Business process remaining time prediction method and system based on hybrid architecture

Abstract

The invention discloses a business process remaining time prediction method and system based on a hybrid architecture, and belongs to the field of artificial intelligence and process mining. The method comprises the following steps: acquiring an original event log of a business process to generate a to-be-predicted event sequence prefix set; a position coding vector and a time difference vector are introduced into the event embedding model; performing multi-scale local feature representation by adopting a multi-scale one-dimensional convolutional neural network; a weight attenuation factor based on a time difference between events is introduced into attention calculation of a time perception Transform encoder; and aggregating and predicting the remaining time of the business process at a full-connection output layer. The system comprises an event prefix acquisition module, an embedding and position coding module, a convolution feature extraction module, a Transform coding module and a remaining time prediction module. According to the method, event attributes, a multi-scale local mode and global time sequence dependence are fused, and the prediction precision and the model robustness are remarkably improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of artificial intelligence and process mining, and particularly relates to a business process remaining time prediction method and system based on a hybrid architecture. BACKGROUND

[0002] With the rapid development of process mining and predictive process monitoring, business process remaining time prediction has become a key requirement in process management. In existing research, most methods model event sequences and predict remaining time based on recurrent neural networks (such as LSTM, GRU). Such methods can capture certain temporal dependencies, but still have deficiencies in the balanced modeling of local patterns and global features: for example, LSTM has an advantage in capturing long-range dependencies, but is not sensitive enough to local patterns between adjacent events; convolutional neural networks (CNN) can efficiently extract local features, but lack the ability to model long-range dependencies. In recent years, models based on Transformer have significantly enhanced the ability to model global context through multi-head self-attention mechanisms, but are still limited in extracting short-term local patterns. In summary, existing methods have deficiencies in balancing local short-term features and global context dependencies, so there is an urgent need to propose a new method that can integrate local and global features to further improve the accuracy and robustness of business process remaining time prediction. SUMMARY

[0003] To solve the above problems, the application proposes a business process remaining time prediction method and system based on a hybrid architecture, which is a CNN-Transformer hybrid model (CNN-TF). The hybrid model uses a structure of "multi-scale convolutional neural network (CNN) to extract local features first, and then time-aware Transformer encoder to model long-range dependencies", so that the local patterns of the sequence are fully strengthened before entering the Transformer, thereby significantly improving the quality of global dependency modeling and time sensitivity. The application predicts the remaining completion time of the event sequence of the business process instance being executed using a deep learning model, thereby improving the accuracy and real-time performance of business process monitoring and management.

[0004] The technical solutions of the application are as follows: A business process remaining time prediction method based on a hybrid architecture, the hybrid architecture comprising an event embedding model, a multi-scale one-dimensional convolutional neural network, a time-aware Transformer encoder, and a fully connected output layer. The method specifically comprises the following steps: Step 1: Obtain the original event log of the business process, extract the prefix of each trajectory in the original event log, and generate a set of event sequence prefixes to be predicted; Step 2, vectorize and embed each event in the event sequence prefix set based on the event embedding model, and introduce a position encoding vector and a time difference vector to obtain a time-aware embedding representation sequence that fuses position information and time information; Step 3, input the time-aware embedding representation sequence into a multi-scale one-dimensional convolutional neural network for convolution operation, extract multi-level local features through convolution operation with different convolution kernel sizes, and fuse them to obtain a multi-scale local feature representation sequence of the event sequence prefix set; Step 4, input the multi-scale local feature representation sequence into the time-aware Transformer encoder to learn the global context dependency in the event sequence prefix set through the multi-head self-attention mechanism; introduce a weight decay factor based on the time difference between events during attention calculation, and finally calculate a global feature representation sequence containing time-dependent information; Step 5, aggregate the global feature representation sequence and calculate the remaining time prediction value of the business process through a fully connected output layer.

