A multi-source disaster expert mixed fine-tuning method and system for an emergency management large model

Abstract

The present application relates to the technical field of multi-source disaster management, and provides a multi-source disaster expert hybrid fine-tuning method and system for an emergency management large model, comprising the following steps: S1, acquiring data, acquiring multi-source data of emergency management; S2, standardizing, field annotating and task annotating the multi-source data, and constructing training samples; S3, constructing an expert set of the emergency management large model; S4, constructing an expert router, wherein the expert router calculates the activation score of each expert in the expert set according to the sample characteristics of the training samples; S5, determining a target expert set; S6, performing expert selection type fine-tuning; S7, performing multi-task joint training; and S8, in the reasoning stage, outputting disaster risk assessment results, risk explanations and emergency disposal suggestions, and then performing constraint checking on the high-risk suggestions by a safety checking expert. The fine-tuning process of the emergency management large model is converted from unified parameter updating to field expert selection type updating, and the knowledge interference between different data sources is reduced.

Description

Technical Field

[0001] This invention relates to the field of multi-source disaster management technology, and in particular to a multi-source disaster expert hybrid fine-tuning method and system for large-scale emergency management models. Background Technology

[0002] In emergency management scenarios for natural disasters such as typhoons, rainstorms, and floods, disaster risk assessment and personnel evacuation decisions typically require the integration of multiple types of data, including climate, water resources, population, geography, transportation, supplies, and emergency plans. Climate data reflects changes in rainfall, wind speed, air pressure, and typhoon paths; water resources data reflects river levels, reservoir levels, flow rates, rise rates, and warning water levels; population data reflects the degree of population exposure in affected areas and the distribution of key population groups; and supplies data reflects the capacity of shelters, the stockpile of relief supplies, and transportation capacity.

[0003] While existing large language models possess strong text understanding, summarization, and generation capabilities, directly inputting the aforementioned multi-source disaster data into a single large model for unified fine-tuning can easily lead to problems such as domain knowledge mixing, negative parameter transfer, and unstable decision interpretation. For example, disaster severity prediction relies more on climate and water resources signals, personnel evacuation strategies rely more on population exposure, geographical risks, and emergency response rules, and material dispatch recommendations rely more on inventory, transportation capacity, and shelter capacity. The knowledge, reasoning processes, and safety constraints required for different tasks are not the same.

[0004] Furthermore, emergency management data typically suffers from limited sample sizes, scarce extreme disaster samples, uneven data quality across different regions, and high costs associated with manual annotation. Performing full-scale, uniform fine-tuning of a large model is not only computationally expensive but also risks improving performance in some areas while compromising capabilities in others. Particularly in high-risk scenarios involving personnel relocation, the opening of shelters, and resource allocation, recommendations generated by large models must have clear sources of evidence and safety constraints, and cannot rely on unconstrained generation.

[0005] Unlike general hybrid expert models, general MoE (Mixture of Experts) are usually designed for training corpora from the same source or relatively uniform task distributions. Their routing objectives are mainly computational sparsity, expert load balancing, or expanding model capacity. However, climate, water conservancy, population, and material data in emergency management scenarios have different physical meanings, sampling frequencies, reliability boundaries, and security constraints. Direct load balancing routing cannot solve the problems of cross-domain knowledge conflicts, incomplete disaster evidence, and unreliable high-risk decision outputs. Summary of the Invention

[0006] To overcome or alleviate one or more of the above-mentioned technical problems, the purpose of this invention is to provide a multi-source disaster expert hybrid fine-tuning method and system for large-scale emergency management models. This invention employs selective fine-tuning, aiming at negative migration suppression and security risk control of multi-source heterogeneous disaster data. It directly links expert selection to the disaster domain, task objectives, risk thresholds, and evidence integrity, rather than simply applying a general MoE structure. Based on different disaster samples, different input sources, and different task objectives, this method dynamically selects corresponding domain experts to participate in training and inference, ensuring that information from different domains such as climate, water resources, population, and materials is processed by the corresponding experts, and constrains high-risk outputs through security checks.

[0007] This invention provides the following technical solution:

[0008] A multi-source disaster expert hybrid fine-tuning method for large-scale emergency management models includes the following steps: S1, data acquisition.

[0009] Acquire multi-source emergency management data, including one or more of the following: climate data, water resources data, population data, material data, geographic data, transportation data, historical disaster records, and emergency plan texts;

[0010] S2, Constructing training samples

[0011] The multi-source data is standardized, domain-labeled, and task-labeled to construct training samples. Each training sample includes input data, domain identifier, disaster type, regional information, time window, task type, and supervision label.

[0012] S3, Constructing an Expert Set

[0013] The expert ensemble for constructing a large-scale emergency management model shall include at least climate experts, water conservancy experts, population exposure experts, material dispatching experts, and safety verification experts.

