Method and system for dynamically allocating computing resources of intelligent control chip

Abstract

The present application relates to the technical field of intelligent control chips, and discloses a method and system for dynamically allocating computing resources of intelligent control chips, the method comprising: collecting characteristic parameters and performing weighted calculations on computing cores in the intelligent control chip to obtain a task priority matrix and a resource evaluation vector; performing calculations through a timing prediction channel and a weight allocation channel in a dual-channel neural network model to obtain resource utilization data of each computing core; performing process dependency analysis and spectral clustering operations to obtain task grouping data; inputting the task grouping data into a dual-layer learning network structure scheduling plan to obtain an initial scheduling plan; performing parallel verification processing on the initial scheduling plan, performing conflict detection and coordination optimization on the local scheduling of each core task through set process rules, generating a multi-core collaborative scheduling strategy, and thereby realizing automatic grouping of related tasks, taking into account task response time and execution efficiency while ensuring load balancing.

Description

Technical Field

[0001] The present application relates to the technical field of intelligent control chips, and in particular to a method and system for dynamically allocating computing resources of an intelligent control chip. Background Art

[0002] As the core control unit in the manufacturing process of electronic components, the efficient allocation and scheduling of computing resources of intelligent control chips have become key factors affecting manufacturing efficiency and product quality. Traditional resource allocation methods mainly use static priority scheduling strategies, which are difficult to cope with complex and changing manufacturing scenarios and dynamic task loads, resulting in uneven computing resource utilization and increased task response delays.

[0003] Most existing computing resource management methods are based on a single scheduling algorithm, lacking a systematic analysis of manufacturing process characteristics and task dependencies, and unable to accurately describe dynamic task load characteristics to balance computing resource allocation. In addition, due to the strict constraints of process parameters and quality requirements in the industrial manufacturing process, traditional methods are difficult to take into account both the real-time nature of manufacturing tasks and the consistency of process requirements when scheduling resources. Summary of the invention

[0004] The present application provides a method and system for dynamically allocating computing resources of an intelligent control chip, thereby realizing automatic grouping of related tasks, taking into account task response time and execution efficiency while ensuring load balancing.

[0005] The first aspect of the present application provides a method for dynamically allocating computing resources of an intelligent control chip, and the method for dynamically allocating computing resources of an intelligent control chip includes: Perform characteristic parameter collection and weighted calculation on the computing core in the intelligent control chip to obtain the task priority matrix and resource evaluation vector; Inputting the task priority matrix and the resource evaluation vector into a dual-channel neural network model, performing calculations through a timing prediction channel and a weight allocation channel in the dual-channel neural network model, and obtaining resource utilization data of each computing core; Performing process dependency analysis on the resource utilization data to obtain a process correlation matrix, and performing spectral clustering operation on the process correlation matrix to obtain task grouping data; The task grouping data is input into a two-layer learning network structure, wherein the first layer of the network uses a dual-depth Q network to globally optimize resource allocation, and the second layer of the network uses a multi-agent deep deterministic policy gradient network to schedule and plan each core task, thereby obtaining an initial scheduling plan; The initial scheduling scheme is verified in parallel, and the local scheduling of each core task is conflict-detected and coordinated and optimized according to the set process rules to generate a multi-core collaborative scheduling strategy.

[0006] The second aspect of the present application provides a computing resource dynamic allocation system for an intelligent control chip, and the computing resource dynamic allocation system for an intelligent control chip includes: The acquisition module is used to collect characteristic parameters and perform weighted calculations on the computing core in the intelligent control chip to obtain the task priority matrix and resource evaluation vector; A calculation module, used for inputting the task priority matrix and the resource evaluation vector into a dual-channel neural network model, performing calculations through a timing prediction channel and a weight allocation channel in the dual-channel neural network model, and obtaining resource utilization data of each computing core; An analysis module, used to perform process dependency analysis on the resource utilization data to obtain a process correlation matrix, and perform spectral clustering operation on the process correlation matrix to obtain task grouping data; A planning module is used to input the task grouping data into a two-layer learning network structure, wherein the first layer of the network in the two-layer learning network structure uses a dual-depth Q network to perform global optimization on resource allocation, and the second layer of the network uses a multi-agent deep deterministic policy gradient network to perform scheduling planning on each core task to obtain an initial scheduling plan; The generation module is used to perform parallel verification processing on the initial scheduling plan, perform conflict detection and coordination optimization on the local scheduling of each core task through the set process rules, and generate a multi-core collaborative scheduling strategy.

[0007] The third aspect of the present application provides an electronic device, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the electronic device executes the above-mentioned method for dynamically allocating computing resources of the intelligent control chip.

[0008] A fourth aspect of the present application provides a computer-readable storage medium, wherein the computer-readable storage medium stores instructions, which, when executed on a computer, enable the computer to execute the above-mentioned method for dynamically allocating computing resources of an intelligent control chip.

[0009] Compared with the prior art, the present application has the following beneficial effects: by constructing a dual-channel neural network model, the task characteristics are deeply analyzed from the two dimensions of timing prediction and weight allocation, thereby improving the perception of the dynamic characteristics of manufacturing tasks and making resource allocation decisions more accurate; by combining the dual-depth Q network and the multi-agent deep deterministic policy gradient network in the double-layer learning network structure, the local scheduling efficiency of each computing core is guaranteed while achieving global resource optimization; a parallel verification mechanism based on process rules is designed, and conflict detection and coordinated optimization are used to ensure that the scheduling scheme meets the manufacturing process requirements and enhance the stability of system operation; a spectral clustering algorithm is introduced to analyze the process correlation matrix, thereby realizing automatic grouping of related tasks and optimizing the rationality of task allocation; the final scheduling scheme is selected by combining multi-objective optimization and hierarchical analysis, which takes into account task response time and execution efficiency while ensuring load balancing. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0011] The structures, proportions, sizes, etc. illustrated in the drawings of this specification are only used to match the contents disclosed in the specification so as to facilitate understanding and reading by persons familiar with this technology. They are not used to limit the conditions under which the present invention can be implemented, and therefore have no substantive technical significance. Any structural modification, change in proportion or adjustment of size, without affecting the effects and purposes that can be achieved by the present invention, should still fall within the scope of the technical contents disclosed by the present invention.

[0012] Figure 1 It is a flow chart of a method for dynamically allocating computing resources of an intelligent control chip provided by an embodiment of the present invention; Figure 2 It is a schematic block diagram of the structure of a computing resource dynamic allocation system of an intelligent control chip provided by an embodiment of the present invention; Figure 3 It is a schematic block diagram of the structure of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0013] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are 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.

[0014] The flowcharts shown in the accompanying drawings are only examples and do not necessarily include all the contents and operations / steps, nor must they be executed in the order described. For example, some operations / steps may be decomposed, combined or partially merged, so the actual execution order may change according to actual conditions.

[0015] It should also be understood that the terms used in this application specification are only for the purpose of describing specific embodiments and are not intended to limit the application. As used in this application specification and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include plural forms.

[0016] It should be further understood that the term "and / or" used in the specification and appended claims of this application refers to any combination and all possible combinations of one or more of the associated listed items, and includes these combinations. Figure 1 , an embodiment of the method for dynamically allocating computing resources of the intelligent control chip in the embodiment of the present application includes: Step 100: collecting characteristic parameters and performing weighted calculation on the computing core in the intelligent control chip to obtain a task priority matrix and a resource evaluation vector; It is understandable that the execution subject of the present application may be a computing resource dynamic allocation system of an intelligent control chip, or may be a terminal or a server, which is not specifically limited here. The present application embodiment is described by taking a server as the execution subject as an example.

