Knowledge engineering-based discrete manufacturing system anomaly prediction method and system regulation and control method

Abstract

The invention discloses a discrete manufacturing system anomaly prediction method and system regulation and control method based on knowledge engineering, and the method comprises the steps: processing the data of a manufacturing system into a triple form, drawing a knowledge graph, and carrying out the representation of entities and relationships in the knowledge graph through a TransR model; training a graph attention network KGAT by using the represented knowledge graph, and outputting node weighted representation by the graph attention network KGAT; and carrying out abnormal event early warning and severity evaluation on the discrete manufacturing system by constructing an early warning model based on node weighted representation. The invention further provides a self-adaptive regulation and control method of the discrete manufacturing system based on the abnormal prediction result, the manufacturing system can dynamically adjust process parameters according to real-time data, and flexible self-adaptive response of the manufacturing process is achieved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent manufacturing technology, and in particular to a discrete manufacturing system abnormality prediction method and a system control method based on knowledge engineering. Background Art

[0002] In modern discrete manufacturing systems, equipment failures and process parameter anomalies seriously affect production efficiency and product quality, so it is crucial to establish an efficient anomaly warning mechanism. Current anomaly monitoring mainly relies on traditional data analysis and empirical judgment methods, which are often difficult to adapt to complex and dynamic manufacturing environments. Especially when faced with diverse production processes and equipment, existing monitoring methods are difficult to accurately identify potential anomalies, resulting in untimely responses, thereby increasing the risk of production downtime and economic losses.

[0003] In recent years, with the development of artificial intelligence technology, especially the application of graph neural networks (GNN) and KGAT, anomaly detection and prediction of manufacturing systems have gradually transformed into intelligence and automation. These advanced deep learning technologies can achieve more accurate anomaly identification and early warning by modeling the complex relationship between equipment and process parameters. However, current research still faces some challenges, such as how to effectively integrate multi-source data to improve the accuracy of the model, and how to achieve efficient early warning while taking into account the severity assessment of abnormal events and optimizing decisions.

[0004] Although existing research has made some progress in the field of abnormal monitoring and early warning, there is still a lack of systematic solutions for discrete manufacturing systems. Existing early warning mechanisms often fail to combine equipment performance and historical abnormal characteristics for comprehensive evaluation, which makes it difficult for companies to formulate reasonable optimization strategies when abnormalities occur, thus affecting the rational allocation of resources and production efficiency. Therefore, developing intelligent abnormality prediction and control methods based on knowledge engineering to improve the responsiveness and resource utilization efficiency of manufacturing systems has become a technical problem that needs to be overcome urgently. Summary of the invention

[0005] In order to address the deficiencies in the prior art, this application proposes a discrete manufacturing system abnormality prediction method and system control method based on knowledge engineering. This method can accurately characterize potential abnormal nodes in the manufacturing process, and enable the manufacturing system to dynamically adjust process parameters according to real-time data, thereby achieving flexible and adaptive response of the manufacturing process.

[0006] The technical solution adopted by the present invention is as follows:

[0007] A discrete manufacturing system abnormality prediction method based on knowledge engineering includes the following steps:

[0008] Step 1: collect data from the manufacturing system, process the data into triples and draw them into a knowledge graph. The triples are in the form of “head entity, relationship, tail entity”.

[0009] Step 2: After the knowledge graph is constructed, the TransR model is used to represent the entities and relationships in the knowledge graph.

[0010] Step 3, use the represented knowledge graph to train the graph attention network KGAT, which is represented by the weighted output nodes of the graph attention network KGAT;

[0011] Step 4: Based on the node weighted representation output by the trained graph attention network (KGAT), an early warning model is constructed to perform abnormal event warning and severity assessment on the discrete manufacturing system.

[0012] Further, the specific process of step 1 is:

[0013] Step 1.1, collect equipment information, process parameters and historical abnormal event data during the system manufacturing process;

[0014] Step 1.2, preprocess the collected data;

[0015] Step 1.3, identify each equipment information, process parameter and abnormal event row as an "entity", and convert it into a triple form according to "head entity, relationship, tail entity", expressed as (h, r, t), where h is the head entity and r is the relationship between entities;

[0016] Step 1.4, convert "entities" as nodes and "relationships" between "entities" as edges into a knowledge graph.

