College intelligent system based on cloud native and multi-modal data fusion

Abstract

The embodiment of the invention provides a college intelligent system based on cloud native and multi-modal data fusion, and the system comprises a multi-modal data collection layer which is used for obtaining a multi-modal behavior data sequence supporting a to-be-tested object to be in a specific behavior state through a campus global data collection platform in a current time period; the data fusion and intelligent analysis layer is used for analyzing the multi-modal behavior data sequence by adopting a depth evidence uncertainty model to obtain belief quality and uncertainty of supporting the to-be-detected object to be in a specific behavior state at each time point in the current time period; and the micro-service business middle platform is used for predicting the belief trend of the to-be-tested object in the specific behavior state in the future time period according to the belief quality and uncertainty of the to-be-tested object in the specific behavior state supported at each time point in the current time period, and managing the to-be-tested object in the future time period according to the belief trend. According to the invention, the refinement, intelligence and individuation level of campus management can be improved.

Description

Technical Field

[0001] This application relates to the interdisciplinary field of educational informatization and cloud computing, and involves, but is not limited to, a smart university system based on cloud-native and multimodal data fusion. Background Technology

[0002] With the deepening of smart campus construction, university management increasingly relies on data-driven intelligent decision-making. Traditional systems typically depend on single, deterministic data analysis models, which often have significant shortcomings when fusing and analyzing multi-source, heterogeneous, multimodal data from multiple sources such as cameras, sensors, and academic records. These systems struggle to quantify the uncertainties inherent in the data itself and in the model's judgments, such as conflicts between different information sources, missing data, or noise interference. This results in fragile decision-making bases and an inability to effectively assess the reliability of prediction results. Especially when dealing with dynamically evolving scenarios such as student behavior recognition and academic warnings, static models cannot adapt to the continuous changes in data distribution, highlighting their rigid decision-making and lack of reliability.

[0003] To address these challenges, some probability-based fusion methods have begun to be introduced into existing technologies. However, most of these methods still remain within the traditional probability framework and fail to fundamentally model and separate cognitive uncertainty. Furthermore, how to use dynamically changing uncertainty as a core decision factor to predict future trends remains a difficult problem in the construction of smart campuses. Summary of the Invention

[0004] Based on the above problems, this application provides a university intelligent system based on cloud-native and multimodal data fusion. It aims to quantify uncertainty based on cloud-native-driven full-domain data collection, and improve the refinement, intelligence and personalization of campus management based on the quantified uncertainty.

[0005] The technical solution of this application embodiment is implemented as follows: This application provides a university intelligent system based on cloud-native and multimodal data fusion, the system comprising: The multimodal data acquisition layer is used to acquire multimodal behavioral data sequences representing the specific behavioral states of the test object within the current time period using the campus-wide data acquisition platform; the campus-wide data acquisition platform is an acquisition platform built on cloud-native infrastructure. The data fusion and intelligent analysis layer is used to analyze multimodal behavioral data sequences using a deep evidence uncertainty model to obtain the quality and uncertainty of beliefs supporting the test subject's specific behavioral state at each time point within the current time period. The deep evidence uncertainty model is constructed based on the Dirichlet distribution and subjective logic theory, and is trained and updated using a playback-based incremental learning strategy. The microservice business platform is used to predict the belief trend of the test object in a specific behavioral state in the future time period based on the belief quality and uncertainty of the test object at each point in time within the current time period, and to manage the test object in the future time period according to the belief trend.

[0006] In some embodiments, the data fusion and intelligent analysis layer includes: an evidence vector generation module, used to extract features from each modal behavioral data in the multimodal behavioral data sequence according to each time point within the current time period, to obtain an evidence vector set for each time point within the current time period; a multimodal data conflict processing module, used to utilize a deep evidence uncertainty model, based on subjective logic theory, to perform Dirichlet distribution parameter transformation on the evidence vector set for each time point within the current time period, to obtain a Dirichlet concentration parameter supporting the test object being in a specific behavioral state at each time point within the current time period; and a belief quality and uncertainty calculation module, used to input the Dirichlet concentration parameter supporting the test object being in a specific behavioral state at each time point within the current time period into a belief quality and uncertainty calculation formula determined based on subjective logic theory for calculation, to obtain the belief quality and uncertainty supporting the test object being in a specific behavioral state at each time point within the current time period.

[0007] In some embodiments, the multimodal data conflict processing module is specifically used to utilize a deep evidence uncertainty model to assign weights to each evidence vector in the evidence vector set at each time point within the current time period, based on the weights of each evidence vector in the evidence vector set, to obtain the weights of each evidence vector in the evidence vector set; and to perform weighted fusion of each evidence vector in the evidence vector set according to the weights of each evidence vector in the evidence vector set to obtain evidence values ​​supporting the test object being in a specific behavioral state at each time point within the current time period, and to convert the evidence values ​​into Dirichlet concentration parameters supporting the test object being in a specific behavioral state at each time point within the current time period based on subjective logic theory.

[0008] In some embodiments, the multimodal data conflict processing module is further configured to determine the reliability of the data source corresponding to each evidence vector in the evidence vector set for each time point within the current time period as the weight of each evidence vector in the evidence vector set.

[0009] In some embodiments, the system further includes: a model initialization layer, used to build a network model whose output layer is configured as an evidence layer based on the Dirichlet distribution and subjective logic theory, and to train the network model using an initial multimodal training dataset corresponding to the campus-wide data and an evidence loss function to obtain a deep evidence uncertainty model; and a model update layer, used to adjust the network parameters of the deep evidence uncertainty model based on incremental learning of the replay buffer pool to obtain an updated deep evidence uncertainty model.

[0010] In some embodiments, the model update layer is specifically configured to, in response to receiving new multimodal data, select initial multimodal training sub-data from the initial multimodal training dataset, and, based on the new multimodal data and the initial multimodal training sub-data, incrementally learn the deep evidence uncertainty model using an overall loss function that combines a new and old knowledge retention mechanism, to obtain an updated deep evidence uncertainty model. The overall loss function is: in, Total loss; This is the balance coefficient; The prediction loss of the deep evidence uncertainty model on new multimodal data; This represents the prediction loss of the deep evidence uncertainty model on the initial multimodal training subdata.

