Multi-terminal cooperation-oriented digital intelligent screen station interaction scene scheduling method and system

Abstract

The invention provides a multi-terminal cooperation-oriented digital intelligent screen platform interaction scene scheduling method and system, and relates to the technical field, and the method comprises the steps: receiving an access request of multiple terminal devices, and determining an optimal task distribution strategy based on device parameter information through combining a deep learning model and an improved Hungary algorithm; acquiring real-time load data, and performing real-time scheduling by adopting a dynamic load balancing algorithm; and the state synchronization of the interactive execution result is realized through the incremental state synchronization packet. According to the invention, efficient cooperation among multiple terminal devices can be realized, the distribution efficiency of interaction tasks is improved, the continuity of interaction experience is ensured, and the consistency of the states of the terminal devices is ensured.

Description

Technical Field

[0001] The present invention relates to digital intelligent screen technology, and in particular to a method and system for scheduling interactive scenes of digital intelligent screens for multi-terminal collaboration. Background Art

[0002] With the popularization of smart terminal devices and the development of Internet of Things technology, multi-terminal collaborative interaction has become an important trend in digital transformation. In the application scenario of digital smart screens, multiple terminal devices need to interact and work together in real time through wireless networks to achieve efficient information transmission and seamless business connection. At present, multi-terminal collaboration mainly uses centralized servers for task scheduling and data synchronization, and unified communication protocols are used between terminal devices for data exchange and status synchronization.

[0003] First, the existing technology lacks an intelligent task allocation mechanism for different terminal device characteristics. Traditional task allocation methods often use fixed rules or simple polling strategies, and cannot optimize allocation based on the hardware configuration, processing power and other parameter characteristics of the terminal device, resulting in low resource utilization efficiency and affecting the overall coordination effect.

[0004] Secondly, the existing load balancing solutions generally adopt static configuration, which cannot effectively cope with the load changes of terminal devices during operation. When some terminal devices are overloaded, it is easy to cause system response delays, reduced interactive experience and other problems, while the resources of terminal devices with lighter loads are not fully utilized.

[0005] Finally, in the multi-terminal collaboration process, due to factors such as network delay and data packet loss, there are consistency issues in the state synchronization between the terminal devices. Existing technologies mostly use full data synchronization, which not only increases network transmission overhead, but also may lead to inconsistent states between terminal devices, affecting the reliability and stability of collaborative interaction. Summary of the invention

[0006] The embodiments of the present invention provide a method and system for scheduling interactive scenes of digital intelligent screens for multi-terminal collaboration, which can solve the problems in the prior art.

[0007] According to a first aspect of the embodiments of the present invention,

[0008] Receive access requests sent by multiple terminal devices through a wireless network, wherein the multiple terminal devices include a master terminal device and a slave terminal device; obtain device parameter information of the multiple terminal devices, and determine the optimal task allocation strategy under different terminal device combinations according to the device parameter information, in combination with a pre-trained deep learning model and an improved Hungarian algorithm, wherein the improved Hungarian algorithm realizes task matching optimization according to alternating path construction and zero element priority calculation;

[0009] Collecting real-time load data of multiple terminal devices, the real-time load data includes processor usage, memory usage, storage usage, and network bandwidth usage; based on the real-time load data, using a dynamic load balancing algorithm to perform real-time scheduling of interactive tasks between the multiple terminal devices, the dynamic load balancing algorithm dynamically adjusts the task allocation plan on the premise of ensuring the continuity of the interactive experience by calculating the comprehensive load index of the terminal device;

[0010] According to the adjusted task allocation plan, corresponding interaction instructions are respectively issued to the multiple terminal devices; the interaction execution results returned by the multiple terminal devices are received, and if there are terminal devices with inconsistent states, the state of the interaction execution results is synchronized through an incremental state synchronization package.

[0011] According to the device parameter information, combined with the pre-trained deep learning model and the improved Hungarian algorithm, the optimal task allocation strategy under different terminal device combinations is determined, including:

[0012] The deep learning model includes a bidirectional long short-term memory network module, a self-attention mechanism module, a multi-head attention mechanism module and a fully connected layer module, wherein the bidirectional long short-term memory network module is used to extract the time series characteristics of the feature vector corresponding to the device parameter information, the self-attention mechanism module generates the associated weights by performing a dot product operation between the query matrix and the key value matrix and performing a scale transformation, the multi-head attention mechanism module performs feature fusion on the associated weights, and the fully connected layer module generates a task assignment probability distribution;

[0013] Constructing a combined loss function, the combined loss function includes a weighted combination of a task allocation accuracy loss value, a performance indicator loss value, and a load balancing loss value; using the Adam optimization algorithm to train the deep learning model, and processing intermediate layer features through batch normalization operations during the training process;

[0014] Calculating a capability score of the terminal device based on the deep learning model, wherein the capability score is obtained by weighted summation of device parameters and corresponding weights;

[0015] A task adaptation matrix is ​​constructed according to the capability score and the task resource requirement vector, and load balancing constraints and resource utilization efficiency constraints are introduced. The task adaptation matrix is ​​solved by using an improved Hungarian algorithm to determine the optimal task allocation strategy under different terminal device combinations; wherein the load balancing constraint ensures the balance of task allocation among terminal devices, and the resource utilization efficiency constraint ensures the maximization of overall resource utilization.

[0016] The task adaptation matrix is ​​constructed according to the capability score and the task resource requirement vector, and load balancing constraints and resource utilization efficiency constraints are introduced. The task adaptation matrix is ​​solved by using the improved Hungarian algorithm to determine the optimal task allocation strategy under different terminal device combinations, including:

[0017] The capability score and the corresponding dimension of the task resource requirement vector are similarly calculated, and the similarity calculation adopts the ratio of the minimum value to the maximum value of the corresponding dimension to obtain the standardized matching degree between the device and the task; the standardized matching degree is multiplied by a preset load balancing coefficient to obtain a task adaptation matrix, wherein the preset load balancing coefficient is obtained by the ratio of the load value of a single terminal device to the load value of all terminal devices;

[0018] An optimization objective function is constructed based on the task fitness matrix combined with the weight value assigned to the task fitness matrix, and load balancing constraints and resource utilization efficiency constraints are introduced to construct a constraint matrix; the optimization objective function and the constraint matrix are jointly solved using an improved Hungarian algorithm.

