Underground magnetic induction positioning method and earthquake emergency disaster reduction intelligent agent system

Abstract

The invention discloses an underground magnetic induction positioning method and an earthquake emergency disaster reduction intelligent agent system, and belongs to the field of natural disaster intelligent emergency response. A magnetic induction signal follows the physical characteristic of intensity attenuation in a near field to invert a target distance; in order to adapt to a complex buried medium, a layered medium model is introduced, a generalized transmission coefficient is obtained through interlayer superposition, attenuation and multipath effects caused by permeability, conductivity and thickness are accurately described, and high-precision penetrating type positioning of an underground target can be achieved on the premise of overcoming strong interference and adapting to complex medium distribution. On the basis, a virtual multi-view collaborative sensing mechanism is introduced, and data alignment and feature complementation rules among view angles of space-time fusion, virtual-real staggering and far-near alternation are designed, so that the system can generate a comprehensive decision-making scheme considering multiple targets of risk early warning, life rescue, damage assessment and the like when coping with the earthquake whole-cycle disaster situation.

Description

Technical Field

[0001] This invention belongs to the field of intelligent emergency response to natural disasters, and more specifically, relates to an underground magnetic induction positioning method and an intelligent system for earthquake emergency disaster reduction. Background Technology

[0002] The rapid urbanization process has led to a high concentration of population and economic factors, which has amplified natural disasters such as earthquakes causing building collapses, paralysis of lifeline projects, and secondary disasters, posing a serious threat to people's lives and property. In particular, magnetic induction positioning of underground targets has become an essential part of building an efficient and accurate emergency rescue system after a natural disaster.

[0003] Existing underground disaster relief positioning methods generally suffer from poor environmental adaptability: one type of method relies on far-field electromagnetic wave propagation, such as WiFi and ZigBee, which rapidly attenuates in conductive / high-loss media such as soil, concrete, and reinforced concrete due to strong absorption, multipath propagation, and non-line-of-sight propagation, leading to link instability or even penetration failure; GPS and other satellite positioning methods cannot cover underground scenarios; UWB, RFID, or radar-based short-range solutions are easily affected by eddy current interference and multipath propagation in mixed reinforced concrete-soil environments, with errors often exceeding 0.8 m, making it difficult to meet the accuracy requirements of rescue operations; acoustic and optical methods are limited by obstruction, multipath propagation, and impermeable media, easily leading to positioning ambiguity, and lacking real-time performance and reliability. In addition, although existing underground magnetic induction (MI) positioning has the advantage of near-field penetration, most of them are still based on the channel assumption of "homogeneous medium / simplified scenario," ignoring the layered non-homogeneous medium and the effective near-field coupling boundary, resulting in model mismatch with the field. Therefore, there is an urgent need to develop an underground magnetic induction positioning technology that can overcome strong interference, adapt to complex media distribution, and achieve high-precision penetrating positioning. Summary of the Invention

[0004] In view of the above-mentioned defects or improvement needs of the existing technology, the present invention provides an underground magnetic induction positioning method and an earthquake emergency disaster reduction intelligent system, the purpose of which is to achieve high-precision penetrating positioning under the premise of overcoming strong interference and adapting to complex media distribution.

[0005] To achieve the above objectives, in a first aspect, the present invention provides an underground magnetic induction positioning method for locating a target in an underground space carrying a coil device capable of transmitting and receiving magnetic induction signals; wherein M magnetic induction sensors are deployed at a predetermined location in the underground space; M is a positive integer; the magnetic induction sensors include: three coils that are orthogonal to each other in pairs; the first... The coil device carried on the target To the The first of the magnetic induction sensors coils A transmission link is formed between them. ; D; ; D represents the target quantity; The aforementioned underground magnetic induction positioning methods include: Multiple transmission signals S of different frequencies are transmitted into the underground space, and frequency response signals from each magnetic induction sensor are acquired. Calculate each frequency separately Downlink length ; R is a preset constant; R is a preset impedance; For frequency The generalized transmission coefficient is as follows: ; For frequency The corresponding angular frequency; N is the number of dielectric layers in the near-field region; the near-field region is the distance from the ground surface in the underground space. The area; ; ; and , respectively, are the minimum and maximum frequencies of the emitted signal S; c is the speed of light; and The first Layer and First Interlayer dielectric interface at angular frequency Transmission coefficient and equivalent Fresnel reflection coefficient at the given conditions; For the first Layered dielectric at angular frequency The complex propagation constant under these conditions; For the first The thickness of the medium layer; Frequency extracted from the frequency response signal Downlink The true RSSI value of the received signal from the upper magnetic induction sensor; Calculate transmission link final length ;in, The set of frequencies for transmitted signals; For frequency The corresponding weights; ; length , and The weighted summation result is used as the first The first goal and the first The distance between the magnetic induction sensors is used to achieve positioning.

[0006] More preferably, in terms of length , and When performing weighted summation, , and The corresponding weights are the normalized results of the link quality factor of the corresponding transmission link; Among them, transmission link The link quality factor is:

[0007] The equivalent transmission coefficient amplitude, ; For transmission links The shadow fading variance.

[0008] More preferably, Determined in the following ways: The frequency to be transmitted into underground space is obtained. The transmitted signal S was measured at time S. Received voltage and coil The emission current, and the ratio of the two is used as the frequency. Downlink The near-field coupling eigenvalues; Near-field coupling eigenvalues ​​and The product of and is used as frequency Downlink The frequency domain gain is then used to obtain the frequency. Downlink The predicted RSSI value of the received signal from the upper magnetic induction sensor; Will Subtracting the corresponding predicted value from the actual value yields the frequency. Downlink shadow disturbance Then, the corresponding shadow fading weights are calculated. ; The standard deviation of shadow fading; A graph is constructed, in which nodes include anchor nodes and target nodes. Anchor nodes represent magnetic induction sensors in underground space; target nodes represent coil devices carried on targets. Edges in the graph represent signal associations between nodes. The edges between anchor nodes and target nodes in the graph carry shadowing fading weights of each transmission link between the anchor node and the target node at different frequencies, as well as the true RSSI value of the received signal at the anchor node. For each anchor node in the graph, when the deviation of the true RSSI value of its received signal at any frequency from the mean of the true RSSI values ​​of its received signal at the same frequency from the same frequency of its one-hop neighbor anchor node exceeds a preset deviation, the true RSSI value of its received signal at that frequency is updated to the average of the true RSSI values ​​of the received signal at the same frequency of its one-hop neighbor anchor node. The graph data is input into a graph neural network. Through the message passing mechanism of the graph neural network, features from all transmission links are converged and fused according to frequency, resulting in fused features of each transmission link at different frequencies. Among these features, the transmission links... In frequency The fusion features are denoted as ; Calculate the score for each frequency, and then normalize it to obtain the weight for each frequency; where, frequency Corresponding score ; This represents a multilayer perceptron.

[0009] More preferably, the above-mentioned underground magnetic induction positioning method further includes: conducting infrared detection on the underground space to further determine the location of the target.

[0010] Secondly, the present invention provides an intelligent system for earthquake emergency response and disaster reduction, used to respond to earthquake emergencies in underground spaces where earthquake zones are located; wherein, M magnetic induction sensors are deployed at preset locations in the underground space; M is a positive integer; the magnetic induction sensors include: three coils that are orthogonal to each other in pairs; The aforementioned earthquake emergency disaster reduction intelligent system includes: an emergency response module for use after an earthquake, which executes the underground magnetic induction positioning method provided in the first aspect of the present invention to locate underground targets.

[0011] More preferably, the system further includes: a pre-disaster risk warning module, used to provide pre-disaster risk warnings before an earthquake occurs, including: The system acquires real-time three-component seismic wave data and corresponding metadata from seismic stations corresponding to the earthquake region. The metadata includes: the address information of the seismic stations and the recorded theoretical arrival time of the P-waves. For each component of the three-component seismic wave data, after preprocessing, the P-wave arrival time of that component is picked using the long-short time window mean ratio method. Furthermore, extracting from this component wave... The subband of the initial preset time period is taken as the corresponding P-band; The metadata and the P-bands corresponding to each component of the three-component seismic wave data are input into a pre-trained deep learning model to obtain the predicted results of earthquake magnitude, focal depth and time difference of arrival of seismic waves, and to provide pre-disaster risk warning.

[0012] More preferably, pre-disaster risk warning before an earthquake also includes: Historical earthquake catalog data for the earthquake region is obtained and preprocessed. The historical earthquake catalog data includes the time, location, focal depth, and magnitude of the earthquake. The preprocessing includes sorting the historical earthquake catalog data in chronological order and interpolating to complete any missing data. The earthquake region is divided into spatial grids, and the data of earthquakes whose locations are in the same spatial grid in the preprocessed historical earthquake catalog data are aggregated to obtain historical earthquake catalog data under different spatial grids; Feature extraction is performed on the historical earthquake catalog data under each spatial grid. After obtaining the corresponding earthquake features, PCA dimensionality reduction is performed to obtain the corresponding candidate features. The earthquake features include: the b-value of the slope used to describe the earthquake magnitude distribution in the Gutenberg-Richter relation, recurrence period features, strain energy release rate, spatial clustering, fault distance, and historical aftershock attenuation. Among them, the recurrence period features are: the mean, variance, and quantile of the interval between adjacent events in the historical earthquake catalog data. The candidate features corresponding to each spatial grid are input into a pre-trained machine learning model to obtain the probability of an earthquake occurring in that spatial grid. Based on the probability of an earthquake occurring in each spatial grid, a heat map and trend curve of earthquake risk for the seismic region are generated.

