Process knowledge graph dynamic reasoning method and system based on sparse activation

Abstract

The invention provides a process knowledge graph dynamic reasoning method and system based on sparse activation in order to solve the problems that when an existing reasoning method carries out process reasoning based on a large-scale and dynamically-changed and updated process knowledge graph, the operation time is long, the needed calculation cost is high, and the reasoning result is inaccurate. According to an expected target of actual production, constructing a relationship associated with the expected target in the knowledge graph, extracting a sub-knowledge graph associated with the expected target, calculating a relationship weight matrix of the sub-knowledge graph, and endowing the relationship weight matrix into the sub-knowledge graph to obtain a multi-relationship weight model; extracting a sparse function activation sub-graph from the multi-relation weight model by utilizing a sparse activation function, removing redundant information which is not greatly associated with the current actual production in the knowledge graph to greatly reduce nodes and the relation number of the knowledge graph, and finally operating an inference algorithm in the sparse activation function sub-graph to obtain the knowledge graph. The operation time of the reasoning algorithm is greatly reduced, and the required computing resources are greatly reduced.

Description

Technical Field

[0001] The present invention relates to a method and system for dynamic reasoning of a process knowledge graph. Background Art

[0002] Process knowledge is the key to ensuring the quality of complex products. A process knowledge graph is a structured semantic knowledge base that stores and represents entities (such as materials, equipment, process parameters, etc.) and their relationships (such as the relationship between materials and process parameters, the relationship between equipment and process parameters, etc.) in the form of a graph. Therefore, the process knowledge graph can help enterprises and researchers better understand and manage complex process knowledge, promote process innovation and optimization, and provide support for ensuring and improving the quality of complex products.

[0003] To meet the actual industrial needs, the process knowledge graph for complex products is characterized by large-scale and dynamic changes and updates. When using existing reasoning methods to perform reasoning based on such a process knowledge graph in order to obtain reasoning results that can guide process design, the following problems exist:

[0004] 1. The process reasoning algorithm takes a long time to run;

[0005] Based on a large-scale process knowledge graph, existing process reasoning algorithms search and run on a huge number of nodes and complex relationships in the global process knowledge graph. Since all nodes and relationships in the global graph need to be traversed, even for simple reasoning tasks, it still leads to global traversal of the process reasoning algorithm, low efficiency, and long time consumption.

[0006] 2. The computational cost required by the process reasoning algorithm is high;

[0007] As the scale of the process knowledge graph gradually expands, the number of nodes in the process knowledge graph continues to increase, resulting in an exponential growth in the number of relationships. This makes the reasoning algorithm need to process complex and huge amounts of relationships during operation, especially for multi-hop reasoning, which requires higher-performance computing resources, resulting in a high required computational cost.

[0008] 3. The reasoning results of the process reasoning algorithm are inaccurate;

[0009] Since most process knowledge graphs are static graphs and existing reasoning is mostly static reasoning, on the one hand, it cannot fit the actual production, and on the other hand, when applying static reasoning to a dynamic knowledge graph, it will lead to a decrease in the accuracy of the reasoning results. Summary of the Invention

[0010] In order to solve the technical problems that existing reasoning methods have when performing process reasoning based on a large-scale, dynamically changing and updated process knowledge graph, such as long running time consumption, high required computational cost, and inaccurate reasoning results, the present invention proposes a method and system for dynamic reasoning of a process knowledge graph based on sparse activation.

[0011] The technical solution of the present invention is as follows:

[0012] A dynamic reasoning method for a process knowledge graph based on sparse activation, which is characterized in that it includes the following steps:

[0013] Step 1: Construct a multi-relationship weight model;

[0014] Step 1.1: Set the expected goal, and construct multi-relations associated with the expected goal in the current processing process knowledge graph and add relationship attributes;

[0015] Step 1.2: Calculate the weight matrix of the multi-relationship between any two different nodes in the current processing process knowledge graph that is related to the multi-relationship constructed in Step 1.1;

[0016] Step 1.2.1: Extract a sub-knowledge graph from the current processing process knowledge graph that is only associated with the multi-relationship, and construct a diagonal matrix and an attribute feature matrix between any two different nodes in the sub-knowledge graph and an attribute feature matrix

[0017]

