5g and ai multi-source fusion decision system for earthquake rescue

Abstract

This invention discloses a 5G and AI multi-source fusion decision-making system for earthquake rescue, belonging to the field of intelligent emergency rescue decision-making technology. The data acquisition module acquires multi-source raw data with spatiotemporal identifiers with low latency through a 5G network; the stage determination module extracts features and generates stage identifiers; based on the stage identifiers, the credibility assessment module and the data reconstruction module respectively perform weighted and reconstructed processing on the data to generate a data representation structure adapted to the current stage; the scheduling and control module receives the stage identifiers and generates trigger signals, controlling the first analysis module and / or the second analysis module to respond to the triggers, respectively analyzing and generating life probability data and environmental accessibility level data from the data representation structure; the fusion decision-making module generates rescue decision instructions based on the stage identifiers through a pre-trained multi-agent reinforcement learning model. This invention improves the timeliness, accuracy, and resource utilization efficiency of earthquake rescue decision-making.

Description

Technical Field

[0001] This invention relates to the field of intelligent emergency rescue decision-making technology, specifically a 5G and AI multi-source fusion decision-making system for earthquake rescue. Background Technology

[0002] Intelligent emergency rescue decision-making technology for earthquakes is an important development direction in the field of disaster emergency management. It aims to generate precise and dynamic rescue action plans by rapidly integrating multi-source heterogeneous information and applying intelligent algorithms. In recent years, with the advancement of 5G communication technology, the Internet of Things, and artificial intelligence algorithms, rescue decision-making is developing towards data-driven and real-time response.

[0003] Existing technologies primarily integrate various monitoring data, such as satellite remote sensing, drone aerial photography, and ground sensors, combined with path planning or resource scheduling algorithms, to generate static or quasi-dynamic rescue guidance plans. These methods can, to some extent, utilize information from different sources to form preliminary judgments on signs of life or road damage in disaster areas, and provide effective auxiliary decision support when data is relatively stable, demonstrating a certain level of information integration and application capabilities. However, earthquake disasters themselves exhibit significant dynamic evolutionary characteristics. From the "golden rescue" phase to the "high incidence of secondary disasters" and then to "post-disaster recovery," the core contradictions, key risks, and data reliability undergo fundamental changes at each stage. Existing solutions are mostly designed with static or single-objective optimization in their architecture, lacking a core mechanism capable of real-time perception and accurate characterization of the current rescue stage. Furthermore, they fail to dynamically and adaptively coordinate the subsequent data processing weights, analytical model focuses, and final decision-making objectives based on this stage of understanding. This makes it difficult for the system to make the optimal balance between multiple objectives such as "racing against time to rescue lives" and "avoiding risks to ensure safety" when facing rapidly changing and complex disaster situations. The timeliness, accuracy and overall rescue efficiency of its output instructions face significant bottlenecks. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a 5G and AI multi-source fusion decision-making system for earthquake rescue.

[0005] To achieve the above objectives, the technical solution of the present invention is as follows:

[0006] This invention discloses a 5G and AI multi-source fusion decision-making system for earthquake rescue, comprising:

[0007] The data acquisition module is used to acquire multi-source raw data with spatiotemporal identifiers with low latency through the 5G network;

[0008] The stage determination module is used to extract stage features from the multi-source raw data and generate a stage identifier to represent the current rescue stage.

[0009] The credibility assessment module is used to assess the credibility of the multi-source raw data based on the stage identifier and generate a weighted data set.

[0010] The data reconstruction module is used to reconstruct the weighted data set based on the stage identifier to generate a data representation structure that adapts to the current rescue stage.

[0011] The scheduling control module is used to receive the stage identifier and generate a trigger signal based on the stage identifier;

[0012] The first analysis module is used to respond to the trigger signal, perform life-related information analysis based on the data representation structure, and generate life existence probability data associated with geographical location;

[0013] The second analysis module is used to respond to the trigger signal, perform environmental access-related information analysis based on the data representation structure, and generate access level data associated with geographical location;

[0014] The fusion decision module is used to jointly analyze the probability data of life presence and the access level data through a pre-trained multi-agent reinforcement learning model based on the stage identifier, and generate rescue decision instructions.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0016] 1. This invention establishes a global awareness and adaptive collaborative mechanism for the dynamic evolution of the rescue system by introducing a stage identifier with a leading role. Based on the actual evolution of rescue work, it can automatically reconstruct the data processing focus and decision-making objectives, overcoming the inherent shortcomings of traditional static architectures in adapting to multi-stage, multi-objective dynamic disasters, and achieving intrinsic synchronization between the system's operational logic and external disaster changes.

[0017] 2. This invention achieves precision and efficiency across the entire chain from data preprocessing to intelligent analysis by using stage identifiers as a technical hub. Through collaborative operation with stage identifiers as unified instructions, the data processing resources and analytical computing power of the entire system can be precisely focused on solving the most pressing rescue sub-problems, thereby improving the targeting and overall efficiency of the information processing process.

[0018] 3. This invention generates action instructions that are more in line with the actual dynamic needs of rescue by integrating intelligent decision-making based on stage perception. The rescue decision instructions integrate the trade-offs of multi-dimensional goals such as timeliness, safety and resource constraints that change with the stage, so that the output plan is no longer a simple application of a fixed strategy, but an intelligent plan with dynamic adaptability and overall coordination, thus having higher operability and effectiveness in actual command. Attached Figure Description

[0019] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. In the drawings, the same reference numerals are used to refer to the same parts. Wherein:

[0020] Figure 1 This is a system module connection diagram of the present invention;

[0021] Figure 2 This is a flowchart of the workflow steps of the present invention;

[0022] Figure 3 This is a flowchart illustrating the steps involved in the credibility assessment of this invention. Detailed Implementation

[0023] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention.

[0024] In existing technologies, intelligent decision-making systems for earthquake rescue largely rely on the integration of multi-source static data or optimization based on a single-objective model, making it difficult to address the multi-stage and multi-objective coordination challenges brought about by the dynamic evolution of the disaster situation. Traditional methods often employ fixed data processing strategies and analysis objectives, potentially leading to delays in life-saving efforts in the early stages of rescue due to excessive focus on environmental safety, or secondary injuries caused by neglecting secondary risks in later stages. Existing architectures lack a core mechanism for real-time and accurate identification of rescue stages, and are unable to dynamically adjust the system's processing logic according to stage changes. This results in insufficient adaptability of the overall solution to rapidly changing disaster situations, making it difficult to generate decision instructions that truly meet the core needs of each stage.

[0025] To address these issues, the study discovered an inherent correlation between core characteristics of the rescue phase (such as aftershock frequency and distribution of rescue forces) and data processing strategies (such as weights of different data sources and focus of analytical models). By establishing an adaptive coordination mechanism centered on phase awareness, global decision optimization can be achieved. Furthermore, the study found that phase identification not only guides data credibility assessment and reconstruction but also effectively controls the on-demand scheduling and activation of different analytical modules, thereby precisely focusing limited computing and communication resources on the most pressing sub-problems of the rescue effort.

[0026] After introducing the basic concept of the present invention, the embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0027] Example:

[0028] like Figure 1 As shown, the 5G and AI multi-source fusion decision-making system for earthquake rescue includes:

[0029] The data acquisition module is used to acquire multi-source raw data with spatiotemporal identifiers with low latency through the 5G network;

[0030] The phase determination module is used to extract phase features from multi-source raw data and generate a phase identifier to represent the current rescue phase.

