A hotel information management method based on a cloud platform

Abstract

The application discloses a hotel information management method based on a cloud platform and belongs to the technical field of electric digital data processing. The method comprises the following steps: constructing a cloud-edge-end three-layer collaborative network, collecting data on the end side, performing dynamic weight mixed scheduling and local inference on the edge side edge server, performing federal learning training and knowledge graph construction on the cloud side cloud platform, calculating the comprehensive priority according to the user emergency degree, network load and calculation complexity of the edge server, dividing the task into local real-time, local batch processing and collaborative processing queues, training a global user preference prediction model by using a federal learning framework, constructing a hotel operation knowledge graph, inferring potential service demand by using a graph neural network, fusing the model prediction and graph inference results to generate an individualized strategy and executing the strategy, and realizing closed-loop feedback optimization through real-time monitoring and deviation hierarchical processing. The application realizes resource elastic scheduling, cross-domain knowledge sharing under privacy protection and self-adaptive optimization.

Description

Technical Field

[0001] This invention relates to the field of electronic digital data processing technology, and in particular to a hotel information management method based on a cloud platform. Background Technology

[0002] With the conglomeration of the hotel industry, achieving intelligent and unified management of multiple hotels has become a major challenge. The integration of cloud platform technology, IoT technology, and AI technology offers new solutions for hotel information management. However, existing technologies still face the following technical challenges when handling complex data from multiple hotels, devices, and users: In the prior art, CN112862294A discloses a cloud platform management system for a hotel group, which adopts a two-level architecture of cloud platform and hotel management server. The edge nodes only perform data forwarding and lack local real-time processing capabilities, resulting in slow response to high-latency sensitive tasks and poor user experience. CN112947166A discloses a smart management system for hotel rooms, which realizes unified management of smart devices, but task scheduling uses fixed priorities and cannot be dynamically adjusted according to network load and computing resources. This results in critical tasks not being processed first when the network is congested or under high load, leading to uneven resource utilization. CN113742048A discloses a hotel cloud service system, which introduces the collaboration of edge service network and cloud processing network, but adopts a centralized... Chinese-style data training requires each hotel's original data to be uploaded to the cloud, posing a risk of user privacy leakage and hindering secure data sharing across hotels, thus limiting the model's generalization ability. CN114782221A discloses a hotel room service management system that uses a device parameter prediction model for regulation, but it only predicts based on users' historical consumption information and behavioral habits, failing to build deep connections between entities and lacking sufficient reasoning ability for users' potential service needs. CN117151938A discloses a hotel digital operation management method that optimizes energy consumption through power status and environmental monitoring, but it lacks an execution feedback loop, and the system cannot adaptively evolve based on user manual adjustments and device execution deviations, resulting in a decline in strategy accuracy after long-term operation.

[0003] Therefore, there is an urgent need for a hotel information management solution that can take into account real-time response, elastic resource scheduling, privacy protection, deep reasoning, and adaptive optimization. Summary of the Invention

[0004] To address the technical problems existing in the prior art, this invention provides a hotel information management method based on a cloud platform, comprising the following steps: S101. Construct a three-layer collaborative network architecture of "cloud-edge-device". The device side includes smart device clusters and room access units deployed in guest rooms, the edge side includes edge servers deployed in each independent hotel, and the cloud side includes the group management cloud platform. S102. Dynamic weighted hybrid scheduling: The edge server receives data tasks from multiple room access units, extracts the feature parameters of each task, calculates the comprehensive scheduling priority based on the dynamically learned weight coefficients, and divides the tasks into local real-time processing queues, local batch processing queues, or collaborative processing queues according to the priority. S103. Global User Preference Prediction Model Training: The group management cloud platform receives de-identified data reported by each edge server, trains a global user preference prediction model using a federated learning framework, and distributes the model parameters to each edge server. S104. Construct and reason about a hotel operation knowledge graph. The group management cloud platform constructs a hotel operation knowledge graph. The knowledge graph includes users, guest rooms, equipment, service items and energy consumption data entities and their interrelationships. It then uses a graph neural network to reason and output the user's potential service needs and abnormal energy consumption point identification results. S105. Generate and execute personalized service strategies. When a user checks in, the edge server generates a preliminary prediction value based on the locally cached user preference prediction model, and combines it with the potential service needs to generate a personalized service strategy through a fusion mechanism, which is then sent to the corresponding room access unit for execution. S106. Closed-loop feedback optimization: The edge server collects feedback data in real time during the strategy execution process, calculates the deviation between the actual execution result and the expected strategy, and performs hierarchical processing according to the magnitude of the deviation value to update the local model or trigger the global model retraining of the cloud platform.

[0005] As a preferred embodiment of this application, in step S102, the feature parameters include at least the user urgency identifier, the current network load, and the computational complexity. The weight coefficients of the dynamic learning are adjusted periodically by an online optimizer with the goal of minimizing the average response time of high-priority tasks.

[0006] As a preferred embodiment of this application, in step S103, the federated learning framework includes: each edge server uses local data to train the received initial model and uploads encrypted gradients; the group management cloud platform uses an attention-based aggregation algorithm to perform weighted aggregation of gradients to generate a new global model.

[0007] As a preferred embodiment of this application, the encryption gradient is encrypted using a homomorphic encryption algorithm, and the attention mechanism dynamically allocates aggregation weights based on the similarity between the gradients of each edge server and historical gradients.