[0018] Among them, the diagonal matrix has a size of m ij ×m ij , where m ij is the number of efficiency preference relationships between nodes i and j; the size of the attribute feature matrix is m ij ×1; x 1 , x 2 , …, are respectively the attribute feature values of the first efficiency preference relationship, the second efficiency preference relationship, …, the m ij th efficiency preference relationship between nodes i and j, which are known quantities;

[0019] Step 1.2.2: Calculate the product of the diagonal matrix and the attribute feature matrix to obtain a matrix H ij with a size of m ij ×1;

[0020] Step 1.2.3: Normalize each element in the matrix H ij , that is, update each element value in the matrix H ij to the reciprocal value of the element to obtain the matrix H i ′ j ;

[0021] Step 1.2.4: Subtract 1 from each element of matrix H i ′ j to obtain the weight matrix W of the efficiency preference relationship ij ;

[0022] Step 1.3: Construct a multi-relationship weight model;

[0023] Assign the element values in the weight matrix of the efficiency preference relationship obtained in Step 1.2 to the corresponding relationships in the current sub-knowledge graph respectively, and update the current sub-knowledge graph. The sub-knowledge graph containing relationship weight information obtained after the update is the multi-relationship weight model;

[0024] Step 2: Set a weight threshold, use the weight threshold and the multi-relationship weight model as the input of the sparse activation function, and use the sparse activation function to extract the sparse function activation sub-graph from the multi-relationship weight model obtained in Step 1;

[0025] Step 3: Select an inference algorithm according to the expected goal set in Step 1.1, and run the selected inference algorithm in the sparse function activation sub-graph to obtain an inference result;

[0026] Step 4: Determine whether the preferred process plan corresponding to the current inference result meets the production requirements. If it meets, the process ends; if it does not meet, go to Step 5;

[0027] Step 5: Dynamically update the current large-scale process knowledge graph according to the actual situation of the current production workshop, including materials, tools, machining features, and equipment. After the update is completed, return to Step 1.

[0028] Further, the expected goal in Step 1.1 is efficiency preference, quality preference, or both efficiency and quality; correspondingly, the multi-relationship is efficiency preference relationship, quality preference relationship, or both efficiency and quality relationship.

[0029] Further, the weight threshold in Step 2 is 0.7 - 0.9.

[0030] Further, the method for the sparse activation function to extract the sparse function activation sub-graph from the multi-relationship weight model obtained in Step 1 is as follows:

[0031] The sparse activation function compares the weight values of each relationship in the multi-relationship weight model with a preset weight threshold respectively, and judges whether to activate the attribute feature values of each relationship according to the comparison results:

[0032] If the weight value of a certain relationship is greater than or equal to the weight threshold, multiply the weight value of this relationship by 1 as the updated weight value to activate its attribute feature value;

[0033] If the weight value of a certain relationship is less than the weight threshold, then multiply the weight value of this relationship by 0 as the updated weight value to deactivate its attribute feature value;

[0034] When the weight values of all relationships in the multi-relationship weight model have been updated, the multi-relationship weight model with updated weight information at this time is the sparse function activation subgraph.

[0035] Further, in step 3, if the expected goal is efficiency optimization, the inference algorithm is the efficiency optimization algorithm; if the expected goal is quality optimization, the inference algorithm is the quality optimization algorithm; if the expected goal is to balance efficiency and quality, the inference algorithm is the algorithm for balancing efficiency and quality optimization.

[0036] Further, the efficiency optimization algorithm is Dijkstra's shortest path algorithm; the quality optimization algorithm is the PageRank algorithm; the algorithm for balancing efficiency and quality optimization is Dijkstra's shortest path algorithm or the PageRank algorithm.

[0037] Further, the method for determining whether the preferred process plan corresponding to the current inference result meets the production requirements in step 6 is as follows:

[0038] If the expected goal is efficiency optimization, then combined with the actual processing efficiency of the current production workshop, estimate whether the total time spent on the preferred process plan corresponding to the current inference result meets the delivery time required by the customer. If so, it indicates that the current inference result meets the production requirements; otherwise, it indicates that the current inference result does not meet the production requirements;

[0039] If the expected goal is quality optimization, then combined with the actual processing capacity of the current production workshop, estimate whether the current production workshop can meet the processing accuracy required by the preferred process plan corresponding to the current inference result, and determine whether the processing accuracy of this preferred process plan can meet the processing accuracy required by the customer. If all of the above are met, it indicates that the current inference result meets the production requirements; otherwise, it indicates that the current inference result does not meet the production requirements;

[0040] If the expected goal is to balance efficiency and quality, then combined with the actual processing efficiency of the current production workshop, estimate whether the time spent on the preferred process plan corresponding to the current inference result meets the delivery time required by the customer. At the same time, combined with the actual processing capacity of the current production workshop, estimate whether the current production workshop can meet the processing accuracy required by this preferred process plan, and determine whether the processing accuracy of this preferred process plan can reach the processing accuracy required by the customer. If all of the above are met, it indicates that the current inference result meets the production requirements; otherwise, it indicates that the current inference result does not meet the production requirements.