[0005] Further, the specific process of step 1 is as follows: Step 1.1, obtain the original event log from the public dataset website, the format of the original event log is a structured file conforming to the XES specification, containing event activity name, timestamp, event ID, event associated performer, organizational role, resource type, event lifecycle identifier, and other optional business attributes; The original event log is a collection of several traces, and a trace is a sequence composed of several event activities; Define the original event log as: (1); Wherein, is the original event log; is the total number of traces; is the th trace, specifically: (2); Wherein, is the th event in the th trace; Each event consists of an activity, a timestamp, and a context attribute embedding set, which is derived from the event context attributes recorded in the original event log, including event-associated performers, organizational roles, resource types, event lifecycle identifiers, and other optional business attributes. The context attribute embedding set is composed of multiple context attributes associated with the current event in the event log, and when a certain type of attribute is missing, it is represented by a preset placeholder; Step 1.2: For each track in the original event log, extract all tracks of length [length missing] in non-decreasing order of event timestamps. The trajectory prefix, where ; For the first The number of events contained in a trajectory; The trajectory prefix is ​​defined as: (3); in, Prefix for trajectory; For the first The first of the trajectories One event, Corresponding trajectory prefix length; The event sequence prefix set is composed of all prefixes of all trajectories. : (4).

[0006] Furthermore, the specific process of step 2 is as follows: Step 2.1: Use an event embedding model to vectorize and embed each event in the trajectory prefix, obtaining the event embedding vector: (5); in, Embedded model for events; For the first The first of the trajectories The activities of this event; For the first The first of the trajectories The context attributes corresponding to the event are embedded in the collection of the first event. Embedded representation of class context attributes; For the first The first of the trajectories Embedding vectors of events, For dimensions; For the first The first of the trajectories The context attributes corresponding to each event are embedded in a collection; Step 2.2: Introduce positional encoding into the event embedding vector to obtain the positional encoding vector. The positional encoding vector is... Dimensional components and the first The dimensional components are calculated using the following formulas: (6); (7); in, For the first The first of the trajectories Location encoding vectors for each event; For location index, ; Step 2.3: Calculate the time interval between adjacent events and map it into a time difference vector using a time embedding function: (8); (9); in, For the first The first of the trajectories The time interval between adjacent events; , The first The first of the trajectories The first event, the The time when the event occurred; For the first The first of the trajectories The time difference vector between adjacent events; For time embedding functions; Step 2.4: Add the embedding vector of each event to its corresponding location encoding vector and time difference vector element-wise to obtain the event embedding representation that fuses location and time information. (10); in, For the first The first of the trajectories Embedded representation of an event; The embedded representations of all events form the time-aware embedded representations of the trajectory prefix sequence. , No. The embedded representation of the event corresponds to the first event placed in the time-aware embedding sequence. One position.

[0007] Furthermore, the specific process of step 3 is as follows: Step 3.1: For the time-aware embedded representation sequence, perform parallel convolution operations using a one-dimensional convolutional neural network with multiple kernel sizes to extract multi-level local feature values; for the first... The convolution operation with one convolution kernel is calculated using the following formula: (11); in, For time-aware embedding representation of the sequence, the first... A vector representation of each event activity; For the first Each convolutional kernel is located at... Weight parameters at the location; For the first Local features generated by each convolutional kernel; It is a non-linear activation function; For the time-aware embedding representation of the input sequence, the first... The embedding vector of the i-th event, after the i-th Local features obtained after computation by each convolutional kernel; For the first The size of each convolutional kernel; Step 3.2, place the first The local features generated by all convolutional kernels of each event are concatenated along the feature dimension to obtain the event sequence prefix set. Multi-scale local feature vectors of an event: (12); in, Indicates the number of multi-scale convolution kernels; For the time-aware embedding representation of the input sequence, the first... The embedding vector of the i-th event, after the i-th Local features obtained after computation by each convolutional kernel; For the first Multi-scale local feature vectors after the fusion of several events; The multi-scale local feature vectors of all events constitute a multi-scale local feature representation sequence. ,in is the length of the feature sequence output by the convolution.

