A post-disaster aftershock perception and psychological defense line combined guardian system and method

Abstract

The application discloses a post-disaster aftershock sensing and psychological defense line combined guarding system and method, and relates to the technical field of disaster rescue Internet of Things, which comprises the following steps: collecting microseismic time sequence waveform flow at a ruin edge node, performing an adaptive denoising algorithm and a seismic phase picking algorithm on the microseismic time sequence waveform flow, extracting aftershock features, and forming an environmental threat index; capturing face visual flow data of trapped personnel, calling a lightweight convolutional neural network to perform face action unit coding on the face visual flow data, inferring multidimensional emotion probability of the trapped personnel, and mapping the multidimensional emotion probability into a core psychological stress state vector through feature dimension reduction; performing feature space fusion and probability renormalization processing on the environmental threat index and the core psychological stress state vector, and constructing a trapped person comprehensive situation assessment model. The application can improve the accuracy and timeliness of aftershock feature extraction by performing adaptive denoising and seismic phase picking on the microseismic time sequence waveform flow at the ruin edge node.

Description

Technical Field

[0001] This invention relates to the field of Internet of Things (IoT) technology for disaster relief, and in particular to a system and method for jointly protecting post-disaster aftershocks and psychological defenses. Background Technology

[0002] In post-disaster emergency rescue, it is usually necessary to obtain information on aftershock environment, status of trapped personnel, and on-site intervention feedback. To improve the information acquisition capabilities at disaster sites, relevant rescue systems will deploy microseismic detection equipment, video monitoring equipment, and edge nodes to collect data such as microseismic waveforms, images, or videos. These data will be combined with perception models, assessment models, and linkage devices to complete environmental monitoring, status assessment, and intervention response. With the development of edge computing, IoT communication, and intelligent identification technologies, applying multi-source perception, fusion analysis, and intervention linkage to post-disaster rescue has become an important development direction for related technologies.

[0003] However, existing methods have some shortcomings when dealing with extreme conditions such as low light, strong noise, and complex media in ruins. There is insufficient correlation between different sensing links, and it is difficult to form a continuous linkage between environmental monitoring, psychological recognition, and intervention response. The original waveform and visual data are mostly processed in parallel and independently. When the computing power of edge devices is limited, it is easy to cause high computing power consumption, response delay, and difficulty in aligning heterogeneous time-series data. At the same time, there is a lack of effective fusion evaluation, anti-shake correction, and error suppression mechanisms. Single-mode errors are easily amplified by subsequent processing, affecting the rescue judgment and intervention timeliness. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a method for jointly protecting the psychological defense line and sensing after a disaster to solve the problems of insufficient correlation in the sensing process, heavy data processing burden, and easy amplification of errors.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for jointly protecting the psychological defense line and sensing after a disaster, which includes: collecting microseismic time-series waveform streams at the edge nodes of the ruins, and performing an adaptive denoising algorithm and a phase picking algorithm on the microseismic time-series waveform streams to extract aftershock features and form environmental threat indicators; capturing facial visual stream data of trapped personnel, calling a lightweight convolutional neural network to encode facial action units on the facial visual stream data, inferring the multidimensional emotional probabilities of the trapped personnel, and mapping them into a core psychological stress state vector through feature dimensionality reduction; performing feature space fusion and probability renormalization processing on the environmental threat indicators and the core psychological stress state vector, and constructing a comprehensive situation assessment model for the trapped personnel; monitoring the stress extremes and physical offsets in the comprehensive situation assessment model for the trapped personnel, and triggering a graded intervention mechanism when a specific component of the core psychological stress state vector exceeds a set threshold, generating execution instructions, and feeding them back to the edge nodes of the ruins.

[0007] As a preferred embodiment of the post-disaster aftershock perception and psychological defense protection method described in this invention, the specific steps for forming environmental threat indicators are as follows: An adaptive wavelet soft thresholding denoising process based on the Lipschitz singularity index is performed on the microseismic time-series waveform stream to obtain a clean denoised signal. The STA / LTA ratio algorithm is used to calculate the feature ratio of the clean, denoised signal; When the characteristic ratio exceeds the preset dynamic threshold, the initial arrival of the aftershock P wave is locked, forming an environmental threat indicator.

[0008] As a preferred embodiment of the post-disaster aftershock perception and psychological defense method described in this invention, the formation of environmental threat indicators further includes: after the feature ratio exceeds a preset dynamic threshold, embedding the Akaike Information Criterion (AIC) based on an autoregressive model within the trigger interval, calculating the logarithmic weighted sum of the variances of the two preceding and following sequences, and accurately locking the absolute first arrival time of the aftershock P-wave by finding the global minimum point.

[0009] As a preferred embodiment of the post-disaster aftershock perception and psychological defense protection method described in this invention, the specific steps of mapping the core psychological stress state vector through feature dimensionality reduction are as follows: Regions of interest are extracted using a Haar feature cascade classifier; If the size of the region of interest is not standard, an interpolation algorithm is called to reshape the region of interest into a single-channel grayscale tensor; If region of interest extraction fails, an exception capture is triggered and the current frame is silently skipped. Facial visual stream data is input into the improved Mini-Xception network, which outputs the original 7-dimensional emotion probability vector. The original 7-dimensional emotion probability vector is transformed into a core psychological stress state vector through feature dimensionality reduction mapping.

[0010] As a preferred embodiment of the post-disaster aftershock perception and psychological defense protection method described in this invention, the core psychological stress state vector refers to the state vector obtained by transforming the original 7-dimensional emotional probability vector through feature dimensionality reduction mapping. The state vector includes calm state components, anxiety state components, panic state components, and despair state components; The calm state component is the sum of the probability of happiness, the probability of neutrality, and the probability of surprise; The anxiety state component is the sum of the probability of anger and the probability of aversion; The panic state component is the absolute value of the fear probability; The despair state component is the absolute value of the probability of sadness.

[0011] As a preferred embodiment of the post-disaster aftershock perception and psychological defense protection method described in this invention, the specific steps for performing feature space fusion and probability renormalization processing on environmental threat indicators and core psychological stress state vectors, and constructing a comprehensive situation assessment model for trapped individuals, are as follows: Feature space fusion is performed on environmental threat indicators and core psychological stress state vectors to obtain feature space fusion results; The core psychological stress state vector is subjected to probability renormalization to generate a standard confidence matrix; A comprehensive situation assessment model for trapped individuals is constructed based on the feature space fusion results and the standard confidence matrix.

[0012] As a preferred embodiment of the post-disaster aftershock perception and psychological defense method described in this invention, the probability renormalization process refers to constructing a normalization ratio by summing the unnormalized component of any one dimension with the sum of all state components and the zero-prevention division bias constant, and using the normalization ratio as the final normalized confidence level of the current dimension.