[0041] The present invention also provides a dynamic inference system for a process knowledge graph based on sparse activation, including a processor and a storage medium, and a computer program is stored on the storage medium; the special feature is that: when the computer program is run, it executes the above-mentioned dynamic inference method for the process knowledge graph based on sparse activation.

[0042] Compared with the prior art, the beneficial effects of the present invention are:

[0043] 1. According to the expected goals of actual production, the present invention constructs (adds) relationships associated with the expected goals in the knowledge graph (such as efficiency optimization relationships, quality optimization relationships, or relationships that take into account both efficiency and quality), then extracts the sub-knowledge graph associated with the constructed expected goals, calculates the relationship weight matrix of the sub-knowledge graph, and assigns the elements (essentially weight values) in the relationship weight matrix to the sub-knowledge graph to obtain a multi-relationship weight model; then uses a sparse activation function to extract a sparse function activation sub-graph from the multi-relationship weight model, removes redundant information in the process knowledge graph that has little association with the current actual production, greatly reduces the number of nodes and relationships in the process knowledge graph, and finally runs an inference algorithm in the extracted sparse activation function sub-graph, so that the inference algorithm does not need to traverse the entire knowledge graph, so the running time of the inference algorithm is greatly reduced and the required computing resources are greatly reduced.

[0044] 2. After obtaining the inference result, the present invention will judge whether the inference result meets the production requirements. If not, it will dynamically update the current process knowledge graph according to the actual production, and then perform process inference on the basis of the updated process knowledge graph, and so on in a cyclic iteration, improving the accuracy of the inference and finally obtaining an optimized process plan that meets the production requirements. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 is a flowchart of the method of the present invention.

[0046] Figure 2 is a multi-relationship weight model obtained in an embodiment of the present invention when the expected goal is efficiency optimization (the pink lines with arrows represent the efficiency optimization relationship, and the numbers on the pink lines are the weights of the efficiency optimization relationship).

[0047] Figure 3 is a multi-relationship weight model obtained in an embodiment of the present invention when the expected goal is quality optimization (the blue lines with arrows represent the quality optimization relationship, and the numbers on the blue lines are the weights of the quality optimization relationship).

[0048] Figure 4When the expected goal is to balance efficiency and quality, the multi-relationship weight model obtained in the embodiments of the present invention (the pink lines with arrows represent the efficiency preference relationships, the blue lines with arrows represent the quality preference relationships, and the numbers on the lines are the weights of the relationships).

[0049] Figure 5 The sparse function activation subgraph obtained in the embodiments of the present invention when the expected goal is efficiency preference.

[0050] Figure 6 The sparse function activation subgraph obtained in the embodiments of the present invention when the expected goal is quality preference.

[0051] Figure 7 The sparse function activation subgraph obtained in the embodiments of the present invention when the expected goal is to balance efficiency and quality.

[0052] Figure 8 It is a visualization example diagram of the inference results respectively obtained in the embodiments of the present invention under three different expected goals of the present invention. (a) corresponds to efficiency preference, (b) corresponds to quality preference, and (c) corresponds to balancing efficiency and quality.

[0053] Figure 9 It is a schematic diagram of the principle for dynamically updating the process knowledge graph in the present invention. Detailed implementation manners

[0054] Large-scale process knowledge graphs have the characteristics of numerous nodes and complex relationships. The process knowledge graph of complex aviation products belongs to a large-scale process knowledge graph. Such process knowledge graphs are usually constructed based on processing technology triples with a huge amount of entities and relationships.

[0055] The processing technology triple is a standard expression composed of a head entity, a tail entity, and the relationship between them. The processing technology triple can be expressed in the form of (head entity)-[relationship]->(tail entity), such as (forged part)-[:efficiency preference{weight:0.897}]->(rough turning) and (rough turning)-[:efficiency preference{weight:0.533}]->(heat treatment).