[0011] In some embodiments, the microservice business platform is specifically used to infer the belief trend of the test object in a specific behavioral state in the future time period based on the belief quality and uncertainty of the test object in a specific behavioral state at each time point within the current time period through a time series prediction model; and to manage the test object based on the comparison result between the belief trend and the preset strategy value.

[0012] In some embodiments, the unified service interaction layer is used to set up a customized cockpit carrying the core information of the test object based on the role of the test object in the university campus scene; and in response to receiving the test object's access operation to the customized cockpit, to display the profile of the test object in the university campus scene and personalized services for the test object in real time.

[0013] The beneficial effects of the technical solutions provided in this application include at least the following: This application provides a cloud-native and multimodal data fusion-based intelligent system for universities. The system includes: a multimodal data acquisition layer, used to acquire multimodal behavioral data sequences representing the behavior of a target object in a specific behavioral state within the current time period using a campus-wide data acquisition platform; wherein the campus-wide data acquisition platform is a cloud-native infrastructure-based acquisition platform; thus, real-time acquisition and aggregation of multi-source heterogeneous data across the campus can be achieved through the cloud-native infrastructure-based acquisition platform, ensuring the comprehensiveness and timeliness of the acquired multimodal behavioral data; and a data fusion and intelligent analysis layer, used to analyze the multimodal behavioral data sequences using a deep evidence uncertainty model to obtain the belief quality and uncertainty supporting the target object's behavior in a specific behavioral state at each time point within the current time period; wherein the deep evidence uncertainty model is constructed based on Dirichlet distribution and subjective logic theory. The model is trained and updated using a replay-based incremental learning strategy. This deep evidence uncertainty model enables the fusion analysis of multimodal behavioral data, outputting refined and interpretable belief levels for specific behavioral states. It explicitly provides a measure of uncertainty for each judgment, offering crucial information for subsequent risk assessment and decision-making. The microservice business platform, based on the belief quality and uncertainty of the target object at each point in time within the current period, predicts the belief trend of the target object in a specific behavioral state in the future, and manages the target object in the future based on this belief trend. Thus, it is possible to proactively predict the behavioral trend of the target object in the future based on the belief quality and uncertainty sequence of the current period, thereby shifting from post-event analysis to pre-event warning, enabling personalized and precise intervention measures for the target object. In this way, by collecting data across the entire domain based on cloud-native technology, uncertainty is quantified, and based on the quantified uncertainty, accurate prediction and flexible management are achieved. This not only greatly improves the level of refinement, intelligence and personalization of campus management, enabling dynamic scheduling of resources, data-driven decision-making and personalized services within university campuses, but also enhances the robustness of the system in dealing with complex campus scenarios by leveraging its inherent uncertainty quantification capabilities and continuous learning capabilities.

[0014] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the technical solutions provided in the embodiments of this application. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort, wherein: Figure 1 A schematic diagram illustrating the composition of a university intelligent system based on cloud-native and multimodal data fusion, provided for an embodiment of this application; Figure 2 A schematic diagram illustrating the deployment and governance of microservices in a cloud-native environment built upon a Kubernetes cluster and Istio-system, provided for embodiments of this application; Figure 3 A schematic diagram illustrating the composition of another intelligent university system based on cloud-native and multimodal data fusion, provided as an embodiment of this application; Figure 4 A schematic diagram illustrating the composition of another intelligent university system based on cloud-native and multimodal data fusion, provided as an embodiment of this application; Figure 5 This is a schematic diagram of the architecture of the intelligent university system based on cloud-native and multimodal data fusion provided in the embodiments of this application in a practical application. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. The following embodiments are used to illustrate this application, but are not intended to limit the scope of this application. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0017] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0018] It should be noted that the terms "first, second, and third" used in the embodiments of this application are merely to distinguish similar objects and do not represent a specific ordering of objects. It is understood that "first, second, and third" can be interchanged in a specific order or sequence where permitted, so that the embodiments of this application described herein can be implemented in an order other than that illustrated or described herein.

[0019] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which the embodiments of this application pertain. It should also be understood that terms such as those defined in general dictionaries should be understood to have a meaning consistent with their meaning in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0020] Example 1: See Figure 1 The diagram shown is a schematic representation of the structural composition of a high-efficiency intelligent system based on cloud-native and multimodal data fusion, according to an embodiment of this application. The high-efficiency intelligent system 100 based on cloud-native and multimodal data fusion includes: The multimodal data acquisition layer 102 is used to acquire, within the current time period, multimodal behavioral data sequences representing the test object being in a specific behavioral state using the campus-wide data acquisition platform 101.

[0021] Among them, the campus-wide data collection platform 101 is a collection platform built on cloud-native infrastructure.

[0022] The data fusion and intelligent analysis layer 103 is used to analyze multimodal behavioral data sequences using a deep evidence uncertainty model to obtain the quality and uncertainty of beliefs supporting the test subject's specific behavioral state at each time point within the current time period.

[0023] Among them, the deep evidence uncertainty model is constructed based on the Dirichlet distribution and subjective logic theory, and is trained and updated using a playback-based incremental learning strategy.

[0024] The microservice business middle platform 104 is used to predict the belief trend of the test object in a specific behavioral state in the future time period based on the belief quality and uncertainty of the test object at each point in time within the current time period, and to manage the test object in the future time period according to the belief trend.

[0025] In some embodiments, the multimodal data acquisition layer 102 is responsible for the unified access and quality assurance of multi-source data on the university campus, providing a foundation for subsequent data fusion and intelligent analysis layer 103 to perform data fusion. Specifically, the multimodal data acquisition layer 102 can achieve full coverage access to multiple types of data (such as video data, image data, and IoT sensor data) through the campus-wide data acquisition platform 101. This includes structured data such as academic affairs databases and student affairs databases, unstructured data such as temperature, humidity, light intensity, and energy consumption, and semi-structured data such as campus card consumption logs, library access records, and network access logs.

[0026] Here, to ensure data quality, the multimodal data acquisition layer 102 can also introduce a data integrity verification unit based on Bayesian networks, which is used to automatically detect problems such as missing or abnormal multimodal behavioral data, and generate a supplementary acquisition priority based on the importance of the data, so as to avoid the impact of data quality problems on subsequent analysis results.

[0027] In some embodiments, the current time period may refer to the current hour, the current class, or the current semester, and this application does not make any limitation in this regard; correspondingly, the test object may refer to students, teachers, academic affairs staff, or teaching equipment in colleges and universities, and this application does not make any limitation in this regard either.

