A method for constructing a process step recommendation model in a process plan and application thereof

Abstract

This invention belongs to the field of mechanical parts processing technology planning, specifically involving a method for constructing a process step recommendation model in process planning and its application. The method includes: determining entity types and the types of relationships between entities; collecting data corresponding to each entity type to construct a process knowledge graph and training the embedding of each entity in the graph; training a policy network using deep reinforcement learning to obtain a process step recommendation model; and generating training samples by aggregating the embeddings of the entities corresponding to each selected process step in the graph to obtain contextual information embeddings; extracting features that guide the process steps to be recommended from these embeddings as potential relationship embeddings; fusing the contextual information embeddings and potential relationship embeddings as a knowledge state vector; generating a sequential preference state vector based on the positional information between the currently selected process steps; and fusing the two vectors as a training sample. This invention can improve the accuracy of process planning.

Description

Technical Field

[0001] This invention belongs to the field of mechanical parts processing technology planning, and more specifically, relates to a method for constructing a process step recommendation model in process planning and its application. Background Technology

[0002] The goal of process planning is to develop a feasible, optimal, or near-optimal sequence of processes for a part, and to rationally allocate manufacturing resources so that the part can be manufactured economically and competitively. Existing process planning systems are generally divided into generative and derivational systems. Knowledge graphs can reveal explicit and implicit relationships among various forms of complex process knowledge, making them a research hotspot for solving process planning problems.

[0003] A generative process planning method is disclosed in Chinese invention patent specification CN116976603A. By identifying the processing features of parts, obtaining the feature processing chain, designing a sequence selection hyperheuristic algorithm, setting the optimization objective as minimizing the maximum processing time, and fully considering various constraints, the process planning result of the parts is obtained.

[0004] A derived process planning method is disclosed in Chinese invention patent specification CN114781641A. It utilizes knowledge graph representation learning technology to learn a vector representation for each entity and relation in the knowledge graph, calculates the similarity of processing feature entities using multi-attribute similarity, and further enables the reuse of historical processing chain information.

[0005] Generative process planning methods only consider local information when utilizing knowledge, neglecting global knowledge and experience, which is detrimental to the accuracy of recommendation results. Derivative methods, including the use of knowledge graphs, mainly focus on retrieval through keywords, models, or descriptors, which may limit the solution space. This may result in the inability to cover a wider range of possibilities and variations, thus limiting the innovation and flexibility of process planning. Summary of the Invention

[0006] To address the shortcomings and improvement needs of existing technologies, this invention provides a method for constructing a process step recommendation model in process planning and its application, with the aim of improving the accuracy of process planning.

[0007] To achieve the above objectives, according to one aspect of the present invention, a method for constructing a process step recommendation model in process planning is provided, comprising:

[0008] Part features, tooling resources, and process information are respectively used as entity types in the knowledge graph to be constructed, and the types of relationships between entities are determined to construct a knowledge graph pattern layer; entity data corresponding to each preset entity type is collected and imported into the knowledge graph pattern layer to construct a knowledge graph data layer, thus obtaining a process knowledge graph; all triple samples in the knowledge graph are obtained, and the embedding of each triple entity is trained.

[0009] A process step recommendation model is obtained by constructing and training a policy network based on a training sample set using deep reinforcement learning.

[0010] The generation method for each training sample is as follows: Context information is aggregated from the embeddings of the entities corresponding to the selected process steps in the knowledge graph to obtain context information embeddings; features that have a directional effect on the process steps to be recommended are extracted from the context information embeddings and used as process step potential relationship embeddings; the context information embeddings and process step potential relationship embeddings are fused to obtain a knowledge state vector; a sequential preference state vector is generated based on the position information between the currently selected process steps; the knowledge state vector and the sequential preference state vector are fused to obtain a state vector, which serves as a training sample with the label of the set next process step.

[0011] Furthermore, the implementation method of training the process step recommendation model using deep reinforcement learning is as follows:

[0012] The policy network in deep reinforcement learning is used as the actor network. Its input is the training samples, and its output is a probability vector composed of the probabilities of each available process step being recommended. This vector is used to select the most suitable process step in the next time step. The output of the policy network is changed to the score value of the process step with the highest probability of being recommended. The resulting network is used as the critic network. Its input is the training samples and the most suitable process step in the next time step output by the actor network in the current iteration. This network is used to score the most suitable process step in the next time step and output the score value.

