An individualized information management system based on a smart campus

Abstract

This invention relates to the field of smart campus information management technology, specifically a personalized information management system based on a smart campus. The system includes: simultaneously acquiring student behavior trajectory data, academic interaction data, and environmental perception data through multimodal data acquisition terminals deployed throughout the campus. Based on the multi-type acquired data, the system constructs an initial individual state vector for each student. This vector is input into a spatiotemporal attention network, which intelligently analyzes and outputs the personalized information demand weight distribution corresponding to students at different times and locations. The system dynamically filters and adaptively sorts the original information entries in the campus information resource database based on the demand weight distribution, generating a dedicated personalized information push queue. Finally, the customized information push queue is accurately distributed to the corresponding student's mobile terminal via the campus wireless network. This system achieves fusion modeling and spatiotemporal correlation analysis of multimodal campus data, enabling precise mining of information needs tailored to changes in student scenarios.

Description

Technical Field

[0001] This invention relates to the field of smart campus information management technology, and in particular to a personalized information management system based on smart campuses. Background Technology

[0002] The construction of smart campuses generates a large amount of campus service and teaching-related information resources. Traditional campus information management systems lack a unified multimodal data collection mechanism, making it impossible to simultaneously aggregate data on student behavior trajectories, academic interactions, and campus environment perception. Various types of campus-related data are stored independently on different business platforms, lacking integration and modeling channels, making it difficult to characterize students' overall well-being.

[0003] Existing campus information services mostly adopt a unified content push model, classifying information service types solely based on fixed static attributes. This fails to leverage multi-source fusion data to construct standardized individual student state vectors. Furthermore, the lack of spatiotemporal attention network analysis methods prevents the quantification of students' dynamic information needs from both temporal and spatial dimensions, hindering the generation of demand weight distribution results by time period and region.

[0004] Conventional information distribution mechanisms lack the processing logic to dynamically filter and reorder information resources based on individual needs, making it difficult to generate personalized information push queues tailored to the characteristics of different students. Information transmission and distribution methods lack targeted adaptation capabilities, failing to leverage the campus network for precise delivery of customized content. Traditional management models suffer from low data integration, rigid demand analysis methods, and severe homogenization of information pushes, making it difficult to meet the practical application needs of smart campuses to provide differentiated, scenario-based, and refined personalized information management services for students. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing a personalized information management system based on smart campuses.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a personalized information management system based on a smart campus, comprising: The multimodal acquisition module acquires students' behavioral trajectory data, academic interaction data, and environmental perception data through multimodal acquisition terminals deployed in various areas of the campus. The individual state modeling module constructs an initial individual state vector for each student based on the behavioral trajectory data, the academic interaction data, and the environmental perception data. The demand analysis module inputs the initial individual state vector into a spatiotemporal attention network and outputs the personalized information demand weight distribution of each student at different times and locations. The information filtering and push module dynamically filters and sorts the original information entries in a campus information resource database according to the weight distribution of personalized information demand, and generates a personalized information push queue for each student. The information distribution module distributes the personalized information push queue to the corresponding students' mobile terminals through the campus wireless network.

[0007] As a further aspect of the present invention, the acquisition of student behavior trajectory data, academic interaction data, and environmental perception data through multimodal acquisition terminals deployed in various areas of the campus specifically includes: Bluetooth beacon detectors are installed at the entrance of each classroom, library, canteen, and dormitory building on campus. These detectors capture the Bluetooth signal strength sequence broadcast by students' mobile terminals. Each Bluetooth beacon detector's unique identifier is packaged with the Bluetooth signal strength sequence to generate a beacon detection record with a timestamp; A data scraping agent is deployed on the campus's online learning platform server. The data scraping agent reads each student's course viewing progress, assignment submission time, forum post content, and quiz answer records at fixed time intervals. The course viewing progress, the homework submission time, the forum post content, and the quiz answer records are stored in a structured manner to form a set of academic interaction records for each student; A capacitive pressure sensor array is embedded on the surface of the desks in each classroom to detect the occupancy status of seats and the frequency of micro-movements in the students' sitting posture. Distributed temperature and humidity sensors and carbon dioxide concentration sensors are deployed in the corridors of the teaching building and the study area of ​​the library. The temperature and humidity sensors and the carbon dioxide concentration sensors output environmental parameter sequences at a sampling rate of once per second. The seat occupancy status, the frequency of micro-movements in sitting posture, and the environmental parameter sequence are time-aligned to form an environmental perception record set for each student at each moment. The beacon detection record, academic interaction record set, and environmental perception record set corresponding to the same student at the same time are fused to generate the original multidimensional collection sample of that student at that time.

[0008] As a further aspect of the present invention, the capacitive pressure sensing array identifies the occupancy status of each seat using an anonymous hash value, which is calculated and generated by combining the seat number and the timestamp.

[0009] As a further aspect of the present invention, an initial individual state vector for each student is constructed based on the behavioral trajectory data, the academic interaction data, and the environmental perception data, specifically including: The Bluetooth signal strength sequence over time is extracted from the original multidimensional acquisition samples. Peak detection is performed on the curve, and the time when each peak appears is taken as the time when the student enters the coverage area of ​​the corresponding Bluetooth beacon detector. The time difference between two consecutive moments when the student enters the coverage area of ​​different Bluetooth beacon detectors is taken as the time the student spends moving between the two areas. The student's moving speed is calculated based on the time spent moving and the preset physical distance between the two areas. By concatenating the movement speed and the corresponding area identifier in chronological order, a behavioral trajectory encoding sequence for the student is generated. Extract the student's total course viewing time, number of incomplete assignments, number of words posted on forums, and accuracy rate in answering quizzes from the academic interaction record set in the most recent 24 hours before the current moment. Concatenate the total viewing time, number of incomplete assignments, number of words posted on forums, and accuracy rate into a four-dimensional academic feature vector. The seat occupancy flag, the average frequency of sitting micro-movements over the past ten minutes, the temperature and humidity values, and the carbon dioxide concentration values ​​at the current moment are extracted from the environmental perception record set. The seat occupancy flag, the average frequency of sitting micro-movements, the temperature and humidity values, and the carbon dioxide concentration values ​​are then concatenated into a four-dimensional environmental feature vector. The behavior trajectory encoding sequence is compressed into a fixed-length trajectory embedding vector; The trajectory embedding vector, the four-dimensional academic feature vector, and the four-dimensional environmental feature vector are concatenated end-to-end along the vector dimension to generate the initial individual state vector of each student at the current moment.

[0010] As a further aspect of the present invention, the behavior trajectory encoding sequence is compressed into a trajectory embedding vector of fixed length, specifically including: encoding the behavior trajectory encoding sequence using a pre-trained unidirectional long short-term memory network, and taking the output of the last hidden layer as the trajectory embedding vector.

