A heterogeneous resource adaptive scheduling method and system fusing topology feature enhancement and constraint decision

Abstract

The present application belongs to the technical field of intelligent education resource scheduling and heterogeneous storage resource optimization, and specifically relates to a heterogeneous resource adaptive scheduling method and system combining topology feature enhancement and constraint decision. The method comprises: bidirectional feature modeling of the heterogeneous resource and the physical execution layer to obtain a resource feature matrix and a hardware feature matrix; constructing a resource topology graph oriented to teaching space-time correlation to perform feature enhancement on the resource feature matrix to obtain a high-order resource feature matrix; performing supply and demand feature tensor alignment and adaptive matching based on bidirectional attention on the high-order resource feature matrix and the hardware feature matrix to obtain a global scheduling adaptation degree matrix; based on the global scheduling adaptation degree matrix, combining multi-dimensional hard constraint preposition, performing adaptive discrete decision and physical control execution to realize adaptive scheduling of the heterogeneous resource. The present application realizes end-to-end closed-loop scheduling, and significantly improves the rationality of intelligent teaching resource scheduling and the stability of the underlying storage system.

Description

Technical Field

[0001] This invention belongs to the field of intelligent education resource scheduling and heterogeneous storage resource optimization technology, specifically involving a heterogeneous resource adaptive scheduling method and system that integrates topology feature enhancement and constraint decision-making. Background Technology

[0002] With the ongoing digital transformation of smart education, cloud-based smart teaching platforms have gradually formed multi-source, heterogeneous, and highly dynamic teaching resource scheduling scenarios. The teaching resources carried by these platforms encompass various types of heterogeneous data, including high-definition teaching videos, structured courseware, real-time interactive teaching instructions, and in-class assessment data. These resources differ significantly in data volume, access latency requirements, read / write characteristics, teaching sequence constraints, and spatial distribution features. Simultaneously, the underlying storage media includes various heterogeneous hardware such as high-speed memory cache, solid-state drives, and hard disk drives, forming a two-layer heterogeneous scheduling environment with heterogeneous upper-layer business processes and heterogeneous lower-layer hardware. Against this backdrop, achieving precise matching, efficient placement, and dynamic adaptive scheduling of massive heterogeneous teaching resources and multi-layered heterogeneous storage media has become a core technical challenge for improving the response speed, resource hit rate, and load balancing capabilities of smart teaching platforms.

[0003] Existing intelligent teaching resource scheduling solutions mostly employ traditional static caching strategies, greedy algorithms, or single-feature matching scheduling mechanisms. They generally rely on manual experience thresholds and fixed mapping rules, and can only complete resource selection and storage allocation based on shallow physical characteristics such as resource size and access frequency, exhibiting significant technical limitations. First, existing technologies lack the ability to finely perceive the deep business semantics and temporal-spatial relationships of teaching resources. They cannot uncover the classroom temporal connection relationships, knowledge point relationships, and campus spatial distribution relationships between different teaching resources, ignoring the unique temporal and spatial constraints of teaching business. This results in low adaptability of resource scheduling to the actual teaching rhythm, easily leading to problems such as delayed response to emergency teaching resources and scattered storage of related resources.

[0004] Secondly, traditional scheduling schemes lack a cross-domain alignment mechanism between business characteristics and underlying hardware constraints. They treat resource characteristics, teaching constraints, and hardware performance in isolation, relying solely on a single-dimensional indicator to select storage media. This fails to achieve a tensor-level precise match between teaching business needs and heterogeneous hardware storage capabilities, easily leading to wasted high-speed media resources and congestion on low-speed media, resulting in poor system storage-scheduling coordination and adaptability. Furthermore, existing placement decisions often employ deterministic hard decision mechanisms, neglecting the dynamic changes in real-time media load, remaining capacity, and hardware latency fluctuations. This results in weak anti-interference capabilities, making them prone to localized load imbalances and scheduling jitter in high-concurrency teaching access scenarios.

[0005] More critically, existing intelligent teaching scheduling systems are mostly open-loop scheduling architectures, capable of only completing a one-way scheduling process of feature extraction, resource placement, and physical execution, lacking a complete feedback, error correction, and self-evolutionary optimization mechanism. Traditional algorithms cannot collect actual measured data such as time consumption, media load, and cache hit rate in real physical execution scenarios, cannot quantify the deviation between the scheduling results and the actual execution effect, and are difficult to iteratively optimize front-end feature extraction, topology modeling, and cross-domain matching parameters through back gradient propagation. As a result, after long-term operation, the system cannot adapt to dynamically changing teaching business traffic and hardware aging characteristics, and the scheduling accuracy, response efficiency, and stability continue to decline, with severely insufficient intelligent adaptive capabilities.

[0006] In summary, existing intelligent teaching resource scheduling technologies suffer from core defects such as limited heterogeneous resource cognition, lack of spatiotemporal topological correlation, insufficient cross-domain adaptation accuracy, poor dynamic load adaptability, and lack of closed-loop self-evolution capability. These shortcomings make it difficult to meet the efficient scheduling requirements of large-scale, high-concurrency, and time-constrained intelligent teaching scenarios. Summary of the Invention

[0007] To address the problems of low matching accuracy between heterogeneous teaching resources and hardware devices, lack of spatiotemporal collaborative association, difficulty in obtaining real scheduling labels, inability to implement intelligent matching results in engineering, and unbalanced hardware load in existing smart teaching scenarios, this invention provides a heterogeneous resource adaptive scheduling method and system that integrates topological feature enhancement and constraint decision-making. It aims to complete the standardization of basic scheduling data through multi-source heterogeneous feature fusion and unified feature space alignment; construct spatiotemporal correlation topological modeling and feature enhancement of teaching resources to achieve technical mining of collaborative relationships among resource groups; build a supervised training system based on prior constraints through supply and demand feature tensor alignment and bidirectional attention adaptive matching to output quantitative adaptation results; and combine hardware operation hard technical constraints to complete dynamic load balancing scheduling decision output, ultimately achieving an automated resource scheduling process that balances intelligent matching accuracy, hardware operation feasibility, and global load balancing.

[0008] To achieve the above objectives, the present invention provides the following solution:

[0009] A heterogeneous resource adaptive scheduling method integrating topology feature enhancement and constraint decision-making, the method comprising:

[0010] Step 1: Perform bidirectional feature modeling on heterogeneous resources and physical execution layer in the smart teaching scenario to obtain resource feature matrix and hardware feature matrix;

[0011] Step 2: Construct a resource topology map oriented towards the spatiotemporal correlation of teaching, and perform feature enhancement on the resource feature matrix to obtain a high-order resource feature matrix;

[0012] Step 3: Perform supply and demand feature tensor alignment and adaptive matching based on bidirectional attention on the high-order resource feature matrix and the hardware feature matrix to obtain the global scheduling adaptability matrix;

[0013] Step 4: Based on the global scheduling adaptability matrix and combined with multi-dimensional hard constraints, perform adaptive discrete decision-making and physical control execution to achieve adaptive scheduling of heterogeneous resources.

[0014] Preferably, the method for performing bidirectional feature modeling on heterogeneous resources and the physical execution layer in the smart teaching scenario in step 1 to obtain the resource feature matrix and the hardware feature matrix includes:

[0015] Obtain metadata description text of heterogeneous resources in smart teaching scenarios, call preset business semantic concatenation parsing operators to perform deep word segmentation and semantic representation encoding, and generate a semantic feature matrix of resources;

[0016] Extract the applicable physical parameters for each heterogeneous resource, normalize them, and construct the physical feature matrix of the resource;

[0017] The semantic and physical features of the resource are cross-domain concatenated and spatially aligned using the first projection weight matrix to generate an initial comprehensive feature vector for a single resource, which constitutes the resource feature matrix.

[0018] Simultaneously, through the device underlying driver interface of the physical execution layer and the kernel performance monitoring operator, the underlying physical performance parameters of heterogeneous hardware are collected in real time to construct hardware physical feature vectors;

[0019] Read the hardware device attribute configuration table to generate standardized hardware business description text, and call the business semantic concatenation parsing operator to process and obtain the semantic feature vector of the hardware;

[0020] The physical and semantic features of the hardware are spatially aligned and stitched together using the second projection weight matrix to generate an initial comprehensive feature vector for a single hardware component, thus constructing a hardware feature matrix.

[0021] Preferably, the method for constructing a resource topology map oriented towards the spatiotemporal correlation of teaching in step 2, and performing feature enhancement on the resource feature matrix to obtain a higher-order resource feature matrix includes:

[0022] Using all heterogeneous resources as network nodes, the association weights between any two heterogeneous resources are calculated from three dimensions: content semantics, teaching time sequence, and physical deployment space, and then integrated to generate a comprehensive association weight.

[0023] Based on the comprehensive association weights and combined with the preset association thresholds, effective association relationships are screened to construct a resource topology map oriented towards the spatiotemporal relevance of teaching.

[0024] The resource topology map is subjected to feature aggregation learning and spatial reconstruction. The attributes of individual resources are integrated into the global spatiotemporal correlation constraints of neighboring resources. The resource feature matrix is ​​then enhanced to obtain a high-order resource feature matrix.

[0025] Preferably, the method for obtaining the global scheduling fit matrix by performing supply and demand feature tensor alignment and adaptive matching based on bidirectional attention on the high-order resource feature matrix and the hardware feature matrix in step 3 includes:

[0026] The high-order resource feature matrix and the hardware feature matrix are uniformly mapped to the same matching feature space to achieve dimensional alignment of the feature spaces on both the supply and demand sides, and the aligned feature matrix is ​​output.

[0027] Based on the fine-grained association mining feature mapping matrix, cross-domain interaction modeling is performed on the aligned feature matrix to generate a query matrix, a key matrix, and a value matrix;

[0028] The query matrix and key matrix are scaled and normalized to generate an adaptive joint control weight matrix;

[0029] The adaptive joint control matrix and the value matrix are weighted and fused to generate the final global scheduling adaptability matrix.

