Design task intelligent decomposition method based on deep learning

Abstract

The invention discloses an intelligent design task decomposition method based on deep learning, and relates to the technical field of intelligent design, and the method comprises the steps: carrying out the semantic analysis of task text data, and obtaining a semantic representation; constructing a dynamic cognitive orchestration map based on semantic representation, generating a multi-granularity sub-task candidate set based on the dynamic cognitive orchestration map, and updating the dynamic cognitive orchestration map; based on the updated dynamic cognitive layout map and the sub-task candidate set, planning an optimal sub-task disassembling path sequence through a sequence decision model; and outputting a structured sub-task tree based on the optimal sub-task disassembly path sequence. According to the method, high-precision and high-efficiency task decomposition can be realized in a complex task scene, the defects of a traditional method in the aspects of granularity inconsistency, logic conflicts, task omission and the like are overcome, and the method has relatively high practicability and popularization value.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent design, more particularly, the present application relates to a design task intelligent decomposition method based on deep learning. BACKGROUND

[0002] At present, the intelligent decomposition method of design tasks mainly depends on the combination of natural language processing (NLP) and rule-based reasoning to realize the structured understanding and automatic decomposition of design task texts. The design tasks in the fields of engineering design, software development, product architecture, etc. usually have high abstraction, strong semantic dependency and complex context association. Traditional methods mainly use keyword extraction, template matching, expert rule base reasoning or shallow semantic model-based logical induction to perform task segmentation. However, these methods have obvious limitations in dealing with complex, multi-level and strongly dependent design tasks. Especially in the modeling of internal cognitive structure of tasks, traditional methods cannot effectively capture the multi-granularity relationship and internal logic between tasks, resulting in the inability to meet the needs of multi-stage path planning and structured subtask generation for complex design goals.

[0003] Some studies introduce deep learning techniques based on recurrent neural networks, graph neural networks or BERT pre-trained models to extract task semantic features and local context information, and then assist in task decomposition. However, existing methods fail to fully consider the semantic dependency and structured relationship 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. The path planning in the prior art is usually based on static task states, lacking dynamic perception ability for design goal evolution, task state transition and structural dependency relationship, resulting in difficulty in adapting the path generation strategy to changing needs in complex design tasks. SUMMARY

[0004] In order to overcome the above-mentioned defects of the prior art, the embodiments of the present application provide a design task intelligent decomposition method based on deep learning, which solves the problems of insufficient internal cognitive structure modeling of tasks and lack of dynamic perception in path planning in the prior art, and realizes more efficient and flexible design task decomposition.

[0005] To achieve the above-mentioned purpose, the present application provides the following technical solutions: A design task intelligent decomposition method based on deep learning, comprising the following steps: performing semantic analysis on task text data to obtain semantic representation; constructing a dynamic cognitive arrangement graph based on the semantic representation, generating a multi-granularity subtask candidate set based on the dynamic cognitive arrangement graph, and updating the dynamic cognitive arrangement graph; based on the updated dynamic cognitive arrangement graph and the subtask candidate set, planning an optimal subtask decomposition path sequence through a sequence decision model; and outputting a structured subtask tree based on the optimal subtask decomposition path sequence.

[0006] In a preferred embodiment, the semantic analysis of the task text data to obtain a semantic representation is specifically: parsing the design task text data to obtain semantic units; processing the semantic units using a pre-trained language model to obtain a word-level context representation; extracting target intent features and design constraint features from the context representation or the text data; fusing the target intent features and the design constraint features to form a unified representation vector, and mapping the unified representation vector to the semantic representation.

[0007] In a preferred embodiment, the construction of the dynamic cognitive arrangement graph based on the semantic representation is specifically: clustering analysis of the semantic representation to form a plurality of cognitive operation units, and assigning a node identifier to each unit as a graph node; calculating the semantic correlation degree between the cognitive operation units; connecting the corresponding graph nodes based on the semantic correlation degree to generate a preliminary graph; optimizing the preliminary graph to generate the dynamic cognitive arrangement graph.

