A service flow prediction method based on petri net reachability graph interpretation

Abstract

The application discloses a business process prediction method based on Petri net reachable graph interpretation, and belongs to the technical field of business process management. First, a technical expert generates a reachable graph and a transition sequence set by constructing a Petri net. Second, a recurrent neural network is trained by using the transition sequence set, and a hidden state of the recurrent neural network is obtained, and then a hidden state decoding class is obtained. A mapping relationship is established, and a reachable state set of the Petri net reachable graph is mapped onto the hidden state decoding class of the recurrent neural network. Finally, the predictive business process monitoring is completed by using an interpretable recurrent neural network model and adding a fully connected neural network. The application proposes a new method of recurrent neural network internal state interpretation based on the Petri net reachable graph, effectively improves the interpretability of the recurrent neural network internal state, and realizes the future state prediction of the business process transition sequence.

Description

Technical Field

[0001] This invention belongs to the field of business process management technology, specifically relating to a business process prediction method based on Petri net reachability graph interpretation. Background Technology

[0002] Predictive business process monitoring, or process prediction, aims to predict the future state of an ongoing business process based on incomplete execution trajectories and historical event logs. Currently, methods for process prediction mainly include process model-driven prediction, data-driven prediction, and a combination of both. Numerous studies have shown that data-driven methods based on recurrent neural networks (RNNs) have achieved superior prediction performance in terms of accuracy. While RNNs possess powerful modeling capabilities for regular languages, compared to other machine learning models (Support Vector Regression, Decision Trees, etc.), RNNs are, to some extent, a black box problem; their internal mechanisms are not easily understood and lack interpretability. Therefore, there is an urgent need for a technique for training and interpreting RNN models to enable business process prediction. Summary of the Invention

[0003] To address the aforementioned issues, this invention proposes a business process prediction method based on Petri net reachability graph interpretation. The method generates sequences from the constructed Petri net to train a recurrent neural network model, and then completes the prediction of the business process based on the interpretable recurrent neural network and the fully connected neural network.

[0004] The technical solution of the present invention is as follows:

[0005] A business process prediction method based on Petri net reachability graph interpretation includes the following steps:

[0006] Step 1: Technical experts within the enterprise or organization construct Petri nets based on business needs;

[0007] Step 2: Obtain the reachability graph of the Petri net using the reachability graph construction algorithm, and obtain its set of reachable states;

[0008] Step 3: Randomly generate a set of transition sequences using Petri nets, and generate a set of positive and negative sample prefix transition sequences; then, train a recurrent neural network model using the prefix transition sequence set, and save the hidden state of the recurrent neural network; obtain the hidden state decoding class of the recurrent neural network through a decoding function;

[0009] Step 4: Define the mapping relationship. Map the set of reachable states of the Petri net reachable graph to the hidden state decoding class of the recurrent neural network. Obtain the reachability graph of the hidden state of the recurrent neural network based on the transition relationship between the reachable states of the reachable graph, and interpret the hidden state of the recurrent neural network through the reachability graph of the hidden state of the recurrent neural network.

[0010] Step 5: Using an interpretable recurrent neural network model, add a fully connected neural network layer to predict the future state of the ongoing business process transition sequence, thus completing predictive business process monitoring.

[0011] Furthermore, the process of training the recurrent neural network model to obtain the hidden state decoding class is as follows: First, a set of prefix transition sequences is generated using the set of transition sequences; then, the transition of the last event in the prefix transition sequence is taken as the label, and the prefix transition sequence and its corresponding label are used as the input of the recurrent neural network model; wherein, the recurrent kernel of the recurrent neural network model has memory and extracts information through parameter sharing at different times; when training the recurrent neural network model, the hidden state at the last time of each prefix transition sequence is saved; the decoding function is used to decode all the hidden states to obtain the decoding class of the hidden states.

[0012] Furthermore, the decoding functions used include linear logistic regression, support vector regression machine, and multilayer perceptron regressor; the kernels of the support vector regression machine are "rbf", "linear", "poly", and "sigmoid", respectively.

