A disaster early warning method
By lowering the initial confidence threshold and model training in the cloud-edge collaborative system, the disaster early warning model was optimized, solving the problems of low data processing accuracy and high early warning error rate in the cloud-edge collaborative system, and achieving high-precision, low-latency disaster early warning.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- YANGTZE DEITA GRADUATE SCHOOI OF BEIJING INST OF TECH (JIAXING)
- Filing Date
- 2022-12-16
- Publication Date
- 2026-06-09
AI Technical Summary
Existing cloud-edge collaborative disaster early warning methods suffer from low data processing accuracy and simplistic algorithms, resulting in a high error rate in early warnings.
The target early warning model is selected through the cloud server, the initial confidence threshold is lowered, the result data is aggregated to the edge node for re-labeling and uploaded to the cloud server for training, the target early warning model is updated, and disaster early warning is issued on the edge terminal.
It has improved the accuracy and precision of disaster early warning data processing, reduced delay time and early warning error rate, and enhanced the sensitivity and responsiveness of disaster early warning.
Smart Images
Figure CN116433998B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of disaster early warning technology, and specifically relates to a disaster early warning method. Background Technology
[0002] In recent years, with the instability of global climate change, various natural disasters have emerged one after another, and there are currently many methods for natural disaster early warning. With the development of the Internet, cloud servers play an indispensable role in the natural disaster early warning process. However, although cloud servers can perform large-scale computing tasks in the cloud, they encounter network bandwidth bottlenecks and other problems for applications requiring low latency. To solve this problem, tasks can be computed at the edge, but compared to the computing power of cloud servers, edge computing is limited by the computing power of local edge terminals.
[0003] Since the computational requirements for data processing in different disaster early warning systems vary, cloud-edge collaboration is currently used to process data during various disasters to reduce decision-making time. However, cloud-edge collaboration results in low data processing accuracy for disaster early warnings, and the algorithms used for disaster data detection are limited, leading to a high error rate in early warnings. Summary of the Invention
[0004] The purpose of this invention is to provide a disaster early warning method that can solve the technical problems of low data processing accuracy in disaster early warning under existing cloud-edge collaboration, and the high error rate of early warning due to the use of a single algorithm when detecting disaster data.
[0005] To solve the above-mentioned technical problems, the present invention is implemented as follows:
[0006] This invention provides a disaster early warning method applied to a cloud-edge collaborative system. The cloud-edge collaborative system includes a cloud server, edge nodes, and multiple edge terminals. The cloud server connects to the multiple edge terminals through the edge nodes. The cloud server stores various disaster early warning models. The disaster early warning method includes:
[0007] S101: The cloud server selects the target early warning model from a variety of disaster early warning models based on the nature of the task, and distributes the target early warning model to multiple edge terminals through edge nodes;
[0008] S102: Lower the initial confidence threshold of the target early warning model, aggregate the result data with the predicted target occurrence probability greater than the initial confidence threshold to the edge node, re-label the result data and upload it to the cloud server;
[0009] S103: Use the relabeled data as the training set, retrain the target early warning model on the current edge node, and update the target early warning model;
[0010] S104: Redistribute the updated target warning model to multiple edge terminals to replace the current target warning model;
[0011] S105: Disaster early warning through target early warning models.
[0012] In this embodiment of the invention, when a disaster occurs, the initial confidence threshold of the target early warning model adapted to different scenarios is lowered, and the obtained result data is uploaded to the cloud server as a training set to continuously train and optimize the target early warning model, thereby improving the sensitivity of the target early warning model to the early warning capability when a disaster occurs, improving the accuracy of disaster data processing, and increasing the early warning accuracy. In addition, through cloud-edge collaboration, the disaster data is preprocessed through edge nodes before being uploaded to the cloud server for calculation, reducing data processing latency and lowering the early warning error rate. Attached Figure Description
[0013] Figure 1 This is a flowchart illustrating a disaster early warning method provided in an embodiment of the present invention.
[0014] Figure 2 This is a schematic diagram of the structure of a cloud-edge collaborative system provided in an embodiment of the present invention.
[0015] The realization of the objective, functional characteristics and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation
[0016] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.
[0017] The following description, in conjunction with the accompanying drawings, details a disaster early warning method provided by the present invention through specific embodiments and application scenarios.
[0018] Reference Figure 1 The diagram shows a flowchart of a disaster early warning method provided by an embodiment of the present invention.
[0019] Reference Figure 2 The diagram shows a schematic representation of a cloud-edge collaborative system provided in an embodiment of the present invention.
[0020] This invention provides a disaster early warning method applied to a cloud-edge collaborative system, such as... Figure 2 As shown, the cloud-edge collaborative system includes a cloud server, edge nodes, and multiple edge terminals. The cloud server connects to multiple edge terminals through the edge nodes, and the cloud server stores various disaster early warning models.
