An emergency material storage optimization layout method and system considering dynamic demand

Abstract

This invention discloses a method and system for optimizing the layout of material storage considering dynamic demand in emergency situations, belonging to the field of emergency management technology. The method achieves real-time prediction of material demand through a dynamic probabilistic convolutional neural network. Combining the importance of emergency response, the urgency of materials, and supply chain sensitivity, an adaptive weighting algorithm is used to classify and prioritize materials. Based on dynamic demand and priority, a multi-level partitioning generation and optimization algorithm is employed to achieve dynamic division and optimal matching of warehouse space. Reinforcement learning-driven path optimization improves the efficiency of material transfer. The system combines intelligent sensing and real-time feedback mechanisms to continuously monitor the storage status and uses Bayesian optimization and self-adjusting algorithms to achieve adaptive adjustment of layout and paths. This invention improves the response speed, reliability, and warehouse utilization rate of emergency material dispatch.

Description

Technical Field

[0001] This invention belongs to the field of emergency management technology, and more specifically relates to a method and system for optimizing the layout of material storage considering dynamic needs in emergency situations. Background Technology

[0002] With the continuous advancement of urbanization and the increasing complexity of the social environment, various emergencies, especially major disasters, pose a serious threat to the safety of people's lives and property. As a crucial protective infrastructure for responding to various emergencies, civil defense engineering plays an increasingly prominent role in emergency material supply. Traditional civil defense material storage layouts often rely on static historical data and manual experience for planning, making it difficult to cope with the strong fluctuations and uncertainties in demand during emergencies such as natural disasters. Furthermore, existing storage systems are mostly based on fixed zones, single paths, and periodic manual inspections, lacking the ability to perceive real-time dynamic demands and adaptive scheduling mechanisms. This results in difficulties in timely and accurate distribution of materials to the locations with the most urgent needs during actual emergency response, affecting overall rescue efficiency and personnel safety.

[0003] In recent years, the rapid development of information technologies such as the Internet of Things, artificial intelligence, and big data analytics has provided a technological foundation for the intelligent and refined management of emergency material storage. However, how to integrate dynamic multi-source data, quickly perceive changes in demand, and scientifically optimize warehouse space layout and material allocation under complex emergency situations to ensure the efficient and safe entry and exit of critical materials in special circumstances remains a key challenge that urgently needs to be addressed in the field of civil defense material management. Therefore, there is an urgent need for a system and method that can dynamically perceive emergency needs in real time and intelligently optimize the layout of material storage and transfer routes to improve the material support capabilities and response efficiency of civil defense projects in responding to emergency events. Summary of the Invention

[0004] This invention aims to address the technical challenges faced by civil defense projects in emergency situations, particularly those involving large fluctuations in material demand, high requirements for response timeliness, and a lack of dynamic optimization and intelligent decision support for warehouse layout and material transportation. It enables real-time perception and accurate prediction of changing emergency demands, improves the ability of civil defense material warehouses to optimize material storage space, scientifically zonate, and efficiently and rapidly allocate materials under special emergencies, thereby significantly enhancing the timeliness and security of emergency support.

[0005] To achieve the above objectives, the present invention employs the following technical solution: the method includes:

[0006] Acquire multi-source data, and perform demand prediction based on the multi-source data using a dynamic probabilistic convolutional neural network to obtain dynamic demand perception and prediction results;

[0007] Based on the dynamic demand perception and prediction results, materials are classified and prioritized to obtain the priority of materials. The priority is determined by an adaptive weight ranking algorithm based on the importance of emergency response, urgency of materials and sensitivity of the supply chain.

[0008] Based on the dynamic demand perception and prediction results and the priority of the materials, a multi-level partitioning generation algorithm is used to generate a partition layout, wherein the partition boundaries of the partition layout are adjusted according to real-time demand changes.

[0009] Based on the partitioned layout, a dynamic path optimization algorithm is used to optimize the path of key nodes to obtain the transfer path. The transfer path is dynamically reconfigured according to the actual flow of emergency supplies entering and leaving the warehouse.

[0010] Based on monitoring data acquired by intelligent sensors and data acquisition equipment, the partition layout and the transfer path are iteratively optimized through a self-adjusting feedback algorithm.

[0011] In one approach, acquiring multi-source data and performing demand prediction based on the multi-source data using a dynamic probabilistic convolutional neural network to obtain dynamic demand perception and prediction results includes:

[0012] The multi-source data is obtained by acquiring disaster early warning information, social media public opinion information, historical disaster event records, and information collected by IoT monitoring devices.

[0013] The multi-source data is standardized, missing value imputation is performed, and noise filtering is performed. The text data in the multi-source data is vectorized using a text content recognition algorithm to obtain multimodal fusion features.

[0014] The multimodal fusion features are input into the dynamic probabilistic convolutional neural network to obtain the dynamic demand perception and prediction results.

[0015] In one approach, the hierarchical management and priority ranking of materials based on the dynamic demand perception and prediction results, to obtain the material priorities, includes:

[0016] Obtain the emergency response importance attribute, demand urgency attribute, and supply chain sensitivity attribute for each type of material, and generate a multi-dimensional attribute matrix based on the emergency response importance attribute, the demand urgency attribute, and the supply chain sensitivity attribute;

[0017] Based on the multidimensional attribute matrix, the priority of the materials is determined by an adaptive weight sorting algorithm;

[0018] The adaptive weight ranking algorithm assigns an initial weight to each attribute and adjusts the initial weight in real time according to the dynamic information of the emergency environment. The adaptive weight ranking algorithm is also used to calculate a structural score, which is used to characterize the demand correlation between materials. The initial weight is adaptively adjusted by combining the structural score and the real-time emergency coefficient to obtain the adjusted weight.

