A method and system for allocating human resources for logistics in multiple areas of one hospital

Abstract

The present invention proposes a method and system for configuring human resources in logistics of one hospital with multiple regions. By constructing a knowledge graph of hospital logistics scenarios and configuration requirement templates, the system uses language understanding and multimodal information processing capabilities to achieve intelligent analysis of human resource configuration requirements and constraints in different business scenarios, and combines a multi-objective optimization algorithm to design a set of dynamic resource allocation mechanisms, which can adaptively generate the optimal human resource allocation plan that meets the needs of multi-regional collaboration and dynamic business. Theoretical analysis and experimental results show that this method can significantly improve the intelligent level and management efficiency of hospital logistics human resource configuration, and provide strong support for the high-quality operation of one hospital with multiple regions.

Description

Technical Field

[0001] The present invention relates to the technical field of resource allocation, and in particular to a method and system for allocating logistics human resources in multiple areas of a hospital. Background Art

[0002] One hospital with multiple districts refers to a construction model in which a hospital follows a set of management systems, takes one district as the center, and has multiple districts to provide medical services. This model helps to achieve the dislocation and complementarity of medical resources and the coordinated and efficient operation, but it also brings new challenges to the logistics support work of the hospital. On the one hand, different regions face differentiated business needs and service objects, and need to carry out targeted services in accordance with local conditions under the premise of unified management; on the other hand, the rapid development of clinical business and the dynamic changes in the policy environment require the logistics department to be able to respond flexibly and quickly. In addition, the workload and work intensity of different positions vary greatly, and many factors such as employees' personal wishes, ability endowments, and experience levels must also be considered. How to achieve the reasonable allocation and dynamic optimization and adjustment of cross-regional logistics human resources under complex and changing constraints has become a new issue facing hospital managers.

[0003] Traditional methods are mainly based on mathematical programming models, such as linear programming, integer programming, and goal programming, and solve the optimal personnel combination plan by constructing objective functions and constraint equations. The advantages of this type of method are rigorous modeling and solid theoretical foundation. The limitations are that it is difficult to accurately describe the complex semantic constraints existing in actual business, and the solution efficiency drops sharply as the problem scale increases.

[0004] In view of the complexity of human resource allocation in real-world scenarios, heuristic algorithms, such as genetic algorithms, ant colony algorithms, and particle swarm algorithms, have also been applied to approach the global optimal solution through intelligent search. This type of method has improved the efficiency and quality of the solution to a certain extent, but it still faces the problem that the constraints are difficult to accurately model. Moreover, most of them require manual design of complex encoding and decoding mechanisms, and the generalization ability is insufficient.

[0005] In recent years, with the development of big data and artificial intelligence technology, some studies have begun to try to use machine learning methods to optimize human resource allocation. For example, patent document CN112488543A proposes a smart construction site intelligent scheduling method and system based on machine learning. In view of the problem that there are many types of work and a large number of workers in large-scale engineering projects, and manual scheduling is difficult to meet the needs, a smart construction site intelligent scheduling method based on machine learning is proposed. This method takes into account multiple factors and can help improve the safety and economy of construction site construction and the enthusiasm of workers.

[0006] Although the above methods have achieved certain results in their respective application scenarios, they have not fully considered the practical challenges faced by human resource allocation in the special situation of a hospital with multiple districts, such as multi-regional coordination, dynamic business changes, and diversified configuration constraints. In addition, most existing methods rely on structured input data, lack semantic understanding and flexible modeling capabilities for unstructured business needs, and are difficult to adapt to the actual complex and changing application environment. In view of this, the present invention innovatively introduces , using its powerful language understanding and knowledge extraction capabilities, build a knowledge graph and demand analysis mechanism for the logistics scenario of one hospital and multiple regions, and combine end-to-end semantic understanding with advanced combinatorial optimization algorithms to continuously optimize human resource allocation while dynamically perceiving changes in business context, in order to achieve breakthroughs in intelligence, adaptability and practicality.

[0007] To solve the above problems, the present invention proposes a method for intelligently configuring logistics human resources in multiple areas of a hospital. With advantages in language understanding, multimodal information processing and knowledge reasoning, the hospital accurately maps business requirements and configuration constraints into structured mathematical models by constructing knowledge graphs of hospital logistics scenarios and templates for describing human resource configuration requirements, and then uses advanced heuristic intelligent optimization algorithms to solve the optimal configuration plan. At the same time, in view of the dynamic and changing business requirements of multiple districts in a hospital, a continuous learning optimization mechanism is designed to capture demand changes in a timely manner and dynamically adjust configuration plans to achieve accurate deployment and real-time optimization of human resources. Summary of the invention

[0008] On the one hand, the present invention provides a method for allocating logistics human resources for multiple departments in one hospital, which specifically includes the following steps.

[0009] Step 1: Build a hospital logistics knowledge graph.

[0010] Step 1.1: Define the ontology model layer and standardize the concept categories, attributes and relationships. The core concepts defined in this invention include hospital, region, department, position, employee, shift, skill, etc., which are formally expressed as:

[0011] ;

[0012] in, Indicates concept categories; Indicates A conceptual relationship.