[0013] More preferably, pre-disaster risk warning before an earthquake also includes: weighted summation of video anomaly scores and audio anomaly scores in the earthquake zone to obtain an environmental precursor score; when the environmental precursor score exceeds a preset score, a precursor warning is issued. The video anomaly score for the earthquake zone is calculated as follows: Animal targets are detected in each frame of the acquired animal monitoring video stream of the earthquake area, and the centroid trajectory and motion velocity of each animal target at each frame time are calculated. Frames containing moving targets in animal surveillance video streams whose speed exceeds a preset speed and whose duration exceeds a first preset time are designated as abnormal frames. Frames containing animal groups in the animal monitoring video stream that last for more than a second preset time are considered abnormal frames; the number of animal targets in the animal group exceeds a preset number, and the minimum distance between any two animal targets does not exceed a preset distance; The ratio of the number of abnormal frames in the animal monitoring video stream to the total number of frames is calculated as the video anomaly score. The audio anomaly score for the earthquake zone is calculated as follows: A sliding window is used to divide the acquired environmental audio from the earthquake area into multiple environmental sub-audio segments; the audio features of each environmental sub-audio segment are extracted and concatenated with the three-axis acceleration of the synchronously acquired audio acquisition device to form the multimodal feature vector of the corresponding environmental sub-audio segment; Each environmental sub-audio segment is classified as abnormal or non-abnormal based on multimodal feature vectors; The ratio of the number of anomalous environmental sub-audio to the total number of environmental sub-audio is calculated as the audio anomalous score.

[0014] More preferably, the above-mentioned disaster emergency response module is also used to acquire social media data streams related to the earthquake zone after an earthquake occurs, including: text data and video data; For text data, the text data is tokenized to obtain a token sequence; the token sequence is encoded using BERT to obtain the context vector of each token; the sequence label header is used to predict whether each token is an earthquake-related entity based on its context vector, and the tokens in the token sequence that are classified as earthquake-related entities are extracted as target tokens; A classifier is used to calculate the probability that each target token belongs to the low, medium, and high public opinion urgency categories, respectively, and the probabilities are weighted and summed to obtain the public opinion urgency score of the corresponding target token. The sum of the public opinion urgency scores of all target tokens is taken as the total public opinion urgency score for the earthquake area; For video data, the system classifies whether an earthquake occurred in each frame of the video data, and calculates the ratio of the number of frames in the video data where an earthquake occurred to the total number of frames, which is used as the damage score for the earthquake area. Each target token in the involved locations is designated as a disaster-stricken location, and a disaster-stricken point distribution map of the earthquake zone is constructed. The population density at each disaster point is extracted from the population density heat map of the earthquake area, and after normalization, a population density distribution map of the disaster points in the earthquake area is constructed. For each disaster-stricken location, the weighted sum of its population density, total public opinion urgency score, and damage score for the earthquake region is calculated as an indicator of the disaster-stricken location's rescue needs, thereby constructing a heat map of rescue needs for disaster-stricken locations in the earthquake region.

[0015] More preferably, the system further includes: a post-disaster recovery and reconstruction guidance module; wherein, the post-disaster recovery and reconstruction guidance module includes: The macroscopic change detection unit is used to acquire real-time post-earthquake satellite imagery of the earthquake area, detect areas in the post-earthquake satellite imagery that have changed compared to the pre-earthquake satellite imagery, and mark them to guide post-earthquake recovery and reconstruction. The microscopic damage detection unit is used to acquire real-time data on each building in the seismic area from drones. Images; a classifier is used to classify the damage level of each building image, thereby obtaining the damage level of each building. Damage level quantification score ; The secondary disaster risk simulation unit is used to calculate the risk of each building in an earthquake zone based on geological environment data and historical earthquake case knowledge, using a Bayesian algorithm. The probability of a secondary disaster occurring at present is considered the secondary disaster risk of the building. ; Planning guidance unit, used to calculate each building Comprehensive priority score ;in, , , , All are preset coefficients; For architecture The preset function importance score; For architecture The estimated repair cost; As constraints, by maximizing ,right The solution is then performed to generate a reconstruction plan; among which, , Indicates no architecture Repair. Indicates the architecture Repairs will be carried out.

[0016] In summary, the above-described technical solutions conceived in this invention can achieve the following beneficial effects: 1. A first aspect of the present invention provides an underground magnetic induction positioning method, which utilizes magnetic induction signals to follow near-field... The physical characteristics of intensity attenuation are used to inversely determine the target distance. In order to adapt to complex buried media, a layered media model is introduced. The generalized transmission coefficient is obtained by superimposing layers, which accurately describes the attenuation and multipath effect caused by permeability, conductivity and thickness. Under the premise of overcoming strong interference and adapting to complex media distribution, high-precision penetrating positioning of underground targets can be achieved.

[0017] 2. Furthermore, considering that existing underground magnetic induction (MI) positioning technologies mostly rely on the RSSI intensity characteristics of single-frequency or a small number of links, unstable measurements are easily generated under magnetic field distortion caused by conductors such as steel bars; and the intensity characteristics may have multiple solutions, leading to ambiguity in the positioning solution and making it difficult to balance accuracy and stability. The underground magnetic induction positioning method provided by the first aspect of this invention adopts a multi-scale fusion framework of "local-mesoscale-global layer": In the local layer, anchor nodes exchange RSSI measurements of the same frequency; when the RSSI of a node deviates from the mean of its one-hop neighbor by more than a threshold, the RSSI is updated to the mean of the neighbor to suppress outlier interference; in the mesoscale layer, a graph data is constructed with nodes as vertices and signal associations as edges, and features including multi-frequency shadow fading weights are input into the graph neural network. Through message passing, the fusion features of each transmission link at different frequencies are obtained, and the scores corresponding to each frequency are obtained. After normalization, the weights corresponding to each frequency are obtained; thus, in the global layer, the multi-frequency distance estimation results are weighted and fused to output the final distance, further improving the accuracy and reliability of the positioning results.

[0018] 3. The second aspect of the present invention provides an earthquake emergency disaster reduction intelligent system. After an earthquake occurs, the system executes the underground magnetic induction positioning method provided in the first aspect of the present invention, which can achieve high-precision penetrating positioning of underground targets under the premise of overcoming strong interference and adapting to complex media distribution.

[0019] 4. The earthquake emergency disaster reduction intelligent agent system provided in the second aspect of this invention introduces a virtual multi-view collaborative perception mechanism and designs data alignment and feature complementarity rules between various perspectives that are spatiotemporally integrated, virtual and real, and near and far. This enables the system to generate comprehensive decision-making schemes that take into account multiple objectives such as risk warning, life rescue, and damage assessment when responding to earthquake disasters throughout their entire lifecycle. Furthermore, the intelligent agent cluster enables dynamic optimization and centralized scheduling of computing resources, reducing the data fragmentation and high coordination costs associated with relying on multiple independent monitoring systems in existing technologies. Therefore, it can achieve full-cycle, all-round emergency decision support with a unified intelligent agent architecture.

[0020] 5. Furthermore, the earthquake emergency disaster reduction intelligent system provided in the second aspect of the present invention includes a pre-disaster risk early warning module, which analyzes multimodal data to provide accurate early warning before an earthquake occurs.

[0021] 6. Furthermore, the earthquake emergency disaster reduction intelligent system provided in the second aspect of the present invention includes a post-disaster recovery and reconstruction guidance module, used to conduct a graded assessment of macro-regional damage and micro-building damage through collaborative observation by space-based satellites and airborne drones during the emergency response phase after an earthquake. It can provide forward-looking planning for a seamless transition from emergency response to recovery and reconstruction, and, combined with prior knowledge, predict secondary disaster risks, achieving closed-loop decision support throughout the entire disaster response cycle. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the hierarchical structure of the high-fidelity channel model provided in an embodiment of the present invention; Figure 2 A schematic diagram illustrating the results of disaster-affected population orientation vector determination using infrared directional and magnetic induction ranging modes provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the multimodal architecture of the earthquake emergency disaster reduction intelligent agent system provided in an embodiment of the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0024] Example 1 This embodiment provides an underground magnetic induction positioning method for locating a target in an underground space carrying a coil device capable of transmitting and receiving magnetic induction signals; wherein, M magnetic induction sensors are deployed at predetermined locations in the underground space; M is a positive integer; the magnetic induction sensors include: three coils that are orthogonal to each other in pairs; the first... The coil device carried on the target To the The first of the magnetic induction sensors coils A transmission link is formed between them. ; D; ; D represents the target quantity.

[0025] The aforementioned underground magnetic induction positioning methods include: Multiple different frequencies (in this embodiment, 300kHz) are transmitted into the underground space. The system transmits a signal S at multiple frequencies selected within the 1MHz band and acquires the frequency response signals from each magnetic induction sensor. It should be noted that, under the excitation of the transmission signal S, the coil device on the target transmits a signal to the magnetic induction sensor, thus forming the aforementioned transmission link.

[0026] Calculate each frequency separately Downlink length ; R is a preset constant; R is a preset impedance; For frequency The generalized transmission coefficient is as follows: ; For frequency The corresponding angular frequency; N is the number of dielectric layers in the near-field region; the near-field region is the effective working area of ​​the MI link based on Fresnel near-field theory, specifically manifested as the distance from the ground in underground space. The area; ; ; and , respectively, are the minimum and maximum frequencies of the emitted signal S; c is the speed of light; and The first Layer and First Interlayer dielectric interface at angular frequency Transmission coefficient and equivalent Fresnel reflection coefficient at the given conditions; For the first Layered dielectric at angular frequency The complex propagation constant under these conditions; For the first The thickness of the medium layer; Frequency extracted from the frequency response signal Downlink The true RSSI value of the received signal from the upper magnetic induction sensor. Specifically, the first... Layer and First Interlayer dielectric interface at angular frequency Transmission coefficient and equivalent Fresnel reflection coefficient , No. Layered dielectric at angular frequency The complex propagation constant under the following conditions They are respectively:

[0027]

[0028]

[0029] in, For the first The intrinsic wave impedance of the layered medium is specifically... ; It is the symbol for imaginary numbers; Corresponding to the number The permeability, conductivity, and dielectric constant of the layered medium.

