Petri-Net-GCN-based intelligent manufacturing system deadlock prediction method

Abstract

The application discloses a Petri-Net-GCN-based intelligent manufacturing system deadlock prediction method, which comprises the following steps: dividing an intelligent manufacturing system into a production equipment module, a production process module and a production resource module; performing formal modeling on the modules based on a Petri net; constructing a Petri-Net-GCN deadlock prediction model, which comprises a place-to-transition message transmission module and a transition-to-place message transmission module; bidirectionally aggregating and updating features between nodes in a Petri net graph by using a graph neural network structure; collecting initial token distribution, a correlation matrix and a deadlock label of the intelligent manufacturing system to construct a data set and input the data set into model training; and predicting a deadlock risk in an unknown system state by using the trained model; and the application can rapidly predict and judge the current state of the intelligent manufacturing system, thereby providing prior warning information for production scheduling and resource allocation.

Description

Technical Field

[0001] This invention relates to the field of deadlock detection technology, and specifically to a deadlock prediction method for intelligent manufacturing systems based on Petri-Net-GCN. Background Technology

[0002] In recent years, with the increasing complexity of Intelligent Manufacturing Systems (IMS) and some discrete event systems, deadlock has become a key challenge in system design and operation. For IMS, deadlock typically arises from several main causes: limited system resources, uneven resource allocation, and resources being used in erroneous processes. Accurately describing the actual operation of an IMS is very difficult. To solve the deadlock problem in intelligent manufacturing systems, it is essential to model the entire system to simulate its operation. From a modeling perspective, Petri net data information is analyzed in terms of static structure and dynamic behavior. Petri nets, as a mathematical modeling tool for describing and analyzing discrete event dynamic systems, are very suitable for solving the above problems. However, traditional Petri nets are often limited in manufacturing system modeling due to state space explosion and insufficient dynamic adaptability, making it difficult to effectively address the deadlock prediction requirements of complex intelligent manufacturing systems (IMS). To address this issue, Graph Convolutional Neural Networks (GCNs), as a deep learning model specifically designed for graph-structured data, excel at capturing relationships between nodes and efficiently processing graph data analysis tasks. Petri nets are essentially graph structures with clear relationships between places and transitions, making them suitable for analysis using GCNs. Therefore, this study proposes a deadlock prediction method for intelligent manufacturing systems based on Petri-Net-GCN from the perspective of prediction in deadlock detection and recovery, aiming to overcome the limitations of traditional methods. Summary of the Invention

[0003] Purpose of the invention: The purpose of this invention is to provide a deadlock prediction method for intelligent manufacturing systems based on Petri-Net-GCN, which solves the problems of state space explosion and dynamic adaptability limitations in traditional Petri net modeling of manufacturing systems.

[0004] Technical solution: The present invention provides a deadlock prediction method for intelligent manufacturing systems based on Petri-Net-GCN, comprising the following steps:

[0005] (1) Divide the intelligent manufacturing system into a production equipment module, a production process module, and a production resource module;

[0006] (2) Formal modeling of the modules based on Petri nets includes: mapping the production equipment module to the equipment state represented by the place in the Petri net; mapping the production process module to the task scheduling or process transformation represented by the transition in the Petri net; mapping the production resource module to the resource distribution represented by the token in the Petri net; constructing the Petri net structure through directed arc connections and representing it with an incidence matrix, where positive values ​​represent the output arc from the transition to the place, negative values ​​represent the input arc from the place to the transition, and zero values ​​represent no connection;

[0007] (3) Construct a Petri-Net-GCN deadlock prediction model including a place-to-change message passing module and a change-to-place message passing module; use a graph neural network structure to perform bidirectional aggregation and update of features between nodes in the Petri network graph;

[0008] (4) Collect the initial Token distribution, correlation matrix and deadlock labels of the intelligent manufacturing system to build a dataset, input it into the model for training; use the trained model to predict the deadlock risk in unknown system states.

[0009] Furthermore, in step (1), the production equipment module includes: processing equipment: including the status (idle, processing, faulty, under maintenance) and quantity of machine tools and robotic arms; transportation equipment: including the status (idle, transporting, charging, faulty) and quantity of AGVs; warehousing equipment: including the status (idle, storing, picking, faulty) and quantity of input / output devices; production process module includes: loading and unloading module status (pending loading, loading, pending unloading, unloading, abnormal) and product quantity; product flow module status (pending flow, flow, arrived, abnormal) and product quantity; production resource module includes raw material module status (available, allocated, exhausted) and inventory; intermediate product module status (pending processing, processing, completed) and quantity; finished product module status (produced, awaiting shipment, unqualified) and inventory.

