Teaching equipment management method and device

Abstract

The invention provides a teaching equipment management method, which comprises the following steps: acquiring user identity verification information, performing identity authentication through a face recognition, card swiping or code scanning mode, and generating user identity confirmation data; based on the user identity confirmation data, calling a user historical borrowing record and learning demand data to carry out artificial intelligence (AI) algorithm analysis, and generating a personalized equipment recommendation list; according to the personalized equipment recommendation list, matching calculation is carried out in combination with equipment state information in the storage cabinet, and a borrowable equipment list is determined; based on the borrowable equipment list, executing door opening operation through an electromagnetic door control system to complete an equipment borrowing process; receiving an equipment return request, collecting an equipment image through a close-range camera, carrying out integrity detection in combination with an AI recognition technology, and generating an equipment state evaluation result; according to the equipment state evaluation result, blue light disinfection processing is executed, background equipment management data is updated, and the equipment return process is completed.

Description

Technical Field

[0001] This disclosure belongs to the field of signal processing technology, and more specifically, relates to a teaching equipment management method and device that integrates data acquisition, artificial intelligence analysis and interactive display. Background Technology

[0002] In the current context of educational informatization, various high-tech teaching equipment, such as development boards, intelligent vehicles, and virtual reality devices, are playing an increasingly important role in teaching activities. However, this also brings challenges to the efficient, safe, and convenient management of these devices.

[0003] Traditional teaching equipment management relies heavily on manual operations, such as manual registration, distribution, collection, and inspection by administrators. This model is generally cumbersome, inefficient, and prone to errors. Students or teachers borrowing equipment must go through multiple steps, including filling out forms and manual review, which consumes a significant amount of valuable time and disrupts the flow of teaching activities. Furthermore, manual recording is prone to omissions or errors, leading to discrepancies between equipment information and actual inventory, increasing the risk of equipment loss or damage. Summary of the Invention

[0004] The purpose of this disclosure is to provide a method for managing teaching equipment, which aims to solve the problems of cumbersome, inefficient, and error-prone manual management processes for teaching equipment.

[0005] To achieve the above objectives, this disclosure provides the following technical solution:

[0006] A method for managing teaching equipment includes the following steps:

[0007] Obtain user identity verification information, and perform identity verification through facial recognition, card swiping or QR code scanning to generate user identity confirmation data;

[0008] Based on the user identity verification data, the system retrieves the user's historical borrowing records and learning needs data for artificial intelligence (AI) algorithm analysis to generate a personalized device recommendation list.

[0009] Based on the personalized device recommendation list, a matching calculation is performed in conjunction with the device status information in the locker to determine the list of available devices.

[0010] Based on the list of available equipment, the equipment borrowing process is completed by opening the door through the electromagnetic door control system.

[0011] Upon receiving a device return request, the system captures images of the device using a close-up camera and combines AI recognition technology to perform integrity detection, generating a device status assessment result.

[0012] Based on the equipment status assessment results, perform blue light disinfection and update the background equipment management data to complete the equipment return process.

[0013] As a further optimization of this disclosure, the AI ​​algorithm analysis includes a multi-layer neural network recommendation model for generating a personalized device recommendation list, the specific steps of which include:

[0014] Based on the user identity verification data and historical borrowing records, a user-device interaction matrix is ​​constructed and a temporal feature vector is extracted to generate user behavior sequence data.

[0015] The user behavior sequence data is input into a Long Short-Term Memory (LSTM) network model for temporal pattern mining to extract the time-dependent features of user learning behavior.

[0016] Combining the aforementioned time-dependent features and user learning performance data, an attention mechanism is used to weight and fuse multidimensional features, and a graph convolutional neural network (GCN) is used to model the user-device-course ternary relationship to generate a fused feature representation vector.

[0017] Based on the fused feature representation vector, a variational autoencoder (VAE) is used to learn the distribution of potential user preferences, and a generative adversarial network (GAN) is combined to enhance data diversity, outputting a device recommendation probability distribution.

[0018] Based on the device recommendation probability distribution, a multi-armed slot machine algorithm is used to balance exploration and utilization, and reinforcement learning is combined to optimize the recommendation strategy to generate the personalized device recommendation list.

[0019] Optionally, the Long Short-Term Memory (LSTM) network model includes an adaptive forgetting gate mechanism:

[0020] The extraction of time-dependent features of user learning behavior includes:

[0021] The importance weights of historical behaviors are calculated using the user behavior sequence data and the current timestamp information through a time decay function to generate a time decay weight matrix.

[0022] Based on the time decay weight matrix and user behavior characteristics, an adaptive forgetting gate is constructed to dynamically adjust the retention level of historical information and output the filtered memory state.

[0023] The filtered memory states are combined with the current learning context information, and the cell states are updated through a gated recurrent unit to generate the time-dependent features.

[0024] Optionally, the attention mechanism includes multi-head self-attention computation:

[0025] The method employs an attention mechanism to weightedly fuse multi-dimensional features and models the user-device-course ternary relationship using a graph convolutional neural network (GCN) to generate a fused feature representation vector, including:

[0026] The time-dependent features and user profile data are mapped to query vector Q, key vector K, and value vector V, respectively, and an attention calculation matrix is ​​constructed.

[0027] Based on the attention calculation matrix, the weight distribution of multiple attention heads is calculated in parallel and a scaled dot product attention mechanism is used to generate multi-head attention output;

[0028] The multi-head attention output is subjected to linear transformation and residual connection, and then processed by layer normalization to generate the fused feature representation vector.

[0029] Optionally, the variational autoencoder (VAE) includes a Bayesian inference mechanism:

[0030] The method employs a variational autoencoder (VAE) to learn the distribution of latent user preferences and combines it with a generative adversarial network (GAN) to enhance data diversity, outputting a device recommendation probability distribution, including:

[0031] The fused feature representation vector is mapped to the latent space through an encoder network, and latent variables are sampled using a reparameterization technique to generate the mean and variance of the user's latent preference distribution.

[0032] Based on the mean and variance, the Monte Carlo sampling method is used to sample from the latent distribution, and combined with the Kullback-Leibler Divergence (KL divergence) regularization constraint, latent preference codes are generated.

[0033] The potential preferences are encoded, user preferences are reconstructed through a decoder network, and the objective function is optimized using the Evidence Lower Bound (ELBO) to output the device recommendation probability distribution.

[0034] Optionally, the multi-armed slot machine algorithm includes an upper confidence bound (UCB) strategy:

[0035] The process of generating the personalized device recommendation list by employing a multi-armed slot machine algorithm combined with reinforcement learning to optimize the recommendation strategy includes:

[0036] Based on historical recommendation performance data and user feedback information, the average reward value and confidence interval for each device category are calculated, and the device reward estimate is generated.

[0037] Based on the estimated device reward and number of explorations, the UCB value of each device is calculated using the upper confidence bound formula, and combined with Thompson sampling balance exploration and utilization, the device selection probability is determined.

[0038] Based on the device selection probability, the recommendation strategy is dynamically adjusted and combined with the ε-greedy algorithm to increase random exploration, thereby optimizing the generation of the personalized device recommendation list.

[0039] Optionally, generating the personalized device recommendation list further includes:

[0040] User behavior data is collected from multiple lockers, and personalized model parameters are trained locally at each edge node to generate local model gradients.

[0041] The local model gradients of each node are weighted and fused using a secure aggregation protocol, and user privacy is protected through a differential privacy mechanism to generate global model parameter updates.

[0042] Based on the global model parameter update, the recommendation model is updated synchronously in each locker, and the local recommendation effect is optimized by combining personalized fine-tuning.

[0043] Optionally, based on the user identity verification data, the user's historical borrowing records and learning needs data are retrieved for AI algorithm analysis to generate a personalized device recommendation list, including:

[0044] Based on the user identity verification data, extract the user's historical borrowing records, learning progress and course completion rate from the database to construct a user behavior profile;

[0045] Based on the user behavior profile, collaborative filtering algorithm and deep learning model are used to analyze user learning preferences and ability levels, and device compatibility score is calculated.

[0046] Based on the device compatibility score, and in conjunction with the current curriculum and teaching objectives, a recommendation algorithm is used to sort and optimize the devices, generating the personalized device recommendation list.

[0047] Optionally, based on the personalized device recommendation list and the device status information in the locker, a matching calculation is performed to determine a list of available devices, including:

[0048] Real-time acquisition of equipment type, quantity, and status information for each storage compartment in the locker, and establishment of an equipment inventory mapping table;

[0049] Based on the personalized device recommendation list, an intersection operation is performed with the device inventory mapping table to filter out available devices in the inventory;

[0050] Prioritize and verify the availability of the available devices in the inventory to determine the list of available devices for borrowing.

