Target damage assessment method and system, electronic equipment and storage medium

Abstract

The embodiment of the invention provides a target damage assessment method and system, electronic equipment and a storage medium, and belongs to the technical field of artificial intelligence. The method comprises the following steps: inputting first building sensing data into an initial damage assessment model to obtain first damage classification data and a type classification probability, and training the initial damage assessment model according to the first damage classification data and the type classification probability to obtain a first damage assessment model; inputting second building sensing data related to the target building into the first damage assessment model to obtain second damage classification data, adjusting parameters of the first damage assessment model according to the second damage classification data to obtain a second damage assessment model, and finally performing damage assessment on the target building by using the second damage assessment model. Therefore, the generalization ability, the application efficiency and the accuracy of the target damage assessment method across scenes and target types can be improved.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to a target damage assessment method and system, an electronic device and a storage medium. Background Art

[0002] The target damage assessment method aims to analyze and evaluate the impact of physical damage on the target. It can be applied to assess the degree of damage to buildings such as houses or bridges after natural disasters, man-made damage, etc., and can also be applied to assess the anti-damage ability of buildings. However, since most of the existing target damage assessment methods use target damage assessment models for damage assessment, the target damage assessment models are usually pre-trained based on building types such as bungalows, bridges and high-rise buildings, so that the target damage assessment model has a high accuracy only when applied to the trained building types. If the target damage assessment model is applied to new building types that have not been trained, there is a problem of low generalization ability of the target damage assessment method across scenarios and target types, which leads to low accuracy of damage assessment for new types of buildings. If the accuracy of damage assessment for new types of buildings is to be improved, the target damage assessment model needs to be re-trained based on the new building type, which makes the cross-scenario and cross-target type application of the target damage assessment method inefficient. Summary of the invention

[0003] The main purpose of the embodiments of the present application is to propose a target damage assessment method and system, an electronic device and a storage medium, aiming to improve the generalization ability of the target damage assessment method across scenarios and target types, thereby improving the damage assessment accuracy and application efficiency of the target damage assessment method on different types of buildings.

[0004] To achieve the above-mentioned purpose, a first aspect of an embodiment of the present application provides a target damage assessment method, the method comprising: Acquire a first sample from a preset first sample set; wherein the first sample set includes at least two first samples, the first sample includes first building sensor data and a first building type of a sample building, and the first building types of the two first samples are different; Inputting the first building sensor data into a preset initial damage assessment model for damage assessment to obtain first damage assessment data; wherein the first damage assessment data includes: first damage classification data and type classification probability; Performing loss construction according to the first damage classification data to obtain a first damage classification loss function, and performing loss construction according to the type classification probability of the at least two sample buildings and the first building type to obtain a building type confrontation loss function; wherein the building type confrontation loss function is used to make the at least two sample buildings of different building types approach the same in type classification probability distribution; Training the initial damage assessment model according to the first damage classification loss function and the building type resistance loss function to obtain a first damage assessment model; Acquire a second sample related to the target building from a preset second sample set; wherein the second sample includes second building sensor data of the target building; Inputting the second building sensor data into the first damage assessment model to perform damage assessment to obtain second damage assessment data; wherein the second damage assessment data includes second damage classification data; Performing loss construction according to the second damage classification data to obtain a second damage classification loss function; The target damage classifier of the first damage assessment model is parameter-adjusted according to the second damage classification loss function to obtain a second damage assessment model, so as to use the second damage assessment model to perform damage assessment on the target building.

[0005] In some embodiments, before acquiring the second sample related to the target building from the preset second sample set, the method further includes: Acquiring third building sensor data related to the target building; Inputting the third building sensor data into the first damage assessment model for damage assessment to obtain a prediction label and a prediction confidence; wherein the prediction label represents the degree of damage corresponding to the third building sensor data, and the prediction confidence is the confidence of the prediction label; Comparing the prediction confidence with the preset third sample screening threshold to obtain first confidence comparison information; When the first confidence comparison information indicates that the prediction confidence is higher than the third sample screening threshold, selecting the third building sensor data and the prediction label corresponding to the prediction confidence as a third sample; constructing a third sample set according to the third sample; The second sample set is updated according to the third sample set.

[0006] In some embodiments, before updating the second sample set according to the third sample set, the method further includes: Comparing the prediction confidence with the preset fourth sample screening threshold to obtain second confidence comparison information; When the second confidence comparison information indicates that the prediction confidence is higher than the fourth sample screening threshold, selecting the prediction label and the prediction label corresponding to the prediction confidence as the fourth sample; constructing a fourth sample set according to the fourth sample; Sample sampling is performed on the first sample set, the second sample set and the fourth sample set according to a preset evaluation sample sampling ratio to obtain an evaluation sample set.

[0007] In some embodiments, the evaluation sample set includes: a fusion sample subset; the first sample set has a first sample number, the first sample number is the total number of the first samples, the second sample set has a second sample number, the second sample number is the total number of the second samples, the fourth sample set has a fourth sample number, the fourth sample number is the total number of the fourth samples; The step of sampling the first sample set, the second sample set, and the fourth sample set according to a preset evaluation sample sampling ratio to obtain an evaluation sample set includes: Obtaining the first sample number, the second sample number, and the fourth sample number; Determine a fifth sample sampling number according to the evaluation sample sampling ratio and the first sample number, determine a sixth sample sampling number according to the evaluation sample sampling ratio and the second sample number, and determine a seventh sample sampling number according to the evaluation sample sampling ratio and the fourth sample number; Collect samples from the first sample set according to the fifth sample sampling number to obtain a fifth sample set; Collect samples from the second sample set according to the sixth sample sampling number to obtain a sixth sample set; Collect samples from the fourth sample set according to the seventh sample sampling number to obtain a seventh sample set; The fifth sample set, the sixth sample set and the seventh sample set are fused to obtain the fused sample subset.

[0008] In some embodiments, the evaluation sample set further includes: generating a sample subset, the fourth sample set includes: at least one predicted positive sample and at least one predicted negative sample, the first sample set includes at least one sample building positive sample and at least one sample building negative sample, the second sample set includes at least one target building positive sample and at least one target building negative sample; The step of sampling the first sample set, the second sample set, and the fourth sample set according to a preset evaluation sample sampling ratio to obtain an evaluation sample set further includes: Selecting a target positive sample from the predicted positive samples, the sample building positive samples and the target building positive samples; Selecting a target negative sample from the predicted negative samples, the sample building negative samples and the target building negative samples; wherein the target positive sample and the target negative sample come from different sample sets; Linearly adding the target positive sample and the target negative sample according to a preset linear addition ratio to obtain a generated sample; The generated sample subset is constructed according to the generated samples.

[0009] In some embodiments, after adjusting parameters of the target damage classifier of the first damage assessment model according to the second damage classification loss function to obtain the second damage assessment model, the method further includes: Selecting an evaluation sample from the evaluation sample set; wherein the evaluation sample includes evaluation building sensor data and verification damage classification labels, and the verification damage classification labels represent the actual damage degree of the evaluation sample; Inputting the sensor data of the evaluation building into the second damage assessment model for damage assessment to obtain a quality assessment label and an assessment confidence; wherein the quality assessment label represents the degree of damage obtained by performing damage assessment on the sensor data of the evaluation building, and the assessment confidence is the confidence of the quality assessment label; Determining a receiver operating characteristic curve of the second damage assessment model according to the quality assessment label, the assessment confidence and the verification damage classification label; A target confidence threshold is selected from the receiver operating characteristic curve according to a preset recall rate threshold, and the confidence threshold of the second damage assessment model is updated according to the target confidence threshold.

[0010] In some embodiments, the initial damage assessment model further includes: a feature extraction module and a target type adversarial classifier; The step of inputting the first building sensor data into a preset initial damage assessment model to perform damage assessment to obtain first damage assessment data includes: Extracting features from the first building sensor data by the feature extraction module to obtain a first damage assessment feature; Performing target damage classification on the first damage assessment feature by the target damage classifier to obtain the first damage classification data; The first damage assessment feature is subjected to target type adversarial classification by the target type adversarial classifier to obtain the type classification probability.

[0011] To achieve the above-mentioned purpose, a second aspect of an embodiment of the present application provides a target damage assessment system, the system comprising: A first sample acquisition module is used to acquire a first sample from a preset first sample set; wherein the first sample set includes at least two first samples, the first sample includes first building sensor data and a first building type of a sample building, and the first building types of the two first samples are different; A first damage assessment module is used to input the first building sensor data into a preset initial damage assessment model to perform damage assessment to obtain first damage assessment data; wherein the first damage assessment data includes: first damage classification data and type classification probability; a first loss construction module, configured to perform loss construction according to the first damage classification data to obtain a first damage classification loss function, and perform loss construction according to the type classification probabilities of the at least two sample buildings and the first building type to obtain a building type confrontation loss function; wherein the building type confrontation loss function is configured to make the at least two sample buildings of different building types approach the same in type classification probability distribution; A first training module, configured to train the initial damage assessment model according to the first damage classification loss function and the building type confrontation loss function to obtain a first damage assessment model; A second sample acquisition module, used to acquire a second sample related to the target building from a preset second sample set; wherein the second sample includes second building sensor data of the target building; a second damage assessment module, configured to input the second building sensor data into the first damage assessment model for damage assessment to obtain second damage assessment data; wherein the second damage assessment data includes second damage classification data; A second loss construction module, configured to perform loss construction according to the second damage classification data to obtain a second damage classification loss function; The second training module is used to adjust the parameters of the target damage classifier of the first damage assessment model according to the second damage classification loss function to obtain a second damage assessment model, so as to use the second damage assessment model to perform damage assessment on the target building.