[0010] Furthermore, in step (2), the place represents the device status node, the transition represents the task scheduling operation, and the token represents the resource distribution; the arc connection direction is consistent with the actual flow direction of the resources.

[0011] Furthermore, in step (3), the Petri-Net-GCN model also includes: a feature extraction module for extracting node features from the association matrix and initial labels; an attention layer for calculating node attention weights to enhance the ability to express key states; and a classification layer for outputting deadlock binary classification results based on graph-level embedding representation.

[0012] Furthermore, the message passing module is specifically as follows: In the location-to-transition module, location features... Through the correlation matrix A p_t Weighted update of transition feature X tThe transition to the storage module is achieved through the transpose of the association matrix (A). p_t ) T The transition features are backpropagated to the places; a ReLU activation function is applied after each message-passing layer to enhance nonlinear expressiveness.

[0013] Furthermore, the message passing functions from the library to the transition module are as follows:

[0014] X t =X p ·A p_t ·W t +B t

[0015] The message passing function for transitioning to the storage module is shown below:

[0016] X p =X t ·(A p_t ) T ·W p +B p

[0017] Among them, X p This represents the feature information of all places in the Petri net, and X t This represents the characteristic information of all transitions in a Petri net, and A p_t Let A represent the correlation matrix of a Petri net, and A p_t ∈Z |P|×|T| W t and W p There are two trainable weight matrices, and W t ∈R |T| W p ∈R |P| B t and B p Denotes the bias matrix, and

[0018] Furthermore, in step (4), the dataset includes the initial labels, deadlock labels, and association matrices of the Petri net under different states.

[0019] Furthermore, in step (4), the model training uses accuracy, precision, recall, and F1 score as evaluation metrics; during prediction, the initial identifier vector and correlation matrix of the unknown state are input, and the deadlock probability is output.

[0020] An electronic device according to the present invention includes a memory, a processor, and a computer program stored in the memory, wherein the processor executes the program to implement the steps of any of the methods described herein.

[0021] An electronic device according to the present invention includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the program to implement the steps of any of the methods described herein.

[0022] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: (1) The present invention uses Petri nets to formally model the intelligent manufacturing system and combines graph convolutional neural networks (GCNs) to design a structured message passing mechanism, thereby realizing accurate prediction of deadlock risk in the system state. Compared with the traditional deadlock detection method that relies on reachable graph enumeration or model checking, the present invention can effectively alleviate the state space explosion problem and improve the adaptability and scalability in complex manufacturing systems. (2) The present invention proposes a feature extraction and graph neural network training mechanism oriented towards Petri net structure. By introducing a bidirectional message passing module from place to transition and from transition to place, the dynamic relationship modeling capability between Petri net nodes is strengthened, and the model's accuracy in identifying deadlock states is improved. (3) The Petri-Net-GCN model constructed by the present invention can quickly predict and judge the current state during the manufacturing system operation phase, providing prior warning information for production scheduling and resource allocation. Attached Figure Description

[0023] Figure 1 This is a flowchart of the deadlock prediction method for the intelligent manufacturing system of the present invention;

[0024] Figure 2 This is a schematic diagram of the Petri net-based intelligent manufacturing system model of the present invention;

[0025] Figure 3 This is a diagram illustrating the library-to-transition module of the present invention;

[0026] Figure 4 This is a diagram illustrating the transition to the storage module of the present invention;

[0027] Figure 5 This is an illustration of an example of the Petri-Net-GCN structure of the present invention. Detailed Implementation

[0028] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0029] like Figure 1 As shown, this embodiment of the invention provides a Petri net-based intelligent manufacturing system (IMS) model. The method for deadlock prediction using this IMS includes the following steps:

[0030] S1. Divide the IMS into three modules: production equipment, production processes, and production resources. The production equipment module includes a processing equipment module, a transportation module, and a warehousing module. The processing equipment module includes machine tools and robotic arms with statuses such as idle, processing, faulty, and under maintenance, as well as the number of processing equipment and robotic arms. The transportation module includes automated guided vehicles (AGVs) with statuses such as idle, transporting, charging, and faulty, as well as the number of AGVs. The warehousing module includes statuses such as idle, storing, retrieving, and faulty, as well as the number of input / output devices. The production process module includes a loading / unloading module and a product flow module. The loading / unloading module includes statuses such as pending loading, loading, pending unloading, unloading, and abnormal, as well as the number of products loaded / unloaded. The product flow module includes statuses such as pending flow, flow, arrival, and abnormal, as well as the number of flowed products. The production resources module includes a raw material module, an intermediate product module, and a finished product module. The raw material module includes statuses such as available, allocated, and depleted, as well as the inventory of raw materials. The intermediate product module includes statuses such as pending processing, processing, and completed, as well as the quantity of intermediate products at each process stage. The finished product module includes statuses such as produced, awaiting shipment, and non-conforming, as well as the inventory of finished products.