[0013] As a preferred embodiment of the post-disaster aftershock perception and psychological defense protection method described in this invention, the step of triggering a graded intervention mechanism when a specific component of the core psychological stress state vector exceeds a set threshold, generating an execution command, and feeding it back to the edge node of the ruins, includes the following specific steps: Use the high-level panic threshold and the medium-level despair threshold as trigger thresholds; When the panic component exceeds the high-level panic threshold, the highest level of intervention is triggered, generating an acoustic reassurance instruction and raising the emergency priority to the highest level. When the despair state component exceeds the intermediate-level despair threshold, an intermediate-level intervention is triggered, generating a pulse tapping execution command; The frequency of the pulse tapping command can be dynamically adjusted based on the real-time feedback from the trapped person. When all state components are less than or equal to their corresponding trigger thresholds, a stable state is marked, and the current high-frequency polling monitoring state is maintained.

[0014] As a preferred embodiment of the post-disaster aftershock perception and psychological defense protection method described in this invention, the final normalized confidence of the current dimension is used to generate a standard confidence matrix that fits the radar chart rendering architecture.

[0015] Secondly, this invention provides a post-disaster aftershock perception and psychological defense joint protection system, comprising: a physical environment perception module, a visual emotion capture module, a multimodal fusion decision module, and an intelligent intervention execution module; the physical environment perception module is used to collect microseismic time-series waveform streams at the edge nodes of the ruins, and to perform adaptive denoising algorithms and phase picking algorithms on the microseismic time-series waveform streams to extract aftershock features and form environmental threat indicators; the visual emotion capture module is used to capture facial visual stream data of trapped personnel, to call a lightweight convolutional neural network to encode facial action units on the facial visual stream data, to infer the multidimensional emotional probabilities of trapped personnel, and to map them into a core psychological stress state vector through feature dimensionality reduction; the multimodal fusion decision module is used to perform feature space fusion and probability renormalization processing on the environmental threat indicators and the core psychological stress state vector, and to construct a comprehensive situation assessment model for trapped personnel; the intelligent intervention execution module is used to monitor the stress extremes and physical offsets in the comprehensive situation assessment model for trapped personnel, and to trigger a graded intervention mechanism when a specific component of the core psychological stress state vector exceeds a set threshold, to generate execution instructions, and to feed them back to the edge nodes of the ruins.

[0016] The beneficial effects of this invention are as follows: By adaptively denoising and picking up seismic phases on the microseismic time-series waveform stream at the edge nodes of the rubble, the accuracy and timeliness of aftershock feature extraction can be improved; by performing lightweight convolutional neural network inference and feature dimensionality reduction mapping on the facial visual stream data of trapped personnel, a precise conversion from redundant general expressions to core psychological stress states can be achieved; while ensuring recognition robustness, the computing power performance at the edge is optimized, reducing the overall power consumption and processing latency of the system; at the same time, by performing feature space fusion and probability renormalization processing on environmental threat indicators and core psychological stress state vectors, joint monitoring and comprehensive evaluation of the physical environment and psychological defenses can be achieved; by triggering graded intervention when specific components exceed the threshold and feeding back execution instructions to the edge nodes, a closed-loop protection link of perception, decision-making, intervention and feedback can be formed, thereby improving the efficiency of rescue decision-making and the survival probability of trapped personnel in complex rubble scenarios. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A flowchart for a method of jointly protecting psychological defenses and sensing aftershocks after a disaster.

[0019] Figure 2 This is a schematic diagram of a joint protection system for aftershock perception and psychological defense.

[0020] Figure 3 A flowchart for forming environmental threat indicators and core psychological stress state vectors.

[0021] Figure 4 A flowchart for comprehensive situation assessment and tiered intervention feedback for those in distress. Detailed Implementation

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0025] Reference Figures 1-4 This is one embodiment of the present invention, which provides a method for jointly protecting the psychological defense line and sensing aftershocks, including the following steps: S1. Collect microseismic time-series waveform streams at the edge nodes of the ruins, and perform adaptive denoising and phase picking algorithms on the microseismic time-series waveform streams to extract aftershock characteristics and form environmental threat indicators.

[0026] It should be noted that the edge nodes of the ruins are integrated with multimodal sensing modules; the multimodal sensing modules include, but are not limited to: cameras for capturing the visual flow of the trapped person's face, microelectromechanical microphones for collecting the voice flow of distress calls, and digital accelerometers for monitoring micro-vibrations of the physical environment.

[0027] It should also be noted that the rubble edge nodes in this embodiment are deployed at the forefront of the rubble and can be heterogeneous computing boards equipped with NVIDIA Jetson architecture or ESP32 dual-core microcontrollers. Their function is to isolate and lock the first arrival signal of aftershock P waves in complex mechanical rescue noise environments. This embodiment achieves algorithm and hardware matching through a heterogeneous computing power distribution mechanism. The high computing power branch is used for subsequent computationally intensive tasks such as visual recognition and multimodal analysis to ensure the real-time response of complex models on site. The low-power high-frequency acquisition branch is used for recursive STA / LTA phase picking, signal noise reduction based on the Lipschitz singularity index, and driving of the pulse knocking and audio broadcasting execution units.

[0028] To meet the edge computing needs of extreme disaster sites, the edge nodes of the ruins adopt a heterogeneous computing collaborative architecture. The core computing platform can use an edge AI accelerator card or a low-power dual-core microcontroller. A low-noise-density triaxial digital accelerometer is directly connected via a built-in integrated circuit bus to acquire microseismic time-series waveform streams. The acquired microseismic time-series waveform streams are stored in a circular memory pool within the local microcontroller through a direct memory access mechanism to reduce CPU interrupt overhead and improve high-frequency acquisition stability. In addition to containing environmental aftershock signals, the microseismic time-series waveform streams can also be used as auxiliary outputs through a parallel filter bank. By using bandpass filtering with a specific frequency bandwidth, physiological micro-vibration time-series vectors can be extracted to characterize the cardiopulmonary activity characteristics of trapped individuals and for subsequent optional multimodal fusion analysis.

[0029] Furthermore, an adaptive denoising algorithm is applied to the microseismic time-series waveform stream, employing an adaptive wavelet nonlinear improved threshold shrinkage algorithm based on the Lipschitz singularity index. Specifically, the acquired microseismic time-series waveform stream is decomposed using a Mallat pyramid decomposition through an orthogonal mirror filter bank, down to the fifth level scale, to obtain approximation coefficients and detail coefficients. Daubechies4, which combines tight support and smoothness, is selected as the mother wavelet basis function. During the decomposition process, the approximation coefficients and detail coefficients of the J-th level are obtained by downsampling after discrete convolution of the approximation coefficients of the J-1-th level with the low-pass filter and the high-pass filter, respectively, to remove high-frequency mechanical white noise and extremely low-frequency sensor thermal drift baseline.

