Emergency rescue resource scheduling method based on Internet of Things

Abstract

The invention discloses an emergency rescue resource scheduling method based on the Internet of Things, and the method comprises the steps: collecting disaster area environment parameters, images and personnel distribution data through Internet of Things equipment, and associating historical disaster cases to construct a multi-source heterogeneous data set; extracting disaster characteristics through preprocessing, dynamically distributing data source weights, and generating a disaster assessment matrix; outputting a resource demand peak value based on the time sequence prediction model and performing dynamic correction; constructing a deep reinforcement learning model optimization decision strategy, and generating a scheduling scheme in combination with priority matching and an improved A * algorithm; according to the method, the disaster sensing precision and the demand prediction accuracy are improved, the scheduling decision adaptability is enhanced, the response time of a high-priority region is shortened, and the efficient demand of emergency rescue in a complex disaster is met.

Description

Technical Field

[0001] This invention relates to the field of emergency rescue resource scheduling technology, specifically to an emergency rescue resource scheduling method based on the Internet of Things. Background Technology

[0002] In recent years, natural disasters (such as earthquakes, floods, and typhoons) and public safety incidents have occurred frequently around the world. Emergency rescue, as a key link in ensuring the safety of life and property, has an increasingly urgent need for "rapid response, precise dispatch, and dynamic adaptation." With the gradual application of technologies such as the Internet of Things and artificial intelligence in the emergency field, the traditional resource dispatch model that relies on human experience is no longer able to cope with complex disasters. Current emergency rescue scenarios are characterized by "wide coverage of disasters, complex data types, and rapid changes." They need to process multi-dimensional information such as environmental parameters, personnel distribution, and infrastructure status at the same time. However, traditional methods lack a systematic technical framework to integrate multi-source data, resulting in decision-making efficiency and accuracy that cannot meet the needs of actual combat. Currently, the field of emergency rescue resource dispatch faces key problems such as incomplete and lagging data collection, weak multi-source information integration capabilities, static and rigid demand forecasting, insufficient decision-making adaptability, and lack of dynamic correction in execution. On the one hand, traditional dispatching relies heavily on manual experience or single-type data (such as collecting only environmental parameters), making it difficult to simultaneously acquire multi-source heterogeneous data such as environment, images, and personnel distribution. Furthermore, it does not connect with historical disaster case experience, resulting in low accuracy of disaster assessment and strong subjectivity in decision-making. On the other hand, resource demand forecasting uses static models, which cannot be dynamically corrected according to real-time disaster conditions, easily leading to supply and demand mismatches. At the same time, the lack of trajectory tracking and deviation correction mechanisms in the execution of dispatching plans makes it difficult to adjust in a timely manner in the face of emergencies such as road interruptions and secondary disasters. Ultimately, this results in resource shortages in high-priority areas and delays in rescue response, making it difficult to meet the core requirements of "rapid, accurate, and dynamic" emergency rescue. Summary of the Invention

[0003] The technical solution of this invention is implemented as follows: an emergency rescue resource scheduling method based on the Internet of Things, comprising: S1. Obtain environmental parameters, images, and personnel distribution data of the disaster area through IoT devices, construct multi-source heterogeneous data, and call the historical disaster case database to perform data association with the multi-source heterogeneous data; S2. Preprocess the multi-source heterogeneous data, extract disaster characteristics, dynamically allocate the weights of each data source according to the disaster type, align the processed data in time and space, and generate a disaster assessment matrix. S3. Based on the disaster assessment matrix, the peak resource demand of each disaster-stricken area in the future is output using a time-series prediction model. At the same time, the model parameters are corrected by combining the deviation between the actual demand and the predicted value to generate resource demand prediction results. S4. Based on the resource demand prediction results, construct a deep reinforcement learning model including a state space, action space, and dynamic reward function, and generate an optimized decision strategy by offline training and online fine-tuning. S5. Based on the optimization decision-making strategy, a resource scheduling scheme is generated by combining the priority matching rules of resources and tasks with real-time path planning. S6. Based on the execution data of the resource scheduling scheme, the rescue trajectory is tracked in real time. When the deviation exceeds the threshold, rescheduling is triggered, and a closed-loop optimization mechanism is formed based on the evaluation data.

[0004] Furthermore, in step S1, environmental parameters, images, and personnel distribution data of the disaster area are acquired through IoT devices to construct multi-source heterogeneous data, and the established historical disaster case database is called to perform data association with the multi-source heterogeneous data; Furthermore, in step S1, the specific steps are as follows: An IoT sensing array was constructed, and environmental acquisition devices, image acquisition devices, and positioning acquisition devices were deployed throughout the disaster area to acquire physical environmental parameters, visualization information, and spatial distribution information of people in the disaster area, forming a multi-source heterogeneous raw dataset. For the aforementioned multi-source heterogeneous raw data, outlier removal and data smoothing are performed on environmental parameters, sharpness enhancement, distortion correction and target recognition are performed on image data, and position coordinate calibration and density statistics are performed on personnel distribution data to obtain processed multi-source heterogeneous data. The pre-established historical disaster case database is invoked, and the processed multi-source heterogeneous data is compared with the key parameters in the historical cases using a feature matching algorithm. Historical cases with similarity exceeding a preset threshold are selected, and an association index between real-time data and historical cases is established to form an associated dataset that includes original sensing data and historical reference information.

[0005] Furthermore, in step S2, the multi-source heterogeneous data is preprocessed to extract disaster features, the weights of each data source are dynamically allocated according to the disaster type, the processed data is spatiotemporally aligned, and a disaster assessment matrix is ​​generated. Furthermore, in step S2, the specific steps are as follows: The multi-source heterogeneous data is preprocessed in real time using edge computing nodes. The 3σ criterion is used to identify and remove outliers for environmental parameters, and the mean of nearby time times is used to fill in missing values. The image data is processed using a semantic segmentation algorithm to identify and extract key disaster features, and the extracted results are converted into a binary disaster heat map. Based on the personnel distribution data, terrain information from the digital elevation model is introduced. The positioning deviation caused by terrain occlusion is corrected through coordinate mapping. Clustering algorithms are used to identify densely populated areas, and the number of trapped people is estimated by combining the positioning signal strength. Based on the coordinate system of the geographic information system, the preprocessed environmental parameters, the binarized disaster heat map and the corrected personnel distribution data are spatiotemporally aligned according to the geographical coordinates to construct a three-dimensional data cube including time, space and disaster characteristics. An adaptive weight allocation mechanism is introduced to dynamically adjust the weights of each data source for different disaster types. By weighted fusion of information from a three-dimensional data cube, a multi-dimensional disaster assessment matrix is ​​generated, including environmental hazard level, infrastructure damage level, and the status of trapped personnel.