[0011] As a further aspect of the present invention, the initial individual state vector is input into a spatiotemporal attention network, which outputs the personalized information demand weight distribution for each student at different times and locations, specifically including: Obtain the initial individual state vector for each of the student's past seven consecutive natural days, and divide the initial individual state vector of each day into ninety-six time slot state vectors according to a time slot of fifteen minutes. Stack the state vectors of ninety-six time slots at the same time slot location on different dates to generate a historical state tensor of size seven times ninety-six times the dimension of the state vector. The historical state tensor is input into the temporal attention module in the spatiotemporal attention network. The temporal attention module performs self-attention calculation on the state similarity of each time slot position between different dates and outputs the diurnal cycle importance score of each time slot. The historical state tensor is input into the spatial attention module in the spatiotemporal attention network. The spatial attention module performs attention calculation on the state transition frequency between different Bluetooth beacon detector coverage areas and outputs the spatial importance score of each campus area. The behavior trajectory encoding sequence in the initial individual state vector at the current moment is differentially calculated with the historical trajectory encoding corresponding to the current time slot in the historical state tensor to obtain the trajectory change increment vector. The trajectory change increment vector is passed through a fully connected mapping layer and then multiplied element-wise with the daytime periodic importance score and the spatial importance score to generate a personalized information demand weight distribution for each student at the current moment for different campus areas and different time offsets.

[0012] As a further aspect of the present invention, the time attention module employs a multi-head self-attention mechanism to calculate the importance score of the daytime cycle.

[0013] As a further aspect of the present invention, the original information entries in a campus information resource database are dynamically filtered and sorted according to the personalized information demand weight distribution to generate a personalized information push queue for each student, specifically including: Read all raw information entries to be pushed from the campus information resource database. Each raw information entry carries a publishing area tag, a publishing time tag, and a content type tag. The publishing area tag of each original information entry is matched with the spatial importance score in the personalized information demand weight distribution to calculate the area matching score of the original information entry. The publication time period tag of each original information entry is compared with the time slot position to which the current time belongs. If the importance score of the publication time period tag and the current time period exceeds the preset matching threshold, it is marked as a time period match; otherwise, it is marked as a time period mismatch. The original information entries that are marked as matching the time period are sorted in the first round according to the region matching score from high to low, and the original information entries that are marked as not matching the time period are temporarily stored in a delayed push buffer; For multiple original information entries with the same region matching score in the first round of sorting, the content type label of each original information entry is further obtained, and the content type label is matched with the content interest tendency extracted from the trajectory change increment vector to calculate the content matching score. Based on the content matching score, a second round of sorting is performed on multiple original information entries with the same regional matching score to form a personalized information push queue for each student. Check the waiting time of the original information entries in the delayed push buffer. When the waiting time exceeds the preset maximum delay time, remove the corresponding original information entries from the delayed push buffer and insert them into the end of the personalized information push queue according to the regional matching score.

[0014] As a further aspect of the present invention, the personalized information push queue is distributed to the mobile terminals of the corresponding students via the campus wireless network, specifically including: Obtain the Media Access Control Address of the wireless access point currently accessed by the student's mobile terminal, and determine the current campus area where the mobile terminal is located based on the Media Access Control Address; Read the first-ranked information entry from the personalized information push queue and extract the publishing area tag of the information entry; The publishing area tag of the information entry is compared with the current campus area. If they match, the information entry is marked as an instant push entry; if they do not match, the information entry is marked as a delayed push entry. For the instant push entry, the main content of the instant push entry is compressed into a data packet, the push service registration identifier of the mobile terminal is written into the header of the data packet, and the data packet is sent through the campus wireless network in the manner of User Datagram Protocol. For the delayed push entry, the reference identifier of the information entry and the preset trigger area label are stored in the local push cache table of the mobile terminal, and the local push cache table is maintained in the form of key-value pairs; When the mobile terminal detects a change in the access point, it reads the new campus area corresponding to the new access point and compares the new campus area with the trigger area label in the local push cache table. If they match, the corresponding delayed push entry is marked as a wake-up entry. The complete content of the entry to be woken up is retrieved from the campus information resource database, a wake-up push data packet is generated, and it is sent to the mobile terminal through a new wireless access point.

[0015] As a further aspect of the present invention, the main content of the instant push item is compressed into a data packet, specifically including: semantic compression of the text content of the information item using an encoder based on a deep context model, volume compression of the image or video attachments in the information item using a lossy compression algorithm, and encapsulating the compressed text semantic vector and the media file stream in the same data packet.

[0016] Compared with the prior art, the advantages and positive effects of the present invention are as follows: This system aggregates behavioral trajectory data, academic interaction data, and environmental perception data acquired from multimodal data collection terminals across the campus, uniformly constructing an initial individual state vector for each student. This enables the fusion, standardization, and structured modeling of multi-dimensional campus data. The initial individual state vector is then introduced into a spatiotemporal attention network for feature analysis across both time and space dimensions, quantifying and generating personalized information demand weight distributions for different time periods and locations. Moving away from the single model of classifying information demands based on fixed static attributes, this system fully leverages the inherent correlation features of multi-source data to quantitatively decompose and finely represent students' information demands, providing a standardized weighting basis for information resource selection and ranking.

[0017] Based on the weight distribution rules of personalized information demand, the original information entries in the campus information resource database are dynamically filtered and hierarchically sorted to generate personalized information push queues adapted to different student characteristics. This breaks away from the inherent model of batch pushing uniform content, achieving differentiated selection and orderly arrangement of information resources according to individual needs, thus realizing precise matching between information content and user needs. The customized information push queues are directly distributed to the corresponding student mobile terminals via the campus wireless network, establishing a complete business chain from multimodal data collection, individual status modeling, spatiotemporal demand analysis, intelligent information filtering to targeted wireless distribution. This improves the dynamic adaptability and personalized service capabilities of the smart campus information management system, enhances the utilization efficiency and distribution accuracy of campus information resources, and supports the development of campus information management towards differentiation, refinement, and intelligence. Attached Figure Description

[0018] Figure 1 This is a sequence diagram of a personalized information management system based on a smart campus as described in this invention. Figure 2 A flowchart illustrating the workflow for multimodal data acquisition and fusion; Figure 3 A flowchart illustrating the workflow for generating demand weight distributions for spatiotemporal attention networks. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0020] See Figure 1 This invention provides a personalized information management system based on a smart campus, specifically including: The multimodal acquisition module is responsible for acquiring students' behavioral trajectory data, academic interaction data, and environmental perception data through multimodal acquisition terminals deployed throughout the campus. The individual state modeling module constructs an initial individual state vector for each student based on the acquired behavioral trajectory data, academic interaction data, and environmental perception data. The demand analysis module inputs the initial individual state vector into a spatiotemporal attention network, outputting the personalized information demand weight distribution for each student at different times and locations. The information filtering and push module dynamically filters and sorts the original information entries in a campus information resource database based on the personalized information demand weight distribution, generating a personalized information push queue for each student. The information distribution module distributes the personalized information push queue to the corresponding students' mobile terminals via the campus wireless network.