[0031] The credibility assessment module is used to assess the credibility of multi-source raw data based on stage identifiers and generate a weighted data set.

[0032] The data reconstruction module is used to reconstruct the weighted data set based on the stage identifier, and generate a data representation structure that adapts to the current rescue stage.

[0033] The scheduling and control module is used to receive stage identifiers and generate trigger signals based on the stage identifiers;

[0034] The first analysis module is used to respond to the trigger signal, perform life-related information analysis based on the data representation structure, and generate life existence probability data associated with geographical location.

[0035] The second analysis module is used to respond to the trigger signal, perform environmental access-related information analysis based on the data representation structure, and generate access level data associated with geographical location;

[0036] The fusion decision module is used to jointly analyze the probability of life presence data and access level data through a pre-trained multi-agent reinforcement learning model based on the stage identifier, and generate rescue decision instructions.

[0037] like Figure 2 The diagram shown is a flowchart of the process of this application. The working principle of this application is to coordinate various dedicated modules through a central processing unit (e.g., a high-performance computing unit deployed on an edge computing node or cloud server) and realize low-latency data aggregation and instruction distribution based on a 5G communication network.

[0038] Specifically, the data acquisition module consists of various sensor terminals deployed in the disaster area (such as vibration sensors, infrared thermal imagers, and optical and radar equipment carried by drones), physiological monitoring equipment worn by rescue personnel, and web crawler servers that access public social media data. These heterogeneous data sources, through a built-in 5G communication module, upload the collected raw data (such as vibration waveforms, thermal images, vital signs, and text image information) along with the timestamps and geographic coordinates (i.e., spatiotemporal identifiers) generated by their GPS / BeiDou positioning chips to the system's central processing unit in real time.

[0039] After receiving the aforementioned multi-source raw data, the stage determination module uses the central processing unit to call its pre-loaded machine learning model (e.g., a deep neural network based on time series analysis) to extract features from the input time-series data (such as continuous earthquake magnitude readings and aftershock frequency statistics) and spatial distribution data (such as heat maps of rescue unit locations). By executing the feature extraction algorithm, key patterns that can characterize the evolution of the disaster are identified, such as determining whether the current situation is in the "72-hour golden rescue period" or the "high-incidence period of secondary disasters." Finally, a structured stage identification data object is generated, which not only contains stage classifications (such as the golden rescue stage) but also metadata such as the estimated duration inferred by the model and a list of core risk factors.

[0040] Based on stage identifiers, the credibility assessment module dynamically calls corresponding assessment algorithms to assign different confidence weights to data from different sources according to the core risk factors indicated in the stage identifiers (e.g., if the stage is identified as the golden rescue stage, then "life detection" is a high-risk factor). (For example, increasing the weight of life detection sensor data and decreasing the weight of unverified information from social media). A weighted dataset is generated through weighted calculation, thereby reducing the impact of noise and interference sources before data fusion. The data reconstruction module further selects corresponding feature encoding algorithms based on the information needs implied by the stage identifiers (e.g., strengthening edge and texture feature extraction in images during stages where structural damage is a key concern). The weighted data is then subjected to dimensionality reduction, alignment, and formatting to generate a unified spatiotemporal reference, multi-feature channel data representation structure (e.g., a three-dimensional tensor whose dimensions represent geographic grids, time slices, and feature channels).

[0041] The system's scheduling and control module receives the stage identifier from the stage determination module and, based on the decision logic preset in the processor (such as a lookup table or rule engine), parses out the analysis tasks that should be prioritized for execution in the current stage. For example, during the golden rescue stage, this module generates a "trigger signal" that prioritizes life search and rescue analysis. This signal is essentially a start command or message queue event sent to a specific analysis subprocess.

[0042] The first and second analysis modules are parallel computing units deployed on processors or dedicated AI accelerator cards (such as GPUs and NPUs). They are activated in response to trigger signals issued by the scheduling control module. The first analysis module is configured to focus on life-related information. It loads a pre-trained lightweight convolutional neural network model, receives data representation structures, extracts feature subsets related to living organisms such as infrared thermal radiation and micro-vibration spectra, and outputs a heat map of "probability of life presence data" covering a geographical grid of the disaster area through model inference. Similarly, the second analysis module is configured to analyze environmental accessibility. Its model may combine semantic segmentation results from UAV imagery with digital elevation models to assess the risk of landslides and the degree of road damage in each grid cell, generating another map of "accessibility data". The operation of these two modules is selectively parallel, and their activation is indirectly determined by a stage identifier through the scheduling control module, thereby realizing on-demand allocation of computing resources.

[0043] Finally, the fusion decision-making module receives the stage identifiers and the outputs from the first and second analysis modules. The model's internal decision-making strategy dynamically adjusts based on the stage identifiers—for example, during the golden rescue stage, the model's strategy tends to reward actions that quickly approach high-probability-of-survival targets; while during the high-risk secondary disaster stage, the strategy tends to avoid high-risk areas. By simulating the coordinated actions of multiple rescue agents, and comprehensively considering factors such as target priority, access constraints, and resource consumption, the model outputs rescue decision instructions containing specific action objectives, path suggestions, and resource allocation plans. These instructions are then transmitted via the 5G network to frontline rescue terminals (such as tablets and AR glasses).

[0044] Through the above technical solutions, this application enables the rescue decision instructions output by the system to no longer be based on static or single optimization models, but to adapt to the evolution of the disaster situation and achieve a better balance among multiple dynamic objectives such as life search and rescue, risk avoidance and resource efficiency, thereby providing more practical and operational intelligent support for earthquake rescue command.

[0045] This application further proposes to extract stage features from multi-source raw data and generate stage identifiers to characterize the current rescue stage. The system executes a series of specific processing steps through a stage determination module deployed in the central processing unit. First, the module extracts at least three types of core data closely related to the macro-evolution of the disaster from the aforementioned multi-source raw data stream in real time as input. The first is earthquake magnitude time series data, which can be obtained from sources such as the real-time earthquake catalog published by the China Earthquake Networks Center or the continuous waveform inversion results of a distributed fiber optic sensor network. The second is aftershock frequency data, which is obtained by statistically analyzing the number of aftershocks above a set magnitude lower limit (e.g., magnitude 2.0) within a set time period (e.g., the past hour) and calculating the frequency using a sliding window with a fixed time window (e.g., 10 minutes). The third is data on the distribution of rescue forces, which can be obtained in real time by accessing the resource dispatch platform of the emergency management department to obtain the number of personnel and equipment of each rescue team (e.g., fire, armed police, social rescue organizations) and their GPS positioning information. Spatial statistical methods such as kernel density estimation are used to form a heat map or gridded data characterizing the spatial density of rescue forces. These three types of data need to be preprocessed before input, including spatiotemporal alignment, normalization (such as linearly mapping the magnitude to the [0,1] interval, and dividing the rescue force density value by the theoretical maximum value of the force that can be deployed in the region), and filling in short-term missing values ​​(such as using linear interpolation).

