Production business process activity prediction method based on bidirectional sequence diagram convolutional network

By building a two-way sequence diagram in the production business process and combining the graph convolution network and attention mechanism, the problem that traditional methods are difficult to capture complex relationships is solved, and accurate prediction of the next activity in the production business process is achieved, which improves the prediction Accuracy and robustness.

CN118982093APending Publication Date: 2024-11-19ZHEJIANG UNIV OF TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410965422.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-18
Publication Date
2024-11-19

AI Technical Summary

Technical Problem

In a changing production environment, traditional activity prediction methods are difficult to accurately predict the next activity in the production business process, and often ignore the complex relationship and attribute data of events.

Method used

Using a method based on a two-way sequential graph convolution network, by constructing forward and reverse sequential graph structures, combining graph convolution networks and attention mechanisms, we can deeply analyze the relationship between activities and attributes, dynamically capture the complex relationship between nodes, and use Leaky ReLU activation function enhances the nonlinear expression ability of the model.

Benefits of technology

Significantly improves the accuracy and robustness of activity predictions, provides better performance and interpretability, and provides strong support for production process optimization and decision-making.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118982093A_ABST
    Figure CN118982093A_ABST
Patent Text Reader

Abstract

The invention discloses a production business process activity prediction method based on a bidirectional sequence diagram convolutional network. The method comprises the following steps: firstly, constructing event logs of a production business process into forward and reverse sequence diagram structures, and respectively processing on an activity level and an attribute level; the sequence diagram of the activity level learns feature representation between nodes through a diagram convolution layer, and forward and reverse features are fused by using an attention mechanism to form activity level information; according to the attribute level, the sequence diagram of each attribute is processed, and attribute level information is fused through feature splicing and an attention mechanism; and finally, combining information of the activity and the attribute level, and inputting the information into a classification model to predict the next activity. According to the method, the bidirectional sequence diagram, the diagram convolution network and the attention mechanism are fully utilized, node features are gradually abstracted and combined, the learning ability of the model for complex features is improved, and therefore the next activity in the production business process can be efficiently and accurately predicted.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of business process management, and in particular to a production business process activity prediction method based on a bidirectional sequential graph convolutional network. Background Art

[0002] In modern enterprise management, it is crucial to ensure the efficient operation of production business processes. However, in a changing production environment, accurately predicting activities in business processes is a very challenging task. Predicting the upcoming activities (next activities) in business processes is one of the core issues of business process monitoring. This predictive ability can provide analysts with key information, enabling them to proactively intervene and prevent potential adverse actions.

[0003] Traditional activity prediction methods mostly rely on Recurrent Neural Networks (RNNs), which directly extract prefix sequence information from raw event logs. These methods mainly focus on the activity and timestamp attributes in event logs, and often ignore data describing events from other dimensions.

[0004] With the rise of Graph Neural Networks (GNNs), new possibilities are provided for processing complex structured data. In production business processes, execution logs record the sequence of activities in detail, and this information can be constructed in the form of a graph. By combining Graph Convolutional Networks (GCNs) with a bidirectional sequential graph structure, the complex relationships between activities in business processes can be captured more effectively. In addition, the introduction of the Attention Mechanism further enhances the model's capabilities, enabling it to adaptively identify and focus on the importance of different attributes to activity predictions, thereby significantly improving the accuracy of predictions.