[0030] Abandoning the globally fixed threshold approach, an adaptive threshold dynamic calculation criterion is adopted: the absolute median difference of the high-frequency detail coefficients of the current layer is extracted as a robust estimation benchmark for the noise variance of the J-th layer, and the dynamic contraction threshold at each scale is determined by combining the empirical constant of the standard normal distribution and the signal length; at the same time, the prior signal-to-noise ratio of the disaster site is loaded into the smoothing adjustment factor; the prior signal-to-noise ratio is determined by the environmental silent window adaptive evaluation method, that is, in the background state without triggering phase picking, the root mean square power of the micro-seismic waveform within the sliding time window is extracted as the environmental mechanical noise benchmark variance, and the total signal variance of the current time window is calculated when capturing waveform anomalies, so as to estimate the prior signal-to-noise ratio in real time; the smoothing adjustment factor comes from the nonlinear wavelet contraction theory, and the value range is set to (0, 1] for example, and has an inverse proportional mapping relationship with the prior signal-to-noise ratio: when the prior signal-to-noise ratio is low, the value approaches 1 to enhance the mechanical noise suppression capability; when the prior signal-to-noise ratio is high, the value approaches 0 to retain the transient singularity characteristics of the initial arrival of the aftershock P wave.

[0031] For detail coefficients greater than the dynamic shrinkage threshold, exponential smoothing shrinkage is performed, while for detail coefficients less than the dynamic shrinkage threshold, they are directly set to zero. The system is then reconstructed layer by layer using the inverse discrete wavelet transform operator to output a pure microseismic time-series waveform stream, i.e., a pure denoised signal, thereby eliminating high-frequency impulse noise and zero-point baseline drift.

[0032] Furthermore, a phase-picking algorithm is applied to the clean, denoised signal, employing an improved short-window and long-window average energy ratio algorithm combined with the Akaike information criterion for joint triggering. Specifically, an energy characteristic function of the signal within the sliding time window is defined, which is the weighted sum of the absolute value of the signal amplitude and the absolute value of the first-order difference. The arithmetic mean of the energy characteristic function within the short-window and the arithmetic mean of the energy characteristic function within the long-window are used to construct the characteristic ratio, expressed as follows: ; ; ; in, Indicates the first The characteristic ratio at each discrete sampling time is used to characterize the degree of energy abrupt change; Indicates the first The average energy value of a short time window at each discrete sampling time reflects the instantaneous energy change during aftershocks. Indicates the first The long-term average energy value of discrete sampling times reflects the long-term stable energy of environmental background noise; Indicates the index of the discrete sampling time; This indicates the length of the short-time sliding window, i.e., the number of sampling points it contains; Indicates the first The characteristic function value of each sampling point is, in this embodiment, a weighted sum of the absolute value of the signal amplitude and the absolute value of the first-order difference; Indicates the length of a long-term sliding window; Represents the iteration index for the summation operation.

[0033] Short time windows are used to reflect the instantaneous energy of sudden aftershocks, while long time windows are used to reflect the long-term stable energy of background noise.

[0034] To meet the requirements of extremely low power consumption and ultra-fast inference for edge devices, this embodiment optimizes the traditional integral operation into a recursive difference equation, expressed as follows: ; ; in, Indicates the first The average energy value of a short time window at discrete sampling times; Indicates the first The average energy value over a long time window at discrete sampling times; This represents the historical characteristic function value that slips out of the short-time window boundary; This represents the historical characteristic function value that slips out of the long time window boundary.

[0035] The average energy of the short-term window at the current discrete moment is recursively updated by the average energy of the short-term window at the previous moment, the weighted value of the characteristic function at the current moment, and the weighted value of the historical characteristic function that has slid out of the boundary of the short-term window. Similarly, the average energy of the long-term window is recursively updated by the same method, using the average energy of the long-term window at the previous moment, the weighted value of the characteristic function at the current moment, and the weighted value of the historical characteristic function that has slid out of the boundary of the long-term window. This avoids the edge nodes of the ruins repeatedly summing the long-term window data in each sampling period, thus releasing the computing load of the edge processor.

[0036] Because the rubble edge nodes employ real-time high-frequency acquisition, multiple feature ratios are arranged chronologically to form a feature ratio sequence. When the feature ratio sequence exceeds the adaptively generated dynamic threshold upper limit in time, an aftershock event is preliminarily determined. The dynamic threshold upper limit is determined during the environmental calibration phase after the rubble edge nodes are activated. Specifically, the physical environment sensing module extracts the statistical characteristics of the feature ratio sequence during the silent period, calculates the mean and standard deviation, and establishes the dynamic threshold upper limit based on the mean and standard deviation. The expression is as follows: ; in, Indicates the upper limit of the dynamic threshold; Represents the mean of the characteristic ratio sequence; The standard deviation of the characteristic ratio sequence; This represents the sensitivity gain coefficient, which is used to adjust the sensitivity of the upper limit of the dynamic threshold to the fluctuation amplitude of the feature ratio sequence. The value is set based on the statistical stability of the feature ratio sequence during the environmental quiet period and the balance between the false negative rate and the false positive rate of aftershock event detection. An exemplary value range is 1.5 to 3.0, and 2.0 is preferred.

[0037] An Akaike Information Criterion Executor based on an autoregressive model is embedded within the trigger interval to traverse the potential real waveform arrival point. By obtaining the global minimum point of the variogram sequence, the absolute first arrival time of the aftershock P-wave is accurately locked, and false positives caused by continuous and chaotic environmental knocking sounds are suppressed. For example, the picking error can be compressed to within fifteen milliseconds.

[0038] Through the above adaptive denoising and phase picking processing of the microseismic time-series waveform stream, aftershock characteristics can be extracted from the complex mechanical noise environment of the ruins, and environmental threat indicators can be formed for subsequent joint evaluation with the core psychological stress state vector.

[0039] S2. Capture the facial visual stream data of the trapped person, call a lightweight convolutional neural network to encode facial action units on the facial visual stream data, infer the multidimensional emotional probability of the trapped person, and map it into the core psychological stress state vector through feature dimensionality reduction.

[0040] It should be noted that in the visual branch, the original image signal is acquired through the visual sensor hardware interface integrated on the edge node of the ruins, and the facial visual stream data is continuously captured by combining the underlying video capture handle; a forced dimension alignment operator is applied to the facial visual stream data, that is, the region of interest (ROI) is extracted through the Haar feature cascade classifier; if the size of the ROI is non-standard, the interpolation algorithm is called to reshape it into a 48×48×1 single-channel grayscale tensor; if the ROI extraction fails, abnormal capture is triggered and the current frame is silently skipped to avoid disrupting the continuous decision chain.

[0041] The value of 48×48×1 is based on the fact that the input tensor volume is compressed while preserving the spatial topology of the facial action unit, and is used as the standard input size for the subsequent Mini-Xception network. Using a single-channel grayscale mode can also remove the environmental color redundancy that is not related to emotion recognition, and reduce the memory usage and initial convolution computation load.

[0042] A lightweight convolutional neural network is used to encode facial action units on facial visual stream data. The Mini-Xception network, which is commonly used in this field, is used as the backbone structure to output the original 7-dimensional emotion probability vector. The original 7-dimensional emotion probability vector includes happiness, neutrality, surprise, anger, disgust, fear, and sadness.