[0051] Optionally, based on the list of available equipment, an opening operation is performed through an electromagnetic door control system to complete the equipment borrowing process, including:

[0052] Based on the target device selected by the user, query the location information of the corresponding storage compartment in the list of available devices, and generate an opening command;

[0053] The opening command is sent to the electromagnetic door of the specified storage compartment through a RESTful API (Representational State Transfer) style application programming interface, which controls the electromagnetic lock to release and open the cabinet door.

[0054] After monitoring the status sensor signal of the storage compartment door and confirming that the user has taken out the device and closed the cabinet door, the device borrowing status record is updated to complete the device borrowing process.

[0055] Optionally, images of the device are captured using a close-up camera and combined with AI recognition technology to perform integrity detection, generating a device status assessment result, including:

[0056] Control the close-up camera to capture images of the returned device from multiple angles and obtain image data of the device's appearance;

[0057] Defect detection is performed on the device appearance image data based on a deep learning image recognition model to identify device damage, screen damage, and missing components.

[0058] Perform power-on tests on devices that support automatic detection to check the functionality of the devices and record the test results;

[0059] The device status assessment result is generated by combining the appearance inspection results and the functional test results.

[0060] Optionally, based on the equipment status assessment results, blue light disinfection is performed and the background equipment management data is updated to complete the equipment return process, including:

[0061] The equipment condition is determined based on the equipment condition assessment results. If the equipment is in good condition, the blue light disinfection device is activated to disinfect the equipment.

[0062] The duration of blue light irradiation is set by a timer to ensure that the disinfection effect meets hygiene and safety standards.

[0063] After disinfection is completed, the equipment status, disinfection time, and return record are uploaded to the backend management system through the data synchronization interface to complete the equipment return process.

[0064] Optionally, it also includes AI digital human interaction capabilities:

[0065] When the panoramic camera detects that a user is approaching, the AI ​​digital human program is launched and switched to the welcome screen, playing the digital human's guiding voice.

[0066] It receives user voice input, converts speech into text commands through speech recognition technology, and combines natural language processing algorithms to understand user intent;

[0067] Based on the user's intent, the system matches the corresponding response content, generates a digital human's response voice through speech synthesis technology, and plays it back to achieve intelligent interactive dialogue.

[0068] Optionally, remote management functionality is also included:

[0069] A network connection is established through a gigabit switch to enable data communication between multiple lockers and the back-end management platform;

[0070] The back-end management platform uses WebSocket technology (a network protocol that enables full-duplex (bidirectional) communication over a single TCP connection) to monitor the borrowing and returning status, inventory, and usage statistics of each locker in real time, generating a visual data dashboard.

[0071] When a user clicks the help button on the locker, the administrator can access the corresponding locker's camera through a remote communication system to enable remote video calls and screen sharing, providing real-time technical support.

[0072] This disclosure also provides a teaching equipment management device, including a processor and a memory. The memory stores machine-readable instructions executable by the processor. The processor is used to execute the machine-readable instructions stored in the memory. When the machine-readable instructions are executed by the processor, the processor performs the steps of the teaching equipment management method described above.

[0073] Based on a single implementation method or combination of implementation methods of the above-mentioned teaching equipment management methods or devices, it can be seen that this disclosure, by integrating technologies such as facial recognition, automatic door opening, and intelligent detection, has achieved automation of the entire process of equipment borrowing and returning, greatly simplifying operation steps, shortening user waiting time, and improving equipment turnover efficiency.

[0074] This disclosure embodiment can utilize AI algorithms to deeply analyze users' historical behavior and learning needs, providing users with personalized device recommendations, achieving precise matching of teaching resources, and effectively improving students' interest and efficiency in self-directed learning.

[0075] In this embodiment of the invention, when returning equipment, the integrity of the equipment can be automatically detected by AI image recognition technology, and blue light disinfection can be performed, which effectively ensures the physical safety and hygiene safety of the equipment and reduces the risk of equipment damage and cross-infection.

[0076] Furthermore, the optional embodiments of this disclosure also introduce AI digital human interaction functionality, providing users with more user-friendly and intelligent guidance and Q&A services; the remote management platform provides administrators with real-time monitoring and technical support capabilities, enhancing the overall system's intelligence level. The system can collect and analyze device usage data and user behavior data in real time, providing strong data support for educational institutions' resource allocation and curriculum optimization, forming a sustainable and optimized management loop.

[0077] To more clearly illustrate the technical solutions in the embodiments of this disclosure or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Attached Figure Description

[0078] To more clearly illustrate the technical solutions in the embodiments of this disclosure or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0079] Figure 1 This is a schematic diagram of the overall process of a teaching equipment management method according to an embodiment of the present disclosure;

[0080] Figure 2 This is a detailed flowchart of the AI ​​algorithm analysis steps according to an embodiment of the present disclosure;

[0081] Figure 3a An interactive interface for configuring the appearance of the locker in this embodiment of the disclosure;

[0082] Figure 3b To create an interface for binding the sub-locker type when adding sub-lockers to a locker;

[0083] Figure 3c The interactive interface for adding sub-lockers to the locker;

[0084] Figure 4a This is the teacher login page;

[0085] Figure 4bThis is a schematic diagram of the authorization and authentication interface;

[0086] Figure 4c This is a diagram showing how teachers can click on each storage compartment and select the storage device on the large screen of the storage cabinet.

[0087] Figure 4d This is a diagram illustrating the placement of actual equipment into the storage compartments of a storage cabinet.

[0088] Figure 5 A landing page for AI digital humans;

[0089] Figure 6 Provides an interactive interface for the backend platform to view storage cabinet information online;

[0090] Figure 7 This is a remote assistance interface;

[0091] Figure 8 This is a schematic diagram of a computer device provided for an embodiment of the present disclosure. Detailed Implementation

[0092] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the technical solutions of the embodiments of this disclosure 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 this disclosure, and not all of them. The components of the embodiments of this disclosure described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this disclosure provided in the accompanying drawings is not intended to limit the scope of the claimed disclosure, but merely represents selected embodiments of this disclosure. All other embodiments obtained by those skilled in the art based on the embodiments of this disclosure without inventive effort are within the scope of protection of this disclosure.

[0093] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0094] In this document, the term "and / or" merely describes a relationship, indicating that three relationships can exist. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. Furthermore, the term "at least one" in this document means any combination of at least two of any one or more elements. For example, including at least one of A, B, and C can mean including any one or more elements selected from the set consisting of A, B, and C.

[0095] Research has found that manual management methods for teaching equipment are cumbersome, inefficient, and prone to errors. Furthermore, with increased awareness of health and safety, especially after public health emergencies, the hygiene of teaching equipment has become a significant concern. Traditional management methods often struggle to standardize and fully cover equipment disinfection and cleaning. Manual disinfection is not only inefficient but also cannot guarantee thoroughness and effectiveness each time, posing a risk of cross-infection. More importantly, there is a general lack of intelligent perception and personalized services based on user needs. In manual management of teaching equipment, equipment recommendations and allocations are often based on the subjective experience of administrators or teachers, failing to dynamically and accurately match equipment to students' learning progress, historical behavior, and ability levels. This results in the inefficient allocation of teaching resources, making it difficult for students to access teaching tools best suited to their current stage, thus impacting their motivation and effectiveness in self-directed learning. Therefore, how to utilize artificial intelligence and the Internet of Things (IoT) technologies to overcome these shortcomings, achieve automation, intelligence, and personalization in teaching equipment management, improve management efficiency, ensure equipment safety and hygiene, and enhance user experience is the problem that the following embodiments of this disclosure aim to solve.

[0096] Example 1

[0097] Please see Figure 1 This disclosure provides a method for managing teaching equipment, including the following steps:

[0098] S1: Obtain user authentication information, perform identity authentication through the target method, and generate user identity confirmation data.

[0099] In this step, multimodal user authentication is achieved through panoramic cameras, card readers, and QR code scanning modules deployed on the locker equipment. When a user approaches the locker, their face is automatically recognized; users can also choose to swipe their student card or scan a dynamic QR code generated by a mobile app. The collected information is compared with the backend user database, and upon successful verification, confirmation data containing the user's unique identifier is generated, providing a basis for identity verification for subsequent personalized services.

[0100] This step enables quick, seamless, or convenient user authentication, forming the foundation for subsequent personalized services. Specific authentication methods can be used individually or in combination to adapt to different scenarios and security levels.