[0017] Further, the specific process of step 2 is:

[0018] Step 2.1: Assign a vector representation to each entity and relationship, represented as entity vector and relationship vector

[0019] Step 2.2, map the entity vector from the entity space to the relation space, expressed as

[0020] Step 2.3, use the scoring function to evaluate the rationality of the triples mapped to the relational space;

[0021] Step 2.4, randomly replace the unreasonable triples formed by entities or relations, and evaluate the rationality of the unreasonable triples;

[0022] Step 2.5, construct the loss function By minimizing the loss function Optimize the embedding vector so that the score of reasonable triples is reduced and the score of unreasonable triples is increased, thereby enhancing the model's ability to represent the knowledge graph.

[0023] Further, the specific process of step 3 is as follows:

[0024] Step 3.1, in the embedding layer, each node in the knowledge graph is parameterized as a vector, and the embedding matrix H is formed by the embedding vectors of all nodes, expressed as:

[0025] H=[h1,h2,…,h N ]

[0026] Where N is the total number of nodes, h i is the embedding vector of the i-th node;

[0027] Step 3.2, in the attention embedding propagation layer, the node representation is updated by recursively propagating the embedding information from the node's neighbor nodes and assigning dynamic weights to each neighbor node using the knowledge-based attention mechanism; each node υ i The update formula is expressed as:

[0028]

[0029] Among them, h' i is the embedding vector of the updated node, σ is the activation function, j is the neighbor node index, and φ(i) is the node υ i The set of neighbor nodes, α ij is the knowledge-based attention weight, representing the node υ j For node v i The importance of; W is the learnable weight matrix;

[0030] Step 3.3, in the prediction layer, all the node representations in the attention embedding propagation layer are aggregated to generate the matching score for prediction, which is expressed as:

[0031] S i =Aggregate({h′ j |j∈φ(i)})

[0032] Among them, S i For node v i The final matching score of , Aggregate is the aggregator;

[0033] Step 3.4, construct the loss function of KGAT During the training process of the graph attention network KGAT, if the loss function begins to converge, the training of the graph attention network KGAT is considered to be completed.

[0034] Furthermore, the loss function of KGAT is specifically expressed as follows:

[0035]

[0036] in, is the loss function of TransR, α is a hyperparameter, and a i,j is the knowledge-based attention weight, representing the node υ j For node v i The importance of ε is the set of node pairs in the knowledge graph, where each pair (i, j) is a certain related node pair in the graph.

[0037] Further, the process of step 4 is:

[0038] Step 4.1, set a dynamic threshold τ, when the weighted representation h' of the node i When the threshold τ is exceeded, the node will be marked as a potential abnormal node;

[0039] Step 4.2: Based on potential abnormal nodes and historical data, an early warning model is constructed using machine learning or deep learning methods.

[0040] An adaptive control method for a discrete manufacturing system comprises the following steps:

[0041] S1, the importance of the abnormal results predicted by the above-mentioned discrete manufacturing system abnormality prediction method based on knowledge engineering is scored. k Calculation of

[0042] S2, based on importance score I k As well as production cost, energy consumption and product quality, an adaptive optimization model is constructed. The objective function of the adaptive optimization model is:

[0043]

[0044] in, is the process parameter vector, is the cost function, is the energy consumption function, is the product quality function, α1, α2, β1 are the weight coefficients corresponding to the cost function, energy consumption function, and product quality function respectively. The model achieves the adaptive optimization goal by minimizing the comprehensive cost and energy consumption while maximizing the quality. K refers to the number of objectives considered in the optimization problem.

[0045] S3, using genetic algorithm and particle swarm algorithm to solve the above adaptive optimization model, using genetic algorithm to perform global search, using excellent individuals as initial particles of particle swarm algorithm; then fine-tuning the particle swarm algorithm to output the optimal solution, which is the optimal solution for the system;

[0046] S4, dynamically adjust the operating parameters of the equipment in the system according to the optimal solution;

[0047] S5, continuously monitors key indicators in the system manufacturing process and uses the feedback information to update the optimization model.

