An industrial process multi-view trajectory construction method and system, and a storage medium

Abstract

This application discloses a method, system, and storage medium for constructing a multi-view trace graph of an industrial process, belonging to the field of industrial process mining technology. The method includes: acquiring event logs of the industrial process; based on the event logs, mining and generating an initial behavioral relationship graph reflecting the behavioral dependencies between activities, wherein the initial behavioral relationship graph has activities as nodes and behavioral dependencies between activities as edges; extracting multi-view attribute information associated with activities from the event logs, wherein the multi-view attribute information includes at least one or more of resource attributes, scale attributes, product attributes, and personnel attributes; embedding and mapping the multi-view attribute information to generate a multi-view fusion feature vector for each activity node; combining the initial behavioral relationship graph with the multi-view fusion feature vector to construct a multi-view trace graph, which is used to uniformly represent the interaction between the execution structure and multi-view attributes of the industrial process.

Description

Technical Field

[0001] This application relates to the field of industrial process mining technology, and in particular to a method, system and storage medium for constructing multi-view trace maps of industrial processes. Background Technology

[0002] Industrial process mining aims to extract process execution information from event logs to support process analysis and optimization. Traditional process representation methods are mainly based on the sequential order of activities, constructing directed graphs through direct follow-up relationships, and focusing only on control flow information.

[0003] However, in complex industrial processes, event logs not only contain activity sequence information but also record information from multiple perspectives, such as resources, scale, products, and personnel. Due to the complex interactions and dynamic dependencies between different perspectives, a single activity sequence cannot fully depict the entirety of process behavior. The execution time and subsequent paths of the same activity may vary significantly under different resources and production scales, and traditional graph structures based on activity sequences cannot reflect these differences.

[0004] To integrate information from multiple perspectives, existing research attempts to incorporate node attribute features into process representations, such as attaching resource and time labels to each activity node. However, these methods often employ simple attribute concatenation or weighted superposition, lacking structured modeling of the interactions between different attributes. Furthermore, the embedding spaces of information from different perspectives are inconsistent, and direct fusion can easily lead to information loss or the introduction of noise. Therefore, how to construct a graph structure that can uniformly represent the interaction relationships between the execution structure of industrial processes and multiple perspective attributes has become a pressing technical problem in this field.

[0005] Existing methods attempt to fuse multi-perspective features by embedding and encoding features such as activities, resources, and scale, and then using weighted methods or attention mechanisms within the model to achieve fusion, which can characterize the relationships between multiple attributes to some extent. However, these methods typically directly splice or weightedly superimpose multi-perspective information, lacking structured modeling of the interaction relationships between different attributes. Graph-based process modeling methods, on the other hand, construct flowcharts or behavioral dependency graphs and use graph neural networks to capture the dependencies between activities, resources, and scale, reflecting the topological structure features of information from different perspectives in the process. However, traditional graph models often focus on static dependencies and struggle to describe the non-stationary dynamic temporal changes in industrial processes.

[0006] In summary, existing technologies still have limitations in semantic interaction modeling and time-dependent dynamic capture of multi-perspective information: there is a lack of solutions for constructing a structured data model that uniformly represents the interaction between the execution structure and multi-perspective attributes of industrial processes. Currently, no effective solutions have been proposed to address the problems existing in these technologies. Summary of the Invention

[0007] Embodiments of this disclosure provide a method, system, and storage medium for constructing multi-view trace maps of industrial processes.

[0008] According to one aspect of the present disclosure, a method for constructing a multi-view trace graph of an industrial process is provided, comprising: acquiring an event log of an industrial process, the event log including multiple trajectories, each trajectory consisting of several events, each event including at least an activity identifier; based on the event log, mining and generating an initial behavioral relationship graph reflecting behavioral dependencies between activities, the initial behavioral relationship graph having activities as nodes and behavioral dependencies between activities as edges; extracting multi-view attribute information associated with activities from the event log, the multi-view attribute information including at least one or more of resource attributes, scale attributes, product attributes, and personnel attributes; embedding and mapping the multi-view attribute information to generate a multi-view fusion feature vector for each activity node; combining the initial behavioral relationship graph with the multi-view fusion feature vector to construct a multi-view trace graph, the multi-view trace graph being used to uniformly represent the interaction relationship between the execution structure and multi-view attributes of the industrial process.

[0009] According to another aspect of the present disclosure, a storage medium is also provided, the storage medium including a stored program, wherein the methods described above are executed by a processor when the program is running.

[0010] According to another aspect of the present disclosure, an industrial process multi-view trace graph construction system is also provided, comprising: an event log acquisition module, configured to acquire event logs of an industrial process, the event logs including multiple trajectories, each trajectory consisting of several events, each event including at least an activity identifier; a behavior relationship mining module, configured to mine and generate an initial behavior relationship graph reflecting the behavior dependencies between activities based on the event logs, the initial behavior relationship graph having activities as nodes and the behavior dependencies between activities as edges; a multi-view attribute extraction module, configured to extract multi-view attribute information associated with activities from the event logs, the multi-view attribute information including at least one or more of resource attributes, scale attributes, product attributes, and personnel attributes; a feature vector extraction module, configured to embed and map the multi-view attribute information to generate a multi-view fused feature vector for each activity node; and a multi-view trace graph construction module, configured to combine the initial behavior relationship graph with the multi-view fused feature vector to construct a multi-view trace graph, the multi-view trace graph being used to uniformly represent the interaction relationship between the execution structure and multi-view attributes of the industrial process.

[0011] The beneficial effects of this application are as follows:

[0012] (1) This application proposes a method for constructing a multi-perspective trace graph of an industrial process. An initial behavioral relationship graph is constructed by mining the behavioral dependencies between activities in the event log. Multi-perspective attribute information such as resources, scale, products, and personnel is embedded and mapped to generate a multi-perspective fusion feature vector. The two are then combined to obtain the multi-perspective trace graph. This trace graph can uniformly represent the execution structure of an industrial process and the interaction between multi-perspective attributes, overcoming the shortcomings of traditional methods that rely solely on activity sequences or simple attribute splicing, and providing a structured data foundation for subsequent process analysis tasks.

[0013] (2) In constructing the multi-view trace graph, this application introduces edge type labeling and edge weighting mechanisms. The edge type labeling is classified into direct following relationships, adjacent relationships, causal relationships, or parallel relationships based on the behavior footprint matrix, and the edge weights are calculated based on the co-occurrence frequency and average time interval of the activity pairs in the event log. These designs enable the constructed trace graph to explicitly express the properties and strengths of different behavioral dependencies, enhancing the graph structure's ability to characterize process semantics.

[0014] (3) The effectiveness of the method was verified by ablation experiments. The experimental results show that, compared with the model that removes graph structure constraints, the method proposed in this application reduces the MAE by 28.9% on the Production dataset and by 50.4% on the WMFD dataset; compared with the model that removes the time-aware mechanism, the complete method reduces the MAE by 17.7% and 33.2% on the two datasets, respectively. The above data fully demonstrate the complementary role of graph structure modeling and time-aware mechanism in improving prediction accuracy. The method of this application can maintain high prediction stability and generalization ability under different working conditions and data distributions. Attached Figure Description

[0015] The accompanying drawings, which are included to provide a further understanding of this disclosure and form part of this application, illustrate exemplary embodiments of this disclosure and are used to explain this disclosure, but do not constitute an undue limitation of this disclosure. In the drawings:

[0016] Figure 1 This is a flowchart of the method for constructing multi-view trace maps of industrial processes according to Embodiment 1 of this application;

[0017] Figure 2 This is an architecture diagram of the multi-view trace graph construction module according to Embodiment 1 of this application;

[0018] Figure 3 This is a flowchart of industrial process remaining time prediction according to Embodiment 1 of this application;

[0019] Figure 4 This is an overall framework diagram of industrial process remaining time prediction according to Embodiment 1 of this application;

[0020] Figure 5 This is a framework diagram of the Graph-Transformer remaining time prediction module according to Embodiment 1 of this application;

[0021] Figure 6 The figure shows the ablation experiment results of the Production dataset and WMFD dataset as described in Embodiment 1 of this application. Detailed Implementation

[0022] To enable those skilled in the art to better understand the technical solutions of this disclosure, the technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this disclosure, and not all embodiments. Based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this disclosure.

[0023] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this disclosure described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0024] Example 1

[0025] According to the first aspect of this embodiment, a method for constructing multi-view trace maps of industrial processes is provided. Figure 1 A flowchart illustrating the method is shown below. (Refer to...) Figure 1 As shown, the method includes steps S1 to S5. The following provides a detailed explanation of each step in conjunction with a specific application scenario. It should be noted that this embodiment uses a mechanical manufacturing process as an example for description, but the method of this application is also applicable to other industrial fields such as chemical, electronic, pharmaceutical, and automotive parts processing. The specific activity names and process structures can be adjusted according to the actual application scenario, and this application does not limit them.

[0026] Step S1: Obtain the event log of the industrial process. The event log includes multiple tracks, each track consists of several events, and each event includes at least an activity identifier.