[0006] Furthermore, in step S3, based on the disaster assessment matrix, the peak resource demand of each disaster-stricken area in the future is output using a time-series prediction model. At the same time, the model parameters are corrected by combining the deviation between the actual demand and the predicted value to generate resource demand prediction results. Furthermore, in step S3, the specific steps are as follows: Using the environmental hazard level, infrastructure damage level, and personnel entrapment status characteristic parameters in the disaster assessment matrix as input, a time series prediction model based on LSTM is constructed. A sliding time window is set to capture the time series characteristics of the data, and the peak resource demand of each disaster-stricken area in the future is output. The resource demand includes the quantity of medical supplies, the configuration of rescue personnel, and the priority of equipment deployment. Real-time data on actual resource demand in various disaster-stricken areas is collected, the mean square error and mean absolute error between actual and predicted values ​​are calculated, and the hyperparameters of the time series prediction model are dynamically adjusted through a Bayesian optimization algorithm. A deviation feedback mechanism is established, in which the latest actual demand and prediction deviation data are backpropagated to the time series prediction model at preset intervals, and the model weight parameters are updated by gradient descent to correct the trend deviation of the prediction curve. By integrating forecast results from multiple time scales with bias correction parameters, resource demand forecast results are generated, including demand type, quantity, time point, and confidence level.

[0007] Furthermore, in step S4, based on the resource demand prediction results, a deep reinforcement learning model including a state space, an action space, and a dynamic reward function is constructed, and an optimized decision strategy is generated through offline training and online fine-tuning. Furthermore, in step S4, the specific steps are as follows: Using the resource demand forecast results as input, a multi-dimensional state space of a deep reinforcement learning model is constructed. The multi-dimensional state space includes the location coordinates of the disaster area, the real-time inventory of resources, the demand priority of each area, the road traffic status, the operating condition of transportation equipment, the weather impact coefficient, and the personnel safety risk index. Define the combined action space of the deep reinforcement learning model. The combined action space consists of three basic actions: resource allocation, path planning, and task switching. Resource allocation includes different types and quantities of resource allocation schemes, path planning includes alternative route selection and order adjustment, and task switching includes dynamic changes in scheduling priority. A dynamic reward function is set up, which comprehensively considers the timeliness of resource arrival, the efficiency of demand fulfillment, the safety of rescue personnel, and the avoidance of route risks. The deep reinforcement learning model is trained and optimized using a proximal policy optimization algorithm. The model, which includes the multi-dimensional state space, combined action space and dynamic reward function, is trained offline for multiple rounds using historical disaster relief data. The policy network parameters are iteratively updated by simulating different disaster scenarios until the model converges to the preset accuracy. By combining real-time disaster data and resource status, the trained model is fine-tuned online, and a near-end strategy optimization algorithm is used to dynamically adjust the decision threshold and generate an optimized decision strategy.

[0008] Furthermore, in step S5, a resource scheduling scheme is generated based on the optimization decision strategy, combining the priority matching rules of resources and tasks with real-time path planning. Furthermore, in step S5, the specific steps are as follows: Based on the aforementioned optimization decision-making strategy, the core parameters of resource scheduling are extracted, including the resource demand type, allocation quantity, scheduling priority ranking, and time window constraints of each disaster-stricken area, to form a preliminary scheduling instruction set; Establish a priority matching mechanism for resources and tasks. Using the population density, proportion of seriously injured, length of time trapped, degree of infrastructure damage, and existing resource reserves in the disaster-stricken area as input parameters, determine the weight of each parameter through the analytic hierarchy process, calculate the comprehensive priority coefficient of each area by weighting, and determine the order of resource allocation by sorting the coefficients from high to low. A real-time path planning model is constructed, using the improved A* algorithm as the core algorithm. The road traffic status, road surface bearing capacity, meteorological conditions, terrain slope and potential secondary disaster risk areas are used as path cost evaluation indicators. The shortest time path from each resource reserve point to the target area is calculated, and two or more alternative paths are generated simultaneously. The spatial separation degree between the alternative paths and the main path is not less than a preset threshold. Integrate scheduling instructions, priority ranking, and path planning results to generate a resource scheduling scheme.

[0009] Furthermore, in step S6, based on the execution data of the resource scheduling scheme, the rescue trajectory is tracked in real time, and rescheduling is triggered when the deviation exceeds the threshold, and a closed-loop optimization mechanism is formed based on the evaluation data; Furthermore, in step S6, the specific steps are as follows: The real-time location, speed, and status information of resource transportation are collected by IoT positioning devices and vehicle terminals. Combined with electronic fence technology, the transportation trajectory and deployment progress of rescue resources are tracked to generate a real-time execution data dashboard. The real-time execution data dashboard includes the actual departure time, current location, estimated remaining time, and the amount of resources deployed. Set time, quantity, and path deviation thresholds, where the time deviation threshold is the maximum allowable difference between the actual arrival time and the planned arrival time of the resources, the quantity deviation threshold is the allowable fluctuation ratio between the actual quantity of resources deployed and the planned quantity of resources deployed, and the path deviation threshold is the maximum spatial deviation distance between the actual driving route and the planned route. Establish a deviation monitoring mechanism to compare the execution data with the planned parameters of the scheduling scheme in real time. When any deviation value exceeds the corresponding threshold, the rescheduling process is automatically triggered. The rescheduling process includes recalculating the priority coefficient, adjusting the transportation route, and optimizing the resource allocation ratio of subsequent batches. A scheduling effectiveness evaluation system is constructed, with evaluation indicators including demand satisfaction rate, average response time, resource utilization rate, and transportation safety coefficient. A comprehensive evaluation score is obtained through weighted calculation. The comprehensive evaluation score and the corresponding scheduling strategy are used as training samples and fed back to the deep reinforcement learning model to update the model's state space parameters and reward function weights, forming a closed-loop mechanism from data collection and scheme execution to model optimization.