[0021] Disaster early warning methods include:
[0022] S101: The cloud server selects the target early warning model from a variety of disaster early warning models based on the nature of the task, and distributes the target early warning model to multiple edge terminals through edge nodes.
[0023] Understandably, the target early warning model is already stored on the cloud server, which can be applied to different scenarios and has a certain early warning capability. When deploying the target early warning model, it can be manually distributed to the edge nodes according to specific task requirements, and then sent to the edge terminals through the edge nodes.
[0024] Among them, various disaster early warning models include: forest fire early warning model, factory fire early warning model, and coal mine high temperature early warning model.
[0025] Optionally, the disaster early warning model includes the YOLO (You-Only-Look-Once) v7 algorithm.
[0026] For safe production, the sensitivity of disaster early warning is crucial, requiring a high response rate from the early warning system. YOLO is a one-stage target detection network. Its detection speed is extremely fast; the standard version can process 45 frames per second, easily running in real-time. The YOLO network is still being continuously updated and iterated. YOLOv7 is currently the most advanced algorithm in the YOLO series, surpassing previous YOLO versions in both accuracy and speed. Applying it to target disaster early warning models can significantly enhance the early warning capabilities of cloud-edge collaborative systems and reduce the error rate of early warnings.
[0027] S102: Lower the initial confidence threshold of the target early warning model, aggregate the result data with the predicted target occurrence probability greater than the initial confidence threshold to the edge node, re-label the result data and upload it to the cloud server.
[0028] Understandably, by continuously lowering the initial confidence threshold of the target early warning model, the sensitivity of the target early warning model can be continuously improved. The cloud-edge collaborative architecture has the characteristics of large-scale scalability in similar application scenarios, which can effectively make up for the problem of insufficient effective data samples in specific scenarios. It adopts a "high sensitivity triggering" strategy, that is, when the predicted probability of a certain environmental monitoring target reaches a certain low threshold, an alarm is triggered.
[0029] Optionally, the initial confidence threshold is 0.25.
[0030] It should be noted that setting the initial confidence threshold to 0.25 can effectively exclude irrelevant data that differs significantly from the target early warning model's monitoring objects, thus avoiding the impact of subsequent processing of irrelevant data. On the other hand, a lower initial confidence threshold can improve the sensitivity of the target prediction model to unexpected disasters or factors that are not obvious before the disaster occurs, thereby improving the prediction accuracy of the target prediction model.
[0031] Furthermore, during subsequent training, the alarm threshold will be continuously increased to improve the accuracy of predictions.
[0032] S103: Use the relabeled data as the training set, retrain the target early warning model on the current edge node, and update the target early warning model.
[0033] Understandably, the results data aggregated and labeled at the edge nodes are used as the training set to train the target prediction model of the corresponding edge terminal stored on the cloud server. During the training process of the target prediction model, by setting an initial threshold, the training set with a higher probability of the predicted target can be selected to train the target prediction model. This can accelerate the convergence speed of the target prediction model during the learning process, and the updated target prediction model can react more quickly in terms of disaster early warning.
[0034] S104: Redistribute the updated target warning model to multiple edge terminals to replace the current target warning model.
[0035] Understandably, the updated target prediction model is obtained through continuous training, and its early warning capability is better than the target prediction model that was manually distributed at the beginning. The updated target prediction model is then redistributed to multiple edge terminals through edge nodes, which can avoid the weakening of early warning capability caused by changes in the monitored environmental data.
[0036] S105: Disaster early warning through target early warning models.
[0037] It should be noted that the updated target early warning model is distributed to each edge terminal through the edge node. Each edge terminal can use the optimized target early warning model to provide disaster warnings for the environment in which it is located and eliminate potential dangers.
[0038] S106: Repeat S102 to S104 periodically according to the preset cycle.
[0039] It should be noted that those skilled in the art can select the size of the preset cycle according to the actual situation.
[0040] The target prediction model initially distributed from the cloud server to the edge terminal has a low false negative rate but a high false positive rate. During the short update process of the target prediction model, a lot of environmental data can be obtained. In order to reduce the false negative rate of the target prediction model, the preset period can be set to 5 days or less at the beginning. After the false negative rate of the target prediction model drops to a certain level, the update time of the preset period can be gradually extended to avoid repeated updates and waste of computing resources.
[0041] S107: Classify the training set according to environmental characteristics to obtain environmental data of multiple categories.
[0042] It is understandable that the training set is used by the cloud server to update the target cloud server, and this data can be further divided by category.
[0043] For example, a fire warning for a factory can be further divided into warnings for daytime, nighttime, or various weather conditions.