[0013] Furthermore, the process of obtaining the hidden state reachability graph based on the mapping relationship is as follows: First, based on the reachability graph of the Petri net, obtain the reachability graph of each transition t. i The set of reachable states after triggering is in, Representing the transition t i After triggering, different reachable states are defined, where M is the total set of reachable states and i is the sequence number of each transition. A mapping relationship f: M→V is defined as a one-to-one mapping from the Petri net's reachable state set to the hidden state decoding class of the recurrent neural network. This mapping relationship considers the order, conflict, and concurrent behavior in the transition sequence. Specifically, the mapping relationship is as follows: in, V represents the decoding class of the hidden state, and V is the total set of decoding classes. Based on the transition relationship of the Petri net reachability graph, the reachability relationship between the hidden state decoding classes is added to obtain the reachability relationship graph of the hidden state. Finally, the operation mechanism of the hidden state inside the recurrent neural network is explained by simulating the operation of the reachability relationship graph.

[0014] Furthermore, predictive business process monitoring includes predictions of the remaining time of an ongoing transition sequence, the next transition, or the outcome.

[0015] The beneficial technical effects of this invention are as follows:

[0016] Building Petri nets based on the specific business needs of an enterprise or organization can solve the problem of predicting business processes with concurrent structures, which is more in line with the needs of the enterprise. Using structured Petri net reachability graphs can better explain the internal operating mechanism of recurrent neural networks, thereby better predicting and managing business processes. Based on Petri nets, a set of transition sequences is generated, which is then used to train recurrent neural networks. Finally, recurrent neural networks and fully connected neural networks are used to predict the future state of business processes. Attached Figure Description

[0017] Figure 1 This is a framework diagram of the business process prediction method based on Petri net reachability graph interpretation of the present invention;

[0018] Figure 2 This is a Petri net constructed according to one embodiment of the present invention;

[0019] Figure 3 This is a flowchart illustrating the steps of training a recurrent neural network model to obtain the hidden state decoding class in this invention.

[0020] Figure 4 This is a flowchart illustrating the steps of obtaining the hidden state reachability graph based on the mapping relationship in this invention.

[0021] Figure 5 This is a flowchart of the process prediction steps based on an interpretable recurrent neural network according to the present invention. Detailed Implementation

[0022] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0023] One approach to interpreting recurrent neural network (RNN) models is to use visible tasks to describe invisible tasks within the RNN. In specific scenarios, visible tasks are used to simulate deep neural network models to verify the consistency of RNN execution results. In Petri net-based regular language recognition, process models can serve as one of the visible tasks for interpreting deep learning models and can simulate the internal workings of RNNs. Currently, besides structured finite state machines (FSMs) used to explain RNNs, Petri nets and their reachability graphs are tools capable of identifying sets of transition sequences. Compared to FSMs, Petri nets can handle more complex structures, such as concurrent structures. A Petri net's reachability graph consists of reachable states and the transition relationships between them. A reachable state can be understood as the state after a transition occurs, which is highly similar to the hidden states of a RNN. Based on this, using structured Petri net reachability graphs can better explain the internal mechanisms of RNNs, thereby better predicting and managing business processes.

[0024] This invention primarily uses Petri net reachability graphs to explain the hidden states of recurrent neural networks, thereby optimizing business process prediction. Specifically, Petri nets are established based on business needs, and corresponding transition sequences are generated using Petri nets. Through the ability of recurrent neural networks to identify transition sequences, the next activity (i.e., the next transition) is predicted, and the subsequent configuration of the transition is completed to meet enterprise needs.

[0025] Figure 1 This is a framework diagram of one embodiment of the present invention, including constructing a Petri net, generating a set of positive and negative transition sequences, training a recurrent neural network model, decoding the hidden states of the recurrent neural network using a decoding function, obtaining the decoding class of the hidden states of the recurrent neural network, obtaining the reachability graph of the hidden states of the recurrent neural network, interpreting the internal operating mechanism of the recurrent neural network using the reachability graph of the hidden states of the recurrent neural network, and predicting the future state transition sequence of the running business process through an interpretable recurrent neural network and a fully connected neural network. Specifically, it includes the following steps:

[0026] Step 1: Technical experts within the enterprise or organization construct Petri nets based on business needs. The Petri nets in this invention are, by default, bounded Petri nets.

[0027] Step 2: Obtain the reachability graph of the Petri net using the reachability graph construction algorithm, and obtain its set of reachable states.

[0028] Step 3: Randomly generate a set of transition sequences using Petri nets, and generate a set of positive and negative sample prefix transition sequences. Then, train a recurrent neural network model using the prefix transition sequence set and save the hidden states of the recurrent neural network. Obtain the hidden state decoding class of the recurrent neural network using a decoding function.