[0008] Furthermore, the specific process of step 4 is as follows: Step 4.1: Represent the multi-scale local feature sequence As input to the Transformer encoder, the attention weights between positions in the sequence are calculated using a multi-head self-attention mechanism, and a weight decay factor for the time difference between events is introduced into the attention score. The calculation formula is as follows: (13); in, This is a multi-head self-attention mechanism; They represent sequences from multi-scale local feature representations through linear transformation. The resulting query matrix, key matrix, and value matrix; It is the transpose symbol; The key vector dimension for each attention head; This is the time difference matrix between events; This is the weight decay factor; This indicates that each row of the matrix is ​​normalized. Step 4.2: The feature representation output by the multi-head self-attention mechanism is processed through a feedforward neural network, residual connections, and layer normalization to obtain a global feature representation sequence containing the event sequence prefix set with time-dependent information. ; for The Middle The global feature vector corresponding to each position The total number of positions; the specific process is as follows: The feature representations output by the multi-head self-attention mechanism constitute the multi-head self-attention output matrix. For the multi-head self-attention output matrix, the calculation formula via the feedforward neural network is: (14); in, It is a feedforward neural network; , This is the weight matrix of the feedforward neural network; , This is the bias term for the feedforward neural network; The output of the feedforward neural network is then combined with Residual connections are performed, and after layer normalization, the global feature vector is obtained. : (15); in, This represents a layer normalization operation that normalizes each dimension of the vector. Ultimately, the global feature vectors corresponding to all positions form a global feature representation sequence. .

[0009] Furthermore, the specific process of step 5 is as follows: Step 5.1: Represent the global feature sequence Perform average pooling to obtain the aggregated global feature vector. The calculation formula is: (16); in, for The Middle The global feature vectors corresponding to each location, with the location index corresponding to the event index; Step 5.2: Aggregate the global feature vector Input a fully connected layer and output a layer to calculate the predicted remaining time for the business process. The calculation formula is: (17); in, a weight vector for a fully connected output layer; a bias term.

[0010] A hybrid architecture based business process remaining time prediction system adopting the hybrid architecture based business process remaining time prediction method as described above; the system comprises: an event prefix acquisition module, an embedding and position encoding module, a convolution feature extraction module, a Transformer encoding module, a remaining time prediction module; the event prefix acquisition module is used to acquire a set of event sequence prefixes to be predicted from the original event log of the business process; the embedding and time position encoding module is used to map each event activity in the set of event sequence prefixes and its associated context attributes to a corresponding embedding vector, and introduce position encoding and time-aware encoding based on the time interval between events for each embedding vector, to generate a time-aware embedding representation sequence containing business context, position information and time information; the multi-scale convolution feature extraction module is used to perform one-dimensional convolution operations with different convolution kernel sizes on the time-aware embedding representation sequence to extract multi-level local features, and fuse them to obtain a multi-scale local feature representation sequence of the set of event sequence prefixes; the time-aware Transformer encoding module is used to perform encoding operations based on the multi-head self-attention mechanism on the multi-scale local feature representation sequence, and introduce a weight decay factor based on the time difference between events in the attention calculation, to learn the global temporal dependency in the set of event sequence prefixes, to obtain a global feature representation sequence of the set of event sequence prefixes; the remaining time prediction module is used to calculate the business process remaining time prediction value corresponding to the set of event sequence prefixes according to the global feature representation sequence.