[0056] The processing technology triples with a huge amount of entities and relationships are constructed based on multi-stage, multi-source, and multi-modal process data and knowledge. Among them, multi-stage means that the process data and knowledge involve the part design stage, the processing stage, and the service stage. Multi-source means that the part process data and knowledge come from multiple sources, which can come from designers, process engineers, production workers, quality inspectors, etc. Multi-modal means that the process data knowledge includes different modalities, such as text, table, audio, and video, etc.

[0057] Nodes, relationships, and attributes (relationship attributes, node attributes) are important components of a process knowledge graph. Through the combination of nodes and relationships, the process knowledge graph can comprehensively display complex process flows. In a machining process knowledge graph, nodes are usually represented by circles, and relationships are represented by directed arrows between nodes.

[0058] Next, with reference to the machining process knowledge graph of a certain compressor annular casing in the aviation field as an example for reasoning, the present invention will be described in detail to make the technical solutions and advantages of the present invention clearer and easier to understand.

[0059] Refer to Figure 1 the following process. Based on the machining process knowledge graph of a certain compressor annular casing in the aviation field, the method for process reasoning using the present invention specifically includes the following steps:

[0060] Step 1: Construct a multi-relationship weight model;

[0061] Step 1.1: Construct multi-relationships in the current machining process knowledge graph;

[0062] First, set the expected goal as efficiency optimization, and construct an efficiency optimization relationship in the current machining process knowledge graph of a certain compressor annular casing in the aviation field;

[0063] Then, add attributes to the constructed efficiency optimization relationship. In this embodiment, the attribute of the efficiency optimization relationship is the machining time of the machining process of a certain compressor annular casing part, in seconds.

[0064] Step 1.2: Calculate the weight matrix of the efficiency optimization relationship between any two different nodes related to the efficiency optimization relationship constructed in Step 1.1 in the current machining process knowledge graph;

[0065] Step 1.2.1: Extract from the current machining process knowledge graph a sub-knowledge graph that is only associated with the efficiency optimization relationship. There are n nodes in the extracted sub-knowledge graph. For any two different nodes among these n nodes, construct a diagonal matrix and an attribute feature matrix

[0066]

[0067] where the diagonal matrix has a size of m ij ×m ij , and m ij is the number of efficiency optimization relationships between nodes i and j; the size of the attribute feature matrix is m ij ×1; x 1 , x 2 , …, They are the first efficiency preference relationship, the second efficiency preference relationship, …, the mth efficiency preference relationship between nodes i and j respectively ij The attribute characteristic values of the efficiency preference relationships (i.e., the specific numerical values of the attributes. If the attribute is processing time, the attribute characteristic value is the specific numerical value of the processing time, such as 120 s) are known quantities.

[0068] For the sake of easy understanding, examples are given here for explanation. Assume that there are three nodes related to the efficiency preference relationships, namely A, B, and C. There are 4 efficiency preference relationships between nodes A and B, 2 efficiency preference relationships between nodes B and C, and 1 efficiency preference relationship between nodes A and C. Then, a 4×4 diagonal matrix is constructed for nodes A and B. The elements on the diagonal of this diagonal matrix are all 1 / 4, and the rest of the elements are 0; a 2×2 diagonal matrix is constructed for nodes B and C. The elements on the diagonal of this diagonal matrix are all 1 / 2, and the rest of the elements are 0; a 1×1 diagonal matrix is constructed for nodes A and C. This diagonal matrix has only 1 element, which is 1; since there are 4 efficiency preference relationships between nodes A and B, and the attribute of each efficiency preference relationship is the processing time of the processing process, and each efficiency preference relationship has 1 attribute characteristic value, then the characteristic matrix between nodes A and B is a 4×1 matrix, and the element values of this matrix are respectively the attribute characteristic values of the 4 efficiency preference relationships between nodes A and B; there are 2 efficiency preference relationships between nodes B and C, and the characteristic matrix between nodes B and C is a 2×1 matrix, and the element values of this matrix are respectively the attribute characteristic values of the 2 efficiency preference relationships between nodes B and C; there is 1 efficiency preference relationship between nodes A and C, and the characteristic matrix between nodes A and C is a 1×1 matrix, and the element value of this matrix is the attribute characteristic value of the 1 efficiency preference relationship between nodes A and C.