[0021] In one embodiment of the present invention, see [reference] Figure 2 Bluetooth beacon detectors are installed at the entrances of every classroom, library, cafeteria, and dormitory building on campus. These detectors capture the Bluetooth signal strength sequences broadcast by students' mobile devices. Each Bluetooth beacon detector's unique identifier is packaged with its Bluetooth signal strength sequence to generate a timestamped beacon detection record. A data scraping agent is deployed on the campus's online learning platform server. This agent reads each student's course viewing progress, assignment submission time, forum posts, and quiz answers at fixed time intervals. This data is then structured and stored to form a set of academic interaction records for each student. Capacitive pressure sensor arrays are embedded on the surfaces of desks in each classroom to detect seat occupancy and the frequency of students' posture changes. Distributed temperature and humidity sensors and carbon dioxide concentration sensors are deployed in the corridors of teaching buildings and the study areas of the library. These sensors output environmental parameter sequences at a sampling rate of once per second. The seat occupancy status, posture change frequency, and environmental parameter sequences are time-aligned to form a set of environmental perception records for each student at each moment. The beacon detection records, academic interaction records, and environmental perception records for the same student at the same time are fused to generate the student's original multidimensional data collection sample at that moment. The capacitive pressure sensor array identifies the occupancy status of each seat using an anonymous hash value, which is calculated by combining the seat number and timestamp.

[0022] In practical implementation, the multimodal acquisition module of the personalized information management system based on the smart campus acquires student behavior trajectory data, academic interaction data, and environmental perception data through multimodal acquisition terminals deployed in various areas of the campus. The specific implementation method is as follows: Bluetooth beacon detectors are installed at the entrances of each classroom, library, cafeteria, and dormitory building. Each Bluetooth beacon detector scans for Bluetooth signals in the surrounding environment at a fixed frequency. When a student's mobile device enters the coverage area of ​​the Bluetooth beacon detector, the detector captures the Bluetooth signal strength sequence broadcast by the mobile device. Each Bluetooth signal strength sequence contains multiple signal strength values ​​arranged in chronological order. The unique identifier of each Bluetooth beacon detector is packaged with the captured Bluetooth signal strength sequence to generate a timestamped beacon detection record. The timestamp precisely records the moment the beacon detection record was generated. In some embodiments, a data scraping agent is deployed on the online learning platform server of the campus. This agent reads each student's course viewing progress, assignment submission time, forum post content, and quiz answer records at fixed time intervals. Course viewing progress is represented as the ratio of the student's completed video viewing time to the total course duration. Assignment submission time is represented as the timestamp recorded by the server when the student submits the assignment. Forum post content is stored in text format, and quiz answer records include the correct answer flag for each question and the student's answer. The course viewing progress, assignment submission time, forum post content, and quiz answer records are structured and stored to form a set of academic interaction records for each student. This set is organized chronologically into multiple records, each associated with a student identifier and a timestamp.

[0023] In practice, capacitive pressure sensor arrays are embedded in the surface of desks in each classroom. These arrays consist of multiple capacitive pressure sensor units arranged in a grid pattern on the desk surface. Each sensor unit outputs the corresponding change in capacitance when it detects a pressure change. The capacitive pressure sensor arrays detect seat occupancy and the frequency of students' posture changes. Seat occupancy is represented by binary values, where 1 indicates the seat is occupied and 0 indicates it is vacant. Posture change frequency represents the number of pressure fluctuations generated by the student's body on the seat per unit time. Distributed temperature and humidity sensors and carbon dioxide concentration sensors are deployed in the corridors of the teaching building and the study areas of the library. These sensors continuously output environmental parameter sequences at a sampling rate of once per second. These sequences include time-aligned temperature, relative humidity, and carbon dioxide concentration values. The seat occupancy, posture change frequency, and environmental parameter sequences are time-aligned. Data collected at the same time are combined into an environmental perception record, forming a set of environmental perception records for each student at each moment. Optionally, the capacitive pressure sensor array identifies the occupancy status of each seat using an anonymous hash value. The anonymous hash value is calculated by combining the seat number and the timestamp, and the calculation method for the anonymous hash value uses the following hash function:

[0024] in: Represents an anonymous hash value. This indicates the 256-bit version of the Secure Hash Algorithm. Indicates the seat number. Represents a timestamp. This represents the string concatenation operator.

[0025] In specific implementation, the beacon detection records, academic interaction record sets, and environmental awareness record sets corresponding to the same student at the same time are fused to generate the student's original multidimensional collection sample at that time. The fusion operation matches the student's identity corresponding to the beacon detection record based on the student's mobile terminal's unique device identifier. The academic interaction record set is selected from the academic interaction record set with the timestamp closest to the current time, and the environmental awareness record set is selected from the environmental awareness record set with the timestamp completely consistent with the current time. All fields from the three record sets are then merged into a complete record. In some embodiments, when a beacon detection record at a certain time is missing, linear interpolation is performed using the Bluetooth signal strength sequence from the beacon detection record at the previous time to complete the record. When the academic interaction record set has no corresponding record at the current time, null placeholders are used to fill the relevant academic interaction fields. It is understandable that the beacon detection record contains the unique identifier of the Bluetooth beacon detector and the Bluetooth signal strength sequence; the academic interaction record set contains course viewing progress, assignment submission time, forum post content, and quiz answer records; and the environmental perception record set contains seat occupancy status, frequency of sitting posture micro-movements, temperature value, relative humidity value, and carbon dioxide concentration value. These data fields together constitute the various dimensions of the original multidimensional acquisition sample. It is also understandable that the original multidimensional acquisition sample, as the output of the multimodal acquisition module, is transmitted to the individual state modeling module to construct the initial individual state vector for each student.