[0046] The preprocessed temporal and spatial distribution data are combined into a structured multidimensional feature vector or tensor, which is then input into a pre-trained disaster evolution prediction neural network. This disaster evolution prediction neural network is a temporal-spatial hybrid model specifically designed for this task; for example, it can employ an architecture combining a long short-term memory network and a convolutional neural network. The training process is as follows: a large amount of historical earthquake case data is used as the training set. Each case contains the three types of feature data mentioned above, sliced ​​at fixed time intervals (e.g., 1 hour) from the moment of the mainshock, as well as "real labels" of the rescue stage corresponding to that time slice, annotated by disaster relief experts afterward. The stage division criteria can be: 0-72 hours after the mainshock is the golden rescue stage; 72 hours later until the risk of major secondary disasters (such as large landslide dams or severe epidemics) is the high-incidence period of secondary disasters; and afterwards is the "post-disaster recovery stage." The training objective is to minimize the cross-entropy loss between the stage classification results predicted by the model and the expert annotations. During training, the Adam optimizer can be used, with an initial learning rate set to 0.001, a batch size of 32, and an early termination strategy used to prevent overfitting. Through this training, the model can learn to effectively identify the current macro-stage of the disaster situation from magnitude attenuation patterns, aftershock activity patterns, and rescue response dynamics.

[0047] The disaster evolution prediction neural network outputs structured stage identifiers. Specifically, the output contains a data structure with the following core fields:

[0048] 1. Stage classification: such as the category code for the golden rescue stage, the period of high incidence of secondary disasters, or the "post-disaster recovery stage".

[0049] 2. Estimated Duration of the Current Phase: This is a time value (in hours) predicted by the model based on regression analysis of the current input features. For example, the model predicts the remaining effective time of the current golden rescue phase based on the fit of the "Omori Formula" for aftershock attenuation and the saturation trend of rescue forces. Its prediction formula can be expressed as:

[0050]

[0051] in, For the input feature vector, and The weights and bias parameters obtained during model training The activation function is used. The model outputs the remaining effective time. This value is obtained based on the statistical patterns learned from historical data.

[0052] 3. Key Risk Factor Set: A list of risk factors identified by the model that pose the main threat at the current stage. The model outputs a vector through its attention mechanism or feature importance analysis layer, where each element corresponds to a current significance score (e.g., a value between 0 and 1) for a predefined risk factor (such as "strong aftershocks," "landslides / debris flows," "building collapses," "communication disruptions," or "disease risk"). The system can set a threshold (e.g., 0.5) to determine whether a risk factor should be included in the "key" set. For example, during periods of high secondary disaster incidence, the model might output highly significant "landslides / debris flows" and "disease risk" factors.

[0053] Through the specific implementation methods described above, this application achieves the transformation from raw, low-level observational data to high-level, phased understanding with clear rescue guidance significance. This not only provides the system with a global judgment of "what stage it is currently in," but more importantly, it outputs key meta-information in a structured manner, such as "how long this stage will last approximately" and "what should be the main focus of prevention at this stage." This overcomes the problem of decision-making lag or target defocus caused by the lack of macro-level, quantitative perception of the rescue process in traditional systems, enabling the entire intelligent decision-making system to keep pace with the dynamic evolution of the disaster.

[0054] like Figure 3The diagram shows the flowchart of the credibility assessment steps in this application. This application further proposes to assess the credibility of multi-source raw data based on stage identifiers. Specifically, the received stage identifier is first parsed, with the input being the set of key risk factors contained within that stage identifier. For example, if the stage identifier is the golden rescue stage, its set of key risk factors might be {urgency of life detection, need for rapid response}; if it is a period of high incidence of secondary disasters, the set might be {geological disaster risk, meteorological disaster risk}.

[0055] The module contains a pre-defined data source-risk factor correlation matrix. This matrix defines the support or correlation strength of various data sources (e.g., Category A: life detection sensors; Category B: social media text; Category C: geological sensors; Category D: meteorological sensors, etc.) for each preset risk factor. This correlation matrix can be pre-set based on historical data analysis, expert knowledge, or the physical characteristics of the data itself. For example, life detection sensor data is highly correlated with the "urgency of life detection" factor, and is assigned a high correlation score (e.g., 0.9); social media text data, due to the difficulty in verifying its authenticity and the time lag during the golden rescue stage, is assigned a lower correlation score with "urgency of life detection" (e.g., 0.3); while geological sensor data is strongly correlated with the "geological disaster risk" factor.

[0056] Subsequently, the module dynamically adjusts the weighting strategy for different data sources based on the set of key risk factors identified in the current stage. Using fuzzy hierarchical analysis, the "stage-specific requirements" are quantified into specific weight coefficients for each data source. The calculation process is as follows:

[0057] First, the set of key risk factors identified in the current stage is used as the criterion layer of the analytic hierarchy process (AHP), and the various data source types are used as the solution layer. Then, domain experts (or based on historical data statistics) are invited to compare the importance of each risk factor in the criterion layer to the overall goal of the current stage (obtaining the most reliable information), constructing a fuzzy judgment matrix. For example, in the golden rescue stage, the factor "urgency of life detection" might be evaluated as "significantly important" relative to "rapid response needs." These linguistic evaluations are transformed into triangular fuzzy numbers.

[0058] Next, the fuzzy weight of each risk factor is calculated. Then, using these fuzzy weights, combined with the aforementioned data source-risk factor correlation matrix (as the contribution of the scheme layer to each criterion layer), the comprehensive fuzzy evaluation value of each data source relative to all risk factors is calculated.

[0059] Finally, by using defuzzification methods (such as the centroid method), the fuzzy comprehensive evaluation value of each data source is transformed into a specific weight coefficient. And satisfy .

[0060] The formula for calculating the composite weight of the scheme layer to the target layer can be expressed as:

[0061]

[0062] in, This is the data source weight vector to be determined. It is the weight vector of the risk factors (obtained by fuzzy analytic hierarchy process). This is the data source-risk factor correlation matrix (after normalization). The calculation result is a set of dynamic weight coefficients. According to this method, during the golden rescue phase, life detection sensor data (Category A) will receive higher weight coefficients due to their strong correlation with key factors. (For example, The weighting coefficient for social media text data (Category B) is likely in the range of 0.4-0.6. Then it is lower (for example, (Potentially in the range of 0.05-0.15). Similarly, during periods of high incidence of secondary disasters, the weight of geological and meteorological sensor data (C and D categories) will be increased.

[0063] After obtaining the weighting coefficients for each data source type, the module applies these coefficients to each specific data instance within that type. For example, for each data point from an infrared thermal imager (belonging to type A), its original value is multiplied by the weighting coefficient. For each text report from social media (belonging to Category B), its sentiment or keyword intensity value after natural language processing will be multiplied by a weighting coefficient. By performing this weighting operation on all data instances, the module ultimately generates a weighted data set. In this set, data highly relevant to the core task of the current stage is enhanced, while data with weak relevance or strong interference is relatively suppressed.

[0064] Through the dynamic weight evaluation mechanism based on fuzzy hierarchical analysis, this application achieves precise alignment between data value assessment and core needs in the rescue phase. This effectively reduces the interference of non-critical or unreliable information on the analysis model, improves the overall system's perception quality of dynamic disaster situations and the reliability of subsequent decisions from the data source, and provides high-quality data input to ensure the generation of rescue decisions that are closer to the actual situation at each stage.