[0012] To achieve the above-mentioned purpose, the third aspect of an embodiment of the present application proposes an electronic device, which includes a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the method described in the first aspect is implemented.

[0013] To achieve the above-mentioned purpose, the fourth aspect of an embodiment of the present application proposes a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method described in the first aspect is implemented.

[0014] The target damage assessment method and system, electronic device and storage medium proposed in the present application make the damage classification data output by the first damage assessment model close to the actual damage degree of the building through the first damage classification loss function, and make at least two sample buildings of different building types approach the same in type classification probability distribution through the building type adversarial loss function, so that the first damage assessment model can perform damage assessment on the input sensor data that is only related to the damage degree but not to the building type, thereby improving the damage assessment accuracy of the target damage assessment method on different types of buildings, that is, improving the generalization ability of the target damage assessment method across scenarios and target types on different types of buildings. Furthermore, the target damage assessment method and system, electronic device and storage medium proposed in the present application obtain a first damage assessment model by training an initial damage assessment model according to first samples of different sample buildings, and then obtain a second damage assessment model by training the first damage assessment model according to second samples related to the target building. When the target damage assessment model is deployed to a new building type, only a small amount of second samples need to be collected to adjust the parameters of the first damage assessment model, so as to obtain a second damage assessment model with higher accuracy in damage assessment of the target building, thereby improving the cross-scenario and cross-target type application efficiency of the target damage assessment method. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 This is an optional flow chart of the target damage assessment method provided in the embodiment of the present application; Figure 2 is a schematic diagram of the structure of the initial damage assessment model provided in an embodiment of the present application; Figure 3 yes Figure 1 Flow chart of step S102 in FIG. Figure 4 This is another optional flow chart of the target damage assessment method provided in the embodiment of the present application; Figure 5 This is another optional flow chart of the target damage assessment method provided in the embodiment of the present application; Figure 6 yes Figure 5 Step S504 in an optional flowchart; Figure 7 yes Figure 5 Another optional flowchart of step S504 in; Figure 8 This is another optional flow chart of the target damage assessment method provided in the embodiment of the present application; Fig. 9 It is a schematic diagram of the overall architecture of the target damage assessment method provided in the embodiment of the present application; Fig.10is a schematic diagram of the structure of a target damage assessment system provided in an embodiment of the present application; Fig.11 It is a schematic diagram of the hardware structure of the electronic device provided in the embodiment of the present application. DETAILED DESCRIPTION

[0016] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0017] It should be noted that, although the functional modules are divided in the system schematic diagram and the logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the system or the order in the flowchart. The terms "first", "second", etc. in the specification, claims and the above drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence.

[0018] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art to which this application belongs. The terms used herein are only for the purpose of describing the embodiments of this application and are not intended to limit this application.

[0019] First, some nouns involved in this application are analyzed: Artificial intelligence (AI) is a new technical science that studies and develops theories, methods, technologies and application systems for simulating, extending and expanding human intelligence. AI is a branch of computer science that attempts to understand the essence of intelligence and produce a new intelligent machine that can respond in a similar way to human intelligence. Research in this field includes image recognition, natural language processing and expert systems. AI can simulate the information process of human consciousness and thinking. AI is also a theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results.

[0020] Target damage assessment aims to analyze and evaluate the degree of damage to the target after it has been physically damaged. It can be applied to assess the degree of damage to buildings such as houses or bridges after natural disasters, man-made damage, etc., and can also be applied to assess the anti-damage capability of buildings. Common low-cost target damage assessment methods are usually based on some sensor data deployed on site for judgment and analysis.

[0021] According to different data analysis methods, target damage assessment methods can be divided into the following two categories: judging the target damage degree based on manual experience, and classifying the target damage degree based on machine learning. However, the target damage assessment method based on manual experience relies on manual experience and needs to be constantly adjusted for different targets, scenes and time. It is costly and cannot be scaled. The method of classifying the target damage degree based on machine learning is generally based on existing labeled data to train the target damage assessment model to classify the damage degree of the target through the target damage assessment model. However, the damage assessment model is usually pre-trained based on building types such as bungalows, bridges and high-rise buildings, and then performs damage assessment on target buildings or scenes with the same pre-trained building types. When the damage assessment model is applied to untrained building types, the accuracy of the damage assessment of the building is low, that is, the generalization ability of the target damage assessment method across scenes and target types is low.

[0022] Based on this, the embodiments of the present application provide a target damage assessment method and system, an electronic device and a storage medium, aiming to improve the generalization capability of the target damage assessment method across scenarios and target types, thereby improving the damage assessment accuracy of the target damage assessment method on different types of buildings and the application efficiency of the target damage assessment method across scenarios and target types.

[0023] The target damage assessment method and system, electronic device and storage medium provided in the embodiments of the present application are specifically described through the following embodiments. First, the target damage assessment method in the embodiments of the present application is described.

[0024] The embodiments of the present application can acquire and process relevant data based on artificial intelligence technology. Among them, artificial intelligence (AI) is the theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results.

[0025] AI basic technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing technology, operating / interactive systems, etc. AI software technologies mainly include computer vision technology, robotics technology, speech processing technology, natural language processing technology, and machine learning / deep learning.

[0026] The target damage assessment method provided in the embodiment of the present application relates to the field of artificial intelligence technology. The target damage assessment method provided in the embodiment of the present application can be applied to a terminal, can be applied to a server side, or can be software running in a terminal or a server side. In some embodiments, the terminal can be a smart phone, a tablet computer, a laptop computer, a desktop computer, etc.; the server side can be configured as an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, CDN, and big data and artificial intelligence platforms; the software can be an application that implements the target damage assessment method, etc., but is not limited to the above forms.

[0027] The present application can be used in many general or special computer system environments or configurations. For example: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, minicomputers, mainframe computers, distributed computing environments including any of the above systems or devices, etc. The present application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform specific tasks or implement specific abstract data types. The present application can also be practiced in distributed computing environments, in which tasks are performed by remote processing devices connected through a communication network. In a distributed computing environment, program modules can be located in local and remote computer storage media including storage devices.

[0028] Figure 1 is an optional flow chart of the target damage assessment method provided in the embodiment of the present application. Figure 1 The method can be applied to target damage assessment of a single target building, or can be applied to buildings in a target scene for target damage assessment, where the target scene contains at least one target building. The method may include but is not limited to steps S101 to S108.

[0029] Step S101, obtaining a first sample from a preset first sample set; wherein the first sample set includes at least two first samples, the first sample includes first building sensor data and a first building type of a sample building, and the first building types of the two first samples are different; Step S102, inputting the first building sensor data into a preset initial damage assessment model to perform damage assessment, thereby obtaining first damage assessment data; wherein the first damage assessment data includes: first damage classification data and type classification probability; Step S103, performing loss construction according to the first damage classification data to obtain a first damage classification loss function, and performing loss construction according to the type classification probability of at least two sample buildings and the first building type to obtain a building type confrontation loss function; wherein the building type confrontation loss function is used to make at least two sample buildings of different building types approach the same in type classification probability distribution; Step S104, training the initial damage assessment model according to the first damage classification loss function and the building type confrontation loss function to obtain a first damage assessment model; Step S105, obtaining a second sample related to the target building from a preset second sample set; wherein the second sample includes second building sensor data of the target building; Step S106, inputting the second building sensor data into the first damage assessment model to perform damage assessment to obtain second damage assessment data; wherein the second damage assessment data includes second damage classification data; Step S107, constructing a loss function according to the second damage classification data to obtain a second damage classification loss function; Step S108, adjusting parameters of the target damage classifier of the first damage assessment model according to the second damage classification loss function to obtain a second damage assessment model, so as to use the second damage assessment model to perform damage assessment on the target building.

[0030] In steps S101 to S108 illustrated in the embodiment of the present application, the damage classification data output by the first damage assessment model is made close to the actual damage degree of the building through the first damage classification loss function, and the type classification probability distribution of at least two sample buildings of different building types is made to be close to the same through the building type adversarial loss function, so that the first damage assessment model can perform damage assessment on the input sensor data that is only related to the damage degree but not to the building type, thereby improving the damage assessment accuracy of the target damage assessment method on different types of buildings, that is, improving the generalization ability of the target damage assessment method across scenarios and target types on different types of buildings. Furthermore, the target damage assessment method and system, electronic device and storage medium proposed in the present application obtain a first damage assessment model by training an initial damage assessment model according to first samples of different sample buildings, and then obtain a second damage assessment model by training the first damage assessment model according to second samples related to the target building, so that when the target damage assessment model is deployed to a new building type, only a small amount of second samples need to be collected to adjust the parameters of the first damage assessment model, so as to obtain a second damage assessment model with a higher accuracy in damage assessment of the target building, so that the target damage assessment method can be quickly and massively applied to target buildings of new building types, thereby improving the cross-scenario and cross-target type application efficiency of the target damage assessment method, shortening the training time of the second damage assessment model, that is, reducing the cost of adapting the target damage assessment method to new types of buildings.