[0005] In this regard, the present invention proposes a production business process activity prediction method based on a bidirectional sequential graph convolutional network. This method fully considers the characteristics of the production business process, designs a special structure of the forward and reverse sequential graphs, and also integrates the graph convolutional network and the attention mechanism to achieve accurate prediction of the next activity in the production business process. By deeply mining the associated information in the execution log data, this method has better performance and interpretability, and provides strong support for the optimization and decision-making of the actual production process. Summary of the invention

[0006] The present invention proposes a method for predicting production business process activities based on a bidirectional sequential graph convolutional network. The method effectively utilizes event logs to construct bidirectional sequential graphs, including forward and reverse sequential graphs, to accurately map the relationship between activities and attributes in the production process. By designing graph structures at the activity level and attribute level, the method can deeply analyze the intrinsic connections of the production process. Furthermore, a graph convolutional network and an attention mechanism are introduced to dynamically capture the complex associations between nodes, and the nonlinear expression ability of the model is enhanced through the Leaky ReLU activation function to avoid the problem of neuron saturation. By stacking multiple layers of graph convolutional networks, the present invention can finely model the complex relationships in the production process and significantly improve the accuracy and robustness of the prediction.

[0007] The production business process activity prediction method based on a bidirectional sequential graph convolutional network includes the following steps:

[0008] 1) Convert the activity sequence in the event log of the production business process into a bidirectional sequence graph structure;

[0009] 2) Design an activity-level graph convolution layer to process the forward sequence graph and the reverse sequence graph respectively, using the activity adjacency matrix A 活动 and B 活动 The information passing convolution operation on the active node is as follows:

[0010]

[0011] in, It is the active node representation of layer l of the forward sequence graph. It is the activity node representation of layer l of the reverse sequence diagram. refers to the learning weight parameter of the l-th layer of the forward sequential graph, refers to the learning weight parameter of the lth layer of the reverse sequence graph, and δ is the ReLU activation function. Through this convolution operation, the information of the active layer can be passed to the next layer through the adjacency matrix, thereby obtaining the representation of the active nodes of the next layer and

[0012] 3) Design the attribute-level graph convolution layer to process the forward sequence graph and reverse sequence graph of each attribute i respectively, using the attribute adjacency matrix A 属性i and B 属性i Information is transferred to attribute nodes, and the convolution operation is as follows:

[0013]

[0014] Among them, for the i-th attribute, It is the attribute node representation of the l-th layer of the forward sequence graph. It is the attribute node representation of the l-th layer of the reverse sequence graph. refers to the learning weight parameter of the l-th layer of the forward sequential graph, refers to the learning weight parameter of the lth layer of the reverse sequence graph, and δ is the ReLU activation function. Through this convolution operation, the information of the attribute layer can be passed to the next layer through the adjacency matrix, thereby obtaining the representation of the attribute nodes of the next layer and

[0015] 4) Introduce the attention mechanism to fuse the global feature representation of the forward sequence graph and the reverse sequence graph at the activity level. First, calculate the global feature representation of the forward sequence graph and the reverse sequence graph:

[0016]

[0017] in, is the global feature representation of the forward sequential graph at layer l, is the global feature representation of the reverse order graph at layer l, is the representation of the ith activity on the lth layer, is the representation of the i-th activity on the l-th layer.

[0018] Then, based on the global feature representation of the forward sequence graph and the reverse sequence graph, the attention weight α between the activity node i of the forward sequence graph and the activity node j of the reverse sequence graph at the activity level is calculated: ij :

[0019]

[0020] Among them, softmax is an activation function for category classification problems, which maps a set of values ​​to a numerical range (0,1) representing the probability distribution, and LeakyReLU is a neural network unit activation function. is a learnable weight vector, || represents a connection operation, represents the weight between the active node in the lth layer and all nodes in the l-1th layer on the forward sequence graph, represents the weight between the active node of the lth layer and all nodes of the l-1th layer on the reverse sequence graph, and They are the activity node representations of layer l of the forward sequence diagram and the reverse sequence diagram respectively.

[0021] Finally, the global feature representations of the activity-level forward and reverse sequence graphs are aggregated based on the attention weights:

[0022]

[0023] in, It is a feature representation that integrates the global information of the forward sequence graph and the reverse sequence graph of the lth layer. is the global feature representation of the forward sequential graph at layer l, It is the global feature representation of the reverse sequence graph at layer l.