[0019] In one approach, the step of generating a partition layout using a multi-level partitioning algorithm based on the dynamic demand perception and prediction results and the priority of the materials includes:

[0020] Based on the dynamic demand perception and prediction results and the priority of the materials, the materials are grouped. Combining the material mobility attributes and security requirement attributes, the warehouse space is divided into multiple partitions, where each partition corresponds to a priority and a material attribute.

[0021] Based on the required area and partitioning requirements of the partition, the boundary of the partition is iteratively optimized by using a partition hybrid binary search method combined with a genetic optimization algorithm and a random perturbation heuristic algorithm to obtain the adjusted partition boundary.

[0022] Based on the constraints of the main passage, entrances and exits, and safety facilities, the zoning layout is determined. In the zoning layout, the high-priority material zones are located in the area near the entrances and exits, and the hazardous materials zones and special materials zones are isolated from the main flow area.

[0023] In one approach, the process of optimizing key node paths using a reinforcement learning-driven dynamic path optimization algorithm based on the partitioned layout to obtain a transit path includes:

[0024] The warehousing environment is abstracted as a directed graph, which includes multiple key nodes and multiple channels. Each channel corresponds to dynamic passage delay, congestion level, obstacle status, and priority path marking.

[0025] Based on the directed graph, the transfer path decision is modeled using state space, action space, and immediate reward function to obtain the path decision model;

[0026] The path decision model is input into a deep neural network to obtain the transfer path. The deep neural network is embedded with an attention module and a multi-head output mechanism. The deep neural network is used to autonomously learn and dynamically optimize the multi-target transfer path.

[0027] In one approach, the monitoring data acquired by intelligent sensors and data acquisition devices is used to iteratively optimize the partition layout and the transport path through a self-adjusting feedback algorithm, including:

[0028] The system uses intelligent sensors and data acquisition devices to monitor material distribution information, storage security information, and environmental change information in real time, obtains the monitoring data, and performs data-driven analysis and anomaly warning based on the monitoring data.

[0029] The self-adjusting feedback algorithm is obtained by fusing the Bayesian optimization algorithm with the reinforcement learning self-adjusting strategy.

[0030] Based on the self-adjusting feedback algorithm, the operating parameters and execution status corresponding to the monitoring data are analyzed, and the partition layout and the transfer path are dynamically corrected to achieve parameter reconstruction.

[0031] In one approach, determining the priority of the materials based on the multidimensional attribute matrix using an adaptive weighted sorting algorithm includes:

[0032] The materials are then weighted and comprehensively scored based on the adjusted weights, and the materials are sorted in descending order based on the weighted comprehensive scores to obtain the sorting results.

[0033] Based on the arrangement results, the materials are divided into high priority, medium priority and low priority using the quantile method.

[0034] Furthermore, a material storage optimization layout system that considers dynamic demand in emergency situations, the system being used to implement the method described, the system comprising:

[0035] The dynamic demand perception and prediction module is used to acquire multi-source data, perform demand prediction based on the multi-source data through a dynamic probabilistic convolutional neural network, obtain dynamic demand perception and prediction results, and perform multi-modal feature fusion and environment embedding.

[0036] The material classification and priority management module is connected to the dynamic demand perception and prediction module. It is used to obtain the multi-dimensional attributes of materials, determine the priority of materials based on the multi-dimensional attributes through an adaptive weight sorting algorithm, and dynamically adjust the weights based on the structured scores.

[0037] A multi-level partition layout generation module, connected to the material classification and priority management module, is used to generate a partition layout based on the dynamic demand perception and prediction results and the priority of the materials. The partition boundary is dynamically adjusted by using a partition hybrid binary method combined with a genetic optimization algorithm and a random perturbation heuristic algorithm.

[0038] The key node path optimization module, connected to the multi-level partition layout generation module, is used to abstract the warehousing environment into a directed graph and dynamically optimize and reconfigure the transfer path based on a dynamic path optimization algorithm driven by deep neural networks and reinforcement learning.

[0039] The real-time feedback and iterative optimization module is connected to the multi-level partition layout generation module and the key node path optimization module, respectively. It is used to iteratively optimize the partition layout and transfer path based on the monitoring data obtained by the intelligent sensor and data acquisition device through a self-adjusting feedback algorithm. The self-adjusting feedback algorithm is obtained by fusing the Bayesian optimization algorithm and the reinforcement learning self-adjusting strategy.

[0040] The human-computer interaction and monitoring terminal is connected to each of the above modules to display the warehouse status, optimization suggestions and decision results, and to receive operation instructions and abnormal alarm information.

[0041] Beneficial effects of this invention:

[0042] This invention addresses the issues of material storage layout and allocation in civil defense projects during emergency situations, significantly improving the intelligence level of material management and emergency response capabilities. Compared to traditional warehousing methods that rely on static planning and manual experience, this invention can collect and analyze multi-source emergency data in real time, accurately predict changes in material demand, automatically optimize warehouse space and zoning layout, and formulate efficient material transfer plans.

[0043] This approach not only improves the timeliness and accuracy of material distribution and reduces the risks and error rates in the material allocation process, but also effectively ensures the safety of material supplies for key areas and personnel, and enhances the overall support capacity and management efficiency of civil defense projects under sudden high-pressure emergency situations. In summary, this invention achieves scientific, intelligent, and efficient management of civil defense emergency materials, and has significant social benefits and application value. Attached Figure Description

[0044] Figure 1 This is a flowchart of the method of the present invention;

[0045] Figure 2 This is a flowchart of the path optimization process for key nodes in this invention;

[0046] Figure 3 This is a system block diagram of the present invention. Detailed Implementation

[0047] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Typical embodiments of the invention are shown in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete.