[0017] Specifically, the parameters of the manufacturing process of the electronic components of the computing core in the intelligent control chip are collected. The parameters include the process calculation complexity, the number of process parameters, the execution cycle and the quality control accuracy. These four dimensions describe the characteristics of the manufacturing process from different angles. For example, the computational complexity reflects the computational intensity during task execution, the number of process parameters reflects the complexity of the process steps required for task processing, the execution cycle provides the time information required to complete the task, and the quality control accuracy measures the degree to which the task execution meets the quality standards. Through the comprehensive collection of multi-dimensional parameters, a feature vector describing the process characteristics is constructed. Based on the process feature vector, a computing resource evaluation matrix R is constructed. The dimension of the matrix is ​​defined as ,in Indicates the number of computing cores, represents the dimension of the evaluation parameter. By this design, the matrix Each row corresponds to a computing core, and each column corresponds to one of the above characteristic parameters. The process characteristics and resource requirements of all computing cores are displayed in a matrix representation. Three quantitative indicators are calculated for each computing core in the initial evaluation matrix, namely, process urgency , Resource occupancy rate Relevance to process . Process urgency It is determined by comprehensively considering the importance and execution cycle of the current process task to quantify the urgency of the task; resource utilization rate It measures the proportion of resources occupied by each computing core in the current task; process relevance The calculation is done by analyzing the dependencies between process tasks to reflect the coupling relationship between tasks. By calculating the quantitative indicators, the core evaluation indicator set of each computing core is obtained. The core evaluation indicator set is weighted and the formula is used. ,in , , Respectively represent the weight coefficients of process urgency, resource occupancy and process relevance, and satisfy . Through this formula, we get the priority index , which is used to comprehensively evaluate the task priority of each computing core. After obtaining the priority index PI, it is arranged and combined according to the computing core number to construct a priority relationship matrix, which is used to represent the priority ranking of each computing core task. At the same time, each element in the initial evaluation matrix R is dot-producted with the corresponding priority index PI to obtain weighted evaluation data. The weighted evaluation data is normalized. Through the linear transformation method, the numerical value of the data is mapped to a unified interval, the deviation caused by different numerical scales is eliminated, and standardized evaluation data is obtained. On this basis, the standardized evaluation data is subjected to column vector extraction and splicing operations to generate a resource evaluation vector to characterize the resource demand status of each computing core.

[0018] Step 200: Input the task priority matrix and the resource evaluation vector into the dual-channel neural network model, and calculate through the timing prediction channel and the weight allocation channel in the dual-channel neural network model to obtain the resource utilization data of each computing core; Specifically, the task priority matrix is ​​input into the timing prediction channel of the dual-channel neural network model, and the feature extraction layer of the channel performs deep feature extraction on the task priority matrix. The feature extraction layer consists of three fully connected layers, each of which is sequentially connected to the ReLU activation function and the batch normalization layer. Through the feature mapping capability of the fully connected layer, combined with the nonlinear transformation of the ReLU activation function and the numerical stability of the batch normalization layer, the key features in the task priority matrix are extracted layer by layer. The output of the feature extraction layer is the priority feature map, which represents the global and local information of the task priority. The priority feature map is input into the LSTM (Long Short-Term Memory) layer in the timing prediction channel. The LSTM layer is set with 128 hidden units, and its core function is to extract and memorize the timing features of T consecutive manufacturing cycles through the forget gate, input gate and output gate. The forget gate filters out irrelevant historical information, the input gate injects the important features of the new input into the hidden state, and the output gate generates meaningful timing features by selectively outputting the current state. Through the time correlation analysis of multi-cycle features, the LSTM layer outputs a timing state sequence, which reflects the regularity of the task priority changing over time. Attention calculation is performed on the time series state sequence. The attention mechanism assigns weights according to the importance of different time points to generate weighted time series features. At the same time, the resource evaluation vector is input into the feature transformation layer of the weight allocation channel of the dual-channel neural network model. The feature transformation layer consists of four fully connected layers. Each fully connected layer uses the Tanh activation function to ensure that the output value remains in the (-1,1) interval, and is combined with the Dropout layer to suppress overfitting. The feature transformation layer generates initial weight features through layer-by-layer linear transformation and nonlinear activation. The initial weight features are normalized and converted to a uniform numerical range using a standardization method to ensure the stability of subsequent calculations. The normalized weight features are input into the bidirectional GRU (Gated Recurrent Unit) network layer, which contains 256 hidden units and can capture the global context information in the resource evaluation vector through forward and reverse hidden state updates. The forward GRU extracts sequential information from the input time series, while the reverse GRU extracts additional information from the reverse direction of the sequence. The resource weight vector is generated by concatenating the forward and reverse hidden states. The weighted time series features are fused with the resource weight vector. Feature fusion uses element-by-element weighting or tensor concatenation to combine the dynamic priority information of the time series prediction channel with the static resource evaluation information of the weight allocation channel to form a fused feature tensor. The fused feature tensor is input into the output mapping layer of the dual-channel neural network model for processing. The output mapping layer consists of one or more fully connected layers, which maps high-dimensional fused features to the output space and generates resource utilization data for each computing core. These data reflect the current resource utilization efficiency and future utilization trends of each computing core.

[0019] Step 300: Perform process dependency analysis on resource utilization data to obtain a process correlation matrix, and perform spectral clustering operation on the process correlation matrix to obtain task grouping data; It should be noted that the dependency relationship between manufacturing process tasks is modeled based on resource utilization data to construct a process dependency graph. In the process dependency graph, the vertex set represents the set of manufacturing process tasks, each vertex corresponds to a specific manufacturing task, and the edge set is used to represent the dependency relationship between processes. The dependency relationship comes from data transmission between tasks, shared resources, and execution order constraints of tasks. This modeling method can display the relationship between process tasks in an intuitive graph structure. In this process, the vertices and edges of the initial dependency graph are directly determined using resource utilization data, and preliminary weight information is added to each edge. Data flow analysis is performed on each pair of processes in the initial dependency graph to quantify the actual data dependency between processes. Data flow analysis generates a data dependency intensity matrix by calculating the amount of data transmission and the number of shared resources between processes. Each element in the matrix represents the data flow intensity between two processes. The calculation of data transmission requires the analysis of the amount of data interaction between two processes, such as the data mapping relationship between task input and output, while the number of shared resources is based on resource utilization data, and the common occupation of the same resource by two processes during execution is counted. These data provide the degree of direct coupling between processes at the resource and information levels. At the same time, control flow analysis is performed on each pair of processes in the initial dependency graph. By analyzing the execution order constraints and parallelism between processes, the control dependency strength matrix is ​​obtained. The execution order constraint quantifies the sequential dependency relationship between two processes during execution, such as the constraint that a process can only be started after the completion of another process. Parallelism reflects the degree to which two processes can be executed simultaneously or independently of each other. The data dependency strength matrix and the control dependency strength matrix are weighted and combined to generate the process association matrix. The weight setting is adjusted according to the focus in the actual scenario. For example, if data interaction is the main bottleneck, a higher weight is given to data dependency; if the scheduling and execution order of tasks are more critical, the weight of control dependency is increased. Each element of the process association matrix comprehensively reflects the overall dependency strength between the two processes. The Laplace matrix L is constructed based on the process association matrix. The Laplace matrix is ​​calculated by the formula L=D−W, where D is the degree matrix, which represents the total connection weight of each vertex, and W is the process association matrix generated above. The Laplace matrix is ​​a tool used in graph theory. Its essence is to describe the global structural characteristics of the entire graph through the local connection information of vertices. The Laplacian matrix is ​​used to obtain the feature space mapping matrix by solving its eigenvalues ​​and eigenvectors. In the steps of solving the eigenvalues ​​and eigenvectors, the eigenvectors corresponding to the smaller eigenvalues ​​can effectively reveal the cluster structure of the graph, that is, the potential pattern of task grouping. The feature space mapping matrix maps the initial high-dimensional process dependency data to a low-dimensional space, so that processes with strong correlations are clustered in adjacent positions in the feature space, which is convenient for grouping. Spectral clustering operation is performed on the feature space mapping matrix.Perform traditional clustering on the mapped feature vectors, such as using the k-means algorithm, and assign processes with high correlation to the same group based on the clustering results. The clustering process can ensure that the correlation of tasks within a group is maximized when grouping tasks, while minimizing the degree of coupling between tasks between groups. The task grouping data is obtained, reflecting the logical relationship between process tasks.

