A weak network topology coordination personnel refuge scheduling method under a natural disaster scenario

Abstract

The application discloses a weak network topology coordination personnel refuge scheduling method in a natural disaster scene, comprising the following steps: collecting multi-source sensing data and performing data preprocessing; constructing a lightweight semantic analysis model, combining business instructions and network states to generate refuge business messages; constructing a site topology mapping model, generating topology coordination constraint parameters, realizing personnel positioning and completing differentiated evacuation path planning; outputting visual scheduling results; collecting on-site operation feedback data, combining actual traffic states to adaptively optimize path constraints and crowd traffic cost weights. The application can dynamically adapt to disaster scene changes by continuously iterating topology coordination constraint parameters, differentiated traffic cost weights and refuge paths, breaking through the limitations of traditional open-loop fixed strategies, and significantly enhancing intelligence and robustness.

Description

Technical Field

[0001] This invention belongs to the field of emergency evacuation and dispatching technology, specifically relating to a method for coordinated personnel evacuation and dispatching in weak network topology scenarios under natural disaster conditions. Background Technology

[0002] Natural disasters such as earthquakes, floods, and typhoons can easily damage base stations, block signals, and cause channel congestion, creating a weak network environment characterized by high latency, high packet loss, and strong jitter. Traditional emergency dispatch systems suffer from defects such as disordered communication control, low intelligence in command recognition, large positioning errors, homogenized evacuation routes, and lack of adaptive iteration capabilities. Furthermore, the functional modules are fragmented and lack coordination, making it difficult to meet the needs of precise, efficient, and hierarchical personnel evacuation dispatch in complex disaster scenarios. Specific technical deficiencies are as follows.

[0003] First, traditional sensing and command recognition technologies suffer from poor adaptability to weak networks and insufficient intelligence: Inertial data collected on-site is easily affected by equipment vibration, human movement, and electromagnetic interference. Conventional filtering and noise reduction have limited effectiveness, and direct use of noisy data in calculations reduces scheduling accuracy. Traditional command recognition relies on fixed keyword matching, which cannot adapt to incomplete, colloquial, and non-standardized commands at disaster sites, resulting in weak fault tolerance and generalization capabilities. Furthermore, it is highly dependent on cloud computing power and stable networks, and cannot function properly in weak or offline network scenarios.

[0004] Second, traditional weak network congestion control and bandwidth allocation strategies are rigid and undifferentiated: conventional emergency service transmission resource control uses fixed parameter control, which cannot adapt to the dynamic fluctuation characteristics of disaster networks; at the same time, it does not distinguish the priority of emergency services, and high-priority services such as command and dispatch and casualty transfer are easily squeezed out, resulting in delays, packet loss or even interruption of critical instructions, and failing to guarantee the reliability of core emergency service transmission.

[0005] Third, traditional positioning and disaster avoidance scheduling suffers from low accuracy, weak adaptability, and lack of closed-loop optimization: Pure inertial navigation suffers from severe cumulative drift errors over long-term operation; conventional filtering parameters are fixed and cannot adapt to dynamic personnel movement and weak network fluctuation scenarios, resulting in poor anti-interference capabilities. Path planning adopts a homogeneous and uniform strategy, failing to differentiate the needs of the general public, medical personnel, and people with mobility impairments. Furthermore, traditional systems employ an open-loop fixed strategy, unable to iteratively optimize parameters based on on-site traffic and network operation feedback, resulting in insufficient scenario adaptability.

[0006] In summary, existing technologies suffer from systemic defects such as weak intelligent recognition capabilities, disordered control of weak networks, insufficient positioning accuracy, homogeneous scheduling, and lack of closed-loop iterative mechanisms, making them unsuitable for the refined and adaptive personnel collaborative evacuation scheduling needs in disaster-prone network scenarios. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a method for coordinated personnel evacuation scheduling in weak network topology scenarios during natural disasters. This invention integrates multi-source data preprocessing, knowledge distillation and LoRA lightweight semantic parsing, topology-constrained adaptive particle filter positioning, and differentiated personnel path planning. It achieves the linkage of differentiated control of weak network communication, intelligent identification of emergency commands, high-precision positioning, and hierarchical evacuation scheduling. It can dynamically adapt to complex scenarios of weak network in disasters, significantly improving the stability of emergency transmission and the efficiency of personnel evacuation.

[0008] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0009] The present invention provides a method for scheduling personnel evacuation in a weak network topology scenario under natural disaster conditions, comprising the following steps:

[0010] Step 1: Collect multi-source sensing data and perform data preprocessing;

[0011] Step 2: Construct a lightweight semantic parsing model and generate risk avoidance service messages by combining business instructions and network status;

[0012] Step 3: Construct a site topology mapping model, generate topology coordination constraint parameters, realize personnel positioning, and complete differentiated evacuation route planning;

[0013] Step 4: Output the visual scheduling results;

[0014] Step 5: Collect on-site operational feedback data and adaptively optimize path constraints and crowd passage cost weights based on actual traffic conditions.

[0015] Furthermore, step 1 specifically includes:

[0016] Collect network status data, service instruction data, and pedestrian movement time-series data, and complete categorized preprocessing, partitioned caching, and time-series alignment; specifically as follows:

[0017] Link status data, including latency, packet loss rate, and channel occupancy rate of each communication node, is acquired through network acquisition probes and stored in the network timing cache partition; on-site handling instructions, task levels, and control area codes are received asynchronously through the terminal interface, and stored in the business instruction cache partition after standardized parsing and compliance verification.

[0018] The three-axis acceleration, angular velocity, and attitude angle data of the mobile terminal worn by pedestrians are collected. Local noise reduction is performed using a sliding window mean filter with a fixed window length W to suppress the jitter and motion interference of the mobile terminal. The filtering formula is as follows:

[0019] ;

[0020] ;

[0021] ;

[0022] ;

[0023] In the formula, These are the filtered and smoothed accelerations along the x, y, and z axes at time i, respectively. These are the filtered and smoothed angular velocities along the x, y, and z axes at time i, respectively. The attitude angle after filtering at time i; The length of the sliding window; These are the original accelerations along the x-axis, y-axis, and z-axis at time k within the window, respectively. These are the original angular velocities along the x-axis, y-axis, and z-axis at time k within the window, respectively. The original attitude angle at time k within the window is used; after filtering, smooth and continuous pedestrian motion time series data is obtained and stored in the pedestrian motion time series cache partition.