[0043] As a parallel auxiliary branch complementing facial visual flow recognition, this embodiment can synchronously load an accompanying acoustic signal processing branch on the basis of the visual channel to provide acoustic modal feature compensation when the visual signal fails due to low light or occlusion. Specifically, considering that the cries for help or rapid breathing of trapped personnel have high-frequency resonant nonlinear distortion after propagating through thick concrete media, the acquired accompanying acoustic signal is pre-emphasized through a first-order high-pass digital filter, and then the non-stationary continuous speech signal is framed and a Hamming window function is applied to smooth the frame boundaries and reduce spectral energy leakage. The frame length and step size of the frame processing are dynamically determined according to the balance criterion of physical inertia of the vocal organs and time-frequency resolution. For example, the frame length is preferably 25 milliseconds and the step size is preferably 10 milliseconds, so that the inter-frame overlap rate reaches 60%, thereby ensuring the smooth transition and continuous extraction of the dynamic features of the post-disaster rescue voice on the time axis.

[0044] A radix-2 Fast Fourier Transform is performed on the accompanying acoustic signal of each windowed frame to obtain the discrete power spectrum energy. This discrete power spectrum energy is then input into a set of triangular bandpass filters that are equidistantly distributed on the Mel frequency scale and logarithmically expanded at the physical Hertz frequency. By taking the natural logarithm of the filter output energy, the multiplicative convolution interference in the frequency domain is converted into an additive component. A Discrete Cosine Transform is then performed on the logarithmic energy to extract the first thirteen static Mel frequency cepstral coefficients. The first and second difference coefficients of the first thirteen static Mel frequency cepstral coefficients are further calculated, and the first and second difference coefficients are concatenated and spliced ​​to form a thirty-nine-dimensional emotion acoustic spatiotemporal feature tensor.

[0045] As an alternative implementation method parallel to visual single-modal recognition, or as an enhancement scheme when visual features are impaired, another implementation method constructs a lightweight multimodal fusion network (1D-CNN+Attention) that combines a one-dimensional convolutional network with a scaled dot product attention mechanism. The lightweight multimodal fusion network enables the network to process multi-source temporal features in real time at the edge by combining standard convolutional layers and scaled dot product attention mechanisms for heterogeneous temporal feature processing and lightweight topology reconstruction. Considering that the robustness of single audiovisual modality recognition is limited in strong dust or extremely dark environments, this step uses the physiological micro-vibration temporal vector as a high-confidence physical feature for stress state discrimination. The high-confidence physical feature and the emotional acoustic spatiotemporal feature tensor are concatenated and aligned in the feature channel dimension and then input into a multi-layer cascaded one-dimensional convolutional neural network to further infer the multidimensional emotional probability of the trapped person.

[0046] Following the high-level semantic feature extraction layer of the one-dimensional convolutional neural network, a scaled dot product self-attention mechanism is loaded. This mechanism maps the high-level semantic feature vector into a query matrix, a key matrix, and a value matrix using three sets of parallel learnable linear mapping weight matrices. This projects the original fused features to different feature subspaces and captures deep associations. The inner product operation is performed using the transpose of the query matrix and the key matrix. The ratio of the inner product result to the square root of the key matrix dimension is then used to obtain a probability distribution with a sum of one through the Softmax function. This probability distribution is then used as the product of the dynamic weight matrix and the value matrix to enhance high signal-to-noise ratio modal features and suppress contaminated modal features. The fused high-dimensional tensor is flattened and input into a fully connected classifier, outputting the original 7-dimensional sentiment probability vector.

[0047] Furthermore, the original 7-dimensional emotion probability vector is transformed into a core psychological stress state vector through a feature dimensionality reduction mapping matrix to remove non-specific components that do not have the significance of triggering graded interventions at disaster sites. The original 7-dimensional emotion probability vector includes a happiness probability component, a neutral probability component, a surprise probability component, an anger probability component, an aversion probability component, a fear probability component, and a sadness probability component. The calm state component is the sum of the happiness, neutral, and surprise probability components. The anxiety state component is the sum of the anger and aversion probability components. The panic state component is the absolute value of the fear probability component. The despair state component is the absolute value of the sadness probability component.

[0048] As an optional performance optimization solution, this embodiment can also configure a dual-modal hot-switching control unit in the visual emotion capture module; the "dual-modal" in the dual-modal hot-switching control unit refers to the real-time live streaming mode and the file test mode of facial visual data; the dual-modal hot-switching control unit is used to dynamically schedule the input method of facial visual stream data according to the reliability of the communication environment and the occupancy status of the underlying inference resources; when in the real-time live streaming mode, the underlying VideoCapture object is locked through the asynchronous thread pool, and the computing power is occupied all time to push the streaming media MJPEG and unpack the real-time neural network to continuously acquire facial visual stream data and perform continuous emotion recognition; when in the file test mode, the underlying camera handle is forcibly suspended and released, the offline file inference pipeline is activated, the recognition inference and anchor box redrawing are performed on the incoming local static image or pre-stored video clip, and the result features are updated to the front-end state machine to realize the offline training closed loop; regardless of whether in the real-time live streaming mode or the file test mode, the obtained recognition result is uniformly output as the original 7-dimensional emotion probability vector, and the core psychological stress state vector is formed by feature dimensionality reduction mapping.

[0049] The dual-modal hot-switching control unit ensures that the visual emotion capture module can still obtain facial image inputs that can be used for emotion recognition when communication conditions are unstable at the disaster site, cameras are abnormally occupied, or offline drills and verifications are required. This avoids the failure to generate the core psychological stress state vector due to the interruption of the real-time video stream. At the same time, it enables the real-time live streaming mode and the file test mode to share the same recognition reasoning process and the same output format, thereby improving the visual recognition continuity and deployment adaptability of the ruin edge nodes under weak communication, strong interference, and drill and test scenarios.

[0050] S3. Perform feature space fusion and probability renormalization on environmental threat indicators and core psychological stress state vectors, and construct a comprehensive situation assessment model for trapped individuals.

[0051] It should be noted that this step in this embodiment is executed by the multimodal fusion decision module, which is deployed in a cloud-based high-concurrency decision cabin. This module is responsible for the joint analysis and global situation assessment of environmental threat indicators and core psychological stress state vectors. It synchronously transmits the environmental threat indicators and core psychological stress state vectors via a full-duplex communication link and performs feature space fusion on both. Specifically, it spatially concatenates the environmental threat indicators and core psychological stress state vectors in the feature dimension to construct an original heterogeneous fusion long vector. Then, it performs a linear transformation on the original heterogeneous fusion long vector using a preset modal weight mapping matrix, mapping physical and psychological indicators of different dimensions to the same high-dimensional linear space, thus obtaining the multimodal joint situation vector, i.e., the feature space fusion result.

[0052] The modal weight mapping matrix is ​​a deterministic weight matrix pre-written into the multimodal fusion decision module. In the preset, the environmental threat index, panic state component, despair state component, anxiety state component, and calmness loss quantity are used as five numerical terms to participate in the modal weight mapping. Among them, the environmental threat index comes from the aftershock feature extraction results of the microseismic time series waveform stream, the panic state component, despair state component, and anxiety state component come from the core psychological stress state vector, and the calmness loss quantity is a difference from the calm state component, which is used to characterize the degree of decline in the stability state of the trapped person.

