A deep learning-based design task intelligent decomposition method

Abstract

The application discloses a kind of based on deep learning's design task intelligent decomposition method, it is related to intelligent design technical field, including: to task text data is carried out semantic analysis, obtain semantic representation;Dynamic cognitive arrangement atlas is constructed based on semantic representation, and the candidate set of multiple granularity subtasks is generated based on dynamic cognitive arrangement atlas, and dynamic cognitive arrangement atlas is updated;Based on the updated dynamic cognitive arrangement atlas and subtask candidate set, the optimal subtask disassembly path sequence is planned out through sequence decision model;Based on optimal subtask disassembly path sequence output structured subtask tree.The application can realize high-precision, high-efficiency task decomposition under complex task scene, solve the deficiency of traditional method in the aspect of inconsistent granularity, logic conflict and task omission, with strong practicality and popularization value.

Description

Technical Field

[0001] This invention relates to the field of intelligent design technology, and more specifically, to a method for intelligent decomposition of design tasks based on deep learning. Background Technology

[0002] Currently, intelligent decomposition methods for design tasks largely rely on a combination of Natural Language Processing (NLP) and rule-based reasoning to achieve structured understanding and automatic decomposition of design task text. Design tasks in fields such as engineering design, software development, and product architecture are typically highly abstract, semantically dependent, and involve complex contextual relationships. Traditional methods often employ keyword extraction, template matching, expert rule base reasoning, or logical induction based on shallow semantic models for task segmentation. However, these methods have significant limitations when dealing with complex, multi-level, and highly dependent design tasks. Particularly in modeling the internal cognitive structure of tasks, traditional methods cannot effectively capture the multi-granular relationships and inherent logic between tasks, resulting in an inability to meet the needs of multi-stage path planning and structured subtask generation for complex design goals.

[0003] Some studies have introduced deep learning techniques based on recurrent neural networks, graph neural networks, or BERT-like pre-trained models to extract task semantic features and local contextual information, thereby assisting in task decomposition. However, existing methods fail to fully consider the semantic dependencies and structured relationships between multi-level goals and operations in design tasks, making it difficult to cope with the dynamic changes of complex tasks and the generation of multi-granularity subtasks. Path planning in existing technologies is often based on static task states, lacking the ability to dynamically perceive the evolution of design goals, task state transitions, and structural dependencies, making path generation strategies difficult to adapt to the changing needs of complex design tasks. Summary of the Invention

[0004] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a deep learning-based intelligent decomposition method for design tasks, which solves the problems of insufficient modeling of the internal cognitive structure of tasks and lack of dynamic perception in path planning in the prior art, and achieves more efficient and flexible decomposition of design tasks.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A deep learning-based intelligent decomposition method for design tasks includes the following steps: performing semantic analysis on task text data to obtain semantic representation; constructing a dynamic cognitive orchestration graph based on the semantic representation; generating a multi-granularity subtask candidate set based on the dynamic cognitive orchestration graph and updating the dynamic cognitive orchestration graph; planning the optimal subtask decomposition path sequence through a sequence decision model based on the updated dynamic cognitive orchestration graph and the subtask candidate set; and outputting a structured subtask tree based on the optimal subtask decomposition path sequence.

[0007] In a preferred embodiment, the step of performing semantic analysis on the task text data to obtain a semantic representation specifically involves: parsing the design task text data to obtain semantic units; processing the semantic units using a pre-trained language model to obtain word-level contextual representations; extracting target intent features and design constraint features from the contextual representations or text data; fusing the target intent features and design constraint features to form a unified representation vector; and mapping the unified representation vector to a semantic representation.

[0008] In a preferred embodiment, the step of constructing a dynamic cognitive orchestration graph based on semantic representation specifically involves: performing cluster analysis on the semantic representation to form multiple cognitive operation units, and assigning a node identifier to each unit as a graph node; calculating the semantic correlation degree between cognitive operation units; connecting the corresponding graph nodes based on the semantic correlation degree to generate a preliminary graph; and optimizing the preliminary graph to generate a dynamic cognitive orchestration graph.