[0024] After the three types of data are cached independently, a unified UTC timestamp is added to complete global time-series alignment and form a standardized shared data base.

[0025] Furthermore, step 2 specifically includes:

[0026] A lightweight semantic parsing model is constructed using samples of risk avoidance service instructions. Based on the processed service instruction data and network status data from step 1, semantic parsing, weak network congestion quantitative assessment, and adaptive bandwidth differentiation are performed sequentially to generate risk avoidance service messages; the details are as follows:

[0027] Step 21: Select a large language model and perform structured pruning to obtain the student model; details are as follows:

[0028] Step 211: Quantify and statistically analyze the contribution rate of the network parameter weights of the attention head and the fully connected feedforward network within the model, and use the accuracy of disaster emergency command semantic recognition as an evaluation index to assess the contribution ability of the attention head and the fully connected feedforward network to the extraction of semantic features of disaster emergency commands.

[0029] Step 212: Set a contribution rate threshold and identify idle attention heads with network parameter weight contribution rates lower than the contribution rate threshold, weak semantic representation capabilities, no positive gain for disaster scenario reasoning, and redundant neurons inside the fully connected feedforward network as redundant structures to be pruned.

[0030] Step 213: Using a holistic hierarchical structured pruning method, only redundant structures that have been determined to be completed are removed, while the semantic feature extraction backbone structure for disaster emergency instructions and the semantic mapping relationship of disaster are fully preserved.

[0031] Step 214: Solidify the core parameters of the model, reduce the number of model parameters and computational load, and obtain the student model;

[0032] Step 22: Using the student model described above as the training object, transfer semantic knowledge of natural disaster scenarios through knowledge distillation between teachers and students. The knowledge distillation training adopts a soft-label joint loss function, the formula of which is as follows:

[0033] ;

[0034] In the formula, The overall distillation joint loss function; The hard classification loss between the student model and the real hedging business labels is used to ensure basic recognition accuracy; The distillation loss of the soft probability distribution is output for both the student model and the teacher model, which is used to transfer semantic knowledge of disaster scenarios; is the weight balancing coefficient, with a value range of (0,1), used to balance the training proportion of the two types of loss. It compensates for the accuracy loss caused by pruning through knowledge distillation, thereby improving the student model's ability to recognize the semantics of disaster emergency instructions.

[0035] Step 23: Use LoRA low-rank fine-tuning to adapt the trained student model to natural disaster scenarios, update the model's low-rank matrix, and freeze all core parameters; enable the model to recognize non-standardized emergency commands at natural disaster sites, distinguish four types of core emergency tasks, and output business priorities. The intent label is used to obtain a lightweight semantic parsing model;

[0036] Step 24: Input the business instruction data processed in Step 1 into the lightweight semantic parsing model for semantic parsing, and output the business priority. And the coding information of the risk avoidance and control area;

[0037] The coding information of the risk avoidance and control area is parsed and converted, and the area is mapped by combining it with the whole-domain electronic disaster prevention map to generate risk avoidance and control area constraint data.

[0038] Step 25: Complete the quantitative assessment of weak network congestion based on network status data;

[0039] The communication latency, packet loss rate, and channel occupancy rate indicators of the synchronization sequence are read, normalized to eliminate dimensional differences, and then the degree of weak network congestion is quantified through a weighted scoring model, as shown in the following formula:

[0040] ;

[0041] In the formula, This is a quantified weak network congestion score; the higher the value, the higher the degree of network congestion, jitter, and latency. This refers to the normalized communication latency. The packet loss rate after normalization; This represents the channel occupancy rate after normalization. To preset the weighting coefficients, satisfying This is used to balance the weights of the three types of network indicators on the degree of congestion.

[0042] Step 26: Combine the service priority P output in Step 24 and the weak network congestion score S obtained in Step 25 to perform adaptive bandwidth differentiation allocation, and encapsulate and generate risk avoidance service messages; the details are as follows:

[0043] Step 261: Use the service priority P and the weak network congestion score S to solve the adaptive bandwidth sampling period and the differentiated bandwidth allocation for single service links, and complete the adaptive adjustment calculation of emergency service transmission bandwidth.

[0044] The adaptive bandwidth sampling period is calculated as follows:

[0045] ;

[0046] In the formula, The adaptive bandwidth sampling period; Preset the basic sampling period for the system;

[0047] The solution for differentiated bandwidth allocation for a single service link is as follows:

[0048] ;

[0049] In the formula, Ultimately allocate bandwidth for a single service link; This represents the maximum available bandwidth for a single communication link.

[0050] Step 262: Unify the business priority, risk avoidance control area constraint data, and the adaptive bandwidth sampling period and single business link differentiated allocation bandwidth obtained in step 261 to generate risk avoidance business messages.

[0051] Furthermore, step 3 specifically includes:

[0052] Based on the business priority and evacuation control area constraint data extracted from the evacuation business message in step 2, and the pedestrian movement time series data in step 1, a site topology mapping model is constructed in conjunction with the full-domain disaster prevention electronic map. Topology collaborative constraint parameters are generated, personnel positioning is achieved through adaptive particle filtering, and differentiated path planning is completed based on personnel classification labels; the details are as follows:

[0053] Step 31: Construct a site topology mapping model and generate topology collaborative constraint parameters; details are as follows:

[0054] Step 311: Construct a site topology mapping model G(V,E) based on the global disaster prevention electronic map, where model G(V,E) is used to represent the on-site passage space structure, G is the set of the overall topology network graph; V is the set of topology nodes, corresponding to the site passage endpoints, turning points, and hazard gathering points; E is the set of topology edges, corresponding to the passage segments between nodes that can pass normally.