[0008] In a preferred embodiment, the updating of the dynamic cognitive arrangement graph is specifically: extracting node representations and hierarchical information in the dynamic cognitive arrangement graph; generating high-level semantic representations and low-level structural representations based on the node representations; fusing the high-level semantic representations and the low-level structural representations, and updating the node representations, the fusion including calculating the correlation degree between the high-level semantic representations and the 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; updating the dynamic cognitive arrangement graph using the subtask candidate set.

[0009] In a preferred embodiment, the planning of an optimal subtask disassembly path sequence based on the updated dynamic cognitive arrangement graph and the subtask candidate set through a sequence decision model is specifically: encoding the current dynamic cognitive arrangement graph and the 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 a node from the subtask candidate set according to the action probability distribution, and adding the selected node to the disassembly path; updating the current joint state representation, and iteratively performing the above selection and updating steps until a termination condition is met, and outputting the optimal subtask disassembly path sequence.

[0010] In a preferred embodiment, the sequence decision model is an improved meta-learning strategy network, and the meta-learning strategy network is specifically: introducing a path memory guidance mechanism and a reversible compression mapping structure into the meta-learning strategy network; retaining the structural mode of the optimal disassembly sequence in the historical tasks through the path memory guidance mechanism, and using it for disassembly initialization of similar subgraphs; mapping the dynamic cognitive graph to a low-dimensional latent space through the reversible compression mapping structure and realizing reversible decoding, so that the meta-learning strategy network inductively learns the mapping state of the dynamic cognitive graph in the unified latent space.

[0011] In a preferred embodiment, the current dynamic cognitive arrangement graph and the subtask candidate set are encoded to obtain a joint state representation, specifically: the graph structure vector is generated by topological structure learning of the dynamic cognitive arrangement graph through the graph neural network; the task embedding vector set is generated by multi-feature fusion coding of the subtask candidate set; the graph structure vector and the task embedding vector set are spliced to form the joint state representation.

[0012] In a preferred embodiment, the structured subtask tree is output based on the optimal subtask disassembly path sequence, specifically: the counterfactual disassembly path sequence is generated by structure disturbance of the optimal subtask disassembly path sequence; the respective embedding representations are obtained by inputting the optimal subtask disassembly path sequence and the counterfactual disassembly path sequence into the comparative embedding encoder; and the final structured subtask tree is output based on the self-embedding representation.

[0013] In a preferred embodiment, the counterfactual disassembly path sequence is specifically generated by: obtaining the order of each cognitive operation unit in the subtask disassembly path sequence and its semantic embedding representation; performing a structure disturbance operation on the path sequence to generate multiple candidate path sequences; applying a semantic correlation constraint to each generated candidate path sequence to filter out sequences with a semantic correlation higher than a preset threshold with the original path sequence, and recording them as counterfactual disassembly path sequences.

[0014] In a preferred embodiment, the output structured subtask tree includes: inputting the optimal subtask disassembly path sequence and the counterfactual disassembly path sequence as positive samples and comparative samples, respectively, into the comparative embedding encoder; encoding the cognitive operation unit sequence in each path through the embedding network in the comparative embedding encoder to generate semantic embedding representations at the path level; optimizing the embedding network to maximize the embedding distance between the positive samples and the comparative samples; and generating a structured subtask tree according to a multi-dimensional scoring criterion based on the optimal path semantic embedding representation obtained after optimization.