[0031] In step S101 of some embodiments, the sample building is the data source of the first sample and is used to pre-train the initial damage assessment model. The first sample set is a collection of first samples, and the first sample set contains at least first samples from two sample buildings, and the building types of the two sample buildings are different. The first sample is used to describe the building type, the actual damage degree of the building and the sensor data of the sample building. Therefore, the first sample at least includes but is not limited to: the first building type, the first building sensor data and the first damage label. Among them, the first building type is used to describe the building type of the sample building, the first building sensor data is the sensor data collected from the sample building, and the first damage label is used to describe the actual damage degree of the building corresponding to the first building sensor data. For example, the first building type may include but is not limited to: bridges, high-rise buildings, flat buildings, tunnels, ports, factories and historical buildings, etc., and the first building sensor data may include but is not limited to: target temperature, target vibration information and packet loss rate of networking equipment, etc. Based on this, the present application can obtain the first sample from the first sample set. The specific method of obtaining the first sample from the preset first sample set can be: random sampling, or other acquisition methods that can be thought of by other technical personnel in this field, and the present application does not make specific restrictions. It should be noted that the type of the first damage label is an enumeration type. The first damage label can be a binary label or a multi-classification label. The specific first damage label can be set and labeled according to the actual needs of technical personnel in this field. If the first damage label is a binary label, the first damage label is used to describe whether the sample building is actually a damaged building or an intact building, and the first sample set includes at least one sample building positive sample and at least one sample building negative sample. Among them, if the sample building is a damaged building, the first damage label marks the sample building as damaged, and uses the first sample as the sample building positive sample; if the sample building is an intact building, the first damage label marks the sample building as undamaged, and uses the first sample as the sample building negative sample. If the first damage label is a multi-classification label, the first sample is labeled according to the degree of damage received by the sample building. Taking the three-category label as an example, if the sample building is a complete building, the first sample of the sample building is marked as no damage; if the sample building is a slightly damaged building, the first sample of the sample building is marked as slightly damaged; if the sample building is a severely damaged building, the first sample of the sample building is marked as severely damaged. Taking the four-category label as an example, if the sample building is a complete building, the first sample of the sample building is marked as no damage; if the sample building is a slightly damaged building, the first sample of the sample building is marked as slightly damaged; if the sample building is a moderately damaged building, the first sample of the sample building is marked as moderately damaged; if the sample building is a severely damaged building, the first sample of the sample building is marked as severely damaged.

[0032] In step S102 of some embodiments, the initial damage assessment model is an untrained target damage assessment model, which is used to perform damage assessment on the first building sensor data input into the initial damage assessment model to obtain first damage assessment data. The first damage assessment data is used to describe the damage degree and sample building type of the sample building obtained by the initial damage assessment model through damage assessment according to the first building sensor data. Therefore, the first damage assessment data includes: first damage classification data and type classification probability, wherein the first damage classification data describes the damage degree of the sample building obtained by the initial damage assessment model through damage assessment according to the first building sensor data, and the type classification probability is used to describe the type classification probability obtained by the initial damage assessment model classifying the sample building into different building types according to the first building sensor data. On this basis, the first building sensor data can be input into the preset initial damage assessment model for damage assessment to obtain the first damage assessment data. It should be noted that the type of the first damage classification data is the same as the type of the first damage label. For example, if the first damage label is a binary classification label, the first damage classification data is also binary classification data. In addition, the specific implementation example of the first damage classification data is similar to the specific implementation example of the first damage label. You may refer to the specific implementation example of the first damage label, and this application will not go into details.

[0033] See also Figure 2 and Figure 3 In some embodiments, the initial damage assessment model includes: a feature extraction module, a target damage classifier, and a target type adversarial classifier. The feature extraction module is used to extract a first damage assessment feature from the first building sensor data, the target damage classifier is used to classify the damage degree of the sample building of the first sample according to the first damage assessment feature, and the target type adversarial classifier is used to classify the building type of the sample building of the first sample according to the first damage assessment feature, for example, the sample building belongs to a high-rise building, or the sample building belongs to a bridge building, etc. Specifically, the feature extraction module can be constructed based on a deep learning architecture, and the deep learning architecture can be a convolutional neural network (CNN) and a residual network (ResNet), etc. The target damage classifier includes at least one fully connected layer (Full Connection, FC), and the target type adversarial classifier includes at least one FC layer.

[0034] In some embodiments, step S102 may include but is not limited to steps S301 to S303: Step S301, extracting features from the first building sensor data through a feature extraction module to obtain a first damage assessment feature; Step S302, performing target damage classification on the first damage assessment feature by a target damage classifier to obtain first damage classification data; Step S303: Perform target type adversarial classification on the first damage assessment feature through a target type adversarial classifier to obtain a type classification probability.

[0035] In step S301 of some embodiments, a feature extraction module may be used to extract features from the first building sensor data to obtain a first damage assessment feature. The first damage assessment feature is represented by a feature vector. Specifically, taking an 18-layer residual network (ResNet18) as an example, the feature extraction module may include a cascaded 7×7 convolution layer (Cov), a 3×3 maximum pooling layer (Maxpool), four stages (Stage), a global average pooling layer (Global AveragePooling, GAP) and an FC layer, wherein the number of output channels (output channels, c) of the 7×7 convolution layer is 64, the stride (stride, s) is 2, and the padding (padding, p) is 3. The input size of the 7×7 convolution layer is 3×64×64, which means that the input first building sensor data has 3 channels, each channel has a size of 64×64, and the output size of the 7×7 convolution layer is 64×32×32, wherein 64 is the number of output channels and 32×32 is the spatial size of the feature map. The 3×3 maximum pooling layer is used to downsample the output feature map of the 7×7 convolutional layer to reduce the spatial size of the feature map to 16×16, and output the feature map after the spatial size reduction. The first stage of ResNet18 includes two cascaded residual blocks (ResidualBlocks), each residual block includes two cascaded 3×3 convolutional layers, and the number of output channels of each convolution layer for convolution operation is 64, the stride is 1, and the padding is 1. The second stage of ResNet18 includes two cascaded residual blocks, where the output size of the first residual block is 128×8×8; each residual block includes two cascaded 3×3 convolutional layers, the number of output channels of each convolutional layer for convolution operation is 128, the step size of the first convolutional layer is 2, the step size of the second to fourth convolutional layers is 1, and the padding of each convolutional layer is 1; and in the first residual block, the input feature map is convolved by two cascaded 3×3 convolutional layers to obtain the first output feature map of the second stage, and downsampled by a 1×1 convolutional layer with a step size of 2 to obtain the downsampled output feature map of the second stage, and then the first output feature map of the second stage and the downsampled output feature map of the second stage are used together as the input feature map of the second residual block, and the input feature map and output feature map of the second residual block are used as the input feature map of the next residual block.The third stage of ResNet18 includes two cascaded residual blocks, the first residual block includes two cascaded 3×3 convolutional layers, and the output size of the first residual block is 256×4×4; the second residual block includes a 3×3 convolutional layer, the number of output channels of each convolutional layer for convolution operation is 256, the step size of the first convolutional layer is 2, the step size of the second convolutional layer and the third convolutional layer is 1, and the padding of each convolutional layer is 1; and in the first residual block, the input feature map is convolved by two cascaded 3×3 convolutional layers to obtain the second output feature map of the third stage, and is downsampled by a 1×1 convolutional layer with a step size of 2 to obtain the downsampled output feature map of the third stage, and then the second output feature map of the third stage and the downsampled output feature map of the third stage are used together as the input feature map of the second residual block, and the input feature map and output feature map of the second residual block are used as the input feature map of the next residual block. The fourth stage of ResNet18 includes two cascaded residual blocks, where the output size of the first residual block is 512×2×2; each residual block includes two cascaded 3×3 convolutional layers, the number of output channels of each convolutional layer for convolution operation is 512, the stride of the first convolutional layer is 2, the stride of the second to fourth convolutional layers is 1, and the padding of each convolutional layer is 1; the processing method of the output feature map in the fourth stage is similar to that in the second stage, please refer to the second stage. Then the total output feature map of the fourth stage is converted into a fixed-length feature vector through the cascaded global average pooling layer and FC layer to obtain the feature vector of the first damage assessment feature, where the fixed length of the feature vector is 512.

[0036] In step S302 of some embodiments, target damage classification may be performed on the feature vector of the first damage assessment feature by a target damage classifier composed of at least one FC layer to obtain first damage classification data.

[0037] In step S303 of some embodiments, a target type adversarial classification may be performed on the feature vector of the first damage assessment feature by a target type adversarial classifier composed of at least one FC layer to obtain a type classification probability.

[0038] In steps S301 to S303 shown in the embodiment of the present application, first, a feature extraction module is used to extract features from the first building sensor data to obtain a first damage assessment feature, and then a target damage classifier is used to perform target damage classification on the first damage assessment feature to obtain first damage classification data, and a target type adversarial classifier is used to perform target type adversarial classification on the first damage assessment feature to obtain a type classification probability, so that the target damage assessment method shown in the embodiment of the present application can obtain the first damage classification data and type classification probability of the first sample by the initial damage assessment model, so as to facilitate the subsequent training of the initial damage assessment model to obtain the first damage assessment model.

[0039] In step S103 of some embodiments, the first damage classification loss function is used to describe the difference between the first damage classification data and the first damage label, and can be constructed based on the first damage classification data and the first damage label. The building type adversarial loss function can be constructed based on the type classification probability and the first building type, and is used to decouple the target damage classification task of the initial damage assessment model from the first building type, so as to avoid serious interference of different building types on the target damage classification task. Specifically, the specific way in which the building type adversarial loss function decouples the target damage classification task from the first building type can be: constrain at least two sample buildings of different building types to be close to the same in type classification probability distribution.