[0055] Step 312: Use the risk avoidance control area constraint data in the risk avoidance business message to correct the constraints of the passable area of ​​the site topology mapping model G(V,E), and update the attribute labels of the passable area and the prohibited area inside the model.

[0056] Step 313: Traverse and parse the corrected site topology mapping model G(V,E), extract the topology node set V, the topology edge set E, and the updated traversable / restricted area boundaries and attribute labels, and encapsulate them to obtain the topology collaborative constraint parameters;

[0057] Step 32: Perform personnel localization using topological constraint adaptive particle filtering; details are as follows:

[0058] Step 321: Generate a particle set using the particle state recursive equation, and constrain the particles by combining the road network spatial structure represented by the topological nodes and edges in the topological cooperative constraint parameters to complete the prior state estimation with road network constraints. Then, initialize the particle weights corresponding to each generated particle. The particle state recursive equation is as follows:

[0059] ;

[0060] In the formula, For time t, the first The position and state vector of each particle; This is a recursive function for the pedestrian's motion state, used to predict the current position based on the motion state at the previous moment; for Time of the first The position and state vector of each particle; for Inertial motion noise at all times is used to simulate the random motion disturbance error of personnel.

[0061] Step 322: Construct a topological constraint observation equation using the traversable / restricted region boundaries and attribute labels in the topological cooperative constraint parameters. This topological constraint observation equation performs particle position discrimination: when the predicted particle position is within a traversable region, output a valid probability; when the predicted particle position falls into a restricted region, output a probability of 0, thus suppressing invalid out-of-bounds particles and outputting the topological constraint observation likelihood function. Based on the initialized particle weights, the observation likelihood function is substituted into the particle filter basic weight update formula to complete the particle weight update calculation. The particle filter basic weight update calculation formula is as follows:

[0062] ;

[0063] In the formula, for Time of the first Initial weights for each particle; for Time of the first Individual particle weights; for Traffic constraint observations obtained at any time based on the site topology road network; This is the topological constraint observation likelihood function output by the topological constraint observation equation;

[0064] Based on the basic weight calculation results, a multi-dimensional dynamic adaptation factor is introduced to correct the particle weights. The dynamic weight adaptive update calculation formula is as follows:

[0065] ;

[0066] In the formula, After being corrected by multi-dimensional factors Time of the first Individual particle weights The motion state adaptation factor has a value range of (0,1) and is used to adapt to the full dynamic behavior of personnel, including starting, stopping, walking, turning, and staying. The value is large in low speed / stationary state to enhance positioning stability, and the value is moderate in high speed motion to preserve trajectory continuity. This is the topology fitting factor, with a value range of (0,1), used to filter compliant particles. The weak network state adaptation factor has a value range of (0,1) and is negatively correlated with the weak network congestion score S.

[0067] Normalize all particle weights and output the normalized corrected particle weights.

[0068] Step 323: Based on normalized particle weight resampling, retain high-weight effective particles. After particle screening, obtain the current time-based localization estimation result through weighted fusion, as shown in the following formula:

[0069] ;

[0070] In the formula, The optimal fused position state vector at time t; Preset the total number of particles for the particle filtering algorithm; The normalized particle weights; for Time of the first The position and state vector of each particle; The particle index, with a value range of 1 ≤ ≤ By using weighted summation to filter out high-weight effective particles, accurate and smooth positioning results are output.

[0071] Step 324: Using the 3σ error criterion, instantaneous outliers are eliminated to obtain the filtered personnel positioning coordinates. , to take as Time and position reference , The inertial recursion is performed, and the inertial recursion formula is as follows:

[0072] ;

[0073] In the formula, for Predicted coordinates based on inertia at any given moment; Equivalent to the location coordinates after removing latitude and longitude values ; These are the horizontal and vertical motion velocities after filtering and smoothing. By fixing the sampling time interval for the sensor, continuous trajectory deduction is achieved based on inertial recursion, and final accurate correction is completed by combining topological particle filtering.

[0074] By using multi-dimensional dynamic weight adaptive correction, topological constraint out-of-bounds elimination, and outlier error discrimination correction, data jitter and positioning offset in weak network environments are suppressed, and finally real-time location information of personnel is obtained.

[0075] Step 33: Define three types of personnel adaptation tags and configure differentiated passage cost weights;

[0076] Define a discrete classification label T to distinguish between personnel attributes and risk aversion needs, specifically dividing them into three differentiated population tags:

[0077] For people with mobility impairments;

[0078] For medical staff / on-site crowd control personnel;

[0079] For the general public;

[0080] For each type of label, configure a unique corresponding passage cost weight. Using traffic cost as an indicator to evaluate the quality of a route, traffic cost is divided into road segment traffic cost and route cost, and weights are used to determine the optimal route. The road segment toll costs are adjusted and accumulated to obtain the path cost. This cost is then used to determine the optimal risk avoidance path for different groups of people. The weighting rules for each group of people are as follows:

[0081] People with mobility impairments Matching high-safety, fault-tolerant passage cost weights To reduce the cost of traveling on flat, barrier-free, and low-difficulty road sections, and increase the cost of traveling on staircases, steep slopes, and narrow and complex road sections, so as to avoid high-risk terrain and prioritize the safety of special groups.

[0082] Medical staff / on-site crowd control personnel Matching efficient passage cost weights Reduce the cost of accessing emergency lanes and direct rescue routes, weaken unnecessary safety constraints, and prioritize shortening travel time to ensure the timeliness of emergency operations;

[0083] ordinary people Matching balanced passage cost weights To balance the safety of passageways with the efficiency of overall mass evacuation, so as to ensure an orderly evacuation from the site;

[0084] Step 34: Construct a path cost calculation model to complete differentiated evacuation route planning;

[0085] Using the site topology mapping model G(V,E) and topology collaborative constraint parameters, based on the traversable road segments of the topological road network, and combining the physical attributes of the road segments, real-time traffic status, and differentiated traffic cost weights, a path cost calculation model is constructed as follows:

[0086] ;

[0087] In the formula, This represents the total cost of traveling along a single evacuation route; the smaller the cost, the better the corresponding route. This refers to a single-segment road section after the complete path has been broken down. To iterate through and solve for the complete set of avoidance paths; The physical length of a single channel segment is a fixed constant of the site. The real-time congestion coefficient of the passage is obtained from on-site traffic status statistics.