[0009] In a preferred embodiment, updating the dynamic cognitive orchestration graph specifically involves: extracting node representations and their hierarchical information from the dynamic cognitive orchestration graph; generating high-level semantic representations and low-level structural representations based on the node representations; fusing the high-level semantic representations and low-level structural representations and updating each node representation, wherein the fusion includes calculating the correlation degree between the high-level semantic representations and low-level structural representations and performing weighted aggregation based on the correlation degree; generating a multi-granularity subtask candidate set based on the updated node representations; and updating the dynamic cognitive orchestration graph using the subtask candidate set.

[0010] In a preferred embodiment, the step of planning the optimal subtask decomposition path sequence based on the updated dynamic cognitive orchestration graph and subtask candidate set through a sequence decision model specifically involves: encoding the current dynamic cognitive orchestration graph and subtask candidate set to obtain a joint state representation; inputting the joint state representation into the sequence decision model to obtain an action probability distribution; selecting nodes from the subtask candidate set according to the action probability distribution and adding the selected nodes to the decomposition path; updating the current joint state representation and iteratively executing the above selection and update steps until the termination condition is met, and outputting the optimal subtask decomposition path sequence.

[0011] In a preferred embodiment, the sequence decision model is an improved meta-learning policy network, which specifically involves: introducing a path memory guidance mechanism and a reversible compressed mapping structure into the meta-learning policy network; retaining the structural pattern of the optimal decomposition sequence in the historical task through the path memory guidance mechanism, and using it for the decomposition initialization of similar subgraphs; and mapping the dynamic cognitive graph to a low-dimensional latent space and achieving reversible decoding through the reversible compressed mapping structure, so that the meta-learning policy network can inductively learn the mapping state of the dynamic cognitive graph in a unified latent space.

[0012] In a preferred embodiment, the step of encoding the current dynamic cognitive orchestration graph and the subtask candidate set to obtain a joint state representation specifically involves: learning the topological structure of the dynamic cognitive orchestration graph through a graph neural network to generate a graph structure vector; performing multi-feature fusion encoding on the subtask candidate set to generate a task embedding vector set; and concatenating the graph structure vector and the task embedding vector set to form a joint state representation.

[0013] In a preferred embodiment, the step of outputting a structured subtask tree based on the optimal subtask decomposition path sequence specifically involves: perturbing the optimal subtask decomposition path sequence to generate a counterfactual decomposition path sequence; inputting the optimal subtask decomposition path sequence and the counterfactual decomposition path sequence into an embedding encoder for comparison to obtain their respective embedding representations; and outputting the final structured subtask tree based on the self-embedding representations.

[0014] In a preferred embodiment, the counterfactual decomposition path sequence is generated by the following steps: obtaining the order and semantic embedding representation of each cognitive operation unit in the subtask decomposition path sequence; performing a structural perturbation operation on the path sequence to generate multiple candidate path sequences; applying semantic relevance constraints to each generated candidate path sequence, and selecting sequences whose semantic relevance to the original path sequence is higher than a preset threshold, and recording them as counterfactual decomposition path sequences.

[0015] In a preferred embodiment, the output structured subtask tree includes: inputting the optimal subtask decomposition path sequence and the counterfactual decomposition path sequence as positive samples and contrast samples, respectively, into a contrastive embedding encoder; encoding the cognitive operation unit sequences in each path through the embedding network in the contrastive embedding encoder to generate a path-level semantic embedding representation; optimizing the embedding network to maximize the embedding distance between the positive samples and the contrast samples; and generating a structured subtask tree based on the optimized optimal path semantic embedding representation according to a multidimensional scoring criterion.

[0016] This invention constructs a deep learning-based dynamic cognitive orchestration graph, which comprehensively captures the complex semantic dependencies and hierarchical structures among subtasks in a design task. Through the generation of multi-granularity subtask candidate sets and dynamic updating of the graph, task decomposition strategies can be flexibly adjusted for different design tasks. Compared with traditional methods, especially when dealing with complex tasks with multi-level, multi-dimensional, and strong dependencies, this invention significantly improves the accuracy and efficiency of task decomposition and effectively addresses the limitations of traditional methods in complex design tasks. Furthermore, by introducing a sequence decision model for dynamic planning of task decomposition paths, this invention can optimize the decomposition paths in real time based on the evolution of task objectives, changes in design states, and adjustments to structural dependencies. Compared to the static decision-making of traditional path planning methods, this invention can dynamically perceive and adapt to changes in design objectives and task states, overcoming the static limitations of traditional path planning methods and improving the system's adaptability and ability to handle complex design tasks. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating a deep learning-based intelligent decomposition method for design tasks according to the present invention.