[0014] S4, Building an Expert Router

[0015] The expert router calculates the activation score of each expert in the expert set based on the sample features of the training samples; wherein, the activation score of expert e for sample x is calculated based on domain relevance, task relevance, risk relevance, context integrity, and activation cost, and is expressed as:

[0016] (1)

[0017] Among them, R domain (e,x) represents domain relevance, R task (e,x) represents task relevance, R risk (e,x) represents the risk correlation, Rcontext (e,x) represents context integrity, C(e) represents activation cost, and a, b, c, d, and λ are weighting coefficients.

[0018] S5. Determine the target expert set.

[0019] The target expert set is determined based on the activation scores through threshold selection, Top-K selection, budget constraint selection, or a combination thereof; when the training samples or inference inputs involve personnel transfer, opening of refuge sites, material allocation, or high-level early warning, the security verification experts are forcibly added to the target expert set.

[0020] S6, Implement expert-selective fine-tuning

[0021] Using a general open-source base model, training samples are generated and post-training fine-tuning is performed. For each training sample, only the expert modules and necessary shared parameters in the target expert set participate in parameter updates. The unselected expert modules remain frozen during the parameter update process corresponding to the training sample and do not participate in gradient updates, so as to reduce the negative transfer between multi-source heterogeneous disaster knowledge.

[0022] S7, Perform multi-task joint training

[0023] The emergency management model simultaneously learns one or more of the following tasks: disaster level prediction, risk interpretation generation, personnel evacuation strategy generation, and material dispatch suggestion generation, and imposes safety constraints on outputs related to personnel safety. Disaster level prediction employs either categorical loss or ordinal regression loss; risk interpretation generation, personnel evacuation strategy generation, and material dispatch suggestion generation employ either sequence generation loss, structured field generation loss, or rule matching loss; safety constraints are implemented through rejecting erroneous suggestions, evidence boundary checks, and penalties for violating contingency plan rules. The disaster level prediction calculation is shown in the following formula:

[0024] (4)

[0025] Among them, L total L represents the total loss during multi-task joint training. level L represents the predicted loss due to the disaster level. explain L represents the loss generated by risk interpretation. evac L represents the loss generated by personnel risk avoidance strategies. resource This indicates that the resource allocation suggestion generates a loss, L route L represents the expert routing regularization loss. safe This represents the safety constraint loss, where α1 to α6 are the task weight coefficients.

[0026] S8. Reason and generate results

[0027] During the inference phase, the expert router selects the appropriate expert module based on real-time input data, and the selected expert module outputs disaster risk assessment results, risk interpretations, and emergency response suggestions. Then, the security verification expert performs constraint checks on the high-risk suggestions.

[0028] Preferably, in step S1, the climate data includes one or more of the following: historical rainfall, future rainfall forecasts, short-term heavy rainfall indicators, wind speed, wind direction, air pressure, typhoon path, typhoon intensity, typhoon movement speed, landfall time, and meteorological warning level; the water conservancy data includes one or more of the following: river level, river flow, water level rise rate, warning water level difference, reservoir level, sluice gate and pump operation status, tide level, upstream inflow, and drainage capacity indicators; the population data includes one or more of the following: regional population density, daytime population distribution, nighttime population distribution, distribution of key population groups, schools, hospitals, elderly care institutions, underground spaces, community grids, and historical disaster-affected population information; the material data includes one or more of the following: location of refuge sites, refuge site capacity, food reserves, drinking water reserves, medical supplies, rescue vehicles, rescue teams, drainage equipment, sandbags, power generation equipment, and transportation capacity.

[0029] Preferably, in step S3, the expert set further includes one or more of the following: geographical risk experts, traffic accessibility experts, and emergency response plan experts; the geographical risk experts are used to process information on terrain elevation, low-lying areas, and historical waterlogging points; the traffic accessibility experts are used to process information on road accessibility and evacuation routes; the emergency response plan experts are used to process information on response levels, departmental responsibilities, and handling procedures; and the safety verification experts are used to perform constraint checks on high-risk emergency recommendations.

[0030] Preferably, in step S4, the sample features of the expert router include any one or more of the following: disaster type, data source, task type, regional risk characteristics, time window, prediction target, historical disaster similarity, risk threshold trigger status, and manual rule identifier.

[0031] Preferably, the domain relevance, task relevance, risk relevance, and contextual integrity are obtained from rule matching, threshold triggering, vector similarity, small neural network prediction results, or a combination thereof; wherein, the domain relevance is used to characterize the degree of matching between the input data source and the expert domain, the task relevance is used to characterize the degree of demand for expert capabilities in the current task, the risk relevance is used to characterize whether the sample triggers the disaster risk signal corresponding to the expert, and the contextual integrity is used to characterize whether the data required for the expert to perform the judgment is sufficient.

[0032] Preferably, the number of experts in the target expert set in step S5 is subject to a preset budget constraint, and the expert router prioritizes experts with higher activation scores and sufficient contextual data support when the budget constraint is met; when multiple expert outputs conflict with each other, the security verification expert outputs a prompt to modify, reject, or require manual confirmation.

[0033] Preferably, the expert selection fine-tuning in step S6 is implemented using an expert adapter, a low-rank adapter, prefix parameters, cue parameters, an expert subnetwork, or a combination thereof; for each training sample, only the parameters of the selected expert adapter or expert subnetwork are updated, and the parameters of the shared coding layer, shared decoding layer, normalization layer, or task header are updated; the general open-source base model is selected from Qwen3-32B, GLM-4.5, or DeepSeek-R1.