[0015] The application can comprehensively capture the complex semantic dependency relationship and hierarchical structure between each subtask in the design task by constructing a dynamic cognitive arrangement graph based on deep learning. Through the generation of a multi-granularity subtask candidate set and the dynamic updating of the graph, the task decomposition strategy can be flexibly adjusted for different design tasks. Compared with traditional methods, especially when dealing with complex tasks with multiple levels, multiple dimensions and strong dependency relationships, the application significantly improves the accuracy and efficiency of task decomposition and effectively solves the limitations of traditional methods in complex design tasks. At the same time, the application introduces a sequence decision model to dynamically plan the task decomposition path, which can optimize the decomposition path in real time according to the evolution of the task target, the change of the design state and the adjustment of the structural dependency relationship. Compared with the static decision of traditional path planning methods, the application can dynamically perceive and adapt to the changes of the design target and the task state, solve the static limitations in the path planning of traditional methods, and improve the adaptability and response ability of the system to complex design tasks. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 FIG. 1 is a flowchart of a design task intelligent decomposition method based on deep learning according to the present application.

[0017] Figure 2 FIG. 2 is a flowchart of generating an optimal subtask decomposition path sequence in a design task intelligent decomposition method based on deep learning according to the present application. DETAILED DESCRIPTION

[0018] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0019] Embodiment 1, Figure 1 A design task intelligent decomposition method based on deep learning is provided, which comprises: S1, performing semantic analysis on the task text data to obtain a semantic representation; In this embodiment, the semantic analysis on the task text data to obtain a semantic representation is specifically: The design task text data is received and sentence segmentation processing is performed on the task text. The existing text segmentation tool is used to divide the design task text into multiple sentences, and it is ensured that each sentence has a clear grammatical structure. The grammatical structure of the segmented text is analyzed, and the dependency syntax analysis is used to analyze the grammatical relationship between the words in the text, identify the noun phrases, verb phrases, and modification relationships, and construct basic semantic units. These semantic units can be the target, constraints, and operations in the task, which are the basic components and are used as the basic units for subsequent analysis.

[0020] The semantic units are input into a pre-trained language model, and the context representation of each unit is obtained using the word-level context representation capability of the pre-trained language model. The pre-trained language model has been trained on a large corpus and can capture complex context relationships between words.

[0021] The obtained context representation or text data is classified. The word-level vector output by the pre-trained model is used to extract features from the context representation using a convolutional neural network (CNN) or a long short-term memory network (LSTM). Based on this, the classifier (such as a fully connected layer plus a Softmax layer) extracts semantic segments related to the design task from the context representation and marks them as target intent features. These features include the core target of the design task, such as functional requirements and performance requirements. Through feature marking, these target intent features are separated from the semantic units, providing clear target guidance for subsequent task decomposition.

[0022] The design task text data is continuously analyzed to extract the restrictive conditions therein, such as technical limitations in the design, legal compliance requirements, and resource constraints. These restrictive conditions will affect the direction and feasibility of task decomposition. For each restrictive condition in the design task, information extraction (IE) technology and pattern recognition algorithms, such as named entity recognition (NER) and relationship extraction, are used to automatically identify and mark design constraint features.

[0023] The target intent features and design constraint features extracted are jointly modeled using a multi-channel fusion network (e.g., a dual-channel neural network or a multi-modal network). In this step, the target intent features and design constraint features are processed using different network channels. During processing, the target intent feature channel focuses on modeling the task target, and the design constraint feature channel focuses on modeling the limiting condition. Convolutional layers, pooling layers, and fully connected layers are used in each channel to extract key feature information. Through a fusion strategy (such as feature concatenation, weighted averaging, etc.), the target intent features and design constraint features are integrated into the same unified embedding space to generate a unified representation vector that contains not only the semantic information of the task target but also the feature information of the task constraints.

[0024] 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.

[0025] 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; In this embodiment, the construction of a dynamic cognitive orchestration graph based on semantic representation specifically includes: 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.

[0026] 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.

[0027] By calculating the semantic correlation between each cognitive operation unit, a graph edge is 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 an attention score for any two cognitive operation units to quantify the dependence strength in the semantic context. The obtained attention score is normalized and then added as an edge weight 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.

[0028] The formula for calculating semantic relevance is:

[0029] in, Indicates the first The first cognitive operational unit and the first 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 are summed up.