[0028] It should be noted that a specific behavioral state can refer to a stable and identifiable behavioral pattern or operational status exhibited by the test subject within a specific time period (such as the current time period mentioned in this application). This specific behavioral state is not an instantaneous action, but a continuous state defined by a series of related activities and data patterns. For example, when the test subject is a university student, the specific behavioral state could be: pre-exam cramming state, academic difficulty state, or financial strain state; when the test subject is a university teacher or academic affairs staff, the specific behavioral state could be: professional burnout state or interdisciplinary collaboration state; when the test subject is university teaching equipment, the specific behavioral state could be: regular operation state or environmental adaptation state.

[0029] For example, when the test subject is a college student, the multimodal behavioral data sequence that supports the test subject being in a specific behavioral state can be: a multimodal consumption behavior data sequence that can prove or indicate that college student A is in a state of economic stress; or, a multimodal consumption behavior data sequence that can prove or indicate that college student B is in a state of irregular life (high pressure); or, a multimodal learning behavior data sequence that can prove or indicate that college student C is in a state of efficient focus.

[0030] When the subject of the test is a university teacher, the multimodal behavioral data sequence that supports the subject of the test being in a specific behavioral state can be: a multimodal teaching behavior data sequence that can prove or indicate that university teacher D is in a state of teaching burnout; or a multimodal office data sequence that can prove or indicate that university teacher E is in a state of administrative overload.

[0031] When the object to be tested is campus equipment of a university, the multimodal behavioral data sequence that supports the object to be tested being in a specific behavioral state can be: a multimodal operation data sequence that can prove or indicate that the air conditioner in the classroom on the second floor of Building F of the university campus is in an inefficient and aging state; or a multimodal environment data sequence that can prove or indicate that the experimental equipment in the classroom on the first floor of Building G of the university campus is in a state of reasonable demand.

[0032] In some embodiments, when a multimodal behavioral data sequence representing a test subject (within the current time period) is defined as a multimodal consumption behavior data sequence capable of proving or indicating that college student A is in a state of economic stress, the multimodal behavioral data sequence includes at least: 1. The consumption transaction data sequence of college student A on the college campus during the current time period can be derived from the college campus card system built by the campus-wide data collection platform 101. This consumption transaction data sequence can include: all consumption nodes on the college campus, such as: canteens, convenience stores, coffee shops, etc., and the data dimensions of this consumption transaction data sequence include: consumption timestamp, consumption location, consumption amount, and complete historical records of the type of merchant.

[0033] 2. Video data sequence of college student A in public places such as canteens during the current time period. The video data sequence may include: data sequence of stay and queuing in front of canteen windows of different price ranges, the length of stay in the dining area, and whether there are frequent companions who consume together.

[0034] In some embodiments, the multimodal data acquisition layer 102 is used to first acquire initial multimodal data representing the test object being in a specific behavioral state using the campus-wide data acquisition platform 101 within the current time period; then, through timestamps, perform alignment and synchronization analysis on the acquired initial multimodal data to form a complete data sequence that can reflect the test object being in a specific behavioral state, i.e., a multimodal behavioral data sequence.

[0035] In some embodiments, the campus-wide data acquisition platform 101 is a data acquisition platform built on cloud-native infrastructure within the university intelligent system 100, which integrates cloud-native and multimodal data. This campus-wide data acquisition platform 101 serves as the underlying support for the university intelligent system 100, and its corresponding cloud-native infrastructure utilizes Kubernetes clusters as a powerful container orchestration platform and Istio-system as an advanced microservice connection and security mesh to build an efficient, secure, and easily manageable application deployment and operation environment. In other words, the core objective of this cloud-native infrastructure is to achieve elastic resource scheduling and stable service governance.

[0036] The Kubernetes cluster supports Horizontal Pod Autoscaler (HPA), which dynamically adjusts computing resources based on workload to avoid waste or shortage. The Istio service mesh within the Istio-system enables fine-grained service governance, including circuit breaking, rate limiting, and failover capabilities, ensuring system stability under high concurrency. Furthermore, it uses the mTLS protocol to encrypt inter-service communication, preventing data leakage risks. To further reduce operational costs, this application integrates a DevOps toolchain (e.g., GitLab CI) into the cloud-native infrastructure (campus-wide data acquisition platform 101) to achieve Continuous Integration / Continuous Deployment (CI / CD), automating code submission, testing, and deployment processes and reducing manual intervention.

[0037] For reference here. Figure 2 The diagram shown illustrates the deployment and governance of microservices in a cloud-native environment built upon a Kubernetes cluster and Istio-system, as provided in this embodiment of the application. Figure 2 The focus is on the core technology of the intelligent university system 100 based on cloud-native and multimodal data fusion—microservice governance in a cloud-native environment. It showcases a simplified Kubernetes cluster that uses the Istio-system service mesh to uniformly manage multiple business microservices (such as intelligent academic affairs and student services). Figure 2 It further demonstrates key governance functions such as service discovery and load balancing, elastic policies (circuit breaking / rate limiting), secure communication (mTLS), and observability (monitoring / tracing) to illustrate how the university intelligent system 100 based on cloud-native and multimodal data fusion achieves high availability, high security, and easy observability of services.

[0038] In some embodiments, the data fusion and intelligent analysis layer 103, as the core of the intelligent university system 100 based on cloud-native and multimodal data fusion, can achieve effective fusion and in-depth analysis of multimodal behavioral data sequences acquired by the multimodal data acquisition layer 102 through storage architecture optimization and algorithm innovation. Specifically, in terms of storage: the data fusion and intelligent analysis layer 103 can adopt a data lakehouse model, balancing low-cost storage of raw multimodal data (i.e., multimodal behavioral data sequences) with efficient querying of preprocessed data, achieving full lifecycle management of data; in terms of analytical capabilities, a machine learning platform can be built based on Kubeflow (an open-source Kubernetes native framework), integrating feature engineering, model training, evaluation, and service-oriented tools to support the rapid development and deployment of intelligent models.