[0013] Step 1.2: Extract instance-level information from heterogeneous data sources such as hospital management texts and human resources systems. Use pre-trained language models Learn semantic representations of medical professional words and concept phrases and perform named entity recognition The technology extracts conceptual entities from text, which are formally expressed as:

[0014] ;

[0015] in, is the set of conceptual entities extracted; A collection of relationships between entities.

[0016] Step 1.3: Match the extracted instance-level information with the ontology schema and obtain a standardized knowledge graph after disambiguation ,in and Represent the concept entities and relations after disambiguation, respectively.

[0017] Step 2: Multimodal semantic analysis of scheduling requirements.

[0018] Step 2.1: Exploitation The scheduling needs of heterogeneous forms such as language, voice, and image Mapping to a unified semantic space and learning task-independent contextual representations:

[0019] ;

[0020] Step 2.2: In the unified semantic representation Based on the above, through a series of fine-grained language understanding such as lexical analysis, dependency syntax analysis, and semantic role labeling, the structured scheduling demand elements are extracted and formally expressed as follows:

[0021] ;

[0022] in, It includes multi-dimensional information such as task type, time requirements, job capabilities, etc.

[0023] Step 2.3: Structure your scheduling requirements Hospital Knowledge Graph Match, complete and expand the requirement information to obtain a standardized task request description .

[0024] Step 3: Generate an initial task allocation plan. is the set of tasks to be assigned, A collection of available employees.

[0025] Step 3.1: Assign each task Requirements description and every employee Attribute portrait Enter together , learn the matching distribution of tasks and employees:

[0026] ;

[0027] Step 3.2: According to the matching probability Sampling generates a set of initial allocations ,in Indicates that the task in the initial plan Is it assigned to employees? .

[0028] Step 4: Evaluate the quality of the initial solution. Each solution in enter , and design language guide Evaluate the quality of the solution from multiple dimensions to get a comprehensive score:

[0029] ;

[0030] in, A normalized score representing the quality of the solution, taking into account multiple objectives such as task matching, employee satisfaction, and scheduling fairness.

[0031] Step 5: Iterate and optimize the allocation scheme.

[0032] Step 5.1: Current Allocation Scheme Each solution in ,use The encoder extracts its semantic features :

[0033] ;

[0034] Step 5.2: Compare Perform knowledge-guided perturbations to generate new candidate solution representations :

[0035] ;

[0036] in, pass guide Generated, it contains the scheduling knowledge learned from historical optimization cases.

[0037] Step 5.3: Exploitation The decoder will perturb the Restore to new readable allocation :

[0038] ;

[0039] at the same time, It also accepts constraints described in the form of language , the feasibility requirement is automatically met during the decoding process:

[0040] ;

[0041] Step 5.4: Evaluate new options Quality If it is better than the current solution, add it to the new solution set .

[0042] Step 5.5: Repeat steps 5.1-5.4 until the iteration termination condition is met and the final optimal scheduling solution is obtained. .

[0043] Step 6: Dynamically update the shift schedule.

[0044] Step 6.1: Continuously monitor real-time information flows such as employee feedback, emergent requests, etc. ,use Perform semantic analysis and extract key information:

[0045] ;

[0046] Step 6.2: Update the information As a new constraint, insert it into the iterative optimization process of step 5 and dynamically adjust the optimal solution:

[0047] ,in Indicates at a point in time The optimal scheduling solution, is the scheme quality scoring function to evaluate the scheduling scheme The overall quality score of Indicates the search for The biggest plan , Indicates that the subsequent constraints are satisfied;

[0048] Step 6.3: Repeat steps 6.1-6.2 periodically to achieve dynamic update and continuous optimization of the shift scheduling plan, and use The interpretability of the system can automatically generate language descriptions for adjusting decisions, enabling intelligent management decisions.

[0049] On the other hand, we proposed an intelligent logistics human resource allocation system for a multi-district scenario in one hospital. The structure diagram of the system is shown in the following figure. Figure 1 As shown in the figure, the system consists of the following key modules:

[0050] Knowledge Management Module:

[0051] This module is responsible for the integration and semantic modeling of hospital logistics business information, mainly including:

[0052] (1) Ontology management: Define the conceptual hierarchy of business entities, attributes, and relationships to form a standardized business ontology.

[0053] (2) Knowledge extraction: using pre-trained Automatically extract structured information from heterogeneous data sources such as hospital management systems, human resources systems, and equipment ledgers, and match and map them with the ontology to form a rich and complete logistics knowledge graph :

[0054] ;

[0055] in, Represents a collection of business entities. Represents a set of relationships between entities. Represents a collection of attributes for an entity.

[0056] (3) Knowledge reasoning: Based on the knowledge graph, the ontology reasoning mechanism is used to reveal implicit business rules and constraints, further enriching the semantic associations of knowledge.

[0057] Scheduling requirements understanding module:

[0058] This module utilizes , achieving flexible and natural expression and analysis of scheduling needs.

[0059] (1) Requirement input: Supports unstructured requirement descriptions in various forms such as text, voice, and images .