[0040] In step S104 of some embodiments, the initial damage assessment model may be trained according to the first damage classification loss function and the building type confrontation loss function to obtain the first damage assessment model. Specifically, the Adam method may be used to adjust the learning rate and perform gradient back propagation on the initial damage assessment model. When the first damage classification loss function and the building type confrontation loss function converge, the first damage assessment model is obtained. Therefore, the target damage assessment method illustrated in this embodiment can decouple the target damage classification task and the first building type through the building type confrontation loss function during the process of training the initial damage assessment model according to the first damage classification loss function and the building type confrontation loss function, and continuously weaken the influence of the first building type on feature extraction, so that the second damage assessment feature extracted from the second sensor data by the trained first damage assessment model is only related to the damage classification, but not to the building type, so that the first damage assessment model is not easily disturbed by the building type, thereby improving the robustness of the first damage assessment model.

[0041] See also Figure 4 In some embodiments, before step S105, the target damage assessment method shown in the embodiment of the present application may also include but is not limited to steps S401 to S406: Step S401, acquiring third building sensor data related to the target building; Step S402: input the third building sensor data into the first damage assessment model for damage assessment to obtain a prediction label and a prediction confidence; wherein the prediction label represents the degree of damage corresponding to the third building sensor data, and the prediction confidence is the confidence of the prediction label; Step S403, performing comparison according to a preset third sample screening threshold and prediction confidence to obtain first confidence comparison information; Step S404, when the first confidence comparison information indicates that the prediction confidence is higher than the third sample screening threshold, selecting the third building sensor data and the prediction label corresponding to the prediction confidence as the third sample; Step S405, constructing a third sample set according to the third sample; Step S406: Update the second sample set according to the third sample set.

[0042] Before S401 of some steps, the first damage assessment model needs to be deployed on an online platform, such as a cloud platform, server or edge device such as AWS, Azure, or Alibaba Cloud, and then the input end of the first damage assessment model and the sensor of the target building are connected through a 4G networking device, so that the sensor of the target building can quickly transmit the collected sensor data of the third building to the first damage assessment model, thereby improving the transmission efficiency of the sensor data of the third building. In addition, the deployment cost of the 4G networking device is low, which can save the hardware cost of transmitting the sensor data of the third building to the first damage assessment model.

[0043] In step S401 of some embodiments, the third building sensor data is continuously collected by sensors deployed on the target building. The specific implementation method of collecting the third building sensor data may be to collect the third building sensor data according to a preset collection command, or to collect the third building sensor data according to a specific collection cycle. The specific implementation method of the third building sensor data is similar to the specific real-time method of the second building sensor data. Please refer to the implementation method of the second building sensor data, and this application will not repeat it. It should be noted that the third building sensor data may be the same as the second building sensor data, or the third building sensor data may be different from the second building sensor data.

[0044] In step S402 of some embodiments, the third building sensor data is input into the first damage assessment model for damage assessment, and a prediction label and prediction confidence can be obtained. The prediction label is used to describe the degree of damage corresponding to the third building sensor data. The specific implementation of the prediction label is similar to the specific implementation of the first damage classification data, and the specific implementation of the first damage classification data can be referred to. The prediction confidence is the confidence of the prediction label, which is used to describe the probability that the prediction label is accurate.

[0045] In step S403 of some embodiments, the third sample screening threshold can be pre-stored in the storage space of the online platform or other storage space, and the third building sensor data and the prediction label with a prediction confidence higher than the third sample screening threshold are used as the third sample, which can ensure the authenticity and accuracy of the third sample. Among them, the third sample is used to update the second sample, and the third sample is divided into a positive sample and a negative sample. Specifically, the third sample screening threshold is set according to the type of the prediction label. For example, if the prediction label is a binary classification label, the third sample screening threshold may include: a third positive sample threshold and a third negative sample threshold. The third positive sample threshold is used to screen positive samples in the third sample, and the third negative sample threshold is used to screen negative samples in the third sample. The first confidence comparison information is used to describe the numerical relationship between the third sample screening threshold and the prediction confidence. The first confidence comparison information may include but is not limited to any one of the following: the prediction confidence is greater than the third sample screening threshold, the prediction confidence is equal to the third sample screening threshold, or the prediction confidence is less than the third sample screening threshold.

[0046] In step S404 of some embodiments, when the first confidence comparison information represents that the prediction confidence is higher than the third sample screening threshold, the third building sensor data and prediction label corresponding to the prediction confidence can be selected as the third sample. For example, the third positive sample threshold and the third negative sample threshold can both be set to 98%. If the prediction confidence is higher than 98%, the third building sensor data and prediction label corresponding to the prediction confidence are selected as the third sample. Among them, if the prediction label is a damaged building, the third building sensor data and prediction label are used as positive samples in the third sample, and if the prediction label is a complete building, the third building sensor data and prediction label are used as negative samples in the third sample. The third positive sample threshold and the third negative sample threshold can also be set to other specific values. The third positive sample threshold and the third negative sample threshold can be equal, and the third positive sample threshold and the third negative sample threshold can also be unequal, which is not limited in this application.

[0047] In step S405 of some embodiments, a third sample set may be constructed according to the third samples. Specifically, all second samples may be directly combined into the second sample set, and all third samples may be combined into the third sample set.

[0048] In step S406 of some embodiments, the second sample set may be updated according to the third sample set. Specifically, the total number of the third samples in the third sample set is first obtained as the third sample number, and the third sample number is compared with the preset second sample update threshold to obtain update comparison information, wherein the update comparison information indicates that the third sample number is not less than the second sample update threshold or the third sample number is less than the second sample update threshold. If the update comparison information indicates that the third sample number is not less than the second sample update threshold, the preset second sample update ratio is obtained, and the second sample set and the third sample set are sampled according to the second sample update ratio to obtain the target sample set, so as to replace the second sample set with the samples in the target sample set. It should be noted that before the second sample set is updated for the first time according to the third sample set, the second sample is a real sample of the target building, including the sensor data and the real damage type collected from the target building. Since the prediction confidence of the third sample is higher than the third sample screening threshold, it indicates that the third sample is authentic and reliable, and the third sample can be regarded as a real sample of the target building.

[0049] In some embodiments, the specific implementation of sampling the second sample set and the third sample set according to the second sample update ratio to obtain the target sample set includes: first, obtaining the second sample number from the total number of second samples in the second sample set; then, obtaining the third sample collection number by multiplying the third sample number and the second sample update ratio, and obtaining the second sample collection number by multiplying the second sample number and the second sample update ratio; then, randomly collecting the third sample from the third sample according to the third sample collection number, and randomly collecting the second sample from the second sample according to the second sample collection number, so as to use the randomly collected third sample and the randomly collected second sample as the target sample; finally, constructing the target sample set according to the target sample. For example, if the second sample number is 10, the second sample update threshold is 10, and the second sample update ratio is 50%:50%, if the third sample number is 10, it is determined that the third sample collection number and the second sample collection number are both 5, then 5 third samples are randomly sampled from the third sample set, and 5 second samples are randomly sampled from the second sample set, and the collected 5 third samples and 5 second samples form the target sample set; finally, the target sample set replaces the second sample set. Therefore, the target damage assessment method illustrated in the embodiment of the present application can update the second sample set according to the third sample set to fully utilize the incremental target damage assessment information brought by the third sample, while retaining the verified target damage assessment information in the second sample set, thereby ensuring the stability and accuracy when adjusting the parameters of the first damage assessment model according to the updated second sample set.

[0050] It should be noted that the second sample update threshold can also be multiple values, for example, it can be 10, 20 and 30, so that when the number of third samples reaches 10, 20 and 30 respectively, the second sample set is updated according to the third sample set, and then the parameters of the first damage assessment model are adjusted three times respectively to improve the accuracy of damage assessment performed by the second damage assessment model. It should be noted that the total number of third samples in the third sample set can be obtained according to a specific third sample number acquisition period. The third sample number acquisition period can be one day, one hour, or other specific values, which can be selected according to actual needs. It is also possible to obtain the total number of third samples in the third sample set once after each screening of a specific number of third samples. The specific number can be once, twice, or other specific times.

[0051] In steps S401 to S406 shown in the embodiment of the present application, since the third sample comes from the third building sensor data that is continuously collected in real time, by using the predicted label predicted by the first damage assessment model and the third building sensor data as the third sample, and updating the second sample according to the third sample, it is possible to fully utilize the target damage assessment information increment brought by the third sample, retain the target damage assessment information that has been verified in the second sample set, and at the same time ensure the stability and accuracy when adjusting the parameters of the first damage assessment model according to the updated second sample set.

[0052] See also Figure 5 In some embodiments, before step S406, the target damage assessment method may also include but is not limited to steps S501 to S504: Step S501, performing comparison according to a preset fourth sample screening threshold and prediction confidence to obtain second confidence comparison information; Step S502, when the second confidence comparison information represents that the prediction confidence is higher than the fourth sample screening threshold, selecting the prediction label and the prediction label corresponding to the prediction confidence as the fourth sample; Step S503, constructing a fourth sample set according to the fourth sample; Step S504: Sample sampling is performed on the first sample set, the second sample set, and the fourth sample set according to a preset evaluation sample sampling ratio to obtain an evaluation sample set.