[0034] Preferably, the total loss of the multi-task joint training in step S7 includes one or more of the following: disaster level prediction loss, risk interpretation generation loss, personnel avoidance strategy generation loss, material scheduling suggestion generation loss, expert routing regularization loss, and safety constraint loss; the expert routing regularization loss is used to constrain the number of expert activations, balance the training frequency of different experts, and reduce the probability of irrelevant experts being activated.

[0035] Preferably, in step S8, during the reasoning phase, a structured risk summary is generated based on real-time meteorological, water conservancy, population, and material inputs, and the expert router selects the corresponding expert module. The selected expert module generates domain judgment, evidence description, and candidate suggestions, and generates the final disaster level, main risk causes, suggested evacuation areas, suggested evacuation time windows, and material scheduling suggestions. For outputs involving personnel safety, the safety verification expert checks whether the candidate suggestions meet the contingency plan rules, evacuation capacity constraints, road accessibility constraints, evidence boundary constraints, and manual approval constraints. If not, a prompt for modification, rejection, or manual confirmation is output.

[0036] On the other hand, the present invention provides a multi-source disaster expert hybrid fine-tuning system for large-scale emergency management models, comprising:

[0037] The data acquisition module is used to acquire multi-source data for emergency management, including one or more of the following: climate data, water conservancy data, population data, material data, geographic data, transportation data, historical disaster records, and emergency plan texts.

[0038] A training sample construction module is used to standardize, domain label, and task label the multi-source data, and construct training samples. Each training sample includes input data, domain identifier, disaster type, regional information, time window, task type, and supervision label.

[0039] An expert set construction module is used to construct an expert set for the emergency management big data model. The expert set includes at least climate experts, water conservancy experts, population exposure experts, material dispatching experts, and safety verification experts.

[0040] An expert router module is constructed to calculate the activation scores of each expert in the expert set based on the sample features of the training samples; wherein, the activation score of expert e for sample x is calculated based on domain relevance, task relevance, risk relevance, context integrity, and activation cost, and is expressed as:

[0041] (1)

[0042] Among them, R domain (e,x) represents domain relevance, R task (e,x) represents task relevance, R risk (e,x) represents the risk correlation, R context (e,x) represents context integrity, C(e) represents activation cost, and a, b, c, d, and λ are weighting coefficients.

[0043] The target expert set determination module is used to determine the target expert set based on the activation score through threshold selection, Top-K selection, budget constraint selection, or a combination thereof; when the training samples or inference inputs involve personnel transfer, opening of refuge sites, material allocation, or high-level early warning, the security verification experts are forcibly added to the target expert set;

[0044] The execution of the expert selection fine-tuning module is used to generate training samples through a general open-source base model and perform post-training fine-tuning. For each training sample, only the expert modules in the target expert set and the necessary shared parameters participate in the parameter update. The expert modules that are not selected remain frozen during the parameter update process corresponding to the training sample and do not participate in the gradient update, so as to reduce the negative transfer between multi-source heterogeneous disaster knowledge.

[0045] A multi-task joint training module is executed to enable the emergency management model to simultaneously learn one or more of the following tasks: disaster level prediction, risk interpretation generation, personnel evacuation strategy generation, and material dispatch suggestion generation, and to impose safety constraints on outputs involving personnel safety. The disaster level prediction uses classification loss or ordinal regression loss; the risk interpretation generation, personnel evacuation strategy generation, and material dispatch suggestion use sequence generation loss, structured field generation loss, or rule matching loss; safety constraints are implemented through rejecting erroneous suggestions, evidence boundary checks, and penalties for violating contingency plan rules. The total loss of the multi-task joint training is shown in the following formula:

[0046] (4)

[0047] Among them, L total L represents the total loss during multi-task joint training. level L represents the predicted loss due to the disaster level. explain L represents the loss generated by risk interpretation. evac L represents the loss generated by personnel risk avoidance strategies. resource This indicates that the resource allocation suggestion generates a loss, L route L represents the expert routing regularization loss. safe This represents the safety constraint loss, where α1 to α6 are the task weight coefficients.

[0048] The reasoning and result generation module is used during the reasoning phase. The expert router selects the appropriate expert module based on real-time input data, and the selected expert module outputs disaster risk assessment results, risk interpretations, and emergency response suggestions. Then, the security verification expert performs constraint checks on the high-risk suggestions.

[0049] Compared with the prior art, the present invention has the following beneficial effects:

[0050] 1. This invention transforms the fine-tuning process of the emergency management big data model from a unified parameter update to a domain expert-selective update, reducing knowledge interference between different data sources.

[0051] 2. This invention uses an expert router to dynamically select experts based on sample characteristics, enabling information from different fields such as climate, water conservancy, population, and materials to be processed by corresponding experts, thereby improving the model's ability to model multi-source disaster data.