[0048] Unless otherwise defined, all technical and scientific terms used in this invention have the same meaning as understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. To facilitate understanding, the invention will now be described more fully with reference to the accompanying drawings. Typical embodiments of the invention are shown in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to make the disclosure of the invention more thorough and complete.

[0049] like Figure 1 As shown, a method for optimizing the layout of material storage considering dynamic demand in emergency situations includes:

[0050] Step 1: Dynamic Demand Perception and Predictive Modeling

[0051] First, given the high uncertainty and suddenness of material demand in emergency situations, a real-time dynamic demand perception mechanism is needed. By integrating multi-source data (disaster early warning, social media, historical events, and IoT monitoring information) and using an improved time series forecasting algorithm—Dynamic Probabilistic Convolutional Neural Network (DP-CNN)—the ability to quickly capture changes in demand is enhanced. This algorithm not only focuses on time-series changes but also incorporates embedded modeling of environmental influencing factors, improving the accuracy of dynamic demand prediction and providing a solid data foundation for warehouse layout optimization.

[0052] In emergency situations, the demand for supplies often exhibits high uncertainty and suddenness. Therefore, it is essential to first construct a real-time dynamic demand perception and prediction modeling system.

[0053] In the specific implementation process, the data acquisition module first integrates data from disaster early warning, social media sentiment, historical disaster event records, and Internet of Things (IoT) monitoring devices. These data types include time-series numerical data (such as real-time sensor monitoring), text data (such as social media messages), spatial location information, and structured historical data. The data preprocessing stage employs methods such as standardization, missing value imputation, and noise filtering to ensure that all types of data are analyzed on the same scale. Subsequently, text content recognition algorithms (such as BERT embedding) are used to vectorize the textual situational information and perform multimodal fusion with the time-series data to form a unified input feature matrix.

[0054] For the prediction model, an improved Dynamic Probabilistic Convolutional Neural Network (DP-CNN) is employed. This model essentially combines probabilistic modeling and dynamic embedding mechanisms with a traditional Convolutional Neural Network (CNN) to achieve temporal and environmental correlation perception of resource demand. Its architecture includes multi-layer convolutional units, probabilistic branches, and an environmental embedding module.

[0055] First, the convolutional unit is used to extract local trend features from time series data. The specific formula is as follows:

[0056]

[0057] in Given the input demand sequence, is the kernel weight, b is the bias, and k is the kernel size.

[0058] Based on the features generated by convolution, the probability branch uses Bayesian probability modeling to predict the demand and its confidence interval.

[0059] The process is as follows: For the predicted value at each time step Generate probability distribution in Features extracted by convolution For environmental embedding factors (such as disaster level, geographic information, social media popularity, etc.), N represents a Gaussian distribution.

[0060] The environment embedding module employs a deep embedding strategy, transforming various environmental influencing factors into numerical representations. Specifically, this is achieved by introducing multi-dimensional embedding vectors.

[0061]

[0062] in This represents the numerical representation of the i-th environmental factor at time t. This vector is then fused with the convolutional features using weighted fusion. Among them Dynamic weighting factors are adaptively adjusted through model training.

[0063] Next, supervised learning is performed using historical data and environmental change information. The loss function design not only considers demand forecasting errors (such as mean squared error) but also introduces a confidence interval coverage index to achieve multiple optimizations for uncertainty and dynamism. The total loss function formula is:

[0064]

[0065] in For practical needs, CoverageError is a weight hyperparameter used to measure the reasonableness of the predicted confidence interval.

[0066] After deployment, the model receives real-time input of multi-source data, and uses DP-CNN to predict and feed back the latest demand dynamics and their probability ranges, forming a real-time data stream of dynamic demand for emergency supplies. This result can drive intelligent decision-making in warehouse layout and provide decision-makers with risk warnings and resource mobilization basis.

[0067] Step Two: Material Classification and Prioritization

[0068] After obtaining accurate demand forecasts, materials are categorized and prioritized. Taking into account the importance of emergency response, the urgency of materials, and supply chain sensitivity, an adaptive weighted ranking algorithm classifies materials into high, medium, and low priorities. This step provides strategic guidance for subsequent warehouse layout, ensuring the timeliness and rationality of resource allocation.

[0069] This process plays a fundamental role in ensuring the efficiency and rationality of subsequent warehousing layout. The specific implementation begins with the collection of material attribute data, covering indicators such as the emergency response importance of each type of material (e.g., life-saving supplies, medical supplies, basic necessities), the urgency of demand (demand timeliness, shortage risk, predicted peak time window, etc.), and supply chain sensitivity (difficulty of material acquisition, stability of supply channels, transportation complexity, etc.). A multi-dimensional attribute matrix is ​​then established for all materials.

[0070]

[0071] in, Let represent the score of the i-th type of material on the j-th attribute, k be the type of material, and n be the number of attributes.

[0072] For multi-objective ranking, an adaptive weighted ranking algorithm is adopted. First, an initial weight is assigned to each attribute; the weight vector is... The weights are adjusted in real time based on the current emergency environment needs and dynamic information. A LINE structured score is introduced to uncover structural demand relationships between resources. For example, if medical supplies and cleaning supplies are both in high demand during a disaster, their structured scores will mutually boost each other. The structured score is calculated using embedded vectors, specifically in the form of...