[0030] Preferably, the method for adaptive scheduling of heterogeneous resources in step 4, based on the global scheduling adaptability matrix and combined with multi-dimensional hard constraints, to perform adaptive discrete decision-making and physical control execution, includes:

[0031] The hardware resource carrying capacity limit and resource exclusivity constraint are used as rigid preconditions to perform feasible domain pruning and filtering on the global scheduling adaptability matrix.

[0032] Based on the global scheduling adaptability matrix after feasible region pruning and filtering, construct the optimal route candidate queue;

[0033] Based on the optimal route candidate queue, an adaptive greedy decoding strategy is used to initially discretize and bind resources and hardware to generate an initial decision matrix.

[0034] Under the premise of satisfying rigid preconditions, with the optimization objective of minimizing the variance of each hardware load, the inverse variance fine-tuning state machine is invoked to fine-tune the initial discretization binding, and the final binary scheduling decision matrix is ​​obtained.

[0035] The physical enable flag in the binary scheduling decision matrix is ​​read, and the physical enable flag is converted into a corresponding hardware storage and network routing control command and sent to the physical execution layer. This drives the corresponding heterogeneous resource data stream to be written into the physical sector of the target hardware storage medium, thus completing the dynamic adaptive collaborative scheduling at the physical layer.

[0036] Preferably, the method further includes:

[0037] The bilateral projection weight matrix, bias vector, and attention weight parameters used in the bidirectional feature modeling in step 1 and the supply and demand feature tensor alignment and adaptive matching based on bidirectional attention in step 3 are adaptively optimized.

[0038] Preferably, the method for adaptively optimizing the bilateral projection weight matrix, bias vector, and attention weight parameters includes:

[0039] Based on historical operation logs, hardware security operation constraints, and load limit indicators, hardware performance adaptation score, resource demand matching score, and hardware load balancing score are calculated and weighted and fused to obtain a pseudo-true value control matrix.

[0040] A supervised loss function is constructed by measuring the error between the global scheduling fit matrix and the pseudo-truth control matrix.

[0041] The bilateral projection weight matrix, bias vector, and attention weight parameters are adaptively optimized based on the supervised loss function.

[0042] The present invention also provides a heterogeneous resource adaptive scheduling system that integrates topology feature enhancement and constraint decision-making. The system is used to implement the aforementioned method and includes: a bidirectional feature modeling module, a topology feature enhancement module, a scheduling matching module, and a constraint adaptive decision-making module.

[0043] The bidirectional feature modeling module is used to perform bidirectional feature modeling on heterogeneous resources and physical execution layer in smart teaching scenarios to obtain resource feature matrix and hardware feature matrix.

[0044] The topology feature enhancement module is used to construct a resource topology map oriented towards the spatiotemporal correlation of teaching, and to enhance the resource feature matrix to obtain a higher-order resource feature matrix.

[0045] The scheduling matching module is used to perform supply and demand feature tensor alignment and adaptive matching based on bidirectional attention on the high-order resource feature matrix and the hardware feature matrix to obtain a global scheduling fit matrix.

[0046] The constraint adaptive decision-making module is used to perform adaptive discrete decision-making and physical control execution based on the global scheduling adaptability matrix and combined with multi-dimensional hard constraints in advance, so as to realize the adaptive scheduling of heterogeneous resources.

[0047] Preferably, the system further includes a closed-loop iterative optimization module;

[0048] The closed-loop iterative optimization module is used to adaptively optimize the bilateral projection weight matrix, bias vector, and attention weight parameters used in the process of bidirectional feature modeling by the bidirectional feature modeling module and the process of supply and demand feature tensor alignment and adaptive matching based on bidirectional attention by the feature alignment and matching module.

[0049] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0050] This invention proposes a heterogeneous resource adaptive scheduling method and system that integrates topology feature enhancement and constraint decision-making. Through a seven-order closed-loop architecture, including multi-dimensional resource profiling perception, dynamic spatiotemporal topology modeling, tensor alignment coupling mapping, probabilistic balanced placement decision-making, standardized instruction parsing, physical execution data acquisition, and full-link gradient review optimization, it achieves accurate, dynamic, adaptive, and collaborative closed-loop self-evolutionary intelligent scheduling of heterogeneous teaching resources and heterogeneous storage media. This effectively solves the technical problems of low scheduling accuracy, load imbalance, weak adaptability, and inability to autonomously iteratively optimize in traditional technologies. Attached Figure Description

[0051] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0052] Figure 1 This is a schematic diagram of the heterogeneous resource adaptive scheduling method that integrates topology feature enhancement and constraint decision-making according to an embodiment of the present invention. Detailed Implementation

[0053] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0054] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0055] Example 1

[0056] This invention provides a heterogeneous resource adaptive scheduling method that integrates topology feature enhancement and constraint decision-making, comprising:

[0057] Step 1: Perform bidirectional feature modeling on heterogeneous resources and physical execution layer in the smart teaching scenario to obtain resource feature matrix and hardware feature matrix;

[0058] Step 2: Construct a resource topology map oriented towards the spatiotemporal correlation of teaching, and perform feature enhancement on the resource feature matrix to obtain a high-order resource feature matrix;

[0059] Step 3: Perform supply and demand feature tensor alignment and adaptive matching based on bidirectional attention on the high-order resource feature matrix and the hardware feature matrix to obtain the global scheduling adaptability matrix;

[0060] Step 4: Based on the global scheduling adaptability matrix and combined with multi-dimensional hard constraints, perform adaptive discrete decision-making and physical control execution to achieve adaptive scheduling of heterogeneous resources.

[0061] like Figure 1 As shown, the specific implementation process of the present invention is as follows:

[0062] The entire heterogeneous resource adaptive scheduling system implementing this invention is deployed at the bottom layer of the smart teaching platform and includes five serially linked functional modules:

[0063] (1) Bidirectional feature modeling module (step S1): Extract semantic features and physical features of teaching resources and hardware nodes in parallel to complete standardized fusion representation;

[0064] (2) Topology feature enhancement module (step S2): Construct a resource spatiotemporal correlation topology graph and generate a high-order resource feature matrix through GCN graph convolution;

[0065] (3) Scheduling matching module (step S3): bidirectional attention tensor alignment, unification of feature dimensions and calculation of the global scheduling fit matrix of resources and hardware;

[0066] (4) Constraint Adaptive Decision Module (Step S4): Based on the global scheduling adaptability matrix superimposed with hardware capacity and resource exclusive rigid constraints, iteratively generate a discrete scheduling allocation scheme for balanced load.

[0067] (5) Closed-loop iterative optimization module (step S5): Collect scheduling operation feedback to construct supervision loss, backpropagate to update the full-link model parameters, and realize the self-evolution optimization of the heterogeneous resource adaptive scheduling system.

[0068] S1: Bidirectional Feature Modeling Based on Multi-Source Heterogeneous Resources and Physical Execution Layer: Bidirectional feature modeling is performed on heterogeneous resources and the physical execution layer in a smart teaching scenario to obtain resource feature matrices and hardware feature matrices, including:

[0069] Obtain metadata description text of heterogeneous resources in smart teaching scenarios, call preset business semantic concatenation parsing operators to perform deep word segmentation and semantic representation encoding, and generate a semantic feature matrix of resources;

[0070] Extract the applicable physical parameters for each heterogeneous resource, normalize them, and construct the physical feature matrix of the resource;

[0071] The semantic and physical features of the resource are cross-domain concatenated and spatially aligned using the first projection weight matrix to generate an initial comprehensive feature vector for a single resource, which constitutes the resource feature matrix.

[0072] Simultaneously, through the device underlying driver interface of the physical execution layer and the kernel performance monitoring operator, the underlying physical performance parameters of heterogeneous hardware are collected in real time to construct hardware physical feature vectors;

[0073] Read the hardware device attribute configuration table to generate standardized hardware business description text, and call the business semantic concatenation parsing operator to process and obtain the semantic feature vector of the hardware;

[0074] The physical and semantic features of the hardware are spatially aligned and stitched together using the second projection weight matrix to generate an initial comprehensive feature vector for a single hardware component, thus constructing a hardware feature matrix.

[0075] This step aims to construct a dynamic feature extraction system with both semantic and physical perception capabilities. This system will enable precise characterization of the needs of heterogeneous teaching resources (referred to as heterogeneous resources / resources), including video streams, structured documents, and real-time interactive commands. Simultaneously, it will standardize the representation of the underlying physical execution layer's carrying capacity, providing a unified basis for subsequent cross-domain matching and resource placement. The specific execution steps are as follows:

[0076] First, perform a preliminary mapping of multimodal business semantic features. This involves accessing and capturing metadata description text from heterogeneous resource requests originating from the smart teaching cloud platform. This includes all original attribute information of heterogeneous teaching resources, such as resource type, teaching knowledge point tags, classroom time sequence attributes, teaching dependencies, file size, bitrate, estimated loading time, and storage format. A pre-trained language model (BERT model) is used as the underlying segmentation operator. Through BERT's multi-layer bidirectional Transformer structure, deep word segmentation and semantic encoding are performed on the resource metadata. In this process, the heterogeneous resource adaptive scheduling system proposed in this invention focuses on extracting keyword weights strongly correlated with scheduling attributes. For example, for video stream resources, it focuses on bandwidth-intensive descriptions, while for instruction streams, it focuses on low-latency interactive descriptions. Finally, the last hidden state vector of BERT is taken as the business semantic baseline representation vector of the resource (i.e., the semantic feature vector of the resource). Therefore, for each resource, a metadata description text is required. ,implement:

[0077]

[0078] in, For semantic encoding operators based on pre-trained bidirectional Transformer models, Let be the semantic feature vector of the i-th resource, and D be the total dimension of the BERT embedding space (768 when using BERT-base). This is a text description field containing resource business attributes (i.e., metadata description text). The semantic feature vectors of all resources are stacked row-wise to form a semantic feature matrix for the resource. , where N is the total number of resources in the current scheduling period.