[0048] Furthermore, the importance score of the anomaly is k The calculation formula is:

[0049]

[0050] Among them, F k is the frequency of abnormal occurrence, S k is the impact on the system, P k is the confidence of the prediction, ω1, ω2, and ω3 are the weight coefficients of the frequency of occurrence of anomalies, the degree of impact on the system, and the confidence of the prediction, respectively, which are used to balance the contribution of different indicators to the score; max(F), max(S), and max(C) represent the maximum values ​​of the frequency of occurrence of anomalies, the degree of impact on the system, and the confidence of the prediction, respectively.

[0051] Further, the process of S3 is:

[0052] S3.1, genetic algorithm performs global search through population initialization, fitness evaluation, selection, crossover and mutation, and according to the optimal individual g best Set the initial velocity:

[0053]

[0054] Where r is the constant for velocity update, g best It refers to the global optimal position of the current particle swarm, which guides the moving direction of all particles through the optimal position. is the initial position of particle i;

[0055] S3.2, use the particle swarm algorithm to fine-tune, update the particle speed and position, and introduce the importance score I k , the optimal solution finally outputted It is expressed as:

[0056]

[0057] in, represents the velocity vector of the ith particle in the t+1 generation, ω is the inertia weight, is the historical optimal position of the particle, g best is the global optimal position, c1 and c2 are acceleration factors, r1 and r2 are random numbers in the range of [0,1]. Represents the current position vector of the i-th particle in the t+1th generation.

[0058] Further, in S5, the optimization model is updated based on the feedback information and expressed as:

[0059]

[0060] The manufacturing system is adjusted in real time based on the above formula.

[0061] Beneficial effects of the present invention:

[0062] (1) By constructing a multi-level knowledge graph, the equipment, process parameters and historical abnormal data in the manufacturing system are structured and represented, and combined with the graph attention network (KGAT), the present invention can accurately characterize the potential abnormal nodes and associated process problems in the manufacturing process. The system can achieve high-precision abnormal warning, not only to identify potential faults in a timely manner, but also to quickly locate the root cause of the problem, shorten the response time of problem handling, and greatly reduce the risk of production interruption and failure.

[0063] (2) The present invention not only provides efficient anomaly prediction, but also builds an adaptive optimization model by combining equipment performance prediction data and anomaly characteristics. The model uses genetic algorithms and particle swarm algorithms to intelligently adjust key parameters that affect efficiency and quality in the production process. Adaptive optimization enables the manufacturing system to dynamically adjust process parameters based on real-time data, achieve flexible and adaptive response of the manufacturing process, greatly improve production efficiency, and effectively reduce the need for human intervention.

[0064] (3) The present invention takes into account factors such as production cost and energy consumption by constructing an optimization model. The application of genetic algorithm and particle swarm algorithm can not only realize the dynamic optimization of production parameters, but also maximize the savings of production resources and energy consumption while maintaining stable product quality, help the manufacturing system reduce the overall operating cost, improve the effect of energy conservation and emission reduction, and meet the needs of modern green manufacturing.

[0065] (4) The knowledge graph and graph attention network structure provided by the present invention make the system highly scalable and can easily expand new equipment and new processes according to changes in the manufacturing system. The system supports dynamic updating of the abnormal event library and process parameters in the manufacturing process, can continuously adapt to the evolution of the manufacturing system, and provides long-term sustainable technical support for the intelligent upgrade of discrete manufacturing systems.

[0066] (5) The optimization model proposed in this invention realizes multi-objective solution based on genetic algorithm and particle swarm optimization algorithm, which can comprehensively consider multiple objectives such as production efficiency, cost control, quality assurance and energy consumption. This intelligent decision support mechanism enables the manufacturing system to independently formulate the optimal operation plan when facing complex production tasks, ensuring the global optimal performance under multiple objectives. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 It is the overall technical flow chart of this application method.

[0068] Figure 2 It is a knowledge graph of the anomaly library of discrete manufacturing systems drawn by this application.

[0069] Figure 3 It is the existing TransR representation knowledge graph process.

[0070] Figure 4 This is the knowledge graph training and abnormal warning process including the abnormal knowledge base of this application.