[0013] Based on the most suitable process step for the next moment obtained by the actor network after the current iteration and its corresponding label, calculate the reward value r. t According to the reward value r t The score V output by the critic network after the current iteration. t Calculate the residual TD = r t +γ·V t -V t-1 ; Calculate the loss function of the actor network after the current iteration. actor =-TD·logπ(p) t+1 |S t ), where π(p t+1|S t ) represents the sample S input to the actor network in the current iteration. t Select p below t+1 The probability, p t+1 This represents the most suitable action step selected by the actor network at the next time step t+1 after the current iteration; it also calculates the loss function of the critic network after the current iteration. critic =TD 2 γ represents the coefficient;

[0014] The parameters of the actor network and the critic network are updated according to the loss function; and the most suitable process step selected by the actor network at the next time step t+1 after the current iteration is added to the selected process steps to update the training samples. The above process is repeated until the termination condition is met. The final actor network is the process step recommendation model with a high recommendation index.

[0015] Furthermore, the reward value r t Hit-based reward function and correlation reward function The sum of;

[0016] Wherein, the hit reward function for:

[0017]

[0018] The correlation reward function Cosine similarity is used as the representation, as follows:

[0019]

[0020] In the formula, p t+1 The tag, z t+1 p t+1 Embedding of the corresponding entities in the knowledge graph of the aforementioned process. Indicate z t+1 The tag.

[0021] Furthermore, the context information embedding and the potential relationship embedding are fused by splicing.

[0022] Furthermore, the knowledge state vector and the sequential preference state vector are fused by splicing.

[0023] Furthermore, average pooling is used to aggregate contextual information for the embedded entities corresponding to each selected process step in the knowledge graph.

[0024] Furthermore, a multilayer perceptron is employed to perceive and extract features that have a directional effect on the recommended process steps from the embedded context information.

[0025] Furthermore, using a GRU network, the sequential preference state vector for process planning is derived based on the data of the entities corresponding to each currently selected process step in the knowledge graph.

[0026] The present invention also provides a process planning method, comprising:

[0027] S1. Based on the currently selected process steps, generate the corresponding state vectors using the training sample generation method described above.

[0028] S2. Obtain the process step recommendation model and process knowledge graph constructed by the construction method described above. Input the state vector obtained in S1 into the process step recommendation model to obtain the next process step, and obtain the data of the corresponding entity from the process knowledge graph. Repeat S1 until the process planning is completed.

[0029] The present invention also provides a computer-readable storage medium comprising a stored computer program, wherein, when the computer program is executed by a processor, it controls the device where the storage medium is located to execute a knowledge graph-guided deep reinforcement learning process planning method and / or a process planning method as described above.

[0030] In summary, the above-described technical solutions conceived in this invention can achieve the following beneficial effects:

[0031] (1) This invention provides a method for constructing a process step recommendation model in process planning. This method combines knowledge graphs and deep reinforcement learning to explore the learning potential of process knowledge graphs and make full use of information such as the decision-making order of processes and the contextual relationship between processes to construct state vectors as training samples. The state vectors can describe the attribute details and potential relationships of process knowledge. Then, deep reinforcement learning is used to train the model to obtain a more flexible and accurate process planning result recommendation model. This model has the ability to further enhance the energy of external knowledge information based on the use of historical sequence information, and therefore is expected to be more effective, accurate and interpretable in process recommendation.

[0032] (2) The present invention also proposes a preferred actor-critic model to perform deep reinforcement learning. Compared with other deep reinforcement learning models, the actor-critic model can improve the efficiency of model training and improve the stability of the model during the training process. Attached Figure Description

[0033] Figure 1A flowchart illustrating a method for constructing a process step recommendation model in process planning, provided in an embodiment of the present invention;

[0034] Figure 2 A schematic diagram of a process knowledge graph established based on a process knowledge base, provided for embodiments of the present invention;

[0035] Figure 3 This is a schematic diagram of the network structure of the actor-critic model provided in an embodiment of the present invention;

[0036] Figure 4 A schematic diagram illustrating the generation of processes based on a recommended model according to process steps, provided for an embodiment of the present invention;