[0030] Based on the calculated semantic correlation degree, a directed edge between the nodes in the graph is constructed, forming a preliminary graph. The direction of the directed edge represents the order and dependency of the task operations, and the weight of the edge is determined by the semantic correlation degree. In the preliminary graph, each node represents a cognitive operation unit, and the edge represents the semantic dependency relationship between the operation units. The preliminary graph is optimized, including merging redundant nodes and removing isolated nodes, to generate a dynamic cognitive arrangement graph composed of optimized graph nodes, graph edges, and semantic correlation degrees. The dynamic cognitive arrangement graph can effectively represent each cognitive operation unit and the relationship between them in the design task.

[0031] The dynamic cognitive arrangement graph is updated, specifically: The node representation in the dynamic cognitive arrangement graph is extracted, i.e., the semantic embedding of each cognitive operation unit. At the same time, the hierarchical information of each node is extracted, i.e., the relative position and hierarchical structure of each node in the graph.

[0032] The extracted node representation is processed in a double-channel manner. Specifically, a double-channel neural network structure is used to separate and process the node representation, generating high-level semantic representation and low-level structural representation. High-level semantic representation: used to extract the core function and target information of each node. Through a high-level semantic network (such as Transformer or BERT), the high-order abstract information of the node in the task is extracted. Low-level structural representation: used to extract the specific operation steps, function implementation, and other low-level structural information of the node. Through a low-level feature learning module (such as a convolutional neural network CNN or LSTM), the specific execution structure and operation mode of the node are obtained. By calculating the correlation degree between the high-level semantic representation and the low-level structural representation, a cross-layer attention matrix is constructed. This matrix is used to represent the mutual relationship between high-level semantics and low-level structures and their importance in task decomposition. The correlation degree calculation usually uses 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 goal.

[0033] 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.

[0034] The formula for generating the candidate set of subtasks is:

[0035] 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 The various subtasks 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.

[0036] 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; 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: Specifically, the joint state representation is obtained by encoding the current dynamic cognitive orchestration graph and the candidate set of subtasks, as follows: The current dynamic cognitive arrangement graph is represented at node level and graph level by using a graph neural network to generate a graph structure vector representing the topology of the current graph, and each subtask in the subtask candidate set is semantically encoded and fused with granularity labels, priorities and constraint conditions to form a task embedding vector set, which is combined with the graph structure vector to form a joint state representation by feature splicing.

[0037] The joint state representation is input into a sequence decision model, which includes a state encoding layer, a policy generation layer and an action sampling layer. The state encoding layer represents the joint state in a high-dimensional manner, the policy generation layer generates selection probabilities for each candidate subtask based on the joint state, and the action sampling layer outputs the action candidate set at the current time according to the probability distribution.

[0038] Based on the action candidate set and the output probability distribution of the policy network, the optimal disassembly node is determined as the optimal action in the current iteration. The selection process not only considers the independent probability of a single node, but also introduces path context constraints to ensure that the selected node is logically consistent with the existing disassembly path.

[0039] The selected node is added to the current disassembly path to form a partial subtask sequence, and the joint state representation is updated according to the context dependency information of the new node, including updating the selectability flag of the remaining subtask set, adjusting the activity parameter of the graph node and correcting the weight of the edge to reflect the latest task progress.

[0040] Repeat the state update and action selection to expand the disassembly path round by round until the subtask candidate set is empty or the preset disassembly depth limit is reached. In this process, the sequence decision model ensures that the generated subtask path meets the design requirements in terms of semantic consistency, logical rationality and structural integrity through continuous state updating and action selection. Finally, the complete optimal subtask disassembly path sequence is output as the basis for generating a structured subtask tree.