[0079] Synchronization involves extracting key physical parameters of resources at the underlying execution layer. To achieve dimensional uniformity of physical characteristics across different resource types, this step defines a global set of key physical parameters, including average bitrate of video, bandwidth fluctuation rate, sustained I / O length, response latency limits for interactive commands, transmission frequency, file size of structured documents, and access popularity. For each resource, a pre-defined data preprocessing module extracts applicable physical parameters and fills inapplicable parameters with zero values, ultimately forming a physical feature vector of the resource with a unified dimension of G. ,in, This represents the physical feature vector of the i-th resource. The physical performance indicators of the resources include, but are not limited to, file size, media bitrate, estimated loading time limit, storage format encoding, and concurrent access popularity of heterogeneous teaching resources. G represents the number of key physical parameters. Each component is extracted and then normalized to eliminate dimensional differences. The physical feature vectors are stacked row-wise to form the physical feature matrix of the resources. .

[0080] Next, spatial alignment and unified mapping of cross-domain features are performed. For the i-th resource, the heterogeneous resource adaptive scheduling system utilizes the first projection weight matrix. The semantic and physical feature vectors of a resource are concatenated and spatially aligned across domains, and then projected onto a unified normalized scheduling feature space using the nonlinear mapping function GeLU to generate an initial comprehensive feature vector for each resource. This eliminates dimensional biases between different attributes, achieves deep coupling between business intent and physical constraints, and unifies the mapping of heterogeneous resources to the same high-dimensional feature space. Specifically, it mainly involves concatenating semantic feature vectors and physical feature vectors, and then mapping them to a unified scheduling feature space through a trainable projection weight matrix and a nonlinear activation function to generate an initial comprehensive feature vector for each resource, i.e.:

[0081]

[0082]

[0083] in, This is a tensor splicing operation; This is the intermediate state vector after linear projection following cross-domain splicing. The standard Gaussian cumulative distribution function, and its intermediate state vector Multiplication together constitutes the complete expression of the GeLU function. This is the trainable first projection weight matrix, used to project the concatenated vectors onto a unified scheduling feature space. To unify the scheduling feature dimensions; This is the bias term. It is used to generate the initial composite feature vector of all independent resources within the current scheduling period. Subsequently, the heterogeneous resource adaptive scheduling system vertically stacks and spatially reconstructs the feature vectors of N resources row by row, thereby ultimately forming a resource feature matrix under a unified scheduling space. .

[0084] During the initialization phase of the comprehensive feature vector... and Random initialization is performed using a zero-mean normal distribution; With the Sigmoid function The curves are highly similar in shape, but their slopes and curvatures deviate. Numerical fitting can be used to obtain... This meets the accuracy requirements for model training. During training, the heterogeneous resource adaptive scheduling system constructs the loss function for subsequent steps based on scheduling delay deviation and resource-hardware matching error. It iteratively updates the projection weight matrix and bias terms through the backpropagation algorithm, so that the feature mapping gradually converges to the optimal state, ultimately achieving deep coupling and precise alignment of semantic features and physical features in a unified scheduling space.

[0085] Secondly, to overcome the shortcomings of traditional scheduling models that rely solely on labels and ignore physical differences, this step simultaneously performs symmetrical feature modeling on the underlying M physical execution layers. The heterogeneous resource adaptive scheduling system collects the underlying physical performance parameters of the j-th hardware in batches and in real time through the device-level driver interface of the physical execution layer and the kernel performance monitoring operator. These parameters include access latency, available bandwidth, remaining capacity, concurrent processing capability, and read / write IOPS, thus constructing the hardware's physical feature vector. ,in, Let K represent the physical feature vector of the j-th hardware node, where K is the dimension of the unified key physical parameters. Represents the k-th physical performance metric of a hardware node, 1 For parameter dimensions that are inapplicable to different hardware, zero values ​​are used for padding to ensure that the physical feature vector dimensions of all hardware are uniform. Each component is extracted and then normalized to eliminate the influence of dimensional differences. It is important to note that the key physical parameter dimension K in the physical execution layer is strictly consistent in numerical setting with the number of key physical parameters G extracted from the aforementioned smart teaching resource side, i.e., G=K. This symmetry design aims to construct a complete supply-demand mapping set mathematically. The K-dimensional vector on the resource side represents the current teaching service's demand boundaries for physical performance such as storage space, throughput bandwidth, and access latency, while the K-dimensional vector on the hardware side symmetrically represents the supply capacity that the corresponding underlying storage medium can provide in the same dimension.

[0086] In addition, pre-trained BERT networks are used to process hardware business attribute text descriptions (such as hardware manufacturing type, business role, service priority, and suitability for teaching scenarios) to generate standardized hardware business description text. The semantic encoding operator of the aforementioned pre-trained bidirectional Transformer model is invoked. Process the text to obtain the semantic feature vector of the j-th hardware. :

[0087]

[0088] in, The service description text for the j-th hardware. D is the total dimension of the BERT embedding space.

[0089] Using the second projection weight matrix The physical feature vector and semantic feature vector of a single hardware component are concatenated and spatially aligned across domains to generate the initial comprehensive feature vector of the hardware. This fully characterizes its carrying capacity and business positioning. A detailed description follows:

[0090]

[0091]

[0092] in, This is a tensor splicing operation; This represents the intermediate state vector of the linear projection after the cross-domain feature concatenation of the j-th hardware node; The standard Gaussian cumulative distribution function, and its intermediate state vector Multiplication together constitutes the complete expression of the GeLU function. This is a trainable second projection weight matrix used to project the concatenated vectors onto a unified scheduling feature space; This is the bias term. GeLU possesses the characteristics of being continuously differentiable throughout and having smooth gradients without hard truncation. Therefore, in the physical execution layer, the same nonlinear activation function GeLU as in the resource layer is used to enhance feature representation capabilities. During the model initialization phase, and A zero-mean normal distribution is used for random initialization. During training, the heterogeneous resource adaptive scheduling system constructs a loss function based on scheduling delay deviation and resource-hardware matching error. The projection weight matrix and bias term are iteratively updated using the backpropagation algorithm, gradually converging the feature mapping to the optimal state. This ultimately achieves deep coupling and precise alignment of semantic and physical features within a unified scheduling space, as described in section S4. Subsequently, the combined feature vectors of the M hardware components are stacked row-wise to form a hardware feature matrix. , This matrix includes both the hardware's performance ceiling and the business scenario positioning, providing a unified representation basis for the accurate matching of subsequent resource needs and hardware supply.

[0093] In summary, this step completed standardized feature modeling of both heterogeneous teaching resources and underlying hardware devices. Through the fusion of semantic and physical features and trainable projection mapping, it achieved dimensional alignment and deep coupling of multi-dimensional heterogeneous features. This provides a unified and accurate foundation of basic features for subsequent S2 topology association mining and S3 supply-demand precise matching, thus establishing the technical basis for end-to-end intelligent scheduling.

[0094] S2: Dynamic Topology Modeling and Feature Enhancement of Resources Based on Spatiotemporal Relevance in Teaching: Constructing a resource topology graph based on the spatiotemporal relevance of teaching, and enhancing the resource feature matrix to obtain a high-order resource feature matrix, including:

[0095] Using all heterogeneous resources as network nodes, the association weights between any two heterogeneous resources are calculated from three dimensions: content semantics, teaching time sequence, and physical deployment space, and then integrated to generate a comprehensive association weight.

[0096] Based on the comprehensive association weights and combined with the preset association thresholds, effective association relationships are screened to construct a resource topology map oriented towards the spatiotemporal relevance of teaching.

[0097] The resource topology map is subjected to feature aggregation learning and spatial reconstruction. The attributes of individual resources are integrated into the global spatiotemporal correlation constraints of neighboring resources. The resource feature matrix is ​​then enhanced to obtain a high-order resource feature matrix.

[0098] This step builds upon the bilateral feature modeling results completed in stage S1 for subsequent processing. S1 has achieved individual feature representation of heterogeneous teaching resources and underlying hardware nodes, outputting a resource feature matrix and a hardware feature matrix under a unified scheduling space. The resource feature matrix is... This involves characterizing the semantic needs and physical operational characteristics of individual teaching resources. Furthermore, teaching resources exhibit strong business relevance, temporal sequence, and spatial deployment differences; isolated individual features cannot depict the collaborative scheduling constraints between resources. Therefore, this step requires topological modeling of the teaching resources, mining content associations, temporal associations, and spatial physical deployment associations between resources, constructing a dynamic resource topology network, and using graph convolutional neural networks to aggregate neighborhood features, achieving higher-order enhancement of resource features. This provides resource representations with globally related information for subsequent accurate supply and demand matching.

[0099] To this end, using all heterogeneous resources as network nodes, and integrating three dimensions—content semantic similarity, teaching sequence constraints, and physical deployment space constraints—the strength of the association between each pair of resources is calculated, effective associations are selected, and a dynamic weighted topology graph adapted to the teaching scheduling scenario is constructed, i.e., a resource topology graph is built. , where the graph node set , contains Each node corresponds one-to-one with a heterogeneous resource adaptive scheduling system. Each node's initial features directly reuse the initial comprehensive feature vector of a single resource output by S1, ensuring complete continuity of feature information before and after, and the overall node features correspond to the resource feature matrix. .

[0100] To comprehensively characterize the collaborative constraints of teaching resources, arbitrary resources are calculated from three dimensions: content semantics, teaching sequence, and physical space. The association weights with resource t are used to construct semantic association weights, temporal association weights, and spatial association weights, which are then integrated to obtain a comprehensive association weight.