[0088] With the goal of minimizing the overall passage cost, and combining the differentiated passage weights of the three types of personnel, the optimal evacuation routes that are adapted to different population attributes and match the priority of emergency operations are solved, and the differentiated evacuation route planning is finally completed.

[0089] Furthermore, step 4 specifically includes:

[0090] The data on risk control areas in the risk avoidance business messages in step 2 are integrated with the topological coordination constraint parameters, differentiated evacuation routes, and personnel location information in step 3. The integrated multi-dimensional data is then overlaid and rendered onto the disaster prevention electronic map, and the visualized scheduling results are output.

[0091] Furthermore, step 5 specifically includes:

[0092] A real-time dataset of on-site dynamic operational feedback from personnel is collected and annotated with a unified timestamp. This dataset serves as input for the generation of topology collaborative constraint parameters and iterative optimization of differentiated path planning. The dataset contains the actual on-site travel time for personnel corresponding to each evacuation path. Topological co-constraint parameters , , The respective passage cost weights for the three groups of people This achieves adaptive optimization of path planning constraints and crowd passage cost weights; among which, Used to provide real feedback on the actual adaptation effect of the path planning strategy;

[0093] Based on the dynamic on-site operational feedback data of personnel annotated with a unified timestamp, a global path iterative loss function is constructed. The optimization objective is to minimize the deviation between the predicted passage cost and the actual passage time. The formula for calculating the loss function is as follows:

[0094] ;

[0095] In the formula, This is the global path iterative loss function, used to predict the deviation between the traffic cost and the actual traffic status on site; The predicted travel cost output for path planning; This represents the actual time taken for personnel to pass through the site.

[0096] By minimizing the global mean square error, continuously and iteratively optimize the topological cooperative constraint parameters and , , Based on the optimized parameters, the passage cost weights for the three groups of people are used to solve the topology correction and differentiated evacuation paths. This dynamically adapts to the real-time changes in network fluctuations, terrain constraints, and personnel flow under disaster scenarios, and completes the closed-loop adaptive optimization of evacuation paths, topology coordination constraint parameters, and passage cost weights.

[0097] The beneficial effects of this invention are:

[0098] 1. This invention constructs a multi-source data standardization preprocessing and time-series alignment mechanism to effectively filter out motion, equipment and electromagnetic noise, forming a high-fidelity unified data base, providing reliable input for subsequent intelligent identification, precise positioning and dynamic scheduling.

[0099] 2. This invention achieves offline edge inference through knowledge distillation, model pruning, and LoRA fine-tuning, solving the problems of poor fault tolerance and reliance on the cloud in traditional keyword matching; it adaptively adjusts the sampling period and emptying rhythm based on real-time weak network conditions to achieve differentiated bandwidth allocation based on service priority and network status, ensuring stable transmission of high-priority services, and significantly improving weak network adaptability and communication reliability.

[0100] 3. This invention proposes a topology-constrained adaptive particle filter positioning mechanism, which effectively suppresses inertial navigation cumulative drift and weak network disturbance errors through multi-dimensional dynamic weight updates and 3σ outlier removal, significantly improving positioning accuracy, continuity, and anti-interference capabilities. It also constructs a differentiated path planning mechanism that couples personnel classification with passage cost weights, enabling multi-group hierarchical adaptive risk avoidance scheduling, overcoming the homogeneity defects of traditional paths, taking into account the safety of special groups and the timeliness of rescue operations, and significantly improving the rationality and safety of evacuation.

[0101] 4. This invention continuously iterates topological collaborative constraint parameters, differentiated passage cost weights, and risk avoidance paths based on on-site passage data, which can dynamically adapt to changes in disaster scenarios, breaking through the limitations of traditional open-loop fixed strategies, and significantly enhancing intelligence and robustness. Attached Figure Description

[0102] Figure 1 This is a flowchart illustrating the principle of the method of the present invention. Detailed Implementation

[0103] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to embodiments and accompanying drawings. The content mentioned in the embodiments is not intended to limit the present invention.

[0104] Reference Figure 1 As shown, the present invention provides a method for scheduling personnel evacuation in a weak network topology scenario under natural disaster conditions, comprising the following steps:

[0105] Step 1: Collect multi-source sensing data and perform data preprocessing; specifically including:

[0106] Collect network status data, service instruction data, and pedestrian movement time-series data, and complete categorized preprocessing, partitioned caching, and time-series alignment; specifically as follows:

[0107] Link status data, including latency, packet loss rate, and channel occupancy rate of each communication node, is acquired through network acquisition probes and stored in the network timing cache partition; on-site handling instructions, task levels, and control area codes are received asynchronously through the terminal interface, and stored in the business instruction cache partition after standardized parsing and compliance verification.

[0108] The three-axis acceleration, angular velocity, and attitude angle data of the mobile terminal worn by pedestrians are collected. Local noise reduction is performed using a sliding window mean filter with a fixed window length W to suppress the jitter and motion interference of the mobile terminal. The filtering formula is as follows:

[0109] ;

[0110] ;

[0111] ;

[0112] ;

[0113] In the formula, These are the filtered and smoothed accelerations along the x, y, and z axes at time i, respectively. These are the filtered and smoothed angular velocities along the x, y, and z axes at time i, respectively. The attitude angle after filtering at time i; The length of the sliding window; These are the original accelerations along the x-axis, y-axis, and z-axis at time k within the window, respectively. These are the original angular velocities along the x-axis, y-axis, and z-axis at time k within the window, respectively. The original attitude angle at time k within the window is used; after filtering, smooth and continuous pedestrian motion time series data is obtained and stored in the pedestrian motion time series cache partition.

[0114] After the three types of data are cached independently, a unified UTC timestamp is added to complete global time-series alignment and form a standardized shared data base.