[0052] 3. This invention supports joint training of multiple tasks such as disaster level prediction, risk interpretation, personnel evacuation strategies, and material dispatch suggestions, so that the model output includes not only prediction results, but also interpretable emergency response suggestions.

[0053] 4. This invention activates only relevant experts during the reasoning stage, which can reduce reasoning costs and improve model response efficiency in emergency scenarios.

[0054] 5. This invention introduces a safety verification expert or rule verification module to constrain and check high-risk suggestions involving personnel transfer, opening of refuge sites, and allocation of materials, thereby reducing the risk of generating unreliable emergency suggestions from large models.

[0055] 6. This invention retains expert activation results, input evidence, output suggestions, and verification results, facilitating the formation of an auditable emergency support decision-making chain. Attached Figure Description

[0056] Figure 1 This is a flowchart illustrating a multi-source disaster expert hybrid fine-tuning method for a large-scale emergency management model, as provided in an embodiment of the present invention.

[0057] Figure 2 This is a schematic diagram of the implementation process provided in an embodiment of the present invention.

[0058] Figure 3 A list of expert modules provided for embodiments of the present invention. Detailed Implementation

[0059] This invention provides a multi-source disaster expert hybrid fine-tuning method and system for large-scale emergency management models. The method divides multi-source data in emergency management into multiple domains, including climate, water resources, population, materials, geography, transportation, and contingency plans, and constructs corresponding expert modules for each domain. The system uses an expert router to select one or more specific expert modules to participate in the fine-tuning and inference of the large-scale model based on the disaster type, data source, task objective, regional characteristics, and risk signals of the input sample.

[0060] Unlike general MoE training methods that only aim at expert load balancing or computational efficiency, this invention binds expert routing with domain evidence, task objectives, and security constraints in emergency management. On the one hand, it reduces parameter interference between heterogeneous knowledge such as climate, water resources, population, and materials by freezing unselected experts and sample-level selective updates. On the other hand, it restricts the generation of unfounded high-risk suggestions involving personnel transfer, opening of shelters, and material allocation by forcibly activating security-verified experts and imposing rule constraints, thereby solving the problems of reliability and security in disaster decision-making.

[0061] The present invention will now be described in detail with reference to embodiments and accompanying drawings. However, it should be understood that the embodiments and drawings are for illustrative purposes only and do not constitute any limitation on the scope of protection of the present invention. All reasonable modifications and combinations included within the inventive spirit of the present invention fall within the scope of protection of the present invention.

[0062] Example 1

[0063] like Figure 1 As shown in the figure, this embodiment provides a multi-source disaster expert hybrid fine-tuning method for large-scale emergency management models. The method includes the following steps:

[0064] S1. Obtain multi-source data for emergency management, including one or more of the following: climate data, water conservancy data, population data, material data, geographic data, transportation data, historical disaster records, and emergency plan texts.

[0065] S2. Standardize, domain-label, and task-label the multi-source data to construct training samples. Each training sample includes input data, domain identifier, disaster type, regional information, time window, task type, and supervision label. Supervision labels may include disaster level, risk cause, affected area, personnel evacuation recommendations, material dispatch recommendations, or emergency response level.

[0066] S3. Constructing an expert ensemble for a large-scale emergency management model. This expert ensemble shall include at least climate experts, water conservancy experts, population exposure experts, and material dispatch experts. Optionally, it may also include geographical risk experts, transportation accessibility experts, emergency response plan experts, and safety verification experts.

[0067] S4. Construct an expert router. The expert router calculates the activation score of each expert based on the input features of the training samples, and selects the target expert set based on the activation scores. The input features include one or more of the following: disaster type, data source, task type, regional risk characteristics, time window, prediction target, and similarity to historical disasters.

[0068] S5. Based on the activation scores, determine the target expert set through threshold selection, Top-K selection, budget constraint selection, or a combination thereof; when the training samples or inference inputs involve personnel transfer, opening of refuge sites, material allocation, or high-level early warning, forcibly add the security verification experts to the target expert set;

[0069] S6. Perform expert selection-based fine-tuning. For each training sample, only the expert modules selected by the expert router and the necessary shared parameters are updated. Unselected expert modules do not participate in the parameter update of that sample, thereby reducing interference between knowledge from different domains.

[0070] S7. Perform multi-task joint training. The multi-tasks include disaster level prediction, risk interpretation generation, personnel evacuation strategy generation, and material dispatch suggestion generation. The system sets corresponding loss functions according to different tasks and combines the weighted losses of each task to form the overall training objective.

[0071] S8. During the inference phase, the expert router selects the appropriate expert module based on real-time input data. The activated expert then generates a disaster risk assessment, risk interpretation, and emergency response recommendations. For outputs involving personnel relocation, material allocation, or high-risk warnings, the system further invokes security verification experts or rule verification modules to perform constraint checks.

[0072] In step S3, regarding the construction of the expert set, in some implementations, the expert set includes, for example: Figure 3 The expert modules are listed in Table 1. Each expert module can be implemented using a low-rank adapter, prefix parameters, cue parameters, expert subnetworks, or a combination thereof, or it can be implemented using an independent lightweight model, toolchain, or retrieval enhancement module.