[0069] Step 1.2.2: Calculate the product of the diagonal matrix and the attribute characteristic matrix to obtain a matrix H ij of size m ij ×1:

[0070]

[0071] where

[0072] Step 1.2.3: Normalize each element in the matrix H ij , that is, update the element values in the matrix H ij to the reciprocal values of these elements to obtain the matrix H i ′ j :

[0073]

[0074] Step 1.2.4: Subtract 1 from each element of matrix H i ′ j to obtain the weight matrix W of the efficiency preference relationship ij :

[0075]

[0076] where W ij is the weight matrix of the efficiency preference relationship between nodes i and j, and each element in W ij is the weight of each efficiency preference relationship between nodes i and j, that is is the weight value of the 1st, 2nd,..., m ij th efficiency preference relationships between nodes i and j.

[0077] Step 1.3: Construct a multi-relationship weight model;

[0078] Correspondingly assign the element values in the weight matrix of the efficiency preference relationship obtained in Step 1.2 to the corresponding efficiency preference relationships in the current sub-knowledge graph, and update the current sub-knowledge graph. The sub-knowledge graph containing relationship weight information obtained after the update is the multi-relationship weight model. For example, for nodes A and B, if there are 4 efficiency preference relationships between them, the weight matrix of the efficiency preference relationship between nodes A and B has 4 elements, and these 4 elements are the weight values of the 4 efficiency preference relationships between nodes A and B respectively. Then, assign the 4 element values of the weight matrix to the 4 efficiency preference relationships between nodes A and B respectively; for the remaining nodes, use the same method to assign weight values to each efficiency preference relationship between every two nodes. The finally obtained knowledge graph with relationship weight values is the constructed multi-relationship weight model, as shown in Figure 2 shown.

[0079] Step 2: Extract the sparse function activation subgraph from the multi-relationship weight model obtained in Step 1;

[0080] Step 2.1: Set the weight threshold;

[0081] Considering that if the weight threshold is too large, key information will be missed, resulting in a decrease in the accuracy of the inference result; if the weight threshold is too small, there will be too much redundant information and the inference efficiency will be too low. After a large number of simulations and actual verifications, it is appropriate to take the weight threshold as 0.7 - 0.9. In this embodiment, the weight threshold is taken as 0.8.

[0082] Step 2.2: Apply the sparse activation function to the multi-relationship weight model constructed in Step 1 to obtain the sparse function activation subgraph;

[0083] Take the weight values of each efficiency optimization relationship in the multi-relationship weight model constructed in Step 1 and the weight threshold set in Step 2.1 as the input of a sparse activation function (a well-known function that can be implemented based on one-hot encoding). The sparse activation function compares the weight values of each efficiency optimization relationship with the preset weight threshold respectively, and determines whether to activate the attribute feature values of each efficiency optimization relationship according to the comparison results:

[0084] If the weight value of a certain efficiency optimization relationship is greater than or equal to the weight threshold, then multiply the weight value of this efficiency optimization relationship by 1 as the updated weight value (i.e., the weight value remains unchanged) to activate its attribute feature value;

[0085] If the weight value of a certain efficiency optimization relationship is less than the weight threshold, then multiply the weight value of this efficiency optimization relationship by 0 as the updated weight value (i.e., reset the weight value to 0) to deactivate its attribute feature value;

[0086] After the weight values of all efficiency optimization relationships are updated, the multi-relationship weight model with updated weight information at this time is the sparse function activation subgraph, as Figure 5 shown.

[0087] Step 3: Run the inference algorithm to obtain the inference result;

[0088] Step 3.1: Select the corresponding inference algorithm according to the expected goal set in Step 1.1. Since the expected goal set in Step 1.1 is efficiency optimization, the inference algorithm selected here is the efficiency optimization algorithm, such as Dijkstra's shortest path algorithm (a known algorithm);

[0089] Step 3.2: Run the inference algorithm selected in Step 3.1 in the sparse function activation subgraph obtained in Step 2. The inference algorithm outputs the inference result, and this inference result is the preferred process plan that meets the expected goal;

[0090] Step 3.3: Visualize the inference result obtained in Step 3.2, as exemplarily shown in Figure 8 (a) in. It can be seen from the figure that the sparse activation probability weights of the path from the starting node (forged part) to the ending node (magnetic flaw detection) are 0.886, 0.832, 0.871, and 0.876 respectively, and the sum of the path probability weights has a total value of 3.465. It should be noted that Step 3.3 is not necessary, but just to more intuitively display the inference result.