[0053] When presetting the weights of the modal weight mapping matrix, the weight levels are determined according to the correspondence between the five numerical items and the intervention triggers: the environmental threat indicator is used to characterize the risk of aftershocks or secondary collapse; the panic state component and the despair state component correspond to the trigger conditions for the highest level intervention and the intermediate level intervention, respectively. All three can directly cause the warning to be upgraded or the intervention command to be issued, and are configured as the first level weight; the anxiety state component is used to characterize the development trend of the psychological stress state, but does not trigger the highest level intervention on its own, and is configured as the second level weight; the calmness loss quantity is used to help judge the degree of decline in the stable state, and is not used as a separate intervention trigger condition, and is configured as the third level weight.

[0054] The first-level weights, second-level weights, and third-level weights follow a decreasing order, meaning the first-level weight is greater than the second-level weight, and the second-level weight is greater than the third-level weight. For example, an equally spaced sequential assignment method can be used, assigning an initial weight of 1 to the third-level weight, 2 to the second-level weight, and 3 to the first-level weight. Each initial weight is then normalized, and the normalized weight parameters are written into the modal weight mapping matrix. The normalization process uses a weight sum normalization method to determine the environmental threat index and the state of panic, respectively. The initial weights corresponding to the components of despair, anxiety, and calmness are calculated, and the sum of all initial weights is calculated. The ratio of each initial weight to the sum of all initial weights is calculated to obtain the corresponding normalized weight, so that the sum of all normalized weights is 1. An equal-interval sequential weighting method is adopted, which can maintain the difference in risk levels while avoiding the amplification of the highest level weight, which would cause other components to lose their role in the evaluation. In other implementations, other positive weight combinations that satisfy the first level weight being greater than the second level weight and the second level weight being greater than the third level weight can also be used.

[0055] To address the issues of inconsistent dimensions and semantic redundancy in the expression of emotional states in multi-source heterogeneous data, this paper proposes a probability renormalization process on the core psychological stress state vector. This ensures that the core psychological stress state vector achieves a unified scale expression before entering the subsequent comprehensive situation assessment model for trapped individuals, and improves the comparability and consistency of visualization between different state components. Specifically, the arithmetic sum of each state component in the core psychological stress state vector is calculated. Then, a minimum value is set as a zero-prevention bias constant and summed with the arithmetic sum to construct a normalized denominator. The ratio of the original unnormalized component to the normalized denominator for each dimension is calculated to obtain the normalized confidence score for each dimension. Finally, the normalized confidence scores for each dimension are used to construct a standard confidence matrix.

[0056] Based on the feature space fusion results and the standard confidence matrix, a comprehensive situation assessment model for trapped individuals is constructed, and subsequent situation analysis is supported in the cloud-based high-concurrency decision-making cabin. To support hundreds of edge nodes in the ruins to push heterogeneous data at high frequency, the cloud-based high-concurrency decision-making cabin adopts an asynchronous non-blocking coroutine event pool architecture. After receiving the lightweight feature vector, the backend scheduling center completes message decoding, logical verification, and local persistent storage to reduce the impact of network backpressure and data backlog.

[0057] Specifically, the lightweight feature vector refers to the composite data packet composed of environmental threat indicators and the core psychological stress state vector after probability renormalization. In order to adapt to the weak communication environment that may exist in the disaster area, the edge nodes of the ruins do not transmit the original image stream or microseismic waveform stream to the multimodal fusion decision module. Instead, they encapsulate the environmental threat indicators calculated by the edge side and the core psychological stress state vector after probability renormalization in the form of low-dimensional vectors.

[0058] The front-end presentation layer receives the feature space fusion results in real time and calls the hardware acceleration method based on the underlying graphics library to map the multi-dimensional indicators to the polar coordinate system to draw a polygonal emotion radar map. This allows rescuers to quickly assess the comprehensive survival risk of trapped individuals by analyzing the size of the graphic area and the shift in the center of gravity, enabling a visual comparison between the priorities of psychological intervention and engineering rescue. At the same time, through rendering thinning and smoothing degradation mechanisms, the clock frame flushing frequency is dynamically adjusted according to the terminal hardware load, prioritizing the continuous display of core situational data and the stability of interactive responses during concurrent rendering across multiple nodes.

[0059] In addition, this embodiment also includes extended processing of anti-shaking mechanism and coordinate targeted compensation mechanism; in order to prevent the edge nodes of the ruins from triggering early warning storms due to high-frequency environmental noise, the back-end scheduling center in the multimodal fusion decision module is equipped with a timestamp throttling valve and records the trigger timestamps of various alarms in real time. By comparing the time difference between the current alarm time point to be triggered and the previous alarm time point, it is determined whether the preset anti-shaking time window has been reached; if the anti-shaking time window has not been reached, the high-frequency repeated alarms are regarded as redundant noise generated by environmental jitter and are silently discarded or only injected into the log bus; if the anti-shaking time window has been reached, the current alarm is allowed to enter the hierarchical intervention judgment process.

[0060] To address displacement drift caused by factors such as physical aftershocks and crosswinds at the edge of drones or ruins, a coordinate-targeted compensation mechanism is set up in the backend dispatch center. The calibrated absolute coordinates of the target are obtained by performing matrix addition on the original relative coordinates and the manual intervention compensation parameter matrix. The original relative coordinates are derived from the pixel center coordinates of the region of interest on the trapped person's face in the original image frame, and the manual intervention compensation parameter matrix is ​​derived from the manual correction instructions received by the rescuers from the front-end display layer. The range of the compensation parameters is set according to the resolution of the video acquisition equipment and the estimated physical displacement deviation. For example, the compensation range can be set between -0.5 and +0.5 in the normalized coordinate system.

[0061] In this embodiment, the comprehensive situation assessment model for trapped individuals adopts a multivariate decision matrix structure based on weight mapping. The comprehensive situation assessment model for trapped individuals does not require independent deep learning training. Instead, it calls the normalized weight parameters in the modal weight mapping matrix to perform deterministic weighted reasoning on the fusion features corresponding to environmental threat indicators and core psychological stress state vectors, so as to ensure the interpretability of rescue decisions and the safety of command issuance in extreme disaster scenarios.

[0062] Specifically, the construction and implementation of the comprehensive situation assessment model for trapped individuals is divided into the following two stages: Phase 1: Weighted Score Accumulation; The comprehensive situation assessment model for trapped individuals uses the main indicators from the feature space fusion results as input variables. The main indicators include at least the intensity of aftershock threat and the trapped individuals' panic confidence. In the extended implementation, the cardiopulmonary physiological characteristic offset may also be included. When constructing the comprehensive situation assessment model for trapped individuals, specific risk impact factor weights are preset for each variable. The risk impact factor weights are set in a grade according to the role of each variable in the comprehensive survival risk assessment. Among them, variables that directly represent the degree of life danger, structural instability risk, or the extreme stress state of trapped individuals are assigned the first weight level, and variables that represent the trend of state change, auxiliary physiological characteristics, or supplementary environmental information are assigned the second weight level. The weights corresponding to each variable are called from the normalized weight parameters in the modal weight mapping matrix, or configured according to the same decreasing rule as the weight level. After being normalized by the sum of weights, they are written into the comprehensive situation assessment model for trapped individuals. By multiplying and accumulating the scores of each indicator with their corresponding weights, the comprehensive situation score representing the current overall survival pressure and stress level of the trapped individuals is calculated.