[0027] In this embodiment of the invention, event logs are the fundamental data source for industrial process mining. In actual production, a company's Manufacturing Execution System (MES) or Enterprise Resource Planning System (ERP) records the execution status of each production task at each process step in real time. When it is necessary to construct a multi-view trace map, the original event logs are first obtained from these systems. The event log contains multiple traces, each trace corresponding to a complete production task instance, such as the entire process of a product from material input to finished product warehousing. Each trace consists of several events arranged in chronological order, and each event represents the occurrence of a process activity, such as "material cutting," "rough processing," "finishing," "heat treatment," and "quality inspection." Each event contains at least an activity identifier (e.g., process name or code) to distinguish different processes. In addition, events usually also include information such as timestamps, operators, and equipment numbers, but this information is not all required in step S1; only the presence of the activity identifier is necessary.

[0028] For example, a trajectory in a mechanical manufacturing process might contain the following sequence of events: [material feeding, rough machining, finish machining, heat treatment, quality inspection, packaging], where the activity identifiers for each event are "material feeding," "rough machining," "finish machining," "heat treatment," "quality inspection," and "packaging," respectively. The event log will record hundreds or thousands of such trajectories, reflecting the production history of different batches and product models.

[0029] This step allows us to obtain raw industrial event logs, providing a data foundation for subsequent extraction of behavioral dependencies and attribute information.

[0030] Step S2: Based on the event log, generate an initial behavior relationship graph that reflects the behavioral dependencies between activities. The initial behavior relationship graph uses activities as nodes and the behavioral dependencies between activities as edges.

[0031] In this embodiment of the invention, after acquiring the event log, it is necessary to analyze the behavioral dependencies between activities in all trajectories. Behavioral dependencies refer to the sequential, selection, parallel, and cyclic relationships between activities due to control flow logic. For example, in a machining process, "rough machining" is usually followed immediately by "finishing," which is a sequential dependency; for some product models, it may be necessary to choose between "heat treatment" or "surface treatment" after "finishing," which is a selection dependency; if "quality inspection" finds that the product is unqualified, it may return to "rough machining" for reprocessing, which is a cyclic dependency; in addition, some processes (such as "cleaning" and "inspection") can be performed simultaneously, which is a parallel dependency.

[0032] To automatically extract these dependencies, this step utilizes process mining algorithms (such as Inductive Miner or the alpha algorithm) to mine event logs. This algorithm statistically analyzes the order of occurrence and co-occurrence frequency of activity pairs across all trajectories to determine if direct follow-up or causal relationships exist between each pair. Based on the mining results, a directed graph, called the initial behavioral relationship graph, is constructed. In this graph, each node represents an activity (e.g., "rough processing" or "finishing"), and each directed edge represents a behavioral dependency between two activities. For example, an edge from "rough processing" to "finishing" indicates that "finishing" usually occurs after "rough processing." This graph reflects the "skeleton" or "rulebook" of the industrial process, i.e., the static execution template of the process, independent of specific production instances.

[0033] For example, based on thousands of historical trajectories of a manufacturing process, the initial behavioral relationship graph might contain the following edges: material input → rough machining, rough machining → finish machining, finish machining → heat treatment, finish machining → surface treatment, heat treatment → quality inspection, surface treatment → quality inspection, quality inspection → packaging, quality inspection → rough machining (rework cycle). This graph fully depicts the basic structure of the manufacturing process, providing a topological framework for subsequent attribute fusion.

[0034] This step allows for the automatic extraction of control flow dependencies between activities from event logs, the construction of an initial behavior relationship graph, and the provision of a structured skeleton for subsequent integration of multi-perspective attributes.

[0035] Step S3: Extract multi-perspective attribute information associated with the activity from the event log. The multi-perspective attribute information includes at least one or more of the following: resource attributes, scale attributes, product attributes, and personnel attributes.

[0036] In this embodiment of the invention, the initial behavioral relationship diagram only describes the control flow relationships between activities, but does not include the dynamic attribute information attached to each activity during execution. This multi-perspective attribute information is also recorded in each event of the event log, and therefore needs to be extracted from the original log. Specifically, the multi-perspective attribute information includes at least one or more of the following: resource attributes, scale attributes, product attributes, and personnel attributes. Resource attributes refer to the physical resources consumed or used in performing the activity, such as "machining center A" or "heat treatment furnace B". Scale attributes refer to the production quantity or workload of the batch, such as "1000 pieces" or "5 tons". Product attributes refer to the type or model of the product being processed, such as "type A workpiece" or "type B workpiece". Personnel attributes refer to the identification of the worker or technician operating the process, such as "operator A" or "senior technician B". These attributes have a significant impact on the execution time, resource load, and quality results of the process. For example, in the same "finishing" process, using "machining center A" is more efficient than "machining center B"; processing "type A workpieces" takes longer than processing "type B workpieces"; and operation by "senior technicians" can reduce rework rates.

[0037] This step involves iterating through each event in the event log and extracting the aforementioned attribute values ​​for each activity node. For example, the activity node "Fine Processing" might be associated with different resources (e.g., "Equipment #5" or "Equipment #9"), different scales (e.g., "500 pieces" or "2000 pieces"), different product models (e.g., "Type A" or "Type B"), and different operators (e.g., "A" or "B") in different batches. These extracted attribute values ​​will serve as the raw input for subsequent embedding mapping.

[0038] This step allows us to extract multi-perspective attribute information related to each activity from the event log, such as resources, scale, products, and personnel, providing a data source for embedding the attribute information into graph nodes later.

[0039] Step S4: Embed and map the multi-view attribute information to generate a multi-view fusion feature vector for each active node.

[0040] In this embodiment of the invention, the multi-view attribute information extracted in step S3 is usually discrete categorical data (e.g., resource ID, product model, worker name) or numerical data (e.g., scale and quantity), while computer models require continuous vectors for mathematical operations. Therefore, this step needs to convert these discrete or numerical information into continuous vectors, i.e., embedding vectors. Specifically, firstly, an independent embedding matrix is ​​constructed for each attribute (resource, scale, product, personnel). The number of rows in each embedding matrix corresponds to the number of different values ​​for that attribute, and the number of columns is a preset embedding dimension. Then, for each activity node, the corresponding embedding vector is retrieved from the corresponding embedding matrix based on its associated resource attribute value, scale attribute value, product attribute value, and personnel attribute value. For example, if the resource associated with the "rough processing" node is "equipment #5", then the vector row corresponding to "equipment #5" is retrieved from the resource embedding matrix. After obtaining the embedding vectors of each attribute, they need to be fused into a unified multi-view fusion feature vector.

[0041] This step transforms the embedding vectors of all attributes to the same dimensional space through linear transformation, then concatenates them, and finally maps them through a non-linear activation function (such as ReLU) to obtain a fixed-length fused feature vector. This vector integrates the semantic information of the activity node across all multi-view attributes, reflecting the potential behavioral characteristics of the activity under different resource, scale, product, and personnel combinations.

[0042] For example, in a certain batch, the resource associated with the "rough processing" node is "equipment #5", the scale is "1000 pieces", the product is "type A workpiece", and the personnel is "B". After embedding mapping and fusion, a 128-dimensional fused feature vector is obtained. This vector numerically encodes the efficiency of "equipment #5", the batch effect of "1000 pieces", the processing difficulty of "type A workpiece", and the skill level of "B". Even if the activities are the same in different batches, as long as the multi-view attributes are different, their fused feature vectors will also be different, thus realizing the quantitative representation of attribute differences.

[0043] This step transforms discrete, multi-perspective attribute information into a continuous, unified-dimensional fused feature vector, enabling the attribute information to be effectively processed by computer models while preserving the semantic relationships between different attributes.

[0044] Step S5: Combine the initial behavior relationship graph with the multi-view fusion feature vector to construct a multi-view trace graph, which is used to uniformly represent the interaction relationship between the execution structure of the industrial process and the multi-view attributes.

[0045] In this embodiment of the invention, after completing the initial behavioral relationship graph (describing the structure between activities) and the multi-view fusion feature vector (describing the attribute semantics of each activity node), this step organically combines the two to form a unified multi-view trace graph. Specifically, each activity node in the initial behavioral relationship graph is assigned the multi-view fusion feature vector generated in step S4 as its attribute feature. Simultaneously, the original edge structure (i.e., the behavioral dependencies between activities) in the graph is preserved. After combination, the multi-view trace graph contains both the topological structure of the process and rich multi-view semantic information at each node. This multi-view trace graph is used to uniformly represent the interaction between the execution structure and multi-view attributes of an industrial process.

[0046] It should be noted that, in this invention, the term "execution structure" refers to the inherent behavioral dependencies between various activities in an industrial process, determined by control flow logic. These dependencies include, but are not limited to, sequential, selection, parallel, cyclic, and causal relationships. This execution structure is explicitly modeled using an initial behavioral relationship graph. "Multi-perspective attributes" refer to dynamic information such as resources, scale, products, and personnel. There is a two-way interaction between the two: the execution structure is influenced by multi-perspective attributes (e.g., large-scale batches may trigger parallel branches), and multi-perspective attributes are constrained by the execution structure (e.g., resource consumption accumulates in a cyclic structure). The multi-perspective trace graph explicitly preserves this interaction by unifying structure and attributes within the same graph framework, providing a rich and structured data foundation for subsequent process analysis (such as remaining time prediction, root cause analysis, and simulation optimization).