[0008] As a preferred embodiment of this application, in step S104, the construction of the knowledge graph includes defining entity types and relation types, extracting entity instances and relation instances from the business database and storing them in the graph database; the graph neural network is a relation graph convolutional network, which performs similarity calculation and anomaly detection by learning entity embedding vectors.

[0009] As a preferred embodiment of this application, in step S105, the fusion mechanism includes a weighted voting rule, which assigns higher weight to the model prediction value for observable preferences and higher weight to the knowledge graph reasoning result for potential needs; when the two conflict, a preset conflict resolution rule is invoked to make a decision.

[0010] As a preferred embodiment of this application, in step S105, the personalized service strategy includes control instructions for smart devices and pre-schedule instructions for service robots; the edge server synchronously queries the robot status, pre-selects target robots, and generates pre-schedule instructions when generating the strategy.

[0011] As a preferred embodiment of this application, in step S106, the hierarchical processing includes: if the deviation value belongs to the first range, then update the local user preference model; if the deviation value belongs to the second range, then start local policy re-optimization; if the deviation value belongs to the third range, then report the deviation data to the group management cloud platform as a negative sample to be added to the training set.

[0012] As a preferred embodiment of this application, in step S104, the construction of the knowledge graph further includes privacy protection processing of entities and relationships in the graph; Specifically, k-anonymization technology is used to generalize or suppress the entities and relationships, so that the association information of any user in the knowledge graph cannot be distinguished from the association information of the other k-1 users, thereby preventing the inference of user identity through association information.

[0013] As a preferred embodiment of this application, the conflict resolution rules in step S105 include at least one of the following principles: User real-time operation priority principle: When a user makes a manual adjustment, the system immediately terminates the current strategy and updates the user preference model based on the parameters adjusted by the user. Safety and compliance bottom line principle: When a generated personalized policy violates mandatory fire safety or electrical safety regulations, the system will automatically disable the policy; Contextual correction principle: When there is no real-time user operation, the long-term preference prediction value of the user is dynamically corrected based on objective contextual factors such as the current time, season or weather.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By constructing a three-layer collaborative network architecture of "cloud-edge-device", edge servers are deployed locally in the hotel to undertake local real-time processing tasks, reducing the response time of high-latency sensitive tasks such as door opening and dimming to the millisecond level. This solves the network latency problem caused by traditional centralized cloud processing and greatly improves the real-time performance and smoothness of user interaction.

[0015] 2. By introducing a dynamic weighted hybrid scheduling mechanism, the edge server extracts the user urgency, network load, and computational complexity characteristics of tasks in real time, calculates a comprehensive priority based on dynamically learned weight coefficients, and distributes tasks to local real-time queues, local batch processing queues, or collaborative processing queues. This mechanism enables elastic scheduling of computing and network resources, ensuring that critical tasks are processed first during network congestion or high load, effectively balancing the load pressure on the edge and cloud sides, and improving the utilization rate of system resources.

[0016] 3. By combining a federated learning framework with homomorphic encryption, each edge server only uploads encrypted gradient information to the group's management cloud platform, which then aggregates the models without decryption. This solution achieves global knowledge sharing and collaborative model training across hotels without accessing the original user data from each hotel. This not only improves the generalization ability of the user preference prediction model but also fundamentally eliminates the risk of leakage of user privacy data during transmission and aggregation, thus meeting data compliance requirements.

[0017] 4. By constructing a hotel operation knowledge graph and combining it with graph neural network reasoning, discrete user, room, equipment, and service data are transformed into a network structure with deep semantic relationships. The system can mine implicit associations between users and services based on entity embedding vectors, and proactively infer users' potential service needs (such as the association between parent-child users and children's products), breaking through the limitations of traditional recommendations based on explicit historical behavior and significantly improving the accuracy and intelligence level of service recommendations.

[0018] 5. A closed-loop feedback optimization mechanism based on deviation grading was established. The edge server collects policy execution feedback in real time and triggers local model updates, local policy re-optimization, or cloud-based global model retraining based on the magnitude of the deviation (mild, moderate, severe). This mechanism enables the system to continuously adapt to dynamic changes in user preferences and environmental evolution, forming an adaptive intelligent agent capable of self-evolution and improvement, ensuring the effectiveness of the policy in long-term operation. Attached Figure Description

[0019] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a general flowchart of the hotel information management method provided in the embodiments of the present invention.

[0022] Figure 2 This is a schematic diagram of the three-layer collaborative network architecture provided in an embodiment of the present invention.

[0023] Figure 3 This is a schematic diagram of the dynamic weighted hybrid scheduling process provided in an embodiment of the present invention.

[0024] Figure 4 This is a schematic diagram of the training of the global user preference model provided in an embodiment of the present invention.

[0025] Figure 5 This is a schematic diagram illustrating the construction and reasoning of a hotel operation knowledge graph provided in an embodiment of the present invention.

[0026] Figure 6 This is a flowchart of the generation and execution of personalized service strategies provided in the embodiments of the present invention.

[0027] Figure 7 This is a schematic diagram of the closed-loop feedback optimization mechanism provided in an embodiment of the present invention.

[0028] Figure 8 This is an architecture diagram of the hotel information management system provided in an embodiment of the present invention. Detailed Implementation

[0029] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0030] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present invention are only used to explain the relative positional relationship and movement of each component in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indication will also change accordingly.