[0024] 5) Design an attribute fusion layer to fuse the full-graph feature representations of multiple attributes in the attribute level. First, calculate the global feature representations of the forward sequence graph and the reverse sequence graph:

[0025]

[0026] Among them, for attribute j, is the global feature representation of the forward sequential graph on the lth layer, is the global feature representation of the reverse sequence graph on the lth layer, is the i-th attribute node representation of the forward sequence graph on the l-th layer, is the i-th attribute node representation of the reverse sequence graph on level l, and N is the number of nodes in the graph.

[0027] Next, the global feature representations of the forward sequence graph and the reverse sequence graph of the same attribute are concatenated and fused:

[0028]

[0029] in, It is an information representation of the global features of the forward sequence graph and the reverse sequence graph of the l-th layer attribute j.

[0030] 6) Introducing the attention mechanism based on the attribute fusion layer to integrate the full-image feature fusion representation of all attributes First, the attention weight α of the full-image feature fusion representation of each attribute is calculated j :

[0031]

[0032] Among them, α j is the attention weight of the full-image feature fusion representation vector of attribute j at layer l.

[0033] Then, the full-image feature fusion representation vectors of all attributes are weighted fused based on the attention weights:

[0034]

[0035] in, It is the information representation of the global features of the forward sequence graph and the reverse sequence graph of the lth layer attribute j. It is a feature representation that integrates the global information of all J attributes.

[0036] 7) Next activity prediction. First, the final global feature representations at the activity level and attribute level are concatenated and fused:

[0037]

[0038] in, It is a feature representation that integrates the global information of all J attributes. It is a feature representation that integrates the global information of the forward sequence graph and the reverse sequence graph of the lth layer. It is an information representation of the final global features that concatenates and fuses the activity level and attribute level of the lth layer.

[0039] Next, use the classification model to predict the next activity:

[0040]

[0041] Among them, T i is the predicted value of the corresponding activity i, W r is the weight matrix of the classification model, b r is the bias of the classification model.

[0042] Further, step 1) consists of the following steps:

[0043] 1.1) Constructing the activity layer, each event in the event log is regarded as an activity-level node, and the event dependencies between activity-level nodes constitute the edges of the activity layer;

[0044] 1.2) Construct the attribute layer, and take each attribute of the event, such as timestamp, executor, performance, etc., as a node at the attribute level. Each attribute forms an edge of the internal node of the attribute according to the event dependency relationship;

[0045] 1.3) Construct a forward sequence graph at the activity level, according to the execution order of events in each case in the event log, that is, arrange the activities in order from small to large timestamps, connect the activity nodes in sequence, and the direction of the edge is forward;

[0046] 1.4) Construct a reverse sequence graph at the activity level, according to the execution order of events in each case in the event log, that is, arrange the activities in descending order of timestamps, connect the activity nodes in sequence, and the direction of the edges is reverse;

[0047] 1.5) Construct a forward sequence graph at the attribute level. For each attribute of the event, arrange the attributes in the order of the execution of the events in each case, that is, in the order of timestamps from small to large, and connect the attribute nodes in sequence, with the direction of the edge pointing to the forward direction;

[0048] 1.6) Construct a reverse order graph at the attribute level. For each attribute of the event, arrange the attributes in descending order of timestamps according to the execution order of the events in each case, connect the attribute nodes in sequence, and point in the reverse direction of the edges;

[0049] 1.7) Use one-hot encoding to represent the activity nodes and attribute nodes of each attribute on the forward and reverse sequence diagrams at the activity level and attribute level, respectively, and obtain the activity node representation Attribute node representation

[0050] 1.8) Create the active node adjacency matrix A of the forward sequence graph 活动 and the attribute node adjacency matrix A 属性i , the active node adjacency matrix B of the reverse sequence graph 活动 and the attribute node adjacency matrix B 属性i .