[0019] Using the improved Hungarian algorithm to jointly solve the optimization objective function and the constraint matrix includes:

[0020] Constructing a first reduction matrix and a second reduction matrix according to the optimization objective function and the constraint matrix respectively, wherein the first reduction matrix is ​​obtained by taking the inverse of the function value of the optimization objective function, and the second reduction matrix is ​​obtained by transposing the constraint matrix;

[0021] Identify the positions of zero elements in the first reduction matrix and the second reduction matrix respectively, count the number of zero elements in each row and the number of zero elements in each column, and calculate the priority of the zero elements, where the priority of the zero element is the sum of the number of zero elements in the row where the zero element is located and the number of zero elements in the column where the zero element is located;

[0022] Mark and select zero elements based on the zero element priority, count the matching number of zero elements in each row and the matching number of zero elements in each column, and construct a matching set; count the unmatched rows without zero elements and the unmatched columns without zero elements, and construct an unmatched set;

[0023] Taking the row index in the unmatched set as the starting point, constructing an alternating path based on the matched set and the unmatched set, wherein adjacent vertices in the alternating path correspond alternately to unmatched zero elements and matched zero elements, marking the positions corresponding to the unmatched zero elements on the alternating path as 1, and marking the positions corresponding to the matched zero elements as 0;

[0024] Determine whether the total number of elements marked as 1 in the matching set in the alternating path is equal to the minimum number of matrix rows and columns. If not, reconstruct the matching set until the total number of elements marked as 1 is equal to the minimum number of matrix rows and columns.

[0025] The position marked as 1 in the matching set is used as the corresponding relationship of the task allocation.

[0026] Based on the real-time load data, a dynamic load balancing algorithm is used to schedule the interactive tasks between the multiple terminal devices in real time. The dynamic load balancing algorithm dynamically adjusts the task allocation scheme by calculating the comprehensive load index of the terminal device while ensuring the continuity of the interactive experience, including:

[0027] Calculating the fluctuation standard deviation of the real-time load data and the pre-acquired historical load data, and setting an adaptive weight vector corresponding to the real-time load data according to the fluctuation standard deviation, wherein the adaptive weight vector is determined based on the product of the fluctuation standard deviation and a preset weight reference value;

[0028] Performing weighted calculation on the real-time load vector corresponding to the real-time load data and the adaptive weight vector to obtain a comprehensive load index of each terminal device;

[0029] When the comprehensive load index exceeds a preset load balancing threshold, the task allocation scheme is reallocated.

[0030] According to the adjusted task allocation scheme, issuing corresponding interaction instructions to the multiple terminal devices respectively; receiving the interaction execution results returned by the multiple terminal devices, and if there are terminal devices with inconsistent states, synchronizing the states of the interaction execution results through an incremental state synchronization package includes:

[0031] Extracting an interactive instruction set of each terminal device according to the task allocation scheme, wherein the interactive instruction set includes a device identifier, an interactive type, and execution timing information; extracting timing information of the interactive instructions in the interactive instruction set, and constructing an interactive instruction directed graph according to the timing information, wherein the nodes of the interactive instruction directed graph represent interactive instructions, and the edges represent timing dependencies between instructions;

[0032] Traversing the interactive instruction directed graph to calculate the path length of each interactive instruction, extracting the delay sensitivity parameter of each interactive instruction, and performing weighted calculation on the path length and the delay sensitivity parameter to obtain the instruction priority;

[0033] Arrange the interactive instructions in descending order according to the instruction priority, divide the arranged interactive instructions into multiple instruction groups according to the priority interval, and construct an instruction group execution queue; based on the order of the instruction group execution queue, send the interactive instructions to the corresponding terminal devices group by group;

[0034] Receive the interactive execution result returned by the terminal device, and extract the state vector in the interactive execution result, wherein the state vector includes the device resource state and the task execution state;

[0035] If there are terminal devices with inconsistent states, the state difference field of the terminal devices with inconsistent states and the state vector of the terminal devices with inconsistent states are encapsulated into an incremental state synchronization package, and the incremental state synchronization package is sent to all terminal devices, and the state update confirmation information returned by the terminal devices is received to verify the consistency of the state after the update.

[0036] A second aspect of an embodiment of the present invention provides a digital intelligent screen interactive scene scheduling system for multi-terminal collaboration, including:

[0037] The first unit is used to receive access requests sent by multiple terminal devices through a wireless network, wherein the multiple terminal devices include a master terminal device and a slave terminal device; obtain device parameter information of the multiple terminal devices, and determine the optimal task allocation strategy under different terminal device combinations according to the device parameter information, in combination with a pre-trained deep learning model and an improved Hungarian algorithm, wherein the improved Hungarian algorithm realizes task matching optimization according to alternating path construction and zero element priority calculation;

[0038] The second unit is used to collect real-time load data of multiple terminal devices, wherein the real-time load data includes processor usage, memory usage, storage usage, and network bandwidth usage; based on the real-time load data, a dynamic load balancing algorithm is used to perform real-time scheduling of interactive tasks between the multiple terminal devices, wherein the dynamic load balancing algorithm dynamically adjusts the task allocation scheme while ensuring the continuity of the interactive experience by calculating the comprehensive load index of the terminal device;

[0039] The third unit is used to issue corresponding interaction instructions to the multiple terminal devices respectively according to the adjusted task allocation plan; receive the interaction execution results returned by the multiple terminal devices, and if there are terminal devices with inconsistent states, synchronize the state of the interaction execution results through an incremental state synchronization package.

[0040] A third aspect of the embodiments of the present invention

[0041] An electronic device is provided, comprising:

[0042] processor;

[0043] a memory for storing processor-executable instructions;

[0044] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0045] A fourth aspect of the embodiments of the present invention is:

[0046] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0047] The beneficial effects of this application are as follows:

[0048] The present invention determines the optimal task allocation strategy by combining a deep learning model with an improved Hungarian algorithm, wherein the improved Hungarian algorithm utilizes alternating path construction and zero element priority calculation to perform task matching optimization, which can effectively improve the efficiency and accuracy of task allocation in multi-terminal collaboration scenarios.

[0049] The present invention adopts a dynamic load balancing algorithm to perform task scheduling based on real-time load data, and dynamically adjusts the task allocation scheme by calculating the comprehensive load index of the terminal device, thereby achieving dynamic balance of loads between terminal devices while ensuring the continuity of the interactive experience, thereby improving the overall performance and stability of the system.

[0050] The present invention uses an incremental state synchronization package to synchronize the state of the interactive execution results, effectively solving the state inconsistency problem that may occur in a multi-terminal collaboration scenario, ensuring the reliability of system operation and data consistency, and improving the efficiency and accuracy of multi-terminal collaborative work. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 A schematic diagram of a flow chart of a method for scheduling interactive scenes of digital intelligent screens and platforms for multi-terminal collaboration according to an embodiment of the present invention;

[0052] Figure 2 A schematic diagram showing a comparison of system resource utilization rates in an embodiment of the present invention;

[0053] Figure 3 A schematic diagram showing a comparison of computational overheads under different problem scales in an embodiment of the present invention;

[0054] Figure 4 A schematic diagram of the load balance distribution of the system according to an embodiment of the present invention;

[0055] Figure 5 Schematic diagram of a distributed task scheduling real-time monitoring system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0056] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0057] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0058] Figure 1 FIG. 1 is a flow chart of a method for scheduling interactive scenes of digital intelligent screens and platforms for multi-terminal collaboration according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0059] Receive access requests sent by multiple terminal devices through a wireless network, wherein the multiple terminal devices include a master terminal device and a slave terminal device; obtain device parameter information of the multiple terminal devices, and determine the optimal task allocation strategy under different terminal device combinations according to the device parameter information, in combination with a pre-trained deep learning model and an improved Hungarian algorithm, wherein the improved Hungarian algorithm realizes task matching optimization according to alternating path construction and zero element priority calculation;

[0060] Collecting real-time load data of multiple terminal devices, the real-time load data includes processor usage, memory usage, storage usage, and network bandwidth usage; based on the real-time load data, using a dynamic load balancing algorithm to perform real-time scheduling of interactive tasks between the multiple terminal devices, the dynamic load balancing algorithm dynamically adjusts the task allocation plan on the premise of ensuring the continuity of the interactive experience by calculating the comprehensive load index of the terminal device;

[0061] According to the adjusted task allocation plan, corresponding interaction instructions are respectively issued to the multiple terminal devices; the interaction execution results returned by the multiple terminal devices are received, and if there are terminal devices with inconsistent states, the state of the interaction execution results is synchronized through an incremental state synchronization package.