[0044] S107A: If the amount of environmental data in a certain category is less than the preset amount, augmentation processing is performed on the environmental data in this category to increase the amount of environmental data in this category.
[0045] Those skilled in the art can select the preset quantity based on the actual situation. It is understood that when the environmental data for a certain category obtained from classification is too small, the various results obtained from the analysis of this environmental data will be inaccurate. By augmenting the environmental data, the environmental data can be diversified as much as possible, so that the various results obtained from the augmented data have stronger generalization ability.
[0046] S108: Configure multiple basic models under the disaster early warning model, and each basic model corresponds to a different type of environmental data.
[0047] Optionally, multiple base models can include: daytime base model, nighttime base model, rainy day base model, foggy day base model, and thunderstorm base model. In practical applications, more categories of environmental data can be divided according to actual needs to train different base models, thereby expanding the disaster early warning capability based on cloud-edge collaboration to a smaller range, improving the effectiveness of early warning, and reducing the error rate of early warning.
[0048] It should be noted that a variety of sensors are installed on the edge terminal side.
[0049] S109A: Real-time acquisition of multimodal sensor data from multiple edge terminals.
[0050] Understandably, acquiring data in multiple formats through multimodal sensors can reduce the impact of a single data format on prediction results and facilitate subsequent analysis of different environmental characteristics.
[0051] S109B: Based on multimodal sensor data, an attention network consisting of an LSTM temporal feature extraction network, a ResNet feature extraction network, and a Transformer Encoder is constructed in a cloud server, and the attention network is trained.
[0052] It should be noted that when constructing the self-attention network, factors such as smoke, water vapor, and light will also be considered to prevent the detection effect from being poor during the disaster target detection process.
[0053] The goal of self-attention network learning is to train a network that can process multimodal data, uncover the intrinsic relationships between data, and classify them. The cloud server uses various multimodal sensor data received as a training set to summarize a set of classification rules. When a new sample is input, the network can classify the new sample according to its own classification rules, providing a decision-making basis for subsequent early warning.
[0054] In one possible implementation, S109B specifically includes:
[0055] S1091: Pre-train the LSTM temporal feature extraction network and the ResNet feature extraction network using multimodal sensor data.
[0056] In one possible implementation, S1091 specifically includes:
[0057] S1091A: Using an LSTM time-series feature extraction network to perform time-series prediction tasks on the monitoring dataset of the target monitoring stations of the China Meteorological Administration in previous years, time-series information is obtained.
[0058] The input to the LSTM time-series feature extraction network is the continuous monitoring values of a certain indicator over a certain period of time; the output of the LSTM time-series feature extraction network is the possible values y of a certain indicator in the next period of time. t ;
[0059] The loss function can be expressed as:
[0060] S1091B: This project uses a ResNet feature extraction network to perform image classification on the ImageNet dataset containing historical target monitoring stations from the China Meteorological Administration, obtaining image information. Pre-training was completed using a ResNet feature extraction network on the ImageNet dataset.
[0061] The input is a single image, and the output is the possible classification of the image. S1092: The acquired features are concatenated and input into the Transformer Encoder in a single stream, and the attention network is trained using cross-entropy as the loss function.
[0062] In one possible implementation, S1092 specifically includes:
[0063] S1092A: The temporal information is fed into the LSTM temporal feature extraction network to obtain the feature mapping t.
[0064] S1092B: Input the image information into the ResNet feature extraction network to obtain the feature map v.
[0065] S1092C: Obtained by embedding feature maps t and v respectively. and And and splicing z 0 :
[0066]
[0067] Among them, v type and t type These represent the characteristics of image information and time-series information, respectively.
[0068] in,
[0069] S1092D: z 0 The input is fed into the Transformer Encoder, and the output p is passed to the fully connected layer for classification. The attention network is trained using cross-entropy as the loss function L.
[0070]
[0071] Where N represents the total number of samples, M represents the total number of categories, y represents positive and negative samples, and p represents the predicted probability.
[0072] S109C: Deploy the attention network to edge nodes.
[0073] S109D: Input multimodal sensor data into the attention network.
[0074] S109E: Determine environmental features through attention networks.
[0075] Understandably, the self-attention network built on the cloud server is deployed to the edge nodes. The edge nodes collect multimodal sensor data in real time, and based on the collected environmental data, they use the self-attention network to classify and determine which environmental feature the edge terminal belongs to, and then request the cloud server to send the corresponding target basic model.
[0076] S109: Train the corresponding basic model using environmental data.
[0077] S110: Select a target basic model based on environmental characteristics, and use the target basic model for disaster early warning.