[0065] This application further proposes to reconstruct the weighted dataset based on stage identifiers. Specifically, the module first selects a corresponding feature extraction backbone network based on the stage identifier category (e.g., "golden rescue" or "high incidence of secondary disasters"). These backbone networks are pre-trained deep learning models tailored to the data characteristics and task objectives of different stages. For example, for the golden rescue stage, the core task is to quickly detect signs of life; therefore, the selected backbone network (e.g., a lightweight 3D convolutional neural network) will focus more on extracting and encoding features related to living organisms, such as non-natural heat source features in infrared images, human voice or knocking sound features in the sound spectrum, and non-seismic frequency feature patterns in micro-vibration signals. The network's training data comes from historical earthquake rescue cases, containing a large number of image, audio, and vibration data slices labeled with the presence or absence of life. During training, the network learns to extract highly discriminative feature vectors from the original weighted data (e.g., treating infrared heatmaps, sound spectrum maps, and vibration waveforms as different channels). The backbone network can be trained using mean squared error or cross-entropy loss function, with the Adam optimizer set to a learning rate of 0.0001 and a batch size of 16. Network parameters are adjusted through backpropagation.

[0066] Then, the feature extraction backbone network encodes features from the input weighted dataset. Specifically, weighted data from different sensors, after different preprocessing steps (such as heatmap intensity matrices, acoustic spectrogram matrices, and vibration spectrograms), are fed into the feature extraction backbone network. The final layer of the feature extraction backbone network outputs a low-dimensional, high-level feature encoding vector, which integrates core information from multiple input sources relevant to the current stage of the task. For example, for a backbone network operating during periods of high secondary disaster incidence, its encoding vector will more strongly include fused features related to landslide and debris flow risks, such as terrain slope, lithology, and rainfall intensity.

[0067] By utilizing a Generative Adversarial Network (GAN) associated with stage markers, data augmentation and missing data generation are performed on encoded features. The GAN is pre-trained for specific rescue stages, and its generator is trained to perform data augmentation and feature completion within the feature space. For example, when communication disruptions lead to data loss in a certain area, the generator can reasonably infer the missing feature vector based on the complete feature codes of the surrounding areas and the historical geographical information of that area, thus achieving data generation. Simultaneously, the GAN can generate more diverse feature code variants to enhance the robustness of subsequent analysis models. The training of this GAN is based on a large dataset of complete historical disaster feature codes. During training, the discriminator learns to distinguish between genuine historical feature codes and forged codes generated by the generator, while the generator strives to generate realistic feature codes that can fool the discriminator. The adversarial training process involves alternately optimizing the generator and discriminator, using a standard adversarial loss function. Ultimately, the generator possesses the ability to perform high-quality completion and augmentation of incomplete or sparse features.

[0068] Finally, the module regularizes all feature codes processed by the GAN network (including features directly encoded from the original data and features that have been augmented / enhanced) according to a preset unified spatiotemporal scale to form the final data representation structure. Specifically, the entire disaster area is divided into a geographic grid with a fixed spatial resolution (e.g., 100 meters × 100 meters), and the time axis is divided into fixed step sizes (e.g., every 10 minutes is a time slice). For each spatiotemporal unit (grid-time slice), all multi-source feature encoding vectors falling within that unit are concatenated or pooled (e.g., average pooling) to form a feature vector of a unified dimension. Arranging the feature vectors of all spatiotemporal units in spatial and temporal order constitutes a three-dimensional multi-channel spatiotemporal feature tensor. Its three dimensions represent the number of rows, columns, and time steps of the spatial grid, respectively, while each element in the tensor is itself a feature vector containing multiple channels.

[0069] Through the aforementioned adaptive data reconstruction process, this application transforms the raw, messy, and incomplete multi-source data acquired from the front end into a highly structured, spatiotemporally aligned, feature-unified, and optimized standard data representation for the current stage of the task. Since the reconstruction process itself is based on the needs of the current stage, it ensures that the information input to the analysis module has undergone a phased focus and enhancement, enabling the analysis model to more accurately capture patterns most relevant to the current core rescue task. This lays a high-quality, highly available data foundation for the final fusion decision, improving the data processing efficiency and reliability of the analysis results throughout the intelligent decision-making process.

[0070] This application further proposes that the scheduling control module, in its specific implementation, is manifested as a logical control unit executed by the system's central processing unit, which can be called a central scheduler. When the central scheduler is working, it first receives a stage identifier, parses the key information in the stage identifier, mainly the stage classification (such as the golden rescue stage) and any additional state descriptions that may be included (such as composite labels like "limited rescue resources" and "weak signs of life"). Subsequently, according to a preset rule mapping table, it generates corresponding analysis module trigger signals. The preset rule mapping table is based on deterministic logical rules set according to historical rescue experience and expert knowledge, which clearly stipulates the analysis strategies to be activated under different stage identifiers. Based on the analysis module trigger signals, the system can decide whether to trigger the first analysis module focused on life information analysis alone, the second analysis module focused on environmental accessibility analysis alone, or both in parallel, to simultaneously obtain a full-dimensional situational map of life and the environment.

[0071] Specifically, when the stage indicator suggests that rescue resources are limited and signs of life are weak, the central dispatcher generates a trigger signal that only activates the first analysis module, in order to concentrate all available analytical computing power to quickly locate signs of life. It then performs spatial clustering and path optimization preprocessing on the preliminary, gridded probability data of life presence, and the resulting high-probability life sign cluster information is input as an enhanced feature to the fusion decision module.

[0072] The central dispatcher uses a built-in clustering algorithm (such as the density-based DBSCAN algorithm) to process the probability data of life presence. This algorithm identifies geographic grids with probability values ​​above a certain threshold (e.g., 0.7) as "high-probability points" and aggregates these spatially adjacent points into several high-probability life sign clusters. Each cluster can be described by its centroid coordinates and the number of grids it contains (representing area size or intensity). Next, based on currently known rescue force locations and initial road accessibility information (possibly from pre-disaster maps or initial reconnaissance), the dispatcher runs a fast path optimization algorithm (such as Dijkstra's algorithm or A algorithm) between these clusters and between the rescue force's starting point and each cluster, estimating the approximate cost (such as time or distance) for rescue forces to reach each cluster. Finally, the resulting "cluster" information (including location and intensity) and its corresponding "initial rescue cost" are encapsulated into an enhanced feature vector.

[0073] Through the aforementioned rule-driven and intelligent preprocessing scheduling mechanism, this application achieves dynamic optimization of system analysis resources and preliminary intelligent fusion of information flow. This transforms low-level perception data into high-level tactical information that directly supports decision-making, reducing the burden on the fusion decision-making module for real-time, complex situational understanding, and improving the overall system's efficiency in transforming perception into decision-making, as well as the timeliness and executability of the final decision instructions.

[0074] This application further proposes that the specific steps for performing life-related information analysis based on data representation structures include:

[0075] The received data representation structure extracts a subset of features directly related to vital signs. From the multi-channel tensor, specific channels or combinations of channels representing infrared thermal imaging features, micro-vibration sensing features, and sound spectrum features are precisely separated. These features are not the original pixels or waveforms, but high-order abstract features encoded by the backbone network during the reconstruction stage. For example, the infrared thermal imaging feature channel may contain encoded information such as the intensity of non-natural heat sources that match the body temperature range and the stability of heat source contours; the micro-vibration sensing feature channel may contain the intensity and pattern encoding of abnormal vibration signals within a specific frequency band (such as 1-10Hz) after filtering out the main shock and aftershock frequencies; and the sound spectrum feature channel may contain the recognition probability or spectral envelope features of specific sound patterns such as human cries for help and knocking sounds.