[0018] Figure 2 This is a flowchart illustrating the process of generating the optimal subtask decomposition path sequence in a deep learning-based intelligent decomposition method for design tasks according to the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0020] Example 1, Figure 1 This invention presents an intelligent decomposition method for design tasks based on deep learning, comprising:

[0021] S1, perform semantic analysis on the task text data to obtain a semantic representation;

[0022] In this embodiment, the semantic analysis of the task text data to obtain a semantic representation specifically includes:

[0023] The system receives design task text data and segments it into sentences. Using existing text segmentation tools, the task text is divided into multiple sentences, ensuring each sentence has a clear grammatical structure. The segmented text undergoes grammatical structure parsing, employing dependency parsing to analyze the grammatical relationships between words, identifying noun phrases, verb phrases, modification relationships, etc., to construct basic semantic units. These semantic units can be fundamental components of the task, such as objectives, constraints, and operations, serving as the basic units for subsequent analysis.

[0024] Semantic units are input into a pre-trained language model, which uses its word-level contextual representation capabilities to obtain the contextual representation of each unit. The pre-trained language model has been trained on a large-scale corpus and is able to capture complex contextual relationships between words.

[0025] The obtained context representations or text data are classified. Using word-level vectors output by a pre-trained model, a Convolutional Neural Network (CNN) or a Long Short-Term Memory (LSTM) network is used to extract features from the context representation. Based on this, a classifier (e.g., a fully connected layer plus a softmax layer) is used to extract semantic fragments relevant to the design task from the context representation and label them as target intent features. These features include the core objectives of the design task, such as functional requirements and performance requirements. Through feature labeling, these target intent features are isolated from the semantic units, providing clear target guidance for subsequent task decomposition.

[0026] The analysis continues with the textual data of the design task, extracting restrictive conditions such as technical limitations, legal compliance requirements, and resource constraints. These restrictive conditions affect the direction and feasibility of task decomposition. For each restrictive condition in the design task, information extraction (IE) techniques and pattern recognition algorithms, such as named entity recognition (NER) and relation extraction, are used to automatically identify and mark the design constraint features.

[0027] The extracted target intent features and design constraint features are jointly modeled using a multi-channel fusion network (e.g., a dual-channel neural network or a multimodal network). In this step, different network channels are used to process the target intent features and design constraint features separately. During processing, the target intent feature channel focuses on modeling the task objective, while the design constraint feature channel focuses on modeling the constraints. Each channel uses network structures such as convolutional layers, pooling layers, and fully connected layers to extract key feature information. Through fusion strategies (such as feature concatenation, weighted averaging, etc.), the target intent features and design constraint features are integrated into a unified embedding space to generate a unified representation vector. This vector contains not only the semantic information of the task objective but also the feature information of the task constraints.

[0028] The unified representation vector is input into a nonlinear mapping network, and a nonlinear transformation is performed using an activation function to generate the final semantic representation.

[0029] S2, construct a dynamic cognitive orchestration graph based on semantic representation, generate a multi-granularity subtask candidate set based on the dynamic cognitive orchestration graph, and update the dynamic cognitive orchestration graph;

[0030] In this embodiment, the construction of a dynamic cognitive orchestration graph based on semantic representation specifically includes:

[0031] Clustering algorithms (such as K-means, DBSCAN, etc.) are used to perform clustering analysis on the semantic representation. The purpose of clustering analysis is to aggregate semantic units into multiple cognitive operation units based on semantic similarity. The cognitive operation units are independent modules in the design task, including defining functions, refining modules, selecting constraints, and confirming processes, which serve as the basic cognitive behaviors of the design task.

[0032] Each cognitive operation unit is assigned a unique node identifier, which serves as a graph node in the dynamic cognitive orchestration graph. This node identifier is used to distinguish different cognitive operation units during the subsequent graph construction process and serves as a unique identifier for the graph node.

[0033] By calculating the semantic correlation between each cognitive operation unit, graph edges are constructed. The semantic correlation is obtained by extracting the embedding vector of each cognitive operation unit and inputting it into a relation scoring network based on a multi-head self-attention mechanism. The relation scoring network generates attention scores for any two cognitive operation units to quantify the dependence strength in the semantic context. The obtained attention scores are normalized and then added as edge weights to the dynamic cognitive orchestration graph to reflect the causal or sequential dependency relationship between nodes. The semantic correlation represents the degree of mutual dependence between two cognitive operation units in the task. The higher the semantic correlation, the stronger the semantic correlation between the two operation units, which may require close collaboration during task decomposition.

