-set assessment model of earthquake building damage after applying hyperparameters to obtain the cross-entropy loss function value.

[0104] The fitness value corresponding to the hyperparameter individual is obtained by adding the cross-entropy loss function value to a non-zero constant term (such as 0.0001) and taking the reciprocal.

[0105] In one possible implementation, based on the historical optimal value corresponding to the hyperparameter individual, a dual-strategy tracking strategy of historical and random transition tracking is used to track the hyperparameter individual to obtain the hyperparameter individual after dual-strategy tracking, including:

[0106] Obtain the current training iteration count, and based on the current training iteration count, obtain the dual-strategy decision factor as follows:

[0107]

[0108] in, Indicates the dual-strategy decision factor. This represents the maximum value of the dual-strategy decision factor, which can be set to 0.95; This represents the minimum value of the dual-strategy decision factor, which can be set to 0.001; Represents the sine function. Represents pi (π). Indicates the current number of training iterations. Indicates the maximum number of training iterations;

[0109] Based on the dual-strategy decision factors and the historical optimal values ​​corresponding to the hyperparameter individuals, dual-strategy tracking is performed on the hyperparameter individuals to obtain the hyperparameter individuals after dual-strategy tracking:

[0110]

[0111]

[0112] in, Indicates the first t During the training process, the first k The first hyperparameter of the individual d dimensional hyperparameters, Indicates the first k The first hyperparameter individual after the dual-strategy tracking d dimensional hyperparameters, k =1,2,…,K, where K represents the total number of hyperparameter individuals. d =1,2,…,Dim, where Dim represents the total dimension of hyperparameters in an individual hyperparameter. Indicates the first t During the training process, the first k The historical best value corresponding to the individual hyperparameters d dimensional hyperparameters, Represents the cosine function. This represents the first tracking coefficient, which can be set to a random number between (0, 1). Indicates the second tracking coefficient. Denotes the first hyperparameter of an individual with random hyperparameters. d dimensional hyperparameters, This represents the random transition tracking control coefficient, which can be set to 0.4; Indicates the order of a random transition. This represents the total number of random transitions, which can be set to 3 or 4.

[0113] The historical and random transition tracking strategy provided in this application uses a decision factor that dynamically changes with the training process to intelligently switch between utilizing historical best practices and exploring random unknown regions. In the early stages, it emphasizes exploration to avoid premature convergence; in the later stages, it emphasizes development to accelerate convergence to the optimal solution.

[0114] In one possible implementation, based on the optimal hyperparameter individual and the historical best value corresponding to each hyperparameter individual, a multi-information social tracking strategy is used to track the population information of the hyperparameter individuals after the dual-strategy tracking, obtaining the hyperparameter individuals after the population information tracking, including:

[0115] Based on the fitness value of the optimal hyperparameter individual, the adaptive inertia weight is obtained as follows:

[0116]

[0117] in, Indicates adaptive inertia weights, This represents the maximum value of the adaptive inertia weight, which can be set to 0.95; This represents the minimum value of the adaptive inertia weight, which can be set to 0.001; Indicates the first i The fitness values ​​of hyperparameter individuals after dual-strategy tracking. This represents the fitness value of the individual with the best hyperparameters. This represents the fitness value of the worst-performing hyperparameter individual. This represents the mean fitness value of all hyperparameter individuals after dual-strategy tracking;

[0118] It's worth noting that the optimal and worst hyperparameter individuals can be determined at the start of each training session and remain unchanged throughout that session. Alternatively, whenever a hyperparameter individual changes, the optimal and worst hyperparameter individuals can be updated synchronously.