[0115] Step 2: Construct a lightweight semantic parsing model to generate risk avoidance service messages by combining business commands and network status; specifically including:

[0116] A lightweight semantic parsing model is constructed using samples of risk avoidance service instructions. Based on the processed service instruction data and network status data from step 1, semantic parsing, weak network congestion quantitative assessment, and adaptive bandwidth differentiation are performed sequentially to generate risk avoidance service messages; the details are as follows:

[0117] Step 21: Select the Qwen-2.5-7B-Base open-source Transformer pre-trained base model and perform structured pruning to obtain the student model; details are as follows:

[0118] Step 211: Quantify and statistically analyze the contribution rate of the network parameter weights of the attention head and the fully connected feedforward network (FFN) within the model. Use the accuracy of disaster emergency command semantic recognition as the evaluation index to assess the contribution ability of the attention head and the fully connected feedforward network (FFN) to the extraction of semantic features of disaster emergency commands. In other examples, this base model can also be replaced by similar open-source Transformer pre-trained base models such as LLaMA-3 and Mistral.

[0119] Step 212: Set a contribution rate threshold and identify idle attention heads with network parameter weight contribution rates lower than the contribution rate threshold, weak semantic representation capabilities, no positive gain for disaster scenario reasoning, and redundant neurons inside the fully connected feedforward network as redundant structures to be pruned.

[0120] Step 213: Using a holistic hierarchical structured pruning method, only redundant structures that have been determined to be completed are removed, while the semantic feature extraction backbone structure for disaster emergency instructions and the semantic mapping relationship of disaster are fully preserved.

[0121] Step 214: Solidify the core parameters of the model, reduce the number of model parameters and computational load, and obtain the student model;

[0122] Step 22: Using the student model described above as the training object, semantic knowledge of the natural disaster scenario is transferred through knowledge distillation between teachers and students to compensate for the slight accuracy loss caused by the pruning process. The knowledge distillation training adopts a soft-label joint loss function, the formula of which is as follows:

[0123] ;

[0124] In the formula, The overall distillation joint loss function; The hard classification loss between the student model and the real hedging business labels is used to ensure basic recognition accuracy; The distillation loss of the soft probability distribution is output for both the student model and the teacher model, which is used to transfer semantic knowledge of disaster scenarios; is the weight balancing coefficient, with a value range of (0,1), used to balance the training proportion of the two types of loss. It compensates for the accuracy loss caused by pruning through knowledge distillation, thereby improving the student model's ability to recognize the semantics of disaster emergency instructions.

[0125] Step 23: Use LoRA low-rank fine-tuning to adapt the trained student model to natural disaster scenarios. Update the model's low-rank matrix and freeze all core parameters to avoid the problems of parameter redundancy, overfitting, and increased inference latency caused by full fine-tuning. This enables the model to recognize non-standardized emergency commands at natural disaster sites, distinguish four types of core emergency tasks, and output business priorities. Intent tags (with smaller values ​​indicating higher priority) improve the accuracy and fault tolerance of command recognition in weak network scenarios, thereby obtaining a lightweight semantic parsing model that effectively solves the shortcomings of traditional large models, such as insufficient edge computing power, high inference latency, and inability to run offline.

[0126] Step 24: Input the business instruction data processed in Step 1 into the lightweight semantic parsing model for semantic parsing, and output the business priority. And the coding information of the risk avoidance and control area;

[0127] The coding information of the risk avoidance and control area is parsed and converted, and the area is mapped by combining it with the whole-domain electronic disaster prevention map to generate risk avoidance and control area constraint data.

[0128] The risk avoidance and control area constraint data includes the spatial boundary and restricted / accessible attribute information of the corresponding control area, which is used for subsequent topology model correction.

[0129] Step 25: Complete the quantitative assessment of weak network congestion based on network status data;

[0130] The communication latency, packet loss rate, and channel occupancy rate indicators of the synchronization sequence are read, normalized to eliminate dimensional differences, and then the degree of weak network congestion is quantified through a weighted scoring model, as shown in the following formula:

[0131] ;

[0132] In the formula, This is a quantified weak network congestion score; the higher the value, the higher the degree of network congestion, jitter, and latency. This refers to the normalized communication latency. The packet loss rate after normalization; This represents the channel occupancy rate after normalization. To preset the weighting coefficients, satisfying This is used to balance the weights of the three types of network indicators on the degree of congestion.

[0133] Step 26: Combine the service priority P output in Step 24 and the weak network congestion score S obtained in Step 25 to perform adaptive bandwidth differentiation allocation, and encapsulate and generate risk avoidance service messages; the details are as follows:

[0134] Step 261: Use the service priority P and the weak network congestion score S to solve the adaptive bandwidth sampling period and the differentiated bandwidth allocation for single service links, and complete the adaptive adjustment calculation of emergency service transmission bandwidth.

[0135] The adaptive bandwidth sampling period is calculated as follows:

[0136] ;

[0137] In the formula, The adaptive bandwidth sampling period; A basic sampling period is preset for the system; the higher the congestion score S, the larger the sampling period is adaptively increased to reduce the frequency of invalid sampling and avoid wasting channel resources;

[0138] The solution for differentiated bandwidth allocation for a single service link is as follows:

[0139] ;

[0140] In the formula, Ultimately allocate bandwidth for a single service link; This represents the maximum available bandwidth for a single communication link. The smaller the value, the higher the service priority, and the larger the allocated bandwidth value, thus achieving differentiated bandwidth guarantee for disaster emergency services.

[0141] By combining business priority and network congestion status in a two-dimensional calculation, the system dynamically adapts to fluctuations in weak network environments, allocates communication resources to high-priority emergency services, and ensures the stability and timeliness of disaster emergency command transmission.

[0142] Step 262: Unify the business priority, risk avoidance control area constraint data, and the adaptive bandwidth sampling period and single business link differentiated allocation bandwidth obtained in step 261 to generate risk avoidance business messages.