[0031] Furthermore, the use of terms such as "first" and "second" in this invention is for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, features defined with "first" and "second" may explicitly or implicitly include at least one of those features. Additionally, the technical solutions of the various embodiments can be combined with each other, but only on the basis of being achievable by those skilled in the art. When the combination of technical solutions is contradictory or impossible to implement, such a combination of technical solutions should be considered non-existent and not within the scope of protection claimed by this invention.

[0032] Example 1 This invention provides a cloud-based method for hotel information management. For example... Figure 1 As shown, the specific steps include: S101. Construct a three-layer collaborative network architecture of "cloud-edge-device". This step aims to construct a distributed network architecture that combines physical and logical elements, serving as the foundation for the entire management system. This network architecture consists of three layers: the endpoint, the edge, and the cloud. Figure 2 As shown, the details are as follows: The edge deployment is located in each guest room and consists of a cluster of smart devices and a room access unit. The smart device cluster includes service devices (such as smart air conditioners, smart lighting control systems, smart curtains, smart TVs, smart door locks, etc.) and sensing sensors (such as temperature sensors, humidity sensors, light sensors, infrared human body sensors, smoke detectors, etc.). All devices connect to the room access unit via a short-range wireless communication protocol. The room access unit is an embedded gateway with protocol conversion, data aggregation, and edge computing capabilities. On one hand, the room access unit is responsible for polling or monitoring all connected devices, collecting device status, environmental parameters, and user operation data in real time. On the other hand, the room access unit parses and executes control commands from the upper layer, converting them into signals that the devices can recognize.

[0033] The edge server consists of at least one deployed within each independent hotel. Connected to all room access units via the hotel's local area network, the edge server's hardware configuration must support localized real-time processing, typically equipped with computing and storage resources superior to ordinary commercial servers. The core functions of the edge server include: 1) Data preprocessing: cleaning, filtering, and standardizing the aggregated raw data to unify heterogeneous data and reduce cloud-side pressure; 2) Real-time response: performing millisecond-level calculations and decisions locally for high-latency sensitive tasks such as door opening and dimming, without cloud-side intervention; 3) Lightweight inference: loading a copy of the user preference model to quickly complete the initial calculation of personalized strategies upon check-in.

[0034] The cloud-side is a group management cloud platform deployed at the hotel group's headquarters, consisting of a server cluster, a distributed database, and a big data computing framework. The group management cloud platform connects to all edge servers via an encrypted wide area network and undertakes global tasks: 1) Data aggregation and storage, integrating data from various hotels to form a global data view; 2) Centralized model training and service, using aggregated data to train complex deep learning models and providing algorithmic support to the edge servers; 3) Knowledge graph construction and maintenance, extracting knowledge from business data to build and continuously update a knowledge graph covering entities and their relationships, including users, rooms, equipment, services, and energy consumption. When building and using the knowledge graph, k-anonymization is used to protect entities and relationships within the graph, preventing the inference of user identities through association information.

[0035] Through the above deployment, a three-layer collaborative network architecture is constructed. The edge side is responsible for data acquisition and command execution, the cloud side is responsible for data preprocessing and real-time response, and the cloud side is responsible for global model training and deep knowledge mining. The three layers cooperate with each other to support subsequent intelligent decision-making.

[0036] S102, Dynamic Weighted Hybrid Scheduling Based on the network architecture built on S101, edge servers process massive concurrent data from multiple room access units. This step achieves elastic scheduling of computing and network resources by performing feature extraction, priority calculation, and queue partitioning of tasks, such as... Figure 3 As shown, the specific implementation is as follows: When a task arrives at the edge server, the scheduling module creates a task descriptor for it and extracts three core feature parameters: 1) User urgency indicator, assigned a value based on the interaction status of the task source. Real-time user-triggered operations (such as temperature adjustment via a mobile app) are assigned a value of 1, indicating extremely high response time requirements; periodically reported data (such as sensor readings) are assigned a value of 0, indicating that the task can tolerate a certain delay. 2) Current network load, the scheduling module monitors the bandwidth utilization of the WAN link between the edge server and the group management cloud platform in real time and normalizes it to the [0,1] range. The closer the value is to 1, the more congested the network is, and the greater the cost of transmitting data to the group management cloud platform. 3) Computational complexity, a preset coefficient based on the task type. For example, image processing tasks, which require more computational resources, are set to a higher value, while sensor reporting tasks, which are very lightweight, are set to a lower value.

[0037] Subsequently, the scheduling module calculates the overall scheduling priority of the task based on the dynamic priority calculation mechanism: Priority = α × U + β / N + γ × C, where U is the user's urgency level, N is the current network load, C is the computational complexity, and α, β, and γ are dynamic weight coefficients (α + β + γ = 1). The values ​​of the dynamic weight coefficients are not fixed but are dynamically adjusted by an online optimizer. This optimizer aims to minimize the average response time of high-priority tasks, periodically analyzes the actual response time, completion status, and network conditions of historical tasks, and updates the weight coefficients using a Bayesian optimization algorithm.