[0010] The advantages of the Internet of Things-based emergency rescue resource scheduling method described in this invention are as follows: This invention discloses an emergency rescue resource scheduling method based on the Internet of Things (IoT). It achieves synchronous multi-source data collection through the deployment of a global IoT sensing array, and enhances disaster perception and data integration capabilities by combining historical disaster case correlation analysis, thus solving the problems of incomplete data and low assessment accuracy in traditional methods. Based on an LSTM time-series prediction model and a deviation feedback mechanism, it enables dynamic correction and prediction of resource demand, avoiding supply-demand mismatch caused by static prediction. A deep reinforcement learning model is constructed and trained offline and fine-tuned online to optimize the adaptability of decisions to different disaster types, balancing resource timeliness and personnel safety. Through real-time trajectory tracking, deviation threshold monitoring, and a rescheduling mechanism, it ensures the flexibility of plan execution. Combined with scheduling effect evaluation, a closed-loop optimization is formed, ultimately significantly improving the accuracy and response speed of resource scheduling, reducing the risk of rescue delays, and meeting the high-efficiency needs of emergency rescue under complex disaster situations. Attached Figure Description

[0011] Figure 1 This is a flowchart of an emergency rescue resource scheduling method based on the Internet of Things in an embodiment of the present invention; Figure 2 This is a sub-flowchart of an emergency rescue resource scheduling method based on the Internet of Things in an embodiment of the present invention. Detailed Implementation

[0012] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0013] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0014] Please see Figures 1-2 As shown in the embodiment of the present invention, an emergency rescue resource scheduling method based on the Internet of Things includes: S1. Obtain environmental parameters, images, and personnel distribution data of the disaster area through IoT devices, construct multi-source heterogeneous data, and call the historical disaster case database to perform data association with the multi-source heterogeneous data; S2. Preprocess the multi-source heterogeneous data, extract disaster characteristics, dynamically allocate the weights of each data source according to the disaster type, align the processed data in time and space, and generate a disaster assessment matrix. S3. Based on the disaster assessment matrix, the peak resource demand of each disaster-stricken area in the future is output using a time-series prediction model. At the same time, the model parameters are corrected by combining the deviation between the actual demand and the predicted value to generate resource demand prediction results. S4. Based on the resource demand prediction results, construct a deep reinforcement learning model including a state space, action space, and dynamic reward function, and generate an optimized decision strategy by offline training and online fine-tuning. S5. Based on the optimization decision-making strategy, a resource scheduling scheme is generated by combining the priority matching rules of resources and tasks with real-time path planning. S6. Based on the execution data of the resource scheduling scheme, the rescue trajectory is tracked in real time. When the deviation exceeds the threshold, rescheduling is triggered, and a closed-loop optimization mechanism is formed based on the evaluation data.

[0015] like Figure 2 As shown, in step S1, environmental parameters, images, and personnel distribution data of the disaster area are obtained through IoT devices to construct multi-source heterogeneous data, and the historical disaster case database is called to associate the multi-source heterogeneous data with the data.

[0016] Specifically, in this embodiment, when constructing the IoT sensing array, three types of data acquisition devices are deployed layer by layer across the entire area according to the terrain features and disaster impact range of the disaster zone: environmental data acquisition devices include temperature and humidity sensors, air pressure sensors, and wind speed and direction sensors deployed on the ground surface; soil moisture sensors buried underground; and water quality monitoring sensors and water level sensors deployed in water bodies. These devices form a grid-like monitoring network with intervals of 500 meters, collecting physical environmental parameters such as temperature, humidity, air pressure, wind speed, wind direction, soil moisture content, pH value, dissolved oxygen concentration, and water level in real time; image data acquisition devices include those equipped with high-definition cameras... A swarm of drones (configured at a density of one drone per square kilometer, with a cruising altitude set at 100 meters), panoramic surveillance cameras fixed at high points on buildings, and portable, image-stabilized cameras carried by rescue personnel acquire visual information about building damage, road blockages, and trapped individuals from aerial, fixed, and moving perspectives, respectively. Positioning-related data acquisition devices combine BeiDou positioning terminals with emergency beacons, sending location signals to trapped individuals every 30 seconds via the emergency beacons. Simultaneously, base station signal analysis and drone thermal imaging recognition technology supplement the data by acquiring the location information of individuals without positioning devices, forming spatial distribution data of the population. Data collected by these three types of devices is transmitted back in real-time via a 5G private network, forming a multi-source heterogeneous raw dataset. For multi-source heterogeneous raw data, the following processing methods were adopted: In environmental parameter processing, through sensor self-test signals and data continuity analysis, obviously invalid data was removed, such as fixed values ​​in the offline state of sensors and blank values ​​caused by data transmission interruption. The remaining data was smoothed using the sliding window mean method, with the window size set to 5 minutes to eliminate high-frequency noise interference. In image data processing, the Retinex algorithm was first used to enhance image clarity, then distortion correction was performed based on the camera intrinsic parameter matrix, and finally the YOLOv5 target detection algorithm was used to identify targets such as buildings, roads, vehicles, and people in the image, and the outlines of damaged buildings and areas where people were trapped were marked. In personnel distribution data processing, combined with the high-precision electronic map of the disaster area, the positioning coordinates were Gaussian projection transformation to calibrate the positional deviation, and the kernel density estimation algorithm was used to generate a personnel distribution heat map, and the number and density of people in each 100m × 100m grid were counted. A pre-established historical disaster case database is invoked, which stores disaster cases such as earthquakes, floods, and typhoons that have occurred in the past 10 years. Each case includes information such as disaster type, occurrence time, impact range, environmental parameter change curves, on-site image features, population distribution, and rescue resource dispatch records. Through feature matching algorithms, the processed multi-source heterogeneous data is compared with the historical case data: key features such as temperature change rate and peak wind speed in environmental parameters, building damage rate and road blockage length in image data, and number of densely populated areas in population distribution data are extracted. The cosine similarity between the current data and each historical case is calculated, and 3-5 historical cases with a similarity of more than 85% are selected. An association index between real-time data and historical cases is established to form an associated dataset containing raw sensing data, processed data, and matched historical case information, providing data support for subsequent disaster assessment. like Figure 1 As shown, in step S2, the multi-source heterogeneous data is preprocessed, disaster features are extracted, the weights of each data source are dynamically allocated according to the disaster type, the processed data is spatiotemporally aligned, and a disaster assessment matrix is ​​generated.