[0020] Step 400: Input the task grouping data into a two-layer learning network structure. The first layer of the two-layer learning network structure uses a dual-depth Q network to globally optimize resource allocation. The second layer of the network uses a multi-agent deep deterministic policy gradient network to schedule and plan each core task to obtain an initial scheduling plan. Specifically, the task grouping data is input into the first layer of the two-layer learning network structure, namely the dual-depth Q network. The dual-depth Q network consists of a main network and a target network, where the main network is used to calculate the state-action value estimation under the current state in real time, and the target network is used as a reference benchmark for the optimization target. By inputting the task grouping data, the dual-depth Q network extracts the resource requirements and computing power characteristics corresponding to the task grouping, generates a resource allocation state representation, and reflects the resource distribution and task load status of the current chip. The network evaluates the resource allocation state representation through a defined reward function. The reward function is designed based on indicators such as global resource utilization efficiency, task execution completion, and energy consumption to quantify the pros and cons of the current state and action. Through this evaluation process, the network can generate a state-action value evaluation result, namely the Q value. In order to improve the stability and generalization ability of the model, the dual-depth Q network uses an experience replay mechanism. The evaluation results are stored in an experience replay pool D, where each entry contains information about the state, action, reward, and next state. Samples in the experience replay pool are extracted by random sampling to form training batch data N. This mechanism can break the correlation between samples, prevent the model from overfitting to a specific sequence, and improve the robustness of training. The training batch data is input into the target network to calculate the target Q value. The target Q value is calculated based on the Bellman equation, which estimates the future reward value by combining the reward of the current state with the maximum Q value of the next state. After the target Q value is calculated, the temporal difference error is calculated by comparing the target Q value with the Q value predicted by the main network. Based on this error, the mean square error loss function is constructed as the optimization target. The main network parameters are updated using the back propagation algorithm, and the loss function is minimized by the gradient descent method. In order to ensure the stability of the network, the parameters of the main network are copied to the target network every C steps. The separation design of the main network and the target network can avoid the instability caused by the rapid change of the target value during training. After multiple rounds of iterative optimization, the first layer of the network generates a global resource allocation strategy, which aims to optimize resource utilization efficiency and task execution effect globally. The global resource allocation strategy is passed as input to the second layer of the two-layer learning network structure, namely the multi-agent deep deterministic policy gradient network. In this layer, each computing core is regarded as an independent agent, and the agents collaborate to complete the scheduling of local tasks. The core of each agent consists of two networks: the Actor network and the Critic network. The Actor network is used to generate scheduling actions based on the current state, and the Critic network calculates the direction of parameter update by evaluating the degree of match between the Actor network's actions and the global goal. By inputting the global resource allocation strategy, the multi-agent deep deterministic policy gradient network trains each agent separately.During the training process, the Critic network combines the global resource allocation status and the local task execution status to evaluate the scheduling actions generated by the Actor network, and calculates the parameter update direction through the policy gradient method. The feedback of the Critic network is used to guide the update of the Actor network, so that each agent can maintain a collaborative relationship with other agents while meeting its own task requirements. Since each agent has an independent Actor and Critic network, this method can fully capture the characteristics of local tasks and achieve the best overall performance through a multi-agent collaboration mechanism. The local scheduling strategy is strategically integrated with the global resource allocation strategy. Strategy integration is completed through a multi-agent collaborative learning mechanism, which allows each agent to adjust its actions to adapt to the constraints of global resource allocation based on local optimization. The integration process ensures the global consistency of the resource allocation strategy and resolves conflicts in local scheduling. The two-layer learning network structure outputs an initial scheduling plan, which optimizes resource utilization efficiency globally and achieves efficient scheduling at the local task level, meeting the high performance requirements of intelligent control chips in a dynamic computing environment.

[0021] The global resource allocation strategy is state-decomposed to construct an observation space for each computing core. The design of the observation space needs to comprehensively consider the local manufacturing task information, global resource state, and state information of neighboring cores. The local manufacturing task information contains details such as the load, computing requirements, and execution priority of the current task of the core. The global resource state reflects the distribution of computing resources in the entire system, and the neighboring core state is used to capture the interdependence and competition between cores. By integrating and modeling these information, a core observation vector for each computing core is formed. The core observation vector is input into the state encoding layer of the Actor network, which extracts features from the input data and generates a high-dimensional state feature encoding. The main function of the state encoding layer is to convert the original input into a more expressive feature representation to ensure that the subsequent network layers can effectively use these features. After processing, the state feature encoding is input into the policy generation layer of the Actor network. The policy generation layer consists of two layers of fully connected networks, each of which uses the Tanh activation function to map the output to the action space. This design ensures that the generated actions are continuous and controllable, and their range is limited to a specific interval. The deterministic action finally output by the policy generation layer directly reflects the actor network's choice of the optimal scheduling decision under the current state. The core observation vector and the deterministic action are input into the dual-path feature extraction layer of the critic network. The dual-path feature extraction layer of the critic network performs feature transformation on the state and action respectively to generate state-action features. The state-action features are input into the evaluation layer of the critic network for processing. The evaluation layer uses a series of fully connected networks to calculate the action value function (Q value). The action value function represents the expected cumulative reward that can be obtained by selecting a specific action in the current state, which serves as the basis for the critic network to evaluate the output of the actor network. Based on the calculated action value function, the policy objective function is constructed. The policy objective function takes the expected cumulative reward as the optimization goal, and generates the policy gradient by maximizing the expectation of the objective function. The direction of the policy gradient directly determines the update direction of the actor network parameters, thereby gradually optimizing the agent's policy. In this process, the feedback of the critic network plays a guiding role, enabling the actor network to continuously improve the value of the output action during the optimization process. After completing the policy gradient calculation, the network is trained using the actor-critic alternating optimization method. The update of the Critic network parameters is based on the error of the action value function, using the mean square error as the loss function, calculating the loss gradient and optimizing it. The parameter update of the Actor network is based on the policy gradient fed back from the Critic network, and the policy improvement is completed by maximizing the objective function. In order to improve the training efficiency and stability, the Adam optimizer is used to update the parameters of both. The Adam optimizer has the ability to dynamically adjust the learning rate, which can accelerate convergence and reduce training shocks.In order to prevent the target network from changing dramatically during training, the soft update rule is used to update the target network parameters. The soft update rule gradually integrates the parameters of the current network into the target network by setting a smaller update rate, rather than directly overwriting them. The progressive update method ensures the smoothness of network training while retaining the long-term stability of the target network. After multiple rounds of iterations, the optimized Actor network and Critic network parameters are generated, and their performance is improved through soft updates of the target network. The optimized Actor network can output more accurate and efficient local scheduling strategies. The local scheduling strategy is combined with the global resource allocation strategy to provide targeted task scheduling solutions for each computing core, and to achieve efficient planning of each core task while ensuring overall resource utilization.

[0022] Step 500 , perform parallel verification processing on the initial scheduling plan, perform conflict detection and coordination optimization on the local scheduling of each core task through the set process rules, and generate a multi-core collaborative scheduling strategy.

[0023] Specifically, the process rules are verified for each core local scheduling sequence in the initial scheduling scheme. By checking the execution order constraints, quality control accuracy and process parameter requirements between processes, it is ensured that the task scheduling scheme meets the established manufacturing process standards. For example, the execution order constraint requires that some tasks can only be started after their dependent tasks are completed, while the quality control accuracy requires that the tasks are completed within the specified tolerance range. Through the verification steps, the process consistency verification results are formed. The process consistency verification results are input into the task detection unit. The task detection unit consists of a resource competition detection module and a timing dependency detection module, which respectively analyze the resource usage conflicts and time overlap problems in task execution. The resource competition detection module evaluates whether multiple tasks occupy the same shared resources at the same time, such as memory, I / O channels or processor cores, while the timing dependency detection module determines whether there is a time conflict or parallelism problem by analyzing the overlap of task execution time. Through the analysis of these two modules, a set of conflicting task pairs is generated, and each pair of tasks needs to be further optimized to resolve potential conflicts. For the obtained set of conflicting task pairs, a conflict resolution priority graph is constructed. The priority graph determines the processing priority of the task by calculating the weighted combination value of the task urgency (TD) and the process relevance (TR). Task urgency measures the urgency of tasks, for example, tasks with a close deadline have a higher urgency; while process relevance reflects the interdependence between tasks, and tasks with high relevance are solved first to reduce subsequent conflicts. By combining these two indicators, a task processing priority sequence is generated. The task processing priority sequence is input into the coordination processing module, and the tasks are divided by the task segmentation threshold. The setting of the task segmentation threshold needs to minimize the overhead of task switching, including resource reallocation cost and execution delay. For example, a large task is divided into multiple smaller tasks to improve scheduling flexibility and reduce conflicts with other tasks. The coordination processing module generates a task segmentation scheme based on the set threshold and passes it as input to the resource evaluation unit. In the resource evaluation unit, the task segmentation scheme is evaluated by two key indicators: load balancing and resource utilization. Load balancing reflects the uniformity of task distribution on multiple computing cores. Better load balancing can reduce the risk of core overload, while resource utilization measures the efficiency of the use of system computing resources. These two indicators together constitute the resource evaluation function, and the resource allocation evaluation result is obtained through evaluation. The resource allocation evaluation result is input into the performance evaluation network for processing. Performance evaluation network computing task response time (RT) and execution efficiency (EF) quantify the performance of the scheduling scheme from the perspectives of time and efficiency. Task response time represents the total time from task submission to completion, while execution efficiency is the ratio of task execution time to resource usage, reflecting the effectiveness of resource utilization. Through these two performance indicators, a comprehensive performance indicator is generated. Multi-objective optimization is performed on the comprehensive performance indicator.The concept of non-dominated solution set is used in the optimization process to perform a trade-off analysis on multiple objectives (such as RT and EF). The non-dominated solution set refers to the set of solutions that cannot be further optimized on a certain objective without compromising other objectives. These solutions are considered as candidate solution sets in the optimization process. After obtaining the non-dominated solution set, the hierarchical analysis method is used to sort the solution sets. The hierarchical analysis method combines the weights of each objective to calculate the comprehensive score of each solution, and selects the scheduling solution with the highest comprehensive score as the final multi-core collaborative scheduling strategy.