[0078] It should be noted that by deploying the network to edge nodes, these nodes use a self-attention network to infer environmental features based on real-time collected environmental information from edge terminals. If the inference result differs from the previous result, a new target base model is requested from the cloud server. This prevents the use of an inaccurate early warning model due to changes in environmental features, thereby improving the responsiveness of disaster early warnings, increasing their accuracy, and preventing various disasters and accidents.
[0079] In this embodiment of the invention, when a disaster occurs, the initial confidence threshold of the target early warning model adapted to different scenarios is lowered, and the obtained result data is uploaded to the cloud server as a training set to continuously train and optimize the target early warning model, thereby improving the sensitivity of the target early warning model to the early warning capability when a disaster occurs, improving the accuracy of disaster data processing, and increasing the early warning accuracy. In addition, through cloud-edge collaboration, the disaster data is preprocessed through edge nodes before being uploaded to the cloud server for calculation, reducing data processing latency and lowering the early warning error rate.
[0080] The above description is merely an embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention should be included within the scope of the claims of the present invention.
Claims
1. A disaster early warning method, characterized in that, This is applied to a cloud-edge collaborative system, which includes a cloud server, edge nodes, and multiple edge terminals. The cloud server connects to the multiple edge terminals through the edge nodes. The cloud server stores various disaster early warning models, and the disaster early warning method includes: S101: The cloud server selects a target early warning model from a variety of disaster early warning models according to the nature of the task, and distributes the target early warning model to multiple edge terminals through the edge nodes; S102: Lower the initial confidence threshold of the target early warning model, aggregate the result data where the predicted probability of the target occurrence is greater than the initial confidence threshold to the edge node, re-label the result data and upload it to the cloud server; S103: Using the relabeled data as the training set, retrain the target early warning model on the current edge nodes, and update the target early warning model; S104: The updated target warning model is redistributed to multiple edge terminals to replace the current target warning model; S105: Disaster early warning is conducted using the aforementioned target early warning model; The disaster early warning method further includes: S107: Classify the training set according to environmental characteristics to obtain environmental data of multiple categories; S108: Configure multiple basic models under the disaster early warning model, and the multiple basic models are integrated with multiple... The environmental data described above correspond one-to-one; S109: Train the corresponding basic model using the environmental data; S110: Select a target basic model based on environmental characteristics, and conduct disaster early warning through the target basic model; Prior to S109, the following is also included: S109A: Real-time acquisition of multimodal sensor data from multiple edge terminals; S109B: Based on the multimodal sensor data, an attention network consisting of an LSTM temporal feature extraction network, a ResNet feature extraction network, and a TransformerEncoder is constructed in the cloud server, and the attention network is trained. S109C: Deploy the attention network to the edge node; S109D: Input the multimodal sensor data into the attention network; S109E: Determine the environmental features through the attention network; Specifically, S109B includes: S1091: Pre-train the LSTM temporal feature extraction network and the ResNet feature extraction network using the multimodal sensor data; S1092: The acquired features are concatenated and input into the TransformerEncoder in a single stream, and the attention network is trained using cross-entropy as the loss function; Specifically, S1091 includes: S1091A: Using the LSTM time-series feature extraction network described above, time-series prediction tasks are performed on the monitoring dataset of the target monitoring stations of the China Meteorological Administration in previous years to obtain time-series information; S1091B: Using the ResNet feature extraction network to perform image classification on the imageNet dataset of the China Meteorological Administration's target monitoring stations from previous years, image information is obtained; Specifically, S1092 includes: S1092A: The time-series information is fed into the LSTM time-series feature extraction network to obtain the feature mapping. t ; S1092B: The image information is fed into the ResNet feature extraction network to obtain the feature map. v ; S1092C: For the feature mapping t and the feature mapping v Each is embedded to obtain and and the and splicing : in and These represent the types and characteristics of image information and temporal information, respectively. S1092D: Will The input is fed into the TransformerEncoder, and the output will be... The data is fed into a fully connected layer for classification, using the cross-entropy as the loss function. L The attention network is trained as follows: in, N Represents the total number of samples. M Represents the total number of categories. y Representing positive and negative samples, This represents the probability of the prediction.
2. The disaster early warning method according to claim 1, characterized in that, The initial confidence threshold is 0.
25.
3. The disaster early warning method according to claim 1, characterized in that, The disaster early warning method also includes: S106: Repeat S102 to S104 periodically according to the preset cycle.
4. The disaster early warning method according to claim 1, characterized in that, Following S107, the following is also included: S107A: If the amount of environmental data in a certain category is less than the preset amount, augmentation processing is performed on the environmental data in this category to increase the amount of environmental data in this category.
5. The disaster early warning method according to claim 1, characterized in that, The various basic models include: daytime basic model, nighttime basic model, rainy day basic model, foggy day basic model, and thunderstorm basic model.