[0076] The aforementioned features were input into multiple lightweight convolutional neural networks (CNNs) operating in parallel for preliminary analysis. These CNNs each have different architectural focuses; for example, one network focuses on processing two-dimensional spatial patterns of infrared features, another focuses on analyzing the temporal sequence patterns of sound spectra, and a third focuses on fusing cross-modal correlations of vibration and thermal features. These networks were pre-trained, with training data derived from a large amount of historical disaster relief site data and simulation data, including various sensor data slices labeled "living" and "non-living." During training, each network independently used a binary cross-entropy loss function, employed a stochastic gradient descent optimizer, set the learning rate to 0.01, and the momentum to 0.9. After multiple iterations, the networks learned to judge the probability of life existence from their respective preferred feature dimensions. After processing the input features, each network outputs a preliminary life existence confidence score (a value between 0 and 1) for the current analysis unit (e.g., a 100m × 100m grid).

[0077] This paper utilizes a particle swarm optimization (PSO) algorithm to dynamically weight and fuse the output confidence scores of multiple networks, and optimizes the parameters of the probability distribution model to generate gridded life existence probability data. The goal of this step is to overcome the limitations of a single model and obtain a more robust and accurate estimate of life existence probability. Specifically, each spatiotemporal grid to be analyzed is treated as an independent optimization problem. A particle swarm is initialized, where each particle represents a possible fusion weight vector (the length of which is equal to the number of CNN networks) and a set of probability distribution model parameters. For example, this probability distribution model could be a logistic regression model, whose input is a weighted sum of the confidence scores of multiple CNN networks, and the particle position contains the weights of each CNN and the coefficients of the logistic regression model.

[0078] The optimization process uses the log-likelihood function (or minimizing the prediction error) that maximizes the probability of life existence as the fitness function. In each iteration, particles update their velocity and position based on their individual best historical position and the group's best historical position. After a preset number of iterations (e.g., 50-100), the algorithm converges. At this point, the fusion weight vector represented by the globally optimal particle is the optimal CNN confidence combination for the current grid, and its corresponding probability distribution model parameters are also optimized. Applying this optimal weight to the original confidence output of each CNN in the grid and transforming it through the optimized probability distribution model yields the final, fusion-optimized probability value of life existence for the grid (x,y). Its calculation formula can be summarized as follows:

[0079]

[0080] in, It is the sigmoid function, and N is the number of CNN networks. and These are the optimal weights and bias parameters obtained from particle swarm optimization. It represents the initial confidence level of the nth CNN network for that grid.

[0081] Finally, by repeating the above process of feature extraction, parallel CNN analysis, and particle swarm optimization fusion on all geographic grids in the disaster area, a gridded probability map of life presence covering the entire area of ​​interest can be generated. In this map, each grid corresponds to a precisely calculated probability value of life presence, intuitively reflecting the spatial distribution probability of signs of life.

[0082] Through the aforementioned multi-model parallel analysis and intelligent meta-fusion technology, this application achieves highly robust and accurate detection and probabilistic representation of concealed life signs. This enables the generated life presence probability data to more accurately reflect the complex, variable, and noisy real disaster scene conditions, greatly reducing the risk of false alarms and missed alarms. This highly reliable probabilistic grid map provides solid, reliable, and quantifiable perceptual input for subsequent rescue route planning, precise resource allocation, and agent target setting in the fusion decision module.

[0083] This application further proposes that when the system responds to the trigger signal of the scheduling and control module, it initiates a second analysis module to perform environmental accessibility-related analysis based on the data representation structure. Its core task is to extract key environmental elements affecting the mobility of rescue forces from this tensor. The analysis begins by extracting a subset of features related to accessibility, receiving the data representation structure, and identifying and separating specific feature streams from the multi-channel feature tensor. This includes extracting UAV image features from encoded UAV aerial images, such as semantic segmentation maps or edge feature maps representing road damage (e.g., cracks, collapses), building ruins distribution, vegetation cover, and water bodies (e.g., landslide dams); terrain digital elevation model features extracted from channels fused with geographic information system data, such as slope, aspect, surface roughness, and elevation changes; and real-time weather data extracted from real-time data streams, such as rainfall intensity, wind speed, and visibility. All these features have undergone spatiotemporal alignment and normalization during the front-end reconstruction stage.

[0084] A cost map is constructed with access difficulty as the optimization objective. The system calculates a basic access cost for each geographic grid cell (i, j) (e.g., 100m × 100m). The cost can be calculated based on a pre-defined cost function model, which quantifies and integrates the impact of various features. For example, a simplified linear cost function can be expressed as:

[0085]

[0086] in, It is a normalized slope value (for example, 1 when the slope is greater than 30 degrees). It is a quantitative value of the degree of road damage (e.g., 1 for completely damaged and 0 for intact). It is the influence coefficient of rainfall intensity. These are preset weighting coefficients, which can be set based on expert experience (for example, set to 0.4, 0.4, 0.2 respectively).

[0087] However, static cost maps cannot fully reflect the actual passage risks under dynamic exploration and collaborative action. Therefore, the module adopts a wolf pack algorithm for dynamic risk assessment. This algorithm simulates the collaborative exploration behavior of multiple rescue units (as "wolf packs") on the cost map. Through the division of labor mechanism of the wolf pack, it dynamically assesses and marks the passage risk level of different geographical areas, generating passage level data.

[0088] During the simulation iteration, the system records information such as the frequency of visits to each geographic grid by different roles and numbers of "wolves," and the changes in the average passage cost during visits. Based on this dynamic exploration information, the module dynamically adjusts the static basic passage cost, generating a more comprehensive passage risk assessment value that better reflects the actual difficulties under collaborative action.

[0089] Finally, based on the final assessed access risk value, the system labels each geographic grid with a different access level. For example, a threshold range can be set: risk values ​​below a certain threshold. (e.g., 0.2) represents "unimpeded (Level 1)", which is between and (e.g., 0.5) indicates "slow (level 2)", while values ​​between 0.5 and 0.5 indicate "low speed". and (e.g., 0.8) indicates "high risk (level 3)", and values ​​higher than 0.8 indicate "high risk". The level is designated as "No Entry (Level 4)". This generates a gridded access level data map covering the entire disaster area.

[0090] The second analysis module combines static environmental feature modeling with biomimetic swarm intelligence dynamic exploration simulation to achieve a more realistic and intelligent assessment of accessibility in disaster areas. This allows the generated accessibility data to go beyond simple geographic information overlay, dynamically reflecting the actual accessibility risks under limited resources and collaborative strategies. This provides subsequent rescue route planning, force deployment, and integrated decision-making modules with a more accurate and insightful input of the environmental situation than static maps, enabling the generation of safer and more efficient rescue action plans.