[0034] The formula for calculating semantic relevance is:

[0035]

[0036] in, Indicates the first The first cognitive operational unit and the first The semantic correlation between cognitive operational units It is a cognitive operating unit and cognitive operating units Similarity between them Represents a node The set of adjacent nodes, that is, the set of nodes with which the node is located. Other directly related cognitive operational units, For all nodes Adjacent cognitive operation units and nodes The similarity is summed.

[0037] Based on the calculated semantic relevance, directed edges are constructed between graph nodes to form a preliminary graph. The direction of the directed edges represents the order and dependencies of task operations, while the weight of the edges is determined by the semantic relevance. In the preliminary graph, each node represents a cognitive operation unit, and the edges represent the semantic dependencies between operation units. The preliminary graph is then optimized by merging redundant nodes and removing isolated nodes to generate a dynamic cognitive orchestration graph composed of optimized graph nodes, graph edges, and semantic relevance. The dynamic cognitive orchestration graph can effectively represent the various cognitive operation units in the design task and the relationships between them.

[0038] The updated dynamic cognitive orchestration map specifically includes:

[0039] Extract the representation of each node in the dynamic cognitive orchestration graph, that is, the semantic embedding of each cognitive operation unit. At the same time, extract the hierarchical information of each node, that is, the relative position and hierarchical structure of each node in the graph.

[0040] The extracted node representations undergo dual-channel processing. Specifically, a dual-channel neural network structure is used to separate and process the node representations, generating high-level semantic representations and low-level structural representations. The high-level semantic representation is used to extract the core functions and target information of each node. High-level semantic networks (such as Transformer or BERT) extract high-order abstract information about nodes in the task. The low-level structural representation extracts low-level structural information such as specific operation steps and functional implementations of nodes. Low-level feature learning modules (such as Convolutional Neural Networks (CNNs) or LSTMs) acquire the specific execution structure and operation patterns of nodes. A cross-layer attention matrix is ​​constructed by calculating the correlation between the high-level semantic representation and the low-level structural representation. This matrix represents the relationship between high-level semantics and low-level structure and their importance in task decomposition. The correlation is typically calculated using methods such as dot product or cosine similarity to measure the correlation between high-level semantic information and low-level structural information, reflecting which structural information is most important for understanding and executing the design goals.

[0041] Based on the cross-layer attention matrix, a weighted aggregation of high-level semantic representation and low-level structural representation is performed. Specifically, the two representations are weighted and averaged according to the weights in the attention matrix to generate an updated representation for each node. Based on the updated node representation, a multi-granularity subtask candidate set is generated by further analyzing the hierarchical relationships and operational dependencies in the task. The multi-granularity includes functional, module, and operational levels. The granularity of the subtask candidate set can be dynamically adjusted according to the complexity of the design goals to ensure that the task can be decomposed into operational units of appropriate size. The generated subtask candidate set is then fed back into the original dynamic cognitive orchestration graph. The subtask candidate set is fed back into the dynamic cognitive orchestration graph. Based on the granularity labels and semantic associations in the subtask candidate set, corresponding cognitive operation unit nodes are added to the original dynamic cognitive orchestration graph. Based on the semantic dependencies calculated by the cross-layer attention mechanism, graph edges are established between the old and new nodes, completing the expansion of the node set and edge set, and realizing the structural update of the dynamic cognitive orchestration graph, so that the updated dynamic cognitive orchestration graph contains a multi-level subtask candidate structure.

[0042] The formula for generating the candidate set of subtasks is:

[0043]

[0044] in, Indicates the first The candidate set of subtasks for each task is determined by the semantic representation of the task. All candidate subtasks selected after evaluating the constraints. Indicates the first Each subtask of the task These are weighting coefficients, representing the importance of different subtask candidates. express Subtasks and task nodes semantic representation and constraints The matching degree function between them.