[0101] S2: Based on the user identity verification data, retrieve the user's historical borrowing records and learning demand data for AI algorithm analysis to generate a personalized device recommendation list.

[0102] After verifying a user's identity, the system can automatically retrieve data from the backend database, including the user's historical device borrowing records, current courses, learning progress, historical grades, and relevant skill tags. The AI ​​recommendation engine comprehensively analyzes this multi-dimensional data to uncover the user's learning preferences, skill gaps, and potential needs, thereby dynamically generating a device recommendation list that matches the user's current learning requirements, such as recommending specific Arduino development boards or sensor kits.

[0103] This is a key step in achieving adaptability and intelligence. An AI engine can analyze user behavior patterns, learning progress, and potential needs to recommend the most suitable teaching equipment for the current user, transforming passive borrowing into proactive recommendation.

[0104] S3: Based on the personalized device recommendation list, and combined with the device status information in the locker, a matching calculation is performed to determine the list of devices that can be borrowed.

[0105] Here, the AI-generated recommendation list can be compared with the real-time status of all devices in the lockers. Device status information includes whether the device is in stock, whether it is in good condition, whether it has been disinfected, and whether it is being reserved. Through a matching calculation, the currently available devices in the recommendation list are filtered out and sorted by recommendation priority, ultimately forming a device list that users can directly select.

[0106] By implementing this step, the recommended list can be compared with the real-time inventory, filtering out devices that have been borrowed, are charging, or are under maintenance, and presenting users with an accurate list of available devices, thereby ensuring that the recommended devices are actually available.

[0107] S4: Based on the list of available equipment, the door is opened through the electromagnetic door control system to complete the equipment borrowing process.

[0108] Here, users select the desired device from the list of available devices on the touchscreen. Upon receiving the user's selection, the system sends an opening command to the electromagnetic door control module of the storage compartment containing that device. The electromagnetic lock unlocks immediately upon receiving the command, and the cabinet door automatically opens. After the user removes the device and closes the cabinet door, the door magnetic sensor detects the door is closed, and the system automatically updates the device's status to "borrowed," recording the borrower's information and the time.

[0109] S5: Receives device return request, captures device images through a close-up camera and combines AI recognition technology to perform integrity detection, and generates device status assessment results.

[0110] Here, when returning the device, the user selects "Return" on the interface and places it in the designated storage compartment. The close-up high-definition camera inside the compartment automatically takes multi-angle photos of the device. An AI image recognition model analyzes the captured images and compares them with a standard image of the device's appearance to detect scratches, damage, missing components, etc. For electronic devices that support self-testing, the system also performs a power-on self-test, ultimately generating a comprehensive evaluation result that includes both appearance and functional status.

[0111] S6: Based on the equipment status assessment results, perform blue light disinfection and update the background equipment management data to complete the equipment return process.

[0112] Here, if the device's status assessment result is "intact," the blue LED disinfection device inside the storage compartment will automatically activate to disinfect the device for a set duration, ensuring hygiene and safety. After disinfection, the system will synchronously update the return record, status assessment result, disinfection time, and other information to the backend database. If the device is determined to be "damaged," the system will lock the storage compartment and send an alarm notification to the administrator.

[0113] Example 2

[0114] This embodiment is a detailed refinement of step S2 (AI algorithm analysis) in Embodiment 1. Please refer to... Figure 2 The specific implementation of the AI ​​algorithm analysis can be a multi-layer neural network recommendation model, which further includes:

[0115] S2.1: Based on the user identity verification data and historical borrowing records, construct a user-device interaction matrix and extract time-series feature vectors to generate user behavior sequence data.

[0116] Here, we first organize users' historical behaviors to construct a sparse user-device interaction matrix, where the elements of the matrix can represent the number of times or frequency a user borrows a certain device. At the same time, we sort the users' borrowing behaviors by time to form a behavior sequence containing device identifiers (IDs) and timestamps, and extract time-series features such as borrowing intervals and borrowing durations, which together constitute the user behavior sequence data.

[0117] This step involves preprocessing the raw data. The system treats each user's device borrowing behavior (including device type, borrowing duration, borrowing time, etc.) as an event, arranging them chronologically to form time-series data that reflects the evolution of user behavior. Simultaneously, a sparse matrix is ​​constructed, where rows represent users, columns represent devices, and values ​​in the matrix represent user interaction frequency or preference.

[0118] S2.2: Input the user behavior sequence data into the Long Short-Term Memory (LSTM) network model for temporal pattern mining to extract the time-dependent features of user learning behavior.

[0119] Here, leveraging the powerful time-series data processing capabilities of the LSTM model, we can learn from user behavior sequence data to capture user learning habits, such as whether they prefer concentrated or distributed learning, and the patterns of their interest in different device types over time. Furthermore, we can capture long-term dependencies and short-term interest changes in user device selection at different points in time. For example, we can identify that when a user learns a particular project, they typically borrow devices in the order of "basic kit -> advanced module -> specialized sensor." The hidden states output by the model represent the time-dependent user behavior features.

[0120] S2.3: Combining the time-dependent features and user learning performance data, an attention mechanism is used to weight and fuse multi-dimensional features, and a graph convolutional neural network (GCN) is used to model the user-device-course ternary relationship to generate a fused feature representation vector.

[0121] Here, the time-dependent features extracted by LSTM, the user's basic profile (such as major and grade), and the user's academic performance data are used as different feature inputs. Through an attention mechanism, the system can adaptively assign weights to different features, highlighting the most important features for the current task. Subsequently, a Graph Convolutional Network (GCN) is used to model the complex relationship between "user-device-course," fusing node information and graph structure information to generate a highly condensed and expressive fused feature representation vector. Here, GCN can learn the complex relationship graph formed by users, devices, and courses, for example, which users achieved better grades in relevant courses after using specific devices, thereby uncovering deeper correlations.

[0122] S2.4: Based on the fused feature representation vector, a variational autoencoder (VAE) is used to learn the distribution of potential user preferences, and a generative adversarial network (GAN) is combined to enhance data diversity, outputting a device recommendation probability distribution.

[0123] Here, the fused feature representation vector is input into the VAE. The VAE's encoder maps it to a latent space, learning a probability distribution of user preferences (not just a single point), which increases the model's generalization ability. To address the data sparsity problem, a GAN can be introduced, where a generator produces realistic, but not real-world, user behavior data based on the learned preference distribution, enhancing the diversity of the training set. Finally, the VAE's decoder outputs a probability distribution of user preferences for various candidate devices based on the enhanced and learned latent preferences.

[0124] S2.5: Based on the device recommendation probability distribution, a multi-armed slot machine algorithm is used to balance exploration and utilization, and reinforcement learning is combined to optimize the recommendation strategy to generate the personalized device recommendation list.

[0125] Here, each device to be recommended is treated as a "slot machine arm," and its recommendation performance (such as whether the user borrows it) is considered the "reward." MAB algorithms, such as UCB (Upper Confidence Bound), are employed to balance "utilization" (recommending devices known to perform well) and "exploration" (trying to recommend new devices to discover potential user interests). This process can be modeled using a reinforcement learning framework, dynamically adjusting the recommendation strategy based on real-time user feedback (borrowing, ignoring, etc.). Finally, based on the device recommendation probability distribution and the optimized strategy, a ranked, personalized device recommendation list is generated.

[0126] This step is the decision-making stage for the recommendation strategy. "Utilization" refers to the devices the recommendation system believes the user prefers, while "exploration" refers to recommending new devices that the user might like but the system is not yet certain about. MAB algorithms (such as UCB or Thompson sampling) can effectively balance these two aspects, preventing the recommendation results from getting stuck in local optima. Simultaneously, by combining reinforcement learning, the system can continuously adjust its recommendation strategy based on the user's real-time feedback (whether to adopt the recommendation), achieving continuous optimization.

[0127] Example 3

[0128] This embodiment is a further refinement of step S2.2 (Long Short-Term Memory Network LSTM Model) in Embodiment 2. The Long Short-Term Memory Network LSTM Model may include an adaptive forgetting gate mechanism, and its steps for extracting the time-dependent features of user learning behavior include:

[0129] S2.2.1: Calculate the importance weight of historical behaviors using the user behavior sequence data and current timestamp information through a time decay function to generate a time decay weighting matrix.

[0130] To differentiate the impact of recent and long-term behaviors on current recommendations, a time decay function (such as exponential decay) can be used. A weight value is calculated based on the interval between the time each historical behavior occurred and the current time. Behaviors more recent than the current time have higher weights. The weights of all historical behaviors together form a time decay weight matrix.