[0031] S2. Map the different devices in IMS to the corresponding elements (places and transitions) of the Petri net.

[0032] like Figure 2 As shown, warehouses {p1, p8} represent input and output devices in IMS, respectively, with black dots indicating that the device contains product resources; warehouses {p2, p4, p5, p7} all represent AGVs in IMS; warehouses {p3, p6} represent processing equipment in IMS that processes products, with black dots indicating that the equipment is currently idle and available; warehouses {p9, p... 10 p 11 The symbol {t1, t2, t3, t4, t5, t6, t7, t8} represents a robotic arm device in IMS, with black dots indicating that the device is currently idle and available. Transitions {t1, t2, t3, t4, t5, t6, t7, t8} represent task scheduling or process transition operations in IMS, with the arc connection direction consistent with the resource flow direction. Each transition is triggered only if the preceding storage location contains the resource; only then can the resource successfully enter the next storage location.

[0033] S3. The design includes a module for passing messages from a location to a change and a module for passing messages from a change to a location.

[0034] In Petri nets, the relationship between transitions and places is represented by directed arcs that determine network evolution. To extract as much feature information as possible from the Petri net topology and effectively capture the connections between transitions and their input / output places, these relationships are categorized into two types based on the arc type: arcs from places to transitions, and arcs from transitions to places. By representing each type of relationship as a convolutional kernel, the Petri net topology can be inherited by the neural network. Therefore, two modules, Place-to-Transition (P2T) and Transition-to-Place (T2P), were designed.

[0035] like Figure 3 As shown, in the module from place to transition, given a directed arc from place p to transition t, circles represent place nodes, rectangles represent transition nodes, and shaded small circles represent Petri-Net-GCN neurons. This is a two-layer fully connected neural network. The neurons in the first layer represent the place features of the Petri net, denoted by X. p To represent the transition features of the Petri net, the neurons in the second layer are represented by X. t This is used to represent the message passing function from the library to the transition module.

[0036] X t =X p ·A p_t ·W t +B t

[0037] Among them, X p This represents the feature information of all places in the Petri net, and λ p X represents the total number of initial identifiers at the Petri net input. t This represents the characteristic information of all transitions in a Petri net, and A p_t Let A represent the correlation matrix of a Petri net, and A p_t ∈Z |P|×|T| W t It is a trainable weight matrix, and W t ∈R |T| B t Denotes the bias matrix, and

[0038] like Figure 4 As shown, in the module for transitions to places, given a directed arc from transition t to place p, circles represent place nodes, rectangles represent transition nodes, and shaded small circles represent Petri-Net-GCN neurons. This is a two-layer fully connected neural network. The neurons in the first layer represent the transition features of the Petri net, denoted by X. tTo represent the second layer of neurons, the place features of the Petri net are represented by X. p This is used to represent the transition to the storage module. The message passing function is as follows:

[0039] X p =X t ·(A p_t ) T ·W p +B p

[0040] Among them, X p This represents the feature information of all places in the Petri net, and X t This represents the characteristic information of all transitions in a Petri net, and A p_t Let A represent the correlation matrix of a Petri net, and A p_t ∈Z |P|×|T| W p It is a trainable weight matrix, and W p ∈R |P| B p Denotes the bias matrix, and

[0041] S4. Design a simple Petri-Net-GCN structure model.

[0042] like Figure 5 As shown, a Petri-Net-GCN structural model was designed for a simple Petri net based on the steps described above. This neural network has five layers, including three place layers and two transition layers, consisting of two P2T modules and two T2P modules. To enhance the nonlinearity of Petri-Net-GCN, each module uses a ReLU activation function. Through the P2T and T2P modules, Petri-Net-GCN inherits the internal structure of the Petri net. Therefore, Petri-Net-GCN aggregates and transmits state features according to the structure of the Petri net. Since the behavior of the Petri net is determined by its internal structure, Petri-Net-GCN possesses dynamic behavioral knowledge through its topological structure.