[0038] Based on the calculated overall scheduling priority, the scheduling module divides tasks into three different processing queues according to the overall scheduling priority value. If the overall scheduling priority of a task is greater than a preset first threshold, the scheduling module determines that the task is highly sensitive to time latency and is very important, and allocates it to a queue specifically for local real-time processing. The edge server's central processing unit immediately calls upon the task in this queue for immediate computation and response. If the overall scheduling priority of a task is less than a preset second threshold, the scheduling module determines that the task is a non-urgent, routine task and allocates it to a local batch processing queue, waiting for the system to idle before batch processing, avoiding the monopolization of valuable computing resources during peak periods. If the overall scheduling priority of a task is between the first and second thresholds, the scheduling module determines that the task is a collaborative processing task and allocates it to a collaborative processing queue. The edge server dynamically distributes the workload based on the current central processing unit utilization. If the utilization rate is below 80%, it indicates that the edge server has idle computing power, and the task is processed by the edge server. If the utilization rate is above 80%, it indicates that the edge server is under high load, and the task is forwarded to the group management cloud platform for processing.

[0039] Through dynamic weighted hybrid scheduling, edge servers transform the originally chaotic torrent of data into an ordered task queue with different processing priorities and paths. This scheduling mechanism ensures the response speed of real-time tasks, optimizes local resource utilization, and achieves cloud-edge load balancing.

[0040] S103, Training of Global User Preference Prediction Model After receiving data reported from various edge servers, the group management cloud platform, while protecting data privacy, utilizes global data and employs a federated learning framework to train a global user preference prediction model. Figure 4 As shown, the specific implementation is as follows: First, each edge server performs data anonymization and feature alignment before reporting data: direct identifiers (name, ID number, etc.) are salted and hashed to convert them into irreversible hash values, preventing the group's management cloud platform from deducing the user's true identity from the hash value; precise location information is obfuscated (e.g., retained down to the city level). Simultaneously, each edge server extracts features from the data, converting the raw data into a uniformly structured feature vector containing dimensions such as timestamps, environmental preferences, and device interactions.

[0041] After data anonymization and feature alignment are completed, the federated learning training module initializes a deep neural network model as the base model. The deep neural network model adopts a multilayer perceptron structure, with the number of nodes in the input layer matching the dimension of the feature vectors, and the number of nodes in the output layer corresponding to the probability values ​​of various personalized service recommendations. After initialization, the group's management cloud platform distributes the initial model parameters to each edge server via an encrypted channel.

[0042] After receiving the initial model parameters, the edge server trains the model using local, un-anonymized data and employs mini-batch stochastic gradient descent. The edge server randomly selects a small batch of samples from its local data and inputs them into the model for forward propagation to obtain predicted values. These predicted values ​​are then compared with the true labels of the samples to calculate the prediction error. Finally, the gradient of each model weight is calculated using the backpropagation algorithm. After each iteration, the edge server does not directly upload the updated model weights; instead, it only uploads the calculated gradient information. Before uploading, the edge server encrypts these gradients using a homomorphic encryption algorithm. Homomorphic encryption allows mathematical operations to be performed directly on the ciphertext without decryption. The group's management cloud platform can aggregate the gradients without decryption, thus fundamentally eliminating the risk of user privacy leakage during gradient information transmission and aggregation.

[0043] After collecting the encrypted gradients uploaded by all edge servers, the group's management cloud platform initiates an aggregation algorithm based on an attention mechanism. The algorithm first calculates the similarity between the gradient contributed by each edge server and the gradients contributed by other edge servers historically. If the gradient of an edge server has a high similarity to the gradients of most servers, it indicates that the hotel's local data distribution is general and representative of most hotels; in this case, the algorithm assigns a higher aggregation weight to that gradient. Conversely, if the gradient of an edge server differs significantly from the gradients of other servers, it indicates that the hotel's data may have unique characteristics. For example, if the hotel is located in a tropical region, users' needs for air conditioning temperature may differ significantly from those in other regions; in this case, the algorithm assigns a lower aggregation weight to that gradient to prevent special data from causing excessive interference to the global model. Through this weighted averaging method, the group's management cloud platform aggregates and generates new global model parameters.

[0044] After aggregation, the group's management cloud platform distributes the new global model back to each edge server, triggering the next round of local training. This process repeats continuously, with each iteration improving the performance of the global model. The training process ends when the loss function value of the global model converges below a preset threshold, or when the preset number of iterations is reached. The final global model is a user preference prediction model that integrates all hotel data knowledge. It possesses good generalization ability, adapting to different regions and hotel types, while retaining certain personalized features to provide accurate predictions for subsequent personalized services.

[0045] Federated learning enables knowledge sharing and model collaboration across hotels, while strictly adhering to legal and regulatory requirements for data privacy protection, laying a solid algorithmic foundation for the intelligent management of hotel groups.

[0046] S104. Construct and reason about a hotel operations knowledge graph. While running federated learning to train the global model, the group's management cloud platform simultaneously constructs a hotel operation knowledge graph for deep relational reasoning.

[0047] First, define the knowledge graph ontology, clarifying the entity types and their relationship types. The entity types in the knowledge graph include at least: User entities, with attributes such as anonymized user identifiers, user's permanent residence information, and user check-in frequency level; Guest room entities, with attributes such as room type, floor, room orientation, room area, and a list of installed equipment; Equipment entities, with attributes such as equipment type, manufacturer, energy consumption level, and historical failure rate; Service item entities, with attributes such as service name, service price, average service time, and required personnel; and Energy consumption data entities, with attributes such as data record timestamps, electricity consumption, and water consumption. The types of relationships between entities include at least the following: check-in relationship, connecting the user entity and the room entity, indicating that the user has checked into the room; control relationship, connecting the user entity and the device entity, indicating that the user has controlled the device; containment relationship, connecting the room entity and the device entity, indicating that the room contains the device; generation relationship, connecting the device entity and the energy consumption data entity, indicating that the device generated these energy consumption records within a certain time period; and association relationship, connecting the service item entity and the room entity, indicating that the room can provide this service to guests.