[0047] For example, by combining the initial behavioral relationship graph of a manufacturing process with the fused feature vector of each node, a multi-view trace graph is obtained. In this multi-view trace graph, the "finishing" node not only knows that it may be connected to "heat treatment" or "surface treatment" afterward, but also knows the resource, scale, product, personnel, and other attribute information of "finishing" in the current batch. Downstream prediction models can directly use this graph for message passing, thereby more accurately predicting the remaining processing time.

[0048] This step allows structural and attribute information to be integrated into a single graph structure, creating a multi-view trace graph. This enables a unified representation of the interaction between the industrial process execution structure and multi-view attributes, providing a structured and computable data foundation for subsequent tasks such as process prediction and root cause analysis.

[0049] As described in the background section, in complex industrial processes, event logs not only contain activity sequence information but also record information from multiple perspectives, including resources, scale, products, and personnel. Due to the complex interactions and dynamic dependencies between different perspectives, a single activity sequence cannot fully depict the entirety of the process behavior. The execution time and subsequent paths of the same activity may differ significantly under different resources and production scales, and traditional graph structures based on activity sequences cannot reflect these differences.

[0050] In view of this, this application constructs a multi-view trace graph that can uniformly represent the interaction between the execution structure and multi-perspective attributes of an industrial process. Specifically, this application first uses a process mining algorithm to extract behavioral dependencies between activities from event logs, forming an initial behavioral relationship graph, thus preserving the topological structure of the process. Simultaneously, it extracts multi-perspective attribute information such as resources, scale, products, and personnel from the event logs, and generates a multi-perspective fusion feature vector for each activity node through embedding mapping, achieving a continuous and unified representation of heterogeneous attribute information. Finally, it combines the structure graph with the node feature vectors to obtain the multi-perspective trace graph. This multi-perspective trace graph explicitly expresses the control flow dependencies such as sequence, selection, parallelism, and looping between activities, and also integrates the semantic features of each activity under different combinations of resources, scale, products, and personnel, thus comprehensively depicting the bidirectional interaction between the "execution structure" and "multi-perspective attributes" in an industrial process.

[0051] Therefore, this application effectively solves the technical problems existing in the prior art, namely, "a single activity sequence is difficult to fully depict the overall process behavior", "there is a lack of structured modeling of the interaction relationship between different attributes", and "traditional graph models focus more on static dependencies and are difficult to describe dynamic temporal changes". Compared with methods that only use activity sequences or simple attribute splicing, the multi-view trace graph constructed in this application provides a structured, computable, and semantically rich data foundation for subsequent process analysis tasks (such as remaining time prediction, root cause analysis, and simulation optimization), significantly improving the model's ability to depict complex industrial processes.

[0052] Optionally, the operation of generating an initial behavioral relationship graph reflecting behavioral dependencies between activities based on the event log includes: mining the event log using a process mining algorithm to obtain a corresponding Petri net model; constructing a behavioral footprint matrix based on the behavioral footprint theory and the Petri net model, wherein the behavioral footprint matrix is ​​used to distinguish between direct follow relationships, adjacent relationships, causal relationships, parallel relationships, and irrelevant relationships between activities; constructing a behavioral relationship matrix by setting the matrix elements corresponding to irrelevant relationships between activities as first values ​​and the matrix elements corresponding to the other four types of relationships as second values, based on the behavioral relationship matrix; and constructing the initial behavioral relationship graph by using activities as nodes and activity pairs with corresponding matrix elements of the second value as edges, based on the behavioral relationship matrix.

[0053] In this embodiment of the invention, considering that in complex industrial processes, event logs not only contain activity sequence information but also record multi-perspective information such as resources, scale, products, and personnel, a single activity sequence cannot fully depict the entirety of the process behavior due to the complex interactions and dynamic dependencies between different perspectives. To address the above problems, this application proposes a multi-perspective trace graph construction module, such as... Figure 2 As shown, this module first utilizes the Inductive Miner algorithm to mine industrial event logs, extracting a Petri net model of the logs. Then, based on behavioral footprint theory, it constructs a behavioral footprint matrix of activities in the logs, and further builds a behavioral relationship matrix to extract control flow relationships between process activities, constructing an initial relationship graph with behavioral dependency constraints. Subsequently, multi-perspective information of activities is embedded and mapped to form graph node feature representations, achieving a structured expression of multi-perspective information. Finally, a hierarchical integration mechanism maps dependencies from different perspectives to a unified semantic space, generating a multi-perspective trace graph reflecting the process execution structure and attribute interaction patterns. This module effectively integrates multi-perspective information such as activities, resources, and processing scale, preserving the topological features and semantic context dependencies of the process, providing a structured and computable data foundation for subsequent tasks such as process prediction and root cause analysis.

[0054] The process of generating the initial behavioral relationship diagram will be explained in detail below.

[0055] Let the event log be Each trajectory Composed of several events, namely ,in, , Represents the event log The number of trajectories, Representing the trajectory The number of events in the event. Each event ,in, , Representative events The activity logo, Representative events Resource attributes, Representative events timestamp, Representative events Load or other properties. By iterating through the event log. All trajectories are used to obtain the activity set. ,in The total number of all activities.

[0056] Then, the Inductive Miner algorithm is used to perform process mining on the event log L, obtaining the corresponding Petri net model. Inductive Miner is a process mining algorithm based on the divide-and-conquer approach, capable of automatically discovering process models from the event log and ensuring that the model's behavior is consistent with the log. This algorithm recursively partitions the event log, detects patterns such as sequential, selection, parallel, and cyclical patterns, and ultimately generates a Petri net. This Petri net includes places, transitions, and arcs, visually representing the control flow relationships between activities. For example, for the wheel manufacturing log, Inductive Miner might generate a Petri net with a structure of "material input → rough turning → finish turning → (parallel heat treatment and polishing) → inspection → (packaging or rework)".

[0057] Next, based on behavioral footprint theory, a behavioral footprint matrix is ​​constructed according to the Petri net model. Among them, behavioral footprint theory defines five relationships between any two activities by analyzing their occurrence patterns in logs and models. For any two activities... and (All belong to activity set A), behavioral footprint matrix The definition is as follows:

[0058] ;

[0059] Specifically, when activity b occurs directly after activity a, activities a and b have a direct following relationship, indicating that there is a trajectory that causes b to occur immediately after a; when activities a and b are adjacent in the Petri net model and there are no other activities in between, activities a and b have an adjacent relationship; when the occurrence of activity a leads to the occurrence of activity b, activities a and b have a causal relationship; when activities a and b can be executed simultaneously and are independent of each other, activities a and b have a parallel relationship; when there is no direct control flow dependency between activities a and b, activities a and b have an unrelated relationship.

[0060] Furthermore, this behavioral footprint matrix fully records the behavioral patterns between each pair of activities. For example, if "rough work" is always followed directly by "finishing work," then... If heat treatment and polishing can be performed simultaneously, then If "feeding" and "packaging" are not directly related, then .

[0061] Next, a behavioral relationship matrix is ​​constructed based on the behavioral footprint matrix. The behavior relation matrix is ​​a binary simplification of the behavior footprint matrix. For any two activities a and b, if the behavior footprint matrix shows an unrelated relationship (a ≠ b), then... Setting it to the first value (e.g., 0) indicates that there is no edge connecting a and b; if it is any of the other four relationships (direct following, adjacent, causal, parallel), then... Setting it to the second value (e.g., 1) indicates the existence of a behavioral dependency. The corresponding behavioral relationship matrix can be constructed using the formula below. :

[0062] ;

[0063] This matrix simplifies the complex five relationships into binary connection indicators, facilitating the construction of graph structures. For example, the unrelated relationship between "feeding" and "packaging" corresponds to... The direct following relationship between "rough cart" and "refined cart" corresponds to... .

[0064] Next, based on the behavior relationship matrix, an initial behavior relationship graph is constructed. That is, activities are nodes, and pairs of activities with a value of 1 in the behavior relation matrix are used as edges to construct a directed graph. Specifically:

[0065] Node set Each node One of the corresponding activities in activity set A .

[0066] edge set That is, for each pair of activities with a value of 1 in the behavior relationship matrix At the node and Establish a directed edge between them. This represents the nodes in the behavior relationship matrix constructed based on event log L. and Behavioral dependency indicator value between nodes: if node and If it is an unrelated relationship in event log L, then The value is 0; otherwise The value is 1.

[0067] The direction of the edge is determined based on the relationship direction in the behavior footprint matrix: for direct following relationships (a>b) and causal relationships (a→b), the direction is from a to b; for adjacent relationships (a>b), the direction is from a to b. b), usually following the actual flow sequence or establishing bidirectional edges; for parallel relationships (a) (b) Bidirectional edges can be established, or unidirectional edges can be established according to the time sequence of the first occurrence of the activities. For example, if If, in the initial behavior graph, there exists a directed edge from the "coarse-car" node to the "fine-car" node; if Then a bidirectional edge can be established.

[0068] By following the four steps above, an initial behavioral relationship graph can be automatically generated from the original event log. This graph explicitly represents the control flow structure of the industrial process in the form of a directed graph, including complex dependency patterns such as sequence, selection, parallelism, and loops, providing a structured topological skeleton for subsequent fusion of multi-perspective attribute information.

[0069] Therefore, this application can automatically and accurately construct an initial behavioral relationship graph reflecting the behavioral dependencies between activities from event logs. Compared with traditional methods that only use activity sequences, this graph explicitly preserves multiple control flow relationships, overcoming the problem that sequence models are insufficient in modeling parallel and loop structures, and providing a structured and computable data foundation for subsequent multi-view trace graph construction and tasks such as process prediction and root cause analysis.