[0091] Step 4: Determine whether the preferred process plan corresponding to the current inference result meets the production requirements. If it meets, the process ends; if it does not meet, enter Step 5;

[0092] The method for determining whether the current inference result meets the requirements of efficient and preferred production is as follows:

[0093] For the goal of efficient optimization, if the inference result corresponds to a preferred process plan, then by combining the actual processing efficiency of the current production workshop, estimate whether the total time spent on the current preferred process plan meets the delivery time required by the customer. If so, it indicates that the current inference result meets the production requirements; otherwise, it indicates that the current inference result does not meet the production requirements.

[0094] Step 5: Dynamically update the current large-scale process knowledge graph;

[0095] To better promote process optimization, in this embodiment, according to the actual situation of the current production workshop, including materials, tools, machining features, and equipment, the current large-scale process knowledge graph is dynamically updated, including the addition and deletion of nodes in the graph, and the discovery and disconnection of relationships. For example, for a certain part, when its processing material increases, it will cause an increase in the material entity node in the large-scale process knowledge graph. When a new tool is used for machining, it will cause an increase in the tool entity node in the large-scale process knowledge graph, etc. Exemplarily, as Figure 9 shown Figure 9 Figure (a) in is the global initial graph, which has 3 nodes, namely node A, node B, and node C. There are n relationships between node A and node B, and m relationships between node B and node C; Figure 9 Figure (b) in is obtained by adding node D to the global initial graph shown in Figure (a); Figure 9 Figure (c) in is obtained by discovering a relationship r AD between node A and node D in Figure (b). The initial weight of this relationship r AD is ω l . A relationship r BC is also discovered between node B and node C. The initial weight of this relationship r BC is ω m+1 . Figure 9 Figure (c) in is the new graph obtained after dynamically updating the global initial graph shown in Figure (a) in Figure 9 .

[0096] After the dynamic update of the current large-scale process knowledge graph is completed, return to Step 1.

[0097] The above process is only illustrated by taking efficient optimization as the expected goal as an example, which is the process of performing process inference using the present invention based on the processing process knowledge graph of a certain compressor annular casing in the aviation field.

[0098] When performing process reasoning based on the processing technology knowledge graph of the annular casing of a certain aviation compressor in the above embodiment, the expected goal can also be set to quality optimization, or to take into account both efficiency and quality.

[0099] When the expected goal is set to quality optimization, the method of process reasoning is the same as the above method flow, with the only difference being that the multi-relation constructed in step 1.1 is a quality optimization relationship, and the attributes of the quality optimization relationship are related to the processing quality, such as the surface roughness value of the part processing or the processing error, etc.; the weight matrix calculated in step 1.2 is the weight matrix of the quality optimization relationship. At this time, the multi-relation weight model obtained is as follows: Figure 3 As shown; in step 2.2, the input of the sparse activation function is the weight value of each quality preference relationship and the set weight threshold. At this time, the sparse function activation subgraph obtained is as follows Figure 6 As shown; the reasoning algorithm selected in step 3.1 and run in step 3.2 is a quality optimization algorithm, such as the PageRank algorithm, and the obtained reasoning result is an optimal process solution that meets the desired goal of quality optimization, and the visual display is exemplarily as shown in Figure 8 As shown in (b), it can be seen that the sparse activation probability weights of the path from the starting node forging to the ending node magnetic flaw detection are 0.825, 0.812, 0.881 and 0.836 respectively, and the total value of the path probability weight is 3.354. Correspondingly, the method for determining whether the current reasoning result meets the quality optimization production requirements is: for the quality optimization goal, the reasoning result corresponds to an optimal process plan, then as long as the actual processing capacity of the current production workshop (such as the processing accuracy of the equipment, etc.) is combined, it is estimated whether the current production workshop can meet the processing accuracy required by the optimal process plan, and whether the processing accuracy of the optimal plan can meet the processing accuracy required by the customer. If all of the above are true, it means that the current reasoning result meets the production requirements; otherwise, it means that the current reasoning result does not meet the production requirements.