[0017] Specifically, in this embodiment, edge computing devices deployed at the edge nodes of the disaster area are used to perform localized real-time preprocessing of multi-source heterogeneous data in the associated dataset: For environmental parameter data, the 3σ criterion is used for outlier identification and removal. First, the mean μ and standard deviation σ of each environmental parameter are calculated, and data with values ​​outside the interval [μ-3σ, μ+3σ] are identified as outliers and removed; for missing values ​​that occur during the data acquisition process, the sliding window method is used to select the monitoring mean of the same type of sensor within the nearest 10 minutes to fill in the missing values, ensuring the continuity of the environmental parameter time series; The U-Net semantic segmentation algorithm is used to process image data. Through a pre-trained disaster scene recognition model, pixel-level classification is performed on the input drone aerial images, fixed monitoring images, and images taken by portable devices to identify and extract key disaster features such as damaged building areas, road interruptions, water accumulation areas, and hazardous chemical leak points. The extracted feature areas are mapped to a grid layer according to geographic coordinates to generate a binary disaster heat map, in which the pixel value of the feature area is set to 1 and the pixel value of the non-feature area is set to 0, which intuitively presents the spatial distribution of the disaster. To address the personnel distribution data, a 1:500 scale digital elevation model (DEM) of the disaster area was introduced to extract information such as terrain slope and altitude. A coordinate mapping algorithm was used to match the personnel location coordinates with the DEM data, correcting positioning deviations caused by mountain obstructions and building shielding, ensuring location accuracy within 5 meters. The DBSCAN clustering algorithm was used to perform density clustering on the corrected personnel coordinates, setting clustering parameters of 50 meters radius and a minimum of 5 people per cluster to identify densely populated areas. Combining the positioning signal strength with the cluster area, a model for estimating the number of trapped individuals was established. For every 10 dBm increase in signal strength, the corresponding correction coefficient for the number of people increased by 0.1, and for every 100 square meters increase in area, the baseline number of people increased by 5. Based on the WGS84 coordinate system of Geographic Information System (GIS), the preprocessed environmental parameters, binarized disaster heat map, and corrected personnel distribution data are spatiotemporally aligned. In the time dimension, using the standard timestamp of Beidou satellite time synchronization as the benchmark, all data are unified to a time series with a 1-minute interval. In the spatial dimension, the disaster area is divided into 10m × 10m grid units, and the environmental parameter monitoring values, disaster heat values, and personnel density data are mapped to the corresponding grids respectively, constructing a three-dimensional data cube containing time, longitude, latitude, and disaster characteristic attributes, realizing accurate correspondence of multi-source data in the spatiotemporal dimensions. An adaptive weight allocation mechanism is introduced to assign differentiated weight proportions to image data (F1), personnel data (F2), and environmental parameters (F3) based on the characteristics of different disaster types. The sum of the weights of the three types of data satisfies the normalization constraint. For example, when an earthquake occurs, the extent of building damage and the risk of people being trapped directly determine the rescue priority; therefore, image data (F1) weights are important. Set to 0.3, focusing on building damage identification results; personnel data (F2) weight. Set to 0.25, focusing on the distribution and number of trapped personnel; environmental parameter (F3) weight. The value is set to 0.45 to cover environmental factors such as aftershocks and secondary disasters. When a flood disaster occurs, environmental parameters such as water level changes and water flow velocity play a dominant role in the spread of the disaster; therefore, the weights are the image data (F1) weights. Set to 0.2, focusing on flooded area identification; personnel data (F2) weight. Set to 0.2, focusing on locating trapped personnel; environmental parameter (F3) weight. The value is set to 0.6, which covers key indicators such as water level and dam pressure. For typhoon disasters, environmental parameters such as wind speed, wind direction, and rainfall directly affect the feasibility of rescue efforts; therefore, image data (F1) weighting is crucial. Set to 0.25, focusing on identifying damage to power and communication facilities; personnel data (F2) weight. Set to 0.2, focusing on the distribution of people in the safe zone; environmental parameter (F3) weight. Set to 0.55, covering core parameters such as wind speed, air pressure, and rainfall; When generating the disaster assessment matrix, the value of any element M(i,j,k) in the matrix is ​​calculated by weighted summation of image data F1(i,j,k), personnel data F2(i,j,k), and environmental parameters F3(i,j,k). In the element value M(i,j,k), i and j are spatial dimension parameters, corresponding to the coordinates of the grid unit in the disaster area division, used to locate the specific geographical area. k is the assessment dimension parameter. When k=1, it corresponds to the environmental hazard level dimension; when k=2, it corresponds to the infrastructure damage level dimension; and when k=3, it corresponds to the personnel entrapment status dimension. The value of M(i,j,k) is the disaster quantitative score of the grid unit (i,j) on the assessment dimension k. F1(i,j,k), F2(i,j,k), and F3(i,j,k) are the standardized values ​​of image data, personnel data, and environmental parameters at their respective locations and dimensions, respectively. F1(i,j,k) is the standardized value of image data (such as the feature value of grid (i,j) in a binary disaster heat map) on the evaluation dimension k; F2(i,j,k) is the standardized value of personnel data (such as the estimated number of people trapped in grid (i,j)) on the evaluation dimension k; and F3(i,j,k) is the standardized value of environmental parameters (such as wind speed, water level, and other parameters in grid (i,j)) on the evaluation dimension k. All three are transformed to the [0,1] interval through range standardization to facilitate weighted fusion of cross-type data. The specific calculation is as follows:

[0018] Where M(i,j,k) represents the element value at a specific location in the disaster assessment matrix, and F1(i,j,k), F2(i,j,k), and F3(i,j,k) are the standardized values ​​of image data, personnel data, and environmental parameters at their respective locations and dimensions, respectively. , , These represent the weighting coefficients for image data, personnel data, and environmental parameters, respectively. The disaster assessment matrix, calculated using the weighted fusion formula described above, can quantitatively reflect the severity of the disaster in each grid unit across different dimensions.

[0019] like Figure 1As shown, in step S3, based on the disaster assessment matrix, the peak resource demand of each disaster-stricken area in the future is output using a time-series prediction model. At the same time, the model parameters are corrected by combining the deviation between the actual demand and the predicted value, and the resource demand prediction result is generated.