[0070] Optionally, when constructing the multi-view trace graph, the method further includes: assigning an edge type label to each edge in the initial behavior relationship graph according to the behavior footprint matrix, wherein the edge type label is a direct follow relationship, an adjacent relationship, a causal relationship, or a parallel relationship; forming an edge type set from the edge type labels of each edge, and incorporating the edge type set into the multi-view trace graph.

[0071] In this embodiment of the invention, each edge in the initial behavioral relationship graph only represents a certain behavioral dependency between two activities, but does not distinguish the specific type of dependency. However, different types of dependencies (such as sequential, parallel, and causal) have different meanings for subsequent process analysis. For example, in the remaining time prediction task, sequential dependency means that activities are executed in sequence, and the time is accumulated; while parallel dependency means that activities can be performed simultaneously, and the time is maximized or overlaps. Therefore, this application further assigns an edge type label to each edge to explicitly express the semantics of the dependency.

[0072] Specifically, based on the behavioral footprint matrix constructed in step S2 This allows us to determine the specific type of relationship between any two activities. For each edge in the initial behavior relationship graph... Query the corresponding element in the behavior footprint matrix Value:

[0073] like If the relationship is direct, then set the type label of that edge to "direct follow relationship";

[0074] like If the relationship is adjacent, then set it to "adjacent".

[0075] like (Cause and effect relationship), then set it as "cause and effect relationship";

[0076] like If it is a parallel relationship, then it is set as a "parallel relationship".

[0077] Note the irrelevant relationship ( Since there are no corresponding edges in the initial behavior graph, no labels are needed.

[0078] For example, in the initial behavior relationship diagram of a mechanical manufacturing process, the edge from "rough machining" to "finishing" corresponds to a direct follow relationship in the behavior footprint matrix, so the type label for this edge is "direct follow relationship". The edge from "quality inspection" to "rough machining" corresponds to a rework cycle, which is usually a causal relationship, and the label is set to "causal relationship". If there is a bidirectional edge between "heat treatment" and "surface treatment", and it is marked as a parallel relationship in the behavior footprint matrix, then the label for both edges is "parallel relationship".

[0079] Furthermore, in order to utilize edge type information in subsequent graph neural networks or attention mechanisms, this application converts discrete edge type labels into continuous embedding vectors. First, an edge type set is defined. Based on the actual needs of behavioral dependence, It can include direct follow relationships, immediate adjacency relationships, causal relationships, and parallel relationships. For simplicity, direct follow relationships and immediate adjacency relationships can also be considered as sequential relationships, but this embodiment retains four types to obtain finer-grained semantics. For example, the set of edge types is defined as follows: ,in Represents sequential relationships (including direct following and adjacent relationships). Represents a cause-and-effect relationship. This represents a parallel relationship. In practical implementation, it can be extended as needed.

[0080] Then, construct the edge type embedding matrix. ,in The number of behavioral relationship types. The embedding dimension for edge types is specified (e.g., 32-dimensional). Each row of this matrix corresponds to an embedding representation for one edge type, which is initially randomly initialized and subsequently learned and optimized along with other parameters during model training.

[0081] For the initial behavior relationship diagram Each edge in Let its behavioral relationship type be , ∈ Then, a lookup table operation is used to map this type to an embedding vector. :

[0082] ;

[0083] That is, the vector corresponding to the row is taken from the embedding matrix and used as the type embedding representation of the edge.

[0084] By collecting all the edge type embedding vectors, a set of edge type embeddings can be formed, which, along with the set of edge type labels, can be incorporated into the multi-view trace graph. In subsequent encoders (such as Graph-Transformer), these embedding vectors can directly participate in the calculation of attention weights, enabling the model to perceive the semantic differences in dependencies between different behaviors.

[0085] After assigning a type label to each edge, all edge type labels are collected to form an edge type set T. Each element in this set corresponds to the type information of one edge. Finally, this edge type set is incorporated into the multi-view trace graph as a component of the graph. The multi-view trace graph can be fully represented as follows: , where T is the set of edge types. By introducing the set of edge types, the multi-view trace graph not only records whether there are dependencies between activities, but also clarifies the specific properties of the dependencies, providing richer semantic information for downstream tasks.

[0086] Therefore, this application assigns a type label based on the behavior footprint matrix to each edge in the initial behavior relationship graph and incorporates the edge type set into a multi-view trace graph, enabling the constructed trace graph to explicitly express the semantic differences of different behavioral dependencies (such as direct following, adjacent, causal, and parallel). This design enhances the graph structure's ability to characterize the semantics of industrial process control flow, providing a more refined structured data foundation for subsequent tasks such as process prediction and root cause analysis.

[0087] Optionally, when constructing the multi-view trace graph, the method further includes: assigning weights to the edges in the initial behavior relationship graph, wherein the weights are calculated based on the co-occurrence frequency and / or average time interval of the activity pairs in the event log; constructing an edge weight matrix from the weights of each edge, and incorporating the edge weight matrix into the multi-view trace graph.

[0088] In this embodiment of the invention, the edges in the initial behavioral relationship graph only represent behavioral dependencies between activities, but the strength of the dependency between different activity pairs may vary significantly. For example, in industrial processes, "rough processing" is almost always followed by "finishing processing," indicating a very strong dependency between the two; while "finishing processing" may sometimes lead to "heat treatment" and sometimes to "surface treatment," indicating a relatively weaker dependency. To quantify this dependency strength, this application assigns a weight to each edge in the graph, which comprehensively considers the co-occurrence frequency of activity pairs in the event log and the average time interval between them, to reflect the strength and stability of the dependency between activities.

[0089] Specifically, for an edge in the initial behavior graph The corresponding activity pair is Edge weight The calculation method is as follows:

[0090] ;

[0091] in, Indicates activity and Normalized frequencies of activity pairs appearing in the same trajectory according to behavioral relationships. Specifically, statistical analysis of activity pairs in all trajectories. According to the direction of the edge (i.e. from) arrive The normalized frequency is obtained by dividing the number of occurrences of each activity pair by the total number of occurrences of all activity pairs. This frequency reflects the prevalence of dependencies between activities; a higher frequency indicates a more stable dependency.

[0092] Indicates activity and The average time interval in the log. For each activity pair appearing in the track. Calculate from Completed The initial time difference (e.g., in minutes or hours) is then averaged across all trajectories to obtain... Shorter time intervals indicate tighter connections and stronger dependencies between activities; longer time intervals indicate looser dependencies. To convert time intervals into positive weight contributions, this application employs an exponential decay function. ,in This is a hyperparameter used to adjust for the influence of time information (usually taken as a positive number). Thus, the smaller the average time interval, the closer the exponential term is to 1; the larger the interval, the closer the exponential term is to 0.

[0093] This is a hyperparameter used to balance frequency and time information, and its value ranges from [0,1]. When... When = 1, the weight is determined solely by the co-occurrence frequency; when When = 0, the weight is determined only by the time interval; when When the coefficient is 0.5, both contribute equally. In practical applications, the appropriate coefficient can be selected based on the characteristics of the specific industrial scenario through methods such as cross-validation. value.

[0094] For example, in the event log of a certain manufacturing process, the activity pair "roughing → finishing" appears with a frequency of 0.95 across all trajectories (i.e., finishing is immediately followed by roughing in 95% of the trajectories), with an average time interval of 2 minutes; while the activity pair "finishing → heat treatment" appears with a frequency of 0.6, with an average time interval of 10 minutes. =0.7, =0.1, then the weight of "rough machining → finish machining" is The weight of "finishing → heat treatment" is... It is evident that the former has a higher weight, indicating a stronger dependency.

[0095] Calculate the weight for each edge in the initial behavioral relationship graph. Then, the weights of all edges are arranged in node order to form an edge weight matrix W. The size of this matrix is... (or A sparse matrix is ​​used, with each element corresponding to the weight of an edge. Finally, this edge weight matrix is ​​incorporated into the multi-view trace graph as a component of the graph. Combining the aforementioned set of edge types, the multi-view trace graph can be represented as... , where W is the edge weight matrix.

[0096] Therefore, this application assigns weights to the edges in the initial behavioral relationship graph that comprehensively consider co-occurrence frequency and average time interval, and incorporates these weights into a multi-view trace graph, enabling the constructed trace graph to quantify the strength of dependencies between activities. This design not only reflects the universality (frequency) of dependencies but also characterizes the tightness (time interval) of dependencies, providing a more granular structured data foundation for subsequent tasks such as process prediction and root cause analysis. For example, in remaining time prediction, high-weighted edges imply more certain and rapid transitions between activities, allowing the model to adjust time estimates accordingly.

[0097] Optionally, the operation of embedding and mapping the multi-view attribute information to generate a multi-view fusion feature vector for each active node includes: constructing an independent embedding matrix for each attribute contained in the multi-view attribute information; using the embedding matrix to map the discrete values ​​of the corresponding attributes into embedding vectors; transforming the embedding vectors of all attributes to the same dimensional space through a linear transformation; concatenating the transformed embedding vectors of each attribute and mapping them through a nonlinear activation function to generate the multi-view fusion feature vector.