[0073] In step S4, the expert router is used to determine the set of experts who need to participate in training or inference at the sample level. The expert router can be implemented using a rule function, a lightweight classifier, a retrieval model, a neural network router, or a combination thereof. For a training sample x and an expert e, the expert router calculates an activation score score(e|x).

[0074] In some implementations, the activation score is determined by domain relevance, task relevance, risk relevance, contextual integrity, and activation cost. Domain relevance indicates the degree of matching between the sample input and the expert's domain; task relevance indicates whether the current task requires the expert's participation; risk relevance indicates whether the sample triggers the risk signal corresponding to the expert; contextual integrity indicates whether the required input data for the expert is sufficient; and activation cost indicates the computational cost required to invoke or train the expert.

[0075] (1)

[0076] Among them, R domain (e,x) represents domain relevance, R task (e,x) represents task relevance, R risk (e,x) represents the risk correlation, R context (e,x) represents context integrity, C(e) represents expert activation cost, and a, b, c, d, and λ are weighting coefficients.

[0077] In other implementations, the expert router can employ an activation score calculation method based on vector similarity. The system constructs sample representation vectors h respectively. x and expert domain representation vector h e Combining the task matching term T(e,x), the risk trigger term G(e,x), and the activation cost term C(e), the calculation formula is as follows:

[0078] (2)

[0079] Where η1 to η4 are weighting coefficients. This method is suitable for scenarios where there is domain description text, historical disaster cases, or expert description text. In some other implementations, the expert router can be implemented using a small neural network. The system represents the samples h. x Experts say h e Task characteristics f task Risk characteristics f risk and context integrity feature f ctx After being concatenated, the data is input into a multilayer perceptron to obtain:

[0080] (3)

[0081] Expert routers can determine the target expert set based on threshold selection, Top-K selection, budget constraint selection, or a combination thereof. When training samples involve personnel transfer, opening of refuge sites, allocation of supplies, or high-level early warnings, security verification experts are forcibly added to the target expert set.

[0082] In step S6, during the training phase, the system does not uniformly update all expert modules. Instead, it performs selective fine-tuning based on the target expert set output by the expert router. For each training sample, only the expert parameters and necessary shared parameters of the selected experts are updated. Unselected expert modules remain frozen to avoid interference from irrelevant samples.

[0083] Expert modules can be deployed as expert adapters within a large model, or as multiple low-rank adapters shared on a large model. For a sample x, its target expert set is denoted as S(x), and the model parameters include shared parameters θ_s and expert parameters θ_e. The target expert set is represented as:

[0084]

[0085] Training only on Θ upd The allowed update parameters in (x) perform gradient updates, Θ frozen The expert parameters that are not selected in (x) remain frozen and do not participate in the gradient update of that sample.

[0086] In some implementations, to prevent over-activation of a few experts, the system sets expert routing regularization terms to constrain the number of activated experts and the frequency of expert training. For emergency management scenarios, the system can also set mandatory activation rules and independent security constraint losses for security verification experts.

[0087] Step S7 involves multi-task joint training, supporting joint training for disaster level prediction, risk interpretation generation, personnel evacuation strategy generation, and material dispatch suggestion generation. Disaster level prediction can employ classification loss or ordinal regression loss; risk interpretation, evacuation strategies, and material dispatch suggestions can utilize sequence generation loss, structured field generation loss, or rule matching loss; safety constraints can be implemented using rejection of erroneous suggestions, evidence boundary checks, and penalties for violation of contingency plan rules. See the following formula:

[0088] (4)

[0089] Among them, L total L represents the total loss during multi-task joint training. level L represents the predicted loss due to the disaster level. explainL represents the loss generated by risk interpretation. evac L represents the loss generated by personnel risk avoidance strategies. resource This indicates that the resource allocation suggestion generates a loss, L route L represents the expert routing regularization loss. safe This represents the safety constraint loss, and α1 to α6 are the task weight coefficients.

[0090] In the reasoning phase of step S8, a structured risk summary is first constructed from real-time meteorological, water resources, population, and material inputs. The expert router selects the appropriate expert module based on the risk summary and task objectives. The activated experts generate domain judgments, evidence descriptions, and candidate suggestions, respectively. By summarizing the outputs of each expert, the final disaster level, main risk causes, recommended evacuation areas, recommended evacuation time windows, and material dispatch recommendations are generated.

[0091] For outputs involving personnel safety, safety verification experts check whether candidate suggestions meet contingency plan rules, evacuation capacity constraints, road accessibility constraints, evidence boundary constraints, and manual approval constraints. If a candidate suggestion does not meet the constraints, the system outputs a prompt to modify, reject, or require manual confirmation.

[0092] The system can record the expert activation results, input evidence, output suggestions, verification results and human feedback for each reasoning as an audit log, which can be used for subsequent model iteration and emergency management accountability.