[0060] (2) Semantic analysis: The heterogeneous input Unified mapping to semantic vector :

[0061] ;

[0062] Then, through fine-grained analysis such as information extraction and semantic role labeling, Extracting structured scheduling elements from , including task type, time requirements, capability requirements, etc.:

[0063] ;

[0064] (3) Demand integration: structural demand With Knowledge Graph Perform matching associations, automatically expand the semantic information of the request, correct inconsistencies, and ultimately form a standardized scheduling requirement description .

[0065] Task optimization scheduling module:

[0066] This module requests a given task and human resources , and give a dynamically optimal task allocation solution.

[0067] (1) Initial allocation: using knowledge-enhanced language models , learn matching distribution based on employee profile and task requirements , generate initial task-employee assignment:

[0068] ;

[0069] (2) Iterative Optimization: In the Knowledge Graph In the constraint space of ,from Generate new candidate solutions :

[0070] ;

[0071] Simultaneous use Comprehensively score candidate solutions for different goals , iterative optimization until convergence, and obtain the dynamic optimal solution .

[0072] (3) Solution explanation: Utilization The text generation capability of the system can automatically generate language explanations for scheduling decisions, enhancing the explainability of management.

[0073] Online learning update modules:

[0074] This module is responsible for continuously optimizing the system's scheduling strategy based on real-time feedback.

[0075] (1) Real-time monitoring: By integrating with human resource management systems and employee mobile applications, it can sense employee feedback, task status, environmental changes and other information flows in real time. .

[0076] (2) Incremental learning: using From the information flow Continuously learn new scheduling rules , and incrementally update the graph and matching model , so that the system can adapt to the dynamic changes of business.

[0077] (3) Strategy self-optimization: Periodically trigger steps (2) and (3) to generate a better scheduling strategy under new knowledge constraints. , achieving continuous self-optimization of the system.

[0078] Human-machine collaborative operation module:

[0079] This module utilizes Realize natural interaction between the scheduling system and managers.

[0080] (1) Voice Q&A: Managers can use voice to ask questions related to scheduling, such as "What is the staffing arrangement for the physical examination center next Monday?" Perform semantic analysis on the question and return the answer through knowledge graph retrieval.

[0081] (2) Conversational analysis: Managers can have multiple rounds of conversations with the system to explore the pros and cons of different scheduling strategies. Multiple rounds of reasoning and comparative analysis can be performed based on the dialogue context to provide qualitative and quantitative decision-making references.

[0082] (3) Suggestions for solution optimization: It can proactively provide managers with suggestions for optimizing the current shift schedule and explain the basis for the optimization and expected results. Managers can choose to adopt, modify or reject the suggestions and participate in the closed-loop iteration of the shift schedule decision.

[0083] The above functional modules cooperate with each other to form an intelligent human resource allocation system. This system enables the hospital logistics scheduling to flexibly respond to dynamic and changing business needs, continuously optimize human resource allocation, and significantly improve management efficiency and decision-making level.

[0084] The beneficial technical effects of the present invention are as follows:

[0085] (1) Based on the traditional mathematical programming model, we innovatively introduce , realize intelligent analysis of unstructured business demand descriptions, build knowledge graphs and resource allocation demand templates for hospital logistics scenarios, and greatly improve the modeling capabilities and business adaptability of configuration constraints.

[0086] (2) Aiming at the specific scenario of cross-regional human resource collaborative allocation in hospitals, a multi-objective optimization resource scheduling algorithm is designed to achieve the optimal balance between human resource utilization and employee satisfaction while taking into account complex constraints such as employee personal endowments and multi-position collaboration.

[0087] (3) Innovatively propose a continuous learning optimization mechanism. Through semantic understanding of multi-source heterogeneous information flows, we can continuously capture the dynamic changes in hospital business needs and achieve real-time dynamic optimization and adjustment of human resource allocation plans. BRIEF DESCRIPTION OF THE DRAWINGS

[0088] Figure 1 This is a system structure diagram of the one-hospital-multi-zone human resources allocation system provided by the present invention. DETAILED DESCRIPTION

[0089] In order to solve the problems of cross-regional coordination difficulties, dynamic and changing human resource demand, and complex and diverse configuration constraints in the logistics management of multiple districts in one hospital, this paper introduces , using its powerful language understanding and contextual learning capabilities, it realizes an intelligent closed loop of the entire process.

[0090] Specifically, first use Extract key information from multi-source heterogeneous data such as hospital management system texts, job descriptions, and employee portraits, and perform structured representations of business scenarios, human resource attributes, and associated constraints in the form of knowledge graphs to form a globally unified semantic representation space. On this basis, we use the knowledge graph to represent the scheduling requirements proposed by human resource managers in the form of language. Perform multimodal semantic analysis to extract structured resource requests and constraints, map them to the knowledge graph for matching and reasoning, and form a standardized resource allocation demand description. Then, comprehensively consider various factors such as business work saturation, employee personal endowments and preferences, job synergy effectiveness, etc., take employee satisfaction and unit time output maximization as the objective function, and take business completion constraints and employee preference constraints as the basic constraints to construct a multi-objective combination optimization model, and design a solution algorithm to give the optimal human resource allocation plan. Finally, in order to dynamically adapt to changes in hospital business needs, through Through semantic analysis of abnormal business events and employee feedback, combined with historical scheduling effect evaluation, human resource allocation strategies are continuously optimized, and humanized scheduling explanations are generated to form an explainable and credible closed loop of human-machine collaboration.