[0131] This step assumes that actions more recent in time have a greater impact on predicting future actions. For example, a device borrowed by a user last week is more relevant to predicting future actions than a device borrowed a year ago. A time decay function (such as exponential decay) assigns a weight to each historical action, with the weight decreasing as time progresses.

[0132] S2.2.2: Based on the time decay weight matrix and user behavior characteristics, an adaptive forgetting gate is constructed to dynamically adjust the retention level of historical information and output the filtered memory state.

[0133] In practical implementation, this time decay weight matrix is ​​introduced into the forget gate calculation of standard LSTM. This way, when deciding which historical information to "forget," the forget gate considers not only the input data and the previous state, but also, and more importantly, the time factor. For older information with lower weights, the forget gate is more likely to forget it, thus dynamically and adaptively adjusting the degree of historical information retention.

[0134] Here, the traditional LSTM forget gate treats all historical information equally, while the adaptive forget gate combines the time decay weights calculated in the previous step, enabling the model to intelligently decide which outdated information should be "forgotten" and which important recent information should be "remembered," thus making the model's memory more timely.

[0135] S2.2.3: The filtered memory state is combined with the current learning context information and the cell state is updated through a gated loop unit to generate the time-dependent feature.

[0136] Here, the memory state, filtered by the adaptive forgetting gate, is combined with information representing the current learning context (such as the course or experiment the user is currently undertaking) as input, and then...

[0137] GRU (Gated Recurrent Unit) or other similar recurrent neural network units are used to update the cell state of the LSTM. This final cell state, or hidden state, is a feature vector containing precise time dependencies.

[0138] By combining filtered, more valuable historical information (memory state) with the current user's input (such as information about the course in progress), and updating it through GRU (a recurrent neural network unit that is more lightweight than LSTM), a vector that can accurately represent the user's learning behavior at a specific point in time is finally generated.

[0139] Example 4

[0140] This embodiment is a further refinement of step S2.3 (attention mechanism) in Embodiment 2. The process of using an attention mechanism to weightedly fuse multi-dimensional features may include multi-head self-attention computation, the steps of which include:

[0141] S2.3.1: Map the time-dependent features and user profile data to query vector Q, key vector K, and value vector V respectively, and construct an attention calculation matrix.

[0142] Here, the time-dependent feature vector output by the LSTM and the user's static profile feature vector (such as major, grade, etc.) are concatenated or linearly transformed separately to generate three different vectors: a query vector, a key vector, and a value vector. This step is to enable the model to evaluate the correlation between different features in subsequent calculations.

[0143] This is the standard procedure for attention mechanisms. Q represents the information to be focused on, K represents the information available for reference in the dataset, and V represents the actual numerical values ​​of this information. Different input features (such as the temporal features of the LSTM output, basic user information, etc.) are linearly transformed into Q, K, and V to prepare for subsequent calculations.

[0144] S2.3.2: Based on the attention calculation matrix, the weight distribution of multiple attention heads is calculated in parallel and a scaled dot product attention mechanism is used to generate multi-head attention output.

[0145] Multi-head attention mechanisms allow models to learn information in parallel across different representation subspaces. The system divides the Q, K, and V vectors into multiple "heads," each independently performing scaled dot product attention computation (i.e., calculating the dot product of Q and K, obtaining weights through the Softmax function, and then weighted summing over V). In other words, the multi-head mechanism involves splitting Q, K, and V into multiple sub-parts and calculating attention weights separately in different "subspaces." This allows the model to simultaneously focus on information from different sources. For example, one head might focus on a user's short-term interests, while another head might focus on the user's long-term capabilities. Scaled dot product is an efficient method for calculating attention weights.

[0146] S2.3.3: The multi-head attention output is linearly transformed and residually connected, and then normalized through layers to generate the fused feature representation vector.

[0147] Here, the outputs of all attention heads are concatenated and transformed through a final linear layer. To prevent the vanishing gradient problem in deep networks and accelerate training, residual connections (applying the input directly to the output) and layer normalization are introduced. After this series of processes, the resulting vector is the final, high-quality fused feature representation vector. The residual connections and layer normalization help prevent gradient vanishing or exploding, accelerate model convergence, and improve the quality of the final generated fused feature representation vector.

[0148] Example 5

[0149] This embodiment is a further refinement of step S2.4 (Variational Autoencoder, VAE) in Embodiment 2. The process of using a Variational Autoencoder (VAE) to learn the distribution of latent user preferences may include a Bayesian inference mechanism, the steps of which include:

[0150] S2.4.1: The fused feature representation vector is mapped to the latent space through the encoder network and latent variables are sampled using reparameterization techniques to generate the mean and variance of the user latent preference distribution.

[0151] Here, the fused feature representation vector generated in Example 4 is input into the encoder (a neural network) of the VAE. The encoder does not directly output a latent vector, but instead outputs two parameters describing the latent distribution: the mean (μ) and the log-variance (logσ). 2 This reflects Bayesian inference, which posits that user preferences are not a fixed point, but rather a probability distribution.

[0152] Unlike traditional autoencoders that learn a deterministic code, the VAEs described above learn a probability distribution (usually Gaussian), described by mean and variance. The reparameterization technique is a method to make the sampling process differentiable, thus enabling gradient descent to be used to train the entire network.

[0153] S2.4.2: Based on the mean and variance parameters, the Monte Carlo sampling method is used to sample from the latent distribution and combined with the KL divergence regularization constraint to generate the latent preference code.

[0154] Here, to obtain a specific latent vector from the learned distribution for decoding, the system employs a reparameterization technique (z = μ + ε*σ, where ε is random noise sampled from the standard normal distribution). This allows the gradient to propagate backward. Simultaneously, a KL divergence term is added to the model's loss function to measure the difference between the learned latent distribution and the standard normal prior distribution, acting as a regularization mechanism to prevent overfitting and improve the structure of the latent space.

[0155] S2.4.3: Encode the potential preferences, reconstruct user preferences through a decoder network, and optimize the objective function using ELBO to output the device recommendation probability distribution.

[0156] Here, the sampled latent preference encoding is input into the decoder network. The decoder's task is to reconstruct the original user preference data (e.g., reconstruct a row in the user-device interaction matrix) based on this latent encoding. The model's optimization objective is to maximize the lower bound of evidence (ELBO), which consists of the reconstruction loss and the KL divergence. After training, the decoder can generate the user's preference probabilities for all devices based on the latent preference encoding.

[0157] Here, the decoder network acts as the opposite of the encoder; it attempts to recover the original user preference information (such as a row in the user-device interaction matrix) from the latent preference encoding. By optimizing the ELBO (lower bound of evidence), the reconstructed preferences are made as close as possible to the true preferences, thus training an effective encoder-decoder. The final output of the decoder can then serve as the probability distribution for device recommendations.

[0158] Example 6

[0159] This embodiment is a further refinement of step S2.5 (multi-armed slot machine algorithm) in embodiment two. The process of balancing exploration and exploitation using the multi-armed slot machine algorithm may include the upper confidence bound (UCB) strategy, the steps of which include:

[0160] S2.5.1: Based on historical recommendation performance data and user feedback information, calculate the average reward value and confidence interval for each device category, and generate device reward estimates.

[0161] In practice, the results of each recommendation can be continuously tracked. If a user uses the recommended device, it is recorded as a positive reward; otherwise, it is recorded as a zero or negative reward. Here, an estimate of the "average reward" is maintained for each device (or device category), and a confidence interval for the reward estimate is calculated based on the number of times the device has been recommended. The fewer the number of attempts, the greater the uncertainty, and the wider the confidence interval.

[0162] S2.5.2: Based on the estimated device reward and the number of explorations, calculate the UCB value of each device using the upper confidence bound formula and combine it with Thompson sampling balance exploration and utilization to determine the device selection probability.

[0163] Here, the core of the UCB strategy is to calculate a UCB value for each device, which equals its average reward estimate plus an exploration term related to the confidence interval width. This exploration term decreases as the number of explorations increases. In each round of recommendation, the system prioritizes the device with the highest UCB value. Furthermore, Thompson sampling can be incorporated; instead of directly selecting the optimal value, a value is sampled from the posterior probability distribution of each arm, and the arm with the largest sampled value is selected, further enhancing the exploration capability.

[0164] S2.5.3: Based on the device selection probability, dynamically adjust the recommendation strategy and combine it with the ε-greedy algorithm to increase random exploration, thereby optimizing the generation of the personalized device recommendation list.