[0020] Specifically, in this embodiment, based on the disaster assessment matrix generated in step S2, accurate prediction of resource demand is achieved through a time-series prediction model and a deviation correction mechanism. The quantified feature parameters in the disaster assessment matrix are used as model inputs, specifically including environmental hazard level, infrastructure damage level, and the number of people trapped. The environmental hazard level is determined by an environmental hazard score (0-10 points, with 10 being the highest hazard) for each grid cell (10m x 10m), covering dynamic indicators such as aftershock intensity, flood level rise, and typhoon wind speed. Infrastructure damage level is quantified by road blockage rate (0-100%), building damage rate (0-100%), and the percentage of critical facilities (such as hospitals and communication base stations) that have lost functionality (0-100%) within the grid cell. The number of people trapped includes the estimated number of people trapped within the grid cell (calculated using DBSCAN clustering and signal strength correction) and the percentage of seriously injured (0-100%). Based on the average duration of being trapped (in minutes), a time-series prediction model based on LSTM (Long Short-Term Memory) was constructed. The input layer has 128 neurons, receiving the time-series sequences of the three types of feature parameters mentioned above. The hidden layer consists of two LSTM units, each with 64 neurons, using the ReLU activation function and a dropout rate of 0.3 to avoid overfitting. The output layer has three neurons, corresponding to the peak demand prediction values ​​for medical supplies, rescue personnel, and equipment, respectively. The sliding time window is set to 60 minutes, that is, using the disaster feature sequence of the past 60 minutes (one data point every 5 minutes, for a total of 12 time steps), the peak resource demand of each disaster area (divided by administrative blocks) in the next 12 hours is predicted, with a time granularity of 1 hour. Then, actual demand data collection and error calculation are carried out. Real-time data on actual resource demand is collected through IoT terminals deployed at the rescue site. The consumption of medical supplies such as bandages, hemostatic agents, and oxygen cylinders is recorded by RFID scanning terminals at temporary medical points (uploaded every 15 minutes). Rescue personnel use BeiDou positioning terminals to count the actual number of rescue personnel deployed in each area (divided by tiers). Based on the status feedback from the equipment controller, the actual number of times demolition tools, ambulances, drones, etc., are called and their continuous working time is recorded. Error indicators between actual and predicted values ​​are calculated, including mean squared error (MSE) and mean absolute error (MAE), and an initial error threshold is set: when MSE ≤ 30 and MAE ≤ 5, the model prediction accuracy meets the standard; exceeding the threshold triggers parameter optimization. When optimizing model hyperparameters and correcting biases, a Bayesian optimization algorithm is used to dynamically adjust the hyperparameters of the LSTM model. The optimization targets include the learning rate (initial value 0.001, adjustment range 0.0001-0.01), the time step size of the LSTM unit (initial 60 minutes, adjustable to 30 minutes or 90 minutes), and the number of hidden layer neurons (optimized within the range of 48-80). A bias feedback mechanism is established, performing corrections every 30 minutes to backpropagate the latest actual demand and prediction bias data to the LSTM model. Stochastic gradient descent (SGD) is used to update the model weight parameters, with the learning rate dynamically adjusted according to the bias ratio. When the bias rate is >20%, the learning rate is increased to 0.002; when the bias rate is <5%, the learning rate is decreased to 0.0005. Trend biases in the prediction curve are corrected. For example, to address the non-linear growth trend of material demand with rising water levels during floods, an attention mechanism is added to strengthen the feature weights during high water periods. By integrating prediction results and bias correction parameters from multiple time scales (1 hour, 3 hours, 6 hours, and 12 hours), a structured resource demand forecast report is generated. The report includes demand type (specified to specific categories), demand quantity (quantified to specific units), time nodes (marking the expected time of peak demand), and confidence level (calculated based on MSE and MAE). This resource demand forecast report is then transmitted to the emergency command platform to provide data support for subsequent decision optimization of deep reinforcement learning models.

[0021] like Figure 1 As shown, in step S4, based on the resource demand prediction results, a deep reinforcement learning model including a state space, an action space, and a dynamic reward function is constructed. The decision strategy is optimized through offline training and online fine-tuning to generate an optimized decision strategy.

[0022] Specifically, in this embodiment, the resource demand prediction results generated in step S3 are used as input to construct a multi-dimensional state space for a deep reinforcement learning model. This multi-dimensional state space covers several key elements, including the location coordinates of the disaster-stricken area (accurate to the street or grid unit level, used to locate the spatial range of resource allocation); the real-time inventory of resources at each reserve point and during transportation (refined to specific categories such as medical supplies, rescue equipment, and food, quantified in quantity or capacity units); the demand priority of each disaster-stricken area (divided according to indicators such as the number of people trapped, the degree of casualties, and the importance of infrastructure, set from high to low as 1-5 levels); road traffic status (collected in real time through IoT devices, represented by parameters such as traffic speed, congestion coefficient, and damaged road section identification); the operating condition of transportation equipment (such as vehicle range, drone battery remaining, and ship carrying capacity, reflecting the feasibility of equipment performing tasks); the weather impact coefficient (combining meteorological data such as rainfall, wind force, and visibility, quantifying the degree of impact on transportation with a value of 0-1); and the personnel safety risk index (combining factors such as the probability of secondary disasters and environmental hazard levels to assess the safety risks of rescue personnel performing tasks). A combined action space for the deep reinforcement learning model is defined, consisting of three basic actions: resource allocation, path planning, and task switching. Resource allocation includes resource allocation plans for different disaster-stricken areas, specifying the outgoing and incoming points and quantities of various resources. For example, allocating 50 first-aid kits and 3 demolition machines from reserve A to a priority level 1 area. Path planning involves selecting alternative routes and adjusting their order. Based on real-time traffic conditions, the optimal route is selected from multiple preset routes, and the transportation order can be dynamically adjusted according to changes in road conditions, such as avoiding sudden congestion and prioritizing transportation routes to high-priority areas. Task switching mainly involves the dynamic change of scheduling priorities. When new changes occur in the disaster situation (such as the addition of new severely affected areas or road clearing), the execution order of existing tasks can be adjusted in real time to ensure that critical needs are met first. A dynamic reward function is set up, which comprehensively considers multiple core indicators to guide the model's optimization decisions. These include: resource arrival timeliness, which uses the time difference between resource allocation and delivery as the metric, with shorter time resulting in higher reward values; demand fulfillment efficiency, calculated by the matching degree between the actual allocation amount and the demand forecast, with higher matching degrees resulting in higher reward values; rescue personnel safety, which adjusts the reward weight based on the personnel safety risk index, with lower risk indices resulting in higher reward values; and route risk avoidance, which combines road conditions and weather impact coefficients, with lower route risks resulting in higher reward values. At the same time, the weight of each indicator is dynamically adjusted according to the changes in the disaster situation. For example, in the early stage of the disaster, the timeliness of resource arrival is emphasized, while in the period of high incidence of secondary disasters, the weight of the safety of rescue personnel is increased. Based on the core metrics considered and the adjusted weights, the dynamic reward function is set as follows:

[0023] in, This represents the total output value of the dynamic reward function. Incentives for timely resource delivery Reward efficiency for meeting demand. As a safety reward for rescue personnel, Incentives for mitigating route risks The weighting coefficient for resource arrival timeliness. The weighting coefficients for efficiency in meeting demand. Weighting coefficients for the safety of rescue personnel. The weighting coefficient for route risk avoidance is given by the following constraints: The values ​​can be dynamically adjusted according to the stage of the disaster. A proximal policy optimization algorithm is employed to train and optimize the deep reinforcement learning model. Historical disaster relief data, including resource allocation records, disaster evolution data, and road traffic logs from the past five years of earthquakes, floods, and typhoons, are used to conduct multiple rounds of offline training on the model, which incorporates the aforementioned multi-dimensional state space, combined action space, and dynamic reward function. During training, different disaster scenarios are simulated, such as varying scales of affected areas, sudden road closures, and resource shortages, to iteratively update the policy network parameters. After each training round, the deviation between the model's decision and the actual optimal solution is calculated. When the deviation value is below a preset threshold (e.g., 5%) for 100 consecutive training rounds, the model is considered to have converged to a preset accuracy. The trained model is fine-tuned online by combining real-time disaster data and resource status. During actual rescue operations, newly collected dynamic data on disaster-stricken areas (such as newly added trapped points and road clearing information) and real-time resource consumption data (such as the remaining amount of medical supplies in a certain area) are input into the model. A proximal strategy optimization algorithm is used to dynamically adjust the decision threshold. For example, when the priority of demand in a certain area suddenly increases, the model automatically lowers the decision threshold for resource allocation in that area to speed up the response; when transportation equipment malfunctions, the scheduling weight of alternative equipment is increased. Through continuous online fine-tuning, optimized decision-making strategies adapted to real-time disaster situations are generated.

[0024] like Figure 1 As shown, in step S5, a resource scheduling scheme is generated based on the optimization decision strategy, combined with the priority matching rules of resources and tasks and real-time path planning.

[0025] Specifically, in this embodiment, the core parameters of resource scheduling are extracted from the optimization decision strategy obtained in step S4 to form a preliminary scheduling instruction set. The core parameters include the types of resources required by each disaster-stricken area, such as specific categories like first aid kits, demolition equipment, food and drinking water; the allocation quantity for each type of resource, accurate to the unit of measurement such as individual, unit, or box; the scheduling priority ranking determined based on the urgency of the disaster, such as level 1 being the most urgent and level 5 being generally urgent; and the time window constraints for resource delivery, such as delivery within the next 2 hours or 4-6 hours. The preliminary scheduling instruction set is presented in the form of a structured table, clearly specifying the dispatch reserve point, target area, resource details, and time requirements corresponding to each instruction. Establish a priority matching mechanism for resources and tasks to determine the order of resource allocation. This mechanism uses several key indicators of the disaster-stricken area as input parameters, including: population density in the area (number of people trapped per square kilometer), proportion of seriously injured (the proportion of seriously injured to the total number of trapped people), length of time trapped (the duration from the disaster to the present), degree of infrastructure damage (quantified by the proportion of functional loss of key facilities such as roads, communications, and medical points), and existing resource reserves (the duration that the various resources currently available in the area can support). The weight of each parameter is determined by the analytic hierarchy process (AHP). For example, the weight of personnel density is 0.3, the weight of the proportion of seriously injured people is 0.25, the weight of the length of time trapped is 0.2, the weight of the degree of infrastructure damage is 0.15, and the weight of the existing resource reserves is 0.1. The sum of the weights is 1. Based on the above weights, the parameter values ​​of each region are weighted and calculated to obtain a comprehensive priority coefficient, which ranges from 0 to 10. The higher the value, the higher the priority. The regions are sorted from high to low according to the coefficient to determine the order of resource allocation to each region. A real-time route planning model is constructed, using an improved A algorithm as the core algorithm. This model incorporates multiple factors into the route cost evaluation system, including road traffic conditions (based on real-time monitored traffic speed and congestion coefficients), road surface load-bearing capacity (matching the weight limitations of transportation equipment), weather conditions (such as the impact coefficients of heavy rain and strong winds on traffic), terrain slope (the gradient of mountainous road sections on the added difficulty of transportation), and potential secondary disaster risk areas (such as the risk coefficients of landslide and debris flow high-incidence areas). The improved A algorithm quantifies the above factors into cost values ​​and substitutes them into a heuristic function to calculate the shortest time path from each resource reserve point to the target area. The heuristic function is calculated as follows:

[0026] Where h(n) is used to estimate the total cost from the current node n to the target node goal, and ht(n, goal) is a time cost heuristic. As a heuristic of risk cost, , These are the weight coefficients for time cost heuristics and risk cost heuristics, respectively. Considering the dual requirements of timeliness and safety in emergency rescue scenarios, the following settings are established. =0.6、 =0.4; The time cost heuristic is calculated as follows:

[0027] Where L(n, goal) is the straight-line distance from node n to the target node (unit: km, calculated using the Haversine formula); v base k1 is the basic road speed (unit: km / h, preset according to road type, such as 80 for expressways and 40 for ordinary roads); k2 is the road traffic condition coefficient (0.3 ≤ k1 ≤ 1, converted from real-time congestion coefficient, congestion coefficient 0 corresponds to k1=1, congestion coefficient 1 corresponds to k1=0.3); k3 is the weather influence coefficient (0.2 ≤ k2 ≤ 1, such as 0.4 for rainy days and 1 for sunny days); k4 is the terrain slope coefficient (0.5 ≤ k3 ≤ 1, 0.5 for slope > 15°, and 1 for flat land). The risk cost heuristic is calculated as follows:

[0028] Among them, R disaster Secondary disaster risk index (0≤R) disaster ≤1, determined based on historical data and real-time monitoring (0.8 for high-risk landslide areas, 0 for safe areas), R road Road damage risk index (0≤R) road ≤1, calculated from parameters such as road surface damage rate and bridge load-bearing capacity. Sub-weights ( =0.7, =0.3), with priority given to the risk of secondary disasters; The total heuristic value is calculated using the heuristic function described above, quantifying the estimated total cost (time + risk) from the current node to the target node. This provides a clear search guide for the algorithm. During path planning, the algorithm will prioritize searching for nodes with smaller total heuristic values, avoiding blindly traversing invalid paths and significantly improving search efficiency. For example, when two candidate nodes have total heuristic values ​​of 0.13 and 0.35 respectively, the algorithm will prioritize exploring the path corresponding to the former, focusing on the potential optimal solution more quickly. Meanwhile, to cope with sudden path interruptions, the model generates two or more alternative paths simultaneously, and the spatial separation between the alternative paths and the main path is not lower than a preset threshold, such as the overlap rate of the main road not exceeding 30%, to ensure path independence. By integrating the initial scheduling instruction set, the resource allocation order based on the comprehensive priority coefficient, and the primary and alternative paths output by the real-time path planning model, a complete resource scheduling plan is generated. The plan includes: a detailed list of resources for each batch (type, quantity, and dispatch point), target areas and allocation order, detailed routes of the primary and alternative paths (including road segments, estimated time, and key nodes), transportation configuration, departure and arrival time nodes, and emergency adjustment plans (such as switching mechanisms in case of road interruption). The final plan is presented in the form of a combination of visual maps and text descriptions.