[0091] Knowledge Graph ) is a structured knowledge representation method that intuitively depicts the internal connections of domain knowledge in the form of a graph by defining the categories, attributes and semantic associations of conceptual entities. Information is extracted from heterogeneous data sources such as hospital management system texts, human resources systems, and employee resumes, and semantic integration is performed to ultimately construct a knowledge graph that reflects the logistics business scenarios of multiple regions in a hospital, providing a unified knowledge representation foundation for subsequent intelligent human resource configuration.

[0092] The entire knowledge graph construction process follows a bottom-up approach, from data to model. Information extraction is performed on multi-source heterogeneous texts to identify core concept entities such as positions, employees, regions, and skills, and semantic associations between entities are mined (such as employees have skills and positions belong to regions) to obtain instance layer data of the graph; then, through knowledge ontology definitions, these concept entities and relationships are abstracted and normalized to form a category system and attribute pattern that reflects the core concepts of the field, and the pattern layer of the graph is established; finally, the pattern layer is matched and mapped with the data layer, and the graph is completed using ontology reasoning to form a complete and high-quality knowledge graph.

[0093] Specifically, knowledge graph construction mainly includes the following key steps:

[0094] (1) Construct the graph model layer based on the ontology definition. With the help of domain expert knowledge, define the ontology that reflects the core concepts of logistics business, and standardize the concept categories, attributes, relationships and other elements. The core concepts defined in this invention mainly include hospitals, regions, departments, positions, employees, shifts, skills, etc.

[0095] (2) Using named entity recognition for texts such as hospital management systems, employee rosters, equipment manuals, and job manuals ( ) technology identifies key concepts and their attribute values ​​related to the domain. Most of the methods are based on dictionary matching and rule templates, and their generalization ability is insufficient. This invention introduces a pre-trained language model , through continuous pre-training on hospital text corpus, the generalization effect of named entity recognition has been significantly improved, especially for the discovery of composite entities and new entity types. Examples of recognized entities are as follows:

[0096] [Hospital] XX Hospital, First Campus [Department] Dietetic Department [Position] Nutritionist [Shift] Regular day shift (8:30-17:00) [Staff] Wang Meiling (female) [Skills] Clinical nutrition assessment, dietary formula design;

[0097] [Hospital] XX Hospital Second Campus [Department] Equipment Department [Position] Maintenance Engineer [Shift] On-call shift (24-hour shift) [Employee] Li Daming (male) [Skills] Medical equipment maintenance, circuit fault diagnosis;

[0098] [Hospital] XX Hospital, District 3 [Department] Security Department [Position] Security Captain [Shift] Night shift (18:00-6:00 the next day) [Employee] Zhang Weiguo (male) [Skills] Security prevention, emergency handling;

[0099] Among them, labels such as [hospital] and [department] correspond to concept categories defined in the knowledge graph model layer, and the text after the label is the specific entity word. It can learn distributed representations of concepts and attributes from multiple levels such as lexical, syntactic, and semantic levels, overcoming the limitations of artificial feature engineering and having better transferability and robustness.

[0100] (3) Extract employee resumes, scheduling histories, performance appraisals and other information from the structured records of the hospital human resources management system and integrate them with the unstructured information in step (2) to further enrich the attributes of the conceptual entity.

[0101] (4) Using open information extraction ( ) technology discovers the semantic association between conceptual entities from domain text. Traditional methods based on template matching and heuristic rules are difficult to discover new relationship types. By continuously learning on large-scale corpora, it can discover more implicit and long-distance semantic dependencies and automatically summarize new relationship patterns based on co-occurrence statistics and semantic similarity. Some examples of extracted relationships are as follows:

[0102] Relationship of core medical positions: [Hospital: XX Hospital First Campus]-[Owns]-[Area: Inpatient Department], [Area: Outpatient Building]-[Includes]-[Department: Cashier's Office], [Position: Nurse]-[Requirements]-[Skills: Cardiopulmonary Resuscitation], [Position: Doctor]-[Use]-[Equipment: CT Machine], [Employee: Zhang San]-[Belongs to]-[Department: Outpatient Nursing Station];

[0103] Relationship of logistics support positions: [Department: Dietary Department]-[Management]-[Equipment: Central Kitchen], [Position: Nutritionist]-[Responsible for]-[Business: Clinical Nutrition Consultation], [Department: Dietary Department]-[Execution]-[Standard: Food Hygiene Standard], [Position: Nutritionist]-[Use]-[System: Nutrition Management System], [Employee: Wang Meiling]-[Hold]-[Certificate: Clinical Nutritionist Qualification Certificate];

[0104] Equipment maintenance position relationship: [Department: Equipment Section]-[Maintenance]-[Equipment: Medical Equipment], [Position: Maintenance Engineer]-[Hold]-[Certificate: Medical Equipment Maintenance Qualification], [Employee: Li Daming]-[Specialization]-[Field: Radiological Equipment Maintenance], [Position: Maintenance Engineer]-[Required]-[Skills: Electrician Certificate], [Department: Equipment Section]-[Responsible for]-[Area: Equipment Room];