[0101] (1) Content semantic association weight

[0102] Resource node feature similarity characterizes the degree of content association. Resources within the same teaching unit and similar business scenarios exhibit higher semantic feature similarity, indicating a stronger need for collaborative scheduling. Cosine similarity is used to calculate semantic association weights. :

[0103]

[0104] In the formula, Resources ,resource The initial comprehensive feature vector, The range of values ​​is The larger the value, the closer the semantic relationship of the resource content.

[0105] (2) Weight of teaching sequence

[0106] Teaching processes have fixed time constraints; the playback, loading, and interaction of resources follow a clear sequence. Resources that are closer in time are more likely to be coordinated and scheduled. An exponential decay function is used to construct temporal correlation weights. :

[0107]

[0108] In the formula, Resources ,resource The preset teaching execution timestamp, The total duration of a single complete teaching task. This is the timing decay coefficient, used to adjust the sensitivity of timing correlation. The smaller the time difference, the higher the timing correlation weight, aligning with the timing scheduling rules of teaching operations. This weight is used to guide the heterogeneous resource adaptive scheduling system to predict future I / O request flows and preload time-adjacent resources to the cache medium (SSD) in advance.

[0109] (3) Physical deployment spatial association weight

[0110] For resources deployed on different server nodes and network links, centralized scheduling of resources that are physically adjacent and deployed in the same environment can effectively reduce network transmission latency and improve scheduling efficiency. Spatial correlation weights are calculated based on the physical feature vectors of the resources. :

[0111]

[0112] In the formula, Normalized physical feature vector for stage S1 The spatial dimension sub-features obtained only record the spatial characteristics of the physical space where resources can be deployed, including static spatial attributes such as the physical server ID, network access segment, cluster deployment location, and link topology location. These are used to accurately characterize the proximity of the resources in their actual physical environment. The three types of association weights are normalized to eliminate dimensional differences, and then linearly weighted and fused to obtain the comprehensive association weights between resources.

[0113]

[0114] In the formula, These are weighting factors for semantic, temporal, and spatial dimensions, respectively, satisfying... The contribution ratio of each dimension of constraints is adaptively adjusted according to the teaching scenario.

[0115] To eliminate invalid weak associations, simplify the topology, and reduce subsequent computational complexity, an association threshold is set. After filtering based on the comprehensive correlation weights, the final dimension is obtained as follows: Resource topology adjacency matrix The filtering rules are as follows:

[0116]

[0117] That is, when the overall association weight of a resource is greater than or equal to the association threshold, the topological connection relationship and weight between nodes are retained; otherwise, it is determined to be an invalid association and the topological edge connection is disconnected.

[0118] Edge Incidence Matrix This is used to characterize the existence of topological connections between resource nodes, and is a binary association structure. First, the comprehensive association weights between resources are calculated through semantic, temporal, and spatial constraints. Then, by setting a preset association threshold Determine the valid connection status of edges and define binary edge association identifiers. Only strongly related topologies are retained, with the following specific rules:

[0119]

[0120] In the formula, when resources With resources The overall association weight is greater than or equal to the association threshold. hour, If two resources are determined to have a strong collaborative scheduling relationship, an effective topological edge is constructed; otherwise... If the association is deemed invalid, no topological edge connection is constructed. Finally, the graph node set is used to determine the final result. Edge Incidence Matrix Adjacency matrix of resource topology Collaboratively complete the construction of the dynamic spatiotemporal topology map of resources.

[0121] After the topology is constructed, the original resource features only contain the individual attributes of each node, without incorporating the association constraints of neighboring resources. A Graph Convolutional Neural Network (GCN) is used to learn feature aggregation on the topology graph, enabling each resource node to perceive global spatiotemporal association information and achieving a higher-order evolution from individual features to group collaborative features. First, self-connected units are added to the resource topology adjacency matrix to retain the node's own feature information and avoid the loss of its own attributes during topology aggregation, thus obtaining the self-connected adjacency matrix. :

[0122]

[0123] In the formula, for An identity matrix of order 1. Simultaneously calculate the self-connected adjacency matrix. The corresponding degree matrix Its diagonal elements are the sum of the adjacency weights of the corresponding nodes, and the specific formula for calculation is as follows:

[0124]

[0125] In the formula, Represents a self-connected adjacency matrix The comprehensive association topology weight factor between the i-th resource and the t-th resource; The degree matrix represents The diagonal elements in the i-th row and i-th column are used to perform Laplacian normalization, addressing the feature value divergence problem during graph convolution training. Finally, graph convolution feature evolution operations are performed to complete higher-order feature enhancement.

[0126]

[0127] In the formula, The resource feature matrix output in step 1 is used as the input feature benchmark for graph convolution operation; This represents the high-order resource feature matrix that is finally output after graph convolutional spatial reconstruction and feature aggregation learning, which integrates global spatiotemporal correlation and business collaboration constraints. For a trainable projected weight matrix, Both are bias vectors, and are iteratively optimized through backpropagation of the model. To unify the scheduling feature dimensions for phase S1, This represents the higher-order feature dimension after topological evolution; This is a non-linear activation function used to enhance the non-linear fitting capability of feature representation. This stage completes the dynamic spatiotemporal topology construction of resources through multi-dimensional correlation modeling, and combines graph convolution to achieve neighborhood feature aggregation and feature dimensionality enhancement, ultimately outputting a high-order resource feature matrix that integrates global spatiotemporal correlation and business collaboration constraints. .

[0128] At this point, this step completes the enhancement of topology-aware features on the resource side, making up for the deficiency that single features cannot represent resource collaborative scheduling relationships, forming a bidirectional input form of enhanced resource features and hardware features, and providing complete feature support for supply and demand tensor alignment and adaptive matching scheduling in the S3 stage.

[0129] S3: Supply and Demand Feature Tensor Alignment and Adaptive Matching Based on Bidirectional Attention: The high-order resource feature matrix and the hardware feature matrix are aligned and adaptively matched based on bidirectional attention to obtain a global scheduling adaptability matrix, including:

[0130] The high-order resource feature matrix and the hardware feature matrix are uniformly mapped to the same matching feature space to achieve dimensional alignment of the feature spaces on both the supply and demand sides, and the aligned feature matrix is ​​output.

[0131] Based on the fine-grained association mining feature mapping matrix, cross-domain interaction modeling is performed on the aligned feature matrix to generate a query matrix, a key matrix, and a value matrix;

[0132] The query matrix and key matrix are scaled and normalized to generate an adaptive joint control weight matrix;

[0133] The adaptive joint control matrix and the value matrix are weighted and fused to generate the final global scheduling adaptability matrix.

[0134] This step follows the high-order resource feature matrix output by S2, which integrates teaching business semantics and temporal-spatial collaborative relationships. and S1 hardware feature matrix At this point, the two types of features suffer from inconsistent dimensionality and representation granularity, making it impossible to directly perform supply-demand matching calculations. Furthermore, traditional fixed-weight matching and simple inner-product matching cannot distinguish the fine-grained matching preferences of resources and hardware, making it difficult to adapt to dynamic scheduling scenarios involving heterogeneous teaching resources and hardware. To address this, a bidirectional attention tensor alignment mechanism is introduced. Through unified feature dimension mapping, bidirectional adaptive weight modeling, and global fit scoring, it achieves precise matching between teaching resource needs and hardware capacity, outputting the final resource-hardware scheduling adaptation matrix and completing the core inference process of the scheduling matching module within the overall heterogeneous resource adaptive scheduling system.

[0135] The dimensions of high-order resource features are Hardware feature dimension is Since the feature spaces of the two are independent, dimensional alignment needs to be achieved through a trainable projection weight matrix to map the features of both sides to the same matching feature space, laying the foundation for subsequent attention interaction and adaptation calculation. For high-order resource feature matrices... Introducing a resource-side projection weight matrix With bias ,Will High-order collaborative feature mapping to a unified matching space dimension :

[0136]

[0137]

[0138] In the formula, This represents the intermediate state matrix of the linear projection for cross-domain dimensional alignment on the resource side. It is represented as the alignment resource feature matrix obtained after activation mapping of the nonlinear Gaussian error linear unit (GeLU).

[0139] For hardware feature matrix Introducing a hardware-side projection weight matrix With bias Simultaneously perform dimension alignment to ensure that the output dimensions of features on both sides are consistent:

[0140]

[0141]

[0142] In the formula, The standard Gaussian cumulative distribution function is... To ensure consistent matching of spatial dimensions; Represented as a linear projection intermediate state matrix with cross-domain dimension alignment on the hardware side; This represents the aligned hardware feature matrix obtained after GeLU activation mapping. All projection weight matrices and bias vectors are adaptive learnable parameters, without any manually preset fixed values. The optimal parameters for the teaching scheduling scenario are obtained through adaptive learning.

[0143] Before the overall iterative optimization of the heterogeneous resource adaptive scheduling system is initiated, all projection parameters are standardized and initialized, assigned only basic initial values. At this point, the parameters have no scenario adaptation capability. All projection weight matrices follow a zero-mean normal distribution. ( To ensure that the initial weights are randomly initialized (to account for variance), the initial weight values ​​are smooth and do not disrupt the original feature distribution.

[0144]

[0145] in, Take a minimal constant. All one-dimensional biases are uniformly initialized to an all-zero vector, serving only as the initial reference for feature offsets:

[0146]

[0147] To achieve adaptive optimization of the bilateral projection weight matrix, bias, and attention weights, supervised learning objectives are required for model training. This is primarily described in the generation and training processes of the projection weight matrix and bias. After mapping, a uniformly dimensional supply-demand bilaterally aligned feature matrix is ​​obtained.

[0148]

[0149] This step completely eliminates the problem of mismatched feature dimensions on both sides, and at the same time, it retains the core attribute features of both supply and demand sides by adaptively adapting the distribution of scenario features through a trainable mapping layer.