[0165] In practice, the devices to be prioritized for recommendation next can be determined based on the calculated UCB value or Thompson sampling results. To avoid getting trapped in local optima, an ε-greedy strategy can be introduced, where there is a tiny probability (ε) of completely random selection of a device for recommendation, and a probability (1-ε) of recommendation based on the UCB or Thompson sampling results. This dynamically adjusted strategy ultimately affects the probability distribution of the VAE output, reordering or weighting it to optimize the generation of the final list.

[0166] Generally, the device with the highest UCB value is selected most of the time (greedy strategy), but a small probability ε (e.g., 5%) is reserved to randomly select another device for recommendation. This ε-greedy strategy ensures that the system does not completely stop exploring, helping to discover new user interests and thus continuously optimizing the quality and diversity of the recommendation list.

[0167] Example 7

[0168] This embodiment is an extension of the method described in Embodiment 2, and may further include a federated learning mechanism to improve model performance using data from multiple lockers while protecting user privacy. The steps include:

[0169] Step A: Collect user behavior data from multiple lockers, train personalized model parameters locally at each edge node, and generate local model gradients.

[0170] Here, it is assumed that multiple smart lockers (edge ​​nodes) are deployed within the school. Each locker only uses the user interaction data collected by its own node to train its recommendation model locally (as described in Example 2). The user's original data does not leave the local node. After training locally for one or more rounds, the gradient update values ​​of the model parameters are calculated, rather than uploading the model itself or the original data.

[0171] Step B: The local model gradients of each node are weighted and fused using a secure aggregation protocol, and user privacy is protected through a differential privacy mechanism to generate global model parameter updates.

[0172] In implementation, a central aggregation server is responsible for collecting gradient updates uploaded by all edge nodes. The server employs protocols such as Secure Aggregation to ensure it only receives the sum of all gradients, without knowing the specific gradient values ​​of any individual node. Before aggregation, a suitable amount of precisely calculated noise (differential privacy) can be added to the gradients of each node to further enhance privacy protection. The server then performs a weighted average of the aggregated gradients (e.g., based on the amount of data at each node) to form a single update to the global model.

[0173] Step C: Based on the global model parameter update, the recommendation model is updated synchronously in each locker, and the local recommendation effect is optimized by combining personalized fine-tuning to improve the accuracy of the personalized device recommendation list.

[0174] The central server updates and distributes the calculated global model parameters back to all edge nodes. Each locker, upon receiving this global update, uses it to update its local model. In this way, the local model learns common knowledge from the global model while retaining its own unique characteristics. Based on this, several rounds of fine-tuning can be performed using local data to better adapt to the characteristics of the local user group, thereby improving the overall accuracy of the recommendation list.

[0175] Example 8

[0176] This embodiment is a detailed expansion of step S1 (obtaining user authentication information) in Embodiment 1, and its steps include:

[0177] S1.1: Based on the panoramic camera detecting the user's approach, the face recognition algorithm is activated to extract facial features and generate facial feature data.

[0178] In practice, the panoramic camera on top of the locker operates continuously with low power. When a human figure is detected approaching within its field of view to a predetermined distance, the system is activated and a high-definition camera is launched to accurately capture facial images. Face recognition algorithms (such as MTCNN for detection and alignment, and FaceNet or ArcFace for feature extraction) process the captured face images to extract a high-dimensional facial feature vector that uniquely identifies the user.

[0179] S1.2: Obtain user card swipe information, read the student card's identity code for database comparison, and generate card verification results.

[0180] In practice, users can place their campus card or student ID in the designated NFC / RFID reader area. The reader will read the unique identification code (such as student ID number) stored in the card chip. The system sends this identification code to the backend user database for comparison. If a match is found, a card verification success result containing user information is generated.

[0181] S1.3: Receive QR code scanning information from the mobile phone, parse the user identity code in the QR code to verify permissions, and generate QR code authentication data.

[0182] In practice, users can log in via the school's official app or mini-program to generate a time-sensitive dynamic QR code. They then scan the QR code on their phone screen against the locker's scanner, which reads and parses it. The QR code typically contains an encrypted user identification code and a timestamp. The system decrypts the code, verifies its validity and timeliness, and generates successful authentication data upon successful verification.

[0183] S1.4: Combine the facial feature data, the card verification result, and the QR code authentication data to generate the user identity confirmation data.

[0184] In practice, multiple authentication methods can be provided to users, who can choose one or more for verification. The system will integrate all successful verification results. For example, a strategy can be set up with facial recognition as the primary method and card swiping as a secondary method. As long as one method of verification passes, the system will generate final user identity confirmation data, which includes key information such as the user's student ID, name, and college, used to initiate the subsequent borrowing process.

[0185] Example 9

[0186] This embodiment provides another or simplified implementation of step S2 (AI algorithm analysis) in Embodiment 1, the steps of which include:

[0187] S2.1: Extract the user's historical borrowing records, learning progress, and course completion rate from the database based on the user identity verification data to construct a user behavior profile.

[0188] After user identity is confirmed, the system accesses the backend database and extracts a series of data related to that user, including: a list of all historical borrowed devices and their dates; the courses being studied and their current progress (e.g., which chapter they have reached); and grades or assessments for completed courses. The system integrates this structured and unstructured data to form a multi-dimensional user behavior profile.

[0189] This step creates a multi-dimensional, tagged profile for the user. For example, user A's profile might be: "{Borrowing history: [Arduino, Raspberry Pi], Learning progress: 'Advanced C++ Programming' 80%, Course completion: 'Embedded Systems' 100%, Active time: 'Weekday evenings'}".

[0190] S2.2: Based on the user behavior profile, use collaborative filtering algorithm and deep learning model to analyze user learning preferences and ability levels, and calculate device compatibility score.

[0191] In practice, classic collaborative filtering algorithms (such as user-based or item-based algorithms) can be used to predict preferences by analyzing "what devices other users with similar behaviors like prefer" or "what other devices users who have borrowed this device also prefer." Simultaneously, a simple deep learning model (such as a multilayer perceptron (MLP)) can be combined, using multiple dimensions of the user profile as input to directly predict the user's fit score for each device. This score comprehensively reflects the degree of matching between the user's interests and current skill level and the device.

[0192] Here, collaborative filtering algorithms find other users ("neighbors") with similar behavioral profiles to the current user and recommend devices that these "neighbors" like but that the current user hasn't used. Deep learning models, on the other hand, can learn more complex non-linear relationships from these profiles, comprehensively assessing the user's ability level (e.g., judging from course grades) and learning preferences, and calculating a suitability score for each device.

[0193] S2.3: Based on the device compatibility score, combined with the current curriculum system and teaching objectives, a recommendation algorithm is used to sort and optimize the devices, generating the personalized device recommendation list.

[0194] In practice, after obtaining the compatibility scores of all candidate devices, business rules can be introduced for optimization. For example, based on the user's current course syllabus, devices necessary to complete the course can be given higher weight. Finally, by combining the compatibility scores and business rules, all available devices are sorted in descending order, and the top-ranked devices are selected to form the final personalized list recommended to the user.

[0195] Example 10

[0196] This embodiment is a detailed expansion of step S3 (determining the list of available borrowable equipment) in embodiment one, and its steps include:

[0197] S3.1: Obtain real-time information on the equipment type, quantity, and status of each storage compartment in the locker, and establish an equipment inventory mapping table.

[0198] Here, each storage compartment is equipped with sensors (such as infrared sensors, weight sensors, or RFID readers) to detect whether there is any equipment inside and the type of equipment. The main control system of the lockers periodically polls or uses an event-triggered mechanism to collect the status of all storage compartments and builds a real-time equipment inventory mapping table in memory. This table records each storage compartment number, the ID of the stored equipment, the equipment type, and the current status (available, damaged, being disinfected), etc. For example: "{Compartment 01:{Device:'VR Glasses', Status:'Available'}, Compartment 02:{Device:'Drone', Status:'Charging'},...}".

[0199] S3.2: Based on the personalized device recommendation list, perform an intersection operation with the device inventory mapping table to filter out available devices in the inventory.

[0200] In practice, the personalized device recommendation list (containing recommended device types or specific models) generated in step S2 can be logically intersected with the real-time device inventory mapping table established in the previous step. The purpose of this operation is to filter out devices that are both needed by the user and in stock in the locker with a status of "available".

[0201] For example, comparing the list of AI-recommended devices (e.g., ['VR glasses', 'Development board A', 'Drone']) with devices in the inventory mapping table that are "available" yields an intersection (e.g., ...).

[0202] ['VR glasses']), this intersection is the device that users can theoretically borrow immediately.

