Petri net supervisory control method and controller applied to complex manufacturing system

Abstract

The application provides a Petri net supervisory control method and a controller applied to a complex manufacturing system, and the method comprises the following steps: obtaining system information of the complex manufacturing system, and constructing a Petri net model of the complex manufacturing system according to the system information; constructing a reachable graph of the Petri net model according to the Petri net model; taking the current state and resources of each device as input, and predicting transition enabling values of each device in all reachable states through a trained neural network model; and performing supervisory control on the complex manufacturing system according to the transition enabling values of each device in all reachable states. For the complex manufacturing system, the application can avoid the state explosion problem caused by the calculation of the reachable graph of the complex manufacturing system, greatly reduce the time complexity and calculation complexity caused by the calculation of the reachable graph, and directly realize the deadlock control of the complex Petri net and has high behavior permissibility.

Description

Technical Field

[0001] This invention belongs to the field of engineering manufacturing control technology, specifically relating to a Petri net supervisory control method and controller applied to complex manufacturing systems. Background Technology

[0002] A Flexible Manufacturing System (FMS) is typically a computer-controlled system comprised of CNC machine tools and material handling systems, automating various types of production tasks. Compared to traditional manufacturing systems, FMS offers advantages such as high quality, high flexibility, strong adaptability, and cost savings. FMS is computer-controlled, with no human intervention during processing, and different processing systems compete for the same limited resources. When some processes wait indefinitely for other processes to release resources, deadlock can occur. Once deadlock occurs, the system may malfunction partially or entirely, negating the inherent advantages of FMS and causing unavoidable losses. Accurately describing the operation of a real flexible manufacturing system is very difficult. To solve the deadlock problem in FMS, it is first necessary to model the entire system and simulate its operation. The model then analyzes data on the system's static structure and dynamic behavior.

[0003] Technicians often use Petri nets to accurately describe concurrency, asynchronicity, resource sharing, and conflict behaviors in manufacturing systems (FMS). Petri net reachability graph analysis typically allows for higher behavioral permissibility in network systems and the introduction of state transition separation events and region theory to improve manufacturing system productivity and resource utilization. However, these methods face the problem of state explosion in computationally completing the reachability graph when dealing with complex manufacturing systems, making these control methods difficult to apply to complex systems. Therefore, existing deadlock prevention strategies based on Petri net reachability graph analysis have limitations in application to complex manufacturing systems. Summary of the Invention

[0004] To address the aforementioned problems in the prior art, this invention provides a Petri net supervisory control method and controller applied to complex manufacturing systems. The technical problem to be solved by this invention is achieved through the following technical solution:

[0005] This invention provides a Petri net-based supervisory control method for complex manufacturing systems, comprising:

[0006] S100: Obtain system information of complex manufacturing systems and determine the Petri net model of complex manufacturing systems based on system information;

[0007] S200, Construct the reachability graph of the Petri net model based on the Petri net model of the complex manufacturing system; the reachability graph includes the current state, the next reachable state, and the reachable path from the current state to the next reachable state for each device in the complex manufacturing system;

[0008] S300 takes the current state and resources of each device as input and predicts the transition enable value of each device in all reachable states through a trained neural network model.

[0009] S400 provides supervisory control over complex manufacturing systems based on the transition enable values ​​of each device in all reachable states.

[0010] This invention provides a Petri net supervisory control system for complex manufacturing systems. The controller has a chip installed inside, and when the chip is running, it uses the Petri net supervisory control method for complex manufacturing systems to achieve supervisory control of the complex manufacturing system.

[0011] This invention provides a Petri net supervised control method and controller for complex manufacturing systems. The method involves acquiring system information of the complex manufacturing system and determining its Petri net model based on this information; constructing a reachability graph of the Petri net model; using the current state and resources of each device as input, and predicting the transition enable values ​​of each device across all reachable states using a trained neural network model; and performing supervised control of the complex manufacturing system based on these transition enable values. For complex manufacturing systems, this invention avoids the state explosion problem caused by directly calculating the reachability graph, significantly reducing the time and computational complexity associated with reaching graph calculation. It can directly achieve deadlock control of complex Petri nets with high behavioral permissibility.