[0062] In an optional implementation, according to the device parameter information, combined with a pre-trained deep learning model and an improved Hungarian algorithm, determining the optimal task allocation strategy under different terminal device combinations includes:

[0063] The deep learning model includes a bidirectional long short-term memory network module, a self-attention mechanism module, a multi-head attention mechanism module and a fully connected layer module, wherein the bidirectional long short-term memory network module is used to extract the time series characteristics of the feature vector corresponding to the device parameter information, the self-attention mechanism module generates the associated weights by performing a dot product operation between the query matrix and the key value matrix and performing a scale transformation, the multi-head attention mechanism module performs feature fusion on the associated weights, and the fully connected layer module generates a task assignment probability distribution;

[0064] Constructing a combined loss function, the combined loss function includes a weighted combination of a task allocation accuracy loss value, a performance indicator loss value, and a load balancing loss value; using the Adam optimization algorithm to train the deep learning model, and processing intermediate layer features through batch normalization operations during the training process;

[0065] Calculating a capability score of the terminal device based on the deep learning model, wherein the capability score is obtained by weighted summation of device parameters and corresponding weights;

[0066] A task adaptation matrix is ​​constructed according to the capability score and the task resource requirement vector, and load balancing constraints and resource utilization efficiency constraints are introduced. The task adaptation matrix is ​​solved by using an improved Hungarian algorithm to determine the optimal task allocation strategy under different terminal device combinations; wherein the load balancing constraint ensures the balance of task allocation among terminal devices, and the resource utilization efficiency constraint ensures the maximization of overall resource utilization.

[0067] The processor frequency, memory capacity, network bandwidth, storage space and other parameter information of the terminal device are collected every 5 seconds through the device management interface, and collected 12 times continuously to form a time series data sequence. The processor frequency data collection range is 0-4GHz, the memory capacity is 0-32GB, the network bandwidth is 0-2000Mbps, and the storage space is 0-1TB. Actual collection example: processor frequency 3.2GHz, memory 16GB, bandwidth 1000Mbps, storage 512GB.

[0068] When extracting time series features, the bidirectional long short-term memory network controls the flow of information through the input gate, forget gate, and output gate. The input gate controls the proportion of new data written, the forget gate controls the degree of retention of historical information, and the output gate controls the amount of information output. The network contains two directions, forward and reverse. The number of hidden layer units in each direction is set to 128. After bidirectional splicing, a 256-dimensional feature vector is obtained.

[0069] In the self-attention mechanism, the 256-dimensional feature vector is transformed into a query vector, a key vector, and a value vector through three independent linear transformations. The dot product of the query vector and the key vector is calculated to obtain the attention score, which is then divided by 16 for scaling and normalized to obtain the attention weight.

[0070] The multi-head attention mechanism sets 8 attention heads, and each head independently calculates 32-dimensional features. The outputs of the 8 heads are concatenated into a 256-dimensional vector, and the dimension is kept unchanged through linear transformation. The fully connected layer maps the features to task assignment probabilities. The first layer uses a 256×128 transformation matrix and a ReLU activation function, and the second layer outputs the task assignment probability.

[0071] A combined loss function is constructed, which includes task allocation accuracy loss, performance index loss, and load balancing loss, with weights of 0.4, 0.3, and 0.3 respectively. The Adam optimization algorithm is used to train the model, with the initial learning rate set to 0.001, which decays to 0.1 times the original rate every 50 cycles. During the training process, batch normalization is performed on the intermediate layer features.

[0072] When calculating the device capability score, the normalized values ​​of processor frequency, memory capacity, network bandwidth, and storage space are multiplied by weights 0.4, 0.3, 0.2, and 0.1 respectively, and then summed. In the example, the normalized parameter values ​​of the device are 0.8, 0.5, 0.5, and 0.512, and the calculated capability score is 0.6024.

[0073] Construct a task fitness matrix, where the matrix elements are the ratios of device capability scores to task resource requirements. Introduce load balancing constraints to limit the number of tasks that a single device can execute to no more than 30% of the total number of tasks. Introduce resource utilization efficiency constraints to require that the average utilization rate of device resources is no less than 60%.

[0074] The improved Hungarian algorithm is used to solve the task fitness matrix. Through iterative optimization steps such as row and column normalization, finding the minimum element, and line covering, the optimal matching solution that meets the constraints is obtained.

[0075] Figure 2 This is a schematic diagram of comparing system resource utilization rates in an embodiment of the present invention:

[0076] According to the experimental data comparison results, this technical solution is significantly better than the LSTM model and the traditional CNN solution in all key performance indicators. Specifically, in terms of resource utilization efficiency, the CPU utilization rate of this technical solution reaches 92.5%, which is significantly higher than the 78.3% of the LSTM model and the 72.1% of the traditional CNN; the memory utilization rate reaches 88.7%, which is significantly higher than the 75.4% of the LSTM model and the 70.8% of the traditional CNN. In terms of task processing capability, the task completion rate of this technical solution is as high as 97.5%, which is significantly ahead of the 85.6% of the LSTM model and the 81.2% of the traditional CNN; the system load balance reaches 0.92, which is better than the 0.78 of the LSTM model and the 0.73 of the traditional CNN, indicating that the system resource allocation is more reasonable. Especially in terms of system response performance, the response time of this technical solution is only 45.3 milliseconds, which is 42.4% and 51.0% lower than the 78.6 milliseconds of the LSTM model and the 92.4 milliseconds of the traditional CNN, which fully reflects the advantages of this technical solution in scenarios with high real-time requirements. These data comprehensively reflect the comprehensive advantages of this technical solution in multiple dimensions such as resource utilization efficiency, task processing capability and system response performance, proving that the solution has better performance and higher practical value in practical applications.

[0077] In the prior art, traditional load balancing algorithms mainly use static thresholds and simple resource monitoring methods to schedule tasks, which cannot fully consider the dynamic changes in device resources. The polling scheduling method only assigns tasks according to the preset order, completely ignoring the actual load status and resource utilization of the device, resulting in low resource utilization efficiency. These methods all have problems such as unbalanced resource allocation and unreasonable device load distribution.