[0143] Step 3: Construct a site topology mapping model, generate topology coordination constraint parameters, realize personnel positioning, and complete differentiated evacuation route planning; specifically including:

[0144] Based on the business priority and evacuation control area constraint data extracted from the evacuation business message in step 2, and the pedestrian movement time series data in step 1, a site topology mapping model is constructed in conjunction with the full-domain disaster prevention electronic map. Topology collaborative constraint parameters are generated, personnel positioning is achieved through adaptive particle filtering, and differentiated path planning is completed based on personnel classification labels; the details are as follows:

[0145] Step 31: Construct a site topology mapping model and generate topology collaborative constraint parameters; details are as follows:

[0146] Step 311: Construct a site topology mapping model G(V,E) based on the global disaster prevention electronic map, where model G(V,E) is used to represent the on-site passage space structure, G is the set of the overall topology network graph; V is the set of topology nodes, corresponding to the site passage endpoints, turning points, and hazard gathering points; E is the set of topology edges, corresponding to the passage segments between nodes that can pass normally.

[0147] Step 312: Use the risk avoidance control area constraint data in the risk avoidance business message to correct the constraints of the passable area of ​​the site topology mapping model G(V,E), and update the attribute labels of the passable area and the prohibited area inside the model.

[0148] Step 313: Traverse and parse the corrected site topology mapping model G(V,E), extract the topology node set V, the topology edge set E, and the updated traversable / restricted area boundaries and attribute labels, and encapsulate them to obtain the topology collaborative constraint parameters;

[0149] The topology coordination constraint parameters include: topology nodes, topology edges, traversable / restricted area boundaries, and attribute markers, which are used for subsequent topology constraint calculations for personnel positioning.

[0150] Step 32: Perform personnel localization using topological constraint adaptive particle filtering; details are as follows:

[0151] Step 321: Generate a particle set using the particle state recursive equation, and constrain the particles by combining the road network spatial structure represented by the topological nodes and edges in the topological cooperative constraint parameters to complete the prior state estimation with road network constraints. Then, initialize the particle weights corresponding to each generated particle. The particle state recursive equation is as follows:

[0152] ;

[0153] In the formula, For time t, the first The position and state vector of each particle; This is a recursive function for the pedestrian's motion state, used to predict the current position based on the motion state at the previous moment; for Time of the first The position and state vector of each particle; for Inertial motion noise at all times is used to simulate the random motion disturbance error of personnel.

[0154] Step 322: Construct a topological constraint observation equation using the traversable / restricted region boundaries and attribute labels in the topological cooperative constraint parameters. This topological constraint observation equation performs particle position discrimination: when the predicted particle position is within a traversable region, output a valid probability; when the predicted particle position falls into a restricted region, output a probability of 0, thus suppressing invalid out-of-bounds particles and outputting the topological constraint observation likelihood function. Based on the initialized particle weights, the observation likelihood function is substituted into the particle filter basic weight update formula to complete the particle weight update calculation. The particle filter basic weight update calculation formula is as follows:

[0155] ;

[0156] In the formula, for Time of the first Initial weights for each particle; for Time of the first Individual particle weights; for Traffic constraint observations obtained at any time based on the site topology road network; This is the topological constraint observation likelihood function output by the topological constraint observation equation;

[0157] This invention breaks through the traditional fixed weight update method and adds a multi-dimensional dynamic weight adaptive mechanism to adapt to the dynamic changes of personnel movement, topology fitting and weak network status.

[0158] Based on the basic weight calculation results, a multi-dimensional dynamic adaptation factor is introduced to correct the particle weights. The dynamic weight adaptive update calculation formula is as follows:

[0159] ;

[0160] In the formula, After being corrected by multi-dimensional factors Time of the first Individual particle weights The motion state adaptation factor has a value range of (0,1) and is used to adapt to the full dynamic behavior of personnel, including starting, stopping, walking, turning, and staying. The value is large in low speed / stationary state to enhance positioning stability, and the value is moderate in high speed motion to preserve trajectory continuity. This is the topology fit factor, with a value range of (0,1). It is used to filter compliant passage particles. The closer the particle is to the center of the passable area, the larger the value, which effectively suppresses out-of-bounds noise particles. The weak network state adaptation factor has a value range of (0,1) and is negatively correlated with the weak network congestion score S. When the weak network is poor and the data jitter is severe, the value is large, and the particle state output of the filtering prediction step is given priority to weaken the abnormal observation disturbance; when the network is smooth, the value is small to enhance the topology observation and positioning accuracy.

[0161] Normalize all particle weights and output the normalized corrected particle weights.

[0162] Step 323: Based on normalized particle weight resampling, retain high-weight effective particles. After particle screening, obtain the current time-based localization estimation result through weighted fusion, as shown in the following formula:

[0163] ;

[0164] In the formula, The optimal fused position state vector at time t; Preset the total number of particles for the particle filtering algorithm; The normalized particle weights; for Time of the first The position and state vector of each particle; The particle index, with a value range of 1 ≤ ≤ By using weighted summation to filter out high-weight effective particles, accurate and smooth positioning results are output.

[0165] Step 324: Using the 3σ (3 times standard deviation) error criterion, instantaneous outlier values ​​are eliminated to obtain the filtered personnel positioning coordinates. , to take as Time and position reference , The inertial recursion is performed, and the inertial recursion formula is as follows:

[0166] ;

[0167] In the formula, for Predicted coordinates based on inertia at any given moment; Equivalent to the location coordinates after removing latitude and longitude values ; These are the horizontal and vertical motion velocities after filtering and smoothing. By fixing the sampling time interval for the sensor, continuous trajectory deduction is achieved based on inertial recursion, and final accurate correction is completed by combining topological particle filtering.

[0168] By using multi-dimensional dynamic weight adaptive correction, topological constraint out-of-bounds elimination, and outlier error discrimination correction, data jitter and positioning offset in weak network environments are suppressed, and finally real-time location information of personnel is obtained.