[0012] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0013] Figure 1 This is a flowchart illustrating a Petri net supervisory control method for complex manufacturing systems provided by the present invention.

[0014] Figure 2a This is a control flowchart of a complex Petri net with a large initial state provided by the present invention;

[0015] Figure 2b This is a schematic diagram of the Petri net research object under a specified initial state provided by the present invention;

[0016] Figure 3This is the reachability graph of a Petri net system under a specified initial state, provided by the present invention.

[0017] Figure 4 This is a schematic diagram of a data sample set generated by a Petri net system under a specified initial state, provided by the present invention.

[0018] Figure 5 This is a schematic diagram illustrating the process of the dataset generation algorithm provided by this invention searching for all reachable states;

[0019] Figure 6 This is a flowchart of the training process of the neural network model provided by the present invention;

[0020] Figure 7 This is a graph showing the loss variation of the model provided by this invention using hyperparameter combinations in the training set.

[0021] Figure 8 This is a graph showing the loss variation of the model provided by this invention using hyperparameter combinations on the validation set. Detailed Implementation

[0022] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0023] like Figure 1 As shown, the present invention provides a Petri net-based supervisory control method for complex manufacturing systems, comprising:

[0024] S100: Obtain system information of complex manufacturing systems and determine the Petri net model of complex manufacturing systems based on system information;

[0025] S200, Construct the reachability graph of the Petri net model based on the Petri net model; the reachability graph includes the current state, the next reachable state, and the reachable path from the current state to the next reachable state for each device in the complex manufacturing system;

[0026] S300 takes the current state and resources of each device as input and predicts the transition enable value of each device in the reachable state through a trained neural network model.

[0027] It's worth noting that in the Petri net model, states are represented by identifiers, and reachable paths represent transitions. For each device, the transition enable value for all reachable states is either 0 or 1. A value of 1 indicates that the device can transition to that state in the next time step, and only one reachable state is 1 at a time. Transition enable values ​​of 0 cause the system to deadlock, or are limited by the system's own resources. Complex manufacturing systems can effectively avoid deadlock.

[0028] S400 provides supervisory control over complex manufacturing systems based on the transition enable values ​​of each device in all reachable states.

[0029] As an optional embodiment of the present invention, prior to S300, the Petri net supervisory control method applied to complex manufacturing systems further includes:

[0030] Constructing training and validation sets; the construction of training and testing sets in this invention includes:

[0031] S510, Construct a Petri net model with multiple different initial states that have the same Petri net model structure as the complex manufacturing system;

[0032] Among them, the Petri net models with different initial states contain fewer resources than the Petri net models of complex manufacturing systems;

[0033] by Figure 2b The Petri net model corresponding to the complex manufacturing system shown is used as the research object. Note that the initial state in this Petri net model is not the final research object, but is only used to illustrate its reachability graph and neural network model sample collection rules. The specific research object is... Figure 2b The Petri net model structure shown combines the initial state [8, 0, 0, 0, 0, 0, 0, 8, 3, 4, 4].

[0034] Once the research object was determined, based on the same network structure, different smaller initial states were selected to implement a series of network systems with the same structure but smaller initial states. These systems were used to generate the dataset required for training the neural network model. The training set samples were derived from Petri net systems with smaller initial states. Specifically, the number of tokens in libraries P1 and P8 varied within the range of [3, 6], while P9 and P... 10 P 11 The number of place tokens varies from [1, 3], with a total of 243 different small initial state petri net systems. The purpose of the validation set is to compare the performance of different models. By evaluating the performance of different models on the validation set, the best model and hyperparameter combination can be selected. It can also help detect whether the model is overfitting the training data and evaluate model performance. If the model performs well on the training set but poorly on the validation set, it may indicate that the model is overfitting. This can prompt corresponding measures, such as adjusting the model complexity and increasing regularization, to reduce overfitting. The initial state of the Petri net system that generates the validation set samples is larger than the initial state of the Petri net that generates the training set and is closer to the research object. Specifically, the number of place tokens in P1 and P8 are both 7, and the number of place tokens in P9 and P1 is 3. 10 P 11 The network system consists of different initial states with the number of tokens in the library varying in the range of [2, 3].