[0119] The first and second social tracking factors were obtained as follows:

[0120]

[0121]

[0122] in, Indicates the first i The first social tracking factor corresponding to the hyperparameters of individuals after dual-strategy tracking. Indicates the first i The second social tracking factor corresponding to the hyperparameters of individuals after dual-strategy tracking. Represents the social tracking angle coefficient, and , A This represents the first constant term between [5, 10]. Represents the first random number between (0,1). Represents the social tracking coefficient, and , This represents the tracking period coefficient between [0.5, 2]. Represents the hyperparameters of individuals after all dual-policy tracking. The maximum value in, Represents the hyperparameters of individuals after all dual-policy tracking. The maximum value in;

[0123] Based on the optimal hyperparameter individual, the historical best value corresponding to each hyperparameter individual, the adaptive inertia weight, the first social tracking factor, and the second social tracking factor, population information tracking is performed on the hyperparameter individuals after dual-strategy tracking to obtain the hyperparameter individuals after population information tracking:

[0124]

[0125]

[0126] in, Indicates the first i The first hyperparameter individual after the dual-strategy tracking d dimensional hyperparameters, Indicates the first i The first hyperparameter of an individual after tracking population information. d dimensional hyperparameters, The first hyperparameter of the individual represents the optimal hyperparameter. d dimensional hyperparameters, The first individual represents the mean. d Hyperparameters are the hyperparameters of each dimension of an individual, which are the mean hyperparameters of all individuals in the same dimension after tracking all population information. Represents the convergence coefficient. Indicates the first learning factor. Indicates the second learning factor. This represents the maximum value of the convergence coefficient, which can be set to 1; This represents the minimum convergence coefficient, which can be set to a constant between (0.7, 0.9). represents the natural constant, and s represents the trend adjustment coefficient of the convergence coefficient, which can be set to 5; This represents the adjustment factor for the convergence coefficient, which can be set to 0.1.

[0127] The multi-information social tracking strategy provided in this application not only allows individuals to learn from the best individual in the population, but also introduces population mean information and designs adaptive social learning factors and inertia weights. This enables the population to more comprehensively utilize collective intelligence, enhancing the algorithm's global search capability and convergence stability.

[0128] In one possible implementation, based on the optimal hyperparameter individual, a collision tracking strategy is used to perform collision tracking on the hyperparameter individuals after the population information tracking, to obtain the hyperparameter individuals after collision tracking, including:

[0129] The collision tracking coefficients are obtained as follows:

[0130]

[0131] in, Indicates the collision tracking coefficient;

[0132] Based on the collision tracking coefficient, the base collision tracking quantity corresponding to the hyperparameter individual after the population information tracking is determined as follows:

[0133]

[0134] in, Indicates the first m The first hyperparameter of an individual after tracking population information. d dimensional hyperparameters, express The corresponding base collision tracking data;

[0135] Based on the aforementioned basic collision tracking data, the real-time collision tracking data is obtained as follows:

[0136]

[0137] in, express The corresponding real-time collision tracking data, Indicates the first m The fitness values ​​of hyperparameter individuals after tracking population information. Indicates the first m The Euclidean distance between hyperparameter individuals and the optimal hyperparameter individual after tracking population information;

[0138] Based on the collision tracking coefficient, the basic collision tracking amount, and the real-time collision tracking amount, collision tracking is performed on the hyperparameter individuals after the population information tracking, resulting in the hyperparameter individuals after collision tracking:

[0139]

[0140] in, Indicates the first m The first hyperparameter individual after the first collision tracking d dimensional hyperparameters, Let represent the second random number between (0,1), and exp represent the exponential function with the natural constant e as the base.

[0141] The collision tracking strategy provided in this application simulates the process of a particle colliding with the optimal solution. The intensity and direction of the collision are jointly determined by the distance between the individual and the optimal solution, the fitness value, and the training phase. When an individual gets stuck in a local optimum, this strategy can provide a powerful perturbation to help it escape the predicament and continue to explore a better solution.

[0142] Optionally, to ensure training effectiveness, hyperparameter individuals can be processed for out-of-bounds errors after each strategy is executed. Alternatively, a greedy strategy or simulated annealing strategy can be used to control each tracking strategy.