[0071] Figure 5 This is a schematic diagram of the active optimization process of this application. DETAILED DESCRIPTION

[0072] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0073] Example 1

[0074] like Figures 1 to 3 , a discrete manufacturing system abnormality prediction method based on knowledge engineering is shown, comprising the following steps:

[0075] Step 1: Collect data from the manufacturing system, process the data into triples and draw a knowledge graph. This includes the following specific steps:

[0076] Step 1.1, collect data during the system manufacturing process, including various equipment information (such as equipment model, function, status, etc.), process parameters (such as temperature, pressure, speed, etc.) and historical abnormal event data (such as fault type, occurrence time, impact degree, etc.). The above data can be obtained through sensors, production records and maintenance logs.

[0077] Step 1.2: Clean the collected data, including removing duplicate values, filling missing values, and correcting erroneous data to ensure data accuracy and consistency. Next, convert the data into a unified format, such as standardized numerical ranges and category labels, for subsequent processing.

[0078] Step 1.3, identify each equipment information, process parameter and abnormal event row as an "entity", and convert it into a triple form according to "head entity, relationship, tail entity" to form triple data. The triple is represented by (h, r, t), where h is the head entity and r is the relationship between entities. More specifically, the triple form is such as: "Equipment A uses process parameter B" or "Equipment C has abnormal event D", and the triple form can facilitate the identification of the relationship between them. In this embodiment, the structured information described by the triple can reflect the deep relationship between various equipment, process parameters and events in the production process, and provide complete background data for abnormality detection.

[0079] Step 1.4, use graph database technology (python + Neo4j) to structure the generated triple data into a knowledge graph; in the knowledge graph, "entities" (equipment, process parameters, abnormal events, etc.) are used as nodes, and "relationships" between "entities" are used as edges to achieve information visualization and easy query. In this embodiment, Figure 2 The construction of the knowledge graph shown not only associates all devices within the system with abnormal events, but also provides a basis for subsequent graph representation and graph embedding.

[0080] Step 2: After the knowledge graph is constructed, the TransR model is used to characterize the entities and relationships in the knowledge graph. The TransR model assigns an independent embedding space to each entity and relationship in the knowledge graph, and maps the head entity and the tail entity to the relationship space to achieve accurate characterization. The following specific steps are included:

[0081] Step 2.1, before starting training, you need to initialize the embedding vectors for each entity and each relationship in the knowledge graph. These embedding vectors are usually generated under random initial conditions and are continuously optimized during the training process. More specifically, the TransR model assigns a vector representation to each entity and relationship, that is, the embedding vector representation of each entity h and t obtained by the TransR model is represented as an entity vector and The TransR model is used to obtain the embedding vector of relation r, which is represented as the relation vector Entity vectors are defined in entity space, and relationship vectors are defined in relationship space, so that each entity and relationship has its corresponding high-dimensional space representation.

[0082] Step 2.2, due to the different spaces between entity vectors and relation vectors, the TransR model needs to map entity vectors from entity space to relation space. Specifically, a mapping matrix M is used r The head entity vector Tail entity vector are mapped to relation vectors The space is expressed as:

[0083]

[0084] in, is the head entity mapped to the relational space, is the tail entity direction mapped to the relational space. The mapping matrix M r It is designed independently according to the properties of relation r.

[0085] Step 2.3: In each triple, after the head entity h and tail entity t corresponding to the relation r are mapped to the relation space, TransR uses a scoring function to evaluate the rationality of the triple. The scoring function is expressed as:

[0086]

[0087] The scoring function calculates the mapped head entity vector h r With the tail entity t r The smaller the score value of the distance in the relational space, the more reasonable the triple (h, r, t) is, indicating that the relationship between entities conforms to the rules of the knowledge graph.

[0088] Step 2.4, during the training process, TransR not only scores positive samples (i.e., real triplets), but also scores negative samples (i.e., unreasonable triplets formed by randomly replacing entities or relations). The negative samples in this implementation are generated by randomly replacing the head entity h or the tail entity t to generate negative triplets (h', r, t) or (h, r, t'); the role of negative sampling technology is to enable the model to recognize not only reasonable relations, but also unreasonable relations, and distinguish between the two through training.