[0150] Subsequently, to uncover fine-grained, implicit supply-demand matching relationships between resources and hardware, a query-key-value (QKV) feature mapping matrix based on the classic attention mechanism was introduced. This method performs cross-domain interaction modeling on the bilateral feature matrices in a unified space to uncover implicit adaptation relationships between resources and hardware, addressing the issues of weak generalization and low adaptation accuracy in traditional matching methods. The initial values ​​of the three sets of trainable projection weight matrices are defined as follows:

[0151]

[0152] With the aligned resource feature matrix For querying source, hardware feature matrix As the key-value source, generate the query matrix, key matrix, and value matrix respectively:

[0153]

[0154] Among them, the query matrix The key matrix represents the personalized scheduling needs and preferences of each teaching resource. The value matrix represents the core carrying characteristics of each hardware node. This is a weighted representation of hardware features, used for subsequent feature fusion and matching scoring. To eliminate the interference of feature dimensions on the attention distribution, a scaled inner product is applied to the query matrix and key matrix to match the feature similarity between resource requirements and hardware supply. The better the feature fit and the more closely the supply and demand attributes match, the larger the inner product value, representing a higher degree of scheduling correlation between the two. (Introduction) Scaling avoids the gradient vanishing problem caused by the divergence of inner product values ​​in high-dimensional features, and normalization using Softmax yields a standardized matching weight matrix, quantifying the degree of adaptation between each resource-hardware pair.

[0155]

[0156] In the formula, The bidirectional attention adaptation score matrix, i.e., the adaptive joint control weight matrix, is used. The Softmax function normalizes all hardware adaptation weights corresponding to any resource, and F is the unified hidden layer feature dimension of the resource-hardware supply-demand matching degree. This ensures that each resource can adaptively select the optimal matching hardware, realizing a fine-grained bidirectional interaction of resource-on-demand hardware matching and hardware adaptively accepting resources. The normalized attention weights and hardware value matrix are weighted and fused to obtain the final matching representation that takes into account resource requirements, hardware capabilities, and global adaptation relationships. At the same time, the final global scheduling adaptation degree matrix of the model is generated.

[0157]

[0158] The final output global scheduling fitness matrix Each element in the matrix Representing the The resource and the first The quantitative control direction weight score of each hardware component. The higher the score, the better the hardware's performance, load, and deployment environment match the scheduling requirements of the corresponding resources, providing a precise quantitative control basis for the final discrete scheduling decision of the physical layer execution mechanism in subsequent steps.

[0159] S4: Generation and Training Process of Projection Weight Matrix and Bias Vector: Adaptive optimization is performed on the bilateral projection weight matrix, bias vector, and attention weight parameters used in step 1 (bidirectional feature modeling) and step 3 (supply and demand feature tensor alignment and adaptive matching based on bidirectional attention). This includes:

[0160] Based on historical operation logs, hardware security operation constraints, and load limit indicators, hardware performance adaptation score, resource demand matching score, and hardware load balancing score are calculated and weighted and fused to obtain a pseudo-true value control matrix.

[0161] A supervised loss function is constructed by measuring the error between the global scheduling fit matrix and the pseudo-truth control matrix.

[0162] The bilateral projection weight matrix, bias vector, and attention weight parameters are adaptively optimized based on the supervised loss function.

[0163] To achieve adaptive optimization of the bilateral projection weight matrix, bias vector, and attention weight parameters in stages S1, S2, and S3, supervised learning objectives are required for model training. This is mainly achieved by constructing a loss function and iteratively updating the projection weight matrix and bias terms through backpropagation, gradually converging the feature mapping to the optimal state.

[0164] First, based on the teaching resource scheduling rules and hardware adaptation constraints, a pseudo-truth value control matrix for real-world scenarios is constructed. .matrix Quantitative constraint indicators are constructed from three core engineering dimensions: hardware feasibility, resource demand matching, and hardware load balancing. Adaptation scores for individual resource-hardware pairings are calculated for each pair, and finally, a weighted fusion is performed to obtain a comprehensive pseudo-true value adaptation score. The formula is as follows:

[0165]

[0166] In the formula, the weights satisfy the normalization constraint. To ensure that the scoring range is controllable, the specific construction rules for each quantitative item are as follows:

[0167] (1) Hardware performance compatibility score : Pre-calculate the minimum hardware operating thresholds for all resources (computing power, memory, bandwidth, latency). If the hardware... Various indicators do not meet resource requirements If any operational threshold is met, the system is directly deemed incompatible, and its score is forcibly set to 0. If all threshold requirements are met, the remaining hardware performance margin is calculated and mapped to a normalized value. Within a given range, the more sufficient the hardware redundancy, the higher the score.

[0168]

[0169] These are all static physical parameters or configuration parameters, which can be obtained from the manufacturer's instructions or system parameters. For the j-th hardware, five types of inherent hardware parameters—maximum IOPS, peak bandwidth, maximum concurrent processing count, extreme read / write speed, and maximum cache capacity—are collected in batches through the device's underlying driver interface. After normalization, a maximum hardware performance feature vector is constructed, namely... This represents the maximum resource limit that hardware j can support. The minimum hardware performance threshold required for the normal operation of the i-th resource is obtained from the fault tolerance parameters of the equipment manufacturer and the long-term operation logs of the smart teaching platform, and belongs to the hardware safety operation constraint threshold.

[0170] This represents the maximum performance of all hardware components. =max{ , , } is used to unify the residual normalization scale and ensure that the score range is controllable. In the formula, This represents the maximum performance feature vector of the Mth heterogeneous hardware storage node accessed within the current scheduling cycle in the smart teaching scenario.

[0171] (2) Resource demand matching score Based on the resource business attribute features (initial comprehensive feature vector of resources) extracted by S1 and the hardware adaptation attribute features (initial comprehensive feature vector of hardware), the cosine similarity metric is used to quantify the degree of fit between the two in terms of business scenarios, teaching types, and operating architectures. The higher the similarity, the higher the adaptation score, naturally avoiding the subjective bias of manual scoring. The value range is fixed. This enables the objective quantification of business adaptability.

[0172]

[0173] In the formula, This represents the standardized comprehensive feature vector of the i-th resource, specifically derived from the initial comprehensive feature vector of the i-th resource. The L2 norm normalization process is performed to obtain the value, and its calculation formula is as follows: ; This represents the standardized comprehensive feature vector of the j-th hardware resource, specifically derived from the initial comprehensive feature vector of the j-th hardware resource. The L2 norm normalization process is performed to obtain the value, and its calculation formula is as follows: This similarity calculation can objectively quantify the business compatibility between resources and hardware, enabling fully automated and quantifiable solutions for resource requirement matching scores.

[0174] (3) Hardware load balancing score Statistics on current hardware The remaining available load is calculated by comparing the occupied resources with the maximum capacity limit. The larger the load margin and the more idle the equipment, the higher the score.

[0175]

[0176] In the formula, This represents the maximum resource limit that hardware j can support. The current resource usage of hardware j is obtained through the operating system kernel performance monitoring tool. This score changes dynamically with the scheduling status and can provide real-time feedback on the hardware load. The guidance model prioritizes low-load hardware to ensure overall scheduling balance.

[0177] To ensure uniformity in the range of supervised label values ​​and convergence of the adaptation loss, the obtained composite pseudo-true value adaptation score is... Perform global min-max normalization to uniformly map to The interval is used to obtain the final pseudo-true control matrix for model supervision:

[0178]

[0179] In the formula, It is the minimum value among all elements of the entire matrix. The maximum value among all elements of the entire matrix is ​​used to standardize all fit scores through global extremum scaling. After normalization... The closer the value in the matrix is ​​to 1, the better the overall adaptability of the resource-hardware pairing under multiple objective constraints; the closer the value is to 0, the worse the adaptability of the pairing or the lack of scheduling feasibility.

[0180] It should be noted that due to the large-scale, highly time-series, and high-concurrency dynamic characteristics of smart teaching scenarios, directly applying the above formula to perform a global traversal calculation of massive resource and hardware pairings during the online real-time scheduling phase results in extremely high computational complexity and I / O overhead. The scheduling latency of the heterogeneous resource adaptive scheduling system would reach the second level, failing to meet the stringent technical requirement of millisecond-level response in smart classrooms. Therefore, this invention employs an offline training and online inference architecture, utilizing historical operation logs during the offline phase to construct a pseudo-truth value control matrix using the above formula. This signal is used as a supervisory signal to train the heterogeneous resource adaptive scheduling network model.

[0181] During online scheduling, the heterogeneous resource adaptive scheduling system directly inputs the current resource and hardware characteristics into the trained adaptive control network for forward propagation lightweight inference, thereby enabling instantaneous matrix output in milliseconds. The prediction matrix (i.e., the global scheduling fit matrix) is measured using the model prediction matrix. With the aforementioned pseudo-true value control matrix To construct a supervised loss function based on the error between them. This provides a clear and traceable optimization basis for the reverse iterative update of all learnable parameters of S3. This invention uses mean squared error loss to achieve accurate fitting of continuous adaptation scores; the supervised loss function is defined as follows:

[0182]

[0183] In the formula, The predicted value is automatically generated for the forward propagation of the model, representing the degree of scheduling fit between the i-th resource and the j-th hardware. The objectively optimal matching label, obtained based on real teaching scheduling constraints and hardware adaptation rules, serves as the ground truth for supervised model training. The loss function quantifies the overall model fit bias by globally traversing all resource and hardware pairings. The smaller the loss value, the closer the feature mapping and matching logic constructed by the projection weight matrix, bias vector, and attention weights are to the real scheduling rules, and the higher the overall parameter fit and model matching accuracy; conversely, a larger loss value indicates a bias in the parameter mapping, requiring continuous optimization through backward gradient updates.

[0184] (1) Iterative optimization of projection weight matrix and bias vector in S1, S2 and S3

[0185] S1 contains two sets of learnable projection parameters, one from the resource side and one from the hardware side, and participates in the global iterative optimization throughout the entire process: the resource-side projection weight matrix. Resource-side bias vector Hardware-side projection weight matrix Hardware-side bias vector The above parameters are entirely based on the loss function. Gradient update complete.