[0029] like Figure 1 As shown, in step S6, the rescue trajectory is tracked in real time based on the execution data of the resource scheduling scheme. When the deviation exceeds the threshold, rescheduling is triggered, and a closed-loop optimization mechanism is formed based on the evaluation data.

[0030] Specifically, in this embodiment, real-time tracking of rescue trajectories relies on IoT devices and vehicle-mounted terminals to collect various execution data of resource transportation. Positioning modules and status sensors are deployed on transport vehicles, drones, and other equipment to obtain information such as the real-time location, speed, and remaining range of resources. The arrival time of resources, the actual number of resources deployed, and the receipt confirmation information are recorded by receiving terminals deployed in the disaster area. This data is aggregated to form a real-time execution data dashboard, which intuitively displays the actual departure time, current location, estimated remaining transportation time, and the number of resources that have been deployed for each batch of resources, allowing commanders to fully grasp the execution progress of the dispatch plan. To promptly identify execution deviations, clear deviation thresholds need to be set, including three categories: time deviation, quantity deviation, and route deviation. The time deviation threshold refers to the maximum allowable difference between the actual arrival time and the planned arrival time of resources. For example, if it is set to 30 minutes, and a batch of resources arrives 40 minutes later than planned, it exceeds the time deviation threshold. The quantity deviation threshold is the allowable fluctuation ratio between the actual quantity of resources deployed and the planned quantity. For example, if it is set to ±5%, and only 93 boxes of food are actually deployed when 100 boxes are planned, the fluctuation ratio is 7%, exceeding the quantity deviation threshold. The route deviation threshold is the maximum spatial deviation distance between the actual driving route and the planned route. Assuming it is set to 2 kilometers, if the transport vehicle deviates from the planned route by 3 kilometers due to detours, it exceeds the route deviation threshold. A deviation monitoring mechanism is established to compare the execution data with the planned parameters in the scheduling plan in real time. The system will continuously compare and calculate the actual arrival time, quantity, and route data collected with the corresponding planned values ​​in the plan. When any deviation value exceeds the corresponding threshold, the rescheduling process is automatically triggered. During rescheduling, the priority coefficient of each disaster area is first recalculated, and the resource allocation order is adjusted based on the latest data such as the number of people trapped and the remaining resources. Secondly, the transportation route is replanned according to the new changes in road traffic conditions, which may involve using the alternative routes generated in step S5. At the same time, the allocation ratio of resources in subsequent batches is optimized, increasing the resource allocation for areas with increased priority and appropriately reducing it for areas where demand has been alleviated. A scheduling effectiveness evaluation system is constructed, which comprehensively measures the execution effect of the scheduling plan through multiple indicators. The evaluation indicators include demand satisfaction rate, which is the proportion of the resource demand that has been satisfied to the total demand; average response time, which is the average time from the issuance of the scheduling instruction to the delivery of the resource; resource utilization rate, which is the proportion of the actual amount of resources used to the total amount of resources deployed; and transportation safety coefficient, which is evaluated in conjunction with whether safety accidents occur during transportation and whether the route passes through high-risk areas. These indicators are weighted according to preset weights to obtain a comprehensive evaluation score, thereby quantifying the overall effectiveness of the scheduling plan. A closed-loop optimization mechanism is formed based on the evaluation data. The comprehensive evaluation score and the corresponding scheduling strategy are used as training samples and fed back to the deep reinforcement learning model in step S4. The model will use these new samples to update the state space parameters and the weights of the dynamic reward function. For example, if the resource utilization rate index has a significant impact on the scheduling effect under a certain type of disaster, the model will increase the weight of the index in the reward function. In this way, a closed loop is realized from data collection, scheme execution to model optimization, so that subsequent resource scheduling decisions can better adapt to the actual disaster situation and continuously improve the efficiency and accuracy of emergency rescue.

[0031] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0032] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and rules of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for emergency rescue resource scheduling based on the Internet of Things, characterized in that, Includes the following steps: S1. Obtain environmental parameters, images, and personnel distribution data of the disaster area through IoT devices, construct multi-source heterogeneous data, and call the historical disaster case database to perform data association with the multi-source heterogeneous data; S2. Preprocess the multi-source heterogeneous data, extract disaster characteristics, dynamically allocate the weights of each data source according to the disaster type, align the processed data in time and space, and generate a disaster assessment matrix. S3. Based on the disaster assessment matrix, the peak resource demand of each disaster-stricken area in the future is output using a time-series prediction model. At the same time, the model parameters are corrected by combining the deviation between the actual demand and the predicted value to generate resource demand prediction results. S4. Based on the resource demand prediction results, construct a deep reinforcement learning model, optimize the decision-making strategy through offline training and online fine-tuning, and generate an optimized decision-making strategy. S5. Based on the optimization decision-making strategy, a resource scheduling scheme is generated by combining the priority matching rules of resources and tasks with real-time path planning. S6. Based on the execution data of the resource scheduling scheme, the rescue trajectory is tracked in real time. When the deviation exceeds the threshold, rescheduling is triggered, and a closed-loop optimization mechanism is formed based on the evaluation data.

2. The method for emergency rescue resource scheduling based on the Internet of Things according to claim 1, characterized in that: The S1 step specifically includes: An IoT sensing array was constructed, and environmental acquisition devices, image acquisition devices, and positioning acquisition devices were deployed throughout the disaster area to acquire physical environmental parameters, visualization information, and spatial distribution information of people in the disaster area, forming a multi-source heterogeneous raw dataset. For the aforementioned multi-source heterogeneous raw data, outlier removal and data smoothing are performed on environmental parameters, sharpness enhancement, distortion correction and target recognition are performed on image data, and position coordinate calibration and density statistics are performed on personnel distribution data to obtain processed multi-source heterogeneous data. The pre-established historical disaster case database is invoked, and the processed multi-source heterogeneous data is compared with the key parameters in the historical cases using a feature matching algorithm. Historical cases with similarity exceeding a preset threshold are selected, and an association index between real-time data and historical cases is established to form an associated dataset that includes original sensing data and historical reference information.

3. The method for emergency rescue resource scheduling based on the Internet of Things according to claim 2, characterized in that: In step S1, the environmental data acquisition equipment includes temperature and humidity sensors, air pressure sensors, wind speed and direction sensors, soil moisture sensors, and water quality monitoring sensors, which are deployed at 500-meter intervals to form a grid-like monitoring network. The image acquisition equipment includes a swarm of drones, panoramic monitoring cameras, and portable anti-shake cameras. The drones are configured at one per square kilometer and have a cruising altitude of 100 meters.