[0051] The technical concept of the present invention is: an innovative method for predicting the next activity of a production business process is proposed. The method realizes an in-depth analysis of the relationship between activities and attributes in the production process by constructing forward and reverse sequence graph structures and integrating graph convolutional networks and attention mechanisms. Specifically, the event log is first converted into forward and reverse sequence graph structures at the activity level and attribute level, wherein the graph nodes at the activity level represent activities, and the graph nodes at the attribute level represent attributes. Each attribute independently constructs forward and reverse sequence graphs to comprehensively and accurately reflect the relationship in the production business process. On this basis, the graph convolutional network is used to learn node representations on these sequence graphs, and the degree of association between nodes is dynamically captured through the attention mechanism. In addition, the Leaky ReLU activation function is introduced to avoid the problem of neuron death and enhance the expression ability of the model. By independently modeling and processing the forward and reverse sequence graphs of activities and each attribute with graph convolution layers, the method can fully express the potential relationship in the attributes, realize accurate modeling of complex relationships, and thus significantly improve the accuracy of prediction. The technical concept is not only highly adaptable, but also has a wide range of generalization capabilities, providing strong technical support for business process optimization and decision making.

[0052] The advantages of the present invention are: it adopts forward and reverse sequential graph structures, and combines graph convolutional network technology, further subdividing the graph convolution layer design into activity level and attribute level, thereby giving the model powerful modeling capabilities, and being able to accurately and dynamically predict the next activity in the production business process. In addition, by introducing the attention mechanism, the model can more effectively capture and learn the associations between nodes, and the addition of the Leaky ReLU activation function effectively avoids the problem of neuron death, further improving the learning ability of the model. This method is particularly suitable for processing complex multi-level production business processes. It not only provides a high degree of accuracy, but also has excellent flexibility, providing reliable and intelligent technical support for actual business decisions. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 It is the overall flow chart of the present invention;

[0054] Figure 2 This is an example diagram of the product processing flow event log;

[0055] Figure 3 There are two levels of bidirectional sequence diagrams;

[0056] Figure 4 An example diagram of a bidirectional sequential graph convolutional network. DETAILED DESCRIPTION

[0057] The present invention will be further described below in conjunction with the accompanying drawings.

[0058] refer to Figure 1

[0059] The production business process activity prediction method based on a bidirectional sequential graph convolutional network includes the following steps:

[0060] 1) Converting the activity sequence in the event log of the production business process into a directed sequence graph structure includes the following steps:

[0061] 1.1) Constructing the activity layer, each event in the event log is regarded as an activity-level node, and the event dependencies between activity-level nodes constitute the edges of the activity layer;

[0062] 1.2) Construct the attribute layer, and take each attribute of the event, such as timestamp, executor, performance, etc., as a node at the attribute level. Each attribute forms an edge of the internal node of the attribute according to the event dependency relationship;

[0063] 1.3) Construct a forward sequence graph at the activity level, according to the execution order of events in each case in the event log, that is, arrange the activities in order from small to large timestamps, connect the activity nodes in sequence, and the direction of the edge is forward;

[0064] 1.4) Construct a reverse sequence graph at the activity level, according to the execution order of events in each case in the event log, that is, arrange the activities in descending order of timestamps, connect the activity nodes in sequence, and the direction of the edges is reverse;

[0065] 1.5) Construct a forward sequence graph at the attribute level. For each attribute of the event, arrange the attributes in the order of the execution of the events in each case, that is, in the order of timestamps from small to large, and connect the attribute nodes in sequence, with the direction of the edge pointing to the forward direction;

[0066] 1.6) Construct a reverse order graph at the attribute level. For each attribute of the event, arrange the attributes in descending order of timestamps according to the execution order of the events in each case, connect the attribute nodes in sequence, and point in the reverse direction of the edges;

[0067] 1.7) Use one-hot encoding to represent the activity nodes and attribute nodes of each attribute on the forward and reverse sequence diagrams at the activity level and attribute level, respectively, and obtain the activity node representation Attribute node representation

[0068] 1.8) Create the active node adjacency matrix A of the forward sequence graph 活动 and the attribute node adjacency matrix A 属性i , the active node adjacency matrix B of the reverse sequence graph 活动 and the attribute node adjacency matrix B 属性i .