[0078] This application automatically extracts the timing characteristics of device parameters through a deep learning model, overcoming the limitation of traditional methods that rely only on static thresholds. The use of a bidirectional long short-term memory network can simultaneously consider historical and future timing information and accurately grasp the law of device resource changes. The introduction of a self-attention mechanism realizes dynamic correlation analysis between different resource dimensions, solving the problem that traditional methods are difficult to handle multi-dimensional resource collaborative scheduling. The feature fusion strategy of the multi-head attention mechanism further improves the feature expression capability, making task allocation decisions more accurate and reasonable.

[0079] In the task allocation process, this application improves the traditional Hungarian algorithm and achieves multi-objective optimization by introducing load balancing constraints and resource utilization efficiency constraints. This improvement effectively solves the problem that traditional methods easily cause local resource overload or idleness. Through the device capability scoring mechanism, an accurate matching relationship between task requirements and device resources is established, avoiding the waste of resources caused by simple polling.

[0080] After the above improvements, this application has achieved significant improvements in multiple resource dimensions such as CPU, memory, network bandwidth and storage, which is a qualitative leap compared to traditional load balancing algorithms and polling scheduling methods. The system as a whole shows higher resource utilization efficiency, more balanced load distribution and stronger task processing capabilities. Especially in large-scale task scenarios, the advantages of this application are more obvious, fully demonstrating the adaptability and reliability of this solution in complex environments.

[0081] This technical solution not only solves the problems of low resource utilization and unbalanced load in traditional methods, but also realizes efficient scheduling and reasonable allocation of system resources through intelligent feature extraction and optimized decision-making mechanism, providing a new solution for task scheduling in distributed systems.

[0082] In an optional implementation, a task adaptation matrix is ​​constructed according to the capability score and the task resource requirement vector, and load balancing constraints and resource utilization efficiency constraints are introduced. The task adaptation matrix is ​​solved using an improved Hungarian algorithm to determine the optimal task allocation strategy under different terminal device combinations, including:

[0083] The capability score and the corresponding dimension of the task resource requirement vector are similarly calculated, and the similarity calculation adopts the ratio of the minimum value to the maximum value of the corresponding dimension to obtain the standardized matching degree between the device and the task; the standardized matching degree is multiplied by a preset load balancing coefficient to obtain a task adaptation matrix, wherein the preset load balancing coefficient is obtained by the ratio of the load value of a single terminal device to the load value of all terminal devices;

[0084] An optimization objective function is constructed based on the task fitness matrix combined with the weight value assigned to the task fitness matrix, and load balancing constraints and resource utilization efficiency constraints are introduced to construct a constraint matrix; the optimization objective function and the constraint matrix are jointly solved using an improved Hungarian algorithm.

[0085] First, obtain the capability score of the terminal device and the task resource requirement vector. The capability score of the terminal device includes three dimensions: computing capability, storage capability, and communication capability. The score range of each dimension is 0-100. The task resource requirement vector also includes three dimensions: computing resource requirement, storage resource requirement, and communication resource requirement. The requirement value range of each dimension is 0-100. For example, the capability score of a terminal device is [80, 70, 90], and the resource requirement vector of a task is [60, 50, 70].

[0086] Calculate the similarity between the capability score and the resource requirement vector. For each dimension, take the minimum value of the corresponding value divided by the maximum value to get the standardized matching degree. Taking the above data as an example, the standardized matching degree of the computing capability dimension is 60 / 80=0.75, the storage capability dimension is 50 / 70=0.71, and the communication capability dimension is 70 / 90=0.78. Take the average of the standardized matching degrees of the three dimensions to get the comprehensive matching degree of the device and the task, which is 0.75.

[0087] Calculate the preset load balancing coefficient. Count the current load value of each terminal device, including the weighted average of CPU usage, memory usage, and other indicators. Divide the load value of a single device by the sum of the load values ​​of all devices to get the load balancing coefficient. For example, if the load values ​​of three terminal devices are 0.3, 0.4, and 0.5, respectively, the corresponding load balancing coefficients are 0.25, 0.33, and 0.42.

[0088] Construct the task fitness matrix. Multiply the comprehensive matching degree between each device and task by the corresponding load balancing coefficient to get the final task fitness. For example, if the comprehensive matching degree between a device and three tasks is 0.75, 0.8, and 0.7 respectively, and the load balancing coefficient of the device is 0.25, then the final task fitness is [0.19, 0.2, 0.18].

[0089] The optimization objective function is constructed based on the task fitness matrix. Different weights are assigned to different types of tasks. For example, the weight of tasks with high real-time requirements is 0.6, and the weight of ordinary tasks is 0.4. The task fitness is multiplied by the corresponding weight and the sum is used as the optimization target.

[0090] Introduce constraints. The load balancing constraint requires that the number of tasks undertaken by a single device does not exceed the preset threshold, and the difference in task allocation between devices does not exceed the preset range. The resource utilization efficiency constraint requires that the device resource utilization rate be maintained within a reasonable range to avoid resource waste and overload.

[0091] The improved Hungarian algorithm is used to solve the problem. Constraint judgment is added to the traditional Hungarian algorithm, and the optimal task allocation solution that meets the constraints is obtained through iterative optimization. Finally, the matching relationship matrix between the device and the task is output.

[0092] By calculating the similarity between device capabilities and task requirements and combining the load balancing coefficient to construct a task adaptation matrix, accurate matching of task allocation is achieved and resource scheduling efficiency is improved. The introduction of load balancing and resource utilization efficiency constraints ensures the balance of device load and the rationality of resource utilization, avoiding the problems of device overload and resource waste. The improved Hungarian algorithm is used for solving, which satisfies multiple constraints while ensuring global optimality, improving the feasibility and practicality of the task allocation scheme.

[0093] In an optional implementation, using the improved Hungarian algorithm to jointly solve the optimization objective function and the constraint matrix includes:

[0094] Constructing a first reduction matrix and a second reduction matrix according to the optimization objective function and the constraint matrix respectively, wherein the first reduction matrix is ​​obtained by taking the inverse of the function value of the optimization objective function, and the second reduction matrix is ​​obtained by transposing the constraint matrix;

[0095] Identify the positions of zero elements in the first reduction matrix and the second reduction matrix respectively, count the number of zero elements in each row and the number of zero elements in each column, and calculate the priority of the zero elements, where the priority of the zero element is the sum of the number of zero elements in the row where the zero element is located and the number of zero elements in the column where the zero element is located;

[0096] Mark and select zero elements based on the zero element priority, count the matching number of zero elements in each row and the matching number of zero elements in each column, and construct a matching set; count the unmatched rows without zero elements and the unmatched columns without zero elements, and construct an unmatched set;

[0097] Taking the row index in the unmatched set as the starting point, constructing an alternating path based on the matched set and the unmatched set, wherein adjacent vertices in the alternating path correspond alternately to unmatched zero elements and matched zero elements, marking the positions corresponding to the unmatched zero elements on the alternating path as 1, and marking the positions corresponding to the matched zero elements as 0;

[0098] Determine whether the total number of elements marked as 1 in the matching set in the alternating path is equal to the minimum number of matrix rows and columns. If not, reconstruct the matching set until the total number of elements marked as 1 is equal to the minimum number of matrix rows and columns.