[0045] S3, based on the updated dynamic cognitive orchestration map and the candidate set of subtasks, plans the optimal subtask decomposition path sequence through a sequence decision model;

[0046] In this embodiment, the step of planning the optimal subtask decomposition path sequence based on the updated dynamic cognitive orchestration graph and subtask candidate set using a sequence decision model specifically involves:

[0047] Specifically, the joint state representation is obtained by encoding the current dynamic cognitive orchestration graph and the candidate set of subtasks, as follows:

[0048] A graph neural network is used to learn node-level and graph-level representations of the current dynamic cognitive orchestration graph, generating a graph structure vector that represents the current graph topology. Each subtask in the subtask candidate set is semantically encoded and granular labels, priorities, and constraints are fused to form a task embedding vector set. This set is then combined with the graph structure vector through feature concatenation to form a joint state representation.

[0049] The joint state representation is input into the sequential decision model, which includes a state encoding layer, a policy generation layer, and an action sampling layer. The state encoding layer provides a high-dimensional representation of the joint state, the policy generation layer generates the selection probabilities of each candidate subtask based on this representation, and the action sampling layer outputs the action candidate set for the current time step according to the probability distribution.

[0050] Based on the action candidate set and the output probability distribution of the policy network, the current optimal dismantling node is determined as the optimal action in this iteration. The selection process considers not only the independent probability of a single node but also introduces path context constraints to ensure that the selected node maintains logical consistency with existing dismantling paths.

[0051] The selected nodes are added to the current decomposition path to form a partial subtask sequence. At the same time, the joint state representation is updated based on the context dependency information of the newly added nodes, including updating the optionality flags of the remaining subtask set, adjusting the activity parameters of the graph nodes, and correcting the edge weights to reflect the latest task progress.

[0052] The process of updating the state and selecting actions is repeated round by round, expanding the decomposition path until the candidate set of subtasks is empty or the preset decomposition depth limit is reached. During this process, the sequence decision model ensures that the generated subtask paths meet the design requirements in terms of semantic consistency, logical rationality, and structural integrity through continuous state updates and action selection. Finally, a complete sequence of optimal subtask decomposition paths is output, serving as the basis for subsequently generating a structured subtask tree.

[0053] The optimal subtask decomposition path sequence planning formula is:

[0054]

[0055] in, Indicates at time status Next, execute the action. The value function represents the expected reward of choosing this action in the current state. Indicates at time Execute action The instant reward obtained afterward The discount factor controls the degree of influence on future rewards. The closer the value is to 1, the higher the importance of future rewards. Indicates the next state In the middle, select all possible actions. The maximum value reflects the expected reward of the optimal action in the future state. The learning rate represents the speed at which the current value is updated, controlling the trade-off between new information and historical values. To update the symbols.

[0056] In this embodiment, the sequence decision model is an improved meta-learning policy network, specifically:

[0057] A path memory guidance mechanism and a reversible compressed mapping structure are introduced into the meta-learning strategy network.

[0058] During the training and inference processes of different design tasks, the path memory guidance mechanism can retain the structural patterns of the optimal decomposition sequence in historical tasks and use them as a knowledge memory. When encountering subgraphs with similar semantic structures or consistent local topology, the system can retrieve the corresponding decomposition patterns from the memory to quickly initialize the decomposition process of similar subgraphs, thereby significantly shortening the convergence time of the search path and improving the planning accuracy.

[0059] The reversible compression mapping structure is used to map high-dimensional state representations in a dynamic cognitive orchestration graph to a low-dimensional latent space, while a reversible decoding mechanism ensures lossless information recovery during the mapping process. With this structure, the meta-learning policy network can inductively learn graph states in a unified latent space, effectively reducing the complexity of state representations and enhancing cross-task generalization capabilities.

[0060] This invention enables the system to perform knowledge transfer between similar tasks through an improved meta-learning policy network, thereby enhancing the model's generalization ability and applicability to various design tasks.

[0061] S4 outputs a structured subtask tree based on the optimal subtask decomposition path sequence.

[0062] In this embodiment, the step of outputting a structured subtask tree based on the optimal subtask decomposition path sequence specifically includes:

[0063] The optimal subtask decomposition path sequence is structurally perturbed to generate a counterfactual decomposition path sequence.

[0064] The optimal subtask decomposition path sequence and the counterfactual decomposition path sequence are input into the embedding encoder for comparison to obtain their respective embedding representations.

[0065] Based on self-embedding representation, the final structured subtask tree is output.