[0026] In one embodiment of the present invention, the change curve of Bluetooth signal strength sequence over time is extracted from the original multidimensional acquisition samples. Peak detection is performed on the change curve, and the moment when each peak occurs is taken as the moment when the student enters the coverage area of ​​the corresponding Bluetooth beacon detector. The time difference between two adjacent moments of entering different Bluetooth beacon detector coverage areas is taken as the student's movement time between the two areas. The student's movement speed is calculated based on the movement time and the preset physical distance between the two areas. The movement speed and the corresponding area identifier are concatenated in chronological order to generate the student's behavior trajectory encoding sequence. The total duration of the student's course viewing, the number of unfinished assignments, the number of words posted on forums, and the accuracy rate of quizzes in the most recent 24 hours before the current moment are extracted from the academic interaction record set. The total duration, the number of unfinished assignments, the number of words posted on forums, and the accuracy rate of quizzes are concatenated into a four-dimensional academic feature vector. The following data are extracted from the environmental perception record set: the current seat occupancy marker, the average frequency of seat micro-movements over the past ten minutes, the current temperature and humidity values, and the current carbon dioxide concentration value. These data are then concatenated into a four-dimensional environmental feature vector. The behavioral trajectory encoding sequence is compressed into a fixed-length trajectory embedding vector. The compression process uses a pre-trained unidirectional long short-term memory network to encode the behavioral trajectory encoding sequence, and the output of the last hidden layer is used as the trajectory embedding vector. The trajectory embedding vector, the four-dimensional academic feature vector, and the four-dimensional environmental feature vector are concatenated end-to-end along the vector dimension to generate the initial individual state vector for each student at the current moment.

[0027] In practical implementation, the individual state modeling module constructs an initial individual state vector for each student based on behavioral trajectory data, academic interaction data, and environmental perception data. The specific implementation method is as follows: The Bluetooth signal strength sequence over time is extracted from the original multi-dimensional collected samples. The curve is constructed with time as the horizontal axis and Bluetooth signal strength as the vertical axis. Peak detection is performed on the curve, using a sliding window method to locate local maxima. The moment each peak occurs is taken as the moment the student enters the coverage area of ​​the corresponding Bluetooth beacon detector. The time difference between two adjacent moments of entering different Bluetooth beacon detector coverage areas is taken as the student's movement time between the two areas. The student's movement speed is calculated based on the movement time and the preset physical distance between the two areas. The formula for calculating the movement speed is:

[0028] in: Indicates the student's movement speed. This indicates the preset physical distance between two areas. This represents the time it takes for a student to move between two areas. By concatenating the movement speed and the corresponding area identifier in chronological order, a sequence of student behavior trajectory codes is generated. Each element in the behavior trajectory code sequence is a tuple, where the first component of the tuple is the area identifier, and the second component is the student's movement speed when moving between the current area and the next area.

[0029] In practice, the following data are extracted from the academic interaction record set within the most recent 24 hours: total course viewing time, number of incomplete assignments, number of forum posts, and quiz accuracy rate. Total course viewing time is accumulated in minutes. The number of incomplete assignments is calculated from the assignment submission time list, indicating the number of assignments not yet submitted as of the current time. The number of forum posts is obtained by counting the characters in the posts. The quiz accuracy rate is expressed as the ratio of correct answers to the total number of correct answers. The total time, number of incomplete assignments, number of posts, and accuracy rate are concatenated into a four-dimensional academic feature vector, with the four dimensions arranged in the order of total time, number of incomplete assignments, number of posts, and accuracy rate. In some embodiments, the following are extracted from the environmental perception recording set: the current seat occupancy flag, the average frequency of seated micro-movements over the past ten minutes, the current temperature and humidity values, and the current carbon dioxide concentration. The seat occupancy flag is read from the binary value output by the capacitive pressure sensor array. The average frequency of seated micro-movements over the past ten minutes is obtained by taking the arithmetic mean of the frequency of seated micro-movements at each sampling time over the past ten minutes. The temperature and humidity values ​​include two components: temperature and relative humidity. The carbon dioxide concentration is expressed in parts per million (ppm). The seat occupancy flag, the average frequency of seated micro-movements, the temperature and humidity values, and the carbon dioxide concentration are concatenated into a four-dimensional environmental feature vector. The four dimensions of the four-dimensional environmental feature vector are arranged in the order of seat occupancy flag, average frequency of seated micro-movements, temperature, relative humidity, and carbon dioxide concentration, with the temperature and relative humidity values ​​sharing one dimension.

[0030] In the specific implementation, the behavioral trajectory encoding sequence is compressed into a fixed-length trajectory embedding vector. The compression operation uses a pre-trained unidirectional long short-term memory (LSTM) network to encode the behavioral trajectory encoding sequence. The input dimension of the LSTM network is consistent with the dimension of each element in the behavioral trajectory encoding sequence. The network contains a hidden layer with a fixed number of neurons. Each pair in the behavioral trajectory encoding sequence is input into the LSTM network in chronological order. The network updates the hidden state at each time step. After the entire behavioral trajectory encoding sequence has been input, the output of the last hidden layer is taken as the trajectory embedding vector. Optionally, the pre-trained LSTM network is trained offline using the historical behavioral trajectory encoding sequences of all students on campus, with the training objective being to minimize the trajectory reconstruction error. The trajectory embedding vector, the four-dimensional academic feature vector, and the four-dimensional environmental feature vector are concatenated end-to-end along the vector dimension to generate the initial individual state vector for each student at the current time. The total dimension of the initial individual state vector is equal to the sum of the fixed-length value of the trajectory embedding vector, the dimension of the four-dimensional academic feature vector, and the dimension of the four-dimensional environmental feature vector. It is understood that the initial individual state vector simultaneously contains the student's behavioral movement patterns on campus, academic performance status, and physical parameters of the environment. It is also understood that the initial individual state vector is transmitted to the requirements analysis module as input data for the spatiotemporal attention network. Optionally, when academic interaction records are missing in the original multidimensional sample, all dimensions of the four-dimensional academic feature vector are filled with zero values, and a flag is appended to the end of the initial individual state vector to indicate missing academic data. In some embodiments, when the length of the behavioral trajectory encoding sequence is insufficient for the minimum sequence length required by the pre-trained unidirectional long short-term memory network, the last tuple is repeatedly padded at the end of the behavioral trajectory encoding sequence until the minimum sequence length is reached.

[0031] In one embodiment of the present invention, see [reference] Figure 3The process involves obtaining the initial individual state vector for each of the student's past seven consecutive natural days, dividing each day's initial individual state vector into ninety-six time slot state vectors, each spaced 15 minutes apart. The ninety-six time slot state vectors for the same time slot location on different dates are stacked to generate a historical state tensor of size seven times ninety-six times the state vector dimension. This historical state tensor is input into the temporal attention module of the spatiotemporal attention network. The temporal attention module uses a multi-head self-attention mechanism to calculate the state similarity between different dates for each time slot location, outputting the diurnal periodic importance score for each time slot. The historical state tensor is then input into the spatial attention module of the spatiotemporal attention network. The spatial attention module calculates the attention frequency of state transitions between different Bluetooth beacon detector coverage areas, outputting the spatial importance score for each campus area. Finally, the behavioral trajectory encoding sequence in the initial individual state vector at the current moment is differentially calculated with the historical trajectory encoding corresponding to the current time slot in the historical state tensor to obtain the trajectory change increment vector. The trajectory change increment vector is passed through a fully connected mapping layer and then multiplied element-wise with the daytime periodic importance score and spatial importance score to generate the personalized information demand weight distribution for each student at the current moment for different campus areas and different time offsets.