[0186] Based on the chain rule, the global matching loss is differentiated layer by layer forward to obtain the gradients of the four types of S1 parameters, accurately quantifying the contribution of the S1 mapping parameters to the final matching error. The gradient of the resource-side parameters is as follows:

[0187]

[0188] The gradient of the hardware-side parameters is:

[0189]

[0190] Set a fixed training learning rate for the model ,by This indicates the current training iteration of the model. The gradient descent algorithm is used to iteratively update all parameters of S1, i.e.:

[0191]

[0192] Secondly, based on the chain rule of differentiation, the loss function The error signal propagates forward and backward layer by layer. This is to quantize the graph convolution feature transformation matrix described in S2. and bias vector The contribution of the partial derivatives to the final matching error is calculated by the heterogeneous resource adaptive scheduling system using the loss function. Compared to and The gradient of the partial derivative matrix is ​​specifically expressed as follows:

[0193]

[0194] In the formula, This is the high-order resource feature matrix output by the graph convolutional neural network. Then, the heterogeneous resource adaptive scheduling system is set to a fixed training learning rate. Let t represent the current iteration of the model, and let gradient descent be used to transform the graph convolutional feature matrix. With the bias vector The iterative update of the dynamic execution parameter values ​​is expressed by the following formula:

[0195]

[0196]

[0197] The heterogeneous resource adaptive scheduling system repeatedly performs forward propagation prediction and the aforementioned reverse parameter update until the loss function is defined. The value gradually decreases and converges to the preset steady-state threshold, at which point it is locked. and This refers to the optimal graph convolution transformation parameters that are suitable for smart teaching scenarios.

[0198] Finally, by employing a dual-sided dedicated projection transformation, the problem of inconsistent feature dimensions between resources and hardware heterogeneity was resolved, achieving precise alignment of supply and demand feature spaces. Simultaneously, a unified learning rate was maintained across the models. With iteration number t, the gradient descent algorithm is used to iteratively update all learnable parameters of S3, gradually optimizing the parameter adaptation capability. The update formulas for the resource-side projection weight matrix and bias vector are as follows:

[0199]

[0200]

[0201] Hardware-side projection weight matrix and bias vector update formula:

[0202]

[0203]

[0204] The model iteratively executes forward propagation for prediction and backpropagation for parameter updates. As the number of iterations increases, the matching loss function gradually decreases and stabilizes. When the loss value no longer fluctuates significantly and the model's matching accuracy reaches a stable state, the model training is considered converged. At this point, the iteratively optimized model... , , , These are the optimal parameters for adapting to the scenario of intelligent teaching resource scheduling, and can be formally used for supply and demand matching reasoning.

[0205] (2) Generation and training process of attention weights

[0206] The model uses a single supervised loss function throughout the entire process. Based on the backpropagation chain rule, hierarchical gradient backpropagation is achieved. The error first updates the top-level QKV attention weights, correcting the feature interaction matching logic; then it propagates forward layer by layer, updating the S3 bilateral projection weight matrix and bias vector, correcting the feature alignment mapping rules. The two types of parameters each perform their respective functions and work together to optimize, completing two different tasks—feature space normalization and fine-grained matching modeling—under the same loss constraint. Based on the solved gradient, the attention weight iterative update formula is as follows:

[0207]

[0208] As offline iterations continue, the control loss function The values ​​gradually decrease and tend to stabilize. When the loss value no longer fluctuates significantly and the model matching accuracy reaches a stable state, the network training is considered to have converged. At this point, the iteratively optimized matrices and bias parameters are the optimal parameters for the intelligent teaching resource scheduling scenario. The heterogeneous resource adaptive scheduling system solidifies these parameters and deploys them in the online scheduling engine for inference and execution of real-time supply and demand flow control decisions.

[0209] S5: Adaptive Discrete Decision Making and Physical Control Execution Based on Multi-Dimensional Hard Constraints: Based on the global scheduling adaptability matrix and combined with multi-dimensional hard constraints, adaptive discrete decision making and physical control execution are performed to achieve adaptive scheduling of heterogeneous resources, including:

[0210] The hardware resource carrying capacity limit and resource exclusivity constraint are used as rigid preconditions to perform feasible domain pruning and filtering on the global scheduling adaptability matrix.

[0211] Based on the global scheduling adaptability matrix after feasible region pruning and filtering, construct the optimal route candidate queue;

[0212] Based on the optimal route candidate queue, an adaptive greedy decoding strategy is used to initially discretize and bind resources and hardware to generate an initial decision matrix.

[0213] Under the premise of satisfying rigid preconditions, with the optimization objective of minimizing the variance of each hardware load, the inverse variance fine-tuning state machine is invoked to fine-tune the initial discretization binding, and the final binary scheduling decision matrix is ​​obtained.

[0214] The physical enable flag in the binary scheduling decision matrix is ​​read, and the physical enable flag is converted into a corresponding hardware storage and network routing control command and sent to the physical execution layer. This drives the corresponding heterogeneous resource data stream to be written into the physical sector of the target hardware storage medium, thus completing the dynamic adaptive collaborative scheduling at the physical layer.

[0215] This step follows the global scheduling fit matrix output by the real-time forward inference in step S3. And the dynamic hardware load parameters updated within the current scheduling period. At this time... Let be a continuous variable in the real number field, representing the quantized fit between the i-th resource and the j-th hardware. To transform the continuous matching probability matrix into a discrete scheduling decision scheme that can be directly implemented in a production environment, this step constructs a constraint-pre-constrained two-stage adaptive load balancing decoding and physical control execution mechanism. This mechanism first ensures that the scheduling scheme never goes out of bounds through rigid constraint pre-filtering; then, through greedy initial matching and inverse variance fine-tuning two-stage closed-loop decoding, it achieves global load balancing while ensuring high-performance matching; finally, it drives the physical layer execution mechanism to complete the actual transfer and writing of the underlying data.

[0216] Before decoding, the scheduling matching module first introduces the hardware resource carrying capacity limit constraint and resource exclusivity constraint as rigid preconditions into the decoder, and then applies them to the global scheduling fit matrix. Perform feasible domain pruning and filtering.

[0217] (1) Resource monopoly constraint

[0218] Define the final output binary scheduling decision matrix as follows: Its elements .in =1 represents establishing a scheduling binding relationship between the i-th resource and the j-th hardware, and C(i,j)=0 represents no binding. The heterogeneous resource adaptive scheduling system enforces the constraint that any independent resource can only be allocated to a single hardware node, strictly prohibiting split deployment. Its mathematical boundary is reflected as follows:

[0219]

[0220] (2) Hardware resource carrying capacity limit constraint

[0221] Real-time reading of the amount of hardware resources currently occupied, collected in step S4. During the traversal allocation process, whenever a resource i (whose known dynamic load is...) is allocated... When allocating to hardware j, a pre-verification must be performed:

[0222]

[0223] Among them, let For the first The dynamic load required for the operation of each resource For hardware Current resource usage For hardware The maximum resource limit that can be supported. This represents the sum of all new resource loads determined to be allocated to hardware j within the current scheduling cycle. If the above condition is met, an allocation attempt is allowed; otherwise, it indicates that the hardware does not have sufficient capacity to handle the resource, and the heterogeneous resource adaptive scheduling system directly assigns the corresponding score from the global scheduling fit matrix. (i,j) is forcibly reset to negative infinity, and the overload pair is permanently removed from the candidate set, thus fundamentally eliminating the running lag and failure caused by hardware overload in the teaching system.

[0224] After rigid constraint filtering, the heterogeneous resource adaptive scheduling system performs the first stage of discretization initial allocation. First, the heterogeneous resource adaptive scheduling system calculates the global scheduling fit matrix. All elements are sorted globally in descending order of their scores to construct an optimal route candidate queue. Next, the heterogeneous resource adaptive scheduling system employs an adaptive greedy decoding strategy to sequentially extract the optimal pairing relationship from the top of the queue. If resources Not yet allocated, and hardware If the aforementioned dynamic capacity out-of-bounds filtering constraint is satisfied, then the corresponding enable flag in the final output binary scheduling decision matrix is ​​set to 0. =1. Finally, the heterogeneous resource adaptive scheduling system updates the dynamic load of the hardware in real time: = + The heterogeneous resource adaptive scheduling system iterates through the queue in this manner until all N heterogeneous teaching resources in the current period have completed their initial discretization and binding, generating an initial decision matrix. .

[0225] To overcome the global hardware load imbalance caused by the adaptive greedy decoding strategy focusing only on local optima, a global load balancing optimization objective is introduced, while satisfying the first two rigid constraints. Global balanced scheduling is achieved by minimizing the load variance of each hardware node. The mathematical expression is:

[0226]

[0227] in, This represents the average amount of resources used after all M hardware updates within the current period.

[0228] Initial decision matrix After generation, the heterogeneous resource adaptive scheduling system automatically calculates the current global load variance. The heterogeneous resource adaptive scheduling system sets a global equilibrium variance tolerance threshold. ,like This indicates that the current initial allocation plan has reached the equilibrium standard, and can be directly... As the final binary scheduling decision matrix C output; if This triggers the reverse fine-tuning of the state machine:

[0229] The heterogeneous resource adaptive scheduling system uses matrix scanning to locate the hardware with the highest actual load rate. And the hardware with the lowest actual load rate. .

[0230] Allocated to Search for the global scheduling fit matrix score in the resource subset. The resource ranked last (i.e., with the lowest compatibility) .

[0231] Feasibility of reverse migration in the pre-verification of heterogeneous resource adaptive scheduling system: if the target hardware is under low load. The remaining capacity satisfies + And the new global variance after the swap Then the original greedy decision is revoked, and a reverse fine-tuning is performed, that is, the original greedy decision is revoked, and a reverse fine-tuning is performed. =0, and =1.