[0069] 2) Design an activity-level graph convolution layer to process the forward sequence graph and the reverse sequence graph respectively, using the activity adjacency matrix A 活动 and B 活动 The information passing convolution operation on the active node is as follows:

[0070]

[0071] in, It is the active node representation of layer l of the forward sequence graph. It is the activity node representation of layer l of the reverse sequence diagram. refers to the learning weight parameter of the l-th layer of the forward sequential graph, refers to the learning weight parameter of the lth layer of the reverse sequence graph, and δ is the ReLU activation function. Through this convolution operation, the information of the active layer can be passed to the next layer through the adjacency matrix, thereby obtaining the representation of the active nodes of the next layer and

[0072] 3) Design the attribute-level graph convolution layer to process the forward sequence graph and reverse sequence graph of each attribute i respectively, using the attribute adjacency matrix A 属性i and B 属性i Information is transferred to attribute nodes, and the convolution operation is as follows:

[0073]

[0074] Among them, for the i-th attribute, It is the attribute node representation of the l-th layer of the forward sequence graph. It is the attribute node representation of the l-th layer of the reverse sequence graph. refers to the learning weight parameter of the l-th layer of the forward sequential graph, refers to the learning weight parameter of the lth layer of the reverse sequence graph, and δ is the ReLU activation function. Through this convolution operation, the information of the attribute layer can be passed to the next layer through the adjacency matrix, thereby obtaining the representation of the attribute nodes of the next layer and

[0075] 4) Introduce the attention mechanism to fuse the global feature representation of the forward sequence graph and the reverse sequence graph at the activity level. First, calculate the global feature representation of the forward sequence graph and the reverse sequence graph:

[0076]

[0077] in, is the global feature representation of the forward sequential graph at layer l, is the global feature representation of the reverse order graph at layer l, is the representation of the ith activity on the lth layer, is the representation of the i-th activity on the l-th layer.

[0078] Then, based on the global feature representation of the forward sequence graph and the reverse sequence graph, the attention weight α between the activity node i of the forward sequence graph and the activity node j of the reverse sequence graph at the activity level is calculated: ij :

[0079]

[0080] Among them, softmax is an activation function for category classification problems, which maps a set of values ​​to a numerical range (0,1) representing the probability distribution, and LeakyReLU is a neural network unit activation function. is a learnable weight vector, || represents a connection operation, represents the weight between the active node in the lth layer and all nodes in the l-1th layer on the forward sequence graph, represents the weight between the active node of the lth layer and all nodes of the l-1th layer on the reverse sequence graph, and They are the activity node representations of layer l of the forward sequence diagram and the reverse sequence diagram respectively.

[0081] Finally, the global feature representations of the activity-level forward and reverse sequence graphs are aggregated based on the attention weights:

[0082]

[0083] in, It is a feature representation that integrates the global information of the forward sequence graph and the reverse sequence graph of the lth layer. is the global feature representation of the forward sequential graph at layer l, It is the global feature representation of the reverse sequence graph at layer l.

[0084] 5) Design an attribute fusion layer to fuse the full-graph feature representations of multiple attributes in the attribute level. First, calculate the global feature representations of the forward sequence graph and the reverse sequence graph:

[0085]

[0086] Among them, for attribute j, is the global feature representation of the forward sequential graph on the lth layer, is the global feature representation of the reverse sequence graph on the lth layer, is the i-th attribute node representation of the forward sequence graph on the l-th layer, is the i-th attribute node representation of the reverse sequence graph on level l, and N is the number of nodes in the graph.

[0087] Next, the global feature representations of the forward sequence graph and the reverse sequence graph of the same attribute are concatenated and fused:

[0088]

[0089] in, It is an information representation of the global features of the forward sequence graph and the reverse sequence graph of the l-th layer attribute j.