[0169] Step 33: Define three types of personnel adaptation tags and configure differentiated passage cost weights;

[0170] Define a discrete classification label T to distinguish between personnel attributes and risk aversion needs, specifically dividing them into three differentiated population tags:

[0171] For people with mobility impairments;

[0172] For medical staff / on-site crowd control personnel;

[0173] For the general public;

[0174] For each type of label, configure a unique corresponding passage cost weight. Using traffic cost as an indicator to evaluate the quality of a route, traffic cost is divided into road segment traffic cost and route cost, and weights are used to determine the optimal route. The road segment toll costs are adjusted and accumulated to obtain the path cost. This cost is then used to determine the optimal risk avoidance path for different groups of people. The weighting rules for each group of people are as follows:

[0175] People with mobility impairments Matching high-safety, fault-tolerant passage cost weights To reduce the cost of traveling on flat, barrier-free, and low-difficulty road sections, and increase the cost of traveling on staircases, steep slopes, and narrow and complex road sections, so as to avoid high-risk terrain and prioritize the safety of special groups.

[0176] Medical staff / on-site crowd control personnel Matching efficient passage cost weights Reduce the cost of accessing emergency lanes and direct rescue routes, weaken unnecessary safety constraints, and prioritize shortening travel time to ensure the timeliness of emergency operations;

[0177] ordinary people Matching balanced passage cost weights To balance the safety of passageways with the efficiency of overall mass evacuation, so as to ensure an orderly evacuation from the site;

[0178] Step 34: Construct a path cost calculation model to complete differentiated evacuation route planning;

[0179] Using the site topology mapping model G(V,E) and topology collaborative constraint parameters, based on the traversable road segments of the topological road network, and combining the physical attributes of the road segments, real-time traffic status, and differentiated traffic cost weights, a path cost calculation model is constructed as follows:

[0180] ;

[0181] In the formula, This represents the total cost of traveling along a single evacuation route; the smaller the cost, the better the corresponding route. This refers to a single-segment road section after the complete path has been broken down. To iterate through and solve for the complete set of avoidance paths; The physical length of a single channel segment is a fixed constant of the site. The real-time congestion coefficient of the passage is obtained from on-site traffic status statistics.

[0182] With the goal of minimizing the overall passage cost, and combining the differentiated passage weights of the three types of personnel, the optimal evacuation routes that are adapted to different population attributes and match the priority of emergency operations are solved, and the differentiated evacuation route planning is finally completed.

[0183] Step 4: Output the visual scheduling results; specifically including:

[0184] The data on risk control areas in the risk avoidance business messages in step 2 are integrated with the topological coordination constraint parameters, differentiated evacuation routes, and personnel location information in step 3. The integrated multi-dimensional data is then overlaid and rendered onto the disaster prevention electronic map, and the visualized scheduling results are output.

[0185] Step 5: Collect on-site operational feedback data and adaptively optimize path constraints and crowd passage cost weights based on actual traffic conditions; specifically including:

[0186] A real-time dataset of on-site dynamic operational feedback from personnel is collected and annotated with a unified timestamp. This dataset serves as input for the generation of topology collaborative constraint parameters and iterative optimization of differentiated path planning. The dataset contains the actual on-site travel time for personnel corresponding to each evacuation path. Topological co-constraint parameters , , The respective passage cost weights for the three groups of people This achieves adaptive optimization of path planning constraints and crowd passage cost weights; among which, Used to provide real feedback on the actual adaptation effect of the path planning strategy;

[0187] Based on the dynamic on-site operational feedback data of personnel annotated with a unified timestamp, a global path iterative loss function is constructed. The optimization objective is to minimize the deviation between the predicted passage cost and the actual passage time. The formula for calculating the loss function is as follows:

[0188] ;

[0189] In the formula, This is the global path iterative loss function, used to predict the deviation between the traffic cost and the actual traffic status on site; The predicted travel cost output for path planning; This represents the actual time taken for personnel to pass through the site.

[0190] By minimizing the global mean square error, continuously and iteratively optimize the topological cooperative constraint parameters and , , Based on the optimized parameters, the passage cost weights for the three groups of people are used to solve the topology correction and differentiated evacuation paths. This dynamically adapts to the real-time changes in network fluctuations, terrain constraints, and personnel flow under disaster scenarios, and completes the closed-loop adaptive optimization of evacuation paths, topology coordination constraint parameters, and passage cost weights.

[0191] This invention has many specific applications. The above description is only a preferred embodiment of this invention. It should be noted that for those skilled in the art, several improvements can be made without departing from the principle of this invention, and these improvements should also be considered within the scope of protection of this invention.

Claims

1. A weak network topology coordination personnel evacuation scheduling method under a natural disaster scenario, characterized in that, The steps are as follows: Step 1: Collect multi-source sensing data and perform data preprocessing; Step 2: Construct a lightweight semantic parsing model and generate risk avoidance service messages by combining business instructions and network status; Step 3: Construct a site topology mapping model, generate topology coordination constraint parameters, realize personnel positioning, and complete differentiated evacuation route planning; Step 4: Output the visual scheduling results; Step 5: Collect on-site operational feedback data and adaptively optimize path constraints and crowd passage cost weights based on actual traffic conditions.

2. The personnel evacuation scheduling method in a weak network topology under a natural disaster scenario according to claim 1, characterized in that, Step 1 specifically includes: Collect network status data, service instruction data, and pedestrian movement time-series data, and complete categorized preprocessing, partitioned caching, and time-series alignment; specifically as follows: Acquire link status data including latency, packet loss rate and channel occupancy rate of each communication node, and store it in the network timing cache partition; asynchronously receive on-site handling instructions, task level and control area code, and store them in the business instruction cache partition after standardized parsing and compliance verification; The three-axis acceleration, angular velocity, and attitude angle data of the mobile terminal worn by pedestrians are collected. Local noise reduction is performed using a sliding window mean filter with a fixed window length W to suppress the jitter and motion interference of the mobile terminal. The filtering formula is as follows: ； ； ； ； In the formula, These are the filtered and smoothed accelerations along the x, y, and z axes at time i, respectively. These are the filtered and smoothed angular velocities along the x, y, and z axes at time i, respectively. The attitude angle after filtering at time i; The length of the sliding window; These are the original accelerations along the x-axis, y-axis, and z-axis at time k within the window, respectively. These are the original angular velocities along the x-axis, y-axis, and z-axis at time k within the window, respectively. The original attitude angle at time k within the window is used; after filtering, smooth and continuous pedestrian motion time series data is obtained and stored in the pedestrian motion time series cache partition. After the three types of data are cached independently, a unified UTC timestamp is added to complete global time-series alignment and form a standardized shared data base.