[0073]

[0074] in and Let be the embedding vectors of the i-th and j-th types of materials. The optimized and trained embeddings can reflect the importance of materials and their relevance to demand.

[0075] The weight vector is adaptively adjusted based on the combined structural score and real-time emergency response coefficients (such as the risk of material shortages and the environmental urgency coefficient). The adjustment formula is as follows:

[0076]

[0077] in This refers to real-time emergency response factors (such as shortage index, response time limit, etc.). Weighting of structured scores, and To adjust the parameters.

[0078] Subsequently, a weighted comprehensive score was calculated for each type of material. The scoring formula is as follows:

[0079]

[0080] Based on the overall score, the materials are sorted in descending order to form three priority groups: high, medium, and low. A common approach is to use quantiles, dividing all scores into three segments (e.g., the top 30% are high priority, the middle 40% are medium priority, and the bottom 30% are low priority).

[0081] Ultimately, the material classification and ranking results provide clear guidance for subsequent warehouse layout zoning and material scheduling. This not only ensures priority allocation of critical materials but also guarantees the timeliness and flexibility of material response under dynamic changes. The adaptive weighted ranking algorithm significantly enhances the intelligence and efficiency of emergency response management.

[0082] Step 3: Generating a multi-level partition layout

[0083] To address predicted dynamic demand and material priorities, a multi-level partitioning algorithm is employed. This algorithm partitions warehouse space according to different priorities, liquidity, and security requirements, and automatically generates a preliminary layout plan using a hybrid partitioning binary search method (combining genetic optimization and a random perturbation heuristic algorithm). The algorithm can quickly adjust partition boundaries based on real-time demand changes, achieving optimal matching of space and materials, maximizing warehouse utilization efficiency and emergency response flexibility.

[0084] This step uses the dynamic demand forecasting results and material priority ranking obtained in the previous two steps as input to maximize warehouse space utilization efficiency and emergency response flexibility, achieving efficient storage and rapid allocation of materials. It can automatically partition warehouse space based on priority, liquidity, and security, and dynamically and flexibly adjust these partitions. It integrates the Partition Hybrid Bisection (PHB) method with genetic optimization and random perturbation heuristic algorithms to generate efficient and robust initial layout schemes.

[0085] First, based on the output material classification and priority ranking results, all materials to be stored are automatically grouped into three priority levels: "high," "medium," and "low," taking into account both liquidity (i.e., frequency of inbound and outbound movement) and security requirements. The overall warehouse space is then abstracted into a two-dimensional (or three-dimensional) coordinate layout. The goal at this point is to rationally divide S into several sub-regions. Each sub-area is primarily responsible for storing a type of material with a specific priority and attribute. The area (or volume) required for each type of sub-area can be roughly calculated based on the total dynamic demand for material priority and its safe turnover coefficient, denoted as [missing information].

[0086]

[0087] in, This represents the set of materials in the k-th group. Let i be the demand for the i-th type of material. The area occupied by a single item. This sets a redundancy factor for the safe turnover of the materials. The initial area allocation provides boundary constraints for the generation of subsequent zoning schemes.

[0088] Subsequently, the Hybrid Partitioning (PHB) method was adopted as the backbone of the automatic partitioning. Hybrid partitioning means integrating the standard spatial binary recursive idea with modern intelligent optimization algorithms to achieve flexible, controllable, and efficient partitioning.

[0089] The process is as follows: First, based on the area requirements of each zone, the entire warehouse space is divided into two parts along the longest side using a regular rule. The first sub-zone corresponds to the highest priority materials (for proximity to entrances and exits and rapid distribution). The remaining sub-zones are recursively divided until all priority sub-zones meet the minimum area requirement. The bisection point p can be selected using the following optimization objective:

[0090]

[0091] in, It is an interval The cumulative area of ​​the upper region, where q is the number of the current partition category.

[0092] Traditional rule-based binary search methods are prone to rigid boundaries and spatial fragmentation due to static input. To address this, the algorithm introduces a Genetic Algorithm (GA) unit into the partitioning recursion, encoding the position and orientation of each binary search point into chromosomes. Individual fitness is determined by the objective function.

[0093]

[0094] The weighted composition measures "partition accessibility", "partition security", and "spatial compactness" respectively. Reflects the minimum total path for materials to enter and leave the warehouse. This reflects the distance / isolation indicators between high-security-requirement materials and other zones. The genetic algorithm measures the dispersion of materials within the same group. Through crossover, mutation, and selection operations, iteratively optimizes the partitioning scheme. The crossover process merges different binary tree intervals, while mutation makes small-scale perturbations at some partition nodes or boundaries to enhance layout diversity and flexible adaptability.

[0095] To cope with sudden changes in demand or warehousing constraints, partition boundaries need to have the ability to be quickly optimized and adjusted. Therefore, a random perturbation heuristic search is embedded in the genetic iteration process. When external inputs (such as real-time updates of backend demand forecasts) change significantly, small adjustments are made to local partitions: the partition node most affected by the change is selected, and a small spatial perturbation is applied to its neighborhood for movement.

[0096]

[0097] The fitness of the objective function is reassessed, and the new partitioning scheme with better performance is retained, achieving adaptive optimization of the layout to real-time dynamics. The perturbation amplitude and selection points are automatically scaled based on real-time feedback.

[0098] The multi-level zoning algorithm fully considers key facilities such as entrance / exit locations, main transport routes, and fire sprinklers. Primary zones (high-priority areas) are located in the space closest to the main entrance / exit, with secondary zones progressing inwards. Hazardous / special materials are isolated in well-ventilated zones away from high-traffic areas, maximizing safety and efficiency. For example, the overall objective function can be formed by weighting the material priority with its distance from the main entrance / exit.