[0053] In step S501 of some embodiments, the fourth sample screening threshold is used to use the third building sensor data and prediction label with a prediction confidence higher than the fourth sample screening threshold as the fourth sample to ensure the quality and accuracy of the fourth sample. The fourth sample is used to adjust the confidence threshold of the second damage assessment model. If the prediction label is a binary classification label, the fourth sample can be divided into: a predicted positive sample and a predicted negative sample, wherein the predicted positive sample characterizes the target building as a damaged building, and the predicted negative sample characterizes the target building as a complete building. The fourth sample screening threshold can be pre-stored in the storage space of the online platform or other storage space. Specifically, the fourth sample screening threshold can be set according to the type of prediction label. For example, if the prediction label is a binary classification label, the fourth sample screening threshold may include: a fourth positive sample threshold and a fourth negative sample threshold, wherein the fourth positive sample threshold is used to screen the predicted positive sample, and the fourth negative sample threshold is used to screen the predicted negative sample in the third sample. The second confidence comparison information is used to describe the numerical relationship between the fourth sample screening threshold and the prediction confidence. The second confidence comparison information may include but is not limited to any one of the following: the prediction confidence is greater than the fourth sample screening threshold, the prediction confidence is equal to the fourth sample screening threshold, or the prediction confidence is less than the fourth sample screening threshold.

[0054] In some other embodiments, the sensor data of the fourth building related to the target building may also be collected; wherein the sensor data of the second building, the sensor data of the third building, and the sensor data of the fourth building are all different; and the sensor data of the fourth building is input into the first damage assessment model to obtain a candidate label and a candidate confidence; wherein the candidate label is the degree of damage corresponding to the sensor data of the fourth building, and the candidate confidence is the confidence of the predicted label. Then, the candidate confidence higher than the fourth sample screening threshold is selected as the target label, and the sensor data of the fourth building corresponding to the target label and the candidate confidence are used as the fourth sample.

[0055] In step S502 of some embodiments, when the second confidence comparison information represents that the prediction confidence is higher than the fourth sample screening threshold, the prediction label and the prediction label corresponding to the prediction confidence are selected as the fourth sample. For example, since most of the prediction labels that mark the target building as a complete building are accurate, a lower fourth negative sample threshold can be selected. Therefore, the fourth positive sample threshold can be set to 95%, and the fourth negative sample threshold can be set to 50%. If the prediction confidence is higher than 95% and the prediction label is a damaged building, the third building sensor data and the prediction label corresponding to the prediction confidence are selected as the prediction positive sample; if the prediction confidence is higher than 50% and the prediction label is a complete building, the third building sensor data and the prediction label corresponding to the prediction confidence are selected as the prediction negative sample. The fourth positive sample threshold and the fourth negative sample threshold can also be set to other specific values. The fourth positive sample threshold and the third negative sample threshold can be equal, and the fourth positive sample threshold and the third negative sample threshold can also be unequal, which is not limited in this application.

[0056] In step S503 of some embodiments, a fourth sample set may be constructed according to the fourth samples. Specifically, all fourth samples may be combined to obtain the fourth sample set.

[0057] In step S504 of some embodiments, the evaluation sample sampling ratio is used to collect samples from the first sample set, the second sample set, and the fourth sample to obtain an evaluation sample set. The evaluation sample set is used to update the confidence threshold of the second damage assessment model so that the second damage assessment model outputs more accurate target damage assessment data.

[0058] In steps S501 to S504 shown in the embodiment of the present application, first, a comparison is performed based on the preset fourth sample screening threshold and the prediction confidence to obtain the second confidence comparison information, and when the second confidence comparison information indicates that the prediction confidence is higher than the fourth sample screening threshold, the prediction label and the prediction label corresponding to the prediction confidence are selected as the fourth sample. Then, a fourth sample set is constructed based on the fourth sample, and the first sample set, the second sample set, and the fourth sample set are sampled according to the preset evaluation sample sampling ratio to obtain an evaluation sample set, which can obtain a high-precision evaluation sample set. Moreover, since the evaluation sample set contains sample data of multiple building types, after the evaluation sample set is used to update the confidence threshold of the second damage assessment model, the second damage assessment model can comprehensively perform damage assessment on the target building, thereby improving the accuracy of the target damage assessment model data.

[0059] See also Figure 6In some embodiments, the evaluation sample set includes: a fused sample subset, which is directly constructed based on the evaluation samples sampled from the first sample set, the second sample set, and the fourth sample set. The first sample set has a first sample number, which is the total number of first samples, the second sample set has a second sample number, which is the total number of second samples, and the fourth sample set has a fourth sample number, which is the total number of fourth samples. Step S504 may include, but is not limited to, steps S601 to S605: Step S601, obtaining a first sample number, a second sample number and a fourth sample number; Step S602, determining the fifth sample sampling number according to the evaluation sample sampling ratio and the first sample number, determining the sixth sample sampling number according to the evaluation sample sampling ratio and the second sample number, and determining the seventh sample sampling number according to the evaluation sample sampling ratio and the fourth sample number; Step S603, collecting samples from the first sample set according to the fifth sample sampling number to obtain a fifth sample set; Step S604, collecting samples from the second sample set according to the sixth sample sampling number to obtain a sixth sample set; Step S605, collecting samples from the fourth sample set according to the seventh sample sampling number to obtain a seventh sample set; Step S606: Fuse the fifth sample set, the sixth sample set and the seventh sample set to obtain a fused sample subset.

[0060] In step S602 of some embodiments, the fifth sample sampling number may be determined according to the evaluation sample sampling ratio and the first sample number, the sixth sample sampling number may be determined according to the evaluation sample sampling ratio and the second sample number, and the seventh sample sampling number may be determined according to the evaluation sample sampling ratio and the fourth sample number. If the first sample number is 9, the second sample number is 9, and the fourth sample number is 9, the fifth sample number can be determined to be 3 based on the evaluation sample sampling ratio and the first sample number, the sixth sample number can be determined to be 3 based on the evaluation sample sampling ratio and the second sample number, and the seventh sample number can be determined to be 3 based on the evaluation sample sampling ratio and the fourth sample number. The evaluation sample sampling ratio can also be other specific values, and this application does not make any specific restrictions.

[0061] In step S603 of some embodiments, the first sample set is sampled according to the fifth sample sampling number to obtain the fifth sample set. The specific implementation of sampling the first sample set according to the fifth sample sampling number may be: random sampling. In the following, the specific implementation of sampling the second sample set according to the sixth sample sampling number to obtain the sixth sample set, and the specific implementation of sampling the fourth sample set according to the seventh sample sampling number to obtain the seventh sample set are similar to the specific implementation of sampling the first sample set according to the fifth sample sampling number, and the specific implementation of sampling the first sample set according to the fifth sample sampling number may be referred to. It should be noted that if the first sample number, the second sample number and the fourth sample number are not equal, the samples in the first sample set, the second sample set and the fourth sample set can be resampled or downsampled first, and the resampled or downsampled first sample set is used as the eighth sample set, the resampled or downsampled second sample set is used as the ninth sample set, and the resampled or downsampled fourth sample set is used as the tenth sample set, wherein the total number of samples in the eighth sample set is the eighth sample number, the total number of samples in the ninth sample set is the ninth sample number, the total number of samples in the tenth sample set is the tenth sample number, and the eighth sample number, the ninth sample number and the tenth sample number are equal. Then, the fifth sample sampling number can be determined based on the evaluation sample sampling ratio and the eighth sample number, the sixth sample sampling number can be determined based on the evaluation sample sampling ratio and the ninth sample number, or the seventh sample sampling number can be determined based on the evaluation sample sampling ratio and the tenth sample number. Finally, samples are collected from the eighth sample set according to the fifth sample sampling number to obtain the fifth sample set, samples are collected from the ninth sample set according to the sixth sample sampling number to obtain the sixth sample set, or samples are collected from the tenth sample set according to the seventh sample sampling number to obtain the seventh sample set. Resampling refers to increasing the total number of samples by repeatedly sampling the samples in the sample set, and downsampling refers to reducing the total number of samples by randomly deleting the samples in the sample set.

[0062] In step S606 of some embodiments, the fifth sample set, the sixth sample set and the seventh sample set are fused to obtain a fused sample subset. A specific implementation of fusing the fifth sample set, the sixth sample set and the seventh sample set may be: directly combining the fifth sample set, the sixth sample set and the seventh sample set.

[0063] In steps S601 to S605 shown in the embodiment of the present application, by sampling and fusing the first sample set, the second sample set and the fourth sample set according to the evaluation sample sampling ratio, a fused sample subset is obtained, which can integrate samples from different building types, verified samples and real-time predicted samples, so that the fused sample subset can comprehensively reflect the damage assessment information of different building types, reflect the real-time target damage assessment information increment of the target building, and reflect the verified target damage assessment information of the target building, so as to ensure the stability, reliability and accuracy of the confidence threshold after the confidence threshold is updated according to the second damage assessment model, so as to improve the accuracy of damage assessment by the target damage assessment method.

[0064] In some embodiments, the evaluation sample set further includes: generating a sample subset, wherein the generated sample subset has a generated sample number, which is the total number of generated samples, and the fused sample subset has a fused sample number, which is the total number of fused samples. The fused sample number and the generated sample number may be the same or different. That is, the generated sample subset and the fused sample subset are fused 1:1 to obtain the evaluation sample set. The second sample set includes at least one target building positive sample and at least one target building negative sample. The target building positive sample characterizes that the target building is a damaged building, and the target building negative sample characterizes that the target building is a complete building.