[0090] 6) Introducing the attention mechanism based on the attribute fusion layer to integrate the full-image feature fusion representation of all attributes First, the attention weight α of the full-image feature fusion representation of each attribute is calculated j :

[0091]

[0092] Among them, α jis the attention weight of the full-image feature fusion representation vector of attribute j at layer l.

[0093] Then, the full-image feature fusion representation vectors of all attributes are weighted fused based on the attention weights:

[0094]

[0095] in, It is the information representation of the global features of the forward sequence graph and the reverse sequence graph of the lth layer attribute j. It is a feature representation that integrates the global information of all J attributes.

[0096] 7) Next activity prediction. First, the final global feature representations at the activity level and attribute level are concatenated and fused:

[0097]

[0098] in, It is a feature representation that integrates the global information of all J attributes. It is a feature representation that integrates the global information of the forward sequence graph and the reverse sequence graph of the lth layer. It is an information representation of the final global features that concatenates and fuses the activity level and attribute level of the lth layer.

[0099] Next, use the classification model to predict the next activity:

[0100]

[0101] Among them, T i is the predicted value of the corresponding activity i, W r is the weight matrix of the classification model, b r is the bias of the classification model.

[0102] Figure 2 The following is an example of an event log for a product processing flow. An event log consists of a case ID and attributes, including activities, timestamps, resources, and costs. An event log consists of multiple cases, each of which consists of multiple events, including activity names, timestamps, resources, and costs. For example, Figure 2Two cases in the event log are shown. The first case includes four events. The corresponding activities are "prepare materials", "assemble products", "quality inspection" and "packaging", and the corresponding timestamps are "2024-03-0110:00:00" (indicating that the execution time of the event is 10:00 on March 1, 2024), "2024-03-0110:20:21", "2024-03-0110:35:56", and "2024-03-0110:51:52". The corresponding resources are operator 1, operator 2, quality inspector 1 and operator 3, and the corresponding costs are 50, 40, 80 and 120 yuan respectively.

[0103] Figure 3 for Figure 2 Displays two levels of forward and reverse sequence diagrams corresponding to event logs. Among them, the forward and reverse sequence diagrams at the activity level only contain activity nodes. Figure 3 Taking Figure (a) as an example, both the forward and reverse sequence diagrams contain four activities: "prepare materials", "assemble products", "seal", and "pack". The direction of the forward sequence diagram is from left to right, and the direction of the reverse sequence diagram is from right to left. Next, for the attribute level, a separate forward and reverse sequence diagram is constructed for each attribute corresponding to the activity. Figure 3 Taking Figure (b) as an example, the forward and reverse sequence graphs of timestamp attributes, resource attributes, and cost attributes are constructed. The arrangement of attribute nodes and the activity arrangement order of the activity level are one-to-one corresponding. For example, the attribute nodes "2024-03-0110:00:00" corresponding to "prepare materials", "operator 1" and "50" are all located in the first position in their respective graphs.

[0104] Figure 4The example diagram of the bidirectional sequential graph convolutional network for predicting the next activity of the production business process is shown. The input of the network is the forward and reverse sequence graphs converted from the event log. For activities and attributes, they are divided into activity level and attribute level, and they are processed separately. At the activity level, the connection between activities forms a sequence graph, which can be divided into forward and reverse order according to the direction of the connection; similarly, at the attribute level, for each attribute corresponding to the activity, the connection between attribute nodes and attribute nodes can also form forward and reverse sequence graphs. First, the bidirectional sequence graphs at the activity level and attribute level are input into the graph convolution layers designed respectively to learn the representation of activity nodes and attribute nodes. The graph convolution layer at the activity level processes the bidirectional sequence graph part of the activity, learns the feature representation of nodes between adjacent nodes, and then uses the attention mechanism to weightedly fuse the feature information of the forward and reverse sequence graphs to obtain the final information representation at the activity level. The attribute-level graph convolution layer processes the bidirectional sequential graph of each attribute separately, and then preliminarily fuses the feature information of the forward and reverse sequential graphs of each attribute through feature splicing. Then, the information representation after feature splicing of all attributes is weighted and fused using the attention mechanism to obtain the final attribute-level information representation. In the attention mechanism, the model can adaptively focus on the importance of different nodes, which helps to better capture the relationship and features between nodes. Finally, the final activity-level information representation and attribute-level information representation are feature spliced ​​and input into the classification model for the prediction of the next activity. This method, which combines bidirectional sequential graphs, graph convolutional networks, and attention mechanisms, enables the model to fully mine the multi-level features in event logs and improves the accuracy and robustness of the next activity prediction.