[0099] The position marked as 1 in the matching set is used as the corresponding relationship of the task allocation.

[0100] First, construct the reduction matrix. For the optimization objective function, take the inverse of its function value to get the first reduction matrix. For example, the original objective function value is [[4,2,3], [1,5,2], [3,2,4]], and after taking the inverse, we get the first reduction matrix [[-4,-2,-3], [-1,-5,-2], [-3,-2,-4]]. For the constraint matrix, transpose it to get the second reduction matrix. For example, the original constraint matrix is ​​[[1,0,1], [0,1,0], [1,0,1]], and after transposing, we get the second reduction matrix [[1,0,1], [0,1,0], [1,0,1]].

[0101] Next, identify the locations of zero elements in the two reduced matrices. For each matrix, count the number of zero elements in each row and column. For example, in the first reduced matrix, there are 2 zero elements in the first row and 1 zero element in the second column. Calculate the priority of each zero element, which is the sum of the number of zero elements in the row and the number of zero elements in the column. For example, if a zero element has 2 zeros in its row and 1 zero in its column, its priority is 3.

[0102] Then, mark selection is performed based on the priority of zero elements. Zero elements with high priority are marked as 1 first, indicating a match. The number of matched zero elements in each row and column is counted to build a matching set. At the same time, the number of unmatched rows and columns that do not contain zero elements is counted to build an unmatched set. For example, the matching set is {(1,2), (2,3), (3,1)}, and the unmatched set is {row 4, column 4}.

[0103] Starting from the rows in the unmatched set, construct an alternating path. Adjacent vertices in the path correspond to unmatched zero elements and matched zero elements alternately. For example, the path is: row 4 → column 2 → row 1 → column 3 → row 2. Mark the unmatched zero elements on the path as 1 and the matched zero elements as 0.

[0104] Determine whether the total number of elements marked as 1 in the matching set is equal to the minimum number of rows and columns of the matrix. If not, repeat the process of building the matching set until the number of elements marked as 1 meets the requirement. Finally, the position marked as 1 in the matching set is used as the task allocation plan.

[0105] Figure 3 This is a schematic diagram comparing the computational overheads of different problem scales in an embodiment of the present invention:

[0106] The table shows in detail the performance comparison data of each algorithm under different task scales and the performance improvement of this technical solution compared with the traditional Hungarian algorithm. In the task scenario with the smallest scale of 50×50, the calculation time of this technical solution is 128.6ms, which is 47.6% less than the 245.3ms of the traditional Hungarian algorithm. Although the greedy algorithm only needs 55.7ms, its accuracy is lower. As the task scale increases, the calculation time of the three algorithms shows a nonlinear growth trend, but this technical solution always maintains a significant performance advantage. At the medium scale of 150×150, this technical solution requires 389.5ms, the traditional Hungarian algorithm requires 728.4ms, and the greedy algorithm requires 168.9ms, with a performance improvement of 46.5%. In the scenario with the largest scale of 300×300, the calculation time of this technical solution is 785.9ms, which is 47.1% less than the 1486.3ms of the traditional Hungarian algorithm. Although the greedy algorithm only needs 338.6ms, it is difficult to guarantee the optimality of the allocation scheme. From the overall trend, this technical solution has maintained a stable performance improvement under various task scales, with the average improvement remaining at around 47%, fully demonstrating the significant advantages and good scalability of the solution in large-scale task scheduling scenarios.

[0107] In the prior art, the standard Hungarian algorithm uses a fixed search order to process zero elements, and it needs to traverse the matrix multiple times when constructing the augmenting path, resulting in high computational complexity. The basic iterative algorithm only adjusts the current solution through simple iterative optimization, does not fully utilize the structural characteristics of the problem, and exhibits poor computational efficiency in large-scale problems. These traditional methods often require a large amount of redundant calculations when dealing with high-dimensional task allocation problems, which seriously affects the practicality of the algorithm.

[0108] To address the above problems, this application proposes an improved Hungarian algorithm based on zero element priority and alternating paths. The scheme first analyzes the distribution characteristics of zero elements in the matrix, calculates the number of zero elements in the row and column where each zero element is located, and establishes a zero element priority evaluation mechanism. This strategy can quickly identify zero elements with greater matching potential and avoid blind search in traditional methods.

[0109] In the matching process, this application innovatively introduces a priority-based zero element selection strategy to give priority to those zero elements that are more likely to form effective matches. At the same time, by constructing and maintaining alternating paths, dynamic optimization of matching results is achieved. This improvement makes full use of the structural characteristics of the problem and greatly reduces invalid search and calculation operations.

[0110] Through the above improvements, this application significantly improves the computational efficiency of the algorithm in large-scale problems. Compared with traditional methods, this technical solution shows faster computing speed and better scalability when dealing with high-dimensional task allocation problems. Especially when the problem scale is large, the growth trend of computing time is more moderate, reflecting good performance advantages.

[0111] This technical solution not only solves the problem of low computational efficiency of traditional algorithms in large-scale problems, but also makes full use of computing resources through intelligent priority mechanism and efficient path construction strategy. This provides a new efficient and feasible method for solving complex task allocation problems, which has important practical application value.

[0112] In an optional implementation, based on the real-time load data, a dynamic load balancing algorithm is used to schedule the interactive tasks between the multiple terminal devices in real time, and the dynamic load balancing algorithm dynamically adjusts the task allocation scheme while ensuring the continuity of the interactive experience by calculating the comprehensive load index of the terminal device, including:

[0113] Calculating the fluctuation standard deviation of the real-time load data and the pre-acquired historical load data, and setting an adaptive weight vector corresponding to the real-time load data according to the fluctuation standard deviation, wherein the adaptive weight vector is determined based on the product of the fluctuation standard deviation and a preset weight reference value;

[0114] Performing weighted calculation on the real-time load vector corresponding to the real-time load data and the adaptive weight vector to obtain a comprehensive load index of each terminal device;

[0115] When the comprehensive load index exceeds a preset load balancing threshold, the task allocation scheme is reallocated.

[0116] In the multi-terminal device interaction scenario, first obtain the real-time load data of each terminal device, including CPU usage, memory occupancy, network bandwidth usage and other operating status parameters. At the same time, read the historical load data of the corresponding terminal device in the past period of time (such as the last 24 hours) from the historical database as a reference benchmark.

[0117] For each terminal device, calculate the standard deviation of the real-time load data relative to the historical load data. Specifically, taking the CPU usage rate as an example, compare the real-time CPU usage rate data points in the last 10 minutes with the historical CPU usage rate data points in the same period, calculate the degree of dispersion of the two sets of data, and obtain the standard deviation of the fluctuation. When the standard deviation of the fluctuation is large, it means that there is a large deviation between the current load and the historical law, and the reference weight of the real-time data needs to be reduced; otherwise, the weight should be increased.