[0066] Specifically, the optimal subtask decomposition path sequence is structurally perturbed to generate a counterfactual decomposition path sequence, as follows:

[0067] The sequence information of each cognitive operation unit in the optimal subtask decomposition path sequence is obtained, and its corresponding semantic embedding representation is extracted. The sequence information is used to maintain the logical dependencies of the path, and the semantic embedding representation serves as the semantic reference benchmark for subsequent path perturbation. Structural perturbation operations are performed on the optimal subtask decomposition path sequence, including operation node replacement, sequence exchange, edge relationship adjustment, and node insertion and deletion. To ensure the generated counterfactual path still possesses rationality, semantic relevance constraints are applied to each structural perturbation operation. Specifically, the semantic similarity between the perturbation path and the optimal path in the embedding space is calculated, and the similarity is compared with a preset threshold. Only candidate paths with a relevance higher than the threshold are retained. This avoids invalid paths with excessive semantic deviation from entering the subsequent training process, thereby improving the effectiveness of the comparison samples. For each counterfactual path that meets the constraints, not only is its overall sequence recorded, but the corresponding perturbation type and the position of the modified cognitive operation unit are also clearly labeled. All the selected path sets are summarized to form the counterfactual decomposition path sequence.

[0068] Specifically, the optimal subtask decomposition path sequence and the counterfactual decomposition path sequence are input into the encoder for comparison and embedding to obtain their respective embedding representations:

[0069] To generate a structured subtask tree that accurately reflects the hierarchical logic of the tasks, the optimal subtask decomposition path sequence obtained earlier is first used as the positive sample input, while the counterfactual decomposition path sequence generated through structural perturbation and semantic constraints is used as the contrast sample input. Both are fed into a contrastive embedding encoder, which includes an embedding network and a contrastive optimization module. The embedding network encodes the cognitive operation units in the path sequence. During encoding, not only the semantic embedding features of the units are considered, but also the sequential information in the path and the dependencies between units are combined. Through recursion or attention mechanisms, a complete path-level semantic embedding representation is obtained. This representation can simultaneously reflect the semantic content and structural logic of the path. After obtaining the path-level semantic embedding representations of the positive and contrast samples, the contrastive optimization module optimizes the embedding space through a contrastive learning strategy. Specifically, it maximizes the embedding distance between the optimal subtask decomposition path sequence and the counterfactual decomposition path sequence, ensuring that positive samples remain tightly aggregated in the embedding space, while contrast samples are effectively distinguished. The specific formula is as follows:

[0070]

[0071] in, This represents the contrastive loss function, which measures the distance between positive and counterfactual samples in the embedding space. This represents the total number of samples. Indicates the first The representation of each subtask path (optimal path) in the embedding space. Indicates the first The representation of a counterfactual decomposition path in the embedding space It is the square of the Euclidean distance, used to measure the similarity between positive and contrast samples.

[0072] This optimization process can enhance the discriminativeness and robustness of the embedded representation, thereby ensuring that the generated subtask tree has high semantic consistency and structural rationality.

[0073] After embedding optimization, the system performs multi-dimensional path scoring and hierarchical construction operations based on the embedded representation of the optimal subtask decomposition path sequence. The multi-dimensional scoring criteria include, but are not limited to, semantic consistency, logical coherence, structural integrity, and executability constraints. Based on the comprehensive scoring results, the system organizes the optimal path sequence into a hierarchical tree structure and outputs the final structured subtask tree.

[0074] Example 2:

[0075] To verify the feasibility of this invention in practice, it was applied to a system design project for a new generation of health tracking wristbands undertaken by a large smart wearable device manufacturer in the second half of 2024. In this project, the company proposed a system-level design task requiring the decomposition of multiple module functions, covering various technical sub-directions such as hardware layout, interaction logic, health monitoring functions, power consumption control, security encryption, and system integration. This design task was presented in natural language and jointly written by multiple product managers and system engineers, exceeding 3200 words and including complex constraints, cross-domain intents, and functional intersections. Traditional methods of module division by human experts based on experience are not only inefficient but also frequently suffer from problems such as task omissions, logical conflicts, and inconsistent granularity of division as system complexity increases.