[0063] Phase Two: Tiered Assessment and Decision Mapping; The comprehensive situation assessment model for trapped individuals pre-defines a tiered assessment and decision space, dividing the comprehensive situation score into four logical intervals from low to high, each corresponding to different rescue support actions: First, when the overall situation score is in the low range (e.g., 0 to 0.25), the overall situation assessment model of the trapped person determines it as a routine monitoring situation, maintains low-frequency perception and recognition, and automatically activates the psychological comfort audio playback function at the edge.

[0064] Second, when the overall situation score is in the low to medium range (e.g., 0.25 to 0.5), the overall situation assessment model of the trapped person determines it as a situation of key concern, instructs the edge nodes to increase the sensor sampling frequency, and establishes a full-duplex voice reassurance channel.

[0065] Third, when the comprehensive situation score is in the medium to high range (e.g., 0.5 to 0.75), the comprehensive situation assessment model of the trapped person determines that it is an emergency intervention situation. The cloud-based high-concurrency decision cabin immediately triggers a red warning to the command end and issues the highest priority target location information and advance instructions to the corresponding rescue unit to guide the rescue forces to carry out rapid approach and disposal along the target direction.

[0066] Fourth, when the comprehensive situation score is in the high range (e.g., 0.75 to 1.0), the comprehensive situation assessment model of the trapped person determines that it is an extreme life protection situation. While locking the absolute geographical location of the trapped person, the cloud-based high-concurrency decision cabin sends survival support and communication maintenance instructions to the nodes at the edge of the ruins to prioritize the continuous operation of the communication link, reassurance broadcast and vibration wake-up functions in the target area, and maintain the highest priority communication link occupation.

[0067] Based on a deterministic logic-based comprehensive situation assessment model for trapped nodes, the cloud-based high-concurrency decision-making cabin can support real-time priority ranking and global scheduling of a large number of distributed trapped nodes without the need for large-scale sample training.

[0068] S4. Monitor the stress extremes and physical offsets in the comprehensive situation assessment model of the trapped persons. When a specific component of the core psychological stress state vector exceeds the set threshold, trigger the graded intervention mechanism, generate execution instructions, and feed them back to the edge node of the ruins.

[0069] It should be noted that the stress extreme value refers to the peak data corresponding to the emotional component with the highest confidence score or the largest instantaneous growth rate in the core psychological stress state vector, used to characterize the extreme fluctuations in the psychological state of the trapped person; the physical offset refers to the target displacement drift calculated by combining the displacement data sensed by the three-axis digital accelerometer and the coordinate targeting compensation mechanism, specifically characterized as the difference between the original relative coordinates before calibration and the target absolute coordinates after calibration; when the physical offset exceeds the preset displacement drift threshold, or the instantaneous change rate exceeds the preset safe change rate threshold, it is determined that there is a risk of secondary collapse; among which, the displacement drift threshold is based on the installation position of the edge nodes in the ruins area. The stability of the structure, the amplitude of historical displacement fluctuations, and the baseline of silent displacement during the environmental calibration phase are set. The safe rate of change threshold is set based on the statistical results of the slope of the displacement change curve within the continuous sampling period, which is used to characterize whether there is a sudden increase in displacement in the target area. For example, the displacement drift threshold can be set as the weighted sum of the mean and standard deviation of the silent baseline displacement, and the safe rate of change threshold can be set as the weighted sum of the mean and standard deviation of the instantaneous rate of change during the silent phase. When a specific component in the core psychological stress state vector exceeds the corresponding set threshold, or when it is determined that there is a risk of secondary collapse, the multimodal fusion decision module immediately triggers a graded intervention mechanism.

[0070] The tiered intervention mechanism generates execution instructions based on the situation assessment results, including voice reassurance instructions, survival resource allocation instructions, or high-frequency early warning instructions. These instructions are then fed back to the edge nodes of the ruins in real time via communication links, whereby the edge nodes execute corresponding hardware actions, thus achieving a logical closed loop from situation awareness to proactive intervention.

[0071] In this embodiment, the multimodal fusion decision module presets a multi-level threshold judgment matrix to trigger differentiated intervention instructions based on the real-time emotional state of the person in distress. The high-level panic threshold is set to 0.8, and the medium-level despair threshold is set to 0.6. This setting is based on the fact that after probability renormalization, the core psychological stress state vector has values ​​between 0 and 1 for each state component. 0.8 represents that the panic state component has reached the high-confidence trigger range, and 0.6 represents that the despair state component has reached the medium-to-high-confidence intervention range. Simultaneously, specific evaluation ranges are set for the calm state component and the anxiety state component: an exemplary threshold of greater than 0.5 for the calm state component is used to confirm that the person in distress possesses basic psychological adjustment capabilities; when the calm state component meets the standard, the multimodal fusion decision module maintains routine observation. An exemplary threshold of greater than 0.4 for the anxiety state component is used to capture early signals of emotional fluctuations in the person in distress; when the anxiety state component meets the standard, the multimodal fusion decision module issues a control instruction to the visual emotion capture module to improve the accuracy of emotional feature extraction.

[0072] Regarding the specific content of the execution instructions: When the panic component exceeds the high-level panic threshold, the multimodal fusion decision module triggers the highest level intervention. The generated acoustic reassurance execution instructions specifically include: playing preset low-frequency alpha wave white noise, soothing light music, or voice packets recorded by the family members of the trapped person and sent in real time by the back-end dispatch center. The aim is to reduce the heart rate fluctuation of the trapped person through acoustic resonance and write the trapped person target corresponding to the current edge node of the ruins into the highest priority level of the cloud dispatch queue so that the back-end can prioritize the allocation of rescue resources. When the despair component exceeds the medium-level despair threshold, the multimodal fusion decision module triggers the medium-level intervention. The generated pulse tapping execution instructions specifically include: driving the vibration motor or piezoelectric ceramic plate of the edge node of the ruins to emit physical vibrations with a specific rhythm (such as simulated heartbeat rhythm or Morse code beat). Through tactile perception, the trapped person is guided to stay awake and prevented from entering a state of hypothermia or coma due to loss of consciousness.

[0073] When all state components of the trapped person are less than or equal to the corresponding trigger threshold, the comprehensive situation assessment model of the trapped person determines that the situation is stable. At this time, the cloud decision cabin enters a high-frequency polling monitoring state, that is, it continuously pulls the lightweight feature vector uploaded by the edge nodes of the ruins at a preset polling frequency. The polling frequency is preset according to the stable sampling period of the edge node sensors, the refresh frequency that the cloud communication link can bear, and the anti-jitter time window. For example, the polling frequency can be set to 10 to 20 times per second, corresponding to a polling period of 0.05 seconds to 0.1 seconds. The lightweight feature vector includes environmental threat indicators and core psychological stress state vectors after probability renormalization. In the extended implementation, physiological micro-vibration related features can also be attached. It is the same data object as the aforementioned composite data packet in the polling monitoring scenario, and a unified data encapsulation structure is maintained between the cloud and the edge nodes of the ruins.