[0118] Set the weight base value to 0.6, and multiply the standard deviation of fluctuation by the weight base value to get the adaptive weight. For example, if the standard deviation of CPU usage fluctuation of a terminal device is 0.2, the corresponding adaptive weight is 0.12. The same method is used to calculate the weights of other load indicators such as memory and bandwidth.

[0119] The real-time load vector and the adaptive weight vector are weighted to obtain the comprehensive load index. Taking a terminal device as an example, its real-time CPU usage is 80%, memory usage is 60%, and bandwidth usage is 50%. The corresponding weights are 0.12, 0.15, and 0.18, respectively. The comprehensive load index is (80×0.12 + 60×0.15 + 50×0.18) = 28.2.

[0120] The load balancing threshold is set to 30. When the comprehensive load index of a terminal device exceeds the threshold, task redistribution is triggered. During redistribution, tasks are assigned to terminal devices with lower load indexes first, while taking into account factors such as network delay between devices to ensure the continuity of the interactive experience.

[0121] Figure 4 This is a schematic diagram of the load balance distribution of the system according to an embodiment of the present invention:

[0122] This table comprehensively compares the key indicators of this technical solution with traditional dynamic balancing and static balancing in terms of load balancing performance. From the peak position, this technical solution reaches 0.80, which is significantly better than the traditional dynamic balancing of 0.70 and the static balancing of 0.45, indicating that it can achieve a higher degree of load balancing. In terms of distribution width, the 0.10 of this technical solution is better than the 0.12 of the traditional dynamic balancing and the 0.15 of the static balancing, indicating that the load distribution is more concentrated and stable. In terms of the maximum number of samples, this technical solution reaches 98, which is higher than the 92 of the traditional dynamic balancing and the 95 of the static balancing, indicating that it has better data representativeness. In terms of variance indicators, this technical solution is only 0.012, which is much lower than the 0.018 of the traditional dynamic balancing and the 0.025 of the static balancing, proving that the discrete degree of the system load distribution is smaller. In terms of the comprehensive indicator of stability coefficient, this technical solution reaches 0.95, which is significantly higher than the 0.87 of the traditional dynamic balancing and the 0.82 of the static balancing, which fully reflects the excellent performance of this solution in maintaining system stability. These data comprehensively and intuitively demonstrate the significant advantages of this technical solution in all dimensions of load balancing performance.

[0123] By introducing historical load data as a reference benchmark and dynamically adjusting weights based on the standard deviation of fluctuations, load assessment is made more accurate and reasonable, avoiding the decision-making bias that may be caused by relying solely on real-time data. Adopting an adaptive weight mechanism, the weight ratio of each indicator is automatically adjusted according to the degree of fluctuation of load data, which improves the adaptability and robustness of the load balancing algorithm to load changes. Dynamic scheduling of tasks is achieved while ensuring the continuity of the interactive experience, which not only ensures the overall load balance of the system, but also avoids the impact of frequent task migration on user experience.

[0124] In an optional implementation, according to the adjusted task allocation scheme, corresponding interaction instructions are respectively issued to the multiple terminal devices; receiving the interaction execution results returned by the multiple terminal devices, and if there are terminal devices with inconsistent states, synchronizing the state of the interaction execution results through an incremental state synchronization package includes:

[0125] Extracting an interactive instruction set of each terminal device according to the task allocation scheme, wherein the interactive instruction set includes a device identifier, an interactive type, and execution timing information; extracting timing information of the interactive instructions in the interactive instruction set, and constructing an interactive instruction directed graph according to the timing information, wherein the nodes of the interactive instruction directed graph represent interactive instructions, and the edges represent timing dependencies between instructions;

[0126] Traversing the interactive instruction directed graph to calculate the path length of each interactive instruction, extracting the delay sensitivity parameter of each interactive instruction, and performing weighted calculation on the path length and the delay sensitivity parameter to obtain the instruction priority;

[0127] Arrange the interactive instructions in descending order according to the instruction priority, divide the arranged interactive instructions into multiple instruction groups according to the priority interval, and construct an instruction group execution queue; based on the order of the instruction group execution queue, send the interactive instructions to the corresponding terminal devices group by group;

[0128] Receive the interactive execution result returned by the terminal device, and extract the state vector in the interactive execution result, wherein the state vector includes the device resource state and the task execution state;

[0129] If there are terminal devices with inconsistent states, the state difference field of the terminal devices with inconsistent states and the state vector of the terminal devices with inconsistent states are encapsulated into an incremental state synchronization package, and the incremental state synchronization package is sent to all terminal devices, and the state update confirmation information returned by the terminal devices is received to verify the consistency of the state after the update.

[0130] First, extract the interaction instruction details of each terminal device according to the task allocation plan. The generation rule of the device identification code is: prefix "DEV_" plus the year, month, and day timestamp, plus two device serial numbers. For example, the identification code of the first device on November 15, 2023 is "DEV_2023111501". The division of interaction types is based on task characteristics: data collection type (01) is for sensor data acquisition; status query type (02) is used for device status monitoring; parameter configuration type (03) performs device parameter adjustment; task execution type (04) handles specific business tasks.

[0131] The execution timing information is recorded in a timestamp format accurate to milliseconds. For each instruction, its planned start time and expected completion time are recorded. The determination of time information takes into account the characteristics of instruction type, historical execution data, and current system load. By analyzing historical execution records, the average execution time of different types of instructions is calculated, and combined with the current system resource usage, the reasonable execution time interval of the instruction is estimated.

[0132] When constructing a directed graph, each instruction is used as a node in the graph. Node attributes include information such as instruction type, execution time interval, and device identification. By analyzing the timing dependencies between instructions, directed edge connections between nodes are established. The determination of dependencies is based on the order of instruction execution. For example, parameter configuration instructions must be executed after data collection instructions are completed, and task execution instructions need to wait for parameter configuration to be completed.

[0133] Calculation process of path length: Starting from the starting node of the directed graph, all possible paths are traversed using a depth-first search method. For each path, the execution time of each instruction on the path is accumulated to obtain the total path length. For example, a path contains three instruction nodes, which take 5 seconds, 3 seconds, and 7 seconds respectively, then the total length of the path is 15 seconds. By comparing the lengths of all possible paths, the longest path length of each instruction is determined.

[0134] Determination process of latency sensitivity parameters: First, establish a mapping relationship between instruction type and business importance. Analyze the impact of different types of instructions on system performance and set the baseline sensitivity value. Then make fine adjustments based on specific business scenarios. For example, the baseline sensitivity of data collection instructions is 0.8. If the data collected is critical sensor data, it is increased to 0.9; if it is ordinary status data, it is reduced to 0.7.

[0135] The priority value is calculated in multiple steps: first, the path length is normalized and converted to the range of 0-1; then it is weighted and combined with the delay sensitivity parameter; and finally it is dynamically adjusted according to the current system load. The adjustment process takes into account the processing capacity of the device and the urgency of the task.