[0011] The application brings the beneficial technical effects: the business process residual time prediction method based on the convolutional neural network and the Transformer provided by the application combines the advantages of local feature extraction and global dependence modeling, significantly improves the prediction accuracy and model generalization ability. Through the event embedding model and the position coding, a time-aware embedding representation sequence with time sequence context is constructed to ensure the complete preservation of the event order information; the one-dimensional convolutional neural network is used to effectively capture the local patterns in the prefix sequence, thereby improving the model's perception ability of short-term behavior characteristics; the time-aware Transformer encoder is further introduced to learn the long-range dependence relationship between events by using the multi-head self-attention mechanism, so as to realize the global feature expression; meanwhile, a weight factor based on the time difference between events is introduced in the attention calculation to further improve the modeling effect of the global time sequence dependence; finally, the residual time prediction value is output through the average pooling and the fully connected regression layer, so as to realize the end-to-end fine prediction. The method achieves excellent prediction performance on multiple public logs, has good robustness and adaptability, and is suitable for intelligent prediction scenarios of financial, medical, government and other types of business processes, thereby providing efficient and accurate technical means for process optimization and decision support. BRIEF DESCRIPTION OF DRAWINGS

[0012] Figure 1 The flowchart of the business process residual time prediction method based on the hybrid architecture of the application. DETAILED DESCRIPTION

[0013] The application will be further described in detail in combination with the drawings and the specific embodiments: Firstly, the application obtains the event sequence prefix set of the instance trajectory from the business process event log, embeds the event activities and the associated context attributes (including the executor, the resource, the timestamp and the like) for modeling, and obtains the event embedding vector sequence; meanwhile, the position coding is introduced for each embedding vector, and the time-aware coding is generated in combination with the time interval between events to preserve the sequence order and the time sequence difference information; secondly, the multi-scale one-dimensional convolutional neural network is used to perform convolution operation on the event embedding representation sequence, thereby extracting multi-level local features through different convolution kernel sizes, so as to enhance the model's representation ability of short-term dependence patterns; then, the fused local features are input into the time-aware Transformer encoder, the global context dependence relationship is captured through the multi-head self-attention mechanism, and the weight factor based on the time difference between events is introduced in the attention calculation to further improve the modeling effect of the global time sequence dependence; finally, the global feature sequence output by the Transformer encoder is aggregated, and the regression calculation is performed through the fully connected layer, so as to obtain the business process residual time prediction value corresponding to the event sequence prefix.

[0014] The application combines attribute embedding, time-aware mechanism and multi-scale convolution feature extraction to construct a CNN-TF model, which can simultaneously consider local patterns and global dependencies, and significantly improve the accuracy and robustness of business process remaining time prediction. In experimental evaluation, the model is better than traditional CNN, Transformer and Transformer-CNN (TF-CNN) models on multiple public event logs, and reduces the mean absolute error (MAE) by 17.8% on some data sets. Meanwhile, the application also provides a corresponding system, including an event prefix acquisition module, an embedding and position encoding module, a convolution feature extraction module, a Transformer encoding module and a remaining time prediction module, which has good engineering feasibility and implementability.

[0015] The application provides a business process remaining time prediction method based on a hybrid architecture, which includes an event embedding model, a multi-scale one-dimensional convolutional neural network, a time-aware Transformer encoder and a fully connected output layer, as shown in Figure 1 The method specifically includes the following steps: Step 1, obtain the original event log of the business process, extract the prefix of each trajectory in the original event log, and generate a set of event sequence prefixes to be predicted; the specific process is as follows: Step 1.1, obtain the original event log from a public dataset website, the format of the original event log is a structured file conforming to the XES specification, which contains event activity name, timestamp, event ID, event associated performer, organization role, resource type, event life cycle identifier and other optional business attributes; The original event log is a set of several trajectories, and a trajectory is a sequence composed of several event activities; The original event log is defined as: (1); Wherein, is the original event log; is the total number of trajectories; is the th trajectory, specifically: (2); Wherein, is the th event in the th trajectory; Each event is defined as consisting of an activity, a timestamp, and a set of contextual attribute embeddings. This set of contextual attribute embeddings is derived from the event contextual attributes recorded in the original event log, including the executor associated with the event, organizational role, resource type, event lifecycle identifier, and other optional business attributes. The set of contextual attribute embeddings consists of multiple contextual attributes associated with the current event from the event log. When a certain type of attribute is missing, it is represented by a preset placeholder to ensure the consistency of subsequent model inputs. Step 1.2: For each track in the original event log, extract all tracks of length [length missing] in non-decreasing order of event timestamps. The trajectory prefix, where ; For the first The number of events contained in a trajectory, i.e., the actual length of the trajectory.