[0041] The optimal subtask disassembly path sequence planning formula is:

[0042] wherein, represents the value function of performing action at time in state , which represents the expected return of selecting the action in the current state, represents the immediate reward obtained after performing action at time , is a discount factor that controls the degree of influence of future rewards, The closer the value is to 1, the higher the importance of future rewards, represents the next state Among them, the maximum value of all possible actions reflects the expected return of the optimal action in the future state, is the learning rate, which represents the speed of current value update, controls the trade-off between new information and historical values, is the update symbol.

[0043] In this embodiment, the sequence decision model is an improved meta-learning strategy network, specifically: A path memory guidance mechanism and a reversible compression mapping structure are introduced into the meta-learning strategy network.

[0044] In the training and inference process of different design tasks, the path memory guidance mechanism can retain the structural pattern of the optimal disassembly sequence in the historical task and use it as a knowledge memory bank. When encountering subgraphs with similar semantic structures or consistent local topologies, the system can retrieve the corresponding disassembly pattern from the memory bank for quick initialization of the disassembly process of similar subgraphs, thereby significantly shortening the convergence time of the search path and improving the planning accuracy.

[0045] The reversible compression mapping structure is used to map the high-dimensional state representation in the dynamic cognitive arrangement atlas to a low-dimensional latent space, while ensuring information lossless recovery of the mapping process through a reversible decoding mechanism. With the help of this structure, the meta-learning strategy network can perform inductive learning on the atlas state in a unified latent space, effectively reducing the complexity of the state representation and enhancing the generalization ability across tasks.

[0046] The improved meta-learning strategy network enables the system to perform knowledge transfer between similar tasks, enhancing the model's generalization ability and applicability to multiple design tasks.

[0047] S4, output a structured subtask tree based on the optimal subtask disassembly path sequence.

[0048] In this embodiment, the structured subtask tree is output based on the optimal subtask disassembly path sequence, specifically: The optimal subtask disassembly path sequence is structurally disturbed to generate counterfactual disassembly path sequences; The optimal subtask disassembly path sequence and the counterfactual disassembly path sequence are input into a comparative embedding encoder to obtain their respective embedding representations; Based on the self-embedding representation, the final structured subtask tree is output.

[0049] Among them, the optimal subtask disassembly path sequence is structurally disturbed to generate counterfactual disassembly path sequences, specifically: obtain the sequence information of each cognitive operation unit in the optimal subtask disassembly path sequence, and extract the corresponding semantic embedding representation, the sequence information is used to maintain the logical dependency of the path, and the semantic embedding representation is used as a semantic reference benchmark for subsequent path disturbance, performing a structural disturbance operation on the optimal subtask disassembly path sequence, the disturbance including operation node replacement, sequence exchange, edge relationship adjustment, and node insertion and deletion, in order to ensure that the generated counterfactual path still has reasonableness, a semantic correlation constraint is applied to each structural disturbance operation. Specifically, the semantic similarity of the disturbed path and the optimal path in the embedding space is calculated, and the similarity is compared with a preset threshold, only the candidate paths with a correlation higher than the threshold are retained. In this way, invalid paths with too large semantic deviation can be avoided from entering the subsequent training process, thereby improving the effectiveness of the contrast samples. For each counterfactual path that meets the constraint condition, not only the overall sequence is recorded, but also the disturbance type and the position of the modified cognitive operation unit are labeled in detail. All filtered paths are collected to form a counterfactual disassembly path sequence.

[0050] wherein the optimal subtask disassembly path sequence and the counterfactual disassembly path sequence are input into a contrast embedding encoder to obtain respective embedding representations, specifically: In order to generate a structured subtask tree that can accurately reflect the task-level logic, the optimal subtask disassembly path sequence obtained as described above is input as a positive sample, and the counterfactual disassembly path sequence generated by structural disturbance and semantic constraint is input as a contrast sample, both of which are input into a contrast embedding encoder, the contrast embedding encoder including an embedding network and a contrast optimization module. The embedding network is used to encode the cognitive operation units in the path sequence. When encoding, not only the semantic embedding features of the units are considered, but also the order information in the path and the dependency relationship between the units are combined, and the semantic embedding representation of the complete path level is obtained through recursion or attention mechanism aggregation. The representation can reflect the semantic content and structural logic of the path at the same time. After obtaining the path-level semantic embedding representation of the positive and contrast samples, the contrast optimization module optimizes the embedding space through a contrast learning strategy. Specifically, the embedding distance between the optimal subtask disassembly path sequence and the counterfactual disassembly path sequence is maximized, so that the positive samples are closely aggregated in the embedding space, and the contrast samples are effectively distinguished. The specific formula is:

[0051] wherein, represents a contrast loss function, which measures the distance between the positive sample and the counterfactual sample in the embedding space, represents the total number of samples, represents the embedding representation of the th subtask path (optimal path) in the embedding space, represents the embedding representation of the th subtask path (counterfactual path) in the embedding space, represents the embedding representation of the th subtask path (counterfactual path) in the embedding space. representation of counterfactual disassembly paths in embedding space, is the square of Euclidean distance, used to measure the similarity between positive samples and contrast samples.

[0052] The optimization process can enhance the discriminability and robustness of the embedding representation, thereby ensuring that the generated sub-task tree has high semantic consistency and structural rationality.

[0053] After the embedding optimization is completed, the system performs multi-dimensional path scoring and hierarchical construction operations based on the embedding representation of the optimal sub-task disassembly 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 sub-task tree.

[0054] Embodiment 2: To verify the feasibility of the present application in implementation, the present application is applied to a new generation of health tracking bracelet product system design project carried out by a large intelligent wearable device manufacturing enterprise in the second half of 2024. In this project, the enterprise proposes a system-level design task, which requires the decomposition of multiple module functions, covering hardware layout, interaction logic, health monitoring function, power consumption control, security encryption, system integration, and other technical sub-directions. The design task is given in natural language form, jointly written by multiple product managers and system engineers, with more than 3200 words of content, including complex constraints, cross-domain intentions, and function intersection points. The traditional method of manually dividing modules by experts based on experience not only has low efficiency, but also often has problems such as task omission, logical conflict, and inconsistent division granularity when the system complexity increases.

[0055] In the actual application process, first, the design task text is input into the system of the present application, and the system performs sentence parsing on the task text based on a deep pre-training model, extracts a total of 412 key words related to design functions, 132 subject / predicate phrases related to design goal intentions, and 87 expression structures related to constraint conditions, and constructs a semantic representation with three types of semantic channels. Subsequently, the semantic representation is sent to the cognitive arrangement graph construction module, and the system automatically identifies 72 core cognitive operation units, including defining heart rate detection modules, setting power consumption trigger conditions, and determining communication protocol specifications, etc. In the semantic dependency relationship mining stage, a total of 162 graph edges are generated, and 34 path structures are adjusted according to the dependency strength, forming an initial cognitive arrangement graph.

[0056] After the cognitive mapping is constructed, the system enters the cross-layer representation and reasoning phase. Through the dual-channel representation method, high-level semantic representations (such as "all-weather health monitoring" and "data end-to-end encryption") and low-level structural representations (such as "configure encryption chip interface" and "optimize power supply path") are extracted respectively. The system further calculates the correlation between the two types of representations, constructs a cross-layer attention matrix, and aggregates the node representations based on the weight, finally generating a multi-granularity sub-task candidate set of 89 items. This set is further refined and re-encoded in the feedback to the cognitive mapping process, and the system finally expands the graph nodes to 143, adds 75 graph edges, and increases the graph depth from the original 4 layers to 7 layers, showing stronger task structure hierarchical expression ability.

[0057] In the path planning phase, the system activates an improved meta-learning strategy network. After introducing the path memory guidance mechanism, the system retrieves 234 corresponding path patterns from the history of 11 wearable product design projects, establishing a structural memory embedding space. Under the action of reversible compression mapping mechanism, the current task graph is reduced to 128-dimensional latent space for planning operation. The optimal disassembly path generated by the strategy network after 19 iterations contains 27 key sub-tasks, and the average sub-task planning time is shortened to 0.96 seconds, compared with the average 7.5 minutes required by the traditional handcrafted structure graph assisted disassembly process, the time efficiency is improved by more than 460%.