[0035] S520 generates a dataset that meets the training requirements of a neural network model based on multiple Petri net models with different initial states using a model dataset generation algorithm.

[0036] For a complex Petri net system with a large initial state, the first step is to perform feature analysis and selection, determining the training and validation sets for the neural network model based on the sample features. The trained neural network model can then be used for supervised control of the network system with a large initial state. By inputting the current reachable state, the neural network model processes the data to obtain an accurate prediction of the next action. Based on the prediction, the next action of the system is determined, ensuring that the system reaches as many reachable states as possible without entering a deadlock state, achieving high-activity control. The specific implementation process is as follows: Figure 2a As shown.

[0037] The pre-built neural network model based on a multilayer perceptron structure is iteratively trained using the training set and validation set to obtain the trained neural network model.

[0038] As an optional embodiment of the present invention, the trained neural network model is obtained by iteratively training a pre-built neural network model based on a multilayer perceptron structure using a training set and a validation set, including:

[0039] a) Input the samples in the training set into a pre-built neural network model based on a multilayer perceptron structure, and predict the transition enable values ​​of all reachable states through the neural network model.

[0040] The neural network model of this invention is a multilayer perceptron-based neural network model, comprising an input layer, hidden layers, and an output layer. The number of neurons in the input layer is determined by the initial state and the dimension of the transition sequence of the current Petri net model, while the number of neurons in the output layer is positively correlated with the number of transitions. The number of neurons in the input layer is |P| + |T|, and the number of neurons in the output layer is |T|. The number of sample output vectors output by the output layer neurons is |T|, where each value in the sample output vector is either 0 or 1, representing the enabled state of all transitions in the reachable state of the current Petri net system, indicating whether the corresponding transition is disabled or enabled, respectively. The activation function uses the Sigmoid function; the loss function of the neural network model adopts the binary cross-entropy loss function; and the optimizer of the neural network model uses mini-batch gradient descent combined with the Adam algorithm. This ensures the speed and stability of the loss function descent during model training.

[0041] b. Calculate the loss function of the neural network model by using the predicted transition enable value, and adjust multiple hyperparameters of the neural network model by using mini-batch gradient descent combined with the Adam algorithm.

[0042] c. Validate the neural network model for each iteration using a validation set and select the best-performing combination of hyperparameters;

[0043] d, the neural network model corresponding to the hyperparameter combination is determined as the trained neural network model.

[0044] To enable the neural network model to learn the state transition rules of systems with the same network structure while ensuring the network structure remains unchanged, a series of Petri net systems with the same structure and smaller initial states are implemented by selecting different smaller initial states. The behavior of these Petri net systems with smaller initial states is analyzed under region theory, serving as the training and validation sets for the neural network model. During training, a large number of training and validation datasets are needed to ensure the model's accuracy and generalization ability. The training set contains a large number of samples to fully learn the state transition rules of the network system; the validation set contains relatively fewer samples, but they come from Petri net systems with larger initial states, used to select the best-performing model design and hyperparameter combinations.