[0089] Step 2.5, TransR uses a loss function based on the scoring function By optimizing the training, the model optimizes the embedding vector by minimizing the loss function, so that the score of reasonable triples is reduced and the score of unreasonable triples is increased, thereby enhancing the model's ability to represent the knowledge graph. The specific form of the loss function is:

[0090]

[0091] Among them, γ is a safety margin parameter that ensures that the score of positive samples is lower than that of negative samples. S is the positive sample set, (h, r, t) is the real triple in the positive sample set; S' is the negative sample set, (h', r, t') is the unreasonable triple in the negative sample set.

[0092] like Figure 3As shown in the figure, in each iteration, the TransR model adjusts the embedding vectors of entities and relations according to the gradient of the loss function through the back-propagation algorithm. The model will continuously adjust the embedding so that the energy of positive triples (i.e., the scoring function value) is as small as possible, while the energy of negative triples is as large as possible. As the training continues, the embedding vector gradually stabilizes, enabling the model to better represent the complex structure between entities and relations in the knowledge graph.

[0093] Step 3: Use the represented knowledge graph to train the KGAT, which is represented by the weighted output nodes of the KGAT. The process of training the KGAT includes the following steps:

[0094] Step 3.1, in the embedding layer, each node in the knowledge graph is parameterized as a vector to maintain the integrity of the graph structure. i is initialized to a random vector h i Through learning, the vector representation of nodes is continuously optimized. The embedding matrix H is composed of the embedding vectors of all nodes and is expressed as:

[0095] H=[h1,h2,…,h N ]

[0096] Here, N is the total number of nodes, h i is the embedding vector of the i-th node. The embedding of each entity node represents its semantic and structural features in the graph and ensures that the overall structure of the graph can be maintained in subsequent steps.

[0097] Step 3.2, in the attention embedding propagation layer, the representation of the node is updated by recursively propagating the embedding information from the node’s neighbor nodes and assigning dynamic weights to each neighbor node using a knowledge-based attention mechanism. For each node υ i , and its update formula is expressed as:

[0098]

[0099] Among them, h' i is the embedding vector of the updated node, σ is the activation function (Sigmoid), j is the neighbor node index, and φ(i) is the node υ i The set of neighbor nodes, α ij is the knowledge-based attention weight, representing the node υ j For node v i The importance of the neighbor nodes; W is a learnable weight matrix. In this way, the node can obtain the feature information of the neighbor nodes and perform weighted updates according to the importance of the neighbor nodes in the knowledge graph.

[0100] Step 3.3, in the prediction layer, all the node representations in the attention embedding propagation layer are aggregated to generate a matching score for prediction, which is used to implement anomaly detection or other related tasks. The aggregation operation can be implemented by summing or averaging, and the formula is as follows:

[0101] S i =Aggregate({h′ j |j∈φ(i)})

[0102] Among them, S i For node v i Aggregate is the aggregator; it aggregates the node representations of all abnormal events and abnormal causes, and outputs the matching scores of the predicted abnormal events and abnormal causes.

[0103] Step 3.4, construct the loss function of KGAT During the training process of the KGAT, if the loss function begins to converge, the training of the KGAT is considered to be completed. The loss function of KGAT is specifically expressed as follows:

[0104]

[0105] in, is the loss function of TransR, α is a hyperparameter, and a i,j is the knowledge-based attention weight, representing the node υ j For node v i The importance of ε is the set of node pairs in the knowledge graph, where each pair (i, j) is a node pair with some relationship in the graph; through this loss function It can improve the accuracy of the model in anomaly detection tasks.

[0106] Step 4: Based on the node weighted representation output by the trained KGAT, abnormal event warning and severity assessment are performed on the discrete manufacturing system. The specific steps are as follows:

[0107] Step 4.1, identify potential abnormal nodes based on the node weighted representation output by the KGAT. The weighted representation of each node reflects its importance in the knowledge graph. By setting a dynamic threshold τ, potential abnormal nodes can be identified. For example, when the weighted representation h' i When the threshold τ is exceeded, the node will be marked as a potential abnormal node, which helps to detect abnormal situations in the manufacturing system in time and take corresponding measures.

[0108] Step 4.2, build an early warning model based on potential abnormal nodes and historical data.