[0058] In the counterfactual disassembly path sequence generation link, the system uses structural perturbation operations to construct about 107 counterfactual disassembly path samples, with perturbation strategies covering node replacement, order exchange, edge relationship adjustment, and node insertion and deletion. During the system training process, 18 semantic broken paths are automatically removed, and finally 89 high-quality negative samples are formed. Through positive and negative path comparison learning, the model's path embedding discrimination accuracy improves from 81.2% in the initial training stage to 92.7%, showing stronger path discrimination and anomaly recognition ability.

[0059] After the structured sub-task tree is output, it is verified by 3 system architects and compared with the expert manual division results. The system's division results achieve 98.4% in task integrity, with a logic consistency score of 4.82 (full score 5), providing an average of 2.6 structural reasoning bases or counterfactual elimination explanations for each task node, significantly better than the existing general task division model which only has an average of 0.8 explainable path basis.

[0060] It can be seen from the embodiment that the application not only improves the design task decomposition efficiency and precision, but also shows excellent performance in structure visualization, path interpretability and high complex task graph processing. In the actual product development cycle, the system saves at least 12 person-days of task decomposition workload, provides stable, consistent and traceable design decision support in multiple rounds of iterative modification, and fully verifies the practicability, universality and technical advancement of the system in complex system task decomposition.

[0061] The above formulas are dimensionless values calculated, the formulas are obtained by collecting a large amount of data to simulate a formula of the most recent real situation, and the preset parameters in the formula are set by a person skilled in the art according to the actual situation.

[0062] The above embodiments can be realized wholly or partially by software, hardware, firmware or any combination thereof. When realized by software, the above embodiments can be realized wholly or partially in the form of a computer program product.

[0063] Those skilled in the art can realize that the modules and algorithm steps of the examples described in combination with the embodiments disclosed herein can be realized by electronic hardware or a combination of computer software and electronic hardware. Whether the functions are realized 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 realize the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0064] In addition, each functional module in each embodiment of the present application can be integrated in one processing module, or each module can exist physically alone, or two or more modules can be integrated in one module.

[0065] The above is merely a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art can easily think of changes or replacements within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

[0066] Finally, the above is only a preferred embodiment of the present application and is not used to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application should be included in the protection scope of the present application.