[0039] It should be noted that the deep evidence uncertainty model in the data fusion and intelligent analysis layer 103 is constructed based on the Dirichlet distribution and subjective logic theory, and is trained and updated using a playback-based incremental learning strategy. This Dirichlet distribution and subjective logic theory can effectively handle conflicts in multimodal behavioral data in universities (such as the inconsistency between "sensor shows no one in the classroom" and "video captures lingering personnel"). The playback-based incremental learning strategy can dynamically update the parameters of the deep evidence uncertainty model to highly adapt to the dynamic evolution characteristics of university data (such as course adjustments and changes in student behavior) in order to realize typical intelligent scenarios such as classroom effectiveness evaluation, early warning of abnormal student behavior, and energy consumption optimization in universities.

[0040] In some embodiments, the Deep Evidential Uncertainty (DEU) model is used for resolving conflicts in multimodal behavioral data sequences. Its fundamental principle differs from traditional neural networks that output single-class probabilities; DEU is inspired by the Dirichlet Distribution and Subjective Logic theory. DEU not only outputs prediction results, but more importantly, it outputs quantified values ​​of evidence and uncertainty. Evidence represents the cumulative strength supporting the candidate's state of a specific behavior, derived from the input data. Stronger evidence results in higher prediction confidence. Uncertainty represents the degree of doubt DEU has about the prediction result due to insufficient or conflicting evidence. When data from different modalities provide contradictory evidence, DEU automatically generates a high uncertainty value, rather than forcibly giving a high-confidence erroneous prediction.

[0041] In some embodiments, reference Figure 3 As shown, the data fusion and intelligent analysis layer 103 includes: The evidence vector generation module 1031 is used to extract features from each modal behavior data in the multimodal behavior data sequence according to each time point within the current time period, so as to obtain the evidence vector set for each time point within the current time period.

[0042] The multimodal data conflict processing module 1032 is used to utilize the deep evidence uncertainty model and, based on subjective logic theory, to transform the evidence vector set at each time point in the current time period into Dirichlet distribution parameters, thereby obtaining the Dirichlet concentration parameters at each time point in the current time period that support the test object being in a specific behavioral state.

[0043] The belief quality and uncertainty calculation module 1033 is used to input the Dirichlet concentration parameters that support the test object being in a specific behavioral state at each time point within the current time period into the belief quality and uncertainty calculation formula determined based on subjective logic theory to calculate the belief quality and uncertainty at each time point within the current time period.

[0044] In some embodiments, the data fusion and intelligent analysis layer 103 first uses the evidence vector generation module 1031 to convert the multimodal behavioral data sequence into an evidence vector set for each time point in the current time period; then, using the multimodal data conflict processing module 1032 and with the help of the deep evidence uncertainty model, the evidence vector set for each time point in the previous time period is converted into a Dirichlet concentration parameter for each time point in the current time period that supports the test object being in a specific behavioral state; finally, using the belief quality and uncertainty calculation module 1033, the Dirichlet concentration parameter for each time point in the current time period that supports the test object being in a specific behavioral state is input into the belief quality and uncertainty calculation formula based on subjective logic theory as shown in formula (1), to obtain the belief quality and uncertainty for each time point in the current time period that supports the test object being in a specific behavioral state.

[0045] Formula (1); in, For each point in time within the current time period Supports the test object being in a specific behavioral state. The quality of beliefs; For each point in time within the current time period The total amount of evidence; For each point in time within the current time period Supports the test object being in a specific behavioral state. Dirichlet concentration parameters; For each point in time within the current time period Supports the test object being in a specific behavioral state. Dirichlet concentration parameters; The number of categories for a specific behavioral state; For summation index; For each point in time within the current time period Supports the test object being in a specific behavioral state. Evidence value; For each point in time within the current time period Supports the test object being in a specific behavioral state. Evidence value; For each point in time within the current time period It supports the uncertainty of the test object being in a specific behavioral state.

[0046] In some embodiments, the multimodal data conflict processing module 1032 described above may be further used for: Using a deep evidence uncertainty model, for each time point in the current time period, the weights of each evidence vector in the evidence vector set are assigned accordingly, thus obtaining the weights of each evidence vector in the evidence vector set. Then, the evidence vectors in the evidence vector set are weighted and fused according to their respective weights to obtain the evidence value supporting the test object being in a specific behavioral state at each time point in the current time period. Based on subjective logic theory, the evidence value is converted into a Dirichlet concentration parameter supporting the test object being in a specific behavioral state at each time point in the current time period.

[0047] Here, the multimodal data conflict processing module 1032 is also used to determine the reliability of the data source corresponding to each evidence vector in the evidence vector set for each time point in the current time period as the weight of each evidence vector in the evidence vector set.

[0048] In some embodiments, the multimodal data conflict processing module 1032 is further configured to assign different confidence weights to each data source based on the inherent reliability and accuracy of each data source; for example, video data, because it can directly capture the visual behavior and state of the object under test to provide an objective, continuous and information-rich stream of evidence, has a higher weight than the indirect and discrete behavioral proxy indicator of check-in behavior data.

[0049] In some embodiments, for each point in time within the current time period, the evidence vector generation module 1031 within the data fusion and intelligent analysis layer 103 can receive data from different sources input from the multimodal data input layer 102, which supports the test object being in a specific behavioral state at that point in time. This data includes, but is not limited to, multimodal behavioral data sequences (including but not limited to, sensor data, video feature vectors, and timetable status). The module then processes the data from each modality within the multimodal behavioral data sequence through a corresponding feature extraction network (e.g., convolutional neural networks for video data, recurrent neural networks for sequence data) to generate preliminary evidence. Vectors; Correspondingly, the multimodal data conflict processing module 1032 within the data fusion and intelligent analysis layer 103 can perform uncertainty-aware fusion. Here, it is not a simple averaging or weighted voting. The multimodal data conflict processing module 1032 first assigns a credibility weight to the evidence of each modality. This weight is not preset, but is automatically learned and generated by the quality of the current input data (for example, the credibility of the video modality is automatically reduced when the light is dim, and its weight is reduced). Then, the evidence formed by weighted fusion of all vectors is merged to form a Dirichlet distribution Dir(α) that supports the test object being in a specific behavioral state at that time point.

[0050] Correspondingly, the belief quality and uncertainty calculation module 1032 in the data fusion and intelligent analysis layer 103 can calculate the belief quality b and uncertainty u that support the test object being in a specific behavioral state at that time point based on the Dirichlet distribution Dir(α) that supports the test object being in a specific behavioral state at that time point.