[0091] This application further proposes that the fusion decision module receives multiple inputs from the preceding modules: stage identifiers representing the global situation, gridded life existence probability data output by the first analysis module, and gridded access level data output by the second analysis module. The pre-trained multi-agent reinforcement learning model in the fusion decision module is based on a reward function. Training and decision optimization, reward function Guide intelligent agents (each representing a virtual rescue unit) to make collaborative decisions under dynamic disaster situations. The formula is as follows:

[0092]

[0093] in: The weighting coefficients are dynamically adjusted based on the stage identifier;

[0094] Golden rescue phase: The system will grant Higher values ​​(e.g., 0.7). Medium value (e.g., 0.2) Lower values ​​(such as 0.1) emphasize the race against time in life-saving search and rescue efforts;

[0095] During periods of high incidence of secondary disasters: [This will] increase The value (e.g., increased to 0.5) should be adjusted accordingly. (e.g., reduced to 0.3), and may increase (e.g., increased to 0.2) to emphasize safety avoidance and maintenance of cell state;

[0096] For life-saving benefits, This represents the set of all currently identified targets awaiting rescue. This represents the probability that target i has a life at time t; this term encourages the agent to go to and cover areas with a high life probability.

[0097] This is a penalty item for passing through. This represents the set of action paths for rescue units. This represents the travel cost of path j at time t; this cost is directly mapped from the traffic level data generated by the second analysis module. This item penalizes the selection of high-risk, high-cost paths and encourages the selection of safe and efficient routes.

[0098] For unit safety constraints, Represents the set of rescue unit states. As an indicator function, when the estimated energy consumption of unit k... Below the safety threshold The value is 1 when the condition is met, and 0 otherwise; this is the preset security threshold. It can be set to 80% of the unit's initial energy or working time; this reward is given to agents that maintain their state within a safe range to ensure the sustainability of rescue operations.

[0099] The training of the multi-agent reinforcement learning model is conducted offline in a simulated environment. The simulator, built upon a large amount of historical earthquake data, simulates the spatiotemporal evolution of signs of life, dynamic changes in environmental accessibility (such as aftershocks triggering new landslides), and the actions and expenditures of rescue units at different stages. During training, multiple agents (e.g., 5-10) share a centralized evaluation network (Critic) but have their own policy networks (Actors). Training algorithms can employ multi-agent reinforcement learning algorithms such as Proximal Policy Optimization (PPO) or Deep Deterministic Policy Gradient (DDPG). Specific training conditions include: a discount factor γ of 0.99, a learning rate of 0.0003, the Adam optimizer, a batch size of 512, and a total training step count of up to several million steps until the policy converges to a stable performance in the simulated test.

[0100] After training, the trained model is deployed to the fusion decision module. During real-time decision-making, the model dynamically selects the corresponding strategy weights (or activates the corresponding fine-tuned sub-model) based on the current input stage identifier, and generates action instructions for each agent based on real-time survival probability data, access level data, and the real-time status of the rescue unit, such as "go to target A, proceed along path B, and execute search and rescue operation C". These instructions are integrated to form the system's rescue decision instructions.

[0101] Through the aforementioned multi-agent reinforcement learning decision-making mechanism, this application achieves globally optimal or near-optimal collaborative rescue planning under complex, dynamic, and multi-objective constraints. This enables the generated rescue decision instructions to intelligently and dynamically weigh multiple objectives such as "saving more lives," "ensuring operational safety," and "maintaining rescue forces." The output solution is no longer a patchwork of single-point optimization results, but rather reflects the collaborative behavior of a swarm of intelligent agents that coordinates multiple units and optimizes global resource allocation. This significantly improves the efficiency, safety, and success rate of overall rescue operations in actual command.

[0102] This application further proposes that after generating rescue decision instructions, the system also constructs a closed-loop feedback and online optimization mechanism based on 5G communication. When the fusion decision module generates rescue decision instructions, it distributes these instructions to rescue terminals via the 5G network. These rescue terminals can be smart handheld devices carried by frontline rescue personnel, command tablets mounted on vehicles, or drone control stations. The instructions typically include the action objective, suggested route, task type, and coordination information.

[0103] While executing commands, the rescue terminal continuously receives command execution status data (e.g., actual arrival location, mission progress percentage, remaining battery power) and real-time environmental change data (e.g., images of newly discovered obstacles ahead, reports of sudden local weather changes, and newly discovered signs of secondary disasters) from its built-in sensors (such as GPS, motion sensors, and cameras) and operator input. This data is transmitted back to the system's central processor via the 5G network.

[0104] The system uses instruction execution status data, real-time environmental change data, and updated stage identifiers as inputs to drive online model optimization. An online policy gradient algorithm is employed to dynamically adjust the policy network parameters of the multi-agent reinforcement learning model. Based on the parameter-adjusted multi-agent reinforcement learning model, incremental decision instructions are generated and distributed.

[0105] Specifically, the system maintains a lightweight online learner. This learner constructs one or more "experience tuples" (state, action, reward, new state) from the aforementioned real-time feedback data. The "reward" is determined according to a reward function. Obtained, but weighting coefficients This may have been adjusted with the updated stage identifier. Subsequently, the online learner applies a policy gradient algorithm, such as an online variant of the Asynchronous Advantage Actor-Critic (A3C) algorithm, to compute the policy gradient. Its core update formula can be expressed as:

[0106]

[0107] in, It is a policy network The parameters, It represents the current state (including environmental data, execution status, and stage identifier). These are actions taken by the intelligent agent (rescue unit). It is the advantage function, which estimates the merits of taking that action relative to the average level in the current state. The online learner uses the real-time data fed back to calculate the policy gradient. The policy network parameters are updated with a very small learning rate (e.g., on the order of 0.00001 to 0.0001). This small-step online adjustment aims to enable the model to quickly adapt to local, sudden changes in situation, such as a predetermined path suddenly becoming impassable or a strong new sign of life being discovered in a certain area, without disrupting the global cooperative policy already possessed by the pre-trained model.

[0108] Finally, based on the parameter-adjusted multi-agent reinforcement learning model, the system reassesses and plans the current overall situation, generating incremental decision instructions. These incremental instructions do not overturn the original plan entirely, but rather make partial modifications, supplements, or insert emergency tasks into the original instructions.

[0109] Through the aforementioned online strategy optimization mechanism based on real-time feedback, this application achieves a leap from static or batch decision-making to dynamic adaptive decision-making. This not only enhances the operability and robustness of decision-making instructions at dynamic disaster sites, reducing action errors or delays caused by information lag or inadequate planning, but also, by directly integrating frontline feedback into model optimization, enables the entire system's intelligence level to continuously and online evolve as rescue operations progress. This better addresses the high uncertainty and complexity of disaster evolution, achieving precise and flexible guidance for rescue operations.

[0110] This application further proposes that when acquiring multi-source raw data with spatiotemporal identifiers with low latency through 5G networks, 5G network slicing technology is used. 5G network slicing is a logically independent end-to-end network that provides guaranteed bandwidth (e.g., uplink peak rate not less than 100Mbps), extremely low end-to-end latency (e.g., air interface latency less than 10 milliseconds), and high-priority transmission channels for emergency data streams, ensuring that the transmission of critical monitoring data is not affected by public network congestion.

[0111] Multi-source raw data includes at least the following categories:

[0112] 1. Seismic wave data from a distributed fiber optic sensor network: This is a fiber optic sensor network deployed in critical infrastructure such as urban underground utility tunnels, bridges, and dams. When seismic waves pass through, they cause changes in parameters such as the phase and intensity of the optical signal in the fiber, which in turn reveal physical quantities such as seismic acceleration and strain. This data is uploaded in real time at a high frequency (e.g., 1000 sampling points per second) through nearby 5G gateway devices. Each data point accurately carries the sensor's geographical coordinates and a nanosecond-level timestamp.