[0024] In the embodiment of the present application, by constructing a dual-channel neural network model, the task characteristics are deeply analyzed from the two dimensions of timing prediction and weight allocation, thereby improving the perception ability of the dynamic characteristics of manufacturing tasks and making resource allocation decisions more accurate; a combination of a dual-depth Q network and a multi-agent deep deterministic policy gradient network in a two-layer learning network structure is adopted to achieve global resource optimization while ensuring the local scheduling efficiency of each computing core; a parallel verification mechanism based on process rules is designed, and conflict detection and coordinated optimization are used to ensure that the scheduling plan meets the manufacturing process requirements and enhance the stability of system operation; a spectral clustering algorithm is introduced to analyze the process correlation matrix, realize automatic grouping of related tasks, and optimize the rationality of task allocation; the final scheduling plan is selected by combining multi-objective optimization and hierarchical analysis, while ensuring load balancing, taking into account task response time and execution efficiency.

[0025] In a specific embodiment, the process of executing step 100 may specifically include the following steps: Collect parameters of the manufacturing process of electronic components in the computing core of the intelligent control chip, including the process calculation complexity, the number of process parameters, the execution cycle and the quality control accuracy, and obtain the process feature vector; Construct a computing resource evaluation matrix R for the process feature vector, where the matrix dimension is m×n, m represents the number of computing cores, and n represents the evaluation parameter dimension, to obtain the initial evaluation matrix; For each computing core in the initial evaluation matrix, three quantitative indicators, process urgency TD, resource occupancy rate RU and process relevance TR, are calculated to obtain the core evaluation indicator set; The core evaluation index set is calculated by weighted calculation formula PI = w 1 ×TD + w 2 ×RU + w 3 ×TR is processed, where w 1 、w 2 、w 3 is the weight coefficient and satisfies w 1 + w 2 + w 3 = 1, get the priority index PI; The priority index PI is arranged and combined according to the computing core number to construct a priority relationship matrix to obtain the task priority matrix, and a dot product operation is performed on each element in the initial evaluation matrix and the corresponding priority index to obtain weighted evaluation data; The weighted evaluation data are normalized according to the core number, and the values ​​are mapped to a uniform interval through linear transformation to obtain standardized evaluation data. The standardized evaluation data are then subjected to column vector extraction and concatenation operations to obtain a resource evaluation vector.

[0026] Specifically, the parameters of the electronic component manufacturing process of each computing core in the chip are collected, including the process calculation complexity, the number of process parameters, the execution cycle and the quality control accuracy. These parameters are recorded as , , , ,in Indicates the computing core number, the value range is , Indicates the corresponding process task number, the value range is By collecting these parameters, the feature vector of each process is constructed : ; Summarize the process feature vectors of each core into a computing resource evaluation matrix , the matrix dimension is ,in is the number of computing cores, is the dimension of the evaluation parameters. The matrix The expression is as follows: ; By building , systematically represents the process characteristics of all computing cores. Finally, three quantitative indicators are calculated for each computing core: process urgency , Resource occupancy rate , process relevance These indicators can reflect the status and requirements of the core in the current task environment. Process urgency It is defined as a comprehensive indicator of task execution cycle and quality control accuracy, and the calculation formula is as follows: ; in, Indicates The core The execution cycle of each process; Indicates Core Quality control accuracy of each process (normalized value); Indicates the value of the maximum execution cycle of all processes, used for normalization. Resource occupancy rate The formula reflects the resource occupation of computational complexity and the number of process parameters: ; in, For the Core The computational complexity of each process; For the Core The denominator represents the total computational complexity of all core tasks and is used for normalization. The task coupling degree between each core and other cores is measured, and the formula is: ; in, Indicates other core numbers; For the Core The calculation complexity of each process; the maximum value of the denominator is used for normalization to ensure that the value is within a unified range. After calculating the three quantitative indicators, they are integrated into the core evaluation indicator set, and the priority index of each core is calculated through a weighted formula. : ; in, are the weight coefficients corresponding to the three indicators, satisfying ; Indicates The priority index of each core reflects the relative priority of its task execution. Arrange by core number and build a priority relationship matrix . Combine the priority index with the initial evaluation matrix Perform dot product operation to generate weighted evaluation data matrix : ; Weighted evaluation data matrix Normalization is performed to map the values ​​to a uniform interval [0,1] through linear transformation. The normalization formula is as follows: ; in, and Respectively represent matrices Extract column vectors from the standardized evaluation data matrix and concatenate these column vectors to form a resource evaluation vector : ; Resource Assessment Vector Provides a unified expression of each core's resource requirements for subsequent resource allocation optimization.

[0027] In a specific embodiment, the process of executing step 200 may specifically include the following steps: The task priority matrix is ​​input into the feature extraction layer of the time series prediction channel in the dual-channel neural network model. The feature extraction layer includes three fully connected layers, each of which is followed by a ReLU activation function and a batch normalization layer to obtain a priority feature map. The priority feature map is input into the LSTM layer of the timing prediction channel. The LSTM layer contains 128 hidden units. The timing features of T consecutive manufacturing cycles are extracted through the forget gate, input gate, and output gate to obtain the timing state sequence. The attention calculation is performed on the timing state sequence to obtain the weighted timing features. The resource evaluation vector is input into the feature transformation layer of the weight allocation channel in the dual-channel neural network model. The feature transformation layer includes four fully connected layers, each of which uses the Tanh activation function and the Dropout layer to obtain the initial weight feature. The initial weight features are normalized to obtain standardized weight features, and the standardized weight features are input into the bidirectional GRU network layer. The bidirectional GRU network layer contains 256 hidden units, and the resource weight vector is obtained by concatenating the forward and reverse hidden states. The weighted time series features and resource weight vectors are fused to obtain a fused feature tensor, which is then input into the output mapping layer in the dual-channel neural network model for processing to obtain the resource utilization data of each computing core.

[0028] Specifically, the task priority matrix The time series prediction channel input into the dual-channel neural network model, Each row represents the priority of a computing core, and the column vector represents the priority of a computing core in a continuous manufacturing cycle. The feature extraction layer consists of three fully connected layers, each followed by a ReLU activation function and a batch normalization layer. Assume that the first Listed as , then The layer feature extraction operation is: ; in, is the initial input; and They are Layer weight matrix and bias vector; ReLU is the activation function; BN It is a batch normalization operation. After three layers of extraction, the priority feature map is obtained. ,in Represents the extracted feature dimension. The priority feature map is input to the LSTM layer of the time series prediction channel. The LSTM layer contains 128 hidden units and extracts time series features through the forget gate, input gate, and output gate. The update formula of LSTM is: ; ; ; ; ; in, They are forget gate, input gate and output gate respectively; It is a memory unit; is a hidden state; is the Sigmoid activation function, and © represents element-by-element multiplication. The output of the LSTM layer after processing is a time series state sequence In order to highlight the characteristic information of key moments, the attention mechanism is used to weight the temporal state sequence. Attention weight The calculation is: ; in is a learnable context vector, and the final weighted temporal feature is: ; The resource evaluation vector Input to the feature transformation layer of the weight distribution channel. The feature transformation layer contains 4 fully connected layers, each of which is transformed by the Tanh activation function and the Dropout layer. Layer feature transformation: ; in After 4 layers of transformation, the initial weight feature is obtained ,in is the feature dimension. The initial weight feature is normalized using the standardization formula: ; in and They are The mean and standard deviation of the column features, the normalized features are . It is input into the bidirectional GRU network layer, and the forward and reverse hidden states of the GRU are updated by the following formula: ; ; The final resource weight vector is: ; The weighted time series features and resource weight vector Perform feature fusion to form a fused feature tensor ; in Indicates the concatenation operation. The fused feature tensor is input to the output mapping layer, and after passing through a fully connected network layer, the resource utilization data of each computing core is generated. : ; Resource utilization data Indicates the resource usage of each core under the current task conditions, thereby supporting dynamic allocation optimization of intelligent chips.