[0051] It should be noted that the closer the value of uncertainty u is to 1, the higher the uncertainty (i.e., high conflict or low evidence).

[0052] In some embodiments, this deep evidence uncertainty model can handle data conflict, data uncertainty, and dynamic evolution within university campuses. Data conflict refers to the frequent contradictions that often arise when heterogeneous data from multiple sources within universities describe the same entity or event. For example: 1. Scenario A (Classroom Occupancy Conflict): Infrared sensor data indicates "no one in the classroom," but video surveillance identifies "students remaining," and the timetable shows "no classes currently scheduled." In this situation, it's impossible to determine which data source is more reliable.

[0053] 2. Scenario B (Student Abnormal Behavior Conflict): Consumption data shows a student's recent food spending is extremely low (possibly due to financial difficulties), but online logs show frequent game top-ups (possibly due to game addiction), while access control data shows a normal daily routine. How to comprehensively determine their true situation?

[0054] Data uncertainty arises from noise, missing data, or limitations in accuracy inherent in the data itself. For example, Internet of Things (IoT) sensor readings contain errors, and video data recognition suffers from false alarm rates.

[0055] Dynamic evolution is due to the fact that data patterns in universities change over time. For example, course arrangements differ each semester, and student behavior patterns vary by grade and season (e.g., behavior differs between the final exam preparation period and the beginning of the semester). Traditional static machine learning models can quickly become outdated due to concept drift.

[0056] Here, the deep evidence uncertainty model provided in this application can be used to output the prediction result "Someone is here" (probability 0.6) and uncertainty 0.75 in the case of classroom occupancy conflict (Scenario A). This high uncertainty value (0.75) will trigger subsequent operations of the university intelligent system 100 based on cloud-native and multimodal data fusion (such as notifying the administrator to review or assigning higher weight to sensor data for recalculation), thereby avoiding making wrong decisions based on conflict data (such as wrongly turning off lighting and air conditioning).

[0057] In some embodiments, such as Figure 4 As shown, the university intelligent system 100 based on cloud-native and multimodal data fusion may further include: Model initialization layer 105 is used to build a network model whose output layer is configured as the evidence layer based on the Dirichlet distribution and subjective logic theory. It also uses the initial multimodal training dataset corresponding to the campus-wide data to train the network model using the evidence loss function, thus obtaining a deep evidence uncertainty model.

[0058] Model update layer 106 is used for incremental learning based on the replay buffer pool to adjust the network parameters of the deep evidence uncertainty model and obtain the updated deep evidence uncertainty model.

[0059] In some embodiments, the model initialization layer 105 is used to construct a network based on Dirichlet distribution and subjective logic theory. The output layer of this network is specially designed as an evidence layer, directly regressing non-negative evidence values. Subsequently, the network is trained using an initial multimodal training dataset covering the entire campus and employing an evidence loss function specifically designed for uncertainty modeling, ultimately obtaining a deep evidence uncertainty model capable of quantifying predictive beliefs and uncertainties.

[0060] In some embodiments, the model update layer 106 is specifically used to select initial multimodal training sub-data from the initial multimodal training dataset in response to receiving new multimodal data, and to incrementally learn the deep evidence uncertainty model using an overall loss function that combines a new and old knowledge retention mechanism based on the new multimodal data and the initial multimodal training sub-data, so as to obtain an updated deep evidence uncertainty model.

[0061] In some embodiments, the model update layer 106 is used to achieve real-time updates of the deep evidence uncertainty model. Here, reinforcement learning can be employed to address the dynamic evolution of data. Specifically, when the university intelligent system 100, based on cloud-native and multimodal data fusion, continuously receives new batches of data (e.g., new semester timetables, newly generated student behavior data), a replay-based incremental learning strategy can be used. When training the deep evidence uncertainty model, a portion of old data sampled from a "memory (historical data)" is mixed in with the training. The university intelligent system 100, based on cloud-native and multimodal data fusion, continuously selects representative old data samples (e.g., core samples selected through clustering) and stores them in a fixed-size memory. When a new batch of data arrives, the training algorithm not only uses the new data but also extracts a small portion of old data from the memory to jointly train the model. Here, the overall loss function used in the training phase of the deep evidence uncertainty model... It can be shown in the following formula (2): Formula (2); in, Total loss; This is the balance coefficient; The prediction loss of the deep evidence uncertainty model on new multimodal data; This represents the prediction loss of the deep evidence uncertainty model on the initial multimodal training subdata.

[0062] here, Hyperparameters used to control the degree of attention given to new multimodal data and some historical multimodal data during the optimization process of the deep evidence uncertainty model.

[0063] In some embodiments, after completing the incremental update of the deep evidence uncertainty model, only the parameter increment of the updated deep evidence uncertainty model can be published to the production application environment corresponding to the cloud-native and multimodal data fusion university intelligent system 100, so as to achieve fast, low-bandwidth model update without the need for full retraining and deployment.

[0064] It should be noted that incremental learning enables the deep evidence uncertainty model to learn new knowledge from continuously new data without forgetting old knowledge (i.e., overcoming the "catastrophic forgetting" problem). This means that the deep evidence uncertainty model can continuously learn from new data without forgetting old knowledge, adapt to changes in student behavior patterns, and ensure the long-term effectiveness of the deep evidence uncertainty model, thereby achieving dynamic evolution and smooth upgrading of the deep evidence uncertainty model.

[0065] For example, taking the application of the deep evidence uncertainty model as a student behavior early warning model: after the new students enroll in September, when it is found that the behavior patterns of the new students (such as consumption habits and library usage frequency) are different from those of senior students, an incremental learning algorithm can be used to update the original deep evidence uncertainty model with the new student data, so that the deep evidence uncertainty model can better identify the abnormal behavior of the new students, while not forgetting the ability to judge the behavior patterns of senior students.

[0066] Alternatively, consider the application of deep evidence uncertainty models as smart energy-saving models: Air conditioning energy consumption patterns on university campuses differ significantly between summer and winter. Deep evidence uncertainty models can be incrementally adjusted based on seasonal data, automatically adapting to the optimal energy-saving strategies for different seasons.

[0067] In some embodiments, the microservice business platform 104 is specifically used to infer the belief trend of the test object in a specific behavioral state in a future time period based on the belief quality and uncertainty of the test object in a specific behavioral state at each point in time within the current time period, through a time series prediction model.