[0076] In practical application, the design task text is first input into the system of this invention. The system performs sentence segmentation and parsing of the task text based on a deep pre-trained model, extracting a total of 412 keywords related to the design function, 132 subject / predicate phrases related to the design objective, and 87 expression structures related to the constraints, thus constructing a semantic representation with three semantic channels. Subsequently, this semantic representation is sent to the cognitive orchestration graph construction module. The system automatically identifies 72 core cognitive operation units, including defining a heart rate detection module, setting power consumption trigger conditions, and determining communication protocol specifications. In the semantic dependency mining stage, a total of 162 graph edges are generated, and the structure of 34 paths is adjusted according to the dependency strength, forming an initial cognitive orchestration graph.

[0077] After the cognitive orchestration graph is constructed, the system enters the cross-layer representation and reasoning stage. Using a dual-channel representation method, high-level semantic representations (such as design goals like "all-weather health monitoring" and "end-to-end data encryption") and low-level structural representations (such as specific tasks like "configuring encryption chip interfaces" and "optimizing power supply paths") are extracted. The system further calculates the correlation between the two types of representations, constructs a cross-layer attention matrix, and uses this matrix to weight and aggregate the representations of each node, ultimately generating a multi-granularity sub-task candidate set of 89 items. This set is further refined and re-encoded during the feedback process to the cognitive graph. The system ultimately expands the graph nodes to 143, adds 75 graph edges, and increases the graph depth from the original 4 layers to 7 layers, demonstrating a stronger ability to express the hierarchical structure of tasks.

[0078] During the path planning phase, the system activated an improved meta-learning strategy network. After introducing a path memory guidance mechanism, the system retrieved 234 corresponding path patterns from 11 historical wearable product design projects, establishing a structural memory embedding space. Under the reversible compression mapping mechanism, the current task graph was reduced to a 128-dimensional latent space for planning operations. After 19 iterations, the optimal decomposition path generated by this strategy network contained 27 key subtasks, reducing the average subtask planning time to 0.96 seconds. Compared to the traditional 7.5 minutes required for manual structure graph-assisted decomposition, this represents a time efficiency improvement of over 460%.

[0079] In the counterfactual path sequence generation stage, the system uses structural perturbation operations to construct approximately 107 counterfactual path samples. The perturbation strategies cover node replacement, order swapping, edge relationship adjustment, and node insertion and deletion. During system training, 18 semantically broken paths are automatically removed, ultimately resulting in 89 high-quality negative samples. Through positive and negative path comparison learning, the model's path embedding discrimination accuracy improved from 81.2% in the initial training stage to 92.7%, demonstrating stronger path discrimination and anomaly detection capabilities.

[0080] After the task finally outputs a structured subtask tree, it is manually verified by three system architects and compared with the results of manual partitioning by experts. The system partitioning results achieve 98.4% in task integrity and a logical consistency score of 4.82 (out of 5). On average, each task node provides 2.6 structural reasoning bases or counterfactual elimination explanations, which is significantly better than the performance of existing general task partitioning models, which average only 0.8 explainable path bases.

[0081] As demonstrated in this embodiment, the present invention not only improves the efficiency and accuracy of design task decomposition, but also exhibits excellent performance in structural visualization, path interpretability, and processing of highly complex task graphs. In actual product development cycles, the system saves at least 12 person-days of task decomposition work, and provides stable, consistent, and traceable design decision support through multiple rounds of iterative modifications, fully verifying its practicality, versatility, and technological advancement in the task decomposition of complex systems.

[0082] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0083] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0084] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0085] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0086] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0087] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A deep learning-based intelligent decomposition method for design tasks, characterized in that, Includes the following steps: Semantic analysis is performed on the task text data to obtain semantic representations. Specifically, the task text data is parsed to obtain semantic units; the semantic units are processed using a pre-trained language model to obtain word-level contextual representations; and target intent features and design constraint features are extracted from the contextual representations or text data. By integrating target intent features and design constraint features, a unified representation vector is formed, and the unified representation vector is mapped to a semantic representation. The dynamic cognitive orchestration graph is constructed based on semantic representation. Specifically, the semantic representation is clustered to form multiple cognitive operation units, and each unit is assigned a node identifier as a graph node. Calculate the semantic association degree between cognitive operation units; Based on the semantic correlation, connect the corresponding graph nodes to generate a preliminary graph; The preliminary map is optimized to generate a dynamic cognitive arrangement map; Generate a multi-granularity set of subtask candidates based on the dynamic cognitive orchestration graph, and update the dynamic cognitive orchestration graph; Based on the updated dynamic cognitive orchestration graph and the candidate set of subtasks, the optimal subtask decomposition path sequence is planned through a sequence decision model. Output a structured subtask tree based on the optimal subtask decomposition path sequence.