[0100] When the expected goal is set to take both efficiency and quality into consideration, the method for process reasoning is the same as the principle and process of the above method, with the only difference being that the multi-relation constructed in step 1.1 is a relationship that takes both efficiency and quality into consideration (the efficiency optimization relationship and the quality optimization relationship can be constructed separately and then superimposed, or the efficiency and quality relationship can be directly constructed), and the attributes of the efficiency and quality relationship are related to the processing efficiency and quality, such as the processing time, the surface roughness value of the part processing, or the processing error; the weight matrix calculated in step 1.2 is a weight matrix that takes both efficiency and quality into consideration. At this time, the multi-relation weight model obtained is as follows: Figure 4 As shown; in step 2.2, the input of the sparse activation function is the weight value of each line that takes into account the relationship between efficiency and quality and the set weight threshold. At this time, the sparse function activation subgraph obtained is as follows Figure 7As shown; the desired goal algorithm selected in step 3.1 and run in step 3.2 is an optimization algorithm that takes into account both efficiency and quality, such as the Dijkstra shortest path algorithm or the PageRank algorithm, and the inference result obtained is an optimal process solution that takes into account both efficiency and quality. The visualization is exemplarily shown as follows Figure 8 As shown in (c), it can be seen that the sparse activation probability weights of the magnetic flaw detection path from the starting node forging to the ending node are 0.825, 0.812, 0.871 and 0.836 respectively, and the total value of the probability weight of the path is 3.344. Correspondingly, for the efficiency and quality goals, the reasoning result corresponds to an optimal process plan. Then, as long as the actual processing efficiency of the current production workshop is combined to estimate whether the time spent on the current optimal process plan meets the delivery time required by the customer, and the actual processing capacity of the current production workshop (such as the processing accuracy of the equipment, etc.) is combined to estimate whether the current production workshop can meet the processing accuracy required by the optimal process plan, and judge whether the processing accuracy of the optimal process plan can meet the processing accuracy required by the customer. If all of the above are true, it means that the current reasoning result meets the production requirements; otherwise, it means that the current reasoning result does not meet the production requirements.

[0101] The above is a detailed description of the specific implementation methods of the present invention in combination with the accompanying drawings. Finally, it should be noted that the above description and the accompanying drawings are not limitations on the protection scope of the present invention. Various modifications or variations made by those skilled in the art on the basis of the technical solution of the present invention without the need for creative labor are still within the protection scope of the present invention.

Claims

1. A dynamic reasoning method for process knowledge graph based on sparse activation, characterized in that: The following steps are involved: Step 1: Construct a multi-relationship weight model; Step 1.1: Set the expected goal, build multiple relationships associated with the expected goal in the current machining process knowledge graph according to the expected goal and add relationship attributes; Step 1.2: Calculate the weight matrix of the multi-relationship between any two different nodes related to the multi-relationship constructed in step 1.1 in the current machining technology knowledge graph; Step 1.2.1: Extract the sub-knowledge graph associated only with the multi-relationship from the current processing technology knowledge graph, and construct a diagonal matrix between any two different nodes in the sub-knowledge graph and an attribute feature matrix Among them, the diagonal matrix The size is m ij ×m ij , m ij is the number of efficiency optimization relationships between nodes i and j; the size of the attribute feature matrix is ​​m ij ×1; x1, x2, …, are the first and second efficiency optimization relations between nodes i and j, respectively, and the mth efficiency optimization relations between nodes i and j are ij The attribute characteristic value of the strip efficiency optimization relation is a known quantity; Step 1.2.2: Calculate the diagonal matrix and attribute feature matrix The product of size m ij ×1 matrix H ij ; Step 1.2.3: Transform the matrix H ij Normalize the elements in the matrix H ij Update the value of each element in to the reciprocal value of the element, and get the matrix H i ' j ; Step 1.2.4: Transform the matrix H i ' j Each element of is subtracted from 1 to obtain the weight matrix W of the efficiency optimization relationship. ij ; Step 1.3: Construct a multi-relationship weight model; Assign the element values ​​in the weight matrix of the efficiency optimization relationship obtained in step 1.2 to the corresponding relationships in the current sub-knowledge graph, and update the current sub-knowledge graph. The sub-knowledge graph containing the relationship weight information obtained after the update is the multi-relationship weight model; Step 2: setting a weight threshold, taking the weight threshold and the multi-relation weight model as inputs of a sparse activation function, and using the sparse activation function to extract a sparse function activation subgraph from the multi-relation weight model obtained in step 1; Step 3: Select an inference algorithm according to the desired goal set in step 1.1, run the selected inference algorithm in the sparse function activation subgraph, and obtain an inference result; Step 4: Determine whether the optimal process plan corresponding to the current reasoning result meets the production requirements. If so, the process ends; if not, proceed to step 5; Step 5: According to the actual situation of the current production workshop, including materials, tools, processing features, and equipment, dynamically update the current large-scale process knowledge graph. After the update is completed, return to step 1.