[0037] Figure 5 A flowchart illustrating the generation of state vectors provided in an embodiment of the present invention. Detailed Implementation

[0038] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0039] Example 1

[0040] A method for constructing a process step recommendation model in process planning, such as... Figure 1 As shown, it includes:

[0041] Part features, tooling resources, and process information are respectively used as entity types in the knowledge graph to be constructed, and the types of relationships between entities are determined to construct the knowledge graph schema layer; part feature data, tooling resource data, and process information data in the pre-set part processing process are collected and imported into the knowledge graph schema layer to construct the knowledge graph data layer, thus obtaining the process knowledge graph; all triple samples in the process knowledge graph are obtained, and the embedding of each triple entity is trained.

[0042] A process step recommendation model is obtained by constructing and training a policy network based on a training sample set using deep reinforcement learning.

[0043] The generation method for each training sample is as follows: Context information is aggregated from the embeddings of the entities corresponding to each selected process step in the knowledge graph to obtain context information embeddings; feature vectors that can be used to accurately point to the ideal process step to be recommended are extracted from the context information embeddings as latent relationship embeddings of process steps; the context information embeddings and the latent relationship embeddings of process steps are fused to obtain a knowledge state vector; a sequential preference state vector is generated based on the position information of the currently selected process steps; the knowledge state vector and the sequential preference state vector are fused to obtain a state vector, which serves as a training sample with the label of the set next process step.

[0044] The aforementioned part feature data includes, for example, the roughness and dimensional accuracy of holes or slots. The entity data corresponding to each process step consists of a combination of part material properties, part material dimensions, part feature properties, process content, candidate machines, cutting tools, and tool approach direction (tad). The entity nodes corresponding to each process step reflect the context of the process links and the relationships between them. Furthermore, a knowledge graph embedding algorithm, such as the Trans-H model, is used to perform embedding training operations for each triple entity. The process step recommendation model conforms to a Markov decision process.

[0045] This embodiment develops a process knowledge database, which mainly contains data such as part information, process routes, and related process resources in the actual processing of the project. Based on this, a mapping relationship between entities, relations, and attributes is established to construct a process knowledge graph. Figure 2 This illustration demonstrates the knowledge graph established according to the above method requirements. In the knowledge graph, each leaf node corresponds to a process instance and contains corresponding attributes. Once the model agent (i.e., the policy network or the constructed process knowledge recommendation model) recommends specific process knowledge items to the process engineers, the knowledge details will also be derived through the graph. As an example, the Trans-H method is used for training, thereby obtaining the knowledge graph information for each entity and relation in the dataset.

[0046] Based on the above method, any deep reinforcement learning approach can solve the technical problem of this invention. There is no specific limitation on the deep reinforcement learning approach, but any approach can be used as a preferred implementation method, such as... Figure 3 As shown, the above implementation method for training the process step recommendation model using deep reinforcement learning can be as follows:

[0047] The policy network in deep reinforcement learning is used as the actor network. Its input is the training samples, and its output is a probability vector composed of the probabilities of each available process step being recommended. This vector is used to select the most suitable process step in the next time step. The output of the policy network is changed to the score value of the process step with the highest probability of being recommended. The resulting network is used as the critic network. Its input is the training samples and the most suitable process step in the next time step output by the actor network in the current iteration. This network is used to score the most suitable process step in the next time step and output the score value.

[0048] Based on the most suitable process step for the next moment obtained by the actor network after the current iteration and its corresponding label, calculate the reward value r. t According to the reward value r t The score V output by the critic network after the current iteration. t Calculate the residual TD = r t +γ·V t -V t-1 ; Calculate the loss function of the actor network after the current iteration. actor =-TD·logπ(p) t+1 |S t ), where π(p t+1 |S t ) represents the sample S input to the actor network in the current iteration. t Select p below t+1 The probability, p t+1 This represents the most suitable action step selected by the actor network at the next time step t+1 after the current iteration; it also calculates the loss function of the critic network after the current iteration. critic =TD 2 γ represents the coefficient;

[0049] The parameters of the actor network and the critic network are updated according to the loss function; and the most suitable process step selected by the actor network at the next time step t+1 after the current iteration is added to the selected process steps to update the training samples. The above process is repeated until the termination condition is met. The final actor network is the process step recommendation model with a high recommendation index.