[0105] The contents described in the embodiments of this specification are merely an enumeration of the implementation forms of the inventive concept. The protection scope of the present invention should not be regarded as limited to the specific forms described in the embodiments. The protection scope of the present invention also extends to equivalent technical means that can be conceived by those skilled in the art based on the inventive concept.

Claims

1. Production business process activity prediction method based on bidirectional sequential graph convolutional network, Features: The following steps are involved: 1) Convert the activity sequence in the event log of the production business process into a bidirectional sequence graph structure; 2) Design an activity-level graph convolution layer to process the forward sequence graph and the reverse sequence graph respectively, using the activity adjacency matrix A 活动 and B 活动 The information passing convolution operation on the active node is as follows: in, It is the active node representation of layer l of the forward sequence graph. It is the activity node representation of layer l of the reverse sequence diagram. refers to the learning weight parameter of the l-th layer of the forward sequential graph, refers to the learning weight parameter of the lth layer of the reverse sequence graph, and δ is the ReLU activation function; through this convolution operation, the information of the active layer can be passed to the next layer through the adjacency matrix, thereby obtaining the representation of the active nodes of the next layer and 3) Design the attribute-level graph convolution layer to process the forward sequence graph and reverse sequence graph of each attribute i respectively, using the attribute adjacency matrix A 属性i and B 属性i Information is transferred to attribute nodes, and the convolution operation is as follows: Among them, for the i-th attribute, It is the attribute node representation of the l-th layer of the forward sequence graph. It is the attribute node representation of the l-th layer of the reverse sequence graph. refers to the learning weight parameter of the l-th layer of the forward sequential graph, refers to the learning weight parameter of the lth layer of the reverse sequence graph, and δ is the ReLU activation function; through this convolution operation, the information of the attribute layer is passed to the next layer through the adjacency matrix, thereby obtaining the representation of the attribute nodes of the next layer and 4) Introduce the attention mechanism to fuse the global feature representations of the forward sequence graph and the reverse sequence graph at the activity level: First, calculate the global feature representations of the forward sequence graph and the reverse sequence graph: in, is the global feature representation of the forward sequential graph at layer l, is the global feature representation of the reverse order graph at layer l, is the representation of the ith activity on the lth layer, is the representation of the ith activity on the lth layer, and N represents the total number of activities; Next, based on the global feature representation of the forward sequence graph and the reverse sequence graph, the attention weight α between the activity node i of the forward sequence graph and the activity node j of the reverse sequence graph at the activity level is calculated. ij : Among them, softmax is an activation function for category classification problems, which maps a set of values ​​to a numerical range (0,1) representing the probability distribution, and LeakyReLU is a neural network unit activation function. is a learnable weight vector, || represents a connection operation, represents the weight between the active node in the lth layer and all nodes in the l-1th layer on the forward sequence graph, represents the weight between the active node of the lth layer and all nodes of the l-1th layer on the reverse sequence graph, and They are the activity node representations of the l-th layer of the forward sequence diagram and the reverse sequence diagram respectively; Finally, the global feature representations of the activity-level forward and reverse sequence graphs are aggregated based on the attention weights: in, It is a feature representation that integrates the global information of the forward sequence graph and the reverse sequence graph of the lth layer. is the global feature representation of the forward sequential graph at layer l, It is the global feature representation of the reverse order graph at layer l; 5) Design an attribute fusion layer to fuse the full-graph feature representations of multiple attributes in the attribute level: First, calculate the global feature representations of the forward sequence graph and the reverse sequence graph: Among them, for attribute j, is the global feature representation of the forward sequential graph on the lth layer, is the global feature representation of the reverse sequence graph on the lth layer, is the i-th attribute node representation of the forward sequence graph on the l-th layer, is the i-th attribute node representation of the reverse sequence graph on the l-th level, and N is the number of nodes in the graph; Next, the global feature representations of the forward sequence graph and the reverse sequence graph of the same attribute are concatenated and fused: in, It is the information representation of the global features of the forward sequence graph and the reverse sequence graph of the l-th layer attribute j; 6) Introducing the attention mechanism based on the attribute fusion layer to integrate the full-image feature fusion representation of all attributes First, the attention weight α of the full-image feature fusion representation of each attribute is calculated j : Among them, α j is the attention weight of the full-image feature fusion representation vector of attribute j at layer l, W (l) Represents the weights between all nodes in the lth layer and the l-1th layer of attribute j. Then, the full-image feature fusion representation vectors of all attributes are weighted fused based on the attention weights: in, It is the information representation of the global features of the forward sequence graph and the reverse sequence graph of the lth layer attribute j. It is a feature representation that integrates the global information of all J attributes; 7) Next activity prediction: First, the final global feature representations at the activity level and attribute level are concatenated and fused: in, It is a feature representation that integrates the global information of all J attributes. It is a feature representation that integrates the global information of the forward sequence graph and the reverse sequence graph of the lth layer. It is the information representation of the final global features that concatenates and fuses the activity level and attribute level of the lth layer; Next, use the classification model to predict the next activity: Among them, T i is the predicted value of the corresponding activity i, W r is the weight matrix of the classification model, b r is the bias of the classification model.