[0068] The test object is managed based on the comparison between the belief trend and the preset strategy value.

[0069] In some embodiments, the microservice business platform 104 first converts the belief quality and uncertainty supporting the test object's specific behavioral state at each time point within the current time period into a belief quality-uncertainty sequence arranged by timestamps. Then, it feeds this belief quality-uncertainty sequence into a dedicated time series prediction model (e.g., a model built on a single-layer long short-term memory network, trained with historical data, capable of capturing the periodicity and contextual patterns of belief changes, and ultimately outputting a prediction sequence of belief values ​​for a future period), outputting the belief trend of the test object in a specific behavioral state in the future time period. Finally, it compares the belief trend with a preset strategy value composed of a dynamic judgment rule set to obtain the corresponding early warning rule, thereby managing the test object based on the early warning rule (e.g., sending early warning information to counselors, automatically allocating learning resources, or adjusting students' course selection suggestions), thus achieving accurate and forward-looking management of the test object.

[0070] It should be noted that the training data for this time series prediction model (single-layer long short-term memory network) can be: historical belief quality sequences at multiple time steps as input, belief quality or state labels at a future point in time as training targets, and the loss function involved in the training process can be the mean squared error loss function or the cross-entropy loss function, and training can be performed using the gradient descent method.

[0071] For example, the microservice business platform 104 is used to form a belief quality-uncertainty sequence for each month based on the belief quality and uncertainty supporting student S's "academic difficulty state" at each point in time within the current month. Then, the belief quality-uncertainty sequence for each month is input into a time series prediction model to predict the corresponding result, namely: "In the next month or the next half month, the academic difficulty state of the student S will continue to rise rapidly and exceed the preset strategy value (critical value, such as ≥0.7) after one week, and the prediction uncertainty is low." Correspondingly, based on the prediction result, an early warning email containing the prediction curve and key data can be sent to student S's academic tutor, and / or, relevant learning materials and links to schedule tutor appointments can be automatically pushed to the homepage of the learning portal for student S in the university intelligent system 100 based on cloud-native and multimodal data fusion.

[0072] In some embodiments, the microservice business platform 104 can transform intelligent analysis results into implementable university management services, achieving a connection between technical capabilities and business needs. This microservice business platform 104 can encapsulate core intelligent capabilities into independent services through a microservice architecture, including: intelligent course scheduling service, student growth early warning service, and smart energy-saving service, among other university-specific services. Specifically, the intelligent course scheduling service can integrate historical course selection data, classroom resources, and teacher preferences to generate the optimal course scheduling plan; the student growth early warning service can build student profiles based on multimodal data to achieve early identification of academic and psychological risks; and the smart energy-saving service combines timetables, weather forecasts, and real-time pedestrian flow data to automatically adjust air conditioning / lighting switches.

[0073] In this application, to achieve efficient management and flexible invocation of services, the microservice business platform 104 manages microservice interfaces in a unified manner through an application programming interface (API) gateway, supporting front-end and back-end decoupling and service composition; at the same time, it provides a visual low-code orchestration interface, allowing university administrators to configure business flows by dragging and dropping operators, and adjust business logic without professional development capabilities, greatly reducing the threshold for university use.

[0074] In some embodiments, continue to refer to Figure 4As shown, the university intelligent system 100 based on cloud-native and multimodal data fusion also includes: The unified service interaction layer 107 is used to set up a customized cockpit carrying the core information of the test object based on the role of the test object in the university campus scene.

[0075] It also responds to the user's access to the customized cockpit, displays the user's profile in the university campus scene in real time, and provides personalized services for the user.

[0076] In some embodiments, the unified service interaction layer 107 can serve as the interaction entry point between the cloud-native and multimodal data fusion-based intelligent university system 100 and users. Its core objective is to provide a personalized and convenient service experience. Customized dashboards can be designed for different roles such as students, teachers, and administrators in universities, displaying core information relevant to each role. For example, the student side displays grades, timetables, and to-do items; the teacher side displays teaching tasks, student attendance, and course evaluation data; and the administrator side presents operational information such as system load, number of warnings, and resource usage. Simultaneously, multi-terminal responsive access is supported, such as on personal computers (PCs) and mobile apps / mini-programs, ensuring users can conveniently use system functions in different scenarios and improving the overall user experience.

[0077] Based on the above description, such as Figure 5 As shown, the overall layered architecture of a university intelligent system 100 based on cloud-native and multimodal data fusion, as well as the data flow and control flow between each layer, are presented. The architecture is divided into five layers from top to bottom: cloud-native infrastructure layer, multimodal data acquisition layer, data fusion and intelligent analysis layer, microservice business platform, and unified service interaction layer. This architecture reflects that the university intelligent system 100 based on cloud-native and multimodal data fusion provided in this application is designed with full-stack cloud-native and intelligent features from underlying resources to top-level applications; among which: Cloud-native infrastructure layer: Built on the Kubernetes container orchestration platform, it provides elastic and scalable computing and network resources; internally it involves core operations or data such as persistent storage, container orchestration, service networking, and monitoring logs.

[0078] Multimodal data acquisition layer: Used to access and preprocess structured and unstructured data from university academic affairs, IoT devices, network behavior, etc.; that is, receiving external data sources, including: university campus card / access control, academic affairs / student affairs systems, surveillance cameras, sensor / network logs, etc. Its core internal components revolve around technologies such as message queues, API interfaces, log collection, and IoT gateways to enable processing of the acquired external data sources.

[0079] Data fusion and intelligent analysis layer: It can use machine learning models to fuse and analyze multi-source heterogeneous data, generate intelligent decision support signals, and may further involve machine learning platforms and integrated data lake warehouse technology.

[0080] Microservice business platform: It encapsulates the analysis results into intelligent services that can be deployed and expanded independently; it can provide different intelligent services for campus service objects, such as: smart logistics services, smart academic affairs services, personalized student services, and data analysis services.

[0081] Unified Service Interaction Layer: By leveraging the API gateway, the intelligent services provided by the microservice business platform can be displayed on a personalized interactive interface for users, and multi-terminal responsive access can be supported, such as web front-end, mobile terminal, etc.