[0029] Figure 2 This invention describes a Petri net with a concurrent structure, constructed by technical experts within an enterprise or organization after they have become familiar with the business, based on business needs. In this net, boxes represent transitions, circles represent places, arrows represent flow relationships, and t1, t2, t3, t4, t5, t6, t7, and t8 represent the names of different transitions. Petri nets can describe the concurrent behavior of transition sequences. This invention primarily studies Petri nets with concurrent structures. Using this Petri net, sets of positive and negative sample sequences can be randomly generated, and then a recurrent neural network model can be trained.

[0030] Figure 3 This is a step diagram illustrating the process of training a recurrent neural network (RNN) model to obtain the hidden state decoding class. First, a set of prefix transition sequences is generated using the set of transition sequences; this is the prefix trajectory of the transition sequences. Then, the transition of the last event in the prefix transition sequence is taken as the label, and the prefix transition sequence and its corresponding label are used as input to the RNN model. The recurrent kernel of the RNN model has memory, extracting information through parameter sharing at different time steps. `x` represents the input feature of the RNN model at the current time step; `h` represents the output feature of the recurrent kernel at the current time step. During training the RNN model, the hidden state at the last time step of each prefix transition sequence is saved. The decoding function is used to decode all the hidden states, thus obtaining the hidden state decoding class (specifically, such as...). Figure 4 (As shown in the lower left corner). The decoding functions used mainly include linear logistic regression, support vector regression (whose kernels are "rbf", "linear", "poly", and "sigmoid" respectively), and multilayer perceptron regressor (MLP regressor).

[0031] Step 4: Define the mapping relationship. Map the set of reachable states of the Petri net reachable graph to the hidden state decoding class of the recurrent neural network. Obtain the reachability graph of the hidden state of the recurrent neural network based on the transition relationship between the reachable states of the reachable graph, and interpret the hidden state of the recurrent neural network through the reachability graph of the hidden state of the recurrent neural network.

[0032] Figure 4 This is a flowchart illustrating the steps of obtaining the hidden state reachability graph based on mapping relationships in this invention. First, based on the reachability graph of the Petri net, each transition t is obtained. i The set of reachable states after triggering is in, Representing the transition t i After triggering, different reachable states are defined, where M is the total set of reachable states and i is the sequence number of each transition. The mapping f: M → V is defined as a one-to-one mapping from the Petri net's reachable state set to the hidden state decoding class of the recurrent neural network. This mapping considers the order, conflict, and concurrent behavior in the transition sequence, but does not account for the behavior of recurrent states. Specifically, the mapping is as follows: in, The decoder class represents the hidden state, and V is the total set of decoder classes. This invention uses the management of highway traffic fines as an example. Transition t1 represents creating a fine, transition t2 represents sending a fine, transition t3 represents inserting a fine notification, transition t4 represents inserting a date to appeal to the county government, transition t5 represents appealing to the judge, transition t6 represents payment, transition t7 represents adding a penalty, and transition t8 represents sending a credit collection. Figure 4 The set of reachable states after transition t1 is {t1|→M1}={M1}, the set of reachable states after transition t2 is {t2|→M2}={M2}, the set of reachable states after transition t3 is {t3|→M3,M5}={M3,M5}, the set of reachable states after transition t4 is {t4|→M4,M5}={M4,M5}, the set of reachable states after transition t5 is {t5|→M6}={M6}, the set of reachable states after transition t6 is {t6|→M7,M9}={M7,M9}, the set of reachable states after transition t7 is {t7|→M8,M9}={M8,M9}, and the set of reachable states after transition t8 is {t8|→M1,M9}={M8,M9}. end}={M end}.in addition, This indicates the decoding class corresponding to transition t1. This indicates the decoding class corresponding to transition t2. This indicates the decoding class corresponding to transition t3. This indicates the decoding class corresponding to transition t4. This indicates the decoding class corresponding to transition t5. This indicates the decoding class corresponding to transition t6. This indicates the decoding class corresponding to transition t7 and This indicates the decoding class corresponding to transition t8.

[0033] Based on the transition relationships in the Petri net reachability graph, reachability relationships between hidden state decoding classes are added to obtain the reachability graph of the hidden states. Finally, the operation mechanism of the hidden states inside the recurrent neural network is explained by simulating the operation of the reachability graph. An example illustrating the operation of the hidden states inside the recurrent neural network is provided, given a running prefix transition sequence.<t1,t2,t3,t4,t5> Its trajectory in the reachability graph is represented by dashed lines, that is, its trajectory is...