[0143] In one possible implementation, obtaining the current training iteration count and obtaining the training phase based on the current training iteration count includes:

[0144] Get the current number of training iterations;

[0145] Determine whether the current number of training iterations is greater than or equal to the preset maximum number of training iterations. If so, determine that the current training stage is the training completion stage; otherwise, determine that the current training stage is the training incomplete stage.

[0146] In one possible implementation, the trained earthquake-damage building pre-assessment model is deployed on a drone, and the drone is scheduled to conduct an earthquake-damage building pre-assessment at the earthquake-stricken area, including:

[0147] The trained earthquake-damage assessment model was deployed on a drone.

[0148] The drones are deployed to collect building image data in the earthquake-stricken area, and the collected building image data is identified by a pre-set assessment model of the degree of earthquake damage to buildings to obtain the pre-set assessment results of the degree of earthquake damage to buildings.

[0149] The trained model files can be deployed to the onboard computing unit of a drone. After an earthquake, drones are dispatched to the disaster area. The drones use their onboard high-definition cameras to capture images of buildings. The onboard computing unit uses the evaluation model in real time to analyze the images, identify the degree of damage to each building (e.g., marked with different colored boxes), and transmits the evaluation results (images + labels) back to the command center in real time via a wireless communication module.

[0150] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0151] This application describes embodiments with reference to flowchart illustrations and / or block diagrams of methods, apparatuses, electronic devices, and computer program products according to embodiments of this application. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal device, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0152] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing terminal device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0153] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal equipment, causing a series of operational steps to be performed on the computer or other programmable terminal equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable terminal equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0154] Although preferred embodiments of the present application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the embodiments of the present application.

[0155] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes said element.

[0156] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for pre-evaluating the damage degree of an earthquake house, characterized by, The method comprises the following steps: obtaining sample image data pre-stored by a staff or through human-computer interaction and data labels corresponding to each sample image data; constructing a seismic house damage degree preset evaluation model by using a deep learning algorithm, and generating a super parameter population based on the super parameters of the seismic house damage degree preset evaluation model; training each super parameter individual in the super parameter population by using a multi-level tracking algorithm according to the sample image data and the data labels corresponding to each sample image data, to obtain a target super parameter individual; obtaining the seismic house damage degree preset evaluation model after training based on the target super parameter individual; deploying the seismic house damage degree preset evaluation model after training on a UAV, and scheduling the UAV to perform seismic house damage degree preset evaluation on a site in a seismic disaster area; training each super parameter individual in the super parameter population by using a multi-level tracking algorithm according to the sample image data and the data labels corresponding to each sample image data, to obtain a target super parameter individual, comprising: for each super parameter individual in the super parameter population, obtaining a fitness value corresponding to the super parameter individual according to the sample image data and the data labels corresponding to each sample image data; determining an optimal super parameter individual and a historical optimal value corresponding to each super parameter individual from all super parameter individuals according to the fitness value corresponding to the super parameter individual; performing double-strategy tracking on the super parameter individual by using a historical and random jump tracking strategy according to the historical optimal value corresponding to the super parameter individual, to obtain the super parameter individual after double-strategy tracking; performing population information tracking on the super parameter individual after double-strategy tracking by using a multi-information social tracking strategy according to the optimal super parameter individual and the historical optimal value corresponding to each super parameter individual, to obtain the super parameter individual after population information tracking; performing collision tracking on the super parameter individual after population information tracking by using a collision tracking strategy according to the optimal super parameter individual, to obtain the super parameter individual after collision tracking; obtaining a current training frequency and a training stage according to the current training frequency; wherein the training stage comprises a training completion stage or a training incomplete stage; when the training stage is the training incomplete stage, returning to the step of obtaining the fitness value corresponding to the super parameter individual based on the super parameter individual after collision tracking, and entering the next training step; when the training stage is the training completed stage, obtaining the target super parameter individual according to the super parameter individual after collision tracking.