[0030] Calculate transmission link final length ;in, The set of frequencies for transmitted signals; For frequency The corresponding weights; ; length , and The weighted summation result is used as the first The first goal and the first The distance between the magnetic induction sensors is used to achieve positioning.

[0031] To adapt to complex buried media, this embodiment introduces a layered media model, and obtains the generalized transmission coefficient through interlayer superposition. It is used to characterize the attenuation and multipath effects caused by permeability, conductivity and thickness.

[0032] It should be noted that the above refers to length , and The weights used in weighted summation can be determined based on empirical values ​​or the performance of the transmission link. Preferably, in one optional implementation, when considering the length... , and When performing weighted summation, , and The corresponding weights are the normalized results of the link quality factor of the corresponding transmission link; Among them, transmission link The link quality factor is:

[0033] The equivalent transmission coefficient amplitude, ; For transmission links The shadow fading variance.

[0034] It should be noted that the above weights This can be determined based on empirical values. Considering that existing underground magnetic induction (MI) positioning technologies mostly rely on the RSSI intensity characteristics of single-frequency or limited-link data, unstable measurements are easily generated under magnetic field distortion caused by conductors such as reinforcing bars; furthermore, the intensity characteristics may exhibit multiple solution mappings, leading to ambiguity in the positioning calculation and making it difficult to balance accuracy and stability. Preferably, in one optional implementation, Determined in the following ways: The frequency to be transmitted into underground space is obtained. The transmitted signal S was measured at time S. Received voltage and coil The emission current, and the ratio of the two is used as the frequency. Downlink The near-field coupling eigenvalues; Near-field coupling eigenvalues ​​and The product of and is used as frequency Downlink The frequency domain gain is then used to obtain the frequency. Downlink The predicted RSSI value of the received signal from the upper magnetic induction sensor; Will Subtracting the corresponding predicted value from the actual value yields the frequency. Downlink shadow disturbance Then, the corresponding shadow fading weights are calculated. ; The standard deviation of shadow fading; A graph is constructed, in which nodes include anchor nodes and target nodes. Anchor nodes represent magnetic induction sensors in underground space; target nodes represent coil devices carried on targets. Edges in the graph represent signal associations between nodes. The edges between anchor nodes and target nodes in the graph carry shadowing fading weights of each transmission link between the anchor node and the target node at different frequencies, as well as the true RSSI value of the received signal at the anchor node. For each anchor node in the graph, when the deviation of the true RSSI value of its received signal at any frequency from the mean of the true RSSI values ​​of its received signal at the same frequency from the same frequency of its one-hop neighbor anchor node exceeds a preset deviation, the true RSSI value of its received signal at that frequency is updated to the average of the true RSSI values ​​of the received signal at the same frequency of its one-hop neighbor anchor node. The graph data is input into a graph neural network. Through the message passing mechanism of the graph neural network, features from all transmission links are converged and fused according to frequency, resulting in fused features of each transmission link at different frequencies. Among these features, the transmission links... In frequency The fusion features are denoted as ; Calculate the score for each frequency, and then normalize it to obtain the weight for each frequency; where, frequency Corresponding score ; This represents a multilayer perceptron.

[0035] In one alternative implementation, the above-mentioned underground magnetic induction positioning method further includes: performing infrared detection on the underground space to further determine the location of the target.

[0036] A life detection system based on the fusion of infrared detection and magnetic induction (MI) signals was constructed to achieve three-dimensional spatial positioning of trapped personnel in collapsed and burial environments and complex media. The system provides orientation information through infrared thermal imaging, distance information through MI near-field coupling, and enhances anti-interference capabilities through multi-frequency feature fusion.

[0037] In one specific implementation, infrared azimuth estimation is performed by acquiring thermal images using an infrared detector (such as an uncooled infrared detector). Perform two-point non-uniform correction: , Corresponding gain coefficient, The offset compensation coefficients are applied, and histogram equalization is performed for enhancement. The field of view is divided into several azimuth sectors. Statistical sector average thermal radiation intensity Candidate target regions are obtained by matching human body temperature feature templates. The imaging coordinates of these candidate regions are then calculated using the sub-pixel centroid method. : , Then from the camera's internal parameters K Convert pixel coordinates into normalized view vectors Output the target unit direction vector .

[0038] It should be noted that the aforementioned target quantity D can be determined in advance in various ways, and no limitation is made here. For example, in one optional implementation, it can be determined by infrared detection; in another optional implementation, the signal generated by the coil device carried on the target carries information indicating the device identification, and the number of targets can be determined by statistically analyzing the information indicating the device identification on the magnetic induction signals received from the target by each magnetic induction sensor.

[0039] To further illustrate the underground magnetic induction positioning method provided in this embodiment, the key technologies are described in detail below: 1) Basic structure of MI near-field coupling and three-coil array This invention employs a magnetic induction (MI) near-field coupling mechanism to achieve penetrating ranging and positioning in complex underground / ruin environments. M magnetic induction sensors are deployed at predetermined locations within the underground space; M is a positive integer. Each magnetic induction sensor includes three mutually orthogonal coils to improve directional decoupling and spatial coverage. Each node can transmit an MI signal carrying node identification and link status information, and can adaptively control its transmission power based on burial depth and propagation distance. The nodes include anchor nodes with known locations (such as rescue-side terminals) and target nodes whose locations are to be determined (such as beacons on the trapped side).

[0040] Under magnetically quasi-static (MQS) near-field conditions, the transmitting coil can be equivalent to a magnetic dipole, which at a distance... The magnitude of the magnetic flux density at that location follows an approximately inverse cubic decay law:

[0041] Furthermore, with magnetic dipole moment If the equivalent excitation of the transmitting coil is given, then the spatial magnetic field can be expressed using a dipole model as follows:

[0042] in The permeability of the medium, Let be a unit vector. The induced voltage of the receiving coil can be written as:

[0043] in For transmitting current, For mutual inductance, and related to coil geometry and attitude angles ,distance It is related to the parameters of the medium.

[0044] 2) Fresnel near-field zone delineation and effective working range determination The effective operating region of the MI link is divided according to Fresnel near-field theory: within this region, magnetic flux coupling energy dominates, and beyond this region, the coupling contribution rapidly weakens.

[0045] According to free space wavelength The magnitude provides an engineering criterion for near-field boundaries:

[0046] Where c is the speed of light. The carrier frequency is selected; the MI operating frequency band can be selected from 300kHz to 1MHz to balance the requirements of penetration and bandwidth / anti-interference.

[0047] 3) High-fidelity MI channel model integrating "Fresnel near-field + layered medium transmission" To address the attenuation and multipath problems caused by the heterogeneous, non-uniform, and layered distribution of media such as "soil-rock-concrete-metal components" in ruins / underground endogenous environments, this embodiment constructs a high-fidelity channel model: abstracting the non-uniform medium along the transmission path as... Layered dielectric superposition, the first The parameter set of the layer medium is defined as follows: , representing magnetic permeability, dielectric constant, electrical conductivity, and thickness, respectively, such as Figure 1 As shown. Define the first Complex propagation constant of layered medium With intrinsic impedance To characterize the absorption loss, phase delay, and skin effect of the medium:

[0048]

[0049] Signal at the interface between two adjacent dielectric layers Reflection and transmission occur at a certain point, and its equivalent Fresnel reflection coefficient is... With transmission coefficient for:

[0050] Considering the multiple reflection and superposition effects of signals in layered media, a recursive form of the generalized transmission coefficient is constructed. Assuming the signal passes through... The overall transmission efficiency of the layered medium can be expressed as:

[0051] This formula precisely quantifies the thickness of each dielectric layer. The cumulative effect of electromagnetic parameters on signal amplitude attenuation and phase rotation.

[0052] It should be noted that the multiple reflection superposition effect originates from classical electromagnetic wave propagation theory and Fresnel's formula. In underground or ruined environments, magnetic induction signals are reflected at interfaces (such as the soil-concrete interface) when passing through different media layers such as soil, concrete slabs, and air gaps. Due to the layered structure underground, the signal will be reflected multiple times between layers. The signal received by the receiver is not a signal from a single path, but a vector superposition of signals from countless reflected paths.

[0053] Therefore, the frequency domain gain of the link can be written as:

[0054] The frequency domain gain of the link is used to characterize the attenuation and multipath effects caused by permeability, conductivity, and thickness. It characterizes the pure magnetic field coupling capability under ideal or homogeneous dielectric conditions, determined solely by the geometry, physical parameters, and relative attitude of the transmitting and receiving coils. It is the "reference signal strength" after eliminating interference from reflections in complex multilayered dielectrics.

[0055] A high-fidelity model is a hybrid process of "pre-modeling + on-site parameter calibration": The pre-modeling part: The mathematical formula structure of "Fresnel near field + layered medium transmission" (such as the recursive formula of the generalized transmission coefficient) is derived in advance and written into the algorithm.

[0056] Correlation of field measurements (parameter calibration): Parameters in the formula (such as soil electrical conductivity) Dielectric layer thickness The result was obtained through on-site inversion using a multi-frequency scanning strategy.

[0057] Workflow: 1. Send multi-frequency scanning signals (300kHz-1MHz) into underground space.

[0058] 2. By utilizing the differences in the penetration ability (skin depth) of different frequencies into the medium, the environmental conditions at the scene can be deduced. and .

[0059] 3. Input the parameters obtained from these field measurements into the pre-established formula to calculate the current channel gain.

[0060] 4) Localization feature extraction based on "spatial correlation of shadow fading" To enhance recognizability, a location feature extraction based on the spatial correlation of shadow fading caused by the target is introduced: the target's disturbance to the magnetic field is not only reflected in intensity attenuation, but also presents a recognizable spatial distribution pattern.