[0032] In specific implementation, the demand analysis module inputs the initial individual state vector into a spatiotemporal attention network and outputs the personalized information demand weight distribution for each student at different times and locations. The specific implementation method is as follows: The initial individual state vector for each of the student's past seven consecutive natural days is obtained. Each day's initial individual state vector is divided into ninety-six time slot state vectors, each spaced 15 minutes apart. Each time slot state vector corresponds to the initial individual state vector at the start time of that time slot. The ninety-six time slot state vectors for the same time slot on different dates are stacked to generate a historical state tensor of size seven times ninety-six times the state vector dimension. The three dimensions of the historical state tensor represent the date index, time slot index, and state vector dimension, respectively. In some embodiments, during the generation of the historical state tensor, if a time slot on a certain date lacks a corresponding initial individual state vector, linear interpolation is used to fill the gap using the initial individual state vectors of the same time slot on adjacent dates.

[0033] In specific implementation, the historical state tensor is input into the temporal attention module of the spatiotemporal attention network. The temporal attention module performs self-attention calculation on the state similarity of each time slot across different dates, outputting the diurnal cycle importance score for each time slot. The temporal attention module employs a multi-head self-attention mechanism to calculate the diurnal cycle importance score. The self-attention calculation process treats the state vector of the same time slot under each date in the historical state tensor as a set of query, key, and value vectors, and obtains the attention weight by calculating the dot product of the query vector and the key vector. Optionally, the temporal attention module contains four attention heads. Each attention head independently calculates the attention distribution, concatenates the output results, and performs a linear transformation. The diurnal cycle importance score ranges from 0 to 1, with a higher score indicating a more stable state repetition pattern in the seven-day cycle for that time slot. In some embodiments, before calculating state similarity, the temporal attention module first maps each state vector in the historical state tensor to a lower-dimensional representation space through a learnable linear projection layer.

[0034] In implementation, the historical state tensor is input into the spatial attention module of the spatiotemporal attention network. The spatial attention module calculates the state transition frequency between different Bluetooth beacon detector coverage areas and outputs a spatial importance score for each campus area. The state transition frequency represents the number of times a student moves from one Bluetooth beacon detector coverage area to another, as counted from the historical state tensor. The attention calculation process treats each area as a query and uses the state features of all areas in the historical state tensor as keys and values, aggregating the transition relationships between areas through an attention mechanism. The spatial importance score is a vector, where each element corresponds to a campus area, and the element value represents the relative importance of that area in the student's daily behavior patterns. It can be understood that the historical state tensor input to the spatial attention module contains information about the behavioral trajectory encoding sequence, which records the student's movement order between different Bluetooth beacon detector coverage areas. It can also be understood that the spatial attention module and the temporal attention module operate in parallel within the spatiotemporal attention network, with both modules receiving the same historical state tensor as input.

[0035] In practice, the behavioral trajectory encoding sequence in the initial individual state vector at the current moment is differentially calculated with the historical trajectory encoding corresponding to the current time slot in the historical state tensor to obtain the trajectory change increment vector. The differential calculation uses vector subtraction, subtracting the compressed trajectory embedding vector from the historical trajectory encoding sequence at the current moment from the compressed vector of the historical trajectory encoding in the historical state tensor that corresponds to the same time slot and date within the past seven days. If no matching date type exists within the past seven days, the arithmetic mean of all historical trajectory encoding vectors at the same time slot within the past seven days is used. The dimension of the trajectory change increment vector is consistent with the dimension of the trajectory embedding vector, and each component of the vector represents the degree of deviation between the current behavioral pattern and the historical behavioral pattern in the corresponding dimension. After passing through a fully connected mapping layer, the trajectory change increment vector is multiplied element-wise with the diurnal cycle importance score and the spatial importance score to generate the personalized information demand weight distribution for each student at the current moment for different campus areas and different time offsets. The formula for the element-wise multiplication operation is:

[0036] in: Indicates that the student is in The campus area and the Weighting of personalized information demand at each time offset This represents the transformation function of a fully connected mapping layer. This represents the trajectory change increment vector. The output of the fully connected mapping layer represents the first... One portion, Indicates the first The importance score of the daytime cycle for each time slot. Indicates the first The spatial importance scores of each campus area are assigned. Optionally, the weight distribution of personalized information demand is organized into a two-dimensional matrix, where the rows of the matrix correspond to campus areas, and the columns correspond to different time slot positions shifted backward from the current moment. It can be understood that the output dimension of the fully connected mapping layer equals the total number of campus areas, the dimension of the diurnal periodic importance score equals the number of time slots, and the dimension of the spatial importance score equals the total number of campus areas. The element-wise multiplication operation expands the diurnal periodic importance score and the spatial importance score to the same dimension as the output of the fully connected mapping layer through a broadcast mechanism before performing a dot product.

[0037] In one embodiment of the present invention, all original information entries to be pushed are read from the campus information resource database. Each original information entry carries a publishing area tag, a publishing time period tag, and a content type tag. The publishing area tag of each original information entry is matched with the spatial importance score in the weight distribution of personalized information demand to calculate the area matching score of the original information entry. The publishing time period tag of each original information entry is compared with the time slot position to which the current time belongs. If the daytime cycle importance score of the publishing time period tag and the current time exceeds a preset matching threshold, it is marked as a time period hit; otherwise, it is marked as a time period miss. The original information entries marked as time period hits are sorted in the first round according to the area matching score from high to low, and the original information entries marked as time period misses are temporarily stored in a delayed push buffer. For multiple original information entries with the same area matching score in the first round of sorting, the content type tag of each original information entry is further obtained. The content type tag is matched with the content interest tendency extracted from the trajectory change increment vector to calculate the content matching score. Based on the content matching score, multiple original information entries with the same area matching score are sorted in the second round to form a personalized information push queue for each student. Check the waiting time of the original information entries in the delayed push buffer. When the waiting time exceeds the preset maximum delay time, remove the corresponding original information entries from the delayed push buffer and insert them into the end of the personalized information push queue according to the regional matching score.