[0099]

[0100] in Let be the path distance from the centroid of the k-th partition to the inlet. It is a penalty coefficient based on the priority of materials to ensure that critical materials are closest to the export destination.

[0101] Ultimately, through multiple rounds of optimization and dynamic feedback, the generated multi-level zoning scheme can be mapped onto a complete warehouse layout map, including the space number, boundary coordinates, service material attributes, allocable upper limit, and main receiving and dispatching path indicators for each area. Each sub-area forms an independent or closely connected "rapid access unit," supporting on-demand response and providing a solid foundation for intelligent warehouse scheduling and subsequent emergency material inflow and outflow. Simultaneously, the multi-level zoning scheme can automatically iterate and reallocate space at any time based on real-time demand forecasts and inventory updates, achieving the highest degree of matching between materials and space and ultimate flexibility in resource allocation.

[0102] Step 4: Optimize the path of critical nodes

[0103] Once the layout plan is finalized, a reinforcement learning-driven dynamic path optimization algorithm is employed to ensure efficient material transfer. By simulating the actual flow of emergency supplies in and out of the warehouse, a deep reinforcement learning model (such as a variant of DQN) learns and optimizes the transfer paths between key nodes in real time, fully considering dynamic obstacles and changes in priority paths within the warehouse. This adapts to various unforeseen changes, enabling dynamic path reconfiguration to ensure rapid and accurate distribution of materials to the points of need.

[0104] like Figure 2 As shown, given the high variability of material flow demand, the extreme requirements for transfer timeliness, and the complexity of the warehousing environment in emergency scenarios, traditional static path planning can no longer meet practical needs. Therefore, this step introduces a reinforcement learning-driven dynamic path optimization algorithm, utilizing a deep reinforcement learning (DRL) model to autonomously learn and intelligently optimize the transfer flow between key nodes. Through real-time perception and feedback of logistics flow, dynamic obstacles, and priority changes within the warehouse, this model can achieve dynamic adaptive reconfiguration of paths, ensuring that transfer routes remain in optimal condition and achieving rapid, accurate, and efficient delivery of materials.

[0105] S401 Problem Modeling and Environment Construction

[0106] First, the warehousing environment is abstracted as a directed graph. ,in Representing key nodes (such as storage areas, inbound / outbound ports, and logistics sorting points), E represents feasible paths between nodes. Each edge It is accompanied by several dynamic attributes, such as the current passage delay. Congestion level Obstacle status These parameters are adjusted in real time according to actual operations and environmental changes. Priority markers are also set for priority paths (such as dedicated channels for high-priority materials). This affects the priority of route selection.

[0107] The optimization objective is to determine a set of transfer paths from the starting node to the target node in each round of material inbound / outbound scheduling tasks. The goal is to minimize the total transit delay, maximize path safety, and ensure the reliability of high-priority material transit. This can be characterized by the following multi-objective optimization function:

[0108]

[0109] Among them, Total path delay Penalty for congestion / obstacles Score is accumulated based on the priority of the path. This is the adjustment coefficient.

[0110] S402, Deep Reinforcement Learning (DQN variant) Algorithm Design

[0111] The path decision-making process is modeled as a Markov Decision Process (MDP): state space S, action space A, transition probabilities P, and immediate reward function R. Based on Deep Q-Network (DQN), the following scheme is adopted:

[0112] 1. State-space design

[0113] Each state The system encodes the current location of the material package, its historical transfer path, local environment (accessibility of nearby pathways, obstacle distribution, current priority change flags, etc.), global emergency level information, and inbound / outbound priority queue status. Formalized as:

[0114]

[0115] in, Current node, , , These represent the neighbor edge state, the local obstacle zone state, and the priority channel weight, respectively.

[0116] 2. Motion space

[0117] action This refers to selecting the next transit node from the current node, which is actually the set of all reachable adjacent nodes:

[0118]

[0119] Obstacles / closed passages are not selectable.

[0120] 3. Instant reward function

[0121] The reward mechanism employs a multi-factor weighting to encourage the model to actively learn solutions that minimize latency, prioritize obstacle avoidance, and ensure timely response.

[0122]

[0123] The system penalizes long delays and congestion, while rewarding those using priority lanes. For exceptionally high-priority tasks, global rewards can be added to enhance responsiveness to urgent path conditions.

[0124] 4. Q network architecture and dynamic adjustment

[0125] The Q-function is approximated using a deep neural network. conduct.

[0126] (1) Embed an attention module to dynamically highlight the input dimension when the current obstacle state changes drastically or new priority requirements (such as the activation of emergency passage) to improve the model's path "switching" ability under sudden events;

[0127] (2) Using multi-head output, the Q value of multiple target points (such as different outbound ports and different emergency distribution points) is predicted respectively, supporting real-time optimization of multiple transfer lines at the same time.

[0128] (3) Expand the experience playback mechanism to improve the model’s learning ability for high-priority sudden events with few samples.

[0129] S403, Model Training and Dynamic Online Optimization Strategies

[0130] In the initial stage, the model is pre-trained offline using historical warehousing data and route transfer samples under simulated conditions. After actual deployment, real-time material scheduling flow and sensor data are used as environmental feedback to continuously perform online reinforcement learning and dynamically adjust the strategy.

[0131] The training objective of reinforcement learning is to maximize the cumulative expected return.

[0132]

[0133] Discount factor should be considered when calculating future returns. .