3. The natural disaster scenario weak network topology personnel evacuation dispatching method according to claim 1, characterized in that, Step 2 specifically includes: Step 21: Select a large language model and perform structured pruning to obtain the student model; Step 22: Using the student model described above as the training object, transfer semantic knowledge of natural disaster scenarios through knowledge distillation between teachers and students; Step 23: Adapt the trained student model to natural disaster scenarios, update the model's low-rank matrix, and freeze all core parameters; enable the model to recognize non-standardized emergency instructions at natural disaster sites, distinguish four types of core emergency tasks, and output intent tags for business priorities, thereby obtaining a lightweight semantic parsing model. Step 24: input the processed service instruction data in step 1 into a lightweight semantic parsing model for semantic parsing, and output service priority And hedging control area coding information; The coding information of the risk avoidance and control area is parsed and converted, and the area is mapped by combining it with the whole-domain electronic disaster prevention map to generate risk avoidance and control area constraint data. Step 25: Complete the quantitative assessment of weak network congestion based on network status data; Step 26: Combine the service priority P output in Step 24 and the weak network congestion score S obtained in Step 25 to perform adaptive bandwidth differentiation allocation, and encapsulate and generate risk avoidance service messages.

4. The natural disaster scenario weak network topology personnel evacuation dispatching method according to claim 3, characterized in that, Step 21 specifically includes: Step 211: Quantify and statistically analyze the contribution rate of the network parameter weights of the attention head and the fully connected feedforward network within the model, and use the accuracy of disaster emergency command semantic recognition as an evaluation index to assess the contribution ability of the attention head and the fully connected feedforward network to the extraction of semantic features of disaster emergency commands. Step 212: Set a contribution rate threshold and identify idle attention heads with network parameter weight contribution rates lower than the contribution rate threshold, weak semantic representation capabilities, no positive gain for disaster scenario reasoning, and redundant neurons inside the fully connected feedforward network as redundant structures to be pruned. Step 213: Using a holistic hierarchical structured pruning method, only redundant structures that have been determined to be completed are removed, while the semantic feature extraction backbone structure for disaster emergency instructions and the semantic mapping relationship of disaster are fully preserved. Step 214: Solidify the core parameters of the model, compress the number of model parameters and computational load, and obtain the student model.

5. The natural disaster scenario weak network topology personnel evacuation dispatching method according to claim 4, characterized in that, Step 25 specifically includes: The communication latency, packet loss rate, and channel occupancy rate indicators of the synchronization sequence are read, normalized to eliminate dimensional differences, and then the degree of weak network congestion is quantified through a weighted scoring model, as shown in the following formula: ； In the formula, is the quantified weak network congestion score, and the larger the value is, the higher the network congestion, jitter, and delay degree is; is the normalized communication delay; is the normalized packet loss rate; is the normalized channel occupancy rate; is a preset weighted weight coefficient, and satisfies is used to balance the influence weight of the three types of network indexes on the congestion degree.

6. The natural disaster scenario weak network topology personnel evacuation dispatching method according to claim 5, characterized in that, Step 26 specifically includes: Step 261: Use the service priority P and the weak network congestion score S to solve the adaptive bandwidth sampling period and the differentiated bandwidth allocation for single service links, and complete the adaptive adjustment calculation of emergency service transmission bandwidth. The adaptive bandwidth sampling period is calculated as follows: ； In the formula, is an adaptive bandwidth sampling period; is a system preset basic sampling period; The solution for differentiated bandwidth allocation for a single service link is as follows: ； In the formula, allocating bandwidth to the single service link; is the maximum available bandwidth for the communication single link; Step 262: Unify the business priority, risk avoidance control area constraint data, and the adaptive bandwidth sampling period and single business link differentiated allocation bandwidth obtained in step 261 to generate risk avoidance business messages.

7. The natural disaster scenario weak network topology personnel evacuation scheduling method according to claim 1, characterized in that, Step 3 specifically includes: Step 31: Construct a site topology mapping model and generate topology co-constraint parameters; Step 32: Perform personnel localization using topological constraint adaptive particle filtering; Step 33: Define three types of personnel adaptation tags and configure differentiated passage cost weights; Step 34: Construct a path cost calculation model to complete the differentiated evacuation path planning.

8. The natural disaster scenario weak network topology personnel evacuation dispatching method according to claim 7, characterized in that, Step 4 specifically includes: The data on risk control areas in the risk avoidance business messages in step 2 are integrated with the topological coordination constraint parameters, differentiated evacuation routes, and personnel location information in step 3. The integrated multi-dimensional data is then overlaid and rendered onto the disaster prevention electronic map, and the visualized scheduling results are output.

9. The natural disaster scenario weak network topology personnel evacuation dispatching method according to claim 7, characterized in that, Step 5 specifically includes: The dataset of real-time on-site dynamic operation feedback from personnel is collected and annotated with a unified timestamp. It serves as input data for the generation of topology collaborative constraint parameters and iterative optimization of differentiated path planning. The dataset contains the actual on-site passage time of personnel for each evacuation path, topology collaborative constraint parameters, and passage cost weights for the three types of people, enabling adaptive optimization of path planning constraints and passage cost weights for the population. Based on the on-site dynamic operation feedback data of personnel after adding a unified timestamp, a global path iterative loss function is constructed with the optimization objective of minimizing the deviation between the predicted passage cost and the actual passage time. By minimizing the global mean square error, the topology coordination constraint parameters and the passage cost weights corresponding to the three groups of people are continuously optimized in a targeted iterative manner. Based on the optimized parameters, the topology correction and differentiated evacuation paths are solved, dynamically adapting to network fluctuations, terrain constraints and real-time changes in personnel flow under disaster scenarios, and completing the closed-loop adaptive optimization of evacuation paths, topology coordination constraint parameters and passage cost weights.

Images