[0061] Logarithmic domain statistics of shadow fading: Received signal power It follows a normal distribution over the logarithmic field, and its probability density function can be expressed as:

[0062] in, Distance Average path loss at that location, The standard deviation of shadow fading.

[0063] Using shadow fading weights To describe this effect, and further introduce the transmission coefficient into the formation To more accurately characterize attenuation patterns in underground environments, thereby obtaining high-dimensional and robust localization features. For any link The difference between the measured RSSI and the model-predicted RSSI is used to characterize the shadowing perturbation:

[0064] The RSSI prediction formula (dBm) is as follows:

[0065] Or it can be expanded into voltage form:

[0066] in, This refers to the transmission power. This is the overall system gain / loss term, used to align the theoretical calculations with the actual circuit measurements. This is the generalized transmission coefficient, reflecting the medium loss. This is a near-field coupling eigenvalue, reflecting the range. attenuation.

[0067] By combining the layered medium transmission coefficient, a shadowing fading weight is constructed:

[0068] in, This represents the amplitude of the generalized transmission coefficient for the layered medium. This applies to multiple nodes and multiple frequency points. Row stacking can form a high-dimensional "fading map / fingerprint" feature vector. Used for subsequent positioning estimation.

[0069] Shadow fading may be symmetrically distributed along a single link, leading to localization ambiguity. Therefore, it is preferable to combine multi-link (multi-directional) and multi-frequency scanning to resolve ambiguity during the feature fusion stage.

[0070] 5) Multi-frequency scanning and multi-scale fusion positioning To compensate for the characteristic drift caused by formation inhomogeneity and conductor / reinforcing bar eddy current interference, a multi-frequency scanning strategy is adopted: K frequency points / bands are selected within the 300 kHz to 1 MHz frequency band, the transmitter sends a known excitation sequence, the receiver measures the frequency response, and the key electromagnetic parameters are jointly estimated / updated. Layer thickness This process can be completed naturally during communication without the need for external instruments. Edge nodes are compared with historical parameters, and if an anomaly is detected (such as a sudden change in conductivity), CSI updates, fusion weight adjustments, and node strategy optimizations are automatically triggered to dynamically correct the deviation.

[0071] By introducing graph neural networks (GNNs), message passing and feature aggregation are performed on the node graph topology to learn the frequency characteristics and nonlinear coupling relationships of adjacent regions, thereby improving complementarity and reducing the impact of conflicting features. Graph Neural Network (GNN) Fusion: Constructing Topological Graph Structures , where nodes Representing each sensor node, edge This represents the signal correlation between nodes. Each node includes: anchor node (i.e., the deployed magnetic induction sensor) and target node (i.e., the terminal worn by the trapped personnel, or the buried black box / device).

[0072] The edge (measurement link) between the anchor node and the target node is the most crucial edge in the graph. Physically, this corresponds to the process where the target node (transmitter) transmits a magnetic induction signal to the anchor node (receiver), and the anchor node receives and measures the signal. In graph data, this is represented as: target node and anchor node There exists an edge between them, and the features on the edge include the measured RSSI values ​​of multiple frequency points of the link, shadowing fading weights, etc. This is the basic data used for distance inversion.

[0073] Edges between anchor nodes (negotiated links) are another important type of edge in graph data. Physically, this corresponds to the exchange of RSSI measurements and interference characteristics between adjacent anchor nodes via magnetic induction communication (i.e., the "neighbor negotiation mechanism"). In graph data, this is represented as: anchor nodes and other adjacent anchor nodes There are edges between them. It is through these edges that nodes can obtain information about their "one-hop neighbors" and thus determine whether their own data is an outlier.

[0074] Node set Target node For the device to be positioned; each anchor node It can be expanded into 3 "coil sub-nodes". Used to explicitly represent three orthogonal links.

[0075] edge set Target node With each anchor coil child node Connecting a side , represents a MI transmission link; anchor nodes are connected by spatial proximity or communication neighbor to achieve neighborhood consistency / negotiation.

[0076] Edge characteristics: For each edge Stack the multi-frequency shadow weights into a vector:

[0077] Other link quality metrics (SNR, etc.) can also be spliced ​​together. Statistics wait).

[0078] The fusion algorithm employs a multi-scale fusion framework of "local-mesoscale-global layer": In the local layer, anchor nodes exchange same-frequency RSSI measurements; when the deviation of a node's RSSI from the mean of its one-hop neighbors exceeds a threshold, the RSSI is updated to the neighbor mean to suppress outlier interference; in the mesoscale layer, a graph data structure is constructed with nodes as vertices and signal correlations as edges, and the multi-frequency shadow fading weights / residual features are input into the graph neural network, outputting the fusion score for each frequency point through message passing. The weights are obtained by softmax. It is used to weight and fuse multi-frequency distance estimation results at the global layer and output the final distance / coordinates.

[0079] Finally, the weight corresponding to each frequency is obtained, and the specific process is as follows: go through L After layer message passing, the hidden representation of each edge (link) is obtained. (for each edge) (A learned vector representation) For each frequency point The frequency score is obtained by performing a "global aggregation" of information from all links on that frequency point. :

[0080] For MLP ( The link-level fusion coefficients of the multilayer perceptron mapping represent the reliability weights of the edges.

[0081] The frequency point fusion weights are obtained by normalizing using softmax:

[0082] K represents the total number of frequencies.

[0083] Based on the above key technologies, the entire process of the underground magnetic induction positioning method is as follows: S1: Node Deployment and Network Initialization Deploy a network of magnetic induction (MI) sensor nodes at the boundaries and key locations of the underground or ruined space to be located. Anchor node configuration: Record the three-dimensional coordinates of each anchor node (known location). And its unique device identifier ID. Edge computing deployment: Deploy edge computing nodes on the ground or in secure areas, and aggregate monitoring data from underground MI nodes through long-distance gateways.

[0084] S2: Multi-target recognition and collision avoidance data acquisition Each MI node performs multi-frequency scanning and data exchange within the 300kHz–1MHz frequency band. To determine the number and identity of multiple targets, the following communication protocol procedure is executed: Identity resolution: The receiver demodulates the digital frame structure of the MI signal, parses the Node ID in the frame header, distinguishes different target sources based on different IDs, and counts the number of targets.

[0085] Collision avoidance mechanism: CSMA / CA (Carrier Sense Multiple Access / Collision Avoidance) or TDMA (Time Division Multiple Access) mechanism is adopted. Before transmission, the node listens to the channel magnetic field strength. If the channel is busy, random backoff is performed to ensure that multiple target signals are staggered in the time domain and avoid co-channel interference.

[0086] Data Acquisition: After establishing a stable link, the Received Signal Strength (RSSI), Channel State Information (CSI), and Complex Frequency Response data are collected at each frequency point.

[0087] S3: High-fidelity channel modeling and parameter inversion Edge nodes determine the effective coupling range based on the Fresnel near-field criterion and construct a hierarchical medium generalized transmission model.

[0088] By utilizing the skin depth differences in multi-frequency scan data, the electromagnetic parameters of the environmental medium can be inverted.

[0089] Calculate the reflection and transmission coefficients at the interlayer interface, and recursively generate the generalized transmission coefficient and theoretical link gain that include the multipath superposition effect.

[0090] S4: Shadow Fading Feature Extraction The residuals between the measured RSSI and the model predictions are calculated to quantify the shadow fading perturbation. Shadow fading weights reflecting the degree of medium interference are constructed to form a high-dimensional location feature vector covering multiple nodes and multiple frequency points.

[0091] S5: Multi-scale Feature Deep Fusion The feature vector is input into a "local-mesoscale-global layer" fusion framework. The local layer performs weighted filtering on the time-series data. The mesoscale layer uses a graph neural network (GNN) to perform message passing on the node topology, aggregating neighborhood features. The global layer outputs a unified fused feature vector, which represents the channel confidence of the current link at various frequency points.

[0092] S6: Distance Inversion Based on Fusion Feature Weighting Instead of the traditional single-point direct inversion, we first invert each frequency point, and then calculate the precise distance between the target and each anchor node by fusion of distances according to multi-frequency weights.

[0093] S7: Multi-objective 3D coordinate solution Based on the target ID determined in step S2 and the distance set calculated in step S6. (Target To anchor node (distance): fused with infrared direction vector Then directly through Perform rapid positioning; among which, The origin of the coordinate system is set as the preset coordinate point for the space containing all sensors (including infrared detectors and magnetic sensors). In multi-sensor scenarios, least-squares optimization is established based on multiple ray / distance constraints: or This allows us to obtain the three-dimensional spatial location of the trapped personnel. For example... Figure 2 The image shows the results of the disaster-affected population's orientation vector determination using infrared directional and magnetic induction ranging modes provided in this embodiment.

[0094] S8: Adaptive Feedback and Dynamic Correction Real-time monitoring of inversion residuals and environmental parameters: If a sudden change in environmental conductivity or a positioning residual exceeding a threshold is detected, the CSI fingerprint database is updated. The fusion weight parameters in step S6 and the medium model parameters in step S3 are dynamically adjusted to form a closed-loop feedback to maintain positioning accuracy and stability in complex time-varying environments.

[0095] Example 2 An intelligent system for earthquake emergency response and disaster reduction is provided for earthquake emergency response in underground spaces located in earthquake zones. M magnetic induction sensors are deployed at predetermined locations in the underground space, where M is a positive integer. Each magnetic induction sensor includes three coils that are orthogonal to each other in pairs. The aforementioned earthquake emergency disaster reduction intelligent system includes: an emergency response module for disaster response, which is used to execute the underground magnetic induction positioning method provided in Embodiment 1 of the present invention to locate underground targets after an earthquake occurs.

[0096] The related technical solutions are the same as those provided in Embodiment 1 of this invention for underground magnetic induction positioning, and are not limited here.