[0048] Secondly, knowledge extraction and storage are performed. The group's management cloud platform extracts entity instances and relation instances that conform to the ontology definition from various business systems. For example, it extracts user check-in registration information from the hotel front desk management system database to obtain user entities, room entities, and check-in relationships. It extracts user operation records for equipment such as air conditioners and lights from equipment control log files to obtain equipment entities and the control relationships between users and equipment. It extracts real-time energy consumption data for each device from the energy consumption monitoring system to obtain energy consumption data entities and the generation relationships between devices and energy consumption. The extracted knowledge is organized into triples (head entity, relation, tail entity) and stored in a graph database.

[0049] Next, representation learning is performed using a relational graph neural network. The relational graph convolutional network learns high-order feature representations for each entity by performing convolution operations on the knowledge graph. Specifically, for each entity node in the knowledge graph, the relational graph convolutional network aggregates the feature information of all its neighboring nodes, and through propagation across multiple layers of the network, generates an embedding vector at each node that reflects its structural and attribute information within the entire graph.

[0050] Based on the learned entity embedding vectors, the group's management cloud platform performs potential service demand inference and abnormal energy consumption point identification. For potential service demand inference, when the system needs to provide service suggestions to currently occupied users, the inference engine calculates the embedding vectors of the user entity and their occupied room entity, and performs similarity calculations with the embedding vectors of all service item entities. If the similarity between the service item entity and the combined vectors of the user and room exceeds a preset threshold, the system determines that the user may need this service. For example, if many entities in the knowledge graph labeled as parent-child users have high similarity to the service entity "children's toiletries," when the system identifies the current user as a parent-child user, it can infer that the current user may need children's toiletries and proactively inquire about or prepare them upon check-in. Abnormal energy consumption point identification: The system constructs an energy consumption prediction mechanism for each guest room entity and its associated equipment entities and energy consumption data entities. Based on historical data, it predicts the energy consumption value that the guest room should generate under the current environmental conditions. The predicted value is compared with the actual energy consumption value reported by the sensor. When the deviation exceeds the preset threshold, the system determines that the guest room or the equipment is an abnormal energy consumption point and automatically triggers an alarm to notify the engineering personnel to investigate.

[0051] By constructing and reasoning with knowledge graphs, originally discrete business data is transformed into a dynamic knowledge network that can be reasoned about, providing semantic support for personalized service decisions.

[0052] S105. Generate and execute personalized service policies. When a user checks in, the process of generating and executing personalized service policies is triggered. This process is primarily completed on edge servers to ensure low latency and high efficiency throughout the entire process.

[0053] When a user checks in, the front desk staff enters the user's identity information into the hotel management system. This identity information is synchronized in real time to the hotel's edge server via the hotel's internal network. The edge server queries its local cache using the user's identifier. If a copy of the user's preference model exists, it is loaded directly; otherwise, it requests the generation of a personalized model branch from the group's management cloud platform and distributes it to the edge server. Simultaneously, the edge server initiates a query request to the group's management cloud platform's knowledge graph reasoning module for the current user and the upcoming guest room, inferring potential service needs. Upon receiving the request, the knowledge graph reasoning module performs knowledge graph reasoning, calculating the similarity between the user and guest room entities and each service item entity. It then returns a list of services with similarity exceeding a threshold as the query result to the edge server.

[0054] The edge server's policy generation engine merges predictions from the user preference model with inference results from the knowledge graph to generate a final, executable personalized service policy. The fusion process employs a weighted voting mechanism: historically observable preferences (such as temperature) are given higher weight by the model's predictions, while potential needs (such as pillow type) are given higher weight by the knowledge graph results. When the suggestions from the model's predictions and the knowledge graph's inference conflict—for example, the model predicts that the user prefers a bright environment, but the knowledge graph infers that the current room's orientation might result in glaring afternoon sunlight requiring the curtains to be drawn—the policy generation engine invokes pre-defined conflict resolution rules. The conflict resolution rules are a hierarchical rule base. Its core principles include: 1) Real-time user operations have the highest priority. Any preset strategy should immediately give way after the user makes a manual adjustment, and the user model should be updated based on the new operation; 2) Safety and compliance are the bottom line. No strategy may violate mandatory regulations such as fire protection and electrical safety; 3) Context priority. When there is no real-time operation, long-term preferences are corrected based on objective context factors such as the current time, season, and weather; 4) Default strategy. When the above rules cannot resolve the conflict, the system's preset, most conservative, and comfortable default value is adopted.

[0055] The generated strategy is converted into a structured instruction set and sent to the room access unit of the corresponding guest room. The room access unit parses the instructions and drives various smart devices (such as air conditioners, lights, and curtains) to reach preset states, completing environmental preparation before the user checks in. If the strategy includes physical services (such as item delivery), the edge server will pre-select and schedule suitable service robots to be in place in advance to shorten the user's waiting time.

[0056] S106, Closed-loop feedback optimization The implementation of personalized service strategies is not the end, but a new starting point for continuous system learning and optimization. To ensure that the system can continuously adapt to changes in user needs and evolution of environmental conditions, a closed-loop feedback optimization mechanism is continuously run during and after strategy implementation.