2. The method of claim 1, wherein, constructing a seismic house damage degree preset evaluation model by using a deep learning algorithm, and generating a super parameter population based on the super parameters of the seismic house damage degree preset evaluation model, comprising: constructing a seismic house damage degree preset evaluation model by using a DeepLabV3+ algorithm; randomly initializing the super parameters between the upper limit of the super parameters and the lower limit of the super parameters for the super parameters to be trained in the seismic house damage degree preset evaluation model, and encoding the super parameters after random initialization into a vector to obtain a super parameter individual; The multiple different hyperparameter individuals are repeatedly generated to obtain a hyperparameter population.

3. The method of claim 1, wherein, For each hyperparameter individual in the hyperparameter population, a fitness value corresponding to the hyperparameter individual is obtained according to the sample image data and the data label corresponding to each sample image data, including: For each hyperparameter individual in the hyperparameter population, the hyperparameters contained in the hyperparameter individual are applied to the seismic house damage degree preset evaluation model to obtain the seismic house damage degree preset evaluation model after the hyperparameters are applied. The sample image data is taken as the actual input of the seismic house damage degree preset evaluation model after the hyperparameters are applied, and the data label corresponding to the sample image data is taken as the expected output of the seismic house damage degree preset evaluation model after the hyperparameters are applied, to obtain a cross-entropy loss function value. The cross-entropy loss function value is added to a non-zero constant term, and then the reciprocal is taken to obtain the fitness value corresponding to the hyperparameter individual.