[0038] In practical implementation, the information filtering and push module dynamically filters and sorts the original information entries in a campus information resource database according to the weight distribution of personalized information needs, generating a personalized information push queue for each student. The specific implementation method is as follows: All original information entries to be pushed are read from the campus information resource database. Each original information entry carries a publishing area tag, a publishing time period tag, and a content type tag. The publishing area tag indicates the name of the campus area associated with the original information entry; the publishing time period tag indicates the suitable time period for pushing the original information entry; and the content type tag indicates whether the original information entry belongs to the notification, activity, learning resource, or life service category. The publishing area tag of each original information entry is matched with the spatial importance score in the weight distribution of personalized information needs. The matching operation searches for the importance score of the corresponding campus area in the spatial importance score vector based on the publishing area tag, and this score is directly used as the area matching score of the original information entry. In some embodiments, when the campus area corresponding to the publishing area tag does not exist in the spatial importance score vector, the area matching score is set to zero.

[0039] In practice, the publication time period tag of each original information entry is compared with the time slot position to which the current time belongs. The comparison operation involves extracting the time period range recorded in the publication time period tag, determining whether the time slot to which the current time belongs falls within the time period range, and if so, further obtaining the daytime cycle importance score of the time slot to which the current time belongs. This daytime cycle importance score is compared with a preset matching threshold. If the daytime cycle importance score exceeds the preset matching threshold, it is marked as a time period match; otherwise, it is marked as a time period mismatch. The preset matching threshold ranges from 0 to 1, with a default value of 0.6. The original information entries marked as time period matches are sorted in the first round according to their regional matching scores from high to low. If the regional matching scores are the same, they are arranged according to the generation timestamp of the original information entries from recent to distant. The original information entries marked as time period mismatches are temporarily stored in a delayed push buffer. The delayed push buffer maintains the complete content of each original information entry and its timestamp of entering the buffer using a queue structure. Optionally, the delayed push buffer stores the original information entries that were not hit during the time period in a first-in-first-out order, and each original information entry does not participate in the generation of the push queue while waiting in the buffer.

[0040] In practice, for multiple original information entries with the same region matching score in the first round of sorting, the content type label of each original information entry is further obtained. This content type label is then matched with the content interest tendency extracted from the trajectory change increment vector. The matching operation uses a predefined content type mapping matrix to convert the content type label into a content type encoding vector. The cosine similarity between the content type encoding vector and the content interest tendency sub-vector extracted from the trajectory change increment vector is calculated, and the cosine similarity value is used as the content matching score. The content interest tendency sub-vector is extracted from the trajectory change increment vector by extracting the last four dimensions of the vector. These four dimensions are pre-trained to encode students' interest in four content types: notifications, activities, learning resources, and life services. The formula for calculating the content matching score is:

[0041] in: This indicates the content matching score. Represents the content type encoding vector. This represents the content interest sub-vector extracted from the trajectory change increment vector. This represents the dot product operation of vectors. The Euclidean norm of the content type encoding vector. This represents the Euclidean norm of the content interest tendency subvector. Based on the content matching score, multiple original information items with the same region matching score are sorted in a second round. This second round of sorting arranges them from highest to lowest content matching score, forming a personalized information push queue for each student. This personalized information push queue can be understood as an ordered list, where each element corresponds to an original information item, and the order of the list determines the order in which the information items are pushed. The first and second rounds of sorting together constitute a two-level sorting mechanism, with the region matching score as the primary sorting key and the content matching score as the secondary sorting key.

[0042] In specific implementation, the waiting time of the original information entries in the delayed push buffer is checked. The waiting time is calculated by subtracting the timestamp recorded when the original information entry entered the delayed push buffer from the current time. When the waiting time exceeds a preset maximum delay time, the corresponding original information entry is removed from the delayed push buffer and inserted into the end of the personalized information push queue according to the regional matching score. The preset maximum delay time is set to 30 minutes. The insertion operation calculates the regional matching score of the removed original information entry, scans backward from the end of the personalized information push queue, finds the first position where the regional matching score is less than the regional matching score of the removed entry, and inserts the removed entry after that position; if the regional matching scores of all entries in the queue are greater than or equal to the regional matching score of the removed entry, the removed entry is inserted at the front of the queue; if the regional matching scores of all entries in the queue are less than the regional matching score of the removed entry, the removed entry is directly appended to the end of the queue. In some embodiments, when the waiting time of multiple original information entries in the delayed push buffer exceeds the preset maximum delay time at the same time, the removal and insertion operations are performed sequentially according to the order in which the entries entered the buffer. Optionally, original information entries in the delayed push buffer whose waiting time does not exceed the preset maximum delay time are retained in the buffer and will be re-evaluated during the next check. Optionally, after the personalized information push queue is generated, the information filtering and push module transmits the personalized information push queue to the information distribution module.

[0043] In one embodiment of the present invention, the medium access control address of the wireless access point currently accessed by the student mobile terminal is obtained, and the current campus area where the mobile terminal is located is determined based on the medium access control address. The first-ranked information entry is read from the personalized information push queue, and the publishing area tag of the information entry is extracted. The publishing area tag of the information entry is compared with the current campus area; if they match, the information entry is marked as an instant push entry; otherwise, it is marked as a delayed push entry. For instant push entries, the main content of the instant push entry is compressed into a data packet. The compression process includes semantic compression of the text content of the information entry using a deep context model-based encoder, volume compression of images or video attachments in the information entry using a lossy compression algorithm, and encapsulating the compressed text semantic vector and media file stream in the same data packet. The push service registration identifier of the mobile terminal is written into the header of the data packet, and the data packet is sent via the campus wireless network using the User Datagram Protocol (UDP). For delayed push entries, the reference identifier of the information entry and the preset trigger area tag are stored in the local push cache table of the mobile terminal, which is maintained in key-value pair format. When the mobile terminal detects a change in the access point, it reads the new campus area corresponding to the new access point and compares it with the trigger area tag in the local push cache table. If they match, the corresponding delayed push entry is marked as a wake-up entry. The complete content of the wake-up entry is retrieved from the campus information resource database, a wake-up push data packet is generated, and it is sent to the mobile terminal through the new access point.

[0044] In practice, the information distribution module distributes personalized information push queues to the corresponding students' mobile terminals via the campus wireless network. The specific implementation is as follows: The medium access control address (MAC address) of the wireless access point currently connected to the student's mobile terminal is obtained. The MAC address is a unique hardware identifier for the wireless access point in the network. By querying a pre-established mapping table of campus wireless access point MAC addresses and campus area names, the current campus area of ​​the mobile terminal is determined. The first-ranked information entry is read from the personalized information push queue, and its publishing area tag is extracted. The publishing area tag represents the campus area name associated with the information entry's content. The publishing area tag of the information entry is compared with the current campus area. The comparison operation uses a string exact match method. If the two strings are the same, the information entry is marked as an immediate push entry; if the two strings are different, the information entry is marked as a delayed push entry.