[0057] The edge server's real-time monitoring and feedback acquisition module continuously collects various execution feedback data. This data includes: the deviation between the actual device status and the expected command value, all manual adjustments made by the user, and the actual completion status and evaluation of the robot's service tasks.

[0058] After collecting feedback data, the deviation analysis unit compares the actual feedback values ​​with the expected strategies generated in S105 to calculate the deviation values ​​of each indicator. Based on the magnitude and nature of the deviation values, the system performs tiered processing. If the deviation is minor, such as a user fine-tuning the temperature, the local user model is directly updated with this new data. If the deviation is moderate, such as a robot response delay, local strategy re-optimization (e.g., rescheduling) is initiated, for example, replacing the robot with a closer one or optimizing the robot's route. The adjusted strategy is then reissued for execution, and the execution effect continues to be monitored. If the deviation is severe, such as a complete model prediction failure or a hardware malfunction, a complete deviation report is submitted to the group's management cloud platform.

[0059] The group's management cloud platform aggregates severe deviation reports from various edge servers, adds them to the training set as negative samples, analyzes common causes, and, based on the analysis results, initiates a new round of federated learning training. After updating the global model, it is distributed to all edge servers, thereby iteratively improving the intelligence level of the entire system.

[0060] By establishing a complete closed-loop mechanism encompassing strategy generation, execution, deviation monitoring, and feedback optimization, a hotel information management system capable of continuous learning and adaptive evolution has been constructed. The system is no longer a static program, but an adaptive intelligent agent that continuously improves itself based on data and feedback, providing efficient and precise management services to hotel groups in a stable and long-term manner.

[0061] Example 2 This invention also provides a cloud-based hotel information management system, employing a three-tier cloud-edge-device architecture. It consists of a group management cloud platform, multiple edge server clusters, and a massive number of room access units. These three layers are connected via network communication, forming an organically collaborative whole. Specifically, it includes: Group Management Cloud Platform The group management cloud platform is the highest decision-making level of the entire system. It consists of multiple functional modules, each of which undertakes specific responsibilities and together supports the complex functions of the cloud platform.

[0062] The Federated Learning Training Module is the core component for achieving cross-hotel collaborative intelligence. Internally, it includes a model initialization unit, an encrypted gradient aggregation unit, and a model distribution and version management unit. The model initialization unit is responsible for building the initial architecture of the deep neural network at the start of training, setting hyperparameters such as the number of layers, the number of nodes per layer, and the type of activation function, and generating initial weight parameters. The encrypted gradient aggregation unit is the key executor of federated learning; it receives homomorphically encrypted model gradients uploaded from each edge server. This unit incorporates an attention-based aggregation algorithm that dynamically assigns aggregation weights to gradients from different edge servers and performs weighted averaging directly without decrypting the encrypted data, generating new global model parameters. The model distribution and version management unit is responsible for distributing the aggregated new global model to all participating edge servers through an encrypted communication channel. It also manages the model's historical versions, recording the time and performance metrics of each update, and supports rolling back the model to a previous stable version when necessary.

[0063] The knowledge graph construction and reasoning module is the foundation for achieving deep semantic understanding and potential demand mining. Internally, this module comprises a knowledge extraction unit, a graph database storage unit, and a relational graph convolutional network reasoning unit. The knowledge extraction unit connects to various business databases within the hotel group, including the front desk management system database, customer relationship management system database, and equipment log database. Through pre-defined extraction rules and natural language processing technology, it extracts entities, attributes, and their relationships from these heterogeneous data sources, forming structured knowledge triples. The graph database storage unit uses a dedicated graph database system to store these triples, constructing a knowledge graph containing entities such as users, rooms, equipment, services, and energy consumption, along with their complex relationships. The relational graph convolutional network reasoning unit periodically reads the graph data, runs relational graph convolutional network algorithms, calculates the embedding vector for each entity node, and performs link prediction and anomaly detection tasks based on these embedding vectors, outputting the identification results of potential user service needs and abnormal energy consumption points.

[0064] The Global Strategy Management module is responsible for formulating and issuing management strategies that do not involve real-time responses but have global guiding significance. Internally, the Global Strategy Management module includes an Energy Consumption Optimization Strategy unit and an Operations and Maintenance Scheduling Strategy unit. The Energy Consumption Optimization Strategy unit receives abnormal energy consumption point identification results from the knowledge graph inference unit and, combined with the global energy consumption model, generates energy consumption optimization strategies for different regions and time periods. For example, it can generate dynamic dimming strategies for public area lighting, automatically adjusting the brightness of corridor and lobby lights based on outdoor light intensity and passenger flow density; it can also generate global temperature setting thresholds for air conditioning, specifying the minimum set temperature for all guest room air conditioners when unoccupied, to save energy. The Operations and Maintenance Scheduling Strategy unit analyzes equipment failure rates and maintenance work order data reported by various edge servers, generates preventative maintenance plans, optimizes cross-hotel scheduling schemes for maintenance personnel, and ensures that critical equipment is always in good operating condition.

[0065] Edge server cluster The edge servers deployed within each individual hotel serve as the bridge and nerve center connecting the cloud platform and guest room equipment. Each edge server logically consists of multiple functional modules that work together to perform local data processing and real-time decision-making tasks.