4. The method of claim 1, wherein, According to the historical optimal value corresponding to the hyperparameter individual, a history and random jump tracking strategy is adopted to perform double-strategy tracking on the hyperparameter individual to obtain a hyperparameter individual after double-strategy tracking, including: A current training number is obtained, and a double-strategy decision factor is obtained according to the current training number: wherein, denotes a double strategy decision factor, denotes a maximum value of the double strategy decision factor, denotes a minimum value of the double strategy decision factor, denotes a sine function, denotes a circle constant, denotes a current training number, denotes a maximum training number; According to the double-strategy decision factor and the historical optimal value corresponding to the hyperparameter individual, the hyperparameter individual is double-strategy tracked to obtain a hyperparameter individual after double-strategy tracking: wherein, represents the i-th hyperparameter of the j-th hyperparameter individual in the i-th training process, t k d represents the i-th hyperparameter of the j-th hyperparameter individual in the i-th training process, k d represents the i-th hyperparameter of the j-th hyperparameter individual in the i-th training process, k = 1, 2, …, K, K represents the total number of hyperparameter individuals, d = 1, 2, …, Dim, Dim represents the total dimension of hyperparameters in the hyperparameter individual, represents the i-th hyperparameter of the j-th hyperparameter individual in the i-th training process, t k d represents the i-th hyperparameter of the j-th hyperparameter individual in the i-th training process, represents a cosine function, represents a first tracking coefficient, represents a second tracking coefficient, represents the i-th hyperparameter of the j-th hyperparameter individual in the i-th training process, d represents a random hyperparameter individual, represents a random jump tracking control coefficient, represents a random jump order, represents a total random jump order.​​​​​​ 5. The method of claim 4, wherein, According to the optimal hyperparameter individual and the historical optimal value corresponding to each hyperparameter individual, a multi-information social tracking strategy is adopted to perform population information tracking on the hyperparameter individual after double-strategy tracking to obtain a hyperparameter individual after population information tracking, including: According to the fitness value of the optimal hyperparameter individual, an adaptive inertia weight is obtained: wherein, denotes the adaptive inertia weight, denotes the maximum value of the adaptive inertia weight, denotes the minimum value of the adaptive inertia weight, denotes the fitness value of the hyperparameter individual after the i double strategy tracking, denotes the fitness value of the best hyperparameter individual, denotes the fitness value of the worst hyperparameter individual, denotes the mean value of the fitness values of all hyperparameter individuals after the double strategy tracking; A first social tracking factor and a second social tracking factor are obtained: wherein, denotes the first social tracking factor corresponding to the hyperparameter individual after the i th double-strategy tracking, denotes the second social tracking factor corresponding to the hyperparameter individual after the i th double-strategy tracking, denotes the social tracking angle coefficient, and , A denotes the first constant term between [5, 10], denotes the first random number between (0, 1), denotes the social tracking quantity coefficient, and , denotes the tracking period coefficient between [0.5, 2], denotes the maximum value of the corresponding to the hyperparameter individual after all double-strategy tracking, denotes the maximum value of the corresponding to the hyperparameter individual after all double-strategy tracking. According to the optimal hyperparameter individual, the historical optimal value corresponding to each hyperparameter individual, the adaptive inertia weight, the first social tracking factor and the second social tracking factor, the hyperparameter individual after double-strategy tracking is population information tracked to obtain a hyperparameter individual after population information tracking: wherein, represents the j-th hyperparameter of the i-th double strategy tracking after the hyperparameter individual, i d represents the j-th hyperparameter of the hyperparameter individual after the i-th population information tracking, i represents the j-th hyperparameter of the hyperparameter individual after the i-th population information tracking, d represents the j-th hyperparameter of the optimal hyperparameter individual, d represents the j-th hyperparameter of the mean individual, d represents a convergence coefficient, represents a first learning factor, represents a second learning factor, represents a maximum value of the convergence coefficient, represents a minimum value of the convergence coefficient, represents a natural constant, and s represents a change trend adjustment coefficient of the convergence coefficient, represents a size adjustment factor of the convergence coefficient.​​​​​ 6. The method of claim 5, wherein, According to the optimal hyperparameter individual, a collision tracking strategy is adopted to perform collision tracking on the hyperparameter individual after population information tracking to obtain a hyperparameter individual after collision tracking, including: A collision tracking coefficient is obtained: wherein, represents the collision tracking coefficient; According to the collision tracking coefficient, a basic collision tracking amount corresponding to the hyperparameter individual after population information tracking is determined: wherein, represents the m th hyperparameter individual after the population information tracking of the d th dimension hyperparameter, represents the corresponding basic collision tracking quantity; According to the basic collision tracking amount, a real-time collision tracking amount is obtained: wherein, denotes a corresponding real-time collision tracking quantity, denotes the m fitness value of the hyperparameter individual after the tracking of the i-th population information, denotes the m Euclidean distance between the hyperparameter individual after the tracking of the i-th population information and the optimal hyperparameter individual. According to the collision tracking coefficient, the basic collision tracking amount and the real-time collision tracking amount, the hyperparameter individual after population information tracking is collision tracked to obtain a hyperparameter individual after collision tracking: wherein, represents the i-th hyperparameter of the i-th hyperparameter individual after the j-th collision trace, m represents the i-th hyperparameter of the i-th hyperparameter individual after the j-th collision trace, d dimensional hyperparameter, represents a second random number between (0, 1), and exp represents an exponential function with base of the natural constant e.

7. The method of claim 6, wherein, A current training number is obtained, and a training stage is obtained according to the current training number, including: The current training number is obtained. Determine whether the current training frequency is greater than or equal to a preset maximum training frequency, if yes, determine that the current training phase is a training completion phase, otherwise determine that the current training phase is a training incomplete phase.

8. The method of claim 1, wherein, Based on the target hyperparameter individual, an earthquake house damage degree preset evaluation model after training is obtained, including: taking the hyperparameters in the target hyperparameter individual as the final hyperparameters of the earthquake house damage degree preset evaluation model, to obtain the earthquake house damage degree preset evaluation model after training.

9. The seismic building damage extent pre-assessment method according to claim 8, characterized in that, Deploy the earthquake house damage degree preset evaluation model after training on a UAV, and dispatch the UAV to perform earthquake house damage degree preset evaluation on the scene of a seismic disaster area, including: Deploying the earthquake house damage degree preset evaluation model after training on a UAV; Dispatching the UAV to collect house image data on the scene of a seismic disaster area, and identifying the collected house image data through the deployed earthquake house damage degree preset evaluation model to obtain an earthquake house damage degree preset evaluation result.

Images