[0113] 2. Multimodal physiological data from rescue personnel's wearable devices: Each frontline rescuer is equipped with a smart wearable device integrating multiple biosensors, such as a smart helmet, vest, or wristband. These devices continuously collect physiological parameters such as heart rate, blood oxygen saturation, body temperature, and posture / movement status, as well as acceleration and angular velocity data recorded by the built-in IMU (Inertial Measurement Unit). All physiological and motion data are transmitted back in real time through the 5G individual soldier terminal carried by the rescuer, and automatically bound to the terminal's location information (from GPS / BeiDou) and precise timestamp.

[0114] 3. Disaster Scene Text and Image Data from Social Media: The system accesses public social media platforms in real time via web crawlers and open API interfaces. For specific geographical areas of the disaster zone (e.g., 50 kilometers around the epicenter), the crawler retrieves user-generated content related to the disaster scene, including geolocation markers or text-based location data. This includes text describing the disaster (e.g., "Road X collapsed"), on-site photos, or short videos. The system records the posting time (if applicable) and the geographical coordinates at the time of posting when this data is acquired.

[0115] After the data is aggregated to the central processor via the 5G network, the data acquisition module immediately performs preprocessing operations including spatiotemporal alignment and redundancy removal. Spatiotemporal alignment refers to unifying all data from different sources to a common time and geographic reference system based on their timestamps and spatial coordinates. For example, all data timestamps are unified to Coordinated Universal Time (UTC), and all spatial coordinates are converted to the same geographic coordinate system (such as WGS-84). This process ensures that data from fiber optic sensors, rescue worker locations, and social media photo locations can be accurately overlaid and analyzed on the same map.

[0116] Redundancy removal is the process of deduplication and noise reduction for unstructured data streams such as those from social media. It utilizes natural language processing and image similarity comparison techniques to identify and merge duplicate or highly similar content from different users that describes the same event (e.g., "XX building collapsed"). For content that is clearly inaccurate, contains irrelevant advertising, or is emotionally charged and lacks substantive information, the system can perform preliminary filtering based on preset keyword blacklists, image feature models, or confidence thresholds (e.g., posts with extremely negative sentiment and no specific location or disaster description can be filtered). This step aims to reduce the interference of invalid data on subsequent analysis.

[0117] Through the aforementioned data acquisition mechanism, this application constructs a disaster site perception layer that is highly timely, reliable, and rich in information dimensions. It achieves comprehensive, multi-scale data fusion and collection, ranging from macroscopic seismic wave monitoring to microscopic individual physiological states and spontaneous public reports. In particular, 5G network slicing technology ensures the priority and stable transmission of critical sensor data under adverse communication conditions, gaining a valuable time window for subsequent real-time analysis and decision-making. Furthermore, the pre-emptive spatiotemporal alignment and redundancy removal improve the regularity and quality of the data from the source, laying a solid and reliable data foundation for the efficient and accurate operation of the entire intelligent decision-making system.

[0118] The following is a specific implementation of a 5G and AI multi-source fusion decision-making system for earthquake relief:

[0119] A 6.5-magnitude earthquake struck a region, with its epicenter located in a mountainous township. Roads were severely damaged, and aftershocks were frequent. The rescue command center activated this system, integrating multi-source data through a 5G network to dynamically generate rescue decision-making instructions. The data acquisition module, using 5G network slicing technology (uplink peak rate 100Mbps, end-to-end latency <10ms), collected three types of core data in real time: seismic wave data recorded by a distributed fiber optic sensor network (sampling rate 1000Hz) (aftershock frequency 0.8 times / minute within 2 hours after the main shock); multimodal physiological data from rescue personnel's wearable devices (smart helmet + IMU sensor) (heart rate 75-110 beats / minute, blood oxygen saturation 95%-98%); and disaster site text and image data from social media (230 pieces of valid information were captured within 1 hour, including location information such as "houses collapsed in XX village" and "bridge collapsed in XX"). After spatiotemporal alignment (unified to WGS-84 coordinate system and UTC time) and redundancy removal (filtering out 35 duplicate messages), a raw dataset with spatiotemporal identifiers was generated.

[0120] The stage determination module extracts features from the raw data and inputs them into a pre-trained disaster evolution prediction neural network (LSTM+CNN hybrid architecture). The model is based on earthquake magnitude time series data (main shock 6.5, maximum aftershock 4.2), aftershock frequency data (12 aftershocks within 1 hour), and rescue force deployment distribution data (the first batch of 300 people are distributed within 5 kilometers of the epicenter). The output stage identifier is: Golden Rescue Stage (estimated duration 48 hours, key risk factors {urgency of life detection, need for rapid response}).

[0121] The credibility assessment module dynamically adjusts the data source weights based on the stage identifier. Using fuzzy hierarchical analysis, the weight of life detection sensor data is increased to 0.6 (while the weight of social media text data is reduced to 0.2), generating a weighted dataset. The data reconstruction module selects a lightweight 3D convolutional neural network (optimized for life detection) to encode the features of the weighted data, and then uses a generative adversarial network (GAN) for data augmentation (completing 15 minutes of missing infrared thermal imaging data for a certain area due to communication interruption), forming a multi-channel spatiotemporal feature tensor with a unified spatiotemporal scale (10-minute time step, 100m×100m spatial resolution).

[0122] Upon receiving the phase marker, the scheduling and control module triggers the parallel operation of the first analysis module (life information analysis) and the second analysis module (environmental accessibility analysis). The first analysis module extracts infrared thermal imaging features (temperature difference from non-natural heat sources >5℃), micro-vibration sensing features (abnormal signals in the 1-10Hz frequency band), and sound spectrum features (human voice / knocking sound patterns), inputting them into three lightweight CNN networks (parallel inference time <2 seconds). The network output confidence scores are weighted and fused using a particle swarm optimization algorithm (50 iterations) to generate gridded life probability data: a village area northeast of the epicenter has a probability value of 0.85, and the mountainous area southwest has a probability value of 0.6-0.7. The second analysis module extracts features from UAV imagery (road damage rate 60%), terrain digital elevation model features (areas with slope >30° account for 25%), and real-time weather data (light rain, wind speed 3m / s) to construct a cost map. The wolf pack algorithm (simulating 10 rescue units) was used to dynamically assess the risk of passage and generate passage level data: the path from the epicenter to the northeastern village was "slow (level 2)" and the southwestern mountainous area was "high risk (level 3)".

[0123] The fusion decision-making module, based on the stage identifier (golden rescue stage), calls a pre-trained multi-agent reinforcement learning model (5 agents representing rescue teams). The reward function is set as follows: α=0.7 (life-saving benefit), β=0.2 (passage cost penalty), γ=0.1 (unit safety constraint). The model is input with life existence probability data (coordinates of high-probability areas), passage level data (path cost), and rescue unit status (remaining power 80%), and generates rescue decision instructions: "Team 1 proceeds along the slow path to the northeast village (life probability 0.85 area), prioritizing search and rescue; Team 2 stands by, ready to support the southwest mountainous area (life probability 0.7 area)."