[0016] The trajectory prefix is ​​defined as: (3); in, Prefix for trajectory; For the first The first of the trajectories One event, Corresponding trajectory prefix length; The event sequence prefix set is composed of all prefixes of all trajectories. : (4); This prefix set serves as the basis for the subsequent model input sequence.

[0017] Step 2: Based on the event embedding model, each event in the event sequence prefix set is vectorized and embedded, and a position encoding vector and a time difference vector are introduced to obtain a time-aware embedding representation sequence that integrates position and time information; the specific process is as follows: Step 2.1: Use an event embedding model to process the trajectory prefix. Each event in the vector is embedded into a vectorized representation, resulting in 3D event embedding vector: (5); in, Embedded model for events; For the first The first of the trajectories The activities of this event; For the first The first of the trajectories The context attributes corresponding to the event are embedded in the collection of the first event. Embedded representation of class context attributes; For the first The first of the trajectories Embedding vectors of events, For dimensions; For the first The first of the trajectories The context attributes corresponding to each event are embedded in a collection; Step 2.2: Introduce positional encoding into the event embedding vector to obtain the positional encoding vector. The positional encoding vector is... Dimensional components and the first The dimensional components are calculated using the following formulas: (6); (7); in, For the first The first of the trajectories Location encoding vectors for each event; For location index, ; Step 2.3: Calculate the time interval between adjacent events and map it into a time difference vector using a time embedding function: (8); (9); in, For the first The first of the trajectories The time interval between adjacent events; , The first The first of the trajectories The first event, the The time when the event occurred; For the first The first of the trajectories The time difference vector between adjacent events; For time embedding functions; Step 2.4: Add or concatenate the embedding vector of each event with its corresponding location encoding vector and time difference vector element-wise to obtain the event embedding representation that integrates location and time information. (10); in, For the first The first of the trajectories Embedded representation of an event; The embedded representations of all events form the time-aware embedded representations of the trajectory prefix sequence. As input for subsequent convolutional feature extraction and global modeling, the first... The embedding representation of the i-th event corresponds to the i-th position placed in the time-aware embedding sequence.

[0018] Step 3, input the time-aware embedding representation sequence into a multi-scale one-dimensional convolutional neural network for convolution operation, extract multi-level local features through convolution operation of different convolution kernel sizes, and fuse them to obtain a multi-scale local feature representation sequence of the event sequence prefix set; the specific process is as follows: Step 3.1, for the time-aware embedding representation sequence, a one-dimensional convolutional neural network with multiple different convolution kernel sizes is used for parallel convolution operation to extract multi-level local feature values. For the convolution operation of the i-th convolution kernel (the size of the i-th convolution kernel is The calculation formula is as follows: (11) ; Wherein, is the vector representation of the i-th event activity in the time-aware embedding representation sequence; is the weight parameter of the i-th convolution kernel at position is the local feature generated by the i-th convolution kernel; is a nonlinear activation function; is the embedding vector of the i-th event in the input time-aware embedding representation sequence, and the local feature obtained after the i-th convolution kernel calculation; Step 3.2, splice the local features of the i-th event generated by all convolution kernels in the feature dimension to obtain the multi-scale local feature vector of the i-th event of the event sequence prefix set: (12) ; Wherein, represents the number of multi-scale convolution kernels; is the embedding vector of the i-th event in the input time-aware embedding representation sequence, and the local feature obtained after the i-th convolution kernel calculation; is the multi-scale local feature vector of the i-th event after fusion. All multi-scale local feature vectors of events form a multi-scale local feature representation sequence (12) ; Wherein, represents the number of multi-scale convolution kernels; is the embedding vector of the i-th event in the input time-aware embedding representation sequence, and the local feature obtained after the i-th convolution kernel calculation; is the multi-scale local feature vector of the i-th event after fusion. All multi-scale local feature vectors of events form a multi-scale local feature representation sequence

[0019] ​​​​​​​​, as the input of subsequent global feature modeling, where is the length of the feature sequence output by the convolution.