[0050] For example, the operating system of the training environment selected in this embodiment is Windows 11, and the deep learning framework of the network model is PyTorch. The batch size is set to 1024, and the policy network is trained iteratively using the Adam optimizer. Specifically, a single-layer GRU network is used, with the GRU vector embedding size set to 50 and the Trans-H embedding vector size set to 50. For the model's hyperparameters, the learning rate is set to 10. -3The set length T is set to 3, the time difference exponent γ is 0.99, and the loss function hyperparameter λ is 10.

[0051] Meanwhile, the training samples are constructed as follows: For each part, this embodiment sorts the process planning routes of the part in the knowledge base, forming 9864 sets of process sequences. Parts with fewer than five processes are deleted, and each process sequence is augmented by randomly truncating the knowledge sequence. Ultimately, a total of 396,342 subsequences are collected for training. Based on the sorted sequences, the last item of each sequence is used as the test data. To accelerate the evaluation process, for each part, 100 negative processes are randomly selected and sorted together with the real processes.

[0052] During training, the actor network uses a policy function to assign each state s to... t Mapping this onto the action space 'a' yields the action probability distribution, where the softmax function is used to calculate the probability of each available action step. The policy function takes the following form:

[0053]

[0054] In the formula, z p* This represents the embedding of any process step among all available process steps in the knowledge graph into the entity p, where P represents all available process steps, and z (pi) This represents the entity p corresponding to the i-th process step in the knowledge graph. i The embedding, W is the parameter in the model, s t It is the current state embedding vector, and the softmax function has the following form:

[0055] As a preferred implementation method, the above-mentioned reward value r t Hit-based reward function and correlation reward function The sum of;

[0056] Among them, the hit reward function for:

[0057]

[0058] Correlation reward function Cosine similarity is used as the representation, as follows:

[0059]

[0060] In the formula, p t+1 The tag, z t+1 pt+1 Embedding of corresponding entities in the process knowledge graph Indicate z t+1 The tag.

[0061] A hit bonus is awarded only when the most appropriate process step matches the tag value at the next moment. It is used to measure the accuracy of recommendation performance; when the most suitable process step in the next time step has a high similarity to the knowledge embedding vector of the label value, it proves that the two are closely related in the knowledge graph, and it is used to measure the approximation of the recommendation result. The cosine similarity function is a commonly used similarity measurement function, and this function is used here to calculate the relevance reward function.

[0062] Given a sequence of work processes and a pre-trained actor network, a probability vector consisting of the probabilities of each selectable step can be output from the network. The selectable steps are then ranked according to this probability vector, and the step with the highest probability is selected as the recommended step. Figure 4 As shown, technicians retrieve data from the knowledge graph corresponding to the node of the recommended process step. Optionally, they can also retrieve data from adjacent nodes for reference. Through knowledge reuse (i.e., using the retrieved data), a machining process card for the part can be generated. Finally, the system receives feedback (i.e., receives the aforementioned recommended process steps) to form a new sequence of process steps as a new historical process sequence. The agent will then initiate a new round of recommendations based on this sequence until the process planning is completed.

[0063] Table 1 below shows the planning results of an application example of the above methods for process planning.

[0064] Table 1

[0065]

[0066]

[0067] In Table 1, each row represents a time unit, and each number represents a specific machine tool or cutting tool. Therefore, each row also corresponds to a decision step. The process plan is listed in chronological order of nine consecutive decisions. The selected operation and the selected resource number are written into the corresponding cell. This method can obtain all the process steps required for machining any part, as well as the process resources (including machine tools, cutting tools, tool approach direction, etc.) included in each step, and it offers higher accuracy compared to traditional methods.

[0068] As a preferred implementation, a concatenation method can be used to fuse the contextual information embedding and the latent relation embedding. Concatenation does not reduce the dimensionality of the input vector and retains all the information in both vectors, thus preserving more original information compared to other fusion methods (such as addition).

[0069] As a preferred implementation method, the knowledge state vector and the sequential preference state vector can be fused by concatenation. Concatenation does not reduce the dimensionality of the input vectors and retains all the information in both vectors, thus preserving more original information compared to other fusion methods (such as addition).

[0070] This can be used as a preferred implementation method, such as Figure 5 As shown, a simple average pooling method is used to aggregate the embedding information of each selected process step (i.e., the knowledge of the current process) to obtain the context information embedding.