[0029] In a specific embodiment, the process of executing step 300 may specifically include the following steps: The process dependency graph is constructed based on the resource utilization data, where the vertex set represents the manufacturing process task set and the edge set represents the dependency relationship between processes, thus obtaining the initial dependency graph; Perform data flow analysis on each pair of processes in the initial dependency graph, and obtain the data dependency strength matrix by calculating the amount of data transmission and the number of shared resources between processes. Perform control flow analysis on each pair of processes in the initial dependency graph, and obtain the control dependency strength matrix by calculating the execution order constraints and parallelism between processes. The data dependency intensity matrix and the control dependency intensity matrix are weightedly combined to obtain the process correlation matrix, and the Laplace matrix L is constructed based on the process correlation matrix, which is calculated by L = D - W, where D is the degree matrix and W is the process correlation matrix; The eigenvalues ​​and eigenvectors are solved according to the Laplace matrix to obtain the feature space mapping matrix, and the spectral clustering operation is performed on the feature space mapping matrix to obtain the task grouping data.

[0030] Specifically, the process dependency graph is constructed through resource utilization data. Assume that the manufacturing process task set is , where each process task The operation unit corresponding to a computing core. Dependency graph The vertex set represents the process task set, edge set Indicates the dependency between process tasks. Edge weight It is used to describe the dependency strength between two process tasks. The specific weight will be determined in subsequent analysis. The construction of is based on resource utilization data, and the initial connection relationship is generated through the resource sharing and timing constraints of the tasks. Data flow analysis is performed on each pair of process tasks in the initial dependency graph. The core of data flow analysis is to calculate the amount of data transmission and the number of shared resources between processes, thereby generating a data dependency intensity matrix . Each element of the matrix Defined as: ; Its Data Indicates the process and The amount of data transmitted between express and The number of shared resources; the denominator is used for normalization to ensure that the values ​​of the matrix elements are within a uniform scale. Perform control flow analysis on each pair of processes in the initial dependency graph, calculate the execution order constraints and parallelism between processes, and obtain the control dependency strength matrix . Each element of the matrix Defined as: ; Among them, Seq It is a process Must be in The strength of the constraint before completion; yes and Possibility of parallel execution; and is the weight coefficient, which controls the influence of order constraint and parallelism; is a small constant used to avoid the denominator being zero. and the control dependency strength matrix Combine in a weighted manner to generate a process correlation matrix : ; in, is a weight parameter that controls the comprehensive ratio of data dependence and control dependence. Each element of Indicates the process and The comprehensive dependence strength between them. Based on the process correlation matrix , construct the Laplacian matrix The Laplacian matrix is ​​defined as: ; in is the degree matrix, which is a diagonal matrix with diagonal elements is the correlation matrix No. The sum of all elements in a row: ; By solving the Laplace matrix The eigenvalues ​​and eigenvectors of get the feature space mapping matrix .in is the number of clusters, Each column corresponds to The eigenvector is used to map the process task to a low-dimensional space. The smaller the eigenvalue, the better it can reflect the structural information of the graph. Input into the spectral clustering algorithm for clustering. The row vector of Indicates the process The position in the feature space and the clustering algorithm (such as -means) divides these row vectors into Each cluster represents a group of process tasks with strong correlation, and the clustering result is the task grouping data.

[0031] The data dependency intensity matrix and the control dependency intensity matrix are weightedly combined to obtain the process correlation matrix, and the Laplace matrix L is constructed according to the process correlation matrix, including: constructing a nonlinear mapping layer for the data dependency intensity matrix, mapping the matrix elements to the [-1,1] interval through the hyperbolic tangent function and maintaining the data distribution characteristics to obtain a nonlinear data dependency feature matrix; applying a self-attention mechanism to the control dependency intensity matrix, calculating the correlation weights between each pair of processes and performing weighted summation to obtain an enhanced control dependency matrix; inputting the nonlinear data dependency feature matrix and the enhanced control dependency matrix into an adaptive weight generation network, and calculating the dynamic weight coefficient α(t ) and β(t), where t represents the current time step; based on the dynamic weight coefficients α(t) and β(t), tensor fusion operations are performed on the nonlinear data dependency feature matrix and the enhanced control dependency matrix, and the initial process association matrix is ​​obtained by Hadamard product and weighted summation; a multi-scale feature extraction module is applied to the initial process association matrix, and local and global dependency features are extracted through graph convolution layers with different convolution kernel sizes to obtain a multi-scale process association matrix; a weighted adjacency matrix W is constructed according to the multi-scale process association matrix, the original dependency information is retained through residual connections, and the rows are normalized using the Softmax function to obtain a normalized weight matrix; the generalized degree matrix D is calculated for the normalized weight matrix W, where D ii Equal to the i-th row W ij The sum of the pth power, p is a learnable parameter, and the parameterized degree matrix is ​​obtained; according to the formula L=ID (-1 / p) WD (-1 / p) Construct a generalized Laplace matrix, where I is the identity matrix and p is the learnable parameter, and obtain the Laplace matrix.

[0032] In a specific embodiment, the process of executing step 400 may specifically include the following steps: The task grouping data is input into the first layer of the double-depth Q network in the double-layer learning network structure, and the double-depth Q network includes a main network and a target network to obtain a resource allocation state representation; The resource allocation state representation is evaluated through the reward function to obtain the state-action value evaluation result, and an experience replay pool is constructed based on the state-action value evaluation result. The experience replay pool D stores state transition samples, and N samples are extracted through random sampling method to obtain training batch data; Input the training batch data into the target network to calculate the target Q value, construct the loss function through the time difference error, and the loss function is the mean square error loss to obtain the network optimization target; The main network parameters are updated through the back propagation algorithm for the network optimization target, and the main network parameters are copied to the target network every C steps to obtain the global resource allocation strategy; The global resource allocation strategy is input into the multi-agent network of the second layer in the two-layer learning network structure. Each computing core acts as an independent agent to build a deep deterministic policy gradient model including the Actor network and the Critic network, calculate the parameter update direction, and obtain the local scheduling strategy. The local scheduling strategy is integrated with the global resource allocation strategy, and the initial scheduling plan is obtained through a multi-agent collaborative learning mechanism.

[0033] Specifically, group the tasks into data Input the first layer of the two-layer learning network structure, namely the Dual Deep Q Network (D-DQN). Contains one or more strongly associated process tasks. D-DQN includes a main network and a target network. The main network is used to calculate the state-action value in real time. ,in is the current resource allocation status, It's action. are network parameters. The target network is used to calculate the stable target value , which is updated less frequently to ensure the stability of the training process. Encoding the task group data into state representation , describing the current resource allocation, such as the load balancing degree of the core and the progress of the process task completion. The main network inputs the current state and actions , calculate the value estimate of the current state-action The output of the main network is defined as: ; in is the feature map of states and actions, and is a learnable parameter of the network. By designing the reward function To evaluate the resource allocation status, the reward function combines indicators such as resource utilization, task completion, and energy consumption balance, for example: ; in Indicates resource utilization, Indicates the degree of task completion. Indicates energy consumption, , and is the weight coefficient. In order to improve the efficiency and stability of training, the experience replay mechanism is used to build an experience replay pool Each state transition Stored in the replay pool, extracted from the replay pool using random sampling samples, forming a training batch of data . For the target network Calculate target value using Bellman equation : ; in is a discount factor that weighs the impact of future rewards. Based on the target value and the output of the main network , construct the mean square error loss function: ; Optimize the main network parameters through the back-propagation algorithm , every Step 1: Copy the master network parameters to the target network parameters , to ensure the stability of the target network. After the optimization is completed, the global resource allocation strategy is output . Set the global resource allocation strategy Input the second layer of the two-layer learning network, the multi-agent deep deterministic policy gradient network. Each computing core acts as an independent agent , each agent consists of an Actor network and a Critic network. The Actor network generates actions , Critic network evaluates action value In a multi-agent deep deterministic policy gradient network, the update of the Actor network is based on the policy gradient: ; The update of the Critic network is based on minimizing the following loss function: ; in Through multiple rounds of training, the multi-agent deep deterministic policy gradient network generates local scheduling strategies The local scheduling strategy is integrated with the global resource allocation strategy, and a multi-agent collaborative learning mechanism is adopted to ensure that the local optimization of each core meets the global optimization goal. The final output of the initial scheduling plan combines the global resource allocation strategy and the local scheduling strategy to optimize resource utilization and task completion efficiency.