[0045] In terms of model design, this invention adopts a neural network model based on a multilayer perceptron structure. Specifically, the model includes a three-layer structure: an input layer, a hidden layer, and an output layer. The number of neurons in the input layer is determined by the initial state of the current Petri net system and the dimension of the transition sequence. The number of neurons in the output layer is related to the number of transitions, which are 19 and 8 respectively. Each dimension of the model's sample output has a value of 0 or 1. If each transition value is considered a separate label, the training task can be viewed as eight independent binary classification problems. In a single binary classification problem, this invention needs to divide the input data into two categories, representing whether the corresponding transition is enabled. The Sigmoid function is suitable as the activation function, with an output range of (0, 1). This invention can set the output threshold of the Sigmoid function to 0.5. When the output of the Sigmoid function is greater than 0.5, the transition is considered enabled in the current state; otherwise, the transition is not enabled. The binary cross-entropy loss function (BCELoss) will yield better results, and its specific implementation is as follows:

[0046]

[0047] When the true probability value y is 0, the first part of the formula is 0, and the predicted probability value is... The value needs to be as close to 0 as possible to minimize the value of the latter part of the formula. Similarly, when the true probability is 1, the latter part is 0, and the predicted probability needs to be as close to 1 as possible to minimize the value of the former part. This achieves the desired result. The optimizer achieves a result closer to the expected value of y. It employs a mini-batch gradient descent method combined with the Adam algorithm to ensure the speed and stability of the loss function descent during model training.

[0048] As an optional embodiment of the present invention, S520 includes:

[0049] S521, based on each reachable state in the Petri net model with different initial states and the preceding correlation matrix, transitions are emitted sequentially to reach the reachable state;

[0050] S522, If the reachable state has been searched or is an illegal state, then the backtracking process is determined. During the backtracking process, it is determined whether the parent state is an illegal state. If it is, the backtracking continues; if not, other reachable states are searched.

[0051] S523, represent the reachable state of the search using the initial state and the transition sequence, and store all transition enable states in the current state;

[0052] S524, which consists of all the stored reachable states and transition enabling states of the Petri net model with different initial states.

[0053] The input sample features of the neural network model are: a vector of dimension |P|+|T|, and the output is a vector of dimension |T|. The first |P| dimensions are the initial state of the current Petri net system, and the |T| dimensions are the transition sequence vector of the current Petri net system from the initial state to a certain reachable state. The sample output is used to represent the enabling state of all transitions in the reachable state of the current Petri net system.

[0054] S524 includes:

[0055] S5241, find the illegal state reached by a certain transition in the current state, then the transition corresponding to the illegal state is the critical transition;

[0056] S5242 modifies the key transitions in the stored dataset so that the neural network model does not learn the enabling behavior of the key transitions during training.

[0057] S525 divides the dataset into a training set and a validation set.

[0058] To generate a dataset that meets the training requirements of neural network models, this invention designs a model dataset generation algorithm. This algorithm is a depth-first traversal process. Using the reachable states of the Petri net and the prior correlation matrix, it calculates which transitions can be emitted in the current state. Then, it traverses all enabled transitions and emits them sequentially. When encountering a reachable state that has already been collected or a state without enabled transitions, it backtracks to the previous state. If the backtracking reaches an illegal state, it is marked with an illegal flag 'd'. It then determines whether the backtracked state is a bad state; if so, the backtracking operation is repeated; otherwise, it continues searching for other unexplored reachable states. If the backtracking reaches a dangerous state, the critical transitions that would lead to deadlock in that state must be disabled. Dataset sample feature selection involves choosing features relevant to the problem from the original data and excluding irrelevant or duplicate features to improve the model's predictive ability and efficiency.