[0105] Security management position relationship: [Department: Security Office]-[Patrol]-[Area: Outpatient Building], [Position: Security]-[Equipment]-[Equipment: Intercom], [Department: Security Office]-[Responsible for]-[Facilities: Monitoring Room], [Position: Security Captain]-[Management]-[System: Access Control System], [Employee: Zhang Weiguo]-[Holds]-[Certificate: Security Certificate];

[0106] Information technology job relationships: [Department: Information Department]-[Maintenance]-[System: Hospital Information System], [Position: IT Engineer]-[Management]-[Equipment: Server], [Department: Information Department]-[Support]-[Business: Telemedicine], [Position: IT Engineer]-[Required]-[Certification: Information System Engineer Certificate], [Employee: Chen Siyuan]-[Proficient]-[Technology: Network Security];

[0107] Drug management position relationship: [Department: Pharmacy Department]-[Management]-[Region: Pharmacy], [Position: Pharmacist]-[Review]-[Document: Prescription], [Department: Pharmacy Department]-[Execution]-[Specification: Drug Management System], [Position: Pharmacist]-[Operation]-[System: Intelligent Dispensing System], [Employee: Zhao Ming]-[Hold]-[Certificate: Licensed Pharmacist Certificate];

[0108] Medical equipment job relationships: [Department: Medical Engineering Department]-[Maintenance]-[Equipment: Medical Gas System], [Position: Equipment Engineer]-[Calibration]-[Instrument: Medical Equipment], [Department: Medical Engineering Department]-[Formulation]-[Plan: Equipment Maintenance Plan], [Position: Equipment Engineer]-[Use]-[Tools: Testing Instruments], [Employee: Qian Wei]-[Possess]-[Qualification: Medical Equipment Repair Certificate];

[0109] Environmental management position relationship: [Department: Logistics Department]-[Purchasing]-[Supplies: Cleaning supplies], [Position: Cleaning]-[Operation]-[Equipment: Floor scrubber], [Department: Logistics Department]-[Implementation]-[Standard: Disinfection and Isolation Specifications], [Position: Cleaning Supervisor]-[Training]-[Content: Infection Prevention and Control Knowledge], [Employee: Liu Chunhua]-[Responsible for]-[Area: Operating Room Environment];

[0110] (5) Match the extracted conceptual entities, attributes and relationships with the ontology model of the knowledge graph, thereby converting unstructured data into a structured graph representation. The matching process needs to consider the polysemy disambiguation of conceptual words. The present invention adopts a disambiguation method based on representation learning. By continuously pre-training word vectors on large-scale text corpora, the distributed representations of synonymous concepts are closer in the vector space, thereby achieving robust link disambiguation.

[0111] (6) Based on the ontology reasoning mechanism, using the existing entity relationships and combining the axiom constraints defined in the ontology, new implicit associations are inferred to further complete the knowledge graph. For example, if "[Area: Operating Room]-[Equipment]-[Equipment: Medical Floor Scrubber]" and "[Position: Cleaner]-[Responsible for]-[Area: Operating Room]" are known, then a new relationship of "[Position: Cleaner]-[Should Master]-[Skills: Medical Floor Scrubber Operation]" can be derived. The present invention defines a set of domain reasoning rules for hospital scenarios, which can effectively improve the richness and quality of the knowledge graph.

[0112] After the above processing steps, a high-quality knowledge graph that comprehensively covers hospital logistics business scenarios can be constructed. On this basis, a unified semantic representation and reasoning basis can be provided for intelligent human resource allocation, enabling subsequent multimodal demand understanding and task optimization decision-making.

[0113] Traditional human resource scheduling systems usually require managers to submit structured scheduling requirements in fixed templates or special languages. The process of filling out forms is cumbersome and inefficient, and it is difficult to fully express actual business constraints. The present invention innovatively introduces a multimodal demand understanding mechanism, allowing managers to flexibly describe scheduling requirements in natural interactive ways such as text, voice, and images, such as "A medical expert seminar will be held in the West Campus tomorrow morning. Two nutritionists with group meal experience will be arranged to formulate the menu, three skilled chefs will be responsible for the preparation of special meals, and two experienced caterers will be arranged. The event will last for 4 hours and must be prepared 2 hours in advance." The system automatically performs semantic analysis, identifies key information such as the quantity, time, position, and capability requirements of resource requests, and then matches the concept entities in the knowledge graph to form a structured resource demand representation to guide subsequent intelligent scheduling.

[0114] The core of resource demand analysis is to make full use of Multimodal semantic understanding capabilities. Traditional language processing technologies are mostly based on dictionary matching, shallow syntactic analysis and domain templates. It is difficult to accurately grasp the deep semantics of the demand text, and it is not enough to deal with incomplete expressions, references, ambiguities and other phenomena. By pre-training on massive multimodal data, the intrinsic relationship between different modal information can be learned, and an end-to-end mapping from surface text, voice, and image to latent semantics can be established. It is converted into a unified semantic representation vector, and then fine-grained semantic structure analysis is carried out on this basis, including lexical analysis, syntactic analysis, semantic role labeling, coreference resolution, etc., so as to accurately extract the key elements of resource requirements.