[0034] In a specific embodiment, the execution step inputs the global resource allocation strategy into the multi-agent network of the second layer in the two-layer learning network structure, each computing core acts as an independent agent, constructs a deep deterministic policy gradient model including an Actor network and a Critic network, calculates the parameter update direction, and obtains the local scheduling strategy. The process may specifically include the following steps: The global resource allocation strategy is decomposed into states, and an observation space including local manufacturing task information, global resource state and neighboring core state is constructed for each computing core to obtain the core observation vector. Input the core observation vector into the state encoding layer of the Actor network to obtain the state feature encoding; The state feature encoding is processed by the strategy generation layer, which contains a 2-layer fully connected network. The Tanh activation function is used to map the output to the action space to obtain deterministic actions. The core observation vector and deterministic action are input into the dual-path feature extraction layer of the Critic network, and the state and action are transformed respectively to obtain the state-action feature; The state-action feature is input into the evaluation layer of the Critic network for processing to obtain the action value function; Construct the policy objective function based on the action value function, calculate the policy gradient by maximizing the expected cumulative reward, and obtain the parameter update direction; According to the parameter update direction, the network training is performed through Actor-Critic alternating optimization, and the Adam optimizer is used to update the Actor network and Critic network parameters to obtain the optimized network parameters. The optimized network parameters are then updated to the target network through soft update rules to generate a local scheduling strategy.

[0035] Specifically, the global resource allocation strategy is decomposed into states, and an observation space is constructed for each computing core to provide local manufacturing task information, global resource status, and the status of neighboring cores. The design of the observation space determines the task and resource environment that each core can perceive. The observation vector is used Indicates that is the number of the computing core. ,in Is the core Local task information, such as the computational complexity and execution priority of the current task; Indicates the global resource status, including the overall resource utilization and remaining capacity; Indicates the status of neighboring cores, including The observation vector of each core is Input to the state encoding layer of the Actor network, the state encoding layer extracts features through the fully connected layer and maps them to a high-dimensional feature space. Layer, the state encoding formula is: ; in, is the initial input; and It is Layer weights and biases; is the activation function, commonly used ReLU or Sigmoid. After passing through the state encoding layer, the state feature encoding is obtained , used to describe the core The state feature encoding is mapped to the action space through the strategy generation layer. The strategy generation layer consists of two layers of fully connected networks, and the Tanh activation function is used to limit the output value to the interval [-1, 1]. The calculation formula for deterministic actions is: ; in, and are the parameters of the fully connected layer; is the activation function used to map the output. Deterministic actions Indicates the core Scheduling decisions, such as allocating computing resources or adjusting task priorities. and deterministic actions Input to the dual-path feature extraction layer of the Critic network to perform feature transformation on the state and action respectively. The formulas for state feature transformation and action feature transformation are: ; ; in, and is the weight matrix of state and action; and is the bias term. and action features Fusion, get state-action features : ; in Indicates feature concatenation operation. State-action feature Input to the evaluation layer of the Critic network to calculate the action value function : ; in, and are the weights and biases of the evaluation layer. The goal of the Critic network is to Evaluate the quality of the current action and provide feedback to the Actor network. By maximizing the expected cumulative reward , construct the strategy objective function: ; The policy gradient is calculated using the following formula: ; The update of the Critic network is achieved by minimizing the following loss function: ; The target value Defined as: ; By using the Adam optimizer, the parameters of the Actor network and the Critic network are updated separately. The updates of the Actor and Critic network parameters are performed alternately, so that the two networks can jointly optimize the policy performance. In order to ensure the stability of training, the target network parameters are updated using the soft update rule: ; in is the update rate. Through multiple rounds of optimization, a local scheduling strategy is generated ,This strategy can provide an efficient task scheduling scheme for each computing core, and coordinate with the global resource allocation strategy to achieve dynamic computing resource allocation optimization of the intelligent control chip.

[0036] In a specific embodiment, the process of executing step 500 may specifically include the following steps: Verify the process rules of each core local scheduling sequence in the initial scheduling plan, and obtain the process consistency verification results by checking the execution order constraints, quality control accuracy and process parameter requirements between processes; The process consistency verification result is input into the task detection unit. The task detection unit includes a resource competition detection module and a timing dependency detection module. By analyzing the task execution time overlap and shared resource occupancy, a set of conflicting task pairs is obtained. Construct a conflict resolution priority graph for the set of conflicting task pairs, and obtain the task processing priority sequence by calculating the weighted combination value of task urgency TD and process relevance TR; The task processing priority sequence is input into the coordination processing module, and the coordination processing module divides the tasks according to the task segmentation threshold. The task segmentation threshold is determined by minimizing the task switching overhead to obtain a task segmentation scheme; The task partitioning scheme is input into the resource evaluation unit for processing. The resource evaluation unit uses two indicators, load balancing and resource utilization, to construct an evaluation function and obtain a resource allocation evaluation result. Input the resource allocation evaluation results into the performance evaluation network, calculate the task response time RT and execution efficiency EF, and obtain the comprehensive performance index; Multi-objective optimization is performed on the comprehensive performance indicators to obtain the non-dominated solution set, and a hierarchical analysis is performed on the non-dominated solution set to select the scheduling scheme with the highest comprehensive score as the multi-core collaborative scheduling strategy.

[0037] Specifically, the local scheduling sequence of each computing core is extracted from the initial scheduling plan. Each sequence contains a set of process tasks that a computing core needs to complete within a specific time period. For each local scheduling sequence, the execution order constraints, quality control accuracy and process parameter requirements between processes are checked to verify whether it meets the preset process rules. The execution order constraints are represented by a directed graph, and the vertex set represents the process task, the edge set Indicates the dependency between tasks. Must be in If completed before, The verification condition is that the execution order of all tasks must conform to the topological sorting. Quality Control Accuracy and process parameter requirements The minimum threshold requirements for the task must be met: ; After verification, the process consistency verification results are obtained , which contains the verification pass or fail information of each core local scheduling. The process consistency verification result is input into the task detection unit, which consists of a resource competition detection module and a timing dependency detection module. The resource competition detection module analyzes the occupancy of shared resources and uses Indicates the process and If there is overlap in the number of resources occupied at the same time, it is determined that there is a resource conflict. The timing dependency detection module identifies timing conflicts by calculating the overlap of task execution time, which is expressed as: ; in The processes are The start time and end time of , then there is a time conflict. After analysis, the task detection unit outputs a set of conflicting task pairs: , represents all conflicting task pairs. For the set of conflicting task pairs, a conflict resolution priority graph is constructed. The priority of each task is determined by the urgency of the task. Relevance to process The weighted combination value of determines: ; in, Indicates the urgency of the task. For the task Deadline: represents the correlation between tasks, It's a task and The strength of dependence. and is the weight coefficient, satisfying The priority sequence is input into the coordination processing module, and the coordination processing module divides the tasks according to the task segmentation threshold. Task segmentation threshold It is determined by minimizing the task switching overhead, which is calculated using the following formula: ; in, is the time overhead of task switching, is the occupation cost of switching resources. By optimizing the goal: ; The task partitioning scheme is input into the resource evaluation unit, and the resource evaluation unit uses load balancing to and resource utilization The load balancing degree indicates the uniformity of task distribution among cores and is defined as: ; in, It is Core workloads, is the average load of all cores. Resource utilization is defined as: ; The load balancing degree and resource utilization rate are combined to form the resource allocation evaluation function: ; The evaluation results are input into the performance evaluation network to calculate the task response time and execution efficiency The task response time is the average time taken for all tasks to complete: ; Execution efficiency is the ratio of the effective workload to the total workload to complete the task: ; Combination and The comprehensive performance index of the system is constructed by multi-objective optimization to construct a non-dominated solution set. The optimization objective is: ; The non-dominated solution set is analyzed hierarchically, and the optimal scheduling scheme is selected through comprehensive scoring to form a multi-core collaborative scheduling strategy. This strategy achieves a balance in terms of task completion time, resource utilization, and energy consumption, and is the optimal solution for multi-core task scheduling.