[0059] Figure 3 and Figure 4 This demonstrates a comparison between the reachability graph of a Petri net with a specified initial state and training data samples of a neural network model. Figure 3 In this diagram, circles represent locations, squares represent transitions, and numbers within circles represent tokens, i.e., resource sizes. This invention uses the initial state of a Petri net and the sequence of transitions leading to a specified state as input samples, with the enabling states of each transition corresponding to that state as output samples. Feature transformation converts features into a format suitable for neural network model learning. The transition sequence consists of the sequential transitions emitted from the initial state to the specified state, represented by a vector of length equal to the number of all transitions. Each dimension of the vector corresponds to a transition, indicating the number of times that transition was emitted; 0 indicates that the transition was not emitted. See also... Figure 4 In reachable states, the enabling states of transitions are also represented by vectors of length equal to the number of transitions. Each dimension of the vector represents whether a transition is enabled; enabled transitions are marked as 1, and disabled transitions as 0. With the initial state [3, 0, 0, 0, 0, 0, 0, 3, 1, 1, 1], the generated dataset samples contain [3, 0, 0, 0, 0, 0, 0, 3, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 1, 0, 0, 0] and [3, 0, 0, 0, 0, 0, 0, 3, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0]. The selected Petri net research objects include 11 places and 8 transitions; therefore, each sample is represented by a feature vector of length 27. The first 11 features in the sample represent the number of tokens in each place in the initial state, the next 8 features record the vector representation of the transition emission sequence, and the last 8 features record all transition enabled states in a certain state reached from the initial state through the transition emission sequence.

[0060] To prepare training and test set data samples more accurately and quickly, a data sample generation algorithm is proposed based on sample characteristics. This algorithm can directly obtain the transition enable states of a Petri net after emitting a certain transition sequence to reach a certain state under a specified initial state. The dataset generated by combining this algorithm with different initial states can meet the data characteristics and quantity requirements for model training and validation.

[0061] This algorithm is a depth-first search process. Using the Petri net's states and prior correlation matrices, it calculates which transitions can be emitted in the current state. Then, it iterates through all enabled transitions and emits them sequentially. When encountering a reachable state that has already been collected or a state without enabled transitions, it backtracks to the previous state. If backtracking reaches an illegal state, it marks it with an illegal flag 'd' and checks if the backtracked state is bad. If it is, it repeats the backtracking operation; otherwise, it continues searching for other unexplored reachable states. If backtracking reaches a dangerous state, it also needs to disable the critical transitions that would cause deadlock in that state. The specific search and backtracking processes are as follows: Figure 5 As shown.

[0062] Once the structure of the neural network is determined, the model can be trained. First, the training set is input into the neural network for forward propagation to obtain the model's predicted output values. The loss function is then calculated based on the predicted values ​​and the true labels of the training set. Next, the optimizer updates the network parameters based on the calculated gradient information to minimize the loss function. These steps are repeated until the network parameters converge or a preset number of training epochs are reached. Figure 6 It demonstrates the basic process of neural network training.

[0063] Network structure, batch size, learning rate, and number of iterations are called model hyperparameters. Hyperparameters cannot be learned by the model and must be manually set before training. These hyperparameters need to be selected and adjusted according to the specific problem and dataset, and experiments are usually required to determine the optimal combination of hyperparameters. If the selected hyperparameters are inappropriate, it may lead to overfitting or underfitting of the model, thereby affecting the model's performance and generalization ability. Figure 7 The diagram shows the loss curves on the training set during the training of a neural network using different combinations of hyperparameters. Figure 8 This paper demonstrates the loss variation curves of a neural network during training with different combinations of hyperparameters on the validation set. The optimal hyperparameter combination is then identified based on the model's performance on the validation set.

[0064] The trained neural network model is used in a complex Petri net with a large initial state. By inputting its initial state and the transition sequence to a certain state, all the transition enabling states of the corresponding reachable state can be obtained. By issuing the predicted enabling transitions, the system can be guaranteed not to enter a deadlock state, thus achieving a highly active supervisory control effect.

[0065] This invention provides a Petri net supervisory control controller for complex manufacturing systems. The controller contains a chip, and when the chip is running, it uses the Petri net supervisory control method for complex manufacturing systems to achieve supervisory control of the complex manufacturing system.

[0066] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0067] Although this application has been described herein in conjunction with various embodiments, those skilled in the art will understand and implement other variations of the disclosed embodiments by reviewing the accompanying drawings, the disclosure, and the appended claims in carrying out the claimed application. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude a plurality.