[0074] There is a logical constraint relationship between the polling period and the anti-shake time window. That is, the polling period must be less than half of the anti-shake time window to ensure that multiple samplings are completed within one anti-shake period and sufficient sample data is provided for the anti-shake mechanism. In order to avoid high-frequency repeated alarms from interfering with the intervention link, this solution adopts a two-level anti-shake architecture: the first level uses a preset anti-shake time window to perform time-series filtering on the original feature stream to ensure that the situation assessment results are smooth; the second level is set before the hierarchical intervention mechanism is triggered, and the alarm signal is filtered again through a timestamp throttling valve. Only when the time difference between the current alarm to be triggered and the last executed intervention command is greater than the preset logical silence period is a new intervention execution command allowed to be generated and issued.

[0075] The anti-shake time window is preset based on the response cycle of the hardware actuator, with an example value range of 3 to 10 seconds; the logical silence period is not less than the anti-shake time window, and is preferably set to 1 to 2 times the anti-shake time window, so as to avoid repeatedly issuing the same type of intervention command within the same execution cycle.

[0076] When the edge node of the rubble receives a pulse striking command, the microcontroller built into the edge hardware node activates a high-precision hardware pulse width modulation generator. By dynamically adjusting the duty cycle and frequency, it drives an external eccentric wheel micro-vibration motor or electromagnetic striking device to output low-frequency mechanical resonance pulses. To enhance the penetration capability of the rubble medium, the underlying driver calls a medium acoustic impedance matching algorithm. By analyzing the environmental micro-vibration echoes captured in the early stage, it dynamically calculates the optimal low-frequency resonance frequency band of the current rubble medium, preferably between 10 Hz and 50 Hz. The setting is based on the low-attenuation transmission characteristics of low-frequency mechanical waves in heterogeneous rubble media and matches the sensitive frequency range of the human body's deep tactile receptors. The edge node of the rubble outputs a striking signal with an intermittent rhythm of "three short, three long, three short" to achieve this in extreme darkness. It provides effective tactile and auditory arousal stimulation for trapped individuals in confined spaces. Simultaneously, a peak amplitude safety threshold is defined, primarily referencing dynamic stability standards for highly sensitive building structures. An exemplary value is preferably a peak vibration acceleration of 0.05 times the gravitational acceleration (0.05g). This value, verified through experiments, is far below the critical acceleration that induces dynamic landslides from loose rock, ensuring mechanical wave penetration while maintaining at least a structural safety margin of more than one time. The system monitors its own vibration output through a closed-loop system, comparing the feedback value from the built-in accelerometer with the peak amplitude safety threshold in real time. When the built-in accelerometer feedback value reaches the peak amplitude safety threshold, the pulse width modulation duty cycle is automatically reduced to prevent excessive mechanical resonance from triggering secondary micro-collapses in the vulnerable area above the trapped individual.

[0077] When the edge node of the rubble receives an acoustic reassurance command, it retrieves high-fidelity human reassurance voice, white noise audio file, or pink noise audio file stored in local non-volatile memory or serial peripheral interface flash memory, and sends them to a digital-to-analog converter for decoding. Then, it drives a dustproof and waterproof speaker for directional broadcast through an impedance-matched miniature audio power amplifier circuit or a miniature Class-D audio power amplifier. While playing the reassurance audio, the multimodal perception module of the edge node of the rubble remains in a 24-hour working state and uses an echo cancellation algorithm to suppress interference from the reassurance voice being played, so as to ensure that the real-time emotional feedback of the trapped people or the physical displacement caused by secondary collapse can still be monitored synchronously during the broadcast.

[0078] In this embodiment, the multimodal perception module performs differential calculations on the temporal feature stream to dynamically capture psychological regression parameters, including heart rate recovery rate and fear index decline slope. The heart rate recovery rate is obtained by performing a peak extraction algorithm on the physiological micro-vibration time-series vector to obtain the instantaneous heart rate and monitoring the heart rate decline amplitude per unit time after the intervention command is issued. The fear index decline slope is obtained by locking the confidence level of the panic component in the core psychological stress state vector and calculating the first-order rate of change with respect to time.

[0079] By packaging psychological regression parameters in real time and transmitting them back to the cloud along with a lightweight feature vector, the backend scheduling center can quantitatively assess the effectiveness of the current intervention command. Specifically, after the intervention command is issued, the multimodal perception module at the edge of the rubble continuously transmits instantaneous heart rate and panic component confidence scores. The backend scheduling center uses two adjacent transmitted values ​​as a calculation unit, taking the previously transmitted instantaneous heart rate and panic component confidence scores as the previous value and the currently transmitted instantaneous heart rate and panic component confidence scores as the subsequent value; the difference between the previous and subsequent values ​​is calculated and compared with the two values. The ratio of the time intervals between data transmissions yields the heart rate recovery rate and the slope of the fear index decline. When both the heart rate recovery rate and the slope of the fear index decline are positive, the current intervention instruction is deemed to have a retardation effect. When either the heart rate recovery rate or the slope of the fear index decline is not positive, the current intervention instruction is deemed not to have a retardation effect, and the current intervention record is transmitted back to the cloud decision-making cabin as a basis for subsequent adjustments to the soothing sound source or the rhythm of the pulse tapping, thus forming an intelligent closed loop of "perception, decision-making, intervention, and effectiveness feedback evaluation".

[0080] In addition, when monitoring the physical offset in the comprehensive situation assessment model of trapped persons, a coordinate-targeting compensation mechanism can be used to assist in correcting the displacement drift of the lead-in drone or the edge node of the ruins caused by physical aftershocks, so as to improve the target pointing accuracy and on-site handling stability after the execution command is fed back to the edge node of the ruins.

[0081] This embodiment also provides a post-disaster aftershock perception and psychological defense system, including: a physical environment perception module, a visual emotion capture module, a multimodal fusion decision-making module, and an intelligent intervention execution module; The physical environment perception module is used to collect microseismic time-series waveform streams at the edge nodes of the ruins, and to perform adaptive denoising and phase picking algorithms on the microseismic time-series waveform streams to extract aftershock features and form environmental threat indicators. The visual emotion capture module is used to capture the facial visual stream data of trapped personnel. It calls a lightweight convolutional neural network to encode facial action units on the facial visual stream data, infers the multidimensional emotion probability of the trapped personnel, and maps it into a core psychological stress state vector through feature dimensionality reduction. The multimodal fusion decision module is used to perform feature space fusion and probability renormalization processing on environmental threat indicators and core psychological stress state vectors, and to construct a comprehensive situation assessment model for trapped individuals. The intelligent intervention execution module is used to monitor the stress extremes and physical offsets in the comprehensive situation assessment model of the trapped people. When a specific component of the core psychological stress state vector exceeds a set threshold, a graded intervention mechanism is triggered to generate execution instructions and feed them back to the edge node of the ruins.