[0066] The dynamic weighted hybrid scheduling module is a key component of the edge server in handling sudden data surges. Internally, it comprises a task feature parsing unit, a priority calculation engine, and a queue distributor. The task feature parsing unit intercepts all data streams reported from room access units, parses each data packet, and attaches task feature tags, including user urgency, current network load, and computational complexity. The priority calculation engine has a built-in priority calculation formula and stores real-time updated weight coefficients, calculating a comprehensive scheduling priority for each pending task in real time. Based on the priority calculation results, the queue distributor distributes tasks to three different circular buffer queues: a local real-time queue, a local batch processing queue, and a collaborative processing queue. Simultaneously, the queue distributor continuously monitors the utilization of the local CPU, dynamically adjusting the flow of data in the collaborative processing queue, and automatically forwarding some tasks to the cloud platform for processing when local resources are strained.

[0067] The local model caching and inference module is a key component for achieving low-latency response in personalized services. Internally, it comprises a model storage unit, a lightweight inference engine, and a policy fusion unit. The model storage unit uses a high-performance in-memory database to cache copies of user preference models downloaded from the cloud platform. To store as many models as possible within limited storage space, the storage unit employs a least recently used cache eviction strategy, prioritizing models for frequently checked users and automatically evictioning the least recently used models when the cache is full. The lightweight inference engine is a deep learning inference engine optimized for mobile or edge devices. It can quickly load the corresponding model from the model storage unit, perform forward inference calculations using user and room identifiers as input, and output preliminary personalized parameter predictions. The policy fusion unit receives the preliminary prediction results from the local inference engine and simultaneously queries the cloud platform's knowledge graph module via an application programming interface to obtain potential demand inference results. This unit uses a weighted voting algorithm to fuse the two types of results, generating the final executable policy instruction set.

[0068] The real-time monitoring and feedback acquisition module is the sensing component for achieving closed-loop optimization in the system. Internally, it includes a device status tracking unit, a user behavior capture unit, and a deviation analysis unit. The device status tracking unit subscribes to and records the execution status of commands sent to each room's access unit in real time, including key indicators such as whether the command was successfully delivered, whether the device responded successfully, and the response time. For commands that fail to execute, it triggers an automatic retry mechanism to ensure service reliability. The user behavior capture unit captures any manual interventions performed by users during their stay in real time through the room access units, such as manually adjusting the air conditioning temperature, manually turning lights on and off, and manually changing TV channels. These actions are considered implicit feedback to the preset strategy, recorded, and timestamped, serving as training samples for subsequent model updates. The deviation analysis unit compares the captured actual execution status and user manual operations with the expected strategy generated by the strategy fusion unit, calculates the deviation values ​​of various indicators, and, based on preset deviation grading rules, decides whether to trigger a local model update and local strategy re-optimization, or report the deviation data to the cloud platform.

[0069] Room access unit The room access unit, deployed in each guest room, is the terminal nerve of the system's interaction with the physical world. Physically, the room access unit is an embedded gateway with communication and computing capabilities, and logically it consists of multiple functional modules.

[0070] The multi-protocol adaptation and aggregation module solves the problem of inconsistent communication protocols among various smart devices in guest rooms. Internally, this module includes a protocol stack unit and a data aggregation and standardization unit. The protocol stack unit integrates protocol stacks for various short-range wireless communication protocols, including ZigBee, Bluetooth Mesh, and Wi-Fi, enabling simultaneous connections with terminal devices using different protocols and ensuring successful gateway access for all smart devices. The data aggregation and standardization unit polls or listens to the ports of each protocol stack, collecting raw data reported by all terminal devices. This unit has a built-in device description file library that can identify the proprietary data formats of different manufacturers and types of devices, converting them into a standardized data format defined internally by the system, such as JSON using key-value pairs, before packaging and reporting to the edge server.

[0071] The instruction parsing and driver module is responsible for converting control instructions from the upper layer into specific electrical signals that the terminal device can recognize. Internally, this module includes an instruction buffer and parsing unit and a device driver unit. The instruction buffer and parsing unit receives instruction sets from the edge server, performs syntax parsing and parameter validation to ensure correct instruction format and parameters within a reasonable range. The parsed instructions are placed in a first-in-first-out (FIFO) ordered execution queue to ensure sequential execution and avoid conflicts. The device driver unit has a built-in driver for each type of terminal device. The driver converts the parsed instruction parameters into specific control codes conforming to the device's communication protocol and sends them out through the corresponding protocol stack port to drive the device to perform specific operations. For example, for smart lights, the driver might adjust brightness using pulse width modulation signals; for smart air conditioners, the driver might send an infrared code to set the temperature and mode.

[0072] The environmental perception and status reporting module enables continuous monitoring of the guest room's physical environment. Internally, this module includes a sensor data acquisition unit and a heartbeat and status maintenance unit. The sensor data acquisition unit periodically wakes up and reads data from various sensors connected to the gateway, including temperature, humidity, light, infrared human body sensors, and smoke sensors. The acquisition frequency can be dynamically adjusted based on configuration parameters issued by the edge server; for example, the acquisition frequency can be increased during peak daytime hours and decreased at night when there are few guests to save energy. The heartbeat and status maintenance unit periodically sends heartbeat packets to the edge server, reporting the online status of itself and connected terminal devices. When it detects that a device is offline or sensor values ​​exceed normal thresholds, it proactively triggers an alarm event and reports it to the edge server for timely intervention by engineers.