[0082] Thus, this application, through the synergistic effect of its five-layer architecture, brings significant multi-dimensional benefits to university management, comprehensively improving management efficiency and service quality. In terms of scalability, the elastic scaling capabilities of the cloud-native architecture support stable system operation during peak course selection periods (increased concurrency), solving the problem of insufficient high-concurrency handling capacity of traditional architectures. Regarding data value release, multimodal data fusion technology breaks down data silos, increasing data utilization by over 40%, providing ample data support for intelligent decision-making. In terms of intelligent upgrades, the combination of deep evidence uncertainty fusion algorithms improves early warning accuracy, providing precise evidence for teaching evaluation and student management. Regarding operational cost control, DevOps automated operations reduce manual intervention, improve resource utilization, and significantly reduce university operational costs. In terms of user experience optimization, the personalized design of the unified portal shortens the average processing time for faculty and students, thereby improving service convenience and satisfaction.

[0083] This application provides a cloud-native and multimodal data fusion-based intelligent system for universities. The system includes: a multimodal data acquisition layer, used to acquire multimodal behavioral data sequences representing the behavior of a target object in a specific behavioral state within the current time period using a campus-wide data acquisition platform; wherein the campus-wide data acquisition platform is a cloud-native infrastructure-based acquisition platform; thus, real-time acquisition and aggregation of multi-source heterogeneous data across the campus can be achieved through the cloud-native infrastructure-based acquisition platform, ensuring the comprehensiveness and timeliness of the acquired multimodal behavioral data; and a data fusion and intelligent analysis layer, used to analyze the multimodal behavioral data sequences using a deep evidence uncertainty model to obtain the belief quality and uncertainty supporting the target object's behavior in a specific behavioral state at each time point within the current time period; wherein the deep evidence uncertainty model is constructed based on Dirichlet distribution and subjective logic theory. The model is trained and updated using a replay-based incremental learning strategy. This deep evidence uncertainty model enables the fusion analysis of multimodal behavioral data, outputting refined and interpretable belief levels for specific behavioral states. It explicitly provides a measure of uncertainty for each judgment, offering crucial information for subsequent risk assessment and decision-making. The microservice business platform, based on the belief quality and uncertainty of the target object at each point in time within the current period, predicts the belief trend of the target object in a specific behavioral state in the future, and manages the target object in the future based on this belief trend. Thus, it is possible to proactively predict the behavioral trend of the target object in the future based on the belief quality and uncertainty sequence of the current period, thereby shifting from post-event analysis to pre-event warning, enabling personalized and precise intervention measures for the target object. In this way, by collecting data across the entire domain based on cloud-native technology, uncertainty is quantified, and based on the quantified uncertainty, accurate prediction and flexible management are achieved. This not only greatly improves the level of refinement, intelligence and personalization of campus management, enabling dynamic scheduling of resources, data-driven decision-making and personalized services within university campuses, but also enhances the robustness of the system in dealing with complex campus scenarios by leveraging its inherent uncertainty quantification capabilities and continuous learning capabilities.

[0084] The above-mentioned intelligent system for universities based on cloud-native and multimodal data fusion will be described below with reference to a specific embodiment. However, it is worth noting that this specific embodiment is only for better illustration of this application and does not constitute an improper limitation of this application.

[0085] Taking xxx College as an application scenario, this application aims to achieve intelligent upgrade of university management through the implementation of four stages: deployment, data access, service development, and operation and maintenance.

[0086] During the deployment preparation phase: A Kubernetes-based containerized platform was built in the campus data center of xxx College. The corresponding cluster configuration included 3 master nodes (responsible for cluster management) and 10 worker nodes (running business workloads). Istio service mesh was installed for microservice governance, Prometheus was integrated for system monitoring, Grafana was used for data visualization, and GitLab CI / CD tools were used for automated deployment.

[0087] During the data access implementation phase: multiple methods were employed to achieve full data access: the academic affairs / student affairs database was synchronized to the data lake by changing the data capture tool to ensure that the structured data was updated in real time; more than 200 IoT sensors were deployed to cover classrooms, dormitories, and office buildings of xxx College, and environmental data such as temperature, humidity, and light intensity were accessed in real time via the MQTT protocol, with a corresponding sampling frequency of once per minute; in addition, log data from xxx College's card consumption system, library access control system, and campus network gateway were integrated and incrementally synchronized to the data lake warehouse on an hourly basis through the Flume log collection framework.

[0088] During the intelligent service development and application phase: Three core services were implemented to meet the needs of xxx College (intelligent course scheduling service, student growth early warning service, and intelligent energy-saving service). The intelligent course scheduling service is based on historical course selection data (data from the past 3 years, approximately 2 million records), classroom resource information (more than 500 classrooms), teacher preference constraints, and other inputs. It uses a genetic algorithm to optimize the solution, achieving a 60% improvement in course scheduling efficiency and reducing the course conflict rate to below 0.5%.

[0089] The student growth early warning service is based on a deep evidence uncertainty model, which integrates multimodal data such as academic performance, consumption behavior (average daily consumption amount, distribution of consumption locations), library usage (number of visits, types of books borrowed), and online activities (online learning time, game time) to establish a student behavior anomaly detection model, achieving an early warning accuracy rate of 89% and identifying multiple students with academic or psychological risks at an early stage.

[0090] The smart energy-saving service combines timetable information, weather forecast data, and real-time pedestrian flow monitoring to establish an energy consumption prediction model based on time series analysis. It automatically adjusts the classroom air conditioning temperature and lighting switches, reducing classroom energy consumption by 25% and saving approximately 150,000 yuan in electricity costs annually.

[0091] During the operation and delivery phase: GitLab CI pipeline automates the entire process of "code submission → automated testing → deployment", enabling dozens of microservice deployments per day and reducing manual operations; Grafana dashboard displays real-time metrics such as system CPU / memory usage, service response time, and data access volume, and sets up intelligent alarm rules to automatically trigger email and SMS notifications in case of anomalies, ensuring stable system operation.

[0092] Based on the above description, the intelligent management system for universities based on cloud-native and multimodal data fusion provided in this application, on the basis of a five-layer architecture consisting of a cloud-native infrastructure layer, a multimodal data acquisition layer, a data fusion and intelligent analysis layer, a microservice business platform, and a unified service interaction layer to achieve intelligent management of universities, can further adopt a Kubernetes container orchestration platform and an Istio service mesh to achieve automatic elastic scaling, service governance, and secure communication; it can also adopt a data lake warehouse integrated storage and integrate a machine learning platform based on a deep evidence uncertainty model and incremental learning algorithm to handle multimodal data conflicts and dynamic model updates. Correspondingly, it can provide intelligent scheduling services, student growth early warning services, and smart energy-saving services specifically for university management, and manage them uniformly through an API gateway, while providing a visual low-code orchestration interface for configuring business logic.