[0203] S3.3: Prioritize and verify the availability of the available devices in the inventory to determine the list of available devices.

[0204] Here, the selected devices can be sorted according to the original order (i.e., recommendation priority) in the personalized recommendation list. Before finally presenting them to the user, a final availability verification can be performed, such as pinging the device's network status again or performing a quick self-check to ensure everything is in order. Finally, the list of all verified and sorted devices will be displayed on the user interface's touchscreen, forming the final list of available devices.

[0205] Example 11

[0206] This embodiment is a detailed elaboration of step S4 (completing the equipment borrowing process) in embodiment one. The steps include:

[0207] S4.1: Based on the target device selected by the user, query the location information of the corresponding storage compartment in the list of available devices and generate an opening command.

[0208] Here, after a user selects a desired device on the touchscreen, the system can look up the specific storage compartment number of that device from a list of available devices in memory or an inventory mapping table. This storage compartment number can then be encapsulated into a standard-formatted door-opening instruction, such as a JSON object containing the target address and the operation command.

[0209] S4.2: Send the opening command to the electromagnetic door of the specified storage compartment via RESTful API to control the electromagnetic lock to release and open the cabinet door.

[0210] In practice, the main control system of the locker acts as a server, providing an API interface for the controller of each storage compartment. The main control system sends an HTTP POST request via the local area network to the control board of the target storage compartment (usually a microcontroller, such as an ESP32). The request body contains the door opening command generated in the previous step. After receiving and verifying the command, the control board drives a relay or MOSFET to energize (or de-energize, depending on the lock type) the electromagnetic lock, causing the bolt to retract and the lock door to automatically open under the action of a spring.

[0211] Here, the main controller of the locker receives the door opening command via a network (such as Wi-Fi or Ethernet). The main controller then sends a low-level signal to the door lock control board of the target compartment via an internal bus (such as RS485). RESTful API is a standardized network communication interface that ensures the reliability and security of command transmission. After receiving the signal, the electromagnetic lock de-energizes its internal coil, the bolt retracts under the action of the spring, and the lock door automatically opens.

[0212] S4.3: Monitor the storage compartment door status sensor signal, confirm that the user has taken out the device and closed the cabinet door, update the device borrowing status record, and complete the device borrowing process.

[0213] In practice, a door magnetic sensor can be installed on each cabinet door. After the cabinet door opens, the system begins monitoring the sensor's signal. When the user removes the device and manually closes the cabinet door, the door magnetic sensor's status changes from "open" to "closed." The system detects this status change and considers the device to have been successfully removed. At this point, the system immediately updates the background database, marking the device's status as "borrowed" and associating it with the current user's ID and borrowing timestamp. The borrowing process is then complete.

[0214] Here, Hall effect sensors or microswitches on the door monitor the cabinet door's status (open / closed) and whether the device has been removed (infrared sensor). Once the system detects the sequence of events—"door open -> device removed -> door closed"—it determines the borrowing action is complete. At this point, the system automatically updates the device's status to "borrowed" in the background database and updates it to the current user's information.

[0215] ID association, recording the borrowing time.

[0216] Example 12

[0217] This embodiment is a detailed elaboration of step S5 (performing integrity detection) in Embodiment 1. The steps include:

[0218] S5.1: Control the close-up camera to capture images of the returned device from multiple angles and obtain image data of the device's appearance.

[0219] In practice, after a user places the device in the return compartment or a designated storage compartment, they can control the internally installed close-up cameras (there may be more than one, or a single rotatable camera, usually a wide-angle or fisheye lens) to take photos or record a short video from different angles (such as top, front, and side) and under different lighting conditions. For example, a high-definition image can be taken from the top, front, left, and right sides to ensure full coverage of the device's main exterior surfaces.

[0220] S5.2: Based on a deep learning image recognition model, perform defect detection on the device appearance image data to identify device damage, screen damage, and missing components.

[0221] In practice, multiple acquired images can be input into a pre-trained deep learning defect detection model (such as YOLO, Faster R-CNN, etc.). This model is trained to identify specific defect types; for example, it can locate areas such as cracks, scratches, broken casings, and shattered screens in the images. For devices such as development boards, the model can also identify whether important chips or interface components are missing by comparing them with standard images.

[0222] Here, the AI ​​model in the background has been trained on a large number of images of normal devices and various damaged devices. It can automatically locate the main body of the device in the acquired images and detect areas that do not conform to the standard appearance, such as cracks, scratches, dents, stains, etc. For devices with specific components (such as development boards), the model can also perform component integrity checks.

[0223] S5.3: Perform a power-on test on the equipment that supports automatic detection, check the functionality of the equipment, and record the test results.

[0224] In practice, if the returned device is a smart device with self-testing capabilities (e.g., a built-in microcontroller), the interface inside the storage compartment will provide power. The system sends a self-test command to the device via the data interface, and the device will run a preset test program to check whether its core functions (such as sensor readings, network connection, screen display, etc.) are normal, and return the test results to the main control system of the locker.

[0225] Here, for electronic devices with standard interfaces (such as USB and HDMI), an automatic testing module can be integrated into the storage compartment. When the device is placed inside, the testing module will automatically turn on the power and send test signals (such as turning on the screen or detecting sensor readings) through the interface to determine whether the core functions of the device are normal, and record the test result as "normal function" or "abnormal function".

[0226] S5.4: Based on the combined results of appearance inspection and functional testing, generate the equipment status assessment result.

[0227] In practice, the results of appearance defect detection and functional self-test can be combined. These results are then comprehensively evaluated using a pre-defined rule engine. For example, the rule can define "minor scratches" as "light wear," and "screen crack" or "functional test failure" as "severe damage." Ultimately, the system outputs a clear and structured device status assessment result (e.g., intact, light wear, severe damage), along with a description of the specific problems detected.

[0228] For example: "Appearance: Minor scratches; Functionality: Normal; Overall assessment: Fully functional" or "Appearance: Screen cracked; Functionality: Cannot power on; Overall assessment: Severely damaged." This result will be used to decide whether to disinfect or report for repair.

[0229] Example 13

[0230] This embodiment is a detailed expansion of step S6 (completing the equipment return process) in embodiment one, and its steps include:

[0231] S6.1: Determine the integrity of the equipment based on the equipment status assessment results. If the equipment is in good condition, activate the blue light disinfection device to disinfect the equipment.

[0232] The system first checks the equipment status assessment results generated in the previous step. Only if the result is "intact" or "slightly worn" will the normal return process begin. If the result is "severely damaged," the system will immediately pause the process, lock the storage compartment, and send an alarm message to the administrator through the backend management platform, awaiting manual intervention. For intact equipment, the system will power on the blue / ultraviolet (UVC) LED disinfection device inside the storage compartment to begin the disinfection process.

[0233] S6.2: Set the blue light irradiation duration using a time controller to ensure that the disinfection effect meets hygiene and safety standards.

[0234] The irradiation time of the disinfection device can be preset, for example, set to 5 to 10 minutes depending on the size and material of the equipment. The main control system of the locker has a built-in timer that precisely controls the on and off times of the disinfection lamps. This setting is designed to ensure effective killing of common bacteria and viruses attached to the surface of the equipment, meeting the corresponding hygiene and safety standards.

[0235] S6.3: After disinfection is completed, the equipment status, disinfection time, and return record are uploaded to the backend management system through the data synchronization interface to complete the equipment return process.

[0236] After the set disinfection time is completed, the system will automatically turn off the disinfection lamps. Then, it will package all relevant data for this return, including the user ID, return time, device ID, final device status assessment result (e.g., "intact, disinfected"), disinfection time, etc., and upload it to the central management database via the network interface. Once the database is successfully updated, the entire return process is complete, and the device is marked as "in stock, disinfected, available" in the system, ready for the next user to borrow.

[0237] Example 14

[0238] This embodiment is a functional extension of the method described in Embodiment 1, and may further include...

[0239] AI-powered digital human interaction functionality. Its steps include:

[0240] Step A: When the panoramic camera detects the user approaching, the AI ​​digital human program is launched and switched to the welcome screen, playing the digital human's guiding voice.

[0241] In practice, the locker's touchscreen display can show saver information when in standby mode. When the system detects a user approaching, the screen immediately switches to a dynamic AI digital human interface. The digital human will make a welcoming gesture and play a preset welcome message through the built-in speaker, such as "Hello, welcome to the smart teaching equipment locker. How can I help you?"

[0242] Step B: Receive user voice input, convert the voice into text instructions using speech recognition technology, and combine natural language processing algorithms to understand the user's intent.