[0065] See also Figure 7 In some embodiments, step S504 may also include but is not limited to steps S701 to S704: Step S701, selecting a target positive sample from the predicted positive samples, the sample building positive samples and the target building positive samples; Step S702, selecting a target negative sample from the predicted negative samples, the sample building negative samples and the target building negative samples; wherein the target positive sample and the target negative sample come from different sample sets; Step S703, linearly adding the target positive sample and the target negative sample according to a preset linear addition ratio to obtain a generated sample; Step S704: construct a generated sample subset according to the generated samples.

[0066] In step S701 of some embodiments, a target positive sample may be selected from the predicted positive samples, the sample building positive samples, and the target building positive samples by random sampling.

[0067] In step S702 of some embodiments, a target negative sample may be selected from the predicted negative samples, the sample building negative samples and the target building negative samples by random sampling, wherein the target positive sample and the target negative sample come from different sample sets, so as to avoid the bias in the generated samples caused by the excessive concentration of sample sources of the target positive sample and the target negative sample.

[0068] In step S703 of some embodiments, the target positive sample and the target negative sample are linearly added according to a preset linear addition ratio to obtain a generated sample, wherein the linear addition ratio is used to control whether the generated sample is a positive sample or a negative sample. The specific implementation method of obtaining the generated sample by linearly adding the target positive sample and the target negative sample can be expressed by the following formula: ; in, Characterize the generated samples, Characterize the target positive sample, Characterize the target negative sample, a is a linear addition ratio, and the value range of a is [0.05-0.15] or [0.85-0.95]. If the value range of a is [0.85-0.95], the generated sample is a positive sample, and if the value range of a is [0.05-0.15], the generated sample is a negative sample.

[0069] In some embodiments, the value range of a can be determined according to the number of positive samples and negative samples in the generated sample subset to control the number of positive samples and negative samples in the generated sample subset, so that the ratio of the number of positive samples to the number of negative samples in the evaluation sample set is close to 1:1, that is, the positive and negative samples in the evaluation sample set are balanced, thereby enhancing the feature recognition capability of the second damage assessment model for positive and negative samples, so as to improve the accuracy of damage assessment performed by the target damage assessment method after the confidence threshold of the second damage assessment model is updated.

[0070] In step S704 of some embodiments, a generated sample subset may be constructed based on the generated samples. Specifically, all generated samples may be directly combined into the generated sample subset.

[0071] In the steps S701 to S704 shown in the embodiment of the present application, by selecting target positive samples and target negative samples from the first sample set, the second sample set, and the fourth sample set, and linearly mixing the target positive samples and the target negative samples to obtain generated samples, multiple samples required for updating the confidence threshold of the second damage assessment model can be quickly generated, reducing the total number of real samples required for collection in the evaluation sample set, thereby reducing the time cost of collecting real samples, thereby reducing the time cost of updating the confidence threshold of the second damage assessment model, and improving the efficiency of updating the confidence threshold of the second damage assessment model. In addition, for some special types of buildings, since it is difficult to collect their real samples, the generated samples obtained by linearly mixing the target positive samples and the target negative samples can make up for the problem of insufficient real sample quantity, provide rich and accurate training samples for the confidence threshold of the second damage assessment model, and thereby improve the accuracy of the second damage assessment model.

[0072] In step S105 of some embodiments, the target building is the data source of the second sample, which is used to train the first damage assessment model. It should be noted that the building type of the target building is different from the building types of all sample buildings. The second sample is used to describe at least the real damage degree and sensor data of the target building. Therefore, the second sample may include but is not limited to: second building sensor data and second damage label, wherein the second building sensor data is sensor data collected from the target building, and the second damage label is used to describe the real damage degree of the building corresponding to the first building sensor data.

[0073] It should be noted that the type of the second building sensor data is the same as the type of the first building sensor data. For example, if the first building sensor data contains the target temperature, the second building sensor data also needs to contain the target temperature; the type of the second damage tag is the same as the type of the first damage tag. For example, if the first damage tag is a binary tag, the second damage tag is also a binary tag. In addition, the specific implementation example of the second building sensor data can refer to the specific implementation example of the first building sensor data. The specific implementation example of the second damage tag is similar to the first damage tag, and the specific implementation example of the first damage tag can refer to the specific implementation example. On this basis, a second sample related to the target building can be obtained. Specifically, a sensor is set on the target building to collect the second building sensor data. The method of collecting the second building sensor data may include but is not limited to: collecting according to a preset sensor data collection instruction or collecting according to a specific period. The specific period may be once a day, once an hour, or once a minute. In addition, in real life scenarios and scenarios such as building anti-damage capability experiments, buildings are usually complete buildings and the target building can generally only be damaged once, the second damage labels in most of the second samples will mark the second samples as negative samples.

[0074] In other embodiments, the source of the second sample may also be other buildings of the same building type as the target building, rather than being limited to the target building, so as to improve the collection efficiency of the second sample.

[0075] In some other embodiments, the second sample may also describe the building type of the target building. Therefore, the second sample may also include but is not limited to: a second building type, wherein the second building type is used to describe the building type of the target building.

[0076] In step S106 of some embodiments, the second damage assessment data is used to describe the damage classification type obtained by the first damage assessment model according to the second building sensor data when performing damage assessment. The specific implementation of the second damage assessment data is similar to the specific implementation of the first damage assessment data, and reference may be made to the specific implementation of the first damage assessment data, which will not be described in detail in this application. It should be noted that the type of the second damage assessment data is the same as the type of the first damage assessment data. It should be noted that the specific implementation of inputting the second building sensor data into the first damage assessment model for damage assessment to obtain the second damage assessment data is similar to the specific implementation of inputting the first building sensor data into the initial damage assessment model for damage assessment, and the only difference is that the parameters of the first damage assessment model and the initial damage assessment model are different. Therefore, the specific implementation of the second damage assessment data may refer to the specific implementation of the initial damage assessment model, which will not be described in detail in this application.

[0077] In step S107 of some embodiments, the second damage classification loss function is used to describe the difference between the second damage classification data and the second damage label, and can be constructed based on the second damage classification data and the second damage label. It should be noted that the first damage assessment model may include, but is not limited to: a feature extraction module, a target damage classifier, and a target type adversarial classifier. However, since the first damage assessment model is only used to perform damage assessment on the second sample of the target building, that is, it only involves the building type of the target building, and does not involve multiple different building types, it is not necessary to adjust the parameters of the target type adversarial classifier. On this basis, it is only necessary to construct a second damage classification loss function, and adjust the parameters of the first damage assessment model based on the second damage classification loss function.

[0078] In step S108 of some embodiments, the target damage classifier of the first damage assessment model is adjusted according to the second damage classification loss function to obtain a second damage assessment model, and then the target building can be assessed for damage using the second damage assessment model. A specific implementation method of adjusting the parameters of the target damage classifier of the first damage assessment model according to the second damage classification loss function may be: using the Adam method to adjust the learning rate of the first damage assessment model and perform gradient back propagation until the second damage classification loss function converges to obtain the second damage assessment model.

[0079] In some embodiments, the specific implementation method of using the second damage assessment model to perform damage assessment on the target building is: obtaining target sensor data of the target building, and performing damage assessment on the target sensor data through the second damage assessment model to obtain target damage assessment data. It should be noted that since the third building sensor data is continuously collected by sensors deployed on the target building, the second sample can be dynamically updated according to multiple second sample update thresholds, and the second damage classification loss function can be reconstructed according to the dynamically updated second sample, so as to dynamically adjust the parameters of the second damage assessment model according to the reconstructed second damage classification loss function.

[0080] See also Figure 8 In some embodiments, after step S108, the target damage assessment method may also include but is not limited to steps S801 to S804: Step S801, selecting an evaluation sample from an evaluation sample set; wherein the evaluation sample includes evaluation building sensor data and verification damage classification labels, and the verification damage classification labels represent the actual damage degree of the evaluation sample; Step S802: Input the sensor data of the evaluation building into the second damage assessment model for damage assessment to obtain a quality assessment label and an assessment confidence; wherein the quality assessment label represents the degree of damage obtained by performing damage assessment on the sensor data of the evaluation building, and the assessment confidence is the confidence of the quality assessment label; Step S803, determining a receiver operating characteristic curve of a second damage assessment model according to the quality assessment label, the assessment confidence and the verification damage classification label; Step S804: selecting a target confidence threshold from the receiver operating characteristic curve according to a preset recall rate threshold, and updating the confidence threshold of the second damage assessment model according to the target confidence threshold.

[0081] In step S801 of some embodiments, evaluating the sample includes evaluating the building sensor data and verifying that the damage classification label represents the real damage degree of the evaluation sample. The evaluation building sensor data is used to update the confidence threshold of the second damage assessment model. The specific implementation of the evaluation building sensor data is similar to the specific implementation of the first building sensor data, and the specific implementation of the first building sensor data can be referred to. Verifying the damage classification label is used to describe the real damage degree of the evaluation sample. The specific implementation of the verification damage classification label is similar to the first damage label, and the specific implementation of the first damage label can be referred to.

[0082] In step S802 of some embodiments, the evaluation building sensor data is input into the second damage assessment model for damage assessment, and a quality assessment label and an assessment confidence can be obtained. The quality assessment label represents the degree of damage obtained by performing damage assessment on the evaluation building sensor data, and its specific implementation is similar to the specific implementation of the first damage classification data, and reference can be made to the specific implementation of the first damage classification data. The assessment confidence is the confidence of the quality assessment label, which is used to describe the probability that the quality assessment label is accurate.