[0093] Example 2

[0094] like Figure 2 This embodiment provides a specific implementation process of a multi-source disaster expert hybrid fine-tuning method for a large-scale emergency management model based on Embodiment 1 in a certain region:

[0095] During the multi-source emergency data access and sample construction phase, data is acquired from meteorological platforms, water conservancy monitoring platforms, population and facility databases, material management platforms, historical disaster databases, and emergency response plan databases. For a typhoon and rainstorm scenario in a certain area, climate data may include future rainfall forecasts, historical rainfall, typhoon paths, wind speeds, and air pressures; water conservancy data may include river levels, reservoir levels, flow rates, and water level rise rates; population data may include street or community population density, schools, hospitals, elderly care facilities, and underground spaces; and material data may include shelter capacity, rescue vehicles, drinking water, food, medical supplies, and drainage equipment.

[0096] The system aligns the above data according to region and time window to construct training samples. Each sample includes a region identifier, time window, disaster type, input features, task type, and supervision label. Supervision labels can be derived from historical disaster records, manual annotations, emergency response records, or contingency plan rules.

[0097] In the domain expert building phase

[0098] The system constructs multiple expert adapters based on a shared large language model. Climate experts fine-tune the system using weather forecasts, typhoon paths, and weather warning texts; water conservancy experts fine-tune the system using water levels, flow rates, warning thresholds, and water conservancy bulletins; population exposure experts fine-tune the system using population distribution, key facilities, and community risk descriptions; and material dispatch experts fine-tune the system using shelters, rescue teams, material inventory, and allocation records.

[0099] In another implementation, each expert may not be an internal module of the large model, but rather an external expert comprised of specialized prompts, a retrieval library, tool calls, and a lightweight model. Regardless of whether internal or external experts are used, the expert router selects the set of experts to participate in training or inference based on sample features.

[0100] During the expert routing and selective fine-tuning phase

[0101] For a training sample that includes a typhoon path, a 24-hour heavy rainfall forecast, and a rapid rise in river levels, the expert router assigns high activation scores to climate and water conservancy experts. If the task objective of the sample is to generate evacuation strategies and the affected area includes schools, hospitals, or densely populated communities, then population exposure and emergency response experts are also activated. If the output involves personnel relocation or the opening of shelters, then safety verification experts are forcibly activated.

[0102] During training, the system only updates the parameters of the activated experts. For example, for a sample containing only material dispatch records and shelter capacity, the system primarily updates the material dispatch expert and the emergency response plan expert, but not the climate expert and the water resources expert. In this way, knowledge from different domains is assigned to the corresponding experts, reducing interference caused by uniform fine-tuning.

[0103] In the disaster level prediction and risk avoidance strategy generation stage

[0104] During the reasoning process, the system receives a 24-hour rainfall forecast, real-time water level, historical flooding locations, population exposure, and shelter capacity for a given area. The expert router determines that the input involves climate, water resources, geographical risks, population exposure, and emergency response plans, thus activating the corresponding experts. The climate expert outputs the risk of continued heavy rainfall, the water resources expert outputs the risk of water levels exceeding warning levels, the geographical risk expert outputs the risk of low-lying areas, the population exposure expert outputs key relocation areas, and the emergency response plan expert outputs corresponding response actions.

[0105] By integrating the input from various experts, the system generates disaster levels and evacuation strategies. For example, the system might identify a region as high-risk, primarily due to forecasts of heavy rainfall, rapid water level rises, and overlap with historical flooding points. Recommendations include preparing for the evacuation of low-lying communities, verifying the capacity of shelters, and pre-positioning drainage equipment. Safety verification experts then check whether the recommendations exceed the authorized scope of the contingency plan, whether they lack evidence of road accessibility, and whether manual confirmation is required. Finally, the system outputs a prompt for whether the recommendations can be issued or manually reviewed.

[0106] During the security constraints and audit log phase

[0107] Emergency management outputs may affect personnel safety, therefore the system sets safety constraints for high-risk recommendations. If the model recommends mandatory evacuation of personnel, but the input lacks evidence of official warning levels, road accessibility, or evacuation capacity, safety verification experts will mark the recommendation as requiring manual confirmation, rather than outputting it as a definitive instruction directly.

[0108] The system records the input summary, expert activation scores, activated experts, each expert's output, fusion results, security verification results, and human feedback for each inference iteration. This audit log can be used for subsequent model evaluation, expert router optimization, and emergency management accountability.