[0109] Step 4.2.1, after identifying the potential abnormal nodes, further analysis is required on the potential abnormal nodes to understand the background of the abnormal events, and to extract the historical abnormal characteristics related to the potential abnormal nodes (such as event frequency and severity) and the performance indicators of the equipment before and after the abnormality occurs (such as temperature, load, etc.), so as to form a multi-dimensional analysis framework and provide rich contextual information for the subsequent abnormal warning model construction.

[0110] Step 4.2.2, construct an abnormal warning model, and use the weighted representation of the node, historical abnormal characteristics and performance indicators before and after the abnormality occurs to train the abnormal warning model. In this embodiment, machine learning or deep learning methods, such as random forest, support vector machine or neural network, are used to build the model network. The input data of the warning model are the weighted representation of the node and the system equipment performance prediction data. The warning process is to predict the abnormalities that may occur in the system in the future by analyzing the weighted representation of the node and combining the equipment performance prediction data, and give the matching score of the abnormal cause corresponding to the abnormality, where the matching score S is given at the prediction layer.

[0111] Example 2

[0112] In view of the abnormal events identified by the above method, the present invention also proposes an adaptive control method for discrete manufacturing systems, the purpose of which is to continuously improve the system in response to abnormal events occurring in the system, so as to improve the accuracy and timeliness of abnormal warnings, thereby enhancing the stability and reliability of the manufacturing system. Figure 4 , 5 .

[0113] S1, according to the abnormal prediction results, each abnormality is scored for importance I k The score takes into account the frequency of abnormal occurrence F k , Impact on the system S k And the confidence level of the prediction P k The scoring formula is:

[0114]

[0115] Among them, ω1, ω2, and ω3 are weight coefficients for the frequency of anomaly occurrence, the degree of impact on the system, and the confidence of the prediction, which are used to balance the contribution of different indicators to the score. By calculating the importance score, high-importance anomalies can be solved first during the optimization process to improve the solution efficiency.

[0116] S2, builds an adaptive optimization model that can not only minimize the cost and energy consumption in the production process, but also ensure product quality, while also considering the importance score of the anomaly I k The weight influence of the model; the objective function of the model is:

[0117]

[0118] in, is the process parameter vector, is the cost function, is the energy consumption function, is the product quality function, α1, α2, β1 are the weight coefficients corresponding to the cost function, energy consumption function, and product quality function respectively. The model achieves the adaptive optimization goal by minimizing the comprehensive cost and energy consumption while maximizing the quality. K refers to the number of objectives considered in the optimization problem.

[0119] S3, using a hybrid optimization algorithm combining genetic algorithm and particle swarm algorithm to solve the above optimization model. In order to efficiently solve the above optimization model, the present invention designs a hybrid optimization algorithm combining GA and PSO. This combination method utilizes the global search advantage of GA, and then improves the local convergence accuracy through fine-tuning of PSO, thereby realizing the combination of global and local optimization.

[0120] First, GA performs global search through population initialization, fitness evaluation, selection, crossover and mutation. The excellent individuals evolved by GA are used as the initial particles of PSO, and the optimal individuals g are selected according to their evolution. best Set the initial velocity:

[0121]

[0122] Where r is a constant for speed update (usually a scalar) or a random factor. best It refers to the global optimal position of the current particle swarm, that is, the best solution for all particles in the entire particle swarm. The particle swarm guides the movement direction of all particles through this optimal position. is the initial position of particle i, which is the position when the particle starts searching, and represents the coordinates of the particle in the search space at the beginning of the current iteration.

[0123] In the PSO fine-tuning stage, the initial solution provided by GA is further optimized by updating the particle speed and position. The specific formula is:

[0124]

[0125] By introducing the importance score I k ,The PSO optimization process can focus more on the processing of important anomalies and improve the convergence speed and accuracy.

[0126] The optimal solution of the final output for:

[0127]

[0128] in, represents the velocity vector of the ith particle in the t+1 generation, ω is the inertia weight, is the historical optimal position of the particle, g best is the global optimal position. Both values ​​are obtained through the objective function of the model. Calculated, c1, c2 are acceleration factors, r1, r2 are random numbers in the range [0,1], Represents the current position vector of the i-th particle in the t+1th generation.