[0136] The improved quick sort algorithm is used to arrange the instructions in descending order. First, determine the threshold for dividing the priority interval: set a reasonable interval boundary by statistically analyzing the priority distribution characteristics of historical instructions. The threshold of the high priority group is set to 0.8, indicating that instructions with a priority value greater than 0.8 are critical instructions; the threshold of the medium priority group is between 0.5 and 0.8; and the threshold of the low priority group is less than 0.5.

[0137] The construction process of the instruction group execution queue: In each priority group, the instructions are sorted according to the priority value. The queue is implemented with a bidirectional linked list structure, which supports dynamic adjustment of the instruction order. The queue node contains instruction details and execution status mark. When an urgent instruction appears, it can be inserted into the corresponding priority position.

[0138] The instruction issuance process is carried out in batches: first, all instructions in the high-priority group are issued, and the execution results are obtained after the execution is completed. The issuance strategy of subsequent instructions is dynamically adjusted according to the execution results. For example, if a high-priority instruction fails to execute, the related medium-priority instructions need to be delayed or canceled.

[0139] State vector extraction process: The device resource status is obtained in real time through the system interface, including four indicators: processor usage, memory usage, bandwidth usage, and storage usage. Each indicator records the current value, change trend, and threshold setting. The task execution status includes the task progress percentage and status code. The status code uses three digits to indicate different execution stages and abnormal situations.

[0140] Detection method of status inconsistency: Establish a status evaluation indicator system, including three dimensions: degree of deviation, duration, and scope of impact. By setting thresholds of different levels, the severity of the status difference can be judged. For example, if the deviation of memory usage exceeds 20% and lasts for more than 1 minute, it is judged as severe inconsistency.

[0141] The construction process of the incremental synchronization package: First, determine the identification format of the difference field, using the form of "field name_difference type_timestamp". Combine the difference field with the complete state vector and add a checksum. The checksum is calculated by feature extraction of the data content. The size of the incremental package is controlled within a reasonable range through a compression algorithm.

[0142] Status update confirmation process: Receive confirmation information returned by the device, parse the update timestamp and status code. Create a status update log to record the details of each update. Verify the consistency of the update by comparing the status data of different devices. If the update fails, start the retry mechanism, and retry up to 3 times.

[0143] Figure 5 This is a schematic diagram of a distributed task scheduling real-time monitoring system according to an embodiment of the present invention:

[0144] The figure shows the real-time operation status of the task scheduling system: the device status monitoring shows that there are currently 28 / 32 online devices, an online rate of 87.5%, a system load of 78%, and 156 active tasks; the task allocation statistics show that there are 45 tasks to be allocated, 128 tasks in execution, and 1,256 completed tasks; the status synchronization monitoring reflects that the synchronization success rate is 99.8%, with only 5 tasks to be synchronized and 2 synchronization failed tasks; the system resource monitoring panel shows that the CPU utilization rate is 85%, the memory utilization rate is 72%, and the network bandwidth utilization rate is 65%. In the instruction execution queue, the execution status of instructions such as INST_20231115_001 is recorded in detail, and tracked through information such as device identification (such as DEV_2023111501) and instruction type (data acquisition-01). The priority is highlighted in red, and the execution progress is displayed through a dynamic loading icon. The time record is accurate to milliseconds (such as 10:30:00.235 to 10:30:05.235), and the current progress bar shows that the execution progress is 75%. The terminal device status table monitors the operating parameters of each device in real time. Taking DEV_2023111501 as an example, its CPU usage is 75%, memory usage is 60%, network bandwidth is 45%, storage usage is 50%, and it currently undertakes 8 tasks. Each monitoring module is equipped with a progress bar and dynamic icons. The green dot indicates the online status of the device, and the check mark icon indicates that the status synchronization is completed, forming a complete, real-time task scheduling and monitoring system. The table design uses a hover highlight effect, and combines icons and percentage values ​​to provide operators with an intuitive and clear display of the system operation status.

[0145] The instruction set construction adopts precise timing control and dependency analysis to achieve efficient management of complex instruction sequences. The directed graph structure intuitively shows the relationship between instructions and optimizes the planning of execution order. The dynamic timing information recording mechanism improves the accuracy and predictability of instruction execution. The multi-dimensional priority calculation method accurately reflects the importance and timeliness requirements of instructions. The dynamic adjustment capability of the group execution queue enhances the system's response speed to emergency tasks. The perfect execution status monitoring and feedback mechanism ensures the reliability and controllability of the instruction execution process. The incremental state synchronization scheme greatly reduces the synchronization overhead through difference detection and directed update. The multi-dimensional expression of the state vector provides rich system status information. The complete state consistency verification mechanism and exception handling process improve the stability and reliability of the distributed system.

[0146] According to a second aspect of the embodiments of the present invention,

[0147] An electronic device is provided, comprising:

[0148] processor;

[0149] a memory for storing processor-executable instructions;

[0150] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0151] According to a third aspect of the embodiments of the present invention,

[0152] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0153] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0154] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for scheduling interactive scenes of digital intelligent screens and platforms for multi-terminal collaboration, characterized in that: include: Receiving access requests sent by multiple terminal devices through a wireless network, the multiple terminal devices including a master terminal device and slave terminal devices; Obtain device parameter information of the multiple terminal devices, and determine the optimal task allocation strategy under different terminal device combinations based on the device parameter information, in combination with a pre-trained deep learning model and an improved Hungarian algorithm, wherein the improved Hungarian algorithm realizes task matching optimization based on alternating path construction and zero element priority calculation; Collecting real-time load data of multiple terminal devices, the real-time load data includes processor usage, memory usage, storage usage, and network bandwidth usage; based on the real-time load data, using a dynamic load balancing algorithm to perform real-time scheduling of interactive tasks between the multiple terminal devices, the dynamic load balancing algorithm dynamically adjusts the task allocation plan on the premise of ensuring the continuity of the interactive experience by calculating the comprehensive load index of the terminal device; According to the adjusted task allocation plan, corresponding interaction instructions are respectively issued to the multiple terminal devices; the interaction execution results returned by the multiple terminal devices are received, and if there are terminal devices with inconsistent states, the state of the interaction execution results is synchronized through an incremental state synchronization package.

2. The method according to claim 1, characterized in that According to the device parameter information, combined with the pre-trained deep learning model and the improved Hungarian algorithm, the optimal task allocation strategy under different terminal device combinations is determined, including: The deep learning model includes a bidirectional long short-term memory network module, a self-attention mechanism module, a multi-head attention mechanism module and a fully connected layer module, wherein the bidirectional long short-term memory network module is used to extract the time series characteristics of the feature vector corresponding to the device parameter information, the self-attention mechanism module generates the associated weights by performing a dot product operation between the query matrix and the key value matrix and performing a scale transformation, the multi-head attention mechanism module performs feature fusion on the associated weights, and the fully connected layer module generates a task assignment probability distribution; Constructing a combined loss function, the combined loss function includes a weighted combination of a task allocation accuracy loss value, a performance indicator loss value, and a load balancing loss value; using the Adam optimization algorithm to train the deep learning model, and processing intermediate layer features through batch normalization operations during the training process; Calculating a capability score of the terminal device based on the deep learning model, wherein the capability score is obtained by weighted summation of device parameters and corresponding weights; A task adaptation matrix is ​​constructed according to the capability score and the task resource requirement vector, and load balancing constraints and resource utilization efficiency constraints are introduced. The task adaptation matrix is ​​solved by using an improved Hungarian algorithm to determine the optimal task allocation strategy under different terminal device combinations; wherein the load balancing constraint ensures the balance of task allocation among terminal devices, and the resource utilization efficiency constraint ensures the maximization of overall resource utilization.