[0115] Specifically, resource requirement analysis includes the following steps:

[0116] (1) Standardized representation of modal information. Since information of different modalities (such as text, speech, and image) differ greatly in the underlying feature space, unified semantic modeling is required. The powerful language understanding ability of the speech recognition model is firstly used to map the intrinsic relationship between different modal information into a shared semantic space by continuously learning the intrinsic relationship between different modal information. ) and then entered along with the requirements in plain text , and learn a unified distributed semantic representation. For image input (such as a handwritten shift schedule), scene text detection and recognition is first performed ( ) extract key information and then perform semantic mapping. Through multimodal semantic integration, different forms of needs can be uniformly portrayed, laying the foundation for subsequent accurate understanding.

[0117] 2) Fine-grained semantic structure analysis. After obtaining a unified semantic representation, the system uses powerful language modeling capabilities to conduct multi-level semantic structure analysis of resource requirements. First, word segmentation and named entity recognition at the lexical level accurately locate key concepts such as personnel, positions, time, and place; second, dependency analysis at the syntactic level characterizes structured relationships such as modification and parallelism between concepts; third, semantic role annotation at the semantic level reveals the predicate argument structure of the sentence. Taking "arrange 3 Chinese chefs to prepare working meals at the West Hospital Nutrition Department tomorrow" as an example, the annotated semantic roles are as follows: [Time: Tomorrow][Action: Arrange][Quantity: 3][Position: Chinese Chef][Location: West Hospital Nutrition Department][Task: Prepare working meals].

[0118] Semantic role annotation directly describes the key elements of demand and provides an important basis for the automatic extraction of resource attributes. Establishing an end-to-end mapping from vocabulary, syntax to semantic roles can significantly improve the accuracy and coverage of annotations.

[0119] (3) Co-reference resolution. In reality, scheduling requirements often involve pronoun reference and noun double reference, such as "Doctor Zhang San has two surgeries today. He needs to complete a physical examination before the surgery and then go to the operating room to prepare. He is familiar with this process." In this sentence, "he" refers to "Doctor Zhang San" who appeared earlier. In order to accurately understand the complete semantics, it is necessary to link "he" and "Doctor Zhang San" through co-reference resolution. Traditional heuristic rules are difficult to handle long-distance and cross-sentence reference. The present invention utilizes By modeling the long-range semantic associations between words, even if the antecedent and pronoun are far apart, the referential relationship between the two can be accurately revealed through the context, thereby improving the completeness of demand text understanding.

[0120] (4) Graph-based semantic matching. Through the above semantic structure analysis, we can extract structured key information from unstructured text, such as resource type, quantity, task duration, job competency, etc., to form a standardized resource demand representation. Furthermore, we match and link it with the conceptual entities in the knowledge graph. On the one hand, we can perform semantic disambiguation on the demand information and improve the accuracy of understanding; on the other hand, we can also reuse the prior knowledge in the graph and automatically expand the semantic information of the demand. For example, after "respiratory nurse" is mapped to the knowledge graph, the skills required by it, such as "endotracheal intubation" and "non-invasive ventilator", can be associated without explicit instructions from the management staff.

[0121] After obtaining the structured resource demand representation, the next step is to combine the current human resource inventory status to match and schedule the most suitable employees for each task request, while taking into account multiple factors such as employee preferences, costs, fairness, etc., to generate the globally optimal human resource scheduling plan, which can be formalized as a multi-objective combinatorial optimization problem.

[0122] Define that the hospital has m areas, each area has n tasks to be scheduled, involving p job roles, and the system goal is to assign the most suitable employees to these n*p task positions. is a 0-1 decision variable. Assigned to area Task The value is 1 when , otherwise it is 0. The objective function includes:

[0123] (1) Mismatch between resources and tasks :

[0124] ;

[0125] in For employees With area Task The comprehensive matching deviation in time, location, skills, experience and other dimensions is measured by the weighted Euclidean distance between employee attributes and task attributes. The smaller the matching deviation, the more matching the employee's ability is with the task requirements.

[0126] (2) Employee satisfaction :

[0127] ;

[0128] in For employees For Region Task The satisfaction of employees depends on factors such as the fit between the task time and employee preferences and the commuting distance. It is the employees' acceptance of the allocation plan, which is obtained from the historical scheduling data.

[0129] (3) Human cost :

[0130] ;

[0131] in For employees Standard hourly wage, For employees Acceptance area Task Additional costs such as overtime pay, night shift allowance, etc.

[0132] (4) Fairness :

[0133] ;

[0134] in is the standard deviation function, and fairness is measured by the inverse of the variance of the total number of employee shifts, reflecting the balance of total workload among employees.

[0135] In addition, the model contains a variety of complex constraints, including:

[0136] (1) Each task position has only one employee, that is:

[0137] ;

[0138] (2) Employees Total hours assigned Not exceeding the legal limit ,Right now:

[0139] ;

[0140] in For Region Task Standard working hours.

[0141] (3) Employee capabilities Must meet job qualification requirements ,Right now:

[0142] ;

[0143] (4) There are priority constraints between some tasks, which need to be completed in order, formalized as:

[0144] ;

[0145] in The priority relationship of the tasks.