2. The intelligent decomposition method for design tasks based on deep learning according to claim 1, characterized in that, The updated dynamic cognitive orchestration map specifically includes: Extract node representations and their hierarchical information from the dynamic cognitive orchestration graph; Based on node representation, high-level semantic representation and low-level structural representation are generated; The high-level semantic representation and the low-level structural representation are fused, and the representation of each node is updated. The fusion includes calculating the correlation degree between the high-level semantic representation and the low-level structural representation, and performing weighted aggregation based on the correlation degree. Based on the updated node representation, a multi-granularity set of subtask candidates is generated; Update the dynamic cognitive orchestration map using the candidate set of subtasks.

3. The intelligent decomposition method for design tasks based on deep learning according to claim 2, characterized in that, Based on the updated dynamic cognitive orchestration map and the candidate set of subtasks, the optimal subtask decomposition path sequence is planned through a sequence decision model, specifically as follows: Encode the current dynamic cognitive orchestration graph and the candidate set of subtasks to obtain a joint state representation; Inputting the joint state representation into the sequence decision model yields the action probability distribution; Nodes are selected from the subtask candidate set based on the action probability distribution, and the selected nodes are added to the decomposition path; Update the current joint state representation and iteratively execute the above selection and update steps until the termination condition is met, and output the optimal subtask decomposition path sequence.

4. The intelligent decomposition method for design tasks based on deep learning according to claim 3, characterized in that, The sequence decision model is an improved meta-learning policy network, which specifically includes: A path memory guidance mechanism and a reversible compressed mapping structure are introduced into the meta-learning strategy network; The path memory guidance mechanism preserves the structural pattern of the optimal decomposition sequence in historical tasks and uses it for the decomposition initialization of similar subgraphs. By using a reversible compression mapping structure, the dynamic cognitive graph is mapped to a low-dimensional latent space and reversibly decoded, enabling the meta-learning policy network to inductively learn the mapping state of the dynamic cognitive graph in a unified latent space.

5. The intelligent decomposition method for design tasks based on deep learning according to claim 4, characterized in that, The process of encoding the current dynamic cognitive orchestration graph and the subtask candidate set to obtain a joint state representation is as follows: Topological structure learning is performed on the dynamic cognitive arrangement graph using a graph neural network to generate graph structure vectors; Multi-feature fusion encoding is performed on the candidate set of subtasks to generate a set of task embedding vectors; The graph structure vector and the task embedding vector set are concatenated to form a joint state representation.

6. The intelligent decomposition method for design tasks based on deep learning according to claim 5, characterized in that, The structured subtask tree output based on the optimal subtask decomposition path sequence is specifically as follows: The optimal subtask decomposition path sequence is structurally perturbed to generate a counterfactual decomposition path sequence. The optimal subtask decomposition path sequence and the counterfactual decomposition path sequence are input into the embedding encoder for comparison to obtain their respective embedding representations. Based on self-embedding representation, the final structured subtask tree is output.

7. The intelligent decomposition method for design tasks based on deep learning according to claim 6, characterized in that, The specific steps for generating the counterfactual deconstruction path sequence are as follows: Obtain the order and semantic embedding representation of each cognitive operation unit in the subtask decomposition path sequence; Perform a structural perturbation operation on the path sequence to generate multiple alternative path sequences; Semantic relevance constraints are applied to each generated candidate path sequence, and sequences with semantic relevance to the original path sequence that are higher than a preset threshold are selected and recorded as counterfactual decomposition path sequences.

8. The intelligent decomposition method for design tasks based on deep learning according to claim 7, characterized in that, The output structured subtask tree includes: The optimal subtask decomposition path sequence and the counterfactual decomposition path sequence are used as positive samples and contrast samples, respectively, and input into the contrast embedding encoder. By comparing the embedding network in the embedding encoder, the sequences of cognitive operation units in each path are encoded to generate a path-level semantic embedding representation; The embedding network is optimized to maximize the embedding distance between positive and negative samples; Based on the optimized path semantic embedding representation, a structured subtask tree is generated according to the multidimensional scoring criteria.

Images