[0043] S5. Construct a complete Petri-Net-GCN deadlock prediction model.

[0044] The Petri-Net-GCN model mainly consists of the following parts: 1) Feature extraction module: Extracts structural features of each transition from the association matrix and place identifiers, constructing a structure vector for each transition; 2) Dataset construction module: Loads data using CSV files, where place_features.csv represents the place features of the Petri net, indicating the number of tokens for each place; adj_matrix.csv represents the association matrix of the Petri net, indicating the P×T connection structure of the Petri net; label.csv represents the deadlock labels of the Petri net; and a heterogeneous graph structure is automatically constructed, with nodes divided into place and transition, and edges divided into p2... 3) Message Passing Layer: From place to change, realizes message propagation from place to change, uses edge weights to adjust the propagation strength, supports normalization and activation, and realizes message propagation from change to place; 4) Attention Layer: Calculates attention weights for each node in the graph, used for graph-level representation aggregation, and extracts the full-graph-level embedding representation; 5) Classification Layer: Multi-layer fully connected layer for binary classification; 6) Training and Evaluation Module: Model training function, including gradient pruning, loss calculation and metric collection, divides the dataset into training set, validation set and test set, and uses accuracy, precision, recall and F1 score as evaluation metrics.

[0045] S6. Prepare a dataset for the intelligent manufacturing system under different initial states, including the Token distribution, the corresponding correlation matrix, and deadlock labels. This dataset corresponds to the implementation example. Figure 2 The Petri net intelligent manufacturing system is shown in Table 1.

[0046] Table 1 Petri-Net-GCN Deadlock Prediction Dataset

[0047] warehouse characteristics Deadlock tag 1,5,1,0,2,2,5,5,0,4,1 1 1,1,3,5,3,4,1,2,0,4,3 0 0,4,4,1,4,4,5,0,0,2,4 1 1,4,3,5,3,4,1,2,0,4,3 0 1,2,0,0,2,3,0,0,5,2,5 0 2,0,1,2,3,0,1,5,3,3,1 1 0,1,3,5,3,4,1,2,0,4,3 0 5,2,0,0,1,5,5,5,2,5,3 0 2,5,0,0,2,3,0,0,5,2,5 0 3,2,1,1,0,4,0,5,1,3,3 1

[0048] The correlation matrix is ​​as follows:

[0049] [-1,0,0,1,0,0,0,0],

[0050] [1,-1,0,0,0,0,0,0],

[0051] [0,1,-1,0,0,0,0,0],

[0052] [0,0,1,-1,0,0,0,0],

[0053] [0,0,0,0,1,-1,0,0],

[0054] [0,0,0,0,0,1,-1,0],

[0055] [0,0,0,0,0,0,1,-1],

[0056] [0,0,0,0,-1,0,0,1],

[0057] [-1,1,0,0,0,0,-1,1],

[0058] [0,-1,1,0,0,-1,1,0],

[0059] [0,0,-1,1,-1,1,0,0]

[0060] Table 1 shows the features of the locations [1,5,1,0,2,2,5,5,0,4,1], indicating that the Petri net has 11 locations. The numbers from left to right represent the number of tokens in each location. Deadlock labels are 0 and 1, where 1 indicates the presence of deadlock and 0 indicates no deadlock. The association matrix is ​​an 11×8 matrix, representing that the Petri net consists of 11 locations and 8 transitions. 1 represents the output arc from the transition to the location, -1 represents the input arc from the location to the transition, and 0 indicates no connection. Due to the large dataset (1000 data points in total), only 10 data points are shown here.

[0061] S7. Input the prepared dataset as input data into the designed Petri-Net-GCN deadlock prediction model for learning and training. The training results are shown in Table 2.

[0062] Table 2 Training results of the Petri-Net-GCN deadlock prediction model

[0063]

[0064] S8. The trained Petri-Net-GCN deadlock prediction model is used to predict deadlock in an intelligent manufacturing system under unknown conditions. Fifty initial labels of the intelligent manufacturing system under unknown conditions are prepared as input data and fed into the trained prediction model for prediction. The prediction results are shown in Table 3.

[0065] Table 3. Prediction results of Petri-Net-GCN deadlock prediction model

[0066]

[0067]

[0068]

[0069] Table 3 shows that 43 of the initial labels under 50 different states were correctly predicted by the Petri-Net-GCN deadlock prediction model, achieving an accuracy of 86%, which is similar to the test set results of the training model.