3. The method according to claim 2, characterized in that The task adaptation matrix is ​​constructed according to the capability score and the task resource requirement vector, and load balancing constraints and resource utilization efficiency constraints are introduced. The task adaptation matrix is ​​solved by using the improved Hungarian algorithm to determine the optimal task allocation strategy under different terminal device combinations, including: The capability score and the corresponding dimension of the task resource requirement vector are similarly calculated, and the similarity calculation adopts the ratio of the minimum value to the maximum value of the corresponding dimension to obtain the standardized matching degree between the device and the task; the standardized matching degree is multiplied by a preset load balancing coefficient to obtain a task adaptation matrix, wherein the preset load balancing coefficient is obtained by the ratio of the load value of a single terminal device to the load value of all terminal devices; An optimization objective function is constructed based on the task fitness matrix combined with the weight value assigned to the task fitness matrix, and load balancing constraints and resource utilization efficiency constraints are introduced to construct a constraint matrix; the optimization objective function and the constraint matrix are jointly solved using an improved Hungarian algorithm.

4. The method according to claim 3, characterized in that Using the improved Hungarian algorithm to jointly solve the optimization objective function and the constraint matrix includes: Constructing a first reduction matrix and a second reduction matrix according to the optimization objective function and the constraint matrix respectively, wherein the first reduction matrix is ​​obtained by taking the inverse of the function value of the optimization objective function, and the second reduction matrix is ​​obtained by transposing the constraint matrix; Identify the positions of zero elements in the first reduction matrix and the second reduction matrix respectively, count the number of zero elements in each row and the number of zero elements in each column, and calculate the priority of the zero elements, where the priority of the zero element is the sum of the number of zero elements in the row where the zero element is located and the number of zero elements in the column where the zero element is located; Mark and select zero elements based on the zero element priority, count the matching number of zero elements in each row and the matching number of zero elements in each column, and construct a matching set; count the unmatched rows without zero elements and the unmatched columns without zero elements, and construct an unmatched set; Taking the row index in the unmatched set as the starting point, constructing an alternating path based on the matched set and the unmatched set, wherein adjacent vertices in the alternating path correspond alternately to unmatched zero elements and matched zero elements, marking the positions corresponding to the unmatched zero elements on the alternating path as 1, and marking the positions corresponding to the matched zero elements as 0; Determine whether the total number of elements marked as 1 in the matching set in the alternating path is equal to the minimum number of matrix rows and columns. If not, reconstruct the matching set until the total number of elements marked as 1 is equal to the minimum number of matrix rows and columns. The position marked as 1 in the matching set is used as the corresponding relationship of the task allocation.

5. The method according to claim 1, characterized in that Based on the real-time load data, a dynamic load balancing algorithm is used to schedule the interactive tasks between the multiple terminal devices in real time. The dynamic load balancing algorithm dynamically adjusts the task allocation scheme by calculating the comprehensive load index of the terminal device while ensuring the continuity of the interactive experience, including: Calculating the fluctuation standard deviation of the real-time load data and the pre-acquired historical load data, and setting an adaptive weight vector corresponding to the real-time load data according to the fluctuation standard deviation, wherein the adaptive weight vector is determined based on the product of the fluctuation standard deviation and a preset weight reference value; Performing weighted calculation on the real-time load vector corresponding to the real-time load data and the adaptive weight vector to obtain a comprehensive load index of each terminal device; When the comprehensive load index exceeds a preset load balancing threshold, the task allocation scheme is reallocated.

6. The method according to claim 1, characterized in that According to the adjusted task allocation scheme, issuing corresponding interaction instructions to the multiple terminal devices respectively; receiving the interaction execution results returned by the multiple terminal devices, and if there are terminal devices with inconsistent states, synchronizing the states of the interaction execution results through an incremental state synchronization package includes: Extracting an interactive instruction set of each terminal device according to the task allocation scheme, wherein the interactive instruction set includes a device identifier, an interactive type, and execution timing information; extracting timing information of the interactive instructions in the interactive instruction set, and constructing an interactive instruction directed graph according to the timing information, wherein the nodes of the interactive instruction directed graph represent interactive instructions, and the edges represent timing dependencies between instructions; Traversing the interactive instruction directed graph to calculate the path length of each interactive instruction, extracting the delay sensitivity parameter of each interactive instruction, and performing weighted calculation on the path length and the delay sensitivity parameter to obtain the instruction priority; Arrange the interactive instructions in descending order according to the instruction priority, divide the arranged interactive instructions into multiple instruction groups according to the priority interval, and construct an instruction group execution queue; based on the order of the instruction group execution queue, send the interactive instructions to the corresponding terminal devices group by group; Receive the interactive execution result returned by the terminal device, and extract the state vector in the interactive execution result, wherein the state vector includes the device resource state and the task execution state; If there are terminal devices with inconsistent states, the state difference field of the terminal devices with inconsistent states and the state vector of the terminal devices with inconsistent states are encapsulated into an incremental state synchronization package, and the incremental state synchronization package is sent to all terminal devices, and the state update confirmation information returned by the terminal devices is received to verify the consistency of the state after the update.

7. A digital intelligent screen interactive scene scheduling system for multi-terminal collaboration, used to implement the method according to any one of claims 1 to 6, characterized in that: include: A first unit is used to receive access requests sent by multiple terminal devices through a wireless network, wherein the multiple terminal devices include a master terminal device and a slave terminal device; Obtain device parameter information of the multiple terminal devices, and determine the optimal task allocation strategy under different terminal device combinations based on the device parameter information, in combination with a pre-trained deep learning model and an improved Hungarian algorithm, wherein the improved Hungarian algorithm realizes task matching optimization based on alternating path construction and zero element priority calculation; The second unit is used to collect real-time load data of multiple terminal devices, wherein the real-time load data includes processor usage, memory usage, storage usage, and network bandwidth usage; based on the real-time load data, a dynamic load balancing algorithm is used to perform real-time scheduling of interactive tasks between the multiple terminal devices, wherein the dynamic load balancing algorithm dynamically adjusts the task allocation scheme while ensuring the continuity of the interactive experience by calculating the comprehensive load index of the terminal device; The third unit is used to issue corresponding interaction instructions to the multiple terminal devices respectively according to the adjusted task allocation plan; receive the interaction execution results returned by the multiple terminal devices, and if there are terminal devices with inconsistent states, synchronize the state of the interaction execution results through an incremental state synchronization package.

8. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images