[0038] The above describes the method for dynamically allocating computing resources of the intelligent control chip in the embodiment of the present application. The following describes the system 10 for dynamically allocating computing resources of the intelligent control chip in the embodiment of the present application. Figure 2 In the embodiment of the present application, an embodiment of the computing resource dynamic allocation system 10 of the intelligent control chip includes: The acquisition module 11 is used to collect characteristic parameters and perform weighted calculation on the computing core in the intelligent control chip to obtain a task priority matrix and a resource evaluation vector; The computing module 12 is used to input the task priority matrix and the resource evaluation vector into the dual-channel neural network model, and perform calculations through the timing prediction channel and the weight allocation channel in the dual-channel neural network model to obtain the resource utilization data of each computing core; The analysis module 13 is used to perform process dependency analysis on the resource utilization data to obtain a process correlation matrix, and perform spectral clustering operation on the process correlation matrix to obtain task grouping data; The planning module 14 is used to input the task grouping data into a two-layer learning network structure. The first layer of the two-layer learning network structure uses a dual-depth Q network to perform global optimization on resource allocation, and the second layer of the network uses a multi-agent deep deterministic policy gradient network to perform scheduling planning on each core task to obtain an initial scheduling plan. The generation module 15 is used to perform parallel verification processing on the initial scheduling plan, perform conflict detection and coordination optimization on the local scheduling of each core task through the set process rules, and generate a multi-core collaborative scheduling strategy.

[0039] Through the collaborative cooperation of the above components and by building a dual-channel neural network model, the task characteristics are deeply analyzed from the two dimensions of time series prediction and weight allocation, which improves the perception of the dynamic characteristics of manufacturing tasks and makes resource allocation decisions more accurate; the dual-depth Q network and the multi-agent deep deterministic policy gradient network in the double-layer learning network structure are combined to achieve global resource optimization while ensuring the local scheduling efficiency of each computing core; a parallel verification mechanism based on process rules is designed, which ensures that the scheduling plan meets the manufacturing process requirements through conflict detection and coordinated optimization, and enhances the stability of system operation; the spectral clustering algorithm is introduced to analyze the process correlation matrix, realize the automatic grouping of related tasks, and optimize the rationality of task allocation; the final scheduling plan is selected by combining multi-objective optimization and hierarchical analysis, which takes into account task response time and execution efficiency while ensuring load balancing.

[0040] See also Figure 3 , Figure 3This is a schematic block diagram of the structure of an electronic device 300 provided in an embodiment of the present application. The electronic device 300 includes a processor 301 and a memory 302. The processor 301 and the memory 302 are connected via a system bus 303, wherein the memory 302 may include a non-volatile storage medium and an internal memory.

[0041] The non-volatile storage medium can store a computer program. The computer program includes program instructions, and when the program instructions are executed by the processor 301, the processor 301 can execute any of the above-mentioned methods for dynamically allocating computing resources of an intelligent control chip.

[0042] The processor 301 is used to provide computing and control capabilities to support the operation of the entire electronic device 300 .

[0043] The internal memory provides an environment for the operation of the computer program in the non-volatile storage medium. When the computer program is executed by the processor 301, the processor 301 can execute any of the above-mentioned methods for dynamically allocating computing resources of the intelligent control chip.

[0044] Those skilled in the art will understand that Figure 3 The structure shown in the figure is only a block diagram of a partial structure related to the present application scheme, and does not constitute a limitation on the electronic device 300 involved in the present application scheme. The specific electronic device 300 may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0045] It should be understood that the processor 301 may be a central processing unit (CPU), and the processor 301 may also be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among them, the general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0046] It should be noted that technicians in the relevant field can clearly understand that for the convenience and simplicity of description, the specific working process of the electronic device 300 described above can refer to the corresponding process of the dynamic allocation method of computing resources of the aforementioned intelligent control chip, and will not be repeated here.

[0047] An embodiment of the present application also provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by one or more processors, the one or more processors implement the method for dynamically allocating computing resources of an intelligent control chip as provided in an embodiment of the present application.

[0048] The computer-readable storage medium may be an internal storage unit of the electronic device 300 in the aforementioned embodiment, such as a hard disk or memory of the electronic device 300. The computer-readable storage medium may also be an external storage device of the electronic device 300, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), etc. equipped with the electronic device 300.

[0049] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0050] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable an electronic device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk, etc., various media that can store program codes.

[0051] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application 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 make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for dynamically allocating computing resources of an intelligent control chip, characterized in that: The method comprises: Perform characteristic parameter collection and weighted calculation on the computing core in the intelligent control chip to obtain the task priority matrix and resource evaluation vector; Inputting the task priority matrix and the resource evaluation vector into a dual-channel neural network model, performing calculations through a timing prediction channel and a weight allocation channel in the dual-channel neural network model, and obtaining resource utilization data of each computing core; Performing process dependency analysis on the resource utilization data to obtain a process correlation matrix, and performing spectral clustering operation on the process correlation matrix to obtain task grouping data; The task grouping data is input into a two-layer learning network structure, wherein the first layer of the network uses a dual-depth Q network to globally optimize resource allocation, and the second layer of the network uses a multi-agent deep deterministic policy gradient network to schedule and plan each core task, thereby obtaining an initial scheduling plan; The initial scheduling scheme is verified in parallel, and the local scheduling of each core task is conflict-detected and coordinated and optimized according to the set process rules to generate a multi-core collaborative scheduling strategy.

2. The method for dynamically allocating computing resources of an intelligent control chip according to claim 1, characterized in that: The characteristic parameter collection and weighted calculation of the computing core in the intelligent control chip to obtain the task priority matrix and resource evaluation vector include: Collect parameters of the manufacturing process of electronic components in the computing core of the intelligent control chip, including the process calculation complexity, the number of process parameters, the execution cycle and the quality control accuracy, and obtain the process feature vector; Constructing a computing resource evaluation matrix R for the process feature vector, wherein the matrix dimension is m×n, m represents the number of computing cores, and n represents the evaluation parameter dimension, to obtain an initial evaluation matrix; For each computing core in the initial evaluation matrix, three quantitative indicators, namely, process urgency TD, resource occupancy rate RU and process relevance TR, are calculated to obtain a core evaluation indicator set; The core evaluation indicator set is processed by a weighted calculation formula PI = w1×TD + w2×RU + w3×TR, where w1, w2, and w3 are weight coefficients and satisfy w1 + w2 + w3 = 1, to obtain a priority index PI; The priority index PI is arranged and combined according to the calculation core number, a priority relationship matrix is ​​constructed, a task priority matrix is ​​obtained, and a dot product operation is performed on each element in the initial evaluation matrix and the corresponding priority index to obtain weighted evaluation data; The weighted evaluation data is normalized according to the core number, and the numerical value is mapped to a uniform interval through linear transformation to obtain standardized evaluation data, and column vector extraction and splicing operations are performed on the standardized evaluation data to obtain a resource evaluation vector.