[0082] In summary, this invention improves the accuracy and timeliness of aftershock feature extraction by: adaptively denoising and picking up seismic phases on microseismic time-series waveforms at the edge nodes of the rubble; extracting core psychological stress states and reducing computational consumption caused by redundant emotional features by performing lightweight convolutional neural network inference and feature dimensionality reduction mapping on the facial visual data of trapped personnel; achieving joint monitoring and comprehensive evaluation of the physical environment and psychological defenses by performing feature space fusion and probability renormalization on environmental threat indicators and core psychological stress state vectors; and forming a closed-loop protection link of perception, decision-making, intervention, and feedback by triggering graded interventions when specific components exceed thresholds and feeding back execution instructions to edge nodes, thereby improving the efficiency of rescue decision-making and the survival probability of trapped personnel in complex rubble scenarios.

[0083] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for jointly protecting post-disaster aftershock perception and psychological defense, characterized by: include: Microseismic time-series waveforms were acquired at the edge nodes of the ruins, and adaptive denoising and phase picking algorithms were applied to the microseismic time-series waveforms to extract aftershock features and form environmental threat indicators. The specific steps are as follows: An adaptive wavelet soft thresholding denoising process based on the Lipschitz singularity index is performed on the microseismic time-series waveform stream to obtain a clean denoised signal. The STA / LTA ratio algorithm is used to calculate the feature ratio of the clean, denoised signal; When the characteristic ratio exceeds the preset dynamic threshold, the initial arrival of the aftershock P wave is locked, forming an environmental threat indicator. The facial visual stream data of the trapped individuals is captured. A lightweight convolutional neural network is used to encode facial action units on the facial visual stream data to infer the multidimensional emotional probabilities of the trapped individuals. Finally, feature dimensionality reduction is used to map the data into a core psychological stress state vector. The specific steps are as follows: Regions of interest are extracted using a Haar feature cascade classifier; If the size of the region of interest is not standard, an interpolation algorithm is called to reshape the region of interest into a single-channel grayscale tensor; If region of interest extraction fails, an exception capture is triggered and the current frame is silently skipped. Facial visual stream data is input into the improved Mini-Xception network, which outputs the original 7-dimensional emotion probability vector. The original 7-dimensional emotion probability vector is transformed into a core psychological stress state vector through feature dimensionality reduction mapping. The environmental threat indicators and core psychological stress state vectors are subjected to feature space fusion and probability renormalization. The specific steps are as follows: Feature space fusion is performed on environmental threat indicators and core psychological stress state vectors to obtain feature space fusion results; The core psychological stress state vector is subjected to probability renormalization to generate a standard confidence matrix; And construct a comprehensive situation assessment model for those trapped; The system monitors the stress extremes and physical offsets in the comprehensive situation assessment model of trapped individuals. When a specific component of the core psychological stress state vector exceeds a set threshold, a graded intervention mechanism is triggered to generate execution instructions and feed them back to the edge nodes of the ruins.

2. The method for jointly protecting post-disaster aftershock perception and psychological defense as described in claim 1, characterized in that, The formation of environmental threat indicators also includes: after the feature ratio exceeds a preset dynamic threshold, embedding the Akaike Information Criterion (AIC) based on an autoregressive model within the trigger interval, calculating the logarithmic weighted sum of the variances of the two preceding and following sequences, and accurately locking the absolute first arrival time of the aftershock P-wave by finding the global minimum point.

3. The method for jointly protecting post-disaster aftershock perception and psychological defense as described in claim 1, characterized in that, The core psychological stress state vector refers to the state vector obtained by transforming the original 7-dimensional emotion probability vector through feature dimensionality reduction mapping. The state vector includes calm state components, anxiety state components, panic state components, and despair state components; The calm state component is the sum of the probability of happiness, the probability of neutrality, and the probability of surprise; The anxiety state component is the sum of the probability of anger and the probability of aversion; The panic state component is the absolute value of the fear probability; The despair state component is the absolute value of the probability of sadness.

4. The method for jointly protecting post-disaster aftershock perception and psychological defense as described in claim 1, characterized in that, The following are the specific steps for constructing a comprehensive situation assessment model for the trapped individuals: A comprehensive situation assessment model for trapped individuals is constructed based on the feature space fusion results and the standard confidence matrix.

5. The method for jointly protecting post-disaster aftershock perception and psychological defense as described in claim 1, characterized in that, The probability renormalization process refers to constructing a normalization ratio by summing the unnormalized component of any dimension with the sum of all state components and the zero-prevention division by bias constant, and using the normalization ratio as the final normalized confidence level of the current dimension.

6. The method for jointly protecting post-disaster aftershock perception and psychological defense as described in claim 1, characterized in that, When a specific component of the core psychological stress state vector exceeds a set threshold, a tiered intervention mechanism is triggered to generate execution instructions and feed them back to the edge nodes of the ruins. The specific steps are as follows: Use the high-level panic threshold and the medium-level despair threshold as trigger thresholds; When the panic component exceeds the high-level panic threshold, the highest level of intervention is triggered, generating an acoustic reassurance instruction and raising the emergency priority to the highest level. When the despair state component exceeds the intermediate-level despair threshold, an intermediate-level intervention is triggered, generating a pulse tapping execution command; The frequency of the pulse tapping command is dynamically adjusted based on the real-time feedback from the trapped person. When all state components are less than or equal to their corresponding trigger thresholds, a stable state is marked, and the current high-frequency polling monitoring state is maintained.

7. The method for jointly protecting post-disaster aftershock perception and psychological defense as described in claim 5, characterized in that, The final normalized confidence score of the current dimension is used to generate a standard confidence matrix that fits the radar chart rendering architecture.

8. A post-disaster aftershock sensing and psychological defense system, based on the post-disaster aftershock sensing and psychological defense system method according to any one of claims 1 to 7, characterized in that, include: The system includes a physical environment perception module, a visual emotion capture module, a multimodal fusion decision-making module, and an intelligent intervention execution module. The physical environment perception module is used to collect microseismic time-series waveform streams at the edge nodes of the ruins, and to perform adaptive denoising and phase picking algorithms on the microseismic time-series waveform streams to extract aftershock features and form environmental threat indicators. The visual emotion capture module is used to capture the facial visual stream data of trapped personnel. It calls a lightweight convolutional neural network to encode facial action units on the facial visual stream data, infers the multidimensional emotion probability of the trapped personnel, and maps it into a core psychological stress state vector through feature dimensionality reduction. The multimodal fusion decision module is used to perform feature space fusion and probability renormalization processing on environmental threat indicators and core psychological stress state vectors, and to construct a comprehensive situation assessment model for trapped individuals. The intelligent intervention execution module is used to monitor the stress extremes and physical offsets in the comprehensive situation assessment model of the trapped people. When a specific component of the core psychological stress state vector exceeds a set threshold, a graded intervention mechanism is triggered to generate execution instructions and feed them back to the edge node of the ruins.

Images