[0098] In this embodiment of the invention, step S3 extracts multi-perspective attribute information associated with each activity from the event log, including resource attributes, scale attributes, product attributes, and personnel attributes. These attributes are typically discrete categorical data (such as resource ID "equipment #5", product model "type A workpiece", and personnel name "Li Si") or numerical data (such as scale quantity "1000 pieces"), while subsequent graph neural networks or prediction models require continuous vectors as input. Therefore, after completing the initial behavioral relationship graph construction, feature extraction is performed on node attributes to further integrate multi-perspective information from the industrial process event log. Specifically, for the resource, scale, product, and personnel attributes of each node, this application uses an embedding mapping method to map discrete attributes to a continuous vector space. For resource, scale, product, and personnel attributes, corresponding embedding matrices are constructed respectively, mapping each type of attribute to a low-dimensional dense vector representation to capture potential semantic features from different perspectives.

[0099] Because different attributes have different semantic spaces and value ranges, direct mixing can lead to information confusion. This application constructs an independent embedding matrix for each attribute, mapping discrete attribute values ​​to continuous vectors. Specifically:

[0100] Resource attributes reflect the differences in operational capabilities, workload, and execution efficiency among different execution units. For resource attributes, the resource embedding matrix is ​​defined as follows: Where |M| is the total number of different resources in the event log. This represents the dimension of the resource embedding vector (e.g., 16 dimensions). Each row of this matrix corresponds to an embedding representation of a specific resource.

[0101] The scale attribute characterizes the number of processes or the amount of work in a process instance, and has a significant impact on process duration and resource load. For the scale attribute, the scale embedding matrix is ​​defined as follows: , where |S| is the number of different size levels (or the number of buckets after discretizing continuous size values). Embedded dimensions for scale.

[0102] Product attributes reflect the differences in process paths and execution times among different processed objects. For product attributes, the product embedding matrix is ​​defined as follows: Where |P| represents the number of product types. Embed dimensions for products.

[0103] Personnel attributes are used to characterize the impact of differences in operator experience on process execution efficiency. For personnel attributes, the personnel embedding matrix is ​​defined as follows: ,in For the number of operators, Embedded dimensions for personnel.

[0104] The initial values ​​for these embedding matrices can be randomly generated or pre-trained word embeddings (such as those from Word2Vec) can be used and then optimized along with other parameters during model training. The number of rows in each matrix is ​​determined based on the number of different values ​​appearing in the actual logs, and the number of columns (embedding dimension) is a hyperparameter that can be selected based on task complexity and computational resources.

[0105] Step 2: Use the embedding matrix to map the discrete values ​​of the corresponding attributes into embedding vectors.

[0106] For each activity node Based on the resource value associated with it in the event log. Scale value Product value and personnel value The corresponding embedding vectors are found from the corresponding embedding matrices.

[0107] Resource embedding vector: That is, to take the resources in the resource embedding matrix. The corresponding row vector.

[0108] Scale embedding vector: .

[0109] Product embedding vector: .

[0110] Person embedding vector: .

[0111] For example, if the resource associated with the "rough processing" node is "equipment #5", the scale is "1000 pieces", the product is "type A workpiece", and the personnel is "Li Si", then four embedding vectors are extracted, each with a fixed length (e.g., 16-dimensional, 8-dimensional, 16-dimensional, 16-dimensional). These vectors each encode the semantic information of the corresponding attribute, but their dimensions may be different, and they cannot be directly concatenated.

[0112] Step 3: Transform the embedding vectors of all attributes to the same dimensional space through linear transformation.

[0113] Due to different embedding dimensions of attributes They may be inconsistent and need to be mapped to the same common dimensional space d (e.g., 128-dimensional). This application learns a linear transformation matrix for each attribute separately. and bias terms Refer to the following formula:

[0114] ;

[0115] in, These are the aligned embedding vectors, all with dimension d. This linear transformation can be achieved using fully connected layers, with parameters trained along with the rest of the model. This step unifies the embedding vectors of all attributes into the same feature space, facilitating subsequent fusion.

[0116] Step 4: Concatenate the embedded vectors of each attribute after conversion, and generate the multi-view fusion feature vector by mapping through a non-linear activation function.

[0117] After obtaining the four embedding vectors with unified dimensions, this application concatenates them along the feature dimension to form a longer vector:

[0118] ;

[0119] in, This represents a vector concatenation operation. It is the ReLU activation function. and For fusion layer parameters, This is the dimension-aligned embedding vector representation of resource attributes. This is the dimension-aligned embedding vector representation of the scale attribute. This is the dimension-aligned embedding vector representation of product attributes. This layer represents the dimension-aligned embedding vectors of personnel attributes. Through this layer, the model can learn higher-order interactions between different attributes, further compressing the concatenated vectors into a final multi-view fused feature vector. .

[0120] The fused feature vector Integrating event nodes Semantic information across all multi-perspective attributes can reflect the potential behavioral characteristics of an activity under different resource, scale, product, and personnel combinations. For example, for the "rough processing" node, its fused feature vector might encode the high efficiency of "equipment #5," the large-batch effect of "1000 pieces," the processing difficulty of "Type A workpieces," and the high skill level of "Li Si." Even if the activity is the same in different batches, the fused feature vector will differ as long as the multi-perspective attributes are different, thus achieving a quantitative representation of attribute differences.

[0121] Finally, the above operations are performed on all active nodes to obtain the multi-view fused feature vector of each node, which constitutes the node feature matrix. Where n is the number of active nodes. This represents the dimension of the multi-view fusion feature vector. This matrix will be combined with the initial behavior relationship graph to form a multi-view trace map.

[0122] Therefore, this application constructs an independent embedding matrix for each multi-perspective attribute, mapping discrete attribute values ​​to embedding vectors. Then, through linear transformation to unify dimensions, concatenation and fusion, and nonlinear mapping, it generates a multi-perspective fused feature vector for each activity node. This process preserves the independent semantics of each attribute while achieving information fusion between heterogeneous attributes. This enables subsequent models to simultaneously utilize multi-perspective information such as resources, scale, products, and personnel for process analysis and prediction, effectively overcoming the information loss problem caused by simple concatenation or weighted superposition in traditional methods.

[0123] To further illustrate the application of the industrial process multi-view trace graph construction method proposed in this embodiment in specific downstream tasks, the following provides specific implementation steps and process descriptions for applying the industrial process multi-view trace graph construction method to the downstream industrial process remaining time prediction task. Figure 3 A schematic diagram illustrating the prediction of remaining time in an industrial process is shown. Figure 4 A framework diagram for predicting remaining time in industrial processes is shown. (Reference) Figure 3 and Figure 4 As shown, the prediction of remaining time for industrial processes includes:

[0124] Step 1: Obtain the multi-view trace map that has been constructed in advance using the industrial process multi-view trace map construction method described in this embodiment.

[0125] Specifically, prior to the prediction phase, a multi-view trace map has been constructed based on historical event logs according to the industrial process multi-view trace map construction method described in this embodiment. This multi-view trace map contains static knowledge of the industrial process: behavioral dependencies between activities (sequence, selection, parallelism, looping, etc.), and a multi-view fused feature vector attached to each activity node (integrating information such as resources, scale, product, and personnel). When remaining time prediction is required, this pre-constructed and stored multi-view trace map is loaded first. For example, for a machine processing production line already in operation, its historical logs have already been used to construct a multi-view trace map reflecting the standard process of that production line. When a new production task begins, this multi-view trace map can be directly used as a reference for subsequent predictions.

[0126] Step 2: Obtain the trajectory prefix sequence of the industrial process instance to be predicted. The trajectory prefix sequence is composed of several events that have been executed by the industrial process instance in chronological order, and each event contains at least an activity identifier.

[0127] Specifically, in industrial production, when a process instance is being executed, the system has recorded a series of events completed by that instance from its start to the current moment. These executed events, arranged chronologically, form a sequence called the trajectory prefix sequence. This step retrieves this sequence from the Manufacturing Execution System (MES) or a real-time database. Each event in the trajectory prefix sequence contains at least an activity identifier to correspond to a node in the multi-view trajectory map. For example, for a wheel hub order that is being processed and has completed three processes—"material feeding," "rough machining," and "finishing"—its trajectory prefix sequence would be [material feeding, rough machining, finishing]. The length of this sequence may vary depending on the process instance; some may contain only one event, while others may be nearing completion. After retrieving the trajectory prefix sequence, it is possible to determine which activity node the current process instance has reached and which paths it has traversed.

[0128] Step 3: Input the multi-view trace map into a pre-trained encoder. The encoder is used to encode the node features and edge relationships in the multi-view trace map to generate a context feature representation for each active node.

[0129] In this embodiment of the invention, in order to utilize the structural and node attribute information in the multi-view trace graph for prediction, this step employs a pre-trained encoder to process the entire graph. The encoder is a neural network model capable of receiving graph structure data and outputting node-level representations. Taking the multi-view trace graph loaded in step one as input, the encoder encodes each active node in the graph: it not only considers the node's own multi-view fusion feature vector but also aggregates information from neighboring nodes along the edges of the graph, thereby generating a feature representation that integrates local and global context, called a contextual feature representation. This representation implies the active node's role in the overall process, its dependencies on other activities, and the combined influence of multi-view attributes.