3. The method for dynamically allocating computing resources of an intelligent control chip according to claim 2, characterized in that: The task priority matrix and the resource evaluation vector are input into a dual-channel neural network model, and the resource utilization data of each computing core is obtained by performing calculations through a timing prediction channel and a weight allocation channel in the dual-channel neural network model, including: Inputting the task priority matrix into the feature extraction layer of the time series prediction channel in the dual-channel neural network model, the feature extraction layer includes three fully connected layers, each of which is followed by a ReLU activation function and a batch normalization layer to obtain a priority feature map; Inputting the priority feature map into the LSTM layer of the timing prediction channel, the LSTM layer includes 128 hidden units, extracting the timing features of T consecutive manufacturing cycles through a forget gate, an input gate, and an output gate to obtain a timing state sequence, and performing attention calculation on the timing state sequence to obtain a weighted timing feature; Input the resource evaluation vector into the feature transformation layer of the weight allocation channel in the dual-channel neural network model, wherein the feature transformation layer includes four fully connected layers, each of which uses a Tanh activation function and a Dropout layer to obtain an initial weight feature; Normalizing the initial weight features to obtain standardized weight features, and inputting the standardized weight features into a bidirectional GRU network layer, wherein the bidirectional GRU network layer includes 256 hidden units, and a resource weight vector is obtained by concatenating forward and reverse hidden states; The weighted time series feature and the resource weight vector are subjected to feature fusion to obtain a fused feature tensor, and the fused feature tensor is input into the output mapping layer in the dual-channel neural network model for processing to obtain resource utilization data of each computing core.

4. The method for dynamically allocating computing resources of an intelligent control chip according to claim 3, characterized in that: The step of performing process dependency analysis on the resource utilization data to obtain a process correlation matrix, and performing spectral clustering operation on the process correlation matrix to obtain task grouping data includes: Constructing a process dependency graph according to the resource utilization data, wherein a vertex set represents a set of manufacturing process tasks, and an edge set represents a dependency relationship between processes, to obtain an initial dependency graph; Performing data flow analysis on each pair of processes in the initial dependency graph, and obtaining a data dependency strength matrix by calculating the amount of data transmission and the number of shared resources between processes; and performing control flow analysis on each pair of processes in the initial dependency graph, and obtaining a control dependency strength matrix by calculating the execution order constraints and parallelism between processes; The data dependency intensity matrix and the control dependency intensity matrix are weightedly combined to obtain a process correlation matrix, and a Laplace matrix L is constructed according to the process correlation matrix, and is calculated by L = D - W, where D is a degree matrix and W is the process correlation matrix; The eigenvalues ​​and eigenvectors are solved according to the Laplace matrix to obtain a feature space mapping matrix, and a spectral clustering operation is performed on the feature space mapping matrix to obtain task grouping data.

5. The method for dynamically allocating computing resources of an intelligent control chip according to claim 4, characterized in that: The task grouping data is input into a double-layer learning network structure, the first layer of the double-layer learning network structure uses a dual-depth Q network to globally optimize resource allocation, and the second layer of the network uses a multi-agent deep deterministic policy gradient network to schedule and plan each core task, and obtain an initial scheduling plan, including: Inputting the task grouping data into a dual-depth Q network of the first layer in a dual-layer learning network structure, wherein the dual-depth Q network includes a main network and a target network, to obtain a resource allocation state representation; The resource allocation state representation is evaluated by a reward function to obtain a state-action value evaluation result, and an experience replay pool is constructed based on the state-action value evaluation result. The experience replay pool D stores state transition samples, and N samples are extracted by a random sampling method to obtain training batch data; Input the training batch data into the target network to calculate the target Q value, construct a loss function through the temporal difference error, the loss function is the mean square error loss, and obtain the network optimization target; The main network parameters are updated by the back propagation algorithm for the network optimization target, and the main network parameters are copied to the target network every C steps to obtain a global resource allocation strategy; The global resource allocation strategy is input into the multi-agent network of the second layer in the two-layer learning network structure, each computing core is used as an independent agent, a deep deterministic policy gradient model including an actor network and a critic network is constructed, the parameter update direction is calculated, and a local scheduling strategy is obtained; The local scheduling strategy is integrated with the global resource allocation strategy, and an initial scheduling solution is obtained through a multi-agent collaborative learning mechanism.

6. The method for dynamically allocating computing resources of an intelligent control chip according to claim 5, characterized in that: The global resource allocation strategy is input into the multi-agent network of the second layer in the two-layer learning network structure, each computing core is used as an independent agent, a deep deterministic policy gradient model including an Actor network and a Critic network is constructed, and the parameter update direction is calculated to obtain a local scheduling strategy, including: Decomposing the global resource allocation strategy in a state, constructing an observation space including local manufacturing task information, global resource state and neighboring core state for each computing core, and obtaining a core observation vector; Inputting the core observation vector into the state encoding layer of the Actor network to obtain a state feature encoding; The state feature encoding is processed through a strategy generation layer, wherein the strategy generation layer includes a 2-layer fully connected network, and a Tanh activation function is used to map the output to the action space to obtain a deterministic action; Inputting the core observation vector and the deterministic action into the dual-path feature extraction layer of the Critic network, performing feature transformation on the state and the action respectively, and obtaining the state-action feature; The state-action feature is input into the evaluation layer of the Critic network for processing to obtain an action value function; Constructing a policy objective function according to the action value function, calculating the policy gradient by maximizing the expected cumulative reward, and obtaining the parameter update direction; The network training is performed through Actor-Critic alternating optimization according to the parameter update direction, and the Actor network and Critic network parameters are updated using the Adam optimizer to obtain optimized network parameters. The optimized network parameters are updated to the target network through soft update rules to generate a local scheduling strategy.

7. The method for dynamically allocating computing resources of an intelligent control chip according to claim 6, characterized in that: The initial scheduling scheme is verified in parallel, and the local scheduling of each core task is conflict detected and coordinated and optimized through the set process rules to generate a multi-core collaborative scheduling strategy, including: Performing process rule verification on each core local scheduling sequence in the initial scheduling plan, and obtaining process consistency verification results by checking execution order constraints, quality control accuracy and process parameter requirements between processes; Inputting the process consistency verification result into a task detection unit, wherein the task detection unit includes a resource competition detection module and a timing dependency detection module, and obtaining a set of conflicting task pairs by analyzing the task execution time overlap and shared resource occupancy; Constructing a conflict resolution priority graph for the conflict task pair set, and obtaining a task processing priority sequence by calculating a weighted combination value of task urgency TD and process relevance TR; The task processing priority sequence is input into a coordination processing module, and the coordination processing module divides the tasks according to a task segmentation threshold, and the task segmentation threshold is determined by minimizing the task switching overhead to obtain a task segmentation scheme; The task partitioning scheme is input into a resource evaluation unit for processing, wherein the resource evaluation unit constructs an evaluation function using two indicators, load balancing and resource utilization, to obtain a resource allocation evaluation result; Input the resource allocation evaluation result into the performance evaluation network, calculate the task response time RT and execution efficiency EF, and obtain a comprehensive performance index; The comprehensive performance index is optimized by multiple objectives to obtain a non-dominated solution set, and the non-dominated solution set is analyzed hierarchically to select a scheduling solution with the highest comprehensive score as the multi-core collaborative scheduling strategy.

8. A computing resource dynamic allocation system for an intelligent control chip, characterized in that: A method for dynamically allocating computing resources of an intelligent control chip according to any one of claims 1 to 7, wherein the computing resource dynamic allocation system of the intelligent control chip comprises: The acquisition module is used to collect characteristic parameters and perform weighted calculations on the computing core in the intelligent control chip to obtain the task priority matrix and resource evaluation vector; A calculation module, used for inputting the task priority matrix and the resource evaluation vector into a dual-channel neural network model, performing calculations through a timing prediction channel and a weight allocation channel in the dual-channel neural network model, and obtaining resource utilization data of each computing core; An analysis module, used to perform process dependency analysis on the resource utilization data to obtain a process correlation matrix, and perform spectral clustering operation on the process correlation matrix to obtain task grouping data; A planning module is used to input the task grouping data into a two-layer learning network structure, wherein the first layer of the network in the two-layer learning network structure uses a dual-depth Q network to perform global optimization on resource allocation, and the second layer of the network uses a multi-agent deep deterministic policy gradient network to perform scheduling planning on each core task to obtain an initial scheduling plan; The generation module is used to perform parallel verification processing on the initial scheduling plan, perform conflict detection and coordination optimization on the local scheduling of each core task through the set process rules, and generate a multi-core collaborative scheduling strategy.

9. An electronic device, characterized in that: The electronic device comprises: a memory and at least one processor, wherein instructions are stored in the memory; The at least one processor calls the instruction in the memory so that the electronic device executes the method for dynamically allocating computing resources of an intelligent control chip as described in any one of claims 1-7.

10. A computer-readable storage medium having instructions stored thereon, characterized in that: When the instructions are executed by the processor, the method for dynamically allocating computing resources of the intelligent control chip according to any one of claims 1 to 7 is implemented.

Images