[0232] Heterogeneous resource adaptive scheduling system synchronously adds, removes and updates resources. and The physical load values ​​are used to recalculate the global variance. This fine-tuning process is executed iteratively within the current scheduling cycle until the global variance falls below the threshold. Or, after reaching the maximum number of iterations, the final binary scheduling decision matrix C is locked and output.

[0233] Finally, after the final binary scheduling decision matrix C is locked, the heterogeneous resource adaptive scheduling system is removed from the pure digital logic layer. The central controller of the heterogeneous resource adaptive scheduling system directly reads the physical enable flag in matrix C (i.e., C(i,j)=1), converts it into the corresponding hardware storage and network routing control instructions, and sends them to the execution mechanism of the physical execution layer.

[0234] (1) If the device type label corresponding to the target hardware j is high-speed flash memory medium (SSD), the central controller sends a physical switching signal to the data flow channel switching matrix of the smart teaching platform, and routes the corresponding low-latency, highly interactive teaching instruction flow (such as in-class assessment data and classroom interaction instructions) directly to the physical bus interface of the flash memory medium, and calls the underlying driver interface to perform physical writing of continuous sectors.

[0235] (2) If the physical tag corresponding to the target hardware j is a cold data storage medium such as a mechanical hard disk cluster, the central controller drives the physical head control mechanism and the file storage management stack to write the large volume of teaching video or archived courseware data stream with low latency requirements into the corresponding physical storage cluster.

[0236] Through the aforementioned pre-filtering, two-stage discrete decoding, and coordinated control of physical actuators, this invention achieves end-to-end closed-loop scheduling control from top-level AI continuous feature matching to low-level hardware discrete physical actions. Under the premise of ensuring hardware operation safety, it significantly improves the global collaborative scheduling efficiency of smart teaching resources and the stability of the underlying storage system.

[0237] In summary, this invention discloses a heterogeneous teaching resource collaborative scheduling method that integrates topological feature enhancement and constraint decision-making. First, it standardizes the representation of resource and hardware features through business semantic parsing and projection weight matrices. Second, it constructs a multi-dimensional topological association structure of resource content, temporal sequence, and physical space, and enhances spatiotemporal collaborative features using high-order feature aggregation evolution operators. It accurately matches based on supply and demand tensor alignment and association mining mechanisms, outputting a global scheduling fit matrix. It also addresses the training problem of unsupervised scheduling models. A two-stage discrete decoding strategy with pre-constrained adaptive greedy initial matching and inverse variance fine-tuning is adopted, combined with hardware load capacity and resource exclusivity rigid constraints, transforming continuous matching scores into a deployable discrete scheduling decision matrix. Finally, the central controller issues control commands to drive the physical layer execution mechanism to physically write the teaching resource data stream into the physical sectors of the target storage medium. This invention achieves end-to-end closed-loop scheduling, significantly improving the rationality of intelligent teaching resource scheduling and the stability of the underlying storage system.

[0238] Example 2

[0239] In this embodiment, two types of heterogeneous teaching resources with different attributes and two types of underlying storage hardware nodes with drastically different performances are selected as implementation objects in a real teaching scenario, so as to complete the implementation and operation of the entire set of intelligent scheduling schemes for heterogeneous teaching resources described in the aforementioned embodiment.

[0240] There are two heterogeneous teaching resources to be scheduled, with N=2 and M=2 underlying heterogeneous hardware nodes. The two teaching resources to be scheduled are shown in Table 1.

[0241] Table 1 Teaching Resources to be Scheduled

[0242] Resource 1 Working principle of photoelectric sensors.mp4 Key concepts in bandwidth-intensive, high-frequency access, and photoelectric detection 1.2GB, 8Mbps, execution time 45min Resource 2 In-class quiz interaction data.json Extremely low latency, highly interactive, in-class assessment 50KB, response latency ≤20ms, execution time 48min

[0243] Table 2 describes the underlying hardware nodes. The matched and deployed underlying heterogeneous hardware nodes include high-performance edge solid-state drives and central data center hard disk drive clusters. Edge solid-state drives have the advantages of fast read and write speeds, high IOPS, strong concurrency processing capabilities, and low runtime latency. In this implementation, the hardware load is relatively low, making them suitable for carrying interactive and real-time small teaching resources. Hard disk drive clusters have a large storage capacity and lower storage costs, making them suitable for long-term archiving and storage of large-capacity teaching resources. However, they have the disadvantages of high read and write latency and inability to adapt to high-frequency interactive services.

[0244] Table 2. Underlying Hardware Nodes

[0245] Hardware 1 High-performance edge solid-state drives (SSDs) High IOPS, low latency, high concurrency, low current load Real-time interaction, low-latency services Hardware 2 Central data center hard disk drive (HDD) cluster Large capacity, high read / write latency, suitable for archiving Large volume, non-real-time data

[0246] Based on the above description, the heterogeneous resource adaptive scheduling system first completes bidirectional feature modeling, digitally characterizing both teaching resources and underlying hardware nodes. On the resource side, on one hand, it uses a BERT pre-trained language model to parse the textual information of the two teaching resources, including their business descriptions, teaching purposes, and associated knowledge points, outputting a 768-dimensional semantic feature vector to accurately capture the teaching business attributes corresponding to the resources. On the other hand, it directly extracts physical operating parameters such as file size, transmission bitrate, latency threshold, and teaching execution time to construct a resource physical feature vector. Subsequently, through a preset projection weight matrix, the semantic feature vector and the physical feature vector are concatenated and dimensionality-reduced to generate comprehensive feature vectors for both teaching resources. These are then integrated to form a complete resource feature matrix. .

[0247] On the hardware node side, the heterogeneous resource adaptive scheduling system collects real-time hardware operation data such as remaining storage space, real-time bandwidth, current operating load, and read / write speed of the solid-state drive and hard disk drive clusters through the hardware driver interface. Simultaneously, it combines this data with pre-set hardware service adaptation tags, service priorities, and other textual information to complete feature transformation, generate corresponding comprehensive hardware feature vectors, and summarize them to obtain the hardware feature matrix. This quantifies the service capabilities and real-time operating status of all hardware nodes, providing data support for subsequent resource matching.

[0248] After completing bidirectional feature modeling, the heterogeneous resource adaptive scheduling system, relying on graph convolutional networks, performs dynamic topology modeling and feature enhancement of teaching resources, deeply mining the implicit correlation constraints between different teaching resources to align with the logic of real classroom teaching. By comparing the teaching execution time and the teaching knowledge points of two teaching resources, the system calculates that their classroom execution times differ by only 3 minutes, and both belong to the unified teaching unit of "photoelectric detection," demonstrating a high degree of consistency in both temporal and semantic correlation. Further calculations reveal the comprehensive correlation strength between the two resources. This value exceeds the pre-set association threshold of the heterogeneous resource adaptive scheduling system. Therefore, the system automatically creates strongly correlated topological edges for the two resource nodes in the resource topology graph, building a complete resource topology structure. Then, a graph convolutional network is invoked to aggregate and update the neighborhood features of all resources within the topology graph. This allows previously independent video resources and evaluation data resources to integrate their temporal and business collaboration relationships, optimizing and upgrading the high-order resource feature matrix and compensating for the shortcomings of traditional scheduling algorithms that ignore the collaborative attributes of teaching resources.

[0249] Based on the optimized resource feature matrix and the initial hardware feature matrix, the heterogeneous resource adaptive scheduling system performs cross-domain feature tensor alignment and supply-demand matching scoring. To address the issue of inconsistent resource and hardware feature dimensions and the inability to directly interact and match them, this scheme sets up two sets of trainable projection weight matrices. and The system performs dimensional unification mapping on resource features and hardware features respectively, unifying the two types of heterogeneous features into the same matching computation space. After dimensional alignment, a QKV multi-head attention mechanism is introduced to conduct fine-grained cross-domain feature interaction, comprehensively measuring the performance requirements and business needs of different teaching resources and the adaptability of the service capabilities of each hardware node, and finally outputting a standardized supply and demand matching score matrix.

[0250]

[0251] in, (2,1)=0.94, so the real-time in-class assessment data (resource 2) and the edge solid-state drive (SSD) (hardware 1) have a matching score of up to 0.94, which is extremely high.

[0252] (1,2)=0.85, the matching score between the high-definition teaching video (resource 1) and the mechanical hard disk cluster HHD (hardware 2) is 0.85, indicating good compatibility;

[0253] (2,2)=0.12. The in-class assessment data and the mechanical hard disk cluster only scored 0.12, indicating extremely poor compatibility. This result also directly confirms that low-latency interactive resources cannot adapt to the business characteristics of high-latency mechanical storage hardware.

[0254] To continuously improve model matching accuracy and ensure the stability of long-term scheduling operations, the heterogeneous resource adaptive scheduling system performs closed-loop iterative training and optimization operations offline. Based on the platform's massive historical teaching resource scheduling operation logs, combined with hardware operation failure records and classroom teaching feedback data, staff constructed a pseudo-truth value matrix for the adaptive scheduling matching module. This matrix clearly defines the optimal and prohibited combinations of different types of teaching resources and hardware nodes. For example, the pseudo-truth value matrix directly restricts the allocation of in-class assessment interaction data to mechanical hard drive clusters, setting the corresponding adaptation score to 0, thus adding hard business constraints. Simultaneously, mean squared error is used as the model loss function to accurately quantify the attention mechanism matching score matrix. The deviation between the value and the pseudo-truth matrix D is minimized. Gradient backpropagation is performed hierarchically using an adaptive evolutionary convergence operator, with the feature mapping matrix of the attention mechanism being updated synchronously. , and With projection weight matrix parameters , The training is iterated repeatedly until the loss function converges, the optimal model parameters are locked, and the offline adaptive evolutionary optimization of the scheduling model is completed.