[0109] The above embodiments are merely preferred embodiments of the present invention, and the scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A multi-source disaster expert hybrid fine-tuning method for large-scale emergency management models, characterized in that, The steps include: S1, Obtaining data S2. Acquire multi-source emergency management data, including one or more of the following: climate data, water resources data, population data, material data, geographic data, transportation data, historical disaster records, and emergency response plan texts; The multi-source data is standardized, domain-labeled, and task-labeled to construct training samples. Each training sample includes input data, domain identifier, disaster type, regional information, time window, task type, and supervision label; S3, construct an expert set. The expert ensemble for constructing a large-scale emergency management model shall include at least climate experts, water conservancy experts, population exposure experts, material dispatching experts, and safety verification experts. S4, Building an Expert Router The expert router calculates the activation score of each expert in the expert set based on the sample features of the training samples; wherein, the activation score of expert e for sample x is calculated based on domain relevance, task relevance, risk relevance, context integrity, and activation cost, and is expressed as: （1） Among them, R domain (e,x) represents domain relevance, R task (e,x) represents task relevance, R risk (e,x) represents the risk correlation, R context (e,x) represents context integrity, C(e) represents activation cost, and a, b, c, d, and λ are weighting coefficients; S5, Determine the target expert set. Based on the activation scores, a target expert set is determined through threshold selection, Top-K selection, budget constraint selection, or a combination thereof; when the training samples or inference inputs involve personnel transfer, opening of refuge sites, material allocation, or high-level early warning, the security verification experts are forcibly added to the target expert set; S6, Perform expert selection fine-tuning. Using a general open-source foundational large model, training samples are generated and post-training fine-tuning is performed. For each training sample, only expert modules and necessary shared parameters from the target expert set participate in parameter updates. Unselected expert modules remain frozen during the parameter update process corresponding to that training sample and do not participate in gradient updates, thereby reducing negative transfer between multi-source heterogeneous disaster knowledge; S7. Perform multi-task joint training. The emergency management model simultaneously learns one or more of the following tasks: disaster level prediction, risk interpretation generation, personnel evacuation strategy generation, and material dispatch suggestion generation, and imposes safety constraints on outputs related to personnel safety. Disaster level prediction employs either categorical loss or ordinal regression loss; risk interpretation generation, personnel evacuation strategy generation, and material dispatch suggestion generation employ either sequence generation loss, structured field generation loss, or rule matching loss; safety constraints are implemented through rejecting erroneous suggestions, evidence boundary checks, and penalties for violating contingency plan rules. The disaster level prediction calculation is shown in the following formula: （4） Among them, L total L represents the total loss during multi-task joint training. level L represents the predicted loss due to the disaster level. explain L represents the loss generated by risk interpretation. evac L represents the loss generated by personnel risk avoidance strategies. resource This indicates that the resource allocation suggestion generates a loss, L route L represents the expert routing regularization loss. safe This represents the safety constraint loss, where α1 to α6 are the task weight coefficients. S8. Reason and generate results During the inference phase, the expert router selects the appropriate expert module based on real-time input data, and the selected expert module outputs disaster risk assessment results, risk interpretations, and emergency response suggestions. Then, the security verification expert performs constraint checks on the high-risk suggestions.

2. The multi-source disaster expert hybrid fine-tuning method for large-scale emergency management models according to claim 1, characterized in that, In step S1, the climate data includes one or more of the following: historical rainfall, future rainfall forecasts, short-term heavy rainfall indicators, wind speed, wind direction, air pressure, typhoon path, typhoon intensity, typhoon movement speed, landfall time, and meteorological warning level; the water conservancy data includes one or more of the following: river level, river flow, water level rise rate, warning water level difference, reservoir level, sluice gate and pump operation status, tide level, upstream inflow, and drainage capacity indicators; the population data includes one or more of the following: regional population density, daytime population distribution, nighttime population distribution, distribution of key population groups, schools, hospitals, elderly care institutions, underground spaces, community grids, and historical disaster-affected population information; the material data includes one or more of the following: location of shelters, shelter capacity, food reserves, drinking water reserves, medical supplies, rescue vehicles, rescue teams, drainage equipment, sandbags, power generation equipment, and transportation capacity.

3. The multi-source disaster expert hybrid fine-tuning method for large-scale emergency management models according to claim 2, characterized in that, In step S3, the expert set further includes one or more of the following: geographical risk experts, transportation accessibility experts, and emergency response plan experts; the geographical risk experts are used to process information on terrain elevation, low-lying areas, and historical waterlogging points; the transportation accessibility experts are used to process information on road accessibility and evacuation routes; the emergency response plan experts are used to process information on response levels, departmental responsibilities, and handling procedures; and the safety verification experts are used to perform constraint checks on high-risk emergency recommendations.

4. The multi-source disaster expert hybrid fine-tuning method for large-scale emergency management models according to claim 3, characterized in that, In step S4, the sample features of the expert router include any one or more of the following: disaster type, data source, task type, regional risk characteristics, time window, prediction target, historical disaster similarity, risk threshold trigger status, and manual rule identifier.

5. The multi-source disaster expert hybrid fine-tuning method for large-scale emergency management models according to claim 4, characterized in that, The domain relevance, task relevance, risk relevance, and contextual integrity are obtained by rule matching, threshold triggering, vector similarity, small neural network prediction results, or a combination thereof; wherein, the domain relevance is used to characterize the degree of matching between the input data source and the expert domain, the task relevance is used to characterize the degree of demand for expert capabilities in the current task, the risk relevance is used to characterize whether the sample triggers the disaster risk signal corresponding to the expert, and the contextual integrity is used to characterize whether the data required for the expert to perform the judgment is sufficient.

6. The multi-source disaster expert hybrid fine-tuning method for large-scale emergency management models according to claim 5, characterized in that, The number of experts in the target expert set mentioned in step S5 is subject to a preset budget constraint, and the expert router prioritizes experts with higher activation scores and sufficient contextual data support when the budget constraint is met; when multiple expert outputs conflict with each other, the security verification expert outputs a prompt to modify, reject, or require manual confirmation.