[0045] In implementation, for instant push notifications, the main content of each notification is compressed into a single data packet. The compression process consists of two parts: semantic compression of the text content using a deep context model-based encoder, which uses a pre-trained BERT model to convert the text content into a fixed-dimensional semantic vector, which is then used as the compressed text representation; and lossy compression of image or video attachments within the notification, with image attachments using JPEG compression at a quality parameter of 75% and video attachments using H.264 encoding at a bitrate of 500kbps. The compressed text semantic vector and the media file stream are encapsulated within the same data packet using a custom binary protocol. The protocol header includes a data packet type identifier, the length of the compressed text semantic vector, the media file stream length, and a checksum. The push service registration identifier of the mobile terminal is written to the header of the data packet. This identifier is a unique 32-byte string obtained by the mobile terminal from the push server when it first launches the application. The data packet is sent via the campus wireless network using the User Datagram Protocol (UDP), which sends data directly without establishing a connection, using port number 8999. Optionally, after sending a data packet, the system does not wait for confirmation from the receiver. If the data packet is lost during network transmission, the mobile terminal sends a retransmission request to the push server in the next heartbeat cycle.

[0046] In specific implementation, for delayed push entries, the reference identifier and preset trigger area label of the information entry are stored in the local push cache table of the mobile terminal. The reference identifier is a unique number of the information entry in the campus information resource database, with a length of 64 bits. The preset trigger area label is a string copied from the publication area label of the information entry. The local push cache table is maintained in key-value pair format, where the key in the key-value pair is the preset trigger area label, and the value in the key-value pair is a list, where each element contains the reference identifier of the information entry and the content digest hash value of the information entry. The local push cache table is stored in the non-volatile memory of the mobile terminal, and the data is retained even after the mobile terminal restarts. In some embodiments, the maximum capacity of the local push cache table is set to 100 key-value pair entries. When the insertion of a new delayed push entry causes the total number of entries to exceed the maximum capacity, the earliest inserted delayed push entry is deleted according to a first-in, first-out (FIFO) strategy.

[0047] In practice, when a mobile terminal detects a change in the accessed wireless access point, it reads the new campus area corresponding to the new wireless access point. The detection of the wireless access point change is achieved through a network connection callback interface provided by the mobile terminal's operating system. Whenever the wireless network card is associated with a new wireless access point, the callback interface is triggered and returns the media access control address of the new wireless access point. The media access control address of the new wireless access point is converted into a name string for the new campus area using a mapping table. The new campus area is compared with the trigger area tag in the local push cache table. The comparison operation involves traversing all keys in the local push cache table and checking whether each key completely matches the new campus area string. If a match is found, the corresponding delayed push entry is marked as a pending wake-up entry. It can be understood that a new campus area may match multiple trigger area tags in the local push cache table, corresponding to multiple pending wake-up entries. The number of pending wake-up entries depends on the number of delayed push entries in the local push cache table that match the new campus area.

[0048] In practice, the complete content of the item to be woken up is retrieved from the campus information resource database, a wake-up push data packet is generated, and sent to the mobile terminal via a new wireless access point. The complete content retrieval process involves using the reference identifier of the item to be woken up as the query key to send a Hypertext Transfer Protocol (HTTP) request to the campus information resource database. The database returns the complete text content of the information item corresponding to the reference identifier, along with the download Uniform Resource Locator (URL) for the image or video attachment. The wake-up push data packet is generated in the same way as the instant push data packet: semantic compression is performed on the text content using a deep context model-based encoder, and lossy compression is used to compress the image or video attachment. The compressed text semantic vector and the media file stream are encapsulated in the same data packet, and the mobile terminal's push service registration identifier is written to the packet header. The wake-up push data packet is then sent to the mobile terminal via the new wireless access point using the User Datagram Protocol (UDP). Optionally, after the wake-up push data packet is sent, the corresponding delayed push entry is deleted from the local push cache table to free up storage space. In some embodiments, if the mobile terminal detects that the new campus area matches multiple trigger area tags in the local push cache table after switching wireless access points, it generates and sends wake-up push data packets in the order in which the matching trigger area tags are stored in the local push cache table, with the sending interval set to 500 milliseconds to avoid network congestion.

[0049] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A personalized information management system based on a smart campus, characterized by, include: The multimodal acquisition module acquires students' behavioral trajectory data, academic interaction data, and environmental perception data through multimodal acquisition terminals deployed in various areas of the campus. The individual state modeling module constructs an initial individual state vector for each student based on the behavioral trajectory data, the academic interaction data, and the environmental perception data. The demand analysis module inputs the initial individual state vector into a spatiotemporal attention network and outputs the personalized information demand weight distribution of each student at different times and locations. The information filtering and push module dynamically filters and sorts the original information entries in a campus information resource database according to the weight distribution of personalized information demand, and generates a personalized information push queue for each student. The information distribution module distributes the personalized information push queue to the corresponding students' mobile terminals through the campus wireless network.

2. The individualized information management system based on the wisdom campus according to claim 1, characterized in that, The acquisition of student behavior trajectory data, academic interaction data, and environmental perception data through multimodal data acquisition terminals deployed in various areas of the campus specifically includes: Bluetooth beacon detectors are installed at the entrance of each classroom, library, canteen, and dormitory building on campus. These detectors capture the Bluetooth signal strength sequence broadcast by students' mobile terminals. Each Bluetooth beacon detector's unique identifier is packaged with the Bluetooth signal strength sequence to generate a beacon detection record with a timestamp; A data scraping agent is deployed on the campus's online learning platform server. The data scraping agent reads each student's course viewing progress, assignment submission time, forum post content, and quiz answer records at fixed time intervals. The course viewing progress, the homework submission time, the forum post content, and the quiz answer records are stored in a structured manner to form a set of academic interaction records for each student; A capacitive pressure sensor array is embedded on the surface of the desks in each classroom to detect the occupancy status of seats and the frequency of micro-movements in the students' sitting posture. Distributed temperature and humidity sensors and carbon dioxide concentration sensors are deployed in the corridors of the teaching building and the study area of ​​the library. The temperature and humidity sensors and the carbon dioxide concentration sensors output environmental parameter sequences at a sampling rate of once per second. The seat occupancy status, the frequency of micro-movements in sitting posture, and the environmental parameter sequence are time-aligned to form an environmental perception record set for each student at each moment. The beacon detection record, the academic interaction record set, and the environmental perception record set corresponding to the same student at the same time are fused to generate the original multidimensional collection sample of that student at that time.

3. The individualized information management system based on the wisdom campus according to claim 2, characterized in that, The capacitive pressure sensor array identifies the occupancy status of each seat using an anonymous hash value, which is calculated by combining the seat number and a timestamp.