[0083] In step S803 of some embodiments, the receiver operating characteristic curve of the second damage assessment model can be determined based on the quality assessment label, the assessment confidence and the verification damage classification label. The receiver operating characteristic (ROC) curve is used to evaluate the performance of the classification model. Specifically, the evaluation samples are first sorted according to the evaluation confidence to obtain the evaluation sample sequence, and the candidate confidence threshold of the second damage assessment model is obtained; then the true positive rate of the evaluation sample is calculated according to each candidate confidence threshold to obtain the candidate true positive rate, and the false positive rate of the evaluation sample is calculated according to each candidate confidence threshold to obtain the candidate false positive rate; finally, the ROC curve is drawn according to all candidate true positive rates and candidate false positive rates. The true positive rate is also called the recall rate.

[0084] In step S804 of some embodiments, the recall rate threshold is used to select a target confidence threshold from the candidate confidence thresholds. For example, the specific value of the recall rate threshold may be 80%. On this basis, the target confidence threshold may be selected from the receiver operating characteristic curve according to the preset recall rate threshold, and the confidence threshold of the second damage assessment model may be updated according to the target confidence threshold. Specifically, the recall rate threshold is first compared with all candidate true positive rates, and the candidate true positive rate greater than the recall rate threshold is selected as the target true positive rate, and then the candidate confidence threshold corresponding to the target true positive rate is used as the target confidence threshold. Finally, the confidence threshold of the second damage assessment model is replaced with the target confidence threshold to update the confidence threshold of the second damage assessment model.

[0085] It should be noted that, since the third building sensor data is continuously collected by the sensors deployed on the target building, the fourth sample can be dynamically updated according to the third building sensor data collected in different time periods, thereby obtaining a dynamically updated evaluation sample set, and then the confidence threshold of the second damage assessment model is dynamically updated according to the dynamically updated evaluation sample set. In addition, since the time for adjusting the parameters of the first damage assessment model and the second damage assessment model is short after the sensors are deployed on the target building, the number of times the confidence threshold of the second damage assessment model is updated is generally 1-2 times.

[0086] In some embodiments, each time after adjusting the parameters of the first damage assessment model or the second damage assessment model, the confidence threshold of the second damage assessment model may be updated immediately to improve the adaptability of the confidence threshold and parameters of the second damage assessment model, thereby improving the damage assessment accuracy of the second damage assessment model.

[0087] In steps S801 to S804 illustrated in the embodiment of the present application, first, an evaluation sample is selected from the evaluation sample set; wherein the evaluation sample includes evaluation building sensor data and verification damage classification labels, and the verification damage classification labels represent the actual damage degree of the evaluation sample. Then, the evaluation building sensor data is input into the second damage assessment model for damage assessment to obtain a quality assessment label and an assessment confidence; wherein the quality assessment label represents the damage degree obtained by performing damage assessment on the evaluation building sensor data, and the assessment confidence is the confidence of the quality assessment label. Finally, the receiver operating characteristic curve of the second damage assessment model is determined based on the quality assessment label, the assessment confidence and the verification damage classification label, and the target confidence threshold is selected from the receiver operating characteristic curve based on the preset recall rate threshold, so as to update the confidence threshold of the second damage assessment model based on the target confidence threshold. Therefore, the target damage assessment method illustrated in the embodiment of the present application, by determining the target confidence threshold based on the assessment sample and updating the confidence threshold of the second damage assessment model based on the target confidence threshold, can more accurately judge the degree of damage to the target building in the subsequent damage assessment by the second damage assessment model, thereby improving the accuracy and reliability of the second damage assessment model.

[0088] See also Fig. 9The target damage assessment method illustrated in the embodiment of the present application can be divided into the following four processes: training the initial damage classification model, efficient online training of cross-type models, dynamic adjustment of thresholds in the prediction stage, and deployment and application of the model. In the process of training the initial damage classification model, firstly, a first sample is obtained from a preset first sample set; wherein the first sample set includes at least two first samples, and the first sample includes the first building sensor data and the first building type of the sample building, and the first building types of the two first samples are different; then, the first building sensor data is input into the preset initial damage assessment model for damage assessment to obtain first damage assessment data; wherein the first damage assessment data includes: first damage classification data and type classification probability; then, loss is constructed according to the first damage classification data to obtain a first damage classification loss function, and loss is constructed according to the type classification probability of at least two sample buildings and the first building type to obtain a building type adversarial loss function; wherein the building type adversarial loss function is used to make at least two sample buildings of different building types approach the same in type classification probability distribution; finally, the initial damage assessment model is trained according to the first damage classification loss function and the building type adversarial loss function to obtain the first damage assessment model. In the process of efficient online training of cross-type models, first, the sensor data of the third building related to the target building is obtained, and the sensor data of the third building is input into the first damage assessment model for damage assessment to obtain a prediction label and a prediction confidence; wherein the prediction label represents the degree of damage corresponding to the sensor data of the third building, and the prediction confidence is the confidence of the prediction label; then, the third sample is selected according to the preset third sample screening threshold, and then the third sample set is constructed according to the third sample; on this basis, the fourth sample is screened according to the preset fourth sample screening threshold to construct the fourth sample set according to the fourth sample, and the first sample set, the second sample set and the fourth sample set are sampled according to the preset evaluation sample sampling ratio to obtain an evaluation sample set; finally, the second sample set is updated according to the third sample set, and the parameters of the first damage assessment model are adjusted according to the updated second sample set to obtain a second damage assessment model. In the process of dynamically adjusting the threshold in the prediction stage, the evaluation sample is first selected from the evaluation sample set, and then the confidence threshold of the second damage assessment model is updated according to the evaluation sample. In the process of model deployment and application, the second damage assessment model is used to conduct damage assessment on the target building.

[0089] See also Fig.10 The embodiment of the present application further provides a target damage assessment system, which can implement the above-mentioned target damage assessment method, and the system includes: A first sample acquisition module 1001 is used to acquire a first sample from a preset first sample set; wherein the first sample set includes at least two first samples, the first sample includes first building sensor data and a first building type of a sample building, and the first building types of the two first samples are different; The first damage assessment module 1002 is used to input the first building sensor data into a preset initial damage assessment model to perform damage assessment to obtain first damage assessment data; wherein the first damage assessment data includes: first damage classification data and type classification probability; The first loss construction module 1003 is used to perform loss construction according to the first damage classification data to obtain a first damage classification loss function, and perform loss construction according to the type classification probability of at least two sample buildings and the first building type to obtain a building type confrontation loss function; wherein the building type confrontation loss function is used to make at least two sample buildings of different building types approach the same in type classification probability distribution; A first training module 1004 is used to train the initial damage assessment model according to the first damage classification loss function and the building type confrontation loss function to obtain a first damage assessment model; The second sample acquisition module 1005 is used to acquire a second sample related to the target building from a preset second sample set; wherein the second sample includes second building sensor data of the target building; A second damage assessment module 1006 is used to input the second building sensor data into the first damage assessment model to perform damage assessment to obtain second damage assessment data; wherein the second damage assessment data includes second damage classification data; A second loss construction module 1007, configured to construct a loss according to the second damage classification data to obtain a second damage classification loss function; The second training module 1008 is used to adjust the parameters of the target damage classifier of the first damage assessment model according to the second damage classification loss function to obtain a second damage assessment model, so as to use the second damage assessment model to perform damage assessment on the target building.

[0090] The specific implementation of the target damage assessment system is basically the same as the specific implementation of the above-mentioned target damage assessment method, and will not be repeated here.

[0091] The embodiment of the present application also provides an electronic device, the electronic device includes a memory and a processor, the memory stores a computer program, and the processor implements the above-mentioned target damage assessment method when executing the computer program. The electronic device can be any intelligent terminal including a tablet computer, a car computer, etc.

[0092] See also Fig.11 , Fig.11The hardware structure of an electronic device of another embodiment is illustrated, and the electronic device includes: The processor 1101 may be implemented by a general-purpose CPU (Central Processing Unit), a microprocessor, an application-specific integrated circuit (Application Specific Integrated Circuit, ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of the present application; The memory 1102 may be implemented in the form of a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM). The memory 1102 may store an operating system and other application programs. When the technical solution provided in the embodiment of this specification is implemented by software or firmware, the relevant program code is stored in the memory 1102, and the processor 1101 calls and executes the target damage assessment method of the embodiment of this application; Input / output interface 1103, used to implement information input and output; The communication interface 1104 is used to realize the communication interaction between the device and other devices. The communication can be realized through a wired manner (such as USB, network cable, etc.) or a wireless manner (such as mobile network, WIFI, Bluetooth, etc.); The bus 1105 transmits information between the various components of the device (such as the processor 1101, the memory 1102, the input / output interface 1103 and the communication interface 1104); wherein the processor 1101, the memory 1102, the input / output interface 1103 and the communication interface 1104 realize communication connection with each other within the device through the bus 1105.