[0097] In one optional implementation, the system further includes: a pre-disaster risk warning module, used to provide pre-disaster risk warnings before an earthquake occurs, including: The system acquires real-time three-component seismic wave data and corresponding metadata from seismic stations corresponding to the earthquake region. The metadata includes: the address information of the seismic stations and the recorded theoretical arrival time of the P-waves. For each component of the three-component seismic wave data, after preprocessing, the P-wave arrival time of that component is picked using the long-short time window mean ratio method. Furthermore, extracting from this component wave... The subband of the initial preset time period is taken as the corresponding P-band; The metadata and the P-bands corresponding to each component of the three-component seismic wave data are input into a pre-trained deep learning model to obtain the predicted results of earthquake magnitude, focal depth and time difference of arrival of seismic waves, and to provide pre-disaster risk warning.

[0098] It should be noted that the above process is a three-component earthquake first wave multi-task prediction. In a specific implementation, data access, preprocessing, automatic labeling of P-wave first arrival, deep learning multi-task inference and confidence calibration are performed on the physical monitoring signal (standard three-component earthquake wave), and key parameters such as phase arrival time difference, magnitude and focal depth are output.

[0099] (1) Data Access and Integrity Verification: Download the three-component waveform and metadata from the cloud-based PKL data address provided by the user, and verify the file hash / length, field integrity, and sampling rate consistency. Preferably, the waveform matrix of a single sample is... The channels are ordered as Z (vertical), N (north-south), and E (east-west), with a sampling rate of fs = 100 Hz and a window length of 60 s; the metadata vector is... It must include at least the station's latitude and longitude, station number, event number, and P-wave arrival time. Fields such as event directory and magnitude.

[0100] (2) Denoising and normalization: For each component Perform DC removal A digital low-pass filter is used to suppress high-frequency environmental noise. An IIR Butterworth low-pass filter is preferred, and its difference equation can be expressed as:

[0101] in, These are the filter coefficients. After filtering, they are normalized by channel: ,middle , These represent the mean and standard deviation of the channel within the window, respectively. To protect the minimum value for stability.

[0102] (3) Automatic labeling of P-wave first arrival (STA / LTA): Calculate the short-window average STA and long-window average LTA for each component:

[0103] Construct ratio and take .when And continue When sampling points are selected, the corresponding time is determined as the first arrival of the P wave. Then with Cut a fixed-length window for the alignment point (preferably from...) The initial 6000 points are used as model input and written into metadata m. Label value.

[0104] (4) Multi-task CNN real-time inference: The input waveform X and metadata m are fed into a pre-trained deep convolutional neural network (CNN), and feature extraction uses multi-layer one-dimensional convolution:

[0105] Where “*” indicates one-dimensional convolution along the time axis, The activation function is non-linear; then, BatchNorm and a max-pooling layer (pool(·)) are used for downsampling and dimensionality reduction. The network shares a backbone to extract the representation vector z, and sets three regression output heads to output the following: Magnitude estimate and estimated focal depth .

[0106] (5) Confidence rating and output: Uncertainty estimation of the regression output is performed based on the error distribution of the validation set. Preferably, the confidence level is determined separately for each validation set error distribution. Establish error standard deviation and output the confidence level. Or give Confidence intervals in the form of ( (Corresponding to approximately 95% confidence level); or Monte Carlo Dropout can be used to repeat forward inference K times, calculating the mean and variance as the confidence level, which is then used for weight allocation in subsequent multimodal fusion.

[0107] (6) Model training process (completed offline, online inference only): a) Training set construction: Collect three-component station waveforms and manual / automatic phase picking results from historical earthquake events, forming samples by event-station. Each sample contains: waveform (6000×3) Metadata (1×35) and supervision label .

[0108] b) ;in Obtained from the P / S wave arrival times marked in the sample. This is the magnitude truth value in the event catalog. This is the true value of the focal depth in the event catalog.

[0109] c) Data preprocessing: First, merge and unify the original block data and output processed_data / xxx.pkl; then, perform channel normalization, random time jitter (±δt perturbation near the alignment point), amplitude scaling and noise injection enhancement on the training samples to improve the generalization ability to different station instruments and environmental noise.

[0110] d) Dataset partitioning: Partition the dataset by event or by time to form training / validation / test sets that do not leak from each other, with the preferred ratio being 80% / 10% / 10%.

[0111] e) Loss Function and Optimization: Multi-task weighted mean squared error (MSE) is used.

[0112] in, , , The task weights can be dynamically adjusted by the validation set to balance dimensional differences. The Adam optimizer is used to iteratively update the parameters, with optimal training for 300 epochs and early stopping with validation set loss. After training, the final weight file (e.g., final_multi_task_model_*.pth) is obtained for fast inference in the online phase.

[0113] It should be noted that the metadata initially contains This information typically comes from a unified directory provided by the earthquake network center. This is a theoretical arrival time based on comprehensive positioning from multiple stations, or a roughly manually annotated time. For a specific station, it may have slight deviations (e.g., hundreds of milliseconds). Before inputting it into the CNN model, the waveform's "phase alignment" must be ensured. The STA / LTA algorithm calculates the physical signal abrupt change points based on the actual waveforms received by the current station. Writing the precise trigger time calculated by STA / LTA into the metadata is to correct for deviations in the initial directory values, ensuring that the 6000 waveform points input to the neural network are extracted based on the "true P-wave first arrival." In future practical use (inference phase), there will be no "data in the metadata." The system relies entirely on STA / LTA to capture... This is done during the training phase to ensure that the processing flow of training data is consistent with the real-world scenario.

[0114] In one alternative implementation, pre-disaster risk warning before an earthquake also includes: Historical earthquake catalog data for the earthquake region is obtained and preprocessed. The historical earthquake catalog data includes the time, location, focal depth, and magnitude of the earthquake. The preprocessing includes sorting the historical earthquake catalog data in chronological order and interpolating to complete any missing data. The earthquake region is divided into spatial grids, and the data of earthquakes whose locations are in the same spatial grid in the preprocessed historical earthquake catalog data are aggregated to obtain historical earthquake catalog data under different spatial grids; Feature extraction is performed on the historical earthquake catalog data under each spatial grid. After obtaining the corresponding earthquake features, PCA dimensionality reduction is performed to obtain the corresponding candidate features. The earthquake features include: the b-value of the slope used to describe the earthquake magnitude distribution in the Gutenberg-Richter relation, recurrence period features, strain energy release rate, spatial clustering, fault distance, and historical aftershock attenuation. Among them, the recurrence period features are: the mean, variance, and quantile of the interval between adjacent events in the historical earthquake catalog data. The candidate features corresponding to each spatial grid are input into a pre-trained machine learning model to obtain the probability of an earthquake occurring in that spatial grid. Based on the probability of an earthquake occurring in each spatial grid, a heat map and trend curve of earthquake risk for the seismic region are generated.

[0115] The above process is a probability prediction of seismic ripple spatiotemporal spectrum and historical catalog. In a specific implementation, for spatiotemporal analysis data, a spatiotemporal evolution model is constructed by combining historical earthquake catalog to explore the regional seismic activity patterns and predict the probability of occurrence within the future window, thereby realizing medium- and long-term risk assessment and spatial zoning.

[0116] (1) Data access and cleaning: Access the historical earthquake catalog CSV containing fields such as time, latitude and longitude (lat, lon), focal depth, and magnitude (mag); unify the time into "YYYY-MM-DD HH:MM:SS" and sort by time; use Kriging interpolation to complete missing geographic points, and its estimated value is ; For known geographical points; for Actual observed values ​​of focal depth and magnitude; weights The equations are solved using the semivariance function constraint equations; threshold clipping / mean backfilling is applied to the abnormal magnitude and depth, and Z-score standardization is performed.

[0117] (2) Feature construction: Statistical features based on spatial grids (preferably 1km×1km) and sliding time windows (preferably 7–30 days), including: a) the b-value of the Gutenberg–Richter frequency–magnitude relationship: ,in For earthquakes with a magnitude greater than or equal to the magnitude The number of events; maximum likelihood estimation is preferred. ,in This represents the average earthquake magnitude within the window. To complete the magnitude threshold, It serves to correct discrete errors. b) Recurrence cycle characteristics: interval between adjacent events. c) Mean, variance, and quantiles; d) Strain energy release rate: based on the energy-magnitude empirical relationship. (C is a constant), estimate the energy. The release rate is obtained by summing / differentiating over a time window; d) supplementary features such as spatial clustering, fault distance, and historical aftershock attenuation. High-dimensional features are reduced using PCA, and principal components with a cumulative contribution rate ≥95% are retained as model input.

[0118] (3) Supervision of labels and training set: The historical timeline is divided into "feature window (past) - prediction window (future)". The input is the feature vector within the past window; the label is defined as the future window. (Preferred 7 days) Has an earthquake with a magnitude ≥ occurred in this grid? Earthquake events were classified into two categories and labeled. Alternatively, the label can be defined as the maximum magnitude within a future window. Used for regression.

[0119] (4) Model Training and Inference: Construct an ensemble model of XGBoost and Random Forest. Grid search is preferred for tuning hyperparameters such as tree depth, learning rate, and subsampling rate; cross-validation is used to select the optimal combination. The probability of gridding occurs is output during the inference phase. The risk is divided into three levels: low, medium, and high, based on the threshold, and a risk heatmap and trend curve are generated.