[0124] The instructions were distributed to the rescue terminal (smart handheld device) via the 5G network. Frontline feedback showed that Team 1 arrived at the target area 15 minutes faster than traditional route planning and successfully rescued 3 trapped people. While Team 2 was on standby, the system dynamically adjusted the model parameters through an online policy gradient algorithm (learning rate 0.00001). Due to a sudden aftershock in the southwest mountainous area (additional road damage), an incremental instruction was generated: "Team 2 adjusts its route and detours to the alternative route in the southwest mountainous area (accessibility level reduced to 2.5)."

[0125] This embodiment utilizes a phase-aware adaptive mechanism. During the golden rescue phase, the system prioritizes life detection and dynamically increases the weight of key data sources, significantly improving the accuracy of life probability identification. Through a multi-agent reinforcement learning model, it achieves global collaborative scheduling of rescue forces, avoiding time delays caused by improper path selection and reducing the risk of secondary injury to rescue personnel due to environmental risks. This provides precise and efficient intelligent decision support for earthquake rescue.

[0126] The technical scope of this invention is not limited to the content described above. Those skilled in the art can make various modifications and variations to the above embodiments without departing from the technical concept of this invention, and all such modifications and variations should fall within the protection scope of this invention.

Claims

1. A 5G and AI multi-source fusion decision-making system for earthquake rescue, characterized in that, include: The data acquisition module is used to acquire multi-source raw data with spatiotemporal identifiers with low latency through the 5G network; The stage determination module is used to extract stage features from the multi-source raw data and generate a stage identifier to represent the current rescue stage. The credibility assessment module is used to assess the credibility of the multi-source raw data based on the stage identifier and generate a weighted data set. The data reconstruction module is used to reconstruct the weighted data set based on the stage identifier to generate a data representation structure that adapts to the current rescue stage. The scheduling control module is used to receive the stage identifier and generate a trigger signal based on the stage identifier; The first analysis module is used to respond to the trigger signal, perform life-related information analysis based on the data representation structure, and generate life existence probability data associated with geographical location; The second analysis module is used to respond to the trigger signal, perform environmental access-related information analysis based on the data representation structure, and generate access level data associated with geographical location; The fusion decision module is used to jointly analyze the probability data of life presence and the access level data through a pre-trained multi-agent reinforcement learning model based on the stage identifier, and generate rescue decision instructions.

2. The 5G and AI multi-source fusion decision-making system for earthquake rescue according to claim 1, characterized in that, The specific steps for extracting stage features from the multi-source raw data and generating stage identifiers to characterize the current rescue stage include: Obtain at least three types of core data from multi-source raw data, including earthquake magnitude time series data, aftershock frequency data, and data on the distribution of rescue force deployment; The data is input into a pre-trained disaster evolution prediction neural network, which outputs the stage identifier. The stage identifier includes at least the expected duration of the current stage and a set of key risk factors.

3. The 5G and AI multi-source fusion decision-making system for earthquake rescue according to claim 1, characterized in that, The specific steps for assessing the credibility of the multi-source raw data based on the stage identifier include: Based on the set of key risk factors identified at the current stage, dynamically adjust the weighting evaluation strategy for different data sources; If the stage is identified as the golden rescue stage, the weighting coefficient of life detection sensor data is increased, and the weighting coefficient of social media text data is decreased. If the stage is identified as a stage with a high incidence of secondary disasters, the weighting coefficient of geological and meteorological sensor data will be increased; The weighting coefficients are dynamically calculated using fuzzy hierarchical analysis combined with the stage identifiers to generate the weighted data set.

4. The 5G and AI multi-source fusion decision-making system for earthquake rescue according to claim 1, characterized in that, The specific steps for reconstructing the weighted data set based on the stage identifier include: Select the corresponding feature extraction backbone network according to the stage identifier, and perform feature encoding on the weighted data set; Using a generative adversarial network associated with the stage identifier, the encoded features are augmented and missing data is generated to form the data representation structure with a unified spatiotemporal scale; wherein, the data representation structure is a multi-channel spatiotemporal feature tensor constructed under a unified time step and a unified spatial resolution.

5. The 5G and AI multi-source fusion decision-making system for earthquake rescue according to claim 1, characterized in that, The scheduling control module is specifically a central scheduler, and the central scheduler operates in the following ways: The system receives the stage identifier and generates a corresponding trigger signal for the analysis module according to a preset rule mapping table, so as to determine whether to trigger the first analysis module, the second analysis module, or both in parallel. When the stage identifier indicates that the rescue force is limited and the signs of life are weak, the central dispatcher generates a trigger signal that only triggers the first analysis module, and performs spatial clustering and path optimization preprocessing on the probability data of the existence of life. The generated information on high-probability life sign clusters is input as an enhanced feature to the fusion decision module.

6. The 5G and AI multi-source fusion decision-making system for earthquake rescue according to claim 1, characterized in that, The specific steps for performing life-related information analysis based on the aforementioned data representation structure include: Receive the data representation structure and extract infrared thermal imaging features, micro-vibration sensing features, and sound spectrum features; The above features were input into multiple lightweight convolutional neural networks operating in parallel for preliminary analysis. The particle swarm optimization algorithm is used to dynamically weight and fuse the output confidence scores of multiple networks, and the parameters of the probability distribution model are optimized to generate gridded probability data of the existence of life.

7. The 5G and AI multi-source fusion decision-making system for earthquake rescue according to claim 1, characterized in that, The specific steps for analyzing environmental access-related information based on the aforementioned data representation structure include: Receive the data representation structure and extract UAV image features, terrain digital elevation model features, and real-time weather data; Construct a cost map with accessibility as the optimization objective; The wolf pack algorithm is used to simulate the collaborative exploration behavior of multiple rescue units in the cost map. Through the division of labor mechanism of the wolf pack, the access risk level of different geographical areas is dynamically assessed and marked, and the access level data is generated.

8. The 5G and AI multi-source fusion decision-making system for earthquake rescue according to claim 1, characterized in that, The pre-trained multi-agent reinforcement learning model in the fusion decision module is based on a reward function. The training and decision optimization formula is as follows: in: The weighting coefficients are dynamically adjusted based on the stage identifiers. Indicates the assembly of targets to be rescued. This represents the probability that target i exists at time t; This represents the set of action paths for rescue units. This represents the travel cost of path j at time t; Represents the set of rescue unit states. As an indicator function, when the estimated energy consumption of unit k... Below the safety threshold The value is 1 if it is true, and 0 otherwise.

9. The 5G and AI multi-source fusion decision-making system for earthquake rescue according to claim 1, characterized in that, After generating the rescue decision instruction, the system also performs the following steps: The rescue decision instructions are distributed to the rescue terminal via a 5G network; Receive instruction execution status data and real-time environmental change data from the rescue terminal; The instruction execution status data, the real-time environmental change data, and the updated stage identifier are used as inputs, and the policy network parameters of the multi-agent reinforcement learning model are dynamically adjusted using an online policy gradient algorithm. Based on the parameter-adjusted multi-agent reinforcement learning model, incremental decision instructions are generated and distributed.

10. The 5G and AI multi-source fusion decision-making system for earthquake rescue according to claim 1, characterized in that, When acquiring multi-source raw data with spatiotemporal identifiers using 5G networks with low latency, the specific steps include: Using 5G network slicing technology, spatiotemporal alignment and redundancy removal are performed on multi-source raw data; The multi-source raw data includes at least: seismic wave data from a distributed fiber optic sensor network, multimodal physiological data from rescue personnel's wearable devices, and disaster site text and image data from social media.

Images