[0070] Comparative Example 1

[0071] The dataset used is the same as in Example 1. The model used is the Petri-Net-MLP deadlock prediction model, which mainly includes the following parts: 1) Feature extraction of the dataset: mainly extracting labeled features, arc structure features, structural connection features, and potential deadlock indicators; 2) Multilayer perceptron module: the input layer dimension is the extracted feature dimension, with multiple hidden structures, and the output layer outputs two units to represent deadlock and non-deadlock; 3) Model training module: returns the accuracy, precision, recall, and F1 score of the model training results.

[0072] The training results obtained from the Petri-Net-MLP deadlock prediction model are shown in Table 4.

[0073] Table 4 Training results of the Petri-Net-MLP deadlock prediction model

[0074]

[0075] According to the comparison in Table 2, the deadlock prediction method based on Petri-Net-GCN proposed in this invention has better prediction performance than traditional neural networks.

Claims

1. A deadlock prediction method for intelligent manufacturing systems based on Petri-Net-GCN, characterized in that, Includes the following steps: (1) Divide the intelligent manufacturing system into a production equipment module, a production process module, and a production resource module; (2) Formal modeling of the modules based on Petri nets includes: mapping the production equipment module to the equipment state represented by the place in the Petri net; mapping the production process module to the task scheduling or process transformation represented by the transition in the Petri net; mapping the production resource module to the resource distribution represented by the token in the Petri net; constructing the Petri net structure through directed arc connections and representing it with an incidence matrix, where positive values ​​represent the output arc from the transition to the place, negative values ​​represent the input arc from the place to the transition, and zero values ​​represent no connection; (3) The Petri-Net-GCN deadlock prediction model is constructed by including a place-to-change message passing module and a change-to-place message passing module; a graph neural network structure is used to perform bidirectional aggregation and update of the features between nodes in the Petri network graph; the Petri-Net-GCN model also includes: a feature extraction module for extracting node features from the association matrix and initial labels; an attention layer for calculating node attention weights to enhance the expressive power of key states; and a classification layer for outputting deadlock binary classification results based on graph-level embedding representation; the message passing module is as follows: in the place-to-change module, P2T, place features via correlation matrix Weighted update of transition features The transition to T2P in the storage module is achieved through the transpose of the association matrix. The transition features are backpropagated to places; each message-passing layer is followed by a ReLU activation function to enhance nonlinear expressiveness; the message-passing function from places to the transition module is shown below: ； The message passing function for transitioning to the storage module is shown below: ； in, This represents the feature information of all places in the Petri net, and , This represents the characteristic information of all transitions in a Petri net, and , Denotes the incidence matrix of a Petri net, and , and These are two trainable weight matrices, and , , and Denotes the bias matrix, and , ; (4) Collect the initial Token distribution, correlation matrix and deadlock labels of the intelligent manufacturing system to build a dataset, input it into the model for training; use the trained model to predict the deadlock risk in unknown system states.

2. The deadlock prediction method for intelligent manufacturing systems based on Petri-Net-GCN according to claim 1, characterized in that, In step (1), the production equipment module includes the processing equipment module containing the status and quantity of machine tools and robotic arms; the transportation module containing the status and quantity of AGVs; the warehousing module containing the status and quantity of input / output devices; the production process module including the loading and unloading module status and product quantity; the product flow module status and product quantity; and the production resource module including the raw material module status and inventory; the intermediate product module status and quantity; and the finished product module status and inventory.

3. The deadlock prediction method for intelligent manufacturing systems based on Petri-Net-GCN according to claim 1, characterized in that, In step (2), the place represents the device status node, the transition represents the task scheduling operation, and the token represents the resource distribution; the arc connection direction is consistent with the actual flow direction of the resources.

4. The deadlock prediction method for intelligent manufacturing systems based on Petri-Net-GCN according to claim 1, characterized in that, In step (4), the dataset includes the initial labels, deadlock labels and association matrix of Petri nets in different states.

5. The deadlock prediction method for intelligent manufacturing systems based on Petri-Net-GCN according to claim 1, characterized in that, In step (4), the model training uses accuracy, precision, recall, and F1 score as evaluation metrics; during prediction, the initial identifier vector and correlation matrix of the unknown state are input, and the deadlock probability is output.

6. An electronic device comprising a memory, a processor, and a computer program stored in the memory, characterized in that, When the processor executes the program, it implements the steps of the method as described in any one of claims 1-5.

7. An electronic device, characterized in that, It includes a memory and a processor, the memory storing a computer program, and the processor executing the program to implement the steps of the method according to any one of claims 1-5.

Images