[0068] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A Petri net supervisory control method applied to complex manufacturing systems, characterized in that, include: S100, Obtain system information of the complex manufacturing system, and determine the Petri net model of the complex manufacturing system based on the system information; S200, Construct a reachability graph of the Petri net model based on the Petri net model of the complex manufacturing system; the reachability graph includes the current state, the next reachable state, and the reachable path from the current state to the next reachable state for each device in the complex manufacturing system; S300 takes the current state and resources of each device as input and predicts the transition enable value of each device in all reachable states through a trained neural network model. S400, supervises and controls the complex manufacturing system based on the transition enable values ​​of each device in all reachable states; Prior to S300, the Petri net supervisory control method applied to complex manufacturing systems further includes: Construct training and validation sets; The pre-built neural network model based on a multilayer perceptron structure is iteratively trained using the training set and the validation set to obtain the trained neural network model. The construction of the training set and test set includes: S510, Construct multiple Petri net models with different initial states that have the same Petri net model structure as the complex manufacturing system; The resources contained in the library of the Petri net model are less than the resources of the Petri net model of the complex manufacturing system. S520, Based on the Petri net models with multiple different initial states, a dataset that meets the requirements for training the neural network model is generated using a model dataset generation algorithm; The neural network model takes as input a |P|+|T| dimensional vector and outputs a |T| dimensional vector. The first |P| dimensions represent the initial state of the current Petri net system, and the |T| dimensions represent the sequence of transitions from the initial state to a reachable state within the current Petri net system. The output represents the enabling states of all transitions within the reachable state of the current Petri net system. The neural network model has |P|+|T| input layer neurons and |T| output layer neurons. The activation function is the Sigmoid function. The loss function is the binary cross-entropy loss function. The neural network optimizer uses a combination of mini-batch gradient descent and the Adam algorithm.

2. The Petri net supervisory control method for complex manufacturing systems according to claim 1, characterized in that, The neural network model in S300 is a neural network model based on a multilayer perceptron structure, which includes an input layer, a hidden layer, and an output layer.

3. The Petri net supervisory control method for complex manufacturing systems according to claim 1, characterized in that, The S520 includes: S521, according to each reachable state in the Petri net model with different initial states and the preceding correlation matrix, transitions are emitted sequentially to reach the reachable state; S522, If the reachable state has been searched or is an illegal state, then the backtracking process is determined. During the backtracking process, it is determined whether the parent state is an illegal state. If it is, the backtracking continues; if not, other reachable states are searched. S523, represent the reachable state of the search using the initial state and the transition sequence, and store all transition enable states in the current state; S524, which consists of all the stored reachable states and transition enabling states of the Petri net model with different initial states. S525, the dataset is divided into a training set and a validation set.

4. The Petri net supervisory control method for complex manufacturing systems according to claim 3, characterized in that, S524 includes: S5241, find the illegal state reached by a certain transition in the current state, then the transition corresponding to the illegal state is the critical transition; S5242, Modify the key transitions in the stored dataset so that the neural network model does not learn the enabling behavior of the key transitions during training.

5. The Petri net supervisory control method for complex manufacturing systems according to claim 1, characterized in that, The method of iteratively training a pre-built neural network model based on a multilayer perceptron structure using a training set and a validation set to obtain a trained neural network model includes: The samples in the training set are input into a pre-built neural network model based on a multilayer perceptron structure, and the neural network model is used to predict the transition enable values ​​of all reachable states. The loss function of the neural network model is calculated by predicting the transition enable value, and multiple hyperparameters of the neural network model are adjusted by using mini-batch gradient descent combined with the Adam algorithm. The neural network model is validated using a validation set for each iteration, and the best-performing combination of hyperparameters is selected. The neural network model corresponding to the hyperparameter combination is determined as the trained neural network model.

6. A controller for Petri net supervisory control applied to complex manufacturing systems, characterized in that, The controller contains a chip, which, when running, utilizes the Petri net supervisory control method for complex manufacturing systems as described in any one of claims 1 to 5 to achieve supervisory control of the complex manufacturing system.

Images