Claims

1.A deep learning-based intelligent decomposition method for a design task, characterized in that, The method comprises the following steps: performing semantic analysis on the task text data to obtain a semantic representation; constructing a dynamic cognitive arrangement graph based on the semantic representation, generating a multi-granularity subtask candidate set based on the dynamic cognitive arrangement graph, and updating the dynamic cognitive arrangement graph; planning an optimal subtask disassembly path sequence based on the updated dynamic cognitive arrangement graph and the subtask candidate set through a sequence decision model; outputting a structured subtask tree based on the optimal subtask disassembly path sequence. 2.The deep learning based design task intelligent decomposition method according to claim 1, characterized in that, The semantic analysis on the task text data to obtain a semantic representation is specifically: parsing the design task text data to obtain semantic units; processing the semantic units by using a pre-trained language model to obtain word-level context representations; extracting target intent features and design constraint features from the context representations or the text data; fusing the target intent features and the design constraint features to form a unified representation vector, and mapping the unified representation vector to the semantic representation. 3.The deep learning based design task intelligent decomposition method of claim 2, wherein, The construction of the dynamic cognitive arrangement graph based on the semantic representation is specifically: performing cluster analysis on the semantic representation to form a plurality of cognitive operation units, and assigning a node identifier to each unit as a graph node; calculating the semantic correlation between the cognitive operation units; connecting the corresponding graph nodes based on the semantic correlation to generate a preliminary graph; optimizing the preliminary graph to generate a dynamic cognitive arrangement graph. 4.The deep learning based design task intelligent decomposition method of claim 3, wherein, The updating of the dynamic cognitive arrangement graph is specifically: extracting node representations and hierarchical information in the dynamic cognitive arrangement graph; generating high-level semantic representations and low-level structural representations based on the node representations; fusing the high-level semantic representations and the low-level structural representations, and updating the node representations, the fusion including calculating the correlation between the high-level semantic representations and the low-level structural representations, and performing weighted aggregation based on the correlation; generating a multi-granularity subtask candidate set based on the updated node representations; updating the dynamic cognitive arrangement graph by using the subtask candidate set. 5.The deep learning based design task intelligent decomposition method of claim 4, wherein, The planning of the optimal subtask disassembly path sequence based on the updated dynamic cognitive arrangement graph and the subtask candidate set through the sequence decision model is specifically: encoding the current dynamic cognitive arrangement graph and the 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 a node from the subtask candidate set according to the action probability distribution, and adding the selected node to the disassembly path; updating the current joint state representation, and iteratively performing the selection and updating steps until a termination condition is met, and outputting the optimal subtask disassembly path sequence. 6.The deep learning based design task intelligent decomposition method of claim 5, wherein, The sequence decision model is an improved meta-learning strategy network, and the meta-learning strategy network is specifically: introducing a path memory guide mechanism and a reversible compression mapping structure into the meta-learning strategy network; retaining the structural mode of the optimal disassembly sequence in the historical tasks by the path memory guide mechanism, and using it for disassembly initialization of similar subgraphs; mapping the dynamic cognitive graph to a low-dimensional latent space through the reversible compression mapping structure and realizing reversible decoding, so that the meta-learning strategy network inductively learns the mapping state of the dynamic cognitive graph in the unified latent space. 7.The deep learning based design task intelligent decomposition method of claim 6, wherein, The current dynamic cognitive arrangement graph and the subtask candidate set are encoded to obtain a joint state representation, specifically: The graph neural network is used for topology structure learning of the dynamic cognitive arrangement graph to generate a graph structure vector; The subtask candidate set is encoded by multi-feature fusion to generate a task embedding vector set; The graph structure vector and the task embedding vector set are spliced to form a joint state representation. 8.The deep learning based design task intelligent decomposition method of claim 7, wherein, The structured subtask tree is output based on the optimal subtask disassembly path sequence, specifically: The optimal subtask disassembly path sequence is subjected to structural disturbance to generate an counterfactual disassembly path sequence; The optimal subtask disassembly path sequence and the counterfactual disassembly path sequence are input into a comparative embedding encoder to obtain respective embedding representations; Based on the self-embedding representation, the final structured subtask tree is output. 9.The deep learning based design task intelligent decomposition method of claim 8, wherein, The counterfactual disassembly path sequence, specifically the generation steps are: The order and semantic embedding representation of each cognitive operation unit in the subtask disassembly path sequence are obtained; The structural disturbance operation is performed on the path sequence to generate multiple alternative path sequences; The semantic correlation constraint is applied to the generated alternative path sequences, and the sequence with a semantic correlation higher than a preset threshold with the original path sequence is selected and recorded as the counterfactual disassembly path sequence. 10.The deep learning based design task intelligent decomposition method of claim 9, wherein, The output structured subtask tree includes: The optimal subtask disassembly path sequence and the counterfactual disassembly path sequence are input into the comparative embedding encoder as positive samples and comparative samples, respectively; The embedding network in the comparative embedding encoder is used to encode the cognitive operation unit sequence in each path to generate a path-level semantic embedding representation; The embedding network is optimized to maximize the embedding distance between the positive samples and the comparative samples; Based on the optimal path semantic embedding representation obtained after optimization, the structured subtask tree is generated according to the multi-dimensional scoring criteria.

Images