[0120] In one optional implementation, pre-disaster risk warning before an earthquake occurs further includes: weighted summation of video anomaly scores and audio anomaly scores in the earthquake zone to obtain an environmental precursor score; and issuing a precursor warning when the environmental precursor score exceeds a preset score. The video anomaly score for the earthquake zone is calculated as follows: Animal targets are detected in each frame of the acquired animal monitoring video stream of the earthquake area, and the centroid trajectory and motion velocity of each animal target at each frame time are calculated. Frames containing moving targets in animal surveillance video streams whose speed exceeds a preset speed and whose duration exceeds a first preset time are designated as abnormal frames. Frames containing animal groups in the animal monitoring video stream that last for more than a second preset time are considered abnormal frames; the number of animal targets in the animal group exceeds a preset number, and the minimum distance between any two animal targets does not exceed a preset distance; The ratio of the number of abnormal frames in the animal monitoring video stream to the total number of frames is calculated as the video anomaly score. The audio anomaly score for the earthquake zone is calculated as follows: A sliding window is used to divide the acquired environmental audio from the earthquake area into multiple environmental sub-audio segments; the audio features of each environmental sub-audio segment are extracted and concatenated with the three-axis acceleration of the synchronously acquired audio acquisition device to form the multimodal feature vector of the corresponding environmental sub-audio segment; Each environmental sub-audio segment is classified as abnormal or non-abnormal based on multimodal feature vectors; The ratio of the number of anomalous environmental sub-audio to the total number of environmental sub-audio is calculated as the audio anomalous score.

[0121] It should be noted that the above process is for the identification of abnormal animal behavior and multimodal biological precursors. In one specific implementation, for biological perception characteristics, animal behavior videos, animal vocalization audio, and accelerometer data are integrated, and a biological precursor score is generated through anomaly detection and multimodal fusion for comprehensive pre-disaster judgment.

[0122] (1) Video-side anomaly detection: Acquire animal monitoring video streams and perform frame decoding: preferably extract frames at 5 fps and perform grayscale conversion and jitter removal preprocessing. Use Gaussian mixture model (GMM) for background modeling to obtain the foreground mask. Morphological opening and closing operations are performed on the foreground connected components for noise reduction. The centroid trajectory is calculated for each animal target. With speed .

[0123] (2) Behavioral state machine recognition: Define a set of states {normal, continuous agitation, abnormal aggregation}. When And duration The condition is classified as "persistent agitation"; when the number of targets within the field of vision... And there is an average distance between targets. And duration When this occurs, it is classified as "abnormal clustering." The percentage of abnormal frames is statistically analyzed. = (Number of abnormal frames) / (Total number of frames) is used as the video abnormality score.

[0124] (3) Audio anomaly detection: The audio is segmented according to a fixed window length (preferably 1-2s), and the Mel spectrum / Mel spectrum cepstral coefficients (MFCC), spectral flux, and mutation rate acoustic features are extracted for each segment; the synchronously acquired triaxial acceleration AccX / Y / Z is concatenated with the audio features to form a multimodal feature vector. Load the random forest classifier and output the anomaly label for each segment. Statistics on the percentage of abnormal audio data .

[0125] (4) Fusion output: The video and audio scores are weighted and fused to obtain the biological precursor score. ,in It can be adaptively adjusted based on the confidence level of each modality; when When the threshold is exceeded, a biological warning alarm is triggered.

[0126] (5) Model training (offline): The audio training set includes normal ambient sounds (wind, insect chirping, animal rest) and abnormal sounds (frightened barking, hissing, etc.). Each audio segment is labeled. ;by f A random forest classifier is trained on the input, and a threshold τ is selected on the validation set. On the video side, a rule-based state machine can be used for direct discrimination, or a lightweight classifier can be further trained based on manually labeled "abnormal / normal" segments to optimize the threshold.

[0127] In one optional implementation, the above-mentioned disaster emergency response module is also used to acquire social media data streams related to the earthquake zone after an earthquake occurs, including: text data and video data; For text data, the text data is tokenized to obtain a token sequence; the token sequence is encoded using BERT to obtain the context vector of each token; the sequence label header is used to predict whether each token is an earthquake-related entity based on its context vector, and the tokens in the token sequence that are classified as earthquake-related entities are extracted as target tokens; A classifier is used to calculate the probability that each target token belongs to the low, medium, and high public opinion urgency categories, respectively, and the probabilities are weighted and summed to obtain the public opinion urgency score of the corresponding target token. The sum of the public opinion urgency scores of all target tokens is taken as the total public opinion urgency score for the earthquake area; For video data, the system classifies whether an earthquake occurred in each frame of the video data, and calculates the ratio of the number of frames in the video data where an earthquake occurred to the total number of frames, which is used as the damage score for the earthquake area. Each target token in the involved locations is designated as a disaster-stricken location, and a disaster-stricken point distribution map of the earthquake zone is constructed. The population density at each disaster point is extracted from the population density heat map of the earthquake area, and after normalization, a population density distribution map of the disaster points in the earthquake area is constructed. For each disaster-stricken location, the weighted sum of its population density, total public opinion urgency score, and damage score for the earthquake region is calculated as an indicator of the disaster-stricken location's rescue needs, thereby constructing a heat map of rescue needs for disaster-stricken locations in the earthquake region.

[0128] The above process is the social media multi-modal semantic analysis and disaster situation mapping process. In a specific implementation, for online public opinion information, natural language processing and multi-modal visual recognition technologies are used to extract disaster situation elements, evaluate the urgency, and perform geographical mapping on social media texts / images / videos / audio, forming a disaster situation heat map and differentiated emergency response suggestions.

[0129] (1) Data access and diversion: Real-time access to the social media data stream, including texts, pictures, videos, and audio; the system automatically identifies the file type and assigns a processing pipeline. The audio is first transcribed into text by automatic speech recognition (ASR) and then reused in the text pipeline; the video samples key frames at time intervals to form an image sequence and enter the image pipeline.

[0130] (2) Text entity extraction and urgency scoring: On the text side, jieba is used for word segmentation and a custom dictionary in the earthquake field is loaded; the token sequence is encoded by BERT, and "BERT + sequence labeling (optional CRF)" is used to extract earthquake-related entities such as location, magnitude, casualty / damage description, etc. It is preferred to use the BIO annotation system as the training label (such as B-LOC / I-LOC, B-MAG / I-MAG, B-DMG / I-DMG). For urgency / sentiment, a classification head outputs probabilities, which are used as the public opinion urgency scores . The specific technical implementation is as follows: Input layer: Before the "sentence sequence" enters the BERT model, it is segmented into a token sequence by a tokenizer (such as [CLS], Chao, Yang, Qu, You, Lou, Ta, Le, [SEP]). Each token is converted into a vector and input into the model. Output layer: The model classifies each token in the sequence (sequence labeling task). Label definition: The label is indeed an earthquake-related entity label. For example, the BIO annotation method is adopted: B-LOC (start of location), I-LOC (middle of location), B-EVENT (disaster event), O (non-entity). Example: "Chao (B-LOC) Yang (I-LOC) Qu (I-LOC)" is extracted as a "location entity". Original text → Word segmentation / sub-word tokenization → Obtain token sequence ; BERT outputs the context vector of each token , and the sequence labeling head outputs the BIO label (whether it belongs to an entity such as location / magnitude / casualty / damage, etc.) for each token.

[0131] The classification head generally outputs softmax probabilities. The "urgency score" can be defined as an interpretable and reproducible scalar. For example, assume the urgency is divided into 3 categories: low / medium / high, corresponding to weights w = [0, 0.5, 1]. Classification head output: ; then the public opinion urgency score: .

[0132] (3) For video-based damage identification: the video side infers the damage level frame by frame by key frame and uses time consistency voting to obtain the video damage level. Specifically, temporal consistency voting is used to address the issue of flickering or false detections in single frames during video recognition. Video is a series of images (frames). If the model detects "collapse" in frame 10, doesn't detect it in frame 11 (possibly due to focus blur), and then detects it again in frame 12, directly outputting the result will cause the warning signal to flicker. A time window is set (e.g., T = 1 second, containing 25 frames). The model predicts each of these 25 frames, obtaining 25 results (e.g., 20 frames are "collapse," and 5 frames are "normal"). The frequency of occurrence of the "collapse" category is then calculated. Only when the confidence score is greater than a set threshold is the system confirmed that the video for that second contains a "collapse" event, thus filtering out random noise frames and ensuring the robustness of the output results.

[0133] (4) Geocoding and Rescue Demand Index: Geocoding the extracted location entities to obtain their latitude and longitude coordinates. With population density heatmap Overlay; combined with damage level and urgency Calculate the rescue demand index: ; Outputs a heat map showing the distribution of disaster-stricken areas, rescue needs, and differentiated emergency response recommendations.

[0134] (5) Training set and label supplementation: The text-side training set is a disaster corpus, which includes disaster description text and manually labeled entity / urgency labels; the image-side training set is on-site pictures before / after or during the disaster, which include damage level or segmentation mask labels; the video-side training set is video clips labeled with damage level, and the labels can be obtained by keyframe voting or manual review.

[0135] In one optional implementation, the system further includes: a post-disaster recovery and reconstruction guidance module; wherein the post-disaster recovery and reconstruction guidance module includes: The macroscopic change detection unit is used to acquire real-time post-earthquake satellite imagery of the earthquake area, detect areas in the post-earthquake satellite imagery that have changed compared to the pre-earthquake satellite imagery, and mark them to guide post-earthquake recovery and reconstruction. The microscopic damage detection unit is used to acquire real-time data on each building in the seismic area from drones. Images; a classifier is used to classify the damage level of each building image, thereby obtaining the damage level of each building. Damage level quantification score ; The secondary disaster risk simulation unit is used to calculate the risk of each building in an earthquake zone based on geological environment data and historical earthquake case knowledge, using a Bayesian algorithm. The probability of a secondary disaster occurring at present is considered the secondary disaster risk of the building. ; Planning guidance unit, used to calculate each building Comprehensive priority score ;in, , , , All are preset coefficients; For architecture The preset function importance score; For architecture The estimated repair cost; As constraints, by maximizing ,right The solution is then performed to generate a reconstruction plan; among which, , Indicates no architecture Repair. Indicates the architecture Repairs will be carried out.

[0136] Furthermore, when subsequent remote sensing / on-site inspection data is updated to reflect the damage level or road accessibility, the calculation is recalculated. The plan is updated on a rolling basis to achieve dynamic reconstruction and optimization.