[0073] Through the close collaboration of these modules, this system achieves fully automated management of the entire process from data collection, intelligent decision-making, command execution to feedback optimization, providing hotel groups with an implementable and verifiable intelligent management technology solution.

[0074] Example 3 The present invention also provides an electronic device, including: a processor, a transmitting device, an input device, an output device, and a memory. The processor may be implemented using a general-purpose CPU (Central Processing Unit), a microprocessor, an application-specific integrated circuit, or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this application. The memory may be implemented using a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM), and is used to store computer program code. The computer program code includes computer instructions. When the processor executes the computer instructions, the electronic device executes a method as described in any of the above possible implementation methods.

[0075] Example 4 The present invention also provides a computer-readable storage medium storing a computer program, the computer program including program instructions, which, when executed by a processor of an electronic device, cause the processor to perform a method as described in any of the above possible implementations.

[0076] In the description of this specification, the references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0077] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features claimed herein.

Claims

1. A hotel information management method based on a cloud platform, characterized in that, Includes the following steps: S101. Construct a three-layer collaborative network architecture of "cloud-edge-device". The device side includes smart device clusters and room access units deployed in guest rooms, the edge side includes edge servers deployed in each independent hotel, and the cloud side includes the group management cloud platform. S102. Dynamic weighted hybrid scheduling: The edge server receives data tasks from multiple room access units, extracts the feature parameters of each task, calculates the comprehensive scheduling priority based on the dynamically learned weight coefficients, and divides the tasks into local real-time processing queues, local batch processing queues, or collaborative processing queues according to the priority. S103. Global User Preference Prediction Model Training: The group management cloud platform receives de-identified data reported by each edge server, trains a global user preference prediction model using a federated learning framework, and distributes the model parameters to each edge server. S104. Construct and reason about a hotel operation knowledge graph. The group management cloud platform constructs a hotel operation knowledge graph. The knowledge graph includes users, guest rooms, equipment, service items and energy consumption data entities and their interrelationships. It then uses a graph neural network to reason and output the user's potential service needs and abnormal energy consumption point identification results. S105. Generate and execute personalized service strategies. When a user checks in, the edge server generates a preliminary prediction value based on the locally cached user preference prediction model, and combines it with the potential service needs to generate a personalized service strategy through a fusion mechanism, which is then sent to the corresponding room access unit for execution. S106. Closed-loop feedback optimization: The edge server collects feedback data in real time during the strategy execution process, calculates the deviation between the actual execution result and the expected strategy, and performs hierarchical processing according to the magnitude of the deviation value to update the local model or trigger the global model retraining of the cloud platform.

2. The method according to claim 1, characterized in that, In step S102, the feature parameters include at least the user urgency identifier, the current network load, and the computational complexity. The weight coefficients of the dynamic learning are adjusted periodically by an online optimizer with the goal of minimizing the average response time of high-priority tasks.

3. The method according to claim 1, characterized in that, In step S103, the federated learning framework includes: each edge server uses local data to train the received initial model and uploads encrypted gradients; the group management cloud platform uses an attention-based aggregation algorithm to weighted aggregate the gradients and generate a new global model.

4. The method according to claim 3, characterized in that, The encryption gradient is encrypted using a homomorphic encryption algorithm, and the attention mechanism dynamically allocates aggregation weights based on the similarity between the gradients of each edge server and historical gradients.

5. The method according to claim 1, characterized in that, In step S104, the construction of the knowledge graph includes defining entity types and relationship types, extracting entity instances and relationship instances from the business database and storing them in the graph database; the graph neural network is a relational graph convolutional network, which performs similarity calculation and anomaly detection by learning entity embedding vectors.

6. The method according to claim 1, characterized in that, In step S105, the fusion mechanism includes a weighted voting rule, which assigns higher weights to the model prediction values ​​for observable preferences and higher weights to the knowledge graph reasoning results for potential needs. When the two conflict, the preset conflict resolution rules are invoked for adjudication.

7. The method according to claim 1, characterized in that, In step S105, the personalized service strategy includes control instructions for smart devices and pre-schedule instructions for service robots; the edge server synchronously queries the robot status, pre-selects target robots, and generates pre-schedule instructions when generating the strategy.

8. The method according to claim 1, characterized in that, In step S106, the hierarchical processing includes: if the deviation value belongs to the first range, then update the local user preference model; if the deviation value belongs to the second range, then start local policy re-optimization; if the deviation value belongs to the third range, then report the deviation data to the group management cloud platform as a negative sample to be added to the training set.

9. The method according to claim 5, characterized in that, In step S104, the construction of the knowledge graph also includes privacy protection processing of entities and relationships in the graph; Specifically, k-anonymization technology is used to generalize or suppress the entities and relationships, so that the association information of any user in the knowledge graph cannot be distinguished from the association information of the other k-1 users, thereby preventing the inference of user identity through association information.

10. The method according to claim 6, characterized in that, The conflict resolution rules in step S105 include at least one of the following principles: User real-time operation priority principle: When a user makes a manual adjustment, the system immediately terminates the current strategy and updates the user preference model based on the parameters adjusted by the user. Safety and compliance bottom line principle: When a generated personalized policy violates mandatory fire safety or electrical safety regulations, the system will automatically disable the policy; Contextual correction principle: When there is no real-time user operation, the long-term preference prediction value of the user is dynamically corrected based on objective contextual factors such as the current time, season or weather.

Images