[0243] Here, users can interact directly with the digital human via voice, for example, saying, "I'd like to borrow a Raspberry Pi." The locker's microphone array captures the user's voice and performs noise reduction processing. The Automatic Speech Recognition (ASR) engine converts the speech signal into text in real time. Subsequently, the Natural Language Processing (NLU) module analyzes the converted text, extracting the user's core intent (borrowing the device) and key entity (Raspberry Pi).

[0244] Step C: Match the corresponding response content based on the user's intent, generate a digital human's response voice through speech synthesis technology and play it to achieve intelligent interactive dialogue.

[0245] In practice, the system can match user intent with data from a knowledge base or business logic database. If a user wants to borrow a device, the system will trigger a borrowing process; if the user is asking about usage, the system will search for relevant help documentation. After matching the response (text format), the system uses text-to-speech (TTS) technology to convert it into natural and fluent speech, which is then spoken by the AI ​​digital human. Simultaneously, relevant text and image information may be displayed on the screen, thus achieving a friendly and efficient intelligent dialogue experience.

[0246] Example 15

[0247] This embodiment is a functional extension of the method described in Embodiment 1, and may further include remote management functionality. The steps include:

[0248] Step A: Establish a network connection through a gigabit switch to enable data communication between multiple lockers and the back-end management platform.

[0249] Here, each smart locker is connected to the campus LAN via an Ethernet interface or Wi-Fi module. Multiple lockers can be aggregated through gigabit switches on floors or within buildings, ultimately connecting to a back-end management server deployed in the data center. This stable network connection is the foundation for enabling real-time data synchronization and remote control.

[0250] Step B: The back-end management platform uses WebSocket technology to monitor the borrowing and returning status, inventory, and usage statistics of each locker in real time, generating a visual data dashboard.

[0251] In practice, administrators can access the web interface of the backend management platform through a browser. The platform establishes a long-lived WebSocket connection with each locker, allowing the lockers to proactively push real-time status changes (such as borrowing, returning, and alarms) to the backend. The backend processes this real-time and historical data, presenting it in the form of charts and dashboards on a large, visualized data screen, allowing administrators to clearly understand the inventory, turnover frequency, and health status of all devices.

[0252] Step C: When a user clicks the help button on the locker, the administrator accesses the corresponding locker's camera through the remote communication system to enable remote video calls and screen sharing, providing real-time technical support.

[0253] In practice, if a user encounters a problem, they can click the "One-Click Help" button on the touchscreen. This request will be immediately pushed to the backend management platform and a notification will be sent to the online administrator. Once the administrator receives the request, they can establish an audio-visual call with the locker through the platform. The administrator can not only see the situation on-site through the camera and hear the user's description, but also share their operating instructions or help documents to the locker's screen, providing remote, real-time "step-by-step" guidance to the user, thus providing efficient problem-solving as if standing beside the user.

[0254] The user interaction process of this embodiment will be further described below with reference to the interactive interface diagram.

[0255] I. Locker Information Configuration:

[0256] Before using the locker's functions, you need to configure the locker information in its supporting backend platform. This mainly includes: locker appearance, sub-locker types, device list, and locker information. Only after the above information is configured can it be bound to the actual locker to complete information interaction and maintain consistency between data and physical lockers.

[0257] like Figure 3a The image shown is an interactive interface for configuring the appearance of the locker according to an embodiment of this disclosure. Based on this interactive interface, a locker logo image, a digital avatar, an advertising image, and a background image of the main control program can be set. After the main control program starts, it obtains the appearance information by calling a RESTful API and renders it on the program interface.

[0258] like Figure 3bAs shown, after setting up the locker information, you can add sub-lockers. When adding a sub-locker, you need to bind its type. Then, as... Figure 3c As shown, when adding locker information, you can add sub-lockers to the locker.

[0259] like Figures 4a-4d The image shows a schematic diagram of the interface interaction for the borrowing and returning of experimental equipment.

[0260] Before students officially use the lockers, the administrator must prepare the lockers by placing the actual equipment inside and saving the information. Only after the administrator has completed the preparation process can students borrow the equipment from the lockers.

[0261] like Figure 4a As shown, you will be taken to the teacher login page; swipe your card to complete the user login process, as shown below. Figure 4b As shown. Open all cabinet doors. After logging in with their card, teachers can click the "Open All Cabinet Doors" button on the management page. At this time, the electromagnetic doors of all storage compartments in the locker will open, as shown. Figure 4c As shown. On the large screen of the storage cabinet, teacher users click on each storage compartment in sequence and select the device to be stored, such as... Figure 4d As shown. Place the actual device into the storage compartment of the storage cabinet, ensuring that the device placed matches the one selected on the screen earlier. Once all devices are in the storage cabinet, close the electromagnetic door of the storage compartment. Click the "Save to Platform" button on the large screen of the storage cabinet to save the storage cabinet information.

[0262] In implementation, the lockers integrate AI technology to handle facial recognition, equipment checks, and digital human functionality. When a user approaches a locker, a panoramic camera captures their image, AI recognizes the approaching person, and the main control program displays an advertisement page followed by an AI digital human guidance page. Figure 5 As shown, a digital human dialogue page is enabled, supporting user voice input. After AI recognition, the query results are played back via voice. Users input voice commands on the digital human page, and the storage cabinet announces the results. When users return the device, it must be placed in a designated location for inspection. A close-up camera captures the current image of the device, and with the help of AI recognition, students are not allowed to return devices with damage, screen breakage, or other issues.

[0263] Specifically, after a user selects to borrow a device and confirms their identity, the screen displays the names of currently available devices for the user to choose from. After selecting a device, the user proceeds to a confirmation page. Once the user confirms the borrowing, the cabinet door opens, the user retrieves the device, and then closes the storage compartment door, completing the device borrowing process. Similarly, after a user selects to return a device and confirms their identity, the screen displays the device currently being borrowed and to be returned. After the user confirms the return, the device is placed in a designated area for testing. Once the test is passed, the device is placed in the open storage compartment door and closed, completing the device return process.

[0264] In specific implementation, such as Figure 6 As shown, the backend platform supports online viewing of storage cabinet information, displaying relevant statistics on the storage cabinets and equipment in a visually appealing data dashboard format. WebSocket technology ensures the page data is updated in real-time.

[0265] like Figure 7 As shown, the locker remote call mini-program supports viewing device status. When a student clicks the "Help" button on a locker, the corresponding locker's camera can be activated for real-time remote monitoring and communication. Users can also view the locker's touchscreen feed. To effectively utilize resources, the remote call function only requests a connection after the user clicks the "Help" button on the locker, enabling remote communication and screen sharing.

[0266] like Figure 8 As shown in the embodiments of this disclosure, a computer device is also provided, including:

[0267] The processor 81 and the memory 82; the memory 82 stores machine-readable instructions executable by the processor 81, and the processor 81 is used to execute the machine-readable instructions stored in the memory 82. When the machine-readable instructions are executed by the processor 81, the processor 81 performs the process of the teaching equipment management method described in any of the above embodiments.

[0268] The aforementioned memory 82 includes a main memory 821 and an external memory 822. The main memory 821, also known as internal memory, is used to temporarily store the computational data in the processor 81, as well as the data exchanged with external memory 822 such as a hard disk. The processor 81 exchanges data with the external memory 822 through the main memory 821.

[0269] This disclosure also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the steps of the teaching equipment management method described in the above-described method embodiments. The storage medium can be a volatile or non-volatile computer-readable storage medium.

[0270] In addition, this disclosure also provides a computer program product, on which a computer program is stored. When the computer program is run by a processor, it executes the steps of the teaching equipment management method provided in any of the above embodiments of this disclosure. For details, please refer to the above method embodiments, which will not be repeated here.

[0271] The aforementioned computer program product can be implemented through hardware, software, or a combination thereof. In one optional embodiment, the computer program product is specifically embodied in a computer storage medium, which can be a volatile or non-volatile computer-readable storage medium. In another optional embodiment, the computer program product is specifically embodied in a software product, such as a software development kit (SDK), etc.

[0272] In one implementation, the above-mentioned teaching equipment management method can be implemented through software modules, such as including:

[0273] The authentication module is used to obtain user authentication information, perform identity authentication through the target method, and generate user identity confirmation data.

[0274] The recommendation list generation module is used to retrieve the user's historical borrowing records and learning demand data based on the user identity confirmation data, perform artificial intelligence (AI) algorithm analysis, and generate a personalized device recommendation list.