[0071]

[0072] Compared to other aggregation methods (such as max pooling), average pooling focuses more on the overall properties of the input rather than extreme values ​​at specific locations, which helps to retain more information.

[0073] During process planning, the processing methods for different procedures exhibit significant differences. Their implicit relationships can be directly reflected in the vector representations of the currently selected process steps, which can be used as a preferred implementation method to embed contextual information. As input, a multi-layer neural network is used to directly perceive the feature vectors of the implicit pointing relationships between entities corresponding to the selected process steps, which can be used to more accurately point to the ideal process steps to be recommended. This yields the embedded vector representation of the implicit relationships between the selected process steps (i.e., the current process knowledge):

[0074]

[0075] MLP is a multilayer perceptron consisting of a single hidden layer, using tanh as the activation function, Φ mlp This indicates the parameters used in the MLP.

[0076] Given a historical sequence of steps p of a process plan (consisting of selected process steps). j:t Here, j represents the start time step, and t represents the current time step. GRU networks can effectively capture time dependencies and generate sequential representations of process interaction history. Therefore, as a preferred implementation, a GRU network can be used to derive the process planning sequential preference state vector y using the following formula. t .

[0077] γ t =σ[W γ ·(y t-1 ,p t )];

[0078]

[0079] β t =σ[W β ·(y t-1 ,p t )];

[0080]

[0081] Among them, y t-1 and p t Let β be the input vector. t and γ t These are the update gate and the reset gate, respectively. W γ W β W and W represent the weight matrices for the reset gate, update gate, and hidden state, respectively, and the hidden state is... The update function is the previous hidden state y t-1 and a new candidate hidden state p t Linear interpolation.

[0082] The historical state of the process planning can be represented by vector concatenation. t (symbol representation) and process knowledge state representation z t (i.e., the knowledge state vector at the current time t) are combined into a vector to model the state representation vector of the process recommendation at time t under the process planning environment.

[0083] Example 2

[0084] A process planning method, comprising:

[0085] S1. Based on the currently selected process steps, generate the corresponding state vectors using the training sample generation method described above.

[0086] S2. Obtain the process step recommendation model and process knowledge graph constructed by the construction method described above. Input the state vector obtained in S1 into the process step recommendation model to obtain the next process step, and obtain the data of the corresponding entity from the process knowledge graph. Repeat S1 until the process planning is completed.

[0087] To elaborate, for example, the process steps are as follows:

[0088] (1) Take part features, tooling resources and process information as entity types in the knowledge graph to be constructed, and determine the types of relationships between entities to construct the knowledge graph pattern layer; collect the part feature data, tooling resource data and process information data in the pre-set part processing process and import them into the knowledge graph pattern layer to construct the knowledge graph data layer and obtain the process knowledge graph; obtain all triple samples in the process knowledge graph and train to obtain the embedding of each triple entity.

[0089] (2) Aggregate the context information of the entities corresponding to the selected process steps in the knowledge graph to obtain the context information embedding; extract the feature vectors of the implicit pointing relationship between the entities corresponding to the selected process steps from the context information embedding, which can be used to more accurately point to the ideal process steps to be recommended, as the process step potential relationship embedding; fuse the context information embedding and the process step potential relationship embedding as the knowledge state vector.

[0090] (3) Generate a sequential preference state vector based on the position information between the currently selected process steps; merge the knowledge state vector and the sequential preference state vector to obtain a state vector, input it into the constructed process knowledge recommendation model, recommend the next process step, and obtain the data of the corresponding entity from the knowledge graph; repeat (2) until the process planning is completed.

[0091] The relevant technical solutions are the same as in Embodiment 1, and will not be repeated here.

[0092] Example 3

[0093] A computer-readable storage medium includes a stored computer program, wherein, when the computer program is executed by a processor, it controls the device where the storage medium is located to perform a knowledge graph-guided deep reinforcement learning process planning method and / or a process planning method as described above.

[0094] The relevant technical solutions are the same as those in Embodiment 1 and Embodiment 2, and will not be repeated here.