[0130] For example, after inputting the multi-view trace map of wheel hub production into the encoder, the encoder outputs a contextual feature vector for each node. For the "finishing" node, its contextual features not only include the resource and scale attributes of "finishing" itself, but also integrate information from its predecessor "roughing" and subsequent "heat treatment" and "polishing," as well as the dependency strength and semantics reflected by the edge weights and edge types between them. The encoder parameters have been pre-trained on a large amount of historical data, enabling the encoded features to effectively support the prediction of the remaining time. It should be noted that the specific structure of the encoder (such as a graph convolutional network, graph attention network, or graph Transformer) can be selected according to actual needs, and this embodiment does not impose any limitations.

[0131] Step 4: Extract the corresponding process state representation from the encoder output based on the last active node in the trajectory prefix sequence.

[0132] In this embodiment of the invention, after step three described above, the encoder has generated a contextual feature representation for each active node in the multi-view trace map. The current process instance to be predicted has executed to the last event in the trajectory prefix sequence, and the active node corresponding to this event represents the current state of the process. This step extracts the activity identifier of the last event in the trajectory prefix sequence, and then searches for the contextual feature representation corresponding to this active node in the encoder output, using it as the process state representation of the current process instance.

[0133] For example, for the trajectory prefix sequence [feeding, roughing, finishing], the last active node is "finishing". The context feature vector corresponding to the "finishing" node is extracted from the encoder output. This vector encodes all relevant information for the current process instance up to the completion of "finishing", including the attributes of "finishing" itself, the previously traversed paths, possible subsequent choices, and the structural constraints of the entire process. This process state representation will serve as input to the decoder to predict the remaining time.

[0134] Step 5: Input the process status representation into the decoder to predict the remaining time of the industrial process instance.

[0135] In this embodiment of the invention, the process state representation extracted in step four is input into a pre-trained decoder. The decoder is a regression model capable of mapping feature vectors to numerical values ​​(remaining time). The decoder can be a simple multilayer perceptron (MLP) or a more complex time series model. After receiving the process state representation, the decoder outputs a scalar value representing the predicted remaining time required from the current state to the end of the process.

[0136] For example, the context feature vector of the "finishing" node is input into the decoder, and the decoder outputs "35 minutes," indicating that it is estimated that another 35 minutes are needed to complete the remaining heat treatment, quality inspection, packaging, and other processes. This prediction result can be fed back to the production scheduling system in real time to optimize resource allocation or provide early warnings of delays.

[0137] The decoder also needs to be pre-trained on historical data. During training, for each complete historical trajectory, prefixes of different lengths can be extracted as input, and the corresponding true remaining time is the actual duration from the end of that prefix to the end of the trajectory. By minimizing the error between the predicted value and the true value (such as the mean squared error MSE), the decoder learns the mapping relationship from the process state to the remaining time.

[0138] In summary, this embodiment achieves industrial process remaining time prediction based on the multi-view trace graph construction method through steps one to five described above. This prediction first utilizes a pre-constructed multi-view trace graph to provide structured prior knowledge of the process. Then, an encoder encodes the graph structure information into contextual features for each node. Next, the process state representation is extracted based on the last node of the current trajectory prefix. Finally, the decoder regresses to obtain the remaining time.

[0139] Existing methods for predicting remaining time in industrial processes struggle to simultaneously capture both the topological dependencies and dynamic temporal evolution characteristics of the process. On one hand, methods that rely solely on activity sequences ignore complex structures such as parallelism and selection between activities; on the other hand, traditional graph models focus primarily on static dependencies, making it difficult to adapt to the dynamic temporal changes during process execution.

[0140] In view of this, this application proposes a scheme for applying a multi-view trace graph construction method for industrial processes to the task of predicting the remaining time of industrial processes. This scheme utilizes a pre-constructed multi-view trace graph to uniformly represent the interaction between the execution structure and multi-view attributes of the industrial process. An encoder encodes the topology and node attributes in the graph, ensuring that the generated node context feature representation simultaneously contains the topological dependencies and dynamic temporal evolution characteristics of the process. A decoder then maps the current process state representation to the predicted remaining time value. Compared to existing technologies, this scheme can simultaneously utilize control flow structure, multi-view attributes, and dynamic temporal information, significantly improving the accuracy and robustness of prediction. Therefore, it effectively solves the technical problem in existing technologies where "industrial process remaining time prediction methods struggle to simultaneously capture the topological dependencies and dynamic temporal evolution characteristics of the process," providing an efficient and scalable solution for real-time predictive monitoring of industrial processes.

[0141] Optionally, the encoder generates the contextual feature representation of each active node through the following steps: determining the set of neighboring nodes of the active node in the multi-view trace graph; using the multi-view fused feature vector of the active node as the current feature of the active node; within the set of neighboring nodes, calculating the attention weight between the active node and each neighboring node based on the current feature of the active node and the current features of each neighboring node, combined with the weight of the corresponding edge and the edge type label; weighting and aggregating the current features of the neighboring nodes according to the attention weight, and updating the features of the active node through a multi-head attention mechanism, feedforward network, residual connection and layer normalization operation to obtain the contextual feature representation of the active node.

[0142] In this embodiment of the invention, after completing the construction of the multi-view trace graph, this application further designs a remaining time prediction module based on Graph-Transformer to fully explore the structural dependencies, multi-view semantic information, and long-term dynamic features contained in the industrial process. This module takes the multi-view trace graph as input. In the encoding stage, the model uses Graph-Transformer as its core architecture, organically combining GCN and Transformer architectures. While retaining the graph structure modeling capability, it also significantly enhances the model's ability to express long-sequence dependencies and global semantic information, enabling it to more comprehensively characterize the complex behavioral patterns in the industrial process execution. In the decoding stage, the model adopts a time-aware decoding structure composed of additive attention mechanism and multi-layer LSTM to achieve the interactive fusion of multi-level information and temporal dependency modeling, accurately predicting the remaining time of the industrial process. The overall framework of the remaining time prediction module is as follows: Figure 5 As shown. The following will combine... Figure 5It describes in detail the specific steps of the encoder to generate the context feature representation of each active node, as well as the subsequent remaining time prediction and training process.

[0143] Step 1: Determine the set of neighboring nodes.

[0144] For any active node in the multi-view trace graph Its set of neighboring nodes Defined as:

[0145] ;

[0146] Here, E is the set of edges in the multi-view trace graph. This set reflects other activities that have direct behavioral dependencies on the current activity.

[0147] Step 2: Initialize node characteristics.

[0148] Each activity node Multi-view fusion feature vector This serves as its initial feature representation. In the first layer of the encoder, the current feature of a node is... .

[0149] Step 3: Construct query, key, and value vectors using linear mapping.

[0150] For nodes and its neighboring nodes Features are mapped to query vectors using a learnable weight matrix. Key vector Sum value vector :

[0151] ;

[0152] ;

[0153] ;

[0154] in, , , These are the learnable query matrix, key matrix, and value matrix, respectively.

[0155] Step 4: Calculate the attention score for the fused edge weights and edge types.

[0156] Introducing edge weights into attention calculation and edge type embedding The original attention score is obtained:

[0157] ;

[0158] in and To adjust the hyperparameters contributing to structural information, is the dimension of the key vector.

[0159] Step 5: Normalize to obtain attention weights.

[0160] For nodes Calculate the Softmax of all neighboring nodes to obtain the node For neighboring nodes attention weights :

[0161] ;

[0162] Step 6: Weighted aggregation of neighbor features.

[0163] Using attention weights on the value vectors of neighboring nodes Perform a weighted summation to obtain the aggregated result:

[0164] ;

[0165] in, Represents the value vector of single-head attention to neighbor nodes. The aggregated result is obtained by performing a weighted summation.

[0166] Step 7: Multi-head attention fusion.

[0167] A multi-head attention mechanism is used to replicate the above calculation process N times. head N times ( head (The number of heads for multi-head attention is specified, with each head having independent parameters). Then, the outputs of each head are concatenated and passed through an output matrix. Transformation:

[0168] ;

[0169] in, This represents the value vector of the nth head attention to its neighbor nodes. The aggregated result is obtained by performing a weighted summation, where n=1,2,…,N head ; This represents the final fused representation after splicing and linearly transforming the aggregation results of the attention from each head under the multi-head attention mechanism.

[0170] Step 8: Residual connectivity and layer normalization.

[0171] Output of multi-headed attention With activity nodes Multi-view fusion feature vector Add them together, then perform layer normalization to obtain the encoded output. :

[0172] ;

[0173] Wherein, FFN is a feedforward network. This is a layer normalization operation.

[0174] Step 9 (Optional, Enhanced Time Awareness): To further capture the dynamic temporal characteristics between activities, this application introduces a time-aware mechanism. For adjacent activities in the trajectory... and Define time interval :

[0175] ;

[0176] in, Indicates activity The timestamp of the occurrence, Indicates activity The timestamp of the occurrence.

[0177] The Time2Vec function is used to encode time intervals:

[0178] ;

[0179] in, Indicates activity and The time interval between The time-aware embedding vector obtained after encoding by the Time2Vec function is used to characterize the dynamic temporal features between adjacent activities;

[0180] By incorporating time encoding into the attention score, we obtain the time-perceived attention weights. :

[0181] ;

[0182] in, This is the time-aware adjustment coefficient. For the event The neighboring nodes, For nodes The query vector, For neighboring nodes The key vector, For edge ( , The weight of ) For edge ( , Type embedding, For edge ( , The time embedding mechanism enables the model to dynamically adjust the information propagation intensity based on the time interval between activities, thereby characterizing the non-uniform and non-stationary temporal evolution characteristics of industrial processes.