[0146] In summary, the optimal allocation of hospital human resources can be formalized as a multi-objective 0-1 integer programming model:

[0147] ;

[0148] For the above complex combinatorial optimization model, traditional exact algorithms (such as branch and bound method, cutting plane method, etc.) often fail to give satisfactory solutions within an acceptable time when solving large-scale problems. Based on the classic heuristic algorithms (such as genetic algorithm, particle swarm optimization, etc.), this paper proposes a method based on A new method for solving multi-objective optimization. By learning the distribution characteristics of domain knowledge optimization solutions and autonomously summarizing the implicit patterns of optimal scheduling from large-scale historical scheduling cases, we can narrow the search space, guide the evolutionary direction, and avoid blind random exploration, thereby significantly improving the optimization efficiency and solution quality.

[0149] Specifically, the multi-objective optimization solution algorithm includes the following steps:

[0150] (1) Knowledge-guided solution space representation learning. Traditional heuristic algorithms mostly use artificially designed encoding strategies, such as binary strings, permutations, etc., which are difficult to fully reflect the domain characteristics of the scheduling problem. Powerful representation learning capabilities, through continuous pre-training on large-scale historical scheduling data, autonomously construct a continuous latent vector representation of the optimal solution space. The corresponding task requirements and employee attributes Enter together , by modeling the interaction of the three through the self-attention mechanism, a solution vector is learned that comprehensively considers the task characteristics and employee endowments :

[0151] ;

[0152] Learned latent representation It has good semantic properties. Closer points in the vector space correspond to more similar scheduling patterns. Therefore, we can directly conduct heuristic search in this continuous latent space and use random perturbations to find the optimal scheduling pattern. Generate new solutions, avoiding the combinatorial explosion problem of traditional discrete coding schemes. It can also be used as a decoder to transform the latent vector Restore to a readable schedule :

[0153] ;

[0154] The end-to-end representation of the solution space is achieved through equations (9) and (10), which transforms the solution of the optimization model (1)-(8) into a search problem in the latent space.

[0155] (2) Multi-objective fitness scoring. Traditional evolutionary algorithms evaluate the quality of candidate solutions by manually designing complex multi-objective fitness functions, such as weighted summation method and Pareto dominance method, which often require repeated adjustment of weight parameters. Strong contextual understanding, through carefully designed Learn the evaluation knowledge contained in historical data. For example, you can use the following ask Schedule Overall rating:

[0156] Given the following employee scheduling scheme:

[0157] [Description of shift scheduling]

[0158] Taking into account multiple factors such as matching quality, employee satisfaction, cost-effectiveness, and scheduling fairness, please rate the plan (0-100 points) and provide analysis from the perspectives of different stakeholders such as hospital managers, employees, and patients.

[0159] After few-shot training, the impact of different aspects can be weighed and a comprehensive and balanced score can be given. Experiments have found that this language-guided scoring function can well fit the multi-objective characteristics of the scheduling problem and avoid the tedious manual adjustment of parameters. Embedding scoring into the fitness evaluation of genetic algorithms can significantly improve the efficiency of selecting high-quality individuals in the population.

[0160] (3) Mutation operator based on prior knowledge. Mutation is a key link in evolutionary algorithms, which is used to expand the solution space and escape from local optimality. However, traditional random mutation strategies lack domain specificity and are prone to generate a large number of invalid searches. Learn the optimization rules in scheduling cases, summarize the semantic features of high-quality scheduling from historical data, and introduce them as prior knowledge into the mutation operator design. Specifically, given a scheduling plan to be mutated , first extract its latent representation , and then use The text generation capability of Add disturbances that conform to the optimization law on the basis of , and finally the perturbed representation Decode back to readable scheme , the whole mutation process can be expressed as:

[0161] ;

[0162] in, and Respectively encoder and decoder; is the knowledge-guided perturbation, guided by few-shot examples Compared with random perturbations, this prior-guided mutation is more inclined to explore areas that are likely to produce high-quality solutions, avoid blind attempts, and accelerate algorithm convergence.

[0163] (4) Case-based constraint processing. Scheduling problems usually involve complex and diverse constraints, and the solution process needs to ensure the feasibility of the solution at all times. To handle the flexibility of the language, set constraints Formatted as a descriptive text, used as an additional parameter during generation and mutation operations. enter , guiding it to automatically avoid invalid solutions. Formally, record The new solution after perturbation is expressed as follows. The constraint-based decoding process can be expressed as:

[0164] ;

[0165] in, is the language description of the constraint set, express Decoder. Experiments show that this soft constraint injection method can significantly reduce the probability of illegal solutions, maintain the feasibility of the population, and avoid the complex solution repair process in traditional algorithms. At the same time, since the constraint information is encoded in language form, the algorithm can flexibly handle descriptive constraints and has better applicability.

[0166] In summary, based on The core of the combinatorial optimization solution is to Learn prior scheduling knowledge and flexibly inject it into various links such as solution space representation, fitness evaluation, mutation search, constraint processing, etc., so as to build an efficient, explainable and robust solution framework.

[0167] The present invention also conducts manual evaluation on the interpretability of the generated scheduling scheme. Automatically generate language explanations, such as "Nurse Wang worked the night shift last week, so she will be given priority to work the day shift this week to ensure her physical and mental health", "Doctor Zhang's pediatric outpatient number satisfaction is higher than other departments, so the number of his outpatient visits will be appropriately increased", etc., so that managers can intuitively see the considerations and trade-offs behind the scheduling. Qualitative interviews show that 76% of managers believe that the algorithm explanations have greatly enhanced the persuasiveness and acceptability of the scheduling results.