2. The production business process activity prediction method based on a bidirectional sequential graph convolutional network as claimed in claim 1, Features: The step 1) is specifically as follows: 1.1) Constructing the activity layer, each event in the event log is regarded as an activity-level node, and the event dependencies between activity-level nodes constitute the edges of the activity layer; 1.2) Construct the attribute layer, taking each attribute of the event as a node at the attribute level. Each attribute forms the edges of the internal nodes of the attribute according to the event dependency relationship; 1.3) Construct a forward sequence graph at the activity level, according to the execution order of events in each case in the event log, that is, arrange the activities in order from small to large timestamps, connect the activity nodes in sequence, and the direction of the edge is forward; 1.4) Construct a reverse sequence graph at the activity level, according to the execution order of events in each case in the event log, that is, arrange the activities in descending order of timestamps, connect the activity nodes in sequence, and the direction of the edges is reverse; 1.5) Construct a forward sequence graph at the attribute level. For each attribute of the event, arrange the attributes in the order of the execution of the events in each case, that is, in the order of timestamps from small to large, and connect the attribute nodes in sequence, with the direction of the edge pointing to the forward direction; 1.6) Construct a reverse order graph at the attribute level. For each attribute of the event, arrange the attributes in descending order of timestamps according to the execution order of the events in each case, connect the attribute nodes in sequence, and point in the reverse direction of the edges; 1.7) Use one-hot encoding to represent the activity nodes and attribute nodes of each attribute on the forward and reverse sequence diagrams at the activity level and attribute level, respectively, and obtain the activity node representation Attribute node representation 1.8) Create the active node adjacency matrix A of the forward sequence graph 活动 and the attribute node adjacency matrix A 属性i , the active node adjacency matrix B of the reverse sequence graph 活动 and the attribute node adjacency matrix B 属性i .