2. The process knowledge graph dynamic reasoning method based on sparse activation according to claim 1 is characterized in that: The desired goal in step 1.1 is efficiency optimization, quality optimization, or a balance between efficiency and quality; correspondingly, the multiple relationships are efficiency optimization relationships, quality optimization relationships, or a balance between efficiency and quality relationships.

3. The process knowledge graph dynamic reasoning method based on sparse activation according to claim 2 is characterized in that: The weight threshold in step 2 is 0.7-0.

9.

4. The process knowledge graph dynamic reasoning method based on sparse activation according to claim 3 is characterized by: The method for extracting the sparse function activation subgraph from the multi-relation weight model obtained in step 1 by the sparse activation function in step 2 is: The sparse activation function compares the weight value of each relationship in the multi-relationship weight model with the preset weight threshold, and determines whether to activate the attribute feature value of each relationship based on the comparison result: If the weight value of a relationship is greater than or equal to the weight threshold, the weight value of the relationship is multiplied by 1 as the updated weight value to activate its attribute feature value; If the weight value of a relationship is less than the weight threshold, the weight value of the relationship is multiplied by 0 as the updated weight value to deactivate its attribute feature value; When the weight values ​​of all relations in the multi-relation weight model are updated, the multi-relation weight model with updated weight information is a sparse function activation subgraph.

5. The process knowledge graph dynamic reasoning method based on sparse activation according to claim 4 is characterized in that: In step 3, if the expected goal is efficiency optimization, the reasoning algorithm is an efficiency optimization algorithm; if the expected goal is quality optimization, the reasoning algorithm is a quality optimization algorithm; if the expected goal is to balance efficiency and quality, the reasoning algorithm is an efficiency and quality optimization algorithm.

6. The process knowledge graph dynamic reasoning method based on sparse activation according to claim 5 is characterized by: The efficiency optimization algorithm is Dijkstra's shortest path algorithm; the quality optimization algorithm is PageRank algorithm; the efficiency and quality optimization algorithm is Dijkstra's shortest path algorithm or PageRank algorithm.

7. The process knowledge graph dynamic reasoning method based on sparse activation according to claim 6 is characterized by: The method for judging whether the optimal process scheme corresponding to the current reasoning result meets the production requirements in step 6 is: If the expected goal is efficiency optimization, then combined with the actual processing efficiency of the current production workshop, estimate whether the total time spent on the optimal process plan corresponding to the current reasoning result meets the delivery time required by the customer. If so, it means that the current reasoning result meets the production requirements; otherwise, it means that the current reasoning result does not meet the production requirements; If the desired goal is quality optimization, then combined with the actual processing capacity of the current production workshop, estimate whether the current production workshop can meet the processing accuracy required by the optimal process solution corresponding to the current reasoning result, and judge whether the processing accuracy of the optimal process solution can meet the processing accuracy required by the customer. If all of the above are true, it means that the current reasoning result meets the production requirements; otherwise, it means that the current reasoning result does not meet the production requirements; If the expected goal is to balance efficiency and quality, then the actual processing efficiency of the current production workshop is combined to estimate whether the time spent on the preferred process scheme corresponding to the current reasoning result meets the delivery time required by the customer. At the same time, combined with the actual processing capacity of the current production workshop, it is estimated whether the current production workshop can meet the processing accuracy required by the preferred process scheme, and it is judged whether the processing accuracy of the preferred process scheme can meet the processing accuracy required by the customer. If all of the above are true, it indicates that the current reasoning result meets the production requirements; otherwise, it indicates that the current reasoning result does not meet the production requirements.

8. A process knowledge graph dynamic reasoning system based on sparse activation, comprising a processor and a storage medium, wherein a computer program is stored on the storage medium; characterized in that: When the computer program is executed, the process knowledge graph dynamic reasoning method based on sparse activation described in any one of claims 1 to 7 is executed.

Images