[0134] In response to sudden environmental changes (such as temporary closure of a district or a surge in the amount of supplies transferred in), the model triggers online relearning through attention mechanisms and environmental perception modules, continuously refreshing the Q network parameters to achieve rapid adaptive reconfiguration of paths.

[0135] Applying a priority dynamic adjuster can adjust the priority weight component in the reward function when high-priority materials are received in bulk (such as large-scale disaster relief supplies). It can be autonomously amplified to guide more transfer traffic to priority channels.

[0136] The dynamic path optimization process can be described using the following pseudocode:

[0137] Initialize parameters State space S, Action space A

[0138] For each batch of emergency supplies transfer mission:

[0139] Current status , according to Select

[0140] Get shipping feedback , Environment transfer to

[0141] Store memories and replay them periodically, continuously updating parameters.

[0142] If a sudden change in obstacle or priority is detected, the attention-relearning mechanism is triggered to optimize the path online.

[0143] End.

[0144] Through reinforcement learning-driven dynamic path optimization algorithms, the system can adapt to changes in logistics pathways under various emergency scenarios. It can quickly and optimally reconfigure paths in environments with complex dynamic obstacles and path priority adjustments, greatly improving the accuracy, timeliness, and safety of material transfer and supporting the efficient operation of the entire intelligent warehousing emergency system.

[0145] Step 5: Real-time feedback and iterative optimization

[0146] Finally, a comprehensive real-time feedback and iterative optimization mechanism for warehouse layout and transshipment is established. Utilizing intelligent sensors and data acquisition equipment, continuous monitoring of material distribution, storage security, and environmental changes is conducted. A self-regulating feedback algorithm (integrating Bayesian optimization and reinforcement learning self-adjustment strategies) continuously corrects the layout and transshipment paths. This ensures the entire warehousing system possesses dynamic adaptive capabilities, achieving optimal response to sudden demands.

[0147] In the final stage of intelligent emergency supplies warehousing management, the construction of a real-time feedback and iterative optimization mechanism is the core to achieving continuous system self-adaptation and ultimate responsiveness. This step adopts a closed-loop system of "perception-analysis-feedback-optimization," fully integrating IoT intelligent sensing, data-driven analysis, and advanced self-regulating optimization algorithms to achieve dynamic and continuous optimization of warehouse layout and transfer routes. It deeply integrates Bayesian optimization with Reinforcement Learning (RL) to create a high-level mechanism that combines feedback, prediction, and adaptive decision-making.

[0148] First, the real-time monitoring layer utilizes smart sensors (including RFID, infrared imaging, temperature and humidity sensors, environmental and structural safety monitoring devices, etc.) deployed at nodes within the warehouse and along key flow routes to collect multi-dimensional data in real time, including material distribution, inbound and outbound dynamics, temperature and humidity safety environment, equipment usage status, and channel utilization. This data is presented in time-series format. The data flows into the central management system in real time, and is combined with a big data processing and analysis framework to complete situation identification and anomaly warning.

[0149] Building upon this foundation, the feedback and correction phase begins. To address discrepancies between actual warehouse operations and forecasts / contingency plans, a real-time feedback mechanism analyzes current layout bottlenecks in material flow, reduced space utilization, abnormal energy consumption, or safety hazards, automatically issuing adjustment signals. This mechanism is based on a self-regulating feedback algorithm, and its structure includes two main modules: Bayesian optimization and reinforcement learning.

[0150] 1. Bayesian optimization feedback mechanism

[0151] Bayesian optimization is suitable for searching the parameter space of complex, continuous, and globally unanalyzable layouts. Its basic idea is to create a "surrogate model" for layout optimization during historical execution, encoding each layout or path scheme as a parameter vector. Let the performance objective function be... (For example, overall collaborative efficiency, emergency response timeliness, energy consumption, etc.). An approximation of the objective function is established using Bayesian inference methods such as Gaussian processes (GP).

[0152]

[0153] in, Represents the mean function, The kernel function reflects the correlation between different schemes. In each round, based on the latest observation data, the next optimal layout adjustment direction is selected using an acquisition function (e.g., to improve EI).

[0154]

[0155] As more observations and feedback are received, the optimization process continuously approaches the global optimum.

[0156] 2. Reinforce learning self-adjustment strategies

[0157] The reinforcement learning module model for warehouse layout and transshipment optimization is a Markov Decision Process (MDP), whose core state variables are... This includes the current distribution of supplies, available space, traffic bottlenecks, environmental safety status, etc., and actions. This includes strategies such as zoning adjustments, route changes, shelf adjustments, and temporary traffic diversion. Real-time feedback. Measured by actual performance improvement or risk reduction. Algorithms such as Deep Q-Network (DQN), policy gradient, or Actor-Critic are used to design a cumulative reward objective.

[0158]

[0159] in, In order to adjust the strategy, This serves as a discount factor. Through trial-and-error learning and environmental feedback, the RL intelligent agent can automatically discover efficient and continuously optimized solutions. For example, when faced with the pressure of concentrated inbound and outbound operations of high-priority materials, the intelligent agent proactively expands key channel zones, temporarily optimizes routes, and even dynamically adjusts equipment scheduling priorities to eliminate bottlenecks in real time.

[0160] The two optimization mechanisms complement each other: Bayesian optimization focuses on probabilistic exploration of the global parameter space and accelerating convergence, while reinforcement learning efficiently discovers local action sequences and environmental adaptation strategies. In a practical closed-loop deployment, the system framework is as follows:

[0161] 1. The sensing end continuously collects data and calculates anomalies and disturbances in real time (such as sudden large-scale scheduling or drastic environmental changes);

[0162] 2. The analysis and decision-making process uses Bayesian optimization to provide overall parameter fine-tuning suggestions, while the RL strategy optimizes the action sequence for specific spatiotemporal scenarios;

[0163] 3. The execution system reconstructs partition boundaries, allocates resources, and optimizes paths in real time based on the suggestions, and synchronously feeds back the execution results to the data stream, completing the self-evolution of "trial-modification-optimization".