[0183] Step 10: Extract process status representation.

[0184] After obtaining node representations that integrate multi-perspective semantics, process structure, and temporal dynamics, the model further predicts the remaining time. For the current process instance to be predicted, let the last active node in its trajectory prefix sequence be... Then, the context feature representation of that node is extracted from the encoder output. As a representation of process status :

[0185] ;

[0186] Step 11: Decode and predict the remaining time.

[0187] Characterizing process status The input is fed into a decoder (e.g., a multilayer perceptron) to obtain the remaining time prediction:

[0188] ;

[0189] in, This represents the predicted remaining time of an industrial process. and These are learnable parameters.

[0190] During the model training phase, mean squared error (MSE) is used as the loss function:

[0191] ;

[0192] in, The loss value of the model, also known as the mean squared error, is used to measure the overall deviation between the model's predicted remaining time and the actual remaining time. This is the sample index, with values ​​ranging from 1, 2, ..., N. Let i be the true remaining time label for the i-th sample. Let N be the model prediction result for the i-th sample, and N be the number of training samples. By minimizing this loss function, the entire prediction model can be trained end-to-end.

[0193] Thus, through the above steps, this application utilizes a Graph-Transformer encoder combined with edge weights, edge types, and a time-aware mechanism to generate a node context representation rich in topological structure, multi-perspective semantics, and dynamic temporal features. Furthermore, it accurately predicts the remaining time through a decoder, effectively solving the problem in existing technologies that it is difficult to simultaneously capture process topological dependencies and dynamic temporal evolution features.

[0194] The following section will present experimental verification and result analysis of the method proposed in this application.

[0195] This application uses the Production dataset and the WMFD dataset for experiments. The Production dataset contains 4543 data entries, and the WMFD dataset contains 20305 data entries. The training and test sets are divided in an 8:2 ratio. Both datasets are derived from industrial process production logs and face modeling challenges due to multi-perspective feature variations in practical applications. For the Production dataset, its process structure is relatively standardized, but the execution order and resource allocation of different batches of production tasks differ, resulting in dynamic changes in the timing and dependency patterns of the event logs. The WMFD dataset comes from the manufacturing production line of a wheel hub company in China. The production process is affected by multiple factors such as equipment status, worker operation, and raw material batches, resulting in significant differences in process characteristics between different production cycles, exhibiting typical multi-condition characteristics and non-stationarity. Therefore, both datasets suffer from the problem that single-perspective models cannot effectively represent the process evolution law. They are suitable for verifying the effectiveness and generalization ability of the proposed industrial process multi-perspective trace graph construction method when applied to the task of predicting the remaining time of industrial processes (hereinafter referred to as MVT-GT).

[0196] To verify the effectiveness of the MVT-GT method, four methods were selected as a control group:

[0197] (1) Informer: This method builds a business process remaining time prediction framework based on the Informer model, uses BiLSTM to generate dynamic event vectors to capture trajectory context semantics, introduces sparse attention mechanism and dilated causal convolution to enhance long sequence modeling ability, and compresses key features through feature distillation structure to improve prediction efficiency and accuracy.

[0198] (2) EFH: This method proposes a process remaining time prediction framework based on an interpretable feature hierarchical model. It adopts a feature self-selection strategy that combines priority-based backward feature deletion and importance-based forward feature selection to screen attributes that have a positive impact on the prediction task. Through feature hierarchical input and layer-by-layer visualization, an interpretable prediction model is constructed to reveal the intrinsic relationship between feature values ​​and prediction results. The model is implemented using LightGBM and LSTM and its prediction accuracy and interpretability are verified on multiple real event logs.

[0199] (3) transKTCN: This method proposes a remaining time prediction framework based on multi-dimensional feature enhancement. It constructs a Multi-Trace2vec embedding representation by fusing multi-dimensional attributes such as activities and resources to enhance trajectory semantic information. On this basis, a transKTCN prediction model is designed. It combines dynamic sparse KAN for nonlinear feature mapping, uses deep separable TCN to extract multi-scale local features, and introduces Transformer to capture global temporal dependencies, forming a hierarchical feature fusion architecture. Its prediction accuracy and generalization ability are verified on multiple real datasets.

[0200] (4) AC-TAFormer: This method first completes the industrial event log, then introduces a time-aware factor into the Transformer encoder, constructs a time-aware attention mechanism, and encodes the trajectory prefix features to predict the remaining time.

[0201] The specific experimental results of each method in the comparative experiment are shown in Table 1, with mean absolute error (MAE) and root mean square error (RMSE) used as evaluation indicators.

[0202] Table 1 Comparison of prediction results by different methods

[0203]

[0204] As shown in Table 1, on the Production dataset, the MVT-GT method achieves the best performance with an MAE of 8.86 and an RMSE of 10.79 compared to the three comparison methods. On the WMFD dataset, the MVT-GT method achieves the best performance with an MAE of 2.29 and an RMSE of 3.51 compared to the Informer, EFH, and transKTCN methods. Furthermore, the MVT-GT method's predicted values ​​are closest to the actual values ​​on both industrial datasets, achieving the best prediction results. The Informer model's prediction performance on both datasets is significantly lower than the method proposed in this application. This indicates that in the task of predicting the remaining time of industrial processes, introducing structured process trace graph modeling and multi-view feature fusion mechanisms can significantly improve the model's ability to characterize the process execution rules, thus outperforming traditional time series prediction methods in overall prediction accuracy. The EFH method achieves relatively good performance on both datasets; however, the MVT-GT method further improves the prediction accuracy based on EFH by introducing graph structure constraints and a time-aware mechanism, demonstrating the effectiveness of multi-view graph modeling and time series dynamic capture. transKTCN combines the advantages of Transformer and temporal convolutional networks, but its performance on both datasets is inferior to the MVT-GT method. This reflects that simply combining different neural network architectures without explicitly modeling the topology and temporal dynamics of the industrial process is unlikely to achieve optimal prediction results. In contrast, the MVT-GT method, after explicitly modeling the topology and temporal dependencies of the industrial process, incorporates a Graph-Transformer model that integrates graph convolutional neural networks and Transformer for remaining time prediction, achieving better results.

[0205] Based on the comprehensive comparison of experimental results, the MVT-GT method demonstrates significant advantages in predicting the remaining time of industrial processes, validating the rationality and effectiveness of its design. This model not only provides a new technical approach for industrial process forecasting but also offers a valuable solution for handling time-series forecasting problems in complex industrial environments.

[0206] To further explore the contributions of each module in the MVT-GT method, this application designed and conducted ablation experiments, progressively removing key modules from the complete model and observing their impact on model performance. Three model versions were designed: the MVT-GT-noG method, the MVT-GT-noT method, and the complete MVT-GT method. Table 2 and... Figure 6 The MAE prediction results of different methods are shown on two datasets with different production data. The following is a brief description of each ablation method:

[0207] (1) MVT-GT-noG method: This method does not use an explicit graph structure, but treats each trajectory as a sequence of multi-view splicing, and uses a Transformer of the same dimension as an encoder.

[0208] (2) MVT-GT-noT method: This method removes the time-aware adjustment term and does not use time weighting in attention calculation. It only calculates attention weights based on structure and semantics.

[0209] (3) MVT-GT method: This method first models the industrial event log as a multi-level trace graph, and then inputs the trace graph into the Graph-Transformer encoder to predict the remaining time of the industrial process based on multiple perspectives.

[0210] Table 2 Comparison of prediction results for different methods in ablation experiments

[0211]

[0212] Table 2 shows that the MVT-GT method achieves a MAE of 8.86 on the Production dataset and 2.29 on the WMFD dataset, demonstrating the best performance compared to the other two ablation methods. Figure 6 As can be seen, the dark gray bars represent the MAE of the Production dataset across the three methods, and the light gray bars represent the MAE of the WMFD dataset across the three methods. The MVT-GT method achieved the lowest MAE compared to the other two ablation methods. Compared to the MVT-GT-noT method, the MVT-GT method, which incorporates a time-aware mechanism, improved prediction accuracy on both datasets. This is mainly because frequent changes in operating conditions and temporal fluctuations exist in real industrial environments, leading to non-stationary characteristics in process execution rhythm. The introduction of the time-aware mechanism allows the model to dynamically quantify the cumulative impact of activity execution intervals on the overall process, effectively overcoming this problem and significantly reducing prediction errors.

[0213] Compared to the MVT-GT-noG method, the MVT-GT method, which adds graph structure constraints, significantly improves prediction performance on both datasets. This is because the MVT-GT-noG method relies solely on linear sequences for feature modeling, which limits its ability to capture global interaction features between activities and makes it difficult to explicitly utilize the complex parallel relationships and causal dependencies in industrial processes. In contrast, the MVT-GT method introduces graph structure constraints, mapping multi-perspective nodes such as activities, resources, and equipment to a unified semantic space and using graph attention mechanisms to achieve a fusion expression of global context, thereby better characterizing the topological dependencies and behavioral patterns of the process.