[0093] Correspondingly, as shown in Table 1 below, the technologies used in the intelligent university system based on cloud-native and multimodal data fusion provided in this application embodiment are compared with the advantages of existing technologies: Table 1 Comparison of Technical Advantages The intelligent management system for universities based on cloud-native and multimodal data fusion provided in this application, through the deep integration of deep evidence uncertainty model and incremental learning algorithm, is no longer a simple and rigid data processor, but an intelligent agent that can perceive its own cognitive limitations, continuously learn and evolve, and reliably handle the complexity and uncertainty of the real world.

[0094] It should be understood that the phrase "one embodiment" or "an embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "in one embodiment" or "in an embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. It should be understood that in the various embodiments of this application, the sequence numbers of the above-described processes do not imply a sequential order of execution; the execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application. The sequence numbers of the above-described embodiments are merely descriptive and do not represent the superiority or inferiority of the embodiments.

[0095] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0096] In the several embodiments provided in this application, it should be understood that the disclosed system can be implemented in other ways. The system 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, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, and can be electrical, mechanical, or other forms.

[0097] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units. They may be located in one place or distributed across multiple network units. Some or all of the units may be selected to achieve the purpose of the embodiments of this application, depending on actual needs.

[0098] In addition, each functional unit in the various embodiments of this application can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

[0099] Alternatively, if the integrated units described above are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this application, or the parts that contribute to related technologies, 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 the device automatic test line to execute all or part of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROMs, magnetic disks, or optical disks.

[0100] The features disclosed in the several system embodiments provided in this application can be arbitrarily combined without conflict to obtain new system embodiments.

[0101] The above description is merely an embodiment of this application, but the scope of protection of this application 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 application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A university intelligent system based on cloud-native and multimodal data fusion, characterized in that, The system includes: The multimodal data acquisition layer is used to acquire multimodal behavioral data sequences representing the specific behavioral states of the test object within the current time period using the campus-wide data acquisition platform; the campus-wide data acquisition platform is an acquisition platform built on cloud-native infrastructure. The data fusion and intelligent analysis layer is used to analyze multimodal behavioral data sequences using a deep evidence uncertainty model to obtain the quality and uncertainty of beliefs supporting the test subject's specific behavioral state at each time point within the current time period. The deep evidence uncertainty model is constructed based on the Dirichlet distribution and subjective logic theory, and is trained and updated using a playback-based incremental learning strategy. The microservice business platform is used to predict the belief trend of the test object in a specific behavioral state in the future time period based on the belief quality and uncertainty of the test object at each point in time within the current time period, and to manage the test object in the future time period according to the belief trend.

2. The system according to claim 1, characterized in that, The data fusion and intelligent analysis layer includes: The evidence vector generation module is used to extract features from each modal behavior data in the multimodal behavior data sequence according to each time point within the current time period, and obtain the evidence vector set for each time point within the current time period. The multimodal data conflict processing module is used to utilize the deep evidence uncertainty model and, based on subjective logic theory, to transform the evidence vector set at each time point in the current time period into Dirichlet distribution parameters, thereby obtaining the Dirichlet concentration parameters at each time point in the current time period that support the test object being in a specific behavioral state. The Belief Quality and Uncertainty Calculation Module is used to input the Dirichlet concentration parameters supporting the test subject's specific behavioral state at each time point within the current time period into the belief quality and uncertainty calculation formula determined based on subjective logic theory, and calculate the belief quality and uncertainty supporting the test subject's specific behavioral state at each time point within the current time period.

3. The system according to claim 2, characterized in that, The multimodal data conflict processing module is specifically used to utilize the deep evidence uncertainty model to assign weight values ​​to each evidence vector in the evidence vector set at each time point within the current time period, based on the weight of each evidence vector in the evidence vector set, thereby obtaining the weight of each evidence vector in the evidence vector set. Furthermore, the evidence vectors in the evidence vector set are weighted and fused according to their respective weights to obtain the evidence value supporting the test object being in a specific behavioral state at each time point within the current time period. Based on subjective logic theory, the evidence value is converted into the Dirichlet concentration parameter supporting the test object being in a specific behavioral state at each time point within the current time period.

4. The system according to claim 3, characterized in that, The multimodal data conflict handling module is also used to determine the reliability of the data source corresponding to each evidence vector in the evidence vector set for each time point within the current time period as the weight of each evidence vector in the evidence vector set.

5. The system according to claim 1, characterized in that, The system also includes: The model initialization layer is used to build a network model whose output layer is configured as the evidence layer based on the Dirichlet distribution and subjective logic theory. It also uses the initial multimodal training dataset corresponding to the campus-wide data to train the network model using the evidence loss function, thus obtaining a deep evidence uncertainty model. The model update layer is used for incremental learning based on the replay buffer pool to adjust the network parameters of the deep evidence uncertainty model and obtain the updated deep evidence uncertainty model.

6. The system according to claim 5, characterized in that, The model update layer is specifically used to respond to new multimodal data, select initial multimodal training sub-data from the initial multimodal training dataset, and incrementally learn the deep evidence uncertainty model based on the new multimodal data and the initial multimodal training sub-data using an overall loss function that combines the new and old knowledge retention mechanism, so as to obtain the updated deep evidence uncertainty model. The overall loss function is: in, Total loss; This is the balance coefficient; The prediction loss of the deep evidence uncertainty model on new multimodal data; This represents the prediction loss of the deep evidence uncertainty model on the initial multimodal training subdata.

7. The system according to claim 1, characterized in that, The microservice business middle platform is specifically used to infer the belief trend of the test object in a specific behavioral state in the future time period based on the belief quality and uncertainty of the test object at each time point in the current time period. The test object is managed based on the comparison between the belief trend and the preset strategy value.

8. The system according to claim 1, characterized in that, The system also includes: A unified service interaction layer is used to set up a customized cockpit carrying the core information of the test object based on the role of the test object in the university campus scene; It also responds to the user's access to the customized cockpit, displays the user's profile in the university campus scene in real time, and provides personalized services for the user.

Images