[0034] Step 5: Using an interpretable recurrent neural network model, add a fully connected neural network layer to predict the future state of the ongoing business process transition sequence, thus completing predictive business process monitoring. Predictive business process monitoring includes predicting the remaining time of the ongoing transition sequence, predicting the next transition, or predicting the outcome.

[0035] Figure 5 This invention presents a step diagram for business process prediction based on interpretable recurrent neural networks and multilayer perceptrons. Further explanation of the hidden states within the recurrent neural network using a reachability graph demonstrates that the recurrent neural network can effectively recognize the normal language of Petri nets, i.e., the transition sequence. Based on this, the recurrent neural network combined with a multilayer perceptron is applied to predictive business processes, i.e., predictive business process monitoring. Predictive business process monitoring aims to predict the remaining time, next transition, or outcome of an ongoing business process. A multilayer perceptron is a multilayer fully connected neural network model; specifically, it adds a layer of linear fully connected neural network to the recurrent neural network to achieve business process prediction, enabling enterprises or organizations to better manage their business processes.

[0036] Regarding the output of business process prediction, for the currently running prefix transition sequence<t1,t2,t3,t4> Taking the next transition as an example, we can predict that the next transition is t5 by adding a fully connected neural network layer to an interpretable recurrent neural network model.

[0037] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A business process prediction method based on Petri net reachability graph interpretation, characterized in that, Includes the following steps: Step 1: Technical experts within the enterprise or organization construct Petri nets based on business needs; Step 2: Obtain the reachability graph of the Petri net using the reachability graph construction algorithm, and obtain its set of reachable states; Step 3: Randomly generate a set of transition sequences using Petri nets, and generate a set of positive and negative sample prefix transition sequences; Then, a recurrent neural network model is trained using a set of prefix transition sequences, and the hidden states of the recurrent neural network are saved; the decoding class of the hidden states of the recurrent neural network is obtained through a decoding function. Step 4: Define the mapping relationship. Map the set of reachable states of the Petri net reachable graph to the hidden state decoding class of the recurrent neural network. Obtain the reachability graph of the hidden state of the recurrent neural network based on the transition relationship between the reachable states of the reachable graph, and interpret the hidden state of the recurrent neural network through the reachability graph of the hidden state of the recurrent neural network. Step 5: Using an interpretable recurrent neural network model, add a fully connected neural network layer to predict the future state of the ongoing business process transition sequence, thus completing predictive business process monitoring. The process of training a recurrent neural network (RNN) model to obtain the hidden state decoding class is as follows: First, a set of prefix transition sequences is generated using the set of transition sequences; then, the transition of the last event in the prefix transition sequence is taken as the label, and the prefix transition sequence and its corresponding label are used as the input of the RNN model; the recurrent kernel of the RNN model has memory and extracts information through parameter sharing at different times; during the training of the RNN model, the hidden state at the last time of each prefix transition sequence is saved; the decoding function is used to decode all the hidden states to obtain the hidden state decoding class.

2. The business process prediction method based on Petri net reachability graph interpretation according to claim 1, characterized in that, The decoding functions used include linear logistic regression, support vector regression machine, and multilayer perceptron regressor; the kernels of the support vector regression machine are "rbf", "linear", "poly", and "sigmoid", respectively.

3. The business process prediction method based on Petri net reachability graph interpretation according to claim 1, characterized in that, The process of obtaining the hidden state reachability graph based on the mapping relationship is as follows: First, based on the reachability graph of the Petri net, obtain the reachability graph of each transition t. i The set of reachable states after triggering is ,in, Representing the transition t i After triggering, different reachable states are defined, where M is the total set of reachable states and i is the sequence number of each transition. A mapping relationship f: M→V is defined as a one-to-one mapping from the Petri net's reachable state set to the hidden state decoding class of the recurrent neural network. This mapping relationship considers the order, conflict, and concurrent behavior in the transition sequence. Specifically, the mapping relationship is as follows: ,in, V represents the decoding class of the hidden state, and V is the total set of decoding classes. Based on the transition relationship of the Petri net reachability graph, the reachability relationship between the hidden state decoding classes is added to obtain the reachability relationship graph of the hidden state. Finally, the operation mechanism of the hidden state inside the recurrent neural network is explained by simulating the operation of the reachability relationship graph.

4. The business process prediction method based on Petri net reachability graph interpretation according to claim 1, characterized in that, Predictive business process monitoring includes predicting the remaining time of an ongoing transition sequence, predicting the next transition, or predicting the outcome.

Images