[0137] The above process is the satellite / UAV remote sensing disaster assessment and classification process. In a specific implementation, during the post-disaster recovery and reconstruction phase, the core processing module of the intelligent agent performs macroscopic change detection and microscopic building damage classification on the remote sensing image data, and outputs disaster mask, damage level and repair candidate list.

[0138] (1) Macroscopic change detection (satellite perspective): Input pre-disaster / post-disaster satellite images of the same area Change detection is performed using the Siamese-UNet twin network: a dual-branch encoder shares weights to extract multi-scale features. The decoding end performs differential feature analysis. Upsampling reconstruction is performed to generate a change mask. Ideally, PPM pyramid pooling and ASPP hollow spatial pyramids are introduced into the decoder to achieve multi-scale feature extraction, and the CBAM attention mechanism is used to focus on key damaged regions.

[0139] Training set and labels: Public disaster remote sensing datasets (such as xView2) are preferred. Paired image tiles are used as input, and the labels are pixel-level binary damage masks (0-undamaged / background, 1-damaged), which are manually labeled or generated by rasterizing disaster damage vector data.

[0140] Loss function: To address sample imbalance, a weighted combination of Dice Loss and Focal Loss is preferred, and the OHEM online hard sample mining strategy can be incorporated. Output: Macroscopic disaster damage binary segmentation map, damaged area percentage, and confidence heatmap.

[0141] (2) Microscopic building damage grading (UAV perspective): Input close-up UAV images The UNet++ architecture is used for five-level damage classification of buildings. The ResNeSt backbone network is preferred for extracting texture and edge features. The dense skip connection structure of UNet++ enables multi-scale feature reuse, and the system achieves stable convergence through deep supervision (applying loss to the multi-scale output simultaneously), outputting pixel-level / instance-level segmentation results of buildings.

[0142] Training set and labels: A high-resolution drone aerial photography dataset is used, and the preferred annotation file is labels.json, which includes building boundaries (polygons / boxes) and a five-level damage grading mask. (Intact, Minor, Moderate, Severe, Completely Damaged) Loss function: Considering the ordinal relationship of the damage level (0<1<2<3<4), it is preferable to introduce CORAL ordered regression loss or ordered cross-entropy, and class weight boosting can be applied to high damage levels. Output: building-level classification mask, structural integrity index, and building-level repair candidate list (such as Top-N high-priority repair targets).

[0143] By further combining geological environmental data and historical earthquake case knowledge, the risk of secondary disasters is deduced; based on the above assessment results, a reconstruction plan scheme including suggestions on repair priorities is automatically generated.

[0144] By combining geological environmental data with historical earthquake case knowledge, the risk of secondary disasters can be deduced, and risk zoning and early warning prompts can be generated.

[0145] (1) Secondary Disaster Inference Model: A secondary disaster inference model is constructed based on a Bayesian network framework. Slope, geological type, rainfall intensity, fault zone distance, and river network density are used as evidence nodes, while landslides, debris flows, landslide-dammed lakes, fires, gas leaks, and road collapses are used as risk nodes. Specifically, nodes are defined as follows: physical elements affecting secondary disasters are defined as network nodes. Parent nodes: "predicted magnitude," "epicenter distance," "rainfall (meteorological data)," and "topographic slope (GIS data)" obtained from the preceding modules. Child nodes: "landslide risk," "landslide-dammed lake risk," and "debris flow risk." Edges are defined: causal connections are established based on geological knowledge. For example: Magnitude Landslide; Slope Landslide; Rainfall Debris flow. Define a conditional probability table (CPT): to quantify causal relationships. For example: These probability values ​​can be obtained through statistical learning from historical earthquake case data or preset by expert knowledge. Extrapolation calculation: When the system inputs current magnitude and topographic data, the network automatically updates the posterior probabilities of its child nodes. This embodiment constructs a dynamically driven network, injecting "real-time predicted S-wave magnitude (from CNN)" and "real-time interpreted remote sensing geomorphological features (from Siamese-UNet)" as dynamic evidence into the network in real time, realizing the dynamic extrapolation of secondary disaster risks as the mainshock situation evolves.

[0146] (2) Parameter learning and online inference: The conditional probability table is learned through historical earthquake samples; the current environmental factors are input during the online stage. Calculate the posterior probability It also generates a risk heatmap and zoning results.

[0147] (3) Output: When the posterior probability of a certain type of risk exceeds the threshold, output the corresponding early warning prompt and disposal suggestion, and pass the secondary disaster risk quantification value to the reconstruction priority scoring module.

[0148] Reconstruction plan generation: Based on the disaster damage assessment results and the risk of secondary disasters, and taking into account the importance of building functions and resource constraints, the reconstruction plan automatically generates a reconstruction plan that prioritizes repairs and recommends resource allocation.

[0149] like Figure 3 The diagram shown illustrates the multimodal architecture of the earthquake emergency response and disaster reduction intelligent agent system provided in this embodiment. Based on the above design, this embodiment solves the technical problems of difficulty in deep fusion of multi-source heterogeneous data, low efficiency of cross-modal collaboration, and "fragmented truth" caused by single-view observation in existing earthquake emergency responses. By constructing an integrated three-dimensional observation network of "space-air-ground-human-network", it achieves a paradigm shift from passive response to proactive prevention and control.

[0150] In summary, this embodiment provides a multimodal earthquake emergency response and disaster reduction intelligent system based on virtual multi-viewpoints. Addressing the pain points of strong interference from multi-source heterogeneous data, difficulties in cross-modal feature fusion, and delayed full-cycle decision-making response in earthquake disaster scenarios, it innovatively constructs a multi-dimensional observation system encompassing "sky-space-ground-human-network." By integrating nine modal disaster data, including three-component initial wave, seismic ripple spatiotemporal spectrum, abnormal animal reactions, social media, infrared thermal imaging, magnetic induction signals, Tianyu satellite, UAV remote sensing, and prior knowledge, and constructing a multi-viewpoint cognitive and assessment model, it achieves a cross-scale, in-depth understanding of the disaster situation. The system covers the entire lifecycle of earthquake disaster risk warning, emergency response during the disaster, and post-disaster recovery and reconstruction. Pre-disaster, it achieves dynamic disaster prediction through a spatiotemporal fusion perspective; during the disaster, it utilizes a virtual-real interleaved perspective to complete real-time disaster relief and precise life location; post-disaster, it supports rapid disaster area repair and secondary disaster prevention based on alternating near and far perspectives. It effectively solves the problems of fragmented truth in multimodal data and the challenges of multi-task collaborative generalization. With a lightweight architecture, extreme real-time response capabilities, and dynamic resource optimization mechanisms, it provides scientific, accurate, and efficient intelligent decision support for earthquake emergency response and disaster reduction, significantly improving the predictability of disaster prevention and control and the timeliness of rescue and disposal. While significantly improving the comprehensiveness and accuracy of decision-making information, it can ensure the extreme timeliness of emergency response, achieve efficient allocation of limited rescue resources, and overcome the limitations of traditional emergency systems such as single information and delayed response without deploying a large number of expensive specialized equipment. It provides intelligent core support for scientific emergency command under major natural disasters.

[0151] This embodiment, starting from the theory of multi-agent systems and the entire process of emergency management, constructs an interaction and game framework for various virtual perspectives, thereby establishing a complete decision-making chain covering pre-disaster spatiotemporal early warning, virtual and real rescue during disasters, and post-disaster near- and far-field assessment. It reveals the intrinsic relationship between the complementarity of multimodal data and the robustness of the final decision, and achieves precise design of virtual perspective rules and weight parameters. For example, in disaster situations, the fusion of infrared and magnetic induction enables precise life location under complex rubble, providing a solid model foundation for adaptive decision-making in complex disaster situations.

[0152] When earthquakes occur in complex environments such as areas with communication disruptions and densely collapsed buildings, the disaster evolution path and rescue constraints become more complex. This embodiment can dynamically adjust the decision weights and task priorities of each virtual perspective. Based on this, targeted optimizations are made to the simulation model, such as using magnetic induction penetration communication technology to solve signal obstruction problems. This enables the model to better adapt to the unique challenges of different disaster scenarios and generate more operable and personalized rescue and reconstruction plans.

[0153] The multi-agent system provided in this embodiment achieves closed-loop decision support throughout the entire lifecycle, from pre-disaster risk warning and emergency response during the disaster to post-disaster reconstruction, through a virtual multi-view collaborative perception mechanism. Compared with traditional systems, it improves response speed by 60% and decision accuracy by 45%.

[0154] In addition, this embodiment can also realize the interactive mode of "one-screen overview and one-click linkage", providing emergency command personnel with an integrated tool from situation assessment and resource scheduling to plan generation, which significantly improves the intelligence level of emergency response, decision-making efficiency and ease of operation for non-professionals.

[0155] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for underground magnetic induction positioning, characterized in that, This device is used to locate targets in underground spaces that carry coil devices capable of transmitting and receiving magnetic induction signals. M magnetic induction sensors are deployed at predetermined locations within the underground space; M is a positive integer. Each magnetic induction sensor comprises three mutually orthogonal coils. The coil device carried on the target To the The first of the magnetic induction sensors coils A transmission link is formed between them. ; D; ; D represents the target quantity; The underground magnetic induction positioning method includes: Multiple transmission signals S of different frequencies are transmitted into the underground space, and frequency response signals from each magnetic induction sensor are acquired. Calculate each frequency separately Downlink length ; R is a preset constant; R is a preset impedance; For frequency The generalized transmission coefficient is as follows: ; For frequency The corresponding angular frequency; N is the number of dielectric layers in the near-field region; the near-field region is the distance from the ground surface in the underground space. The area; ; ; and , respectively, are the minimum and maximum frequencies of the emitted signal S; c is the speed of light; and The first Layer and First Interlayer dielectric interface at angular frequency Transmission coefficient and equivalent Fresnel reflection coefficient at the given conditions; For the first Layered dielectric at angular frequency The complex propagation constant under these conditions; For the first The thickness of the medium layer; The frequency extracted from the frequency response signal Downlink The true RSSI value of the received signal from the upper magnetic induction sensor; Calculate transmission link final length ;in, The set of frequencies for transmitted signals; For frequency The corresponding weights; ; length , and The weighted summation result is used as the first The first goal and the first The distance between the magnetic induction sensors is used to achieve positioning.