[0164] At the mathematical level, the intrinsic objective of closed-loop optimization is to minimize the multi-stage, multi-objective loss function:

[0165]

[0166] Among them, Spatial allocation imbalance, for Loss of transit efficiency For security risks, Energy consumption, etc., is accumulated by weighting from continuously updated data and evaluation indicators. Parameters It dynamically adjusts in real time according to the environment.

[0167] Ultimately, this mechanism ensures that the warehousing system can flexibly cope with various uncertainties such as surges in demand, equipment malfunctions, and sudden environmental changes. It can achieve system-level self-adaptation and extreme optimization without human intervention, greatly improving the intelligence level and service stability of warehousing and emergency logistics.

[0168] like Figure 3 As shown, a material storage optimization layout system considering dynamic demand in emergency situations is described. The system is applicable to the method described above and includes:

[0169] The dynamic demand perception and prediction module is used to integrate multi-source data and adopt a dynamic probabilistic convolutional neural network (DP-CNN) to realize real-time dynamic demand perception and prediction, as well as multi-modal data feature fusion and environment embedding.

[0170] The material classification and priority management module is used to collect multi-dimensional attributes of materials, adopt an adaptive weight sorting algorithm and structured score analysis to classify and prioritize materials, and dynamically adjust the weight allocation.

[0171] The multi-level partition layout generation module automatically divides warehouse space into partitions based on dynamic needs and material priorities. It adopts a partition hybrid binary search method, genetic optimization, and random perturbation heuristic algorithm, combined with main channel and safety facility constraints, to achieve dynamic optimization and adjustment of partition boundaries.

[0172] The critical node path optimization module abstracts the warehousing environment into a directed graph, models the transfer path decision, and performs path autonomous learning, dynamic optimization and reconfiguration based on deep neural networks and reinforcement learning to adapt to sudden environmental changes.

[0173] The real-time feedback and iterative optimization module, based on intelligent sensors and data acquisition equipment, continuously monitors the warehouse status and environment. It adopts a self-regulating feedback mechanism that combines Bayesian optimization and reinforcement learning to perform real-time correction and adaptive optimization of warehouse layout and transfer routes.

[0174] The human-computer interaction and monitoring terminal is used to display warehouse status, optimization suggestions and decision results, receive operation instructions and abnormal alarms, and assist managers in emergency decision-making and response.

[0175] Example:

[0176] I. Scene Description

[0177] Suppose a major city receives a disaster warning, the Civil Defense Office needs to optimize the layout of materials in underground warehouses to ensure the supply of evacuation and survival materials for residents, and maximize emergency response capabilities and safe storage efficiency.

[0178] II. Dynamic Demand Perception and Predictive Modeling

[0179] 1. Sample form for multi-source data acquisition

[0180]

[0181] 2. Sample table of dynamic demand forecasting results (DP-CNN output)

[0182]

[0183] III. Material Classification and Priority Ranking

[0184] 1. Material Attribute Collection Matrix

[0185]

[0186] 2. Priority sorting and quantile partitioning table

[0187]

[0188] IV. Generation of Multi-level Partition Layout

[0189] 1. Zoning Area and Attribute Allocation Table

[0190]

[0191] V. Optimization of Critical Node Paths

[0192] 1. Warehouse Key Nodes and Path Status Table

[0193]

[0194] VI. Real-time feedback and iterative optimization

[0195] 1. Monitoring and Feedback Data Table

[0196]

[0197] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0198] It should be understood that the above detailed description of the technical solutions of the present invention with reference to preferred embodiments is illustrative and not restrictive. Those skilled in the art can modify the technical solutions described in the embodiments or make equivalent substitutions for some of the technical features based on reading this specification; however, these modifications or substitutions do not cause the essence of the corresponding technical solutions to depart from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for optimizing the layout of material storage considering dynamic demand in emergency situations, characterized in that: The method includes: Acquire multi-source data, and perform demand prediction based on the multi-source data using a dynamic probabilistic convolutional neural network to obtain dynamic demand perception and prediction results; Based on the dynamic demand perception and prediction results, materials are classified and prioritized to obtain the priority of materials. The priority is determined by an adaptive weight ranking algorithm based on the importance of emergency response, urgency of materials and sensitivity of the supply chain. Based on the dynamic demand perception and prediction results and the priority of the materials, a multi-level partitioning generation algorithm is used to generate a partition layout, wherein the partition boundaries of the partition layout are adjusted according to real-time demand changes. Based on the partitioned layout, a dynamic path optimization algorithm is used to optimize the path of key nodes to obtain the transfer path. The transfer path is dynamically reconfigured according to the actual flow of emergency supplies entering and leaving the warehouse. Based on monitoring data acquired by intelligent sensors and data acquisition equipment, the partition layout and the transfer path are iteratively optimized through a self-adjusting feedback algorithm.

2. The method according to claim 1, characterized in that: The process of acquiring multi-source data and performing demand prediction based on the multi-source data using a dynamic probabilistic convolutional neural network to obtain dynamic demand perception and prediction results includes: The multi-source data is obtained by acquiring disaster early warning information, social media public opinion information, historical disaster event records, and information collected by IoT monitoring devices. The multi-source data is standardized, missing value imputation is performed, and noise filtering is performed. The text data in the multi-source data is vectorized using a text content recognition algorithm to obtain multimodal fusion features. The multimodal fusion features are input into the dynamic probabilistic convolutional neural network to obtain the dynamic demand perception and prediction results.