[0275] The available equipment list generation module is used to perform matching calculations based on the personalized equipment recommendation list and the equipment status information in the locker to determine the list of available equipment.

[0276] The borrowing module is used to control the execution: based on the list of available borrowable equipment, the electromagnetic door control system performs the door opening operation to complete the equipment borrowing process.

[0277] The return module is used to receive device return requests, capture device images through a close-up camera and perform integrity detection using AI recognition technology, and generate device status assessment results; based on the device status assessment results, it performs blue light disinfection treatment and updates the background device management data to complete the device return process.

[0278] For details on the implementation of each of the above modules, please refer to the description of the aforementioned method embodiments, which will not be elaborated here.

[0279] The above description is merely a specific embodiment of this disclosure, but the scope of protection of this disclosure is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this disclosure should be included within the scope of protection of this disclosure. Therefore, the scope of protection of this disclosure should be determined by the scope of the claims.

[0280] The above description is merely a preferred embodiment of this disclosure and is not intended to limit this disclosure. Various modifications and variations can be made to this disclosure by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this disclosure should be included within the protection scope of this disclosure.

[0281] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the devices and apparatuses described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. In the several embodiments provided in this disclosure, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Another point is that the displayed or discussed mutual coupling or direct coupling or communication connection may be through some communication interfaces; the indirect coupling or communication connection of devices or units may be electrical, mechanical, or other forms.

[0282] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0283] In addition, the functional units in the various embodiments of this disclosure can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0284] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a processor-executable, non-volatile, computer-readable storage medium. Based on this understanding, the technical solution of this disclosure, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this disclosure. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0285] Finally, it should be noted that the above-described embodiments are merely specific implementations of this disclosure, used to illustrate the technical solutions of this disclosure, and not to limit it. The protection scope of this disclosure is not limited thereto. Although this disclosure has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features, within the scope of the technology disclosed in this disclosure. Such modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this disclosure, and should all be covered within the protection scope of this disclosure. Therefore, the protection scope of this disclosure should be determined by the protection scope of the claims.

Claims

1. A method for managing teaching equipment, characterized in that, include: Obtain user authentication information, perform identity authentication through the target method, and generate user identity confirmation data; Based on the user identity verification data, the user's historical borrowing records and learning needs data are retrieved for artificial intelligence (AI) algorithm analysis to generate a personalized device recommendation list. Based on the personalized device recommendation list, a matching calculation is performed in conjunction with the device status information in the locker to determine the list of available devices. Based on the list of available equipment, the equipment borrowing process is completed by opening the door through the electromagnetic door control system. Upon receiving a device return request, the system captures images of the device using a close-up camera and combines AI recognition technology to perform integrity detection, generating a device status assessment result. Based on the equipment status assessment results, perform blue light disinfection and update the background equipment management data to complete the equipment return process.

2. The method according to claim 1, characterized in that, The AI ​​algorithm includes a multi-layer neural network recommendation model: it retrieves user historical borrowing records and learning demand data for AI algorithm analysis to generate a personalized device recommendation list, including: Based on the user identity verification data and historical borrowing records, a user-device interaction matrix is ​​constructed and a temporal feature vector is extracted to generate user behavior sequence data. The user behavior sequence data is input into a Long Short-Term Memory (LSTM) network model for temporal pattern mining to extract the time-dependent features of user learning behavior. Combining the aforementioned time-dependent features and user learning performance data, an attention mechanism is used to weight and fuse multi-dimensional features, and a graph convolutional neural network (GCN) is used to model the user-device-course ternary relationship to generate a fused feature representation vector. Based on the fused feature representation vector, a variational autoencoder (VAE) is used to learn the distribution of potential user preferences, and a generative adversarial network (GAN) is combined to enhance data diversity, outputting a device recommendation probability distribution. Based on the device recommendation probability distribution, a multi-armed slot machine algorithm combined with reinforcement learning is used to optimize the recommendation strategy and generate the personalized device recommendation list.

3. The method according to claim 2, characterized in that, The Long Short-Term Memory (LSTM) network model includes an adaptive forgetting gate mechanism: The extraction of time-dependent features of user learning behavior includes: The importance weights of historical behaviors are calculated using the user behavior sequence data and the current timestamp information through a time decay function to generate a time decay weight matrix. Based on the time decay weight matrix and user behavior characteristics, an adaptive forgetting gate is constructed to dynamically adjust the retention level of historical information and output the filtered memory state. The filtered memory states are combined with the current learning context information, and the cell states are updated through a gated recurrent unit to generate the time-dependent features.

4. The method according to claim 2, characterized in that, The attention mechanism includes multi-head self-attention computation: The method employs an attention mechanism to weightedly fuse multi-dimensional features and models the user-device-course ternary relationship using a graph convolutional neural network (GCN) to generate a fused feature representation vector, including: The time-dependent features and user profile data are mapped to query vector Q, key vector K, and value vector V, respectively, and an attention calculation matrix is ​​constructed. Based on the attention calculation matrix, the weight distribution of multiple attention heads is calculated in parallel and a scaled dot product attention mechanism is used to generate multi-head attention output; The multi-head attention output is subjected to linear transformation and residual connection, and then processed by layer normalization to generate the fused feature representation vector.

5. The method according to claim 2, characterized in that, The variational autoencoder (VAE) includes a Bayesian inference mechanism: The method employs a variational autoencoder (VAE) to learn the distribution of latent user preferences and combines it with a generative adversarial network (GAN) to enhance data diversity, outputting a device recommendation probability distribution, including: The fused feature representation vector is mapped to the latent space through an encoder network, and latent variables are sampled using a reparameterization technique to generate the mean and variance of the user's latent preference distribution. Based on the mean and variance, the Monte Carlo sampling method is used to sample from the latent distribution, and combined with the KL divergence regularization constraint, a latent preference code is generated. The potential preferences are encoded, user preferences are reconstructed through a decoder network, and the objective function is optimized using the lower bound of evidence ELBO to output the device recommendation probability distribution.

6. The method according to claim 2, characterized in that, The multi-armed slot machine algorithm includes the upper confidence bound UCB strategy: The process of generating the personalized device recommendation list by employing a multi-armed slot machine algorithm combined with reinforcement learning to optimize the recommendation strategy includes: Based on historical recommendation performance data and user feedback information, the average reward value and confidence interval for each device category are calculated, and the device reward estimate is generated. Based on the estimated device reward and number of explorations, the UCB value of each device is calculated using the upper confidence bound formula, and combined with Thompson sampling balance exploration and utilization, the device selection probability is determined. Based on the device selection probability, the recommendation strategy is dynamically adjusted and combined with the ε-greedy algorithm to increase random exploration, thereby optimizing the generation of the personalized device recommendation list.

7. The method according to claim 2, characterized in that, The process of generating the personalized device recommendation list further includes: User behavior data is collected from multiple lockers, and personalized model parameters are trained locally at each edge node to generate local model gradients. The local model gradients of each node are weighted and fused using a secure aggregation protocol, and user privacy is protected through a differential privacy mechanism to generate global model parameter updates. Based on the global model parameter update, the recommendation model is updated synchronously in each locker, and the local recommendation effect is optimized by combining personalized fine-tuning.

8. The method according to claim 1, characterized in that, Based on the user identity verification data, the system retrieves the user's historical borrowing records and learning needs data for AI algorithm analysis, generating a personalized device recommendation list, including: Based on the user identity verification data, extract the user's historical borrowing records, learning progress and course completion rate from the database to construct a user behavior profile; Based on the user behavior profile, collaborative filtering algorithm and deep learning model are used to analyze user learning preferences and ability levels, and device compatibility score is calculated. Based on the device compatibility score, and in conjunction with the current curriculum and teaching objectives, a recommendation algorithm is used to sort and optimize the devices, generating the personalized device recommendation list.

9. The method according to claim 1, characterized in that, Based on the personalized device recommendation list, and combined with the device status information within the locker, a matching calculation is performed to determine the list of available devices, including: Real-time acquisition of equipment type, quantity, and status information for each storage compartment in the locker, and establishment of an equipment inventory mapping table; Based on the personalized device recommendation list, an intersection operation is performed with the device inventory mapping table to filter out available devices in the inventory; Prioritize and verify the availability of the available devices in the inventory to determine the list of available devices for borrowing.

10. A teaching equipment management device, characterized in that, The device includes a processor and a memory, the memory storing machine-readable instructions executable by the processor, the processor executing the machine-readable instructions stored in the memory, and when the machine-readable instructions are executed by the processor, the processor performs the steps of the teaching equipment management method as described in any one of claims 1 to 9.

Images