[0255] During the online scheduling phase, the system combines hardware load constraints with matching scores to complete discretized scheduling decisions and issue physical execution instructions. The heterogeneous resource adaptive scheduling system first verifies the resource attribute constraints of all resources to be scheduled, confirming that high-definition video resources and in-class assessment data are indivisible, exclusive teaching resources, prohibiting fragmented storage allocation. Simultaneously, it monitors the load status of both types of hardware nodes in real time, confirming that neither the SSD nor the HDD cluster has reached its full load threshold and is ready for storage access. Subsequently, a two-stage decoding decision mode is used to complete resource allocation. The first stage relies on an adaptive greedy algorithm to perform initial resource binding based on matching scores from highest to lowest, prioritizing the allocation of the highest-scoring in-class assessment data to the edge SSD. The second stage performs global load variance fine-tuning, calculating the load balancing variance of all hardware nodes under the initial allocation scheme, determining that the variance value of the current allocation scheme is within a preset threshold range, requiring no further adjustment, and finally generating a binary scheduling decision matrix.

[0256]

[0257] This indicates that the video from Resource 1 is allocated to the mechanical hard drive of Hardware 2, and the answer data from Resource 2 is allocated to the SSD of Hardware 1.

[0258] Finally, the underlying central controller reads the decision matrix content, sends read and write control commands to the corresponding hardware, writes large-capacity high-definition teaching videos to the mechanical hard disk cluster for archiving storage, routes low-latency in-class assessment data to the edge solid-state drives, and completes data writing with the help of high-speed bus interfaces, comprehensively meeting the differentiated storage and operation needs of different heterogeneous teaching resources, and realizing intelligent and precise resource scheduling.

[0259] Example 3

[0260] Based on the same inventive concept, the present invention also provides a heterogeneous resource adaptive scheduling system that integrates topology feature enhancement and constraint decision-making, for implementing the method described in the foregoing embodiments. The system includes: a bidirectional feature modeling module, a topology feature enhancement module, a scheduling matching module, and a constraint adaptive decision-making module.

[0261] The bidirectional feature modeling module is used to perform bidirectional feature modeling on heterogeneous resources and physical execution layer in smart teaching scenarios to obtain resource feature matrix and hardware feature matrix.

[0262] The topology feature enhancement module is used to construct a resource topology map oriented towards the spatiotemporal correlation of teaching, and to enhance the resource feature matrix to obtain a higher-order resource feature matrix.

[0263] The scheduling matching module is used to perform supply and demand feature tensor alignment and adaptive matching based on bidirectional attention on the high-order resource feature matrix and the hardware feature matrix to obtain a global scheduling fit matrix.

[0264] The constraint adaptive decision-making module is used to perform adaptive discrete decision-making and physical control execution based on the global scheduling adaptability matrix and combined with multi-dimensional hard constraints in advance, so as to realize the adaptive scheduling of heterogeneous resources.

[0265] Furthermore, in this embodiment, the system also includes a closed-loop iterative optimization module;

[0266] The closed-loop iterative optimization module is used to adaptively optimize the bilateral projection weight matrix, bias vector, and attention weight parameters used in the process of bidirectional feature modeling by the bidirectional feature modeling module and the process of supply and demand feature tensor alignment and adaptive matching based on bidirectional attention by the feature alignment and matching module.

[0267] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A heterogeneous resource adaptive scheduling method integrating topology feature enhancement and constraint decision-making, characterized in that, The method includes: Step 1: Perform bidirectional feature modeling on heterogeneous resources and the physical execution layer in the smart teaching scenario to obtain the resource feature matrix and hardware feature matrix, including: Obtain metadata description text of heterogeneous resources in smart teaching scenarios, call preset business semantic concatenation parsing operators to perform deep word segmentation and semantic representation encoding, and generate a semantic feature matrix of resources; Extract the applicable physical parameters for each heterogeneous resource, normalize them, and construct the physical feature matrix of the resource; The semantic and physical features of the resource are cross-domain concatenated and spatially aligned using the first projection weight matrix to generate an initial comprehensive feature vector for a single resource, which constitutes the resource feature matrix. Simultaneously, through the device underlying driver interface of the physical execution layer and the kernel performance monitoring operator, the underlying physical performance parameters of heterogeneous hardware are collected in real time to construct hardware physical feature vectors; Read the hardware device attribute configuration table to generate standardized hardware business description text, and call the business semantic concatenation parsing operator to process and obtain the semantic feature vector of the hardware; The physical and semantic features of the hardware are spatially aligned and stitched together using the second projection weight matrix to generate an initial comprehensive feature vector for a single hardware component, and a hardware feature matrix is ​​constructed. Step 2: Construct a resource topology map oriented towards the spatiotemporal correlation of teaching, and perform feature enhancement on the resource feature matrix to obtain a high-order resource feature matrix; Step 3: Perform supply and demand feature tensor alignment and adaptive matching based on bidirectional attention on the high-order resource feature matrix and the hardware feature matrix to obtain a global scheduling fit matrix, including: The high-order resource feature matrix and the hardware feature matrix are uniformly mapped to the same matching feature space to achieve dimensional alignment of the feature spaces on both the supply and demand sides, and the aligned feature matrix is ​​output. Based on the fine-grained association mining feature mapping matrix, cross-domain interaction modeling is performed on the aligned feature matrix to generate a query matrix, a key matrix, and a value matrix; The query matrix and key matrix are scaled and normalized to generate an adaptive joint control weight matrix; The adaptive joint control weight matrix and the value matrix are weighted and fused to generate the final global scheduling adaptability matrix; Step 4: Based on the global scheduling adaptability matrix and combined with multi-dimensional hard constraints, perform adaptive discrete decision-making and physical control execution to achieve adaptive scheduling of heterogeneous resources, including: The hardware resource carrying capacity limit and resource exclusivity constraint are used as rigid preconditions to perform feasible domain pruning and filtering on the global scheduling adaptability matrix. Based on the global scheduling adaptability matrix after feasible region pruning and filtering, construct the optimal route candidate queue; Based on the optimal route candidate queue, an adaptive greedy decoding strategy is used to initially discretize and bind resources and hardware to generate an initial decision matrix. Under the premise of satisfying rigid preconditions, with the optimization objective of minimizing the variance of each hardware load, the inverse variance fine-tuning state machine is invoked to fine-tune the initial discretization binding, and the final binary scheduling decision matrix is ​​obtained. The physical enable flag in the binary scheduling decision matrix is ​​read, and the physical enable flag is converted into a corresponding hardware storage and network routing control command and sent to the physical execution layer. This drives the corresponding heterogeneous resource data stream to be written into the physical sector of the target hardware storage medium, thus completing the dynamic adaptive collaborative scheduling at the physical layer.

2. The method according to claim 1, characterized in that, The method for constructing a resource topology map oriented towards the spatiotemporal correlation of teaching in step 2, and performing feature enhancement on the resource feature matrix to obtain a higher-order resource feature matrix includes: Using all heterogeneous resources as network nodes, the association weights between any two heterogeneous resources are calculated from three dimensions: content semantics, teaching time sequence, and physical deployment space, and then integrated to generate a comprehensive association weight. Based on the comprehensive association weights and combined with the preset association thresholds, effective association relationships are screened to construct a resource topology map oriented towards the spatiotemporal relevance of teaching. The resource topology map is subjected to feature aggregation learning and spatial reconstruction. The attributes of individual resources are integrated into the global spatiotemporal correlation constraints of neighboring resources. The resource feature matrix is ​​then enhanced to obtain a high-order resource feature matrix.

3. The method according to claim 1, characterized in that, The method further includes: The bilateral projection weight matrix, bias vector, and attention weight parameters used in the bidirectional feature modeling in step 1 and the supply and demand feature tensor alignment and adaptive matching based on bidirectional attention in step 3 are adaptively optimized.

4. The method according to claim 3, characterized in that, Methods for adaptively optimizing the bilateral projection weight matrix, bias vector, and attention weight parameters include: Based on historical operation logs, hardware security operation constraints, and load limit indicators, hardware performance adaptation score, resource demand matching score, and hardware load balancing score are calculated and weighted and fused to obtain a pseudo-true value control matrix. A supervised loss function is constructed by measuring the error between the global scheduling fit matrix and the pseudo-truth control matrix. The bilateral projection weight matrix, bias vector, and attention weight parameters are adaptively optimized based on the supervised loss function.

5. A heterogeneous resource adaptive scheduling system integrating topology feature enhancement and constraint decision-making, the system being used to implement the method described in any one of claims 1-4, characterized in that, The system includes: a bidirectional feature modeling module, a topology feature enhancement module, a scheduling and matching module, and a constraint adaptive decision-making module; The bidirectional feature modeling module is used to perform bidirectional feature modeling on heterogeneous resources and physical execution layer in smart teaching scenarios to obtain resource feature matrix and hardware feature matrix. The topology feature enhancement module is used to construct a resource topology map oriented towards the spatiotemporal correlation of teaching, and to enhance the resource feature matrix to obtain a higher-order resource feature matrix. The scheduling matching module is used to perform supply and demand feature tensor alignment and adaptive matching based on bidirectional attention on the high-order resource feature matrix and the hardware feature matrix to obtain a global scheduling fit matrix. The constraint adaptive decision-making module is used to perform adaptive discrete decision-making and physical control execution based on the global scheduling adaptability matrix and combined with multi-dimensional hard constraints in advance, so as to realize the adaptive scheduling of heterogeneous resources.

6. The system according to claim 5, characterized in that, The system also includes a closed-loop iterative optimization module; The closed-loop iterative optimization module is used to adaptively optimize the bidirectional projection weight matrix, bias vector, and attention weight parameters used in the bidirectional feature modeling module for bidirectional feature modeling and in the scheduling and matching module for supply and demand feature tensor alignment and adaptive matching based on bidirectional attention.

Images