4. The method for emergency rescue resource scheduling based on the Internet of Things according to claim 1, characterized in that: The S2 step specifically includes: The multi-source heterogeneous data is preprocessed in real time using edge computing nodes. The 3σ criterion is used to identify and remove outliers for environmental parameters, and the mean of nearby time times is used to fill in missing values. The image data is processed using a semantic segmentation algorithm to identify and extract key disaster features, and the extracted results are converted into a binary disaster heat map. Based on the personnel distribution data, terrain information from the digital elevation model is introduced. The positioning deviation caused by terrain occlusion is corrected through coordinate mapping. Clustering algorithms are used to identify densely populated areas, and the number of trapped people is estimated by combining the positioning signal strength. Based on the coordinate system of the geographic information system, the preprocessed environmental parameters, the binarized disaster heat map and the corrected personnel distribution data are spatiotemporally aligned according to the geographical coordinates to construct a three-dimensional data cube including time, space and disaster characteristics. An adaptive weight allocation mechanism is introduced to dynamically adjust the weights of each data source for different disaster types. By weighted fusion of information from a three-dimensional data cube, a multi-dimensional disaster assessment matrix is ​​generated, including environmental hazard level, infrastructure damage level, and the status of trapped personnel.

5. The method for emergency rescue resource scheduling based on the Internet of Things according to claim 4, characterized in that: The semantic segmentation algorithm uses the U-Net model to process image data and generate a binary disaster heat map, wherein the pixel value of the feature region in the binary disaster heat map is set to 1, and the pixel value of the non-feature region is set to 0.

6. The method for emergency rescue resource scheduling based on the Internet of Things according to claim 1, characterized in that: The S3 step specifically includes: Using the environmental hazard level, infrastructure damage level, and personnel entrapment status characteristic parameters in the disaster assessment matrix as input, a time series prediction model based on LSTM is constructed. A sliding time window is set to capture the time series characteristics of the data, and the peak resource demand of each disaster-stricken area in the future is output. The resource demand includes the quantity of medical supplies, the configuration of rescue personnel, and the priority of equipment deployment. Real-time data on actual resource demand in various disaster-stricken areas is collected, the mean square error and mean absolute error between actual and predicted values ​​are calculated, and the hyperparameters of the time series prediction model are dynamically adjusted through a Bayesian optimization algorithm. A deviation feedback mechanism is established, in which the latest actual demand and prediction deviation data are backpropagated to the time series prediction model at preset intervals, and the model weight parameters are updated by gradient descent to correct the trend deviation of the prediction curve. By integrating forecast results from multiple time scales with bias correction parameters, resource demand forecast results are generated.

7. The method for emergency rescue resource scheduling based on the Internet of Things according to claim 1, characterized in that: The S4 step specifically includes: Based on the resource demand forecast results, a multi-dimensional state space of a deep reinforcement learning model is constructed. The multi-dimensional state space includes the location coordinates of the disaster area, the real-time inventory of resources, the demand priority of each area, the road traffic status, the operating condition of transportation equipment, the weather impact coefficient, and the personnel safety risk index. Define the combined action space of the deep reinforcement learning model. The combined action space consists of three basic actions: resource allocation, path planning, and task switching. Resource allocation includes different types and quantities of resource allocation schemes, path planning includes alternative route selection and order adjustment, and task switching includes dynamic changes in scheduling priority. A dynamic reward function is set up, which comprehensively considers the timeliness of resource arrival, the efficiency of demand fulfillment, the safety of rescue personnel, and the avoidance of route risks. The deep reinforcement learning model is trained and optimized using a proximal policy optimization algorithm. The model, which includes the multi-dimensional state space, combined action space and dynamic reward function, is trained offline for multiple rounds using historical disaster relief data. The policy network parameters are iteratively updated by simulating different disaster scenarios until the model converges to the preset accuracy. The trained model is fine-tuned online by combining real-time disaster data and resource status, and the decision threshold is dynamically adjusted by a near-end strategy optimization algorithm to generate an optimized decision strategy.

8. The method for emergency rescue resource scheduling based on the Internet of Things according to claim 1, characterized in that: The S5 step specifically includes: Based on the aforementioned optimization decision-making strategy, the core parameters of resource scheduling are extracted to form a preliminary scheduling instruction set; Establish a priority matching mechanism for resources and tasks. Using the population density, proportion of seriously injured, length of time trapped, degree of infrastructure damage, and existing resource reserves in the disaster-stricken area as input parameters, determine the weight of each parameter through the analytic hierarchy process, calculate the comprehensive priority coefficient of each area by weighting, and determine the order of resource allocation by sorting the coefficients from high to low. A real-time path planning model is constructed, using the improved A* algorithm as the core algorithm. The road traffic status, road surface bearing capacity, meteorological conditions, terrain slope and potential secondary disaster risk areas are used as path cost evaluation indicators. The shortest time path from each resource reserve point to the target area is calculated, and two or more alternative paths are generated simultaneously. The spatial separation degree between the alternative paths and the main path is not less than a preset threshold. Integrate scheduling instructions, priority ranking, and path planning results to generate a resource scheduling scheme.

9. The method for emergency rescue resource scheduling based on the Internet of Things according to claim 8, characterized in that: The resource scheduling plan specifically includes a detailed list of resources for each batch, target areas and allocation order, detailed routes of the main path and alternative paths, configuration of transportation vehicles, departure and arrival time nodes, and emergency adjustment plan.

10. The method for emergency rescue resource scheduling based on the Internet of Things according to claim 1, characterized in that: Step S6 specifically includes: The real-time location, speed, and status information of resource transportation are collected by IoT positioning devices and vehicle terminals. Combined with electronic fence technology, the transportation trajectory and deployment progress of rescue resources are tracked to generate a real-time execution data dashboard. The real-time execution data dashboard includes the actual departure time, current location, estimated remaining time, and the amount of resources deployed. Set time, quantity, and path deviation thresholds, where the time deviation threshold is the maximum allowable difference between the actual arrival time and the planned arrival time of the resources, the quantity deviation threshold is the allowable fluctuation ratio between the actual quantity of resources deployed and the planned quantity of resources deployed, and the path deviation threshold is the maximum spatial deviation distance between the actual driving route and the planned route. Establish a deviation monitoring mechanism to compare the execution data with the planned parameters of the scheduling scheme in real time. When any type of deviation value exceeds the corresponding threshold, the rescheduling process is automatically triggered. A scheduling performance evaluation system is constructed. A comprehensive evaluation score is obtained through weighted calculation. The comprehensive evaluation score and the corresponding scheduling strategy are used as training samples and fed back to the deep reinforcement learning model to update the model's state space parameters and reward function weights.

Images