[0214] Comprehensive analysis reveals that the MVT-GT method demonstrates stable predictive superiority on both datasets, indicating that graph structure modeling and the time-aware mechanism play complementary roles in enhancing model generalization and robustness. Specifically, the time-aware mechanism enhances the model's sensitivity to rhythmic changes and temporal dynamics between events, while graph structure modeling provides structural priors and multi-perspective semantic constraints, enabling the model to extend from local behavioral dependencies to global relational modeling. The synergistic effect of these two mechanisms allows the MVT-GT method to maintain high prediction accuracy and stability under different operating conditions and data distributions, validating the effectiveness and advancement of the proposed method in multi-perspective industrial process remaining time prediction tasks.

[0215] Furthermore, according to a second aspect of this embodiment, a storage medium is provided. The storage medium includes a stored program, wherein, when the program is executed, a processor performs the method described in any of the above embodiments.

[0216] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the present invention is not limited to the described order of actions, because according to the present invention, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to the present invention.

[0217] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods according to the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0218] Example 2

[0219] This embodiment provides an industrial process multi-view trace graph construction system, which corresponds to the industrial process multi-view trace graph construction method according to Embodiment 1. The system includes: an event log acquisition module for acquiring event logs of the industrial process, the event logs including multiple trajectories, each trajectory consisting of several events, and each event including at least an activity identifier; a behavior relationship mining module for mining and generating an initial behavior relationship graph reflecting the behavioral dependencies between activities based on the event logs, the initial behavior relationship graph using activities as nodes and the behavioral dependencies between activities as edges; a multi-view attribute extraction module for extracting multi-view attribute information associated with activities from the event logs, the multi-view attribute information including at least one or more of resource attributes, scale attributes, product attributes, and personnel attributes; a feature vector extraction module for embedding and mapping the multi-view attribute information to generate a multi-view fused feature vector for each activity node; and a multi-view trace graph construction module for combining the initial behavior relationship graph with the multi-view fused feature vector to construct a multi-view trace graph, the multi-view trace graph being used to uniformly represent the interaction relationship between the execution structure and multi-view attributes of the industrial process.

[0220] Optionally, the behavior relationship mining module is specifically used for: mining the event logs using a process mining algorithm to obtain the corresponding Petri net model; constructing a behavior footprint matrix based on the behavior footprint theory and the Petri net model, wherein the behavior footprint matrix is ​​used to distinguish between direct follow relationships, adjacent relationships, causal relationships, parallel relationships, and irrelevant relationships between activities; constructing a behavior relationship matrix by setting the matrix elements corresponding to irrelevant relationships between activities as first values ​​and setting the matrix elements corresponding to the other four types of relationships as second values, and using the behavior relationship matrix as the initial behavior relationship graph with activities as nodes and activity pairs with corresponding matrix elements as second values ​​as edges.

[0221] Optionally, the multi-view trace graph construction module is further configured to: assign an edge type label to each edge in the initial behavior relationship graph according to the behavior footprint matrix, wherein the edge type label is a direct follow relationship, an adjacent relationship, a causal relationship, or a parallel relationship; form an edge type set by the edge type labels of each edge, and incorporate the edge type set into the multi-view trace graph.

[0222] Optionally, the multi-view trace graph construction module is further specifically used to: assign weights to the edges in the initial behavior relationship graph, the weights being calculated based on the co-occurrence frequency and / or average time interval of the activity pairs in the event log; construct an edge weight matrix from the weights of each edge, and incorporate the edge weight matrix into the multi-view trace graph.

[0223] It should be noted that the industrial process multi-view trace map construction system provided in this embodiment can realize all the functions and steps in the above-mentioned industrial process multi-view trace map construction method embodiment, solve the same technical problems, and achieve the same technical effects. The similarities will not be repeated.

[0224] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0225] In the above embodiments of the present invention, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0226] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. An industrial process multi-view footprint construction method, characterized in that, include: Obtain the event log of the industrial process, wherein the event log includes multiple tracks, each track consists of several events, and each event includes at least an activity identifier; Based on the event log, an initial behavior relationship graph reflecting the behavioral dependencies between activities is generated. The initial behavior relationship graph uses activities as nodes and the behavioral dependencies between activities as edges. Extract multi-perspective attribute information associated with the activity from the event log. The multi-perspective attribute information includes at least one or more of the following: resource attributes, scale attributes, product attributes, and personnel attributes. The multi-view attribute information is embedded and mapped to generate a multi-view fusion feature vector for each active node; The initial behavior relationship graph is combined with the multi-view fusion feature vector to construct a multi-view trace graph, which is used to uniformly represent the interaction relationship between the execution structure of the industrial process and the multi-view attributes.

2. The method of claim 1, wherein, Based on the event logs, the operation of mining and generating an initial behavioral relationship graph reflecting the behavioral dependencies between activities includes: The event logs were mined using a process mining algorithm to obtain the corresponding Petri net model; Based on behavioral footprint theory, a behavioral footprint matrix is ​​constructed according to the Petri net model. The behavioral footprint matrix is ​​used to distinguish between direct follow-up relationships, adjacent relationships, causal relationships, parallel relationships and irrelevant relationships between activities. Based on the behavior footprint matrix, the matrix elements corresponding to the irrelevant relationships between activities are set as the first value, and the matrix elements corresponding to the other four types of relationships are set as the second value to construct the behavior relationship matrix. Based on the behavior relationship matrix, an initial behavior relationship graph is constructed with activities as nodes and activity pairs whose corresponding matrix elements are the second values ​​as edges.

3. The method of claim 2, wherein, Constructing the multi-view trace map also includes: Based on the behavior footprint matrix, each edge in the initial behavior relationship graph is assigned an edge type label, which can be a direct follow relationship, an adjacent relationship, a causal relationship, or a parallel relationship. An edge type set is formed by the edge type labels of each edge, and the edge type set is incorporated into the multi-view trace graph.

4. The method of claim 1, wherein, Constructing the multi-view trace map also includes: Weights are assigned to the edges in the initial behavior graph, and the weights are calculated based on the co-occurrence frequency and / or average time interval of the activity pairs in the event log. An edge weight matrix is ​​constructed from the weights of each edge, and the edge weight matrix is ​​incorporated into the multi-view trace graph.

5. The method of claim 1, wherein, The operation of embedding and mapping the multi-view attribute information to generate a multi-view fused feature vector for each active node includes: An independent embedding matrix is ​​constructed for each attribute contained in the multi-view attribute information; The embedding matrix is ​​used to map the discrete values ​​of the corresponding attributes into embedding vectors; The embedding vectors of all attributes are transformed to the same dimensional space through linear transformation; The embedded vectors of each attribute after transformation are concatenated and mapped through a nonlinear activation function to generate the multi-view fused feature vector.

6. A computer readable storage medium, characterized in that, The storage medium includes a stored program, wherein, when the program is executed, the method described in any one of claims 1 to 5 is performed by a processor.

7. An industrial process multi-view footprint construction system, comprising: include: The event log acquisition module is used to acquire event logs of industrial processes. The event logs include multiple tracks, each track consists of several events, and each event includes at least an activity identifier. The behavior relationship mining module is used to mine and generate an initial behavior relationship graph reflecting the behavior dependencies between activities based on the event log. The initial behavior relationship graph has activities as nodes and the behavior dependencies between activities as edges. The multi-perspective attribute extraction module is used to extract multi-perspective attribute information associated with the activity from the event log. The multi-perspective attribute information includes at least one or more of the following: resource attributes, scale attributes, product attributes, and personnel attributes. The feature vector extraction module is used to embed and map the multi-view attribute information to generate a multi-view fusion feature vector for each active node. The multi-view trace graph construction module is used to combine the initial behavior relationship graph with the multi-view fusion feature vector to construct a multi-view trace graph, which is used to uniformly represent the interaction relationship between the execution structure of the industrial process and the multi-view attributes.

8. The system of claim 7, wherein, The behavior relationship mining module is specifically used for: The event logs were mined using a process mining algorithm to obtain the corresponding Petri net model; Based on behavioral footprint theory, a behavioral footprint matrix is ​​constructed according to the Petri net model. The behavioral footprint matrix is ​​used to distinguish between direct follow-up relationships, adjacent relationships, causal relationships, parallel relationships and irrelevant relationships between activities. Based on the behavior footprint matrix, the matrix elements corresponding to the irrelevant relationships between activities are set as the first value, and the matrix elements corresponding to the other four types of relationships are set as the second value to construct the behavior relationship matrix. Based on the behavior relationship matrix, an initial behavior relationship graph is constructed with activities as nodes and activity pairs whose corresponding matrix elements are the second values ​​as edges.

9. The system of claim 8, wherein, The multi-view trace map construction module is also specifically used for: Based on the behavior footprint matrix, each edge in the initial behavior relationship graph is assigned an edge type label, which can be a direct follow relationship, an adjacent relationship, a causal relationship, or a parallel relationship. An edge type set is formed by the edge type labels of each edge, and the edge type set is incorporated into the multi-view trace graph.

10. The system of claim 7, wherein, The multi-view trace map construction module is also specifically used for: Weights are assigned to the edges in the initial behavior graph, and the weights are calculated based on the co-occurrence frequency and / or average time interval of the activity pairs in the event log. An edge weight matrix is ​​constructed from the weights of each edge, and the edge weight matrix is ​​incorporated into the multi-view trace graph.

Images