[0095] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for constructing a process step recommendation model in process planning, characterized in that, include: Part features, tooling resources, and process information are respectively regarded as entity types in the knowledge graph to be constructed, and the types of relationships between entities are determined to construct the knowledge graph pattern layer; entity data corresponding to each preset entity type are collected and imported into the knowledge graph pattern layer to construct the knowledge graph data layer, thus obtaining the process knowledge graph. Obtain all triplet samples in the knowledge graph and train to obtain the embedding of each triplet entity; A process step recommendation model is obtained by constructing and training a policy network based on a training sample set using deep reinforcement learning. The training sample set is generated as follows: the context information of the entity embeddings corresponding to each selected process step in the knowledge graph is aggregated to obtain the context information embedding; features that have a directional effect on the process step to be recommended are extracted from the context information embedding as the potential relationship embedding of the process step. The contextual information embedding and the potential relationship embedding of the process steps are fused together to form a knowledge state vector; a sequential preference state vector is generated based on the position information between the currently selected process steps; the knowledge state vector and the sequential preference state vector are fused together to obtain a state vector, which is used as a training sample with the label of the set next process step.

2. The construction method according to claim 1, characterized in that, The implementation method of training the process step recommendation model using deep reinforcement learning is as follows: The policy network in deep reinforcement learning is used as the actor network. Its input is the training samples, and its output is a probability vector composed of the probabilities of each available process step being recommended. This vector is used to select the most suitable process step in the next time step. The output of the policy network is changed to the score value of the process step with the highest probability of being recommended. The resulting network is used as the critic network. Its input is the training samples and the most suitable process step in the next time step output by the actor network in the current iteration. This network is used to score the most suitable process step in the next time step and output the score value. Based on the most suitable process step for the next moment obtained by the actor network after the current iteration and its corresponding label, calculate the reward value r. t According to the reward value r t The score V output by the critic network after the current iteration. t Calculate the residual TD = r t +γ·V t -V t-1 ; Calculate the loss function of the actor network after the current iteration. actor =-TD·logπ(p) t+1 |S t ), where π(p t+1 |S t ) represents the sample S input to the actor network in the current iteration. t Select p below t+1 The probability, p t+1 This represents the most suitable action step selected by the actor network at the next time step t+1 after the current iteration; it also calculates the loss function of the critic network after the current iteration. critic =TD 2 γ represents the coefficient; The parameters of the actor network and the critic network are updated according to the loss function; and the most suitable process step selected by the actor network at the next time step t+1 after the current iteration is added to the selected process steps to update the training samples. The above process is repeated until the termination condition is met. The final actor network is the process step recommendation model with a high recommendation index.

3. The construction method according to claim 2, characterized in that, The reward value r t For the hit reward function r t 1 and correlation reward function r t 2 The sum of; Wherein, the hit reward function r t 1 for: The correlation reward function r t 2 Cosine similarity is used as the representation, as follows: In the formula, p t+1 The tag, z t+1 p t+1 Embedding of the corresponding entities in the knowledge graph of the aforementioned process. Indicate z t+1 The tag.

4. The construction method according to claim 1, characterized in that, The context information embedding and the potential relationship embedding are fused by splicing.

5. The construction method according to claim 1, characterized in that, The knowledge state vector and the sequential preference state vector are fused by splicing.

6. The construction method according to claim 1, characterized in that, The average pooling method is used to aggregate the contextual information of the entities corresponding to the currently selected process steps in the knowledge graph.

7. The construction method according to claim 1, characterized in that, A multilayer perceptron is used to perceive and extract features that have a directional effect on the recommended process steps from the embedded context information.

8. The construction method according to claim 1, characterized in that, Using a GRU network, the sequential preference state vector for process planning is derived based on the data of the entities corresponding to each currently selected process step in the knowledge graph.

9. A process planning method, characterized in that, include: S1. Based on the currently selected process steps, generate the corresponding state vector using the training sample generation method in the construction method described in any one of claims 1 to 8. S2. Obtain the process step recommendation model and process knowledge graph constructed by the construction method as described in any one of claims 1 to 8, input the state vector obtained in S1 into the process step recommendation model to obtain the next process step, and obtain the data of the corresponding entity from the process knowledge graph; repeat S1 until the process planning is completed.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored computer program, wherein, when the computer program is executed by a processor, it controls the device in which the storage medium is located to perform the construction method as described in any one of claims 1 to 8 and / or the process planning method as described in claim 9.

Images