[0093] The embodiment of the present application also provides a computer-readable storage medium, which stores a computer program, and the computer program implements the above-mentioned target damage assessment method when executed by the processor. The memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer executable programs. In addition, the memory may include a high-speed random access memory, and may also include a non-transitory memory, such as at least one disk storage device, a flash memory device, or other non-transitory solid-state storage device. In some embodiments, the memory may optionally include a memory remotely arranged relative to the processor, and these remote memories may be connected to the processor via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0094] The target damage assessment method and system, electronic device and storage medium proposed in the embodiments of the present application make the damage classification data output by the first damage assessment model close to the actual damage degree of the building through the first damage classification loss function, and make at least two sample buildings of different building types approach the same in type classification probability distribution through the building type adversarial loss function, so that the first damage assessment model can perform damage assessment on the input sensor data that is only related to the damage degree but not to the building type, thereby improving the damage assessment accuracy of the target damage assessment method on different types of buildings, that is, improving the generalization ability of the target damage assessment method across scenarios and target types on different types of buildings. Furthermore, the target damage assessment method and system, electronic device and storage medium proposed in the present application obtain a first damage assessment model by training an initial damage assessment model according to first samples of different sample buildings, and then obtain a second damage assessment model by training the first damage assessment model according to second samples related to the target building, so that when the target damage assessment model is deployed to a new building type, only a small amount of second samples need to be collected to adjust the parameters of the first damage assessment model, so as to obtain a second damage assessment model with a higher accuracy in damage assessment of the target building. This not only enables the target damage assessment method to be quickly and massively applied to target buildings of new building types, thereby improving the cross-scenario and cross-target type application efficiency of the target damage assessment method, but also shortens the training time of the second damage assessment model, that is, reduces the cost of adapting the target damage assessment method to new types of buildings.

[0095] The embodiments described in the embodiments of the present application are intended to more clearly illustrate the technical solutions of the embodiments of the present application and do not constitute a limitation on the technical solutions provided in the embodiments of the present application. Those skilled in the art will appreciate that with the evolution of technology and the emergence of new application scenarios, the technical solutions provided in the embodiments of the present application are also applicable to similar technical problems.

[0096] Those skilled in the art will appreciate that the technical solutions shown in the figures do not constitute a limitation on the embodiments of the present application, and may include more or fewer steps than shown in the figures, or a combination of certain steps, or different steps.

Claims

1. A target damage assessment method, characterized in that: The method comprises: Acquire a first sample from a preset first sample set; wherein the first sample set includes at least two first samples, the first sample includes first building sensor data and a first building type of a sample building, and the first building types of the two first samples are different; Inputting the first building sensor data into a preset initial damage assessment model for damage assessment to obtain first damage assessment data; wherein the first damage assessment data includes: first damage classification data and type classification probability; Performing loss construction according to the first damage classification data to obtain a first damage classification loss function, and performing loss construction according to the type classification probability of the at least two sample buildings and the first building type to obtain a building type confrontation loss function; wherein the building type confrontation loss function is used to make the at least two sample buildings of different building types approach the same in type classification probability distribution; Training the initial damage assessment model according to the first damage classification loss function and the building type resistance loss function to obtain a first damage assessment model; Acquire a second sample related to the target building from a preset second sample set; wherein the second sample includes second building sensor data of the target building; Inputting the second building sensor data into the first damage assessment model to perform damage assessment to obtain second damage assessment data; wherein the second damage assessment data includes second damage classification data; Performing loss construction according to the second damage classification data to obtain a second damage classification loss function; The target damage classifier of the first damage assessment model is parameter-adjusted according to the second damage classification loss function to obtain a second damage assessment model, so as to use the second damage assessment model to perform damage assessment on the target building.

2. The method according to claim 1, characterized in that Before acquiring the second sample related to the target building from the preset second sample set, the method further includes: Acquiring third building sensor data related to the target building; Inputting the third building sensor data into the first damage assessment model for damage assessment to obtain a prediction label and a prediction confidence; wherein the prediction label represents the degree of damage corresponding to the third building sensor data, and the prediction confidence is the confidence of the prediction label; Comparing the prediction confidence with the preset third sample screening threshold to obtain first confidence comparison information; When the first confidence comparison information indicates that the prediction confidence is higher than the third sample screening threshold, selecting the third building sensor data and the prediction label corresponding to the prediction confidence as a third sample; constructing a third sample set according to the third sample; The second sample set is updated according to the third sample set.

3. The method according to claim 2, characterized in that Before updating the second sample set according to the third sample set, the method further includes: Comparing the prediction confidence with the preset fourth sample screening threshold to obtain second confidence comparison information; When the second confidence comparison information indicates that the prediction confidence is higher than the fourth sample screening threshold, selecting the prediction label and the prediction label corresponding to the prediction confidence as the fourth sample; constructing a fourth sample set according to the fourth sample; Sample sampling is performed on the first sample set, the second sample set and the fourth sample set according to a preset evaluation sample sampling ratio to obtain an evaluation sample set.

4. The method according to claim 3, characterized in that The evaluation sample set includes: a fusion sample subset; the first sample set has a first sample number, which is the total number of the first samples; the second sample set has a second sample number, which is the total number of the second samples; the fourth sample set has a fourth sample number, which is the total number of the fourth samples; The step of sampling the first sample set, the second sample set, and the fourth sample set according to a preset evaluation sample sampling ratio to obtain an evaluation sample set includes: Obtaining the first sample number, the second sample number, and the fourth sample number; Determine a fifth sample sampling number according to the evaluation sample sampling ratio and the first sample number, determine a sixth sample sampling number according to the evaluation sample sampling ratio and the second sample number, and determine a seventh sample sampling number according to the evaluation sample sampling ratio and the fourth sample number; Collect samples from the first sample set according to the fifth sample sampling number to obtain a fifth sample set; Collect samples from the second sample set according to the sixth sample sampling number to obtain a sixth sample set; Collect samples from the fourth sample set according to the seventh sample sampling number to obtain a seventh sample set; The fifth sample set, the sixth sample set and the seventh sample set are fused to obtain the fused sample subset.

5. The method according to claim 4, characterized in that The evaluation sample set further includes: generating a sample subset, the fourth sample set includes: at least one predicted positive sample and at least one predicted negative sample, the first sample set includes at least one sample building positive sample and at least one sample building negative sample, the second sample set includes at least one target building positive sample and at least one target building negative sample; The step of sampling the first sample set, the second sample set, and the fourth sample set according to a preset evaluation sample sampling ratio to obtain an evaluation sample set further includes: Selecting a target positive sample from the predicted positive samples, the sample building positive samples and the target building positive samples; Selecting a target negative sample from the predicted negative samples, the sample building negative samples and the target building negative samples; wherein the target positive sample and the target negative sample come from different sample sets; Linearly adding the target positive sample and the target negative sample according to a preset linear addition ratio to obtain a generated sample; The generated sample subset is constructed according to the generated samples.

6. The method according to claim 5, characterized in that After adjusting parameters of the target damage classifier of the first damage assessment model according to the second damage classification loss function to obtain the second damage assessment model, the method further includes: Selecting an evaluation sample from the evaluation sample set; wherein the evaluation sample includes evaluation building sensor data and verification damage classification labels, and the verification damage classification labels represent the actual damage degree of the evaluation sample; Inputting the sensor data of the evaluation building into the second damage assessment model for damage assessment to obtain a quality assessment label and an assessment confidence; wherein the quality assessment label represents the degree of damage obtained by performing damage assessment on the sensor data of the evaluation building, and the assessment confidence is the confidence of the quality assessment label; Determining a receiver operating characteristic curve of the second damage assessment model according to the quality assessment label, the assessment confidence and the verification damage classification label; A target confidence threshold is selected from the receiver operating characteristic curve according to a preset recall rate threshold, and the confidence threshold of the second damage assessment model is updated according to the target confidence threshold.

7. The method according to claim 1, characterized in that The initial damage assessment model also includes: a feature extraction module and a target type adversarial classifier; The step of inputting the first building sensor data into a preset initial damage assessment model to perform damage assessment to obtain first damage assessment data includes: Extracting features from the first building sensor data by the feature extraction module to obtain a first damage assessment feature; Performing target damage classification on the first damage assessment feature by the target damage classifier to obtain the first damage classification data; The first damage assessment feature is subjected to target type adversarial classification by the target type adversarial classifier to obtain the type classification probability.

8. Target damage assessment system, characterized in that: The system comprises: A first sample acquisition module is used to acquire a first sample from a preset first sample set; wherein the first sample set includes at least two first samples, the first sample includes first building sensor data and a first building type of a sample building, and the first building types of the two first samples are different; A first damage assessment module is used to input the first building sensor data into a preset initial damage assessment model to perform damage assessment to obtain first damage assessment data; wherein the first damage assessment data includes: first damage classification data and type classification probability; a first loss construction module, configured to perform loss construction according to the first damage classification data to obtain a first damage classification loss function, and perform loss construction according to the type classification probabilities of the at least two sample buildings and the first building type to obtain a building type confrontation loss function; wherein the building type confrontation loss function is configured to make the at least two sample buildings of different building types approach the same in type classification probability distribution; A first training module, configured to train the initial damage assessment model according to the first damage classification loss function and the building type confrontation loss function to obtain a first damage assessment model; A second sample acquisition module, used to acquire a second sample related to the target building from a preset second sample set; wherein the second sample includes second building sensor data of the target building; a second damage assessment module, configured to input the second building sensor data into the first damage assessment model for damage assessment to obtain second damage assessment data; wherein the second damage assessment data includes second damage classification data; A second loss construction module, configured to perform loss construction according to the second damage classification data to obtain a second damage classification loss function; The second training module is used to adjust the parameters of the target damage classifier of the first damage assessment model according to the second damage classification loss function to obtain a second damage assessment model, so as to use the second damage assessment model to perform damage assessment on the target building.

9. An electronic device, characterized in that The electronic device comprises a memory and a processor, the memory stores a computer program, and the processor implements the target damage assessment method according to any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the target damage assessment method according to any one of claims 1 to 7 is implemented.

Images