[0129] S4, once the optimal solutions are obtained, the next step is to apply these optimal solutions to the manufacturing system. Specifically, the operating parameters are dynamically adjusted through intelligent devices to improve production efficiency. These operating parameters may include the operating speed of the machine, the feed rate of the material, etc. The set parameter update formula can be expressed as:

[0130] p new =p old +Δ p

[0131] Among them, p new is the new operating parameter, p old is the old operating parameter, Δ p It is the adjustment value calculated based on the optimization results.

[0132] S5. Finally, implement monitoring and feedback mechanisms to ensure the effectiveness of the system in a dynamic environment. Continuously monitor key indicators in the manufacturing process (such as production efficiency, cost, and product quality), and use feedback information to update the optimization model. This feedback can be expressed by an update formula:

[0133]

[0134] In this way, manufacturing systems can adjust response strategies in real time, achieve adaptive response and improve efficiency, ultimately reducing the risk of failure and increasing productivity.

[0135] Ultimately, by solving the optimal parameters, the system can dynamically adjust the equipment operating parameters, thereby improving overall production efficiency and reducing the probability of abnormal events and equipment failures. Figure 5 The whole optimization process is demonstrated, from key factor identification to parameter adjustment in detail.

[0136] The above embodiments are only used to illustrate the design ideas and features of the present invention, and their purpose is to enable those skilled in the art to understand the content of the present invention and implement it accordingly. The protection scope of the present invention is not limited to the above embodiments. Therefore, any equivalent changes or modifications made based on the principles and design ideas disclosed by the present invention are within the protection scope of the present invention.

Claims

1. A method for predicting anomalies in discrete manufacturing systems based on knowledge engineering, characterized in that: The steps include: Step 1, collect the data of the manufacturing system, process the data into triples and draw them into a knowledge graph, where the triples are in the form of "head entity, relationship, tail entity"; Step 2: After the knowledge graph is constructed, the TransR model is used to represent the entities and relationships in the knowledge graph. Step 3: Use the represented knowledge graph to train the graph attention network KGAT, which is represented by the weighted output nodes of the graph attention network KGAT; Step 4: Based on the node weighted representation output by the trained graph attention network KGAT, an early warning model is constructed to perform abnormal event warning and severity assessment on the discrete manufacturing system.

2. The method for predicting anomalies of discrete manufacturing systems based on knowledge engineering according to claim 1 is characterized in that: The specific process of step 1 is: Step 1.1, collect equipment information, process parameters and historical abnormal event data during the system manufacturing process; Step 1.2, preprocess the collected data; Step 1.3, identify each equipment information, process parameter and abnormal event row as an "entity", and convert it into a triple form according to "head entity, relationship, tail entity", expressed as (h, r, t), where h is the head entity and r is the relationship between entities; Step 1.4, convert "entities" as nodes and "relationships" between "entities" as edges into a knowledge graph.

3. The method for predicting anomalies of discrete manufacturing systems based on knowledge engineering according to claim 1 is characterized in that: The specific process of step 2 is: Step 2.1, assign a vector representation to each entity and relationship; Step 2.2, map the entity vector from the entity space to the relation space; Step 2.3, use the scoring function to evaluate the rationality of the triples mapped to the relational space; Step 2.4, randomly replace the unreasonable triples formed by entities or relations, and evaluate the rationality of the unreasonable triples; Step 2.5, construct the loss function By minimizing the loss function Optimize the embedding vector so that the score of reasonable triples is reduced and the score of unreasonable triples is increased, thereby enhancing the model's ability to represent the knowledge graph.

4. The method for predicting anomalies of discrete manufacturing systems based on knowledge engineering according to claim 1, characterized in that: The specific process of step 3 is as follows: Step 3.1, in the embedding layer, each node in the knowledge graph is parameterized as a vector, and the embedding matrix H is formed by the embedding vectors of all nodes, expressed as: H=[h1,h2,…,h N ] Where N is the total number of nodes, h i is the embedding vector of the i-th node; Step 3.2, in the attention embedding propagation layer, the node representation is updated by recursively propagating the embedding information from the node's neighbor nodes and assigning dynamic weights to each neighbor node using the knowledge-based attention mechanism; each node υ i The update formula is expressed as: Among them, h' i is the embedding vector of the updated node, σ is the activation function, j is the neighbor node index, and φ(i) is the node v i The set of neighbor nodes, α ij is the knowledge-based attention weight, representing the node v j For node v i The importance of; W is the learnable weight matrix; Step 3.3, in the prediction layer, all the node representations in the attention embedding propagation layer are aggregated to generate the matching score for prediction, which is expressed as: S i =Aggregate({h′ j |j∈φ(i)}) Among them, S i For node v i The final matching score of , Aggregate is the aggregator; Step 3.4, construct the loss function of KGAT During the training process of the graph attention network KGAT, if the loss function begins to converge, the training of the graph attention network KGAT is considered to be completed.