2. The underground magnetic induction positioning method according to claim 1, characterized in that, In terms of length , and When performing weighted summation, , and The corresponding weights are the normalized results of the link quality factor of the corresponding transmission link; Among them, transmission link The link quality factor is: The equivalent transmission coefficient amplitude, ; For transmission links The shadow fading variance.

3. The underground magnetic induction positioning method according to claim 1 or 2, characterized in that, Determined in the following ways: The frequency to be transmitted into underground space is obtained. The measured value of the transmitted signal S Received voltage and coil The emission current, and the ratio of the two is used as the frequency. Downlink The near-field coupling eigenvalues; Near-field coupling eigenvalues ​​and The product of and is used as frequency Downlink The frequency domain gain is then used to obtain the frequency. Downlink The predicted RSSI value of the received signal from the upper magnetic induction sensor; Will Subtracting the corresponding predicted value from the actual value yields the frequency. Downlink shadow disturbance Then, the corresponding shadow fading weights are calculated. ; The standard deviation of shadow fading; A graph is constructed, in which nodes include anchor nodes and target nodes. Anchor nodes represent magnetic induction sensors in underground space; target nodes represent coil devices carried on targets. Edges in the graph represent signal associations between nodes. The edges between anchor nodes and target nodes in the graph carry shadowing fading weights of each transmission link between the anchor node and the target node at different frequencies, as well as the true RSSI value of the received signal at the anchor node. For each anchor node in the graph, when the deviation of the true RSSI value of its received signal at any frequency from the mean of the true RSSI values ​​of its received signal at the same frequency from the same frequency of its one-hop neighbor anchor node exceeds a preset deviation, the true RSSI value of its received signal at that frequency is updated to the average of the true RSSI values ​​of the received signal at the same frequency of its one-hop neighbor anchor node. The graph data is input into a graph neural network. Through the message passing mechanism of the graph neural network, features from all transmission links are converged and fused according to frequency, resulting in fused features of each transmission link at different frequencies. Among these features, the transmission links... In frequency The fusion features are denoted as ; Calculate the score for each frequency, and then normalize it to obtain the weight for each frequency; where, frequency Corresponding score ; This represents a multilayer perceptron.

4. The underground magnetic induction positioning method according to claim 1 or 2, characterized in that, Also includes: Infrared detection is used in underground spaces to determine the location of targets.

5. An intelligent agent system for earthquake emergency response and disaster reduction, characterized in that, Used for earthquake emergency response in underground spaces within earthquake zones; M magnetic induction sensors are deployed at predetermined locations in the underground space; M is a positive integer; the magnetic induction sensors include three coils that are orthogonal to each other in pairs; The earthquake emergency disaster reduction intelligent system includes: an emergency response module for use after an earthquake, which executes the underground magnetic induction positioning method according to any one of claims 1-4 to locate underground targets.

6. The earthquake emergency disaster reduction intelligent agent system according to claim 5, characterized in that, Also includes: The pre-disaster risk early warning module is used to provide pre-disaster risk warnings before an earthquake occurs, including: The system acquires real-time three-component seismic wave data and corresponding metadata from seismic stations corresponding to the earthquake region. The metadata includes: the address information of the seismic stations and the recorded theoretical arrival time of the P-waves. For each component of the three-component seismic wave data, after preprocessing, the P-wave arrival time of that component is picked using the long-short time window mean ratio method. Furthermore, extracting from this component wave... The subband of the initial preset time period is taken as the corresponding P-band; The metadata and the P-bands corresponding to each component of the three-component seismic wave data are input into a pre-trained deep learning model to obtain the predicted results of earthquake magnitude, focal depth and time difference of arrival of seismic waves, and to provide pre-disaster risk warning.

7. The earthquake emergency disaster reduction intelligent agent system according to claim 6, characterized in that, The aforementioned pre-disaster risk warning before an earthquake also includes: Historical earthquake catalog data for the earthquake region is obtained and preprocessed. The historical earthquake catalog data includes the time, location, focal depth, and magnitude of the earthquake. The preprocessing includes sorting the historical earthquake catalog data in chronological order and interpolating to complete any missing data. The earthquake region is divided into spatial grids, and the data of earthquakes whose locations are in the same spatial grid in the preprocessed historical earthquake catalog data are aggregated to obtain historical earthquake catalog data under different spatial grids; Feature extraction is performed on the historical earthquake catalog data under each spatial grid. After obtaining the corresponding earthquake features, PCA dimensionality reduction is performed to obtain the corresponding candidate features. The earthquake features include: the b-value of the slope used to describe the earthquake magnitude distribution in the Gutenberg-Richter relation, recurrence period features, strain energy release rate, spatial clustering, fault distance, and historical aftershock attenuation. Among them, the recurrence period features are: the mean, variance, and quantile of the interval between adjacent events in the historical earthquake catalog data. The candidate features corresponding to each spatial grid are input into a pre-trained machine learning model to obtain the probability of an earthquake occurring in that spatial grid. Based on the probability of an earthquake occurring in each spatial grid, a heat map and trend curve of earthquake risk for the seismic region are generated.

8. The earthquake emergency disaster reduction intelligent agent system according to claim 6, characterized in that, The aforementioned pre-disaster risk warning before an earthquake also includes: The video anomaly score and audio anomaly score of the earthquake area are weighted and summed to obtain the environmental precursor score; when the environmental precursor score exceeds the preset score, a precursor warning is issued. The video anomaly score for the earthquake region is calculated as follows: Animal targets are detected in each frame of the acquired animal monitoring video stream of the earthquake area, and the centroid trajectory and motion velocity of each animal target at each frame time are calculated. Frames containing moving targets in animal surveillance video streams whose speed exceeds a preset speed and whose duration exceeds a first preset time are designated as abnormal frames. Frames containing animal groups in the animal monitoring video stream that last for more than a second preset time are considered abnormal frames; the number of animal targets in the animal group exceeds a preset number, and the minimum distance between any two animal targets does not exceed a preset distance; The ratio of the number of abnormal frames in the animal monitoring video stream to the total number of frames is calculated as the video anomaly score. The audio anomaly score for the earthquake zone is calculated as follows: A sliding window is used to divide the acquired environmental audio from the earthquake area into multiple environmental sub-audio segments; the audio features of each environmental sub-audio segment are extracted and concatenated with the three-axis acceleration of the synchronously acquired audio acquisition device to form the multimodal feature vector of the corresponding environmental sub-audio segment; Each environmental sub-audio segment is classified as abnormal or non-abnormal based on multimodal feature vectors; The ratio of the number of anomalous environmental sub-audio to the total number of environmental sub-audio is calculated as the audio anomalous score.

9. The earthquake emergency disaster reduction intelligent agent system according to claim 5, characterized in that, It is also used to acquire social media data streams related to the earthquake zone after an earthquake, including text data and video data; For text data, the text data is tokenized to obtain a token sequence; the token sequence is encoded using BERT to obtain the context vector of each token; the sequence label header is used to predict whether each token is an earthquake-related entity based on its context vector, and the tokens in the token sequence that are classified as earthquake-related entities are extracted as target tokens; A classifier is used to calculate the probability that each target token belongs to the low, medium, and high public opinion urgency categories, respectively, and the probabilities are weighted and summed to obtain the public opinion urgency score of the corresponding target token. The sum of the public opinion urgency scores of all target tokens is taken as the total public opinion urgency score for the earthquake area; For video data, the system classifies whether an earthquake occurred in each frame of the video data, and calculates the ratio of the number of frames in the video data where an earthquake occurred to the total number of frames, which is used as the damage score for the earthquake area. Each target token in the involved locations is designated as a disaster-stricken location, and a disaster-stricken point distribution map of the earthquake zone is constructed. The population density at each disaster point is extracted from the population density heat map of the earthquake area, and after normalization, a population density distribution map of the disaster points in the earthquake area is constructed. For each disaster-stricken location, the weighted sum of its population density, total public opinion urgency score, and damage score for the earthquake region is calculated as an indicator of the disaster-stricken location's rescue needs, thereby constructing a heat map of rescue needs for disaster-stricken locations in the earthquake region.

10. The earthquake emergency disaster reduction intelligent agent system according to claim 5, characterized in that, Also includes: A post-disaster recovery and reconstruction guidance module; wherein, the post-disaster recovery and reconstruction guidance module includes: The macroscopic change detection unit is used to acquire real-time post-earthquake satellite imagery of the earthquake area, detect areas in the post-earthquake satellite imagery that have changed compared to the pre-earthquake satellite imagery, and mark them to guide post-earthquake recovery and reconstruction. The microscopic damage detection unit is used to acquire real-time data on each building in the seismic area from drones. Images; a classifier is used to classify the damage level of each building image, thereby obtaining the damage level of each building. Damage level quantification score ; The secondary disaster risk simulation unit is used to calculate the risk of each building in an earthquake zone based on geological environment data and historical earthquake case knowledge, using a Bayesian algorithm. The probability of a secondary disaster occurring at present is considered the secondary disaster risk of the building. ; Planning guidance unit, used to calculate each building Comprehensive priority score ;in, , , , All are preset coefficients; For architecture The preset function importance score; For architecture The estimated repair cost; As constraints, by maximizing ,right The solution is then performed to generate a reconstruction plan; among which, , Indicates no architecture Repair. Indicates the architecture Repairs will be carried out.

Images