3. The method according to claim 1, characterized in that: Based on the dynamic demand perception and prediction results, the materials are classified and prioritized to obtain the material priorities, including: Obtain the emergency response importance attribute, demand urgency attribute, and supply chain sensitivity attribute for each type of material, and generate a multi-dimensional attribute matrix based on the emergency response importance attribute, the demand urgency attribute, and the supply chain sensitivity attribute; Based on the multidimensional attribute matrix, the priority of the materials is determined by an adaptive weight sorting algorithm; The adaptive weight ranking algorithm assigns an initial weight to each attribute and adjusts the initial weight in real time according to the dynamic information of the emergency environment. The adaptive weight ranking algorithm is also used to calculate a structural score, which is used to characterize the demand correlation between materials. The initial weight is adaptively adjusted by combining the structural score and the real-time emergency coefficient to obtain the adjusted weight.

4. The method according to claim 1, characterized in that: Based on the dynamic demand perception and prediction results and the priority of the materials, a multi-level partitioning algorithm is used to generate a partition layout, including: Based on the dynamic demand perception and prediction results and the priority of the materials, the materials are grouped. Combining the material mobility attributes and security requirement attributes, the warehouse space is divided into multiple partitions, where each partition corresponds to a priority and a material attribute. Based on the required area and partitioning requirements of the partition, the boundary of the partition is iteratively optimized by using a partition hybrid binary search method combined with a genetic optimization algorithm and a random perturbation heuristic algorithm to obtain the adjusted partition boundary. Based on the constraints of the main passage, entrances and exits, and safety facilities, the zoning layout is determined. In the zoning layout, the high-priority material zones are located in the area near the entrances and exits, and the hazardous materials zones and special materials zones are isolated from the main flow area.

5. The method according to claim 1, characterized in that, Based on the partition layout, a reinforcement learning-driven dynamic path optimization algorithm is used to optimize the paths of key nodes to obtain the transit path, including: The warehousing environment is abstracted as a directed graph, which includes multiple key nodes and multiple channels. Each channel corresponds to dynamic passage delay, congestion level, obstacle status, and priority path marking. Based on the directed graph, the transfer path decision is modeled using state space, action space, and immediate reward function to obtain the path decision model; The path decision model is input into a deep neural network to obtain the transfer path. The deep neural network is embedded with an attention module and a multi-head output mechanism. The deep neural network is used to autonomously learn and dynamically optimize the multi-target transfer path.

6. The method according to claim 1, characterized in that: The monitoring data acquired by the intelligent sensors and data acquisition devices is used to iteratively optimize the partition layout and the transfer path through a self-adjusting feedback algorithm, including: The system uses intelligent sensors and data acquisition devices to monitor material distribution information, storage security information, and environmental change information in real time, obtains the monitoring data, and performs data-driven analysis and anomaly warning based on the monitoring data. The self-adjusting feedback algorithm is obtained by fusing the Bayesian optimization algorithm with the reinforcement learning self-adjusting strategy. Based on the self-adjusting feedback algorithm, the operating parameters and execution status corresponding to the monitoring data are analyzed, and the partition layout and the transfer path are dynamically corrected to achieve parameter reconstruction.

7. The method according to claim 3, characterized in that, The step of determining the priority of the materials based on the multidimensional attribute matrix using an adaptive weighted sorting algorithm includes: The materials are then weighted and comprehensively scored based on the adjusted weights, and the materials are sorted in descending order based on the weighted comprehensive scores to obtain the sorting results. Based on the arrangement results, the materials are divided into high priority, medium priority and low priority using the quantile method.

8. A material storage optimization layout system considering dynamic demand in emergency situations, the system being used to implement the method of any one of claims 1 to 7, characterized in that: The system includes: The dynamic demand perception and prediction module is used to acquire multi-source data, perform demand prediction based on the multi-source data through a dynamic probabilistic convolutional neural network, obtain dynamic demand perception and prediction results, and perform multi-modal feature fusion and environment embedding. The material classification and priority management module is connected to the dynamic demand perception and prediction module. It is used to obtain the multi-dimensional attributes of materials, determine the priority of materials based on the multi-dimensional attributes through an adaptive weight sorting algorithm, and dynamically adjust the weights based on the structured scores. A multi-level partition layout generation module, connected to the material classification and priority management module, is used to generate a partition layout based on the dynamic demand perception and prediction results and the priority of the materials. The partition boundary is dynamically adjusted by using a partition hybrid binary method combined with a genetic optimization algorithm and a random perturbation heuristic algorithm. The key node path optimization module, connected to the multi-level partition layout generation module, is used to abstract the warehousing environment into a directed graph and dynamically optimize and reconfigure the transfer path based on a dynamic path optimization algorithm driven by deep neural networks and reinforcement learning. The real-time feedback and iterative optimization module is connected to the multi-level partition layout generation module and the key node path optimization module, respectively. It is used to iteratively optimize the partition layout and transfer path based on the monitoring data obtained by the intelligent sensor and data acquisition device through a self-adjusting feedback algorithm. The self-adjusting feedback algorithm is obtained by fusing the Bayesian optimization algorithm and the reinforcement learning self-adjusting strategy. The human-computer interaction and monitoring terminal is connected to each module to display the warehouse status, optimization suggestions and decision results, and to receive operation instructions and abnormal alarm information.

Images