5. The method for predicting anomalies of discrete manufacturing systems based on knowledge engineering according to claim 4 is characterized in that: The loss function of KGAT is specifically expressed as follows: in, is the loss function of TransR, α is a hyperparameter, and a i,j is the knowledge-based attention weight, representing the node υ j For node v i The importance of ε is the set of node pairs in the knowledge graph, where each pair (i, j) is a certain related node pair in the graph.

6. The method for predicting anomalies of discrete manufacturing systems based on knowledge engineering according to claim 4 is characterized in that: The process of step 4 is: Step 4.1, set a dynamic threshold τ, when the weighted representation h' of the node i When the threshold τ is exceeded, the node will be marked as a potential abnormal node; Step 4.2: Based on potential abnormal nodes and historical data, an early warning model is constructed using machine learning or deep learning methods.

7. An adaptive control method for a discrete manufacturing system, characterized in that: The steps include: S1, scoring the importance of the abnormal results predicted by the method for predicting anomalies in a discrete manufacturing system based on knowledge engineering in claim 1 k Calculation of S2, based on importance score I k As well as production cost, energy consumption and product quality, an adaptive optimization model is constructed. The objective function of the adaptive optimization model is: in, is the process parameter vector, is the cost function, is the energy consumption function, is the product quality function, α1, α2, β1 are the weight coefficients corresponding to the cost function, energy consumption function, and product quality function respectively. The model achieves the adaptive optimization goal by minimizing the comprehensive cost and energy consumption while maximizing the quality. K refers to the number of objectives considered in the optimization problem. S3, using genetic algorithm and particle swarm algorithm to solve the above adaptive optimization model, using genetic algorithm to perform global search, using excellent individuals as initial particles of particle swarm algorithm; then fine-tuning the particle swarm algorithm to output the optimal solution, which is the optimal solution for the system; S4, dynamically adjust the operating parameters of the equipment in the system according to the optimal solution; S5, continuously monitors key indicators in the system manufacturing process and uses the feedback information to update the optimization model.

8. The adaptive control method of a discrete manufacturing system according to claim 7, characterized in that: Anomaly Importance Rating I k The calculation formula is: Among them, F k is the frequency of abnormal occurrence, S k is the impact on the system, P k is the confidence of the prediction, ω1, ω2, and ω3 are the weight coefficients of the frequency of occurrence of anomalies, the degree of impact on the system, and the confidence of the prediction, respectively, which are used to balance the contribution of different indicators to the score; max(F), max(S), and max(c) represent the maximum values ​​of the frequency of occurrence of anomalies, the degree of impact on the system, and the confidence of the prediction, respectively.

9. The adaptive control method of a discrete manufacturing system according to claim 7, characterized in that: The process of S3 is: S3.1, genetic algorithm performs global search through population initialization, fitness evaluation, selection, crossover and mutation, and according to the optimal individual g best Set the initial velocity: Where r is the constant for velocity update, g best It refers to the global optimal position of the current particle swarm, which guides the moving direction of all particles through the optimal position. is the initial position of particle i; S3.2, use the particle swarm algorithm to fine-tune, update the particle speed and position, and introduce the importance score Ik to finally output the optimal solution It is expressed as: in, represents the velocity vector of the ith particle in the t+1 generation, ω is the inertia weight, is the historical optimal position of the particle, g best is the global optimal position, c1 and c2 are acceleration factors, r1 and r2 are random numbers in the range of [0,1]. Represents the current position vector of the i-th particle in the t+1th generation.

10. The adaptive control method of a discrete manufacturing system according to claim 8, characterized in that: In S5, the optimization model is updated based on the feedback information as follows: The manufacturing system is adjusted in real time based on the above formula.

Images