[0168] The present invention aims at optimizing the configuration of logistics human resources in multiple districts of a smart hospital. Based on a thorough investigation of the actual challenges, the present invention constructs a knowledge graph of hospital scenarios to achieve the semantic integration of task requirements and personnel portraits; realizes the automatic analysis and standardization of unstructured scheduling requirements through multimodal language understanding; realizes the efficient generation and dynamic optimization of scheduling plans through knowledge-guided combinatorial optimization solution; and provides powerful assistance for management decision-making through language interpretation. The innovative achievements of the present invention have broad application prospects. They can not only provide new digital management methods for the smart hospital industry, improve the allocation efficiency and operation quality of medical resources, but also open up new practical paths for technology-driven administrative management reforms.

Claims

1. A method for allocating logistics human resources in multiple areas of a hospital, characterized by: The method comprises the following steps: Step 1: Build a hospital logistics knowledge graph ; Step 2: Multimodal semantic analysis of scheduling requirements; Step 3: Generate an initial task allocation plan; is a set of tasks to be assigned; Step 3.1: Assign each task Requirements description and every employee Attribute portrait Enter together , learn the matching distribution of tasks and employees: ； Step 3.2: According to the matching probability Sampling generates a set of initial allocations ,in Indicates that the task in the initial plan Is it assigned to employees? ; Step 4: Evaluate the quality of the initial solution; allocate the initial solution Each solution in enter , and design language guide Evaluate the quality of the solution from multiple dimensions to get a comprehensive score: ,in, It represents the normalized score of the solution quality, taking into account the following multiple objectives: task matching, employee satisfaction, and scheduling fairness; Step 5: Iterate and optimize the allocation plan; Step 5.1: Current Allocation Scheme Each solution in ,use The encoder extracts its semantic features : ; Step 5.2: Compare Perform knowledge-guided perturbations to generate new candidate solution representations : ;in, pass guide Generates,which contains the scheduling knowledge learned from historical optimization cases; Step 5.3: Exploitation The decoder will perturb the Restore to new readable allocation : ;at the same time, It also accepts constraints described in the form of language , the feasibility requirement is automatically met during the decoding process: ; Step 5.4: Evaluate the new allocation Quality If it is better than the current solution, add it to the new solution set ; Step 5.5: Repeat steps 5.1-5.4 until the iteration termination condition is met and the final optimal scheduling solution is obtained. , Step 6: Dynamically update the shift schedule; Step 6.1: Continuously monitor employee feedback and emergency requests in real-time ,use Perform semantic analysis and extract key information: ; Step 6.2: Update the information As a new constraint, insert it into the iterative optimization process of step 5 and dynamically adjust the optimal solution: , ,in Indicates at a point in time The optimal scheduling solution, is the scheme quality scoring function to evaluate the scheduling scheme The overall quality score of Indicates the search for The biggest plan , Indicates that the constraints that follow are satisfied; Step 6.3: Repeat steps 6.1-6.2 periodically to achieve dynamic update and continuous optimization of the scheduling plan, automatically generate language instructions for adjustment decisions, and enable intelligent management decisions.

2. The method for allocating logistics human resources in multiple areas of a hospital according to claim 1 is characterized in that: The step 1 specifically includes: Step 1.1: Define the ontology model layer and standardize the concept categories, attributes and relationships; the defined concepts include hospital, region, department, position, employee, shift and skill, which are formally expressed as: ； in, Indicates concept categories; Indicates conceptual relationship; Step 1.2: Extract instance-level information from heterogeneous data sources such as hospital management text and human resource systems; use pre-trained language models Learn semantic representations of medical professional words and concept phrases and perform named entity recognition The technology extracts conceptual entities from text, which are formally expressed as: ;in, is the set of conceptual entities extracted; is a set of relationships between entities; Step 1.3: Match the extracted instance-level information with the ontology schema and obtain a standardized knowledge graph after disambiguation ,in and Represent the concept entities and relations after disambiguation, respectively.

3. The method for allocating logistics human resources in multiple areas of a hospital according to claim 1 is characterized in that: The step 2 specifically includes: Step 2.1: Exploitation The scheduling needs of heterogeneous forms such as language, voice and image Mapping to a unified semantic space and learning task-independent contextual representations: ； Step 2.2: In the unified semantic representation Based on the above, through a series of fine-grained language understanding such as lexical analysis, dependency syntax analysis, and semantic role labeling, the structured scheduling demand elements are extracted and formally expressed as follows: ； in, Contains multi-dimensional information on task type, time requirements, and job capabilities; Step 2.3: Structure your scheduling requirements Hospital Knowledge Graph Match, complete and expand the requirement information to obtain a standardized task request description .

4. One hospital multi-district logistics human resources allocation system, characterized by: The system includes the following modules: Knowledge management module; Scheduling demand understanding module; Task optimization scheduling module; Online learning update module; Human-machine collaborative operation module; The above modules cooperate to implement the method described in any one of claims 1-3.

Images