4. The individualized information management system based on the wisdom campus according to claim 2, characterized in that, Based on the behavioral trajectory data, the academic interaction data, and the environmental perception data, an initial individual state vector is constructed for each student, specifically including: The Bluetooth signal strength sequence over time is extracted from the original multidimensional acquisition samples. Peak detection is performed on the curve, and the time when each peak appears is taken as the time when the student enters the coverage area of ​​the corresponding Bluetooth beacon detector. The time difference between two consecutive moments when the student enters the coverage area of ​​different Bluetooth beacon detectors is taken as the time the student spends moving between the two areas. The student's moving speed is calculated based on the time spent moving and the preset physical distance between the two areas. By concatenating the movement speed and the corresponding area identifier in chronological order, a behavioral trajectory encoding sequence for the student is generated. Extract the student's total course viewing time, number of incomplete assignments, number of words posted on forums, and accuracy rate in answering quizzes from the academic interaction record set in the most recent 24 hours before the current moment. Concatenate the total viewing time, number of incomplete assignments, number of words posted on forums, and accuracy rate into a four-dimensional academic feature vector. The seat occupancy flag, the average frequency of sitting micro-movements over the past ten minutes, the temperature and humidity values, and the carbon dioxide concentration values ​​at the current moment are extracted from the environmental perception record set. The seat occupancy flag, the average frequency of sitting micro-movements, the temperature and humidity values, and the carbon dioxide concentration values ​​are then concatenated into a four-dimensional environmental feature vector. The behavior trajectory encoding sequence is compressed into a fixed-length trajectory embedding vector; The trajectory embedding vector, the four-dimensional academic feature vector, and the four-dimensional environmental feature vector are concatenated end-to-end along the vector dimension to generate the initial individual state vector of each student at the current moment.

5. The individualized information management system based on the wisdom campus according to claim 4, characterized in that, The behavior trajectory encoding sequence is compressed into a fixed-length trajectory embedding vector, specifically by: encoding the behavior trajectory encoding sequence using a pre-trained unidirectional long short-term memory network, and taking the output of the last hidden layer as the trajectory embedding vector.

6. A personalized information management system based on a smart campus according to claim 4, characterized in that, The initial individual state vector is input into a spatiotemporal attention network, which outputs the personalized information demand weight distribution for each student at different times and locations, specifically including: Obtain the initial individual state vector for each of the student's past seven consecutive natural days, and divide the initial individual state vector of each day into ninety-six time slot state vectors according to a time slot of fifteen minutes. Stack the state vectors of ninety-six time slots at the same time slot location on different dates to generate a historical state tensor of size seven times ninety-six times the dimension of the state vector. The historical state tensor is input into the temporal attention module in the spatiotemporal attention network. The temporal attention module performs self-attention calculation on the state similarity of each time slot position between different dates and outputs the diurnal cycle importance score of each time slot. The historical state tensor is input into the spatial attention module in the spatiotemporal attention network. The spatial attention module performs attention calculation on the state transition frequency between different Bluetooth beacon detector coverage areas and outputs the spatial importance score of each campus area. The behavior trajectory encoding sequence in the initial individual state vector at the current moment is differentially calculated with the historical trajectory encoding corresponding to the current time slot in the historical state tensor to obtain the trajectory change increment vector. The trajectory change increment vector is passed through a fully connected mapping layer and then multiplied element-wise with the daytime periodic importance score and the spatial importance score to generate a personalized information demand weight distribution for each student at the current moment for different campus areas and different time offsets.

7. A personalized information management system based on a smart campus according to claim 6, characterized in that, The time attention module uses a multi-head self-attention mechanism to calculate the importance score of the daytime cycle.

8. A personalized information management system based on a smart campus according to claim 6, characterized in that, Based on the weight distribution of personalized information demand, the original information entries in a campus information resource database are dynamically filtered and sorted to generate a personalized information push queue for each student, specifically including: Read all raw information entries to be pushed from the campus information resource database. Each raw information entry carries a publishing area tag, a publishing time tag, and a content type tag. The publishing area tag of each original information entry is matched with the spatial importance score in the personalized information demand weight distribution to calculate the area matching score of the original information entry. The publication time period tag of each original information entry is compared with the time slot position to which the current time belongs. If the daytime cycle importance score of the publication time period tag and the current time exceeds the preset matching threshold, it is marked as a time period match; otherwise, it is marked as a time period mismatch. The original information entries that are marked as matching the time period are sorted in the first round according to the region matching score from high to low, and the original information entries that are marked as not matching the time period are temporarily stored in a delayed push buffer; For multiple original information entries with the same region matching score in the first round of sorting, the content type label of each original information entry is further obtained, and the content type label is matched with the content interest tendency extracted from the trajectory change increment vector to calculate the content matching score. Based on the content matching score, a second round of sorting is performed on multiple original information entries with the same regional matching score to form a personalized information push queue for each student. Check the waiting time of the original information entries in the delayed push buffer. When the waiting time exceeds the preset maximum delay time, remove the corresponding original information entries from the delayed push buffer and insert them into the end of the personalized information push queue according to the regional matching score.

9. A personalized information management system based on a smart campus according to claim 8, characterized in that, The personalized information push queue is distributed to the mobile terminals of the corresponding students via the campus wireless network, specifically including: Obtain the Media Access Control Address of the wireless access point currently accessed by the student's mobile terminal, and determine the current campus area where the mobile terminal is located based on the Media Access Control Address; Read the first-ranked information entry from the personalized information push queue and extract the publishing area tag of the information entry; The publishing area tag of the information entry is compared with the current campus area. If they match, the information entry is marked as an instant push entry; if they do not match, the information entry is marked as a delayed push entry. For the instant push entry, the main content of the instant push entry is compressed into a data packet, the push service registration identifier of the mobile terminal is written into the header of the data packet, and the data packet is sent through the campus wireless network in the manner of User Datagram Protocol. For the delayed push entry, the reference identifier of the information entry and the preset trigger area label are stored in the local push cache table of the mobile terminal, and the local push cache table is maintained in the form of key-value pairs; When the mobile terminal detects a change in the access point, it reads the new campus area corresponding to the new access point and compares the new campus area with the trigger area label in the local push cache table. If they match, the corresponding delayed push entry is marked as a wake-up entry. The complete content of the entry to be woken up is retrieved from the campus information resource database, a wake-up push data packet is generated, and it is sent to the mobile terminal through a new wireless access point.

10. A personalized information management system based on a smart campus according to claim 9, characterized in that, The main content of the instant push item is compressed into a data packet, specifically including: semantic compression of the text content of the information item using an encoder based on a deep context model, volume compression of the image or video attachments in the information item using a lossy compression algorithm, and encapsulating the compressed text semantic vector and the media file stream in the same data packet.

Images