7. The multi-source disaster expert hybrid fine-tuning method for large-scale emergency management models according to claim 6, characterized in that, The expert selection fine-tuning described in step S6 is implemented using an expert adapter, a low-rank adapter, prefix parameters, cue parameters, an expert subnetwork, or a combination thereof. For each training sample, only the parameters of the selected expert adapter or expert subnetwork are updated, and the parameters of the shared coding layer, shared decoding layer, normalization layer, or task header are also updated. The general open-source base model is selected from Qwen3-32B, GLM-4.5, or DeepSeek-R1.

8. The multi-source disaster expert hybrid fine-tuning method for large-scale emergency management models according to claim 7, characterized in that, The total loss of the multi-task joint training in step S7 includes one or more of the following: disaster level prediction loss, risk interpretation generation loss, personnel evacuation strategy generation loss, material scheduling suggestion generation loss, expert routing regularization loss, and safety constraint loss; the expert routing regularization loss is used to constrain the number of activated experts, balance the training frequency of different experts, and reduce the probability of irrelevant experts being activated.

9. The multi-source disaster expert hybrid fine-tuning method for large-scale emergency management models according to claim 8, characterized in that, In step S8, during the reasoning phase, a structured risk summary is generated based on real-time meteorological, water conservancy, population, and material inputs. The expert router then selects the corresponding expert module. The selected expert module generates domain judgment, evidence description, and candidate suggestions, and generates the final disaster level, main risk causes, suggested evacuation areas, suggested evacuation time windows, and material scheduling suggestions. For outputs involving personnel safety, safety verification experts check whether candidate suggestions meet the contingency plan rules, evacuation capacity constraints, road accessibility constraints, evidence boundary constraints, and manual approval constraints; if not, they output prompts for modification, rejection, or manual confirmation.

10. A multi-source disaster expert hybrid fine-tuning system for large-scale emergency management models, characterized in that, include: The system includes the following modules: a data acquisition module for acquiring multi-source emergency management data, including climate data, water resources data, population data, material data, geographic data, transportation data, historical disaster records, and emergency plan texts; a training sample construction module for standardizing, domain-specifically labeling, and task-specifically labeling the multi-source data to construct training samples, each including input data, domain identifier, disaster type, regional information, time window, task type, and supervision label; an expert ensemble construction module for constructing an expert ensemble for the emergency management large-scale model, including at least climate experts, water resources experts, population exposure experts, material dispatching experts, and safety verification experts; and an expert router construction module for calculating the activation scores of each expert in the expert ensemble based on the sample characteristics of the training samples. The activation score of expert e for sample x is calculated based on domain relevance, task relevance, risk relevance, contextual integrity, and activation cost, and is expressed as follows: （1） Among them, R domain (e,x) represents domain relevance, R task (e,x) represents task relevance, R risk (e,x) represents the risk correlation, R context (e,x) represents context integrity, C(e) represents activation cost, and a, b, c, d, and λ represent weighting coefficients. A target expert set determination module is used to determine the target expert set based on the activation scores through threshold selection, Top-K selection, budget constraint selection, or a combination thereof. When training samples or inference inputs involve personnel transfer, opening of refuge sites, material allocation, or high-level early warnings, the safety verification experts are forcibly added to the target expert set. An expert selection fine-tuning module is used to apply generated training samples through a general open-source base model and perform post-training fine-tuning. For each training sample, only expert modules and necessary shared parameters in the target expert set participate in parameter updates. Unselected expert modules remain frozen during the parameter update process corresponding to the training sample and do not participate in gradient updates, thereby reducing negative transfer between multi-source heterogeneous disaster knowledge. A multi-task joint training module is executed to enable the emergency management model to simultaneously learn one or more of the following tasks: disaster level prediction, risk interpretation generation, personnel evacuation strategy generation, and material dispatch suggestion generation, and to impose safety constraints on outputs involving personnel safety. The disaster level prediction uses classification loss or ordinal regression loss; the risk interpretation generation, personnel evacuation strategy generation, and material dispatch suggestion use sequence generation loss, structured field generation loss, or rule matching loss; safety constraints are implemented through rejecting erroneous suggestions, evidence boundary checks, and penalties for violating contingency plan rules. The total loss of the multi-task joint training is shown in the following formula: （4） Among them, L total L represents the total loss during multi-task joint training. level L represents the predicted loss due to the disaster level. explain L represents the loss generated by risk interpretation. evac L represents the loss generated by personnel risk avoidance strategies. resource This indicates that the resource allocation suggestion generates a loss, L route L represents the expert routing regularization loss. safe The safety constraint loss is represented by α1 to α6, which are task weight coefficients. The reasoning and result generation module is used in the reasoning stage. The expert router selects the corresponding expert module based on the real-time input data, and the selected expert module outputs the disaster risk assessment results, risk interpretation and emergency response suggestions. Then, the safety verification expert performs constraint checks on the high-risk suggestions.

Images