[0020] Step 4, input the multi-scale local feature representation sequence output by the multi-scale one-dimensional convolutional neural network into the time-aware Transformer encoder, and learn the global context dependency in the event sequence prefix set through the multi-head self-attention mechanism. In the attention calculation process, a weight decay factor based on the time difference between events is introduced to enhance the modeling ability of global temporal dependency, so as to obtain a global feature representation sequence of the event sequence prefix set containing time dependency information; the specific process is as follows: Step 4.1, the multi-scale local feature representation sequence is input into the Transformer encoder, and the multi-head self-attention mechanism is used to calculate the attention weight between each position in the sequence, and a weight decay factor based on the time difference between events is introduced in the attention score to enhance the modeling ability of temporal dependency, and the calculation formula is as follows: (13); wherein, is the multi-head self-attention mechanism; are the query matrix, key matrix and value matrix obtained by linear transformation from the multi-scale local feature representation sequence , respectively; is the transpose symbol; is the key vector dimension of each attention head; is the time difference matrix between events; is the weight decay factor; represents normalizing each row of the matrix.

[0021] Step 4.2, the feature representation output by the multi-head self-attention mechanism is processed through a feed-forward neural network, a residual connection and layer normalization to obtain a global feature representation sequence of the event sequence prefix set containing time dependency information . is the global feature vector corresponding to the th position in , and is the total number of positions. The specific process is as follows: The feature representation output by the multi-head self-attention mechanism constitutes a multi-head self-attention output matrix , and the calculation formula of the multi-head self-attention output matrix through the feed-forward neural network (Feed-Forward Network) is as follows: (14); wherein, is the feed-forward neural network; , is a weight matrix of the feedforward neural network; 、 is a bias term of the feedforward neural network; The output of the feedforward neural network is further connected in residual and normalized by layer to obtain a global feature vector : (15) ; wherein, represents a layer normalization operation for normalizing each dimension of the vector; Finally, the global feature vectors corresponding to all positions form a global feature representation sequence .

[0022] Step 5, aggregate the global feature representation sequence output by the time-aware Transformer encoder, and calculate the remaining time prediction value of the business process through a fully connected output layer. The specific process is as follows: Step 5.1, perform average pooling operation on the global feature representation sequence to obtain an aggregated global feature vector , and the calculation formula is as follows: (16) ; wherein, is the global feature vector corresponding to the th position in , and the position number corresponds to the event number; Step 5.2, input the aggregated global feature vector into a fully connected output layer to calculate the remaining time prediction value of the business process, and the calculation formula is as follows: (17) ; wherein, is a weight vector of the fully connected output layer; is a bias term.

[0023] ​A business process remaining time prediction system based on convolutional neural networks and Transformers, employing the aforementioned business process remaining time prediction method to predict the remaining time of business process event logs, comprises: an event prefix acquisition module, an embedding and location encoding module, a convolutional feature extraction module, a Transformer encoding module, and a remaining time prediction module. The event prefix acquisition module is used to obtain a set of event sequence prefixes to be predicted from the original event logs of the business process. The embedding and time-location encoding module is used to map each event activity and its associated context attributes (including executor, resource, timestamp, etc.) in the event sequence prefix set into a corresponding embedding vector, and introduces location encoding and time-aware encoding based on the time interval between events into each embedding vector, generating a vector containing business context and location information. The system comprises: a time-aware embedded representation sequence with time information; a multi-scale convolutional feature extraction module, used to perform one-dimensional convolution operations with different kernel sizes (or dilation rates) on the time-aware embedded representation sequence to extract multi-level local features and fuse them to obtain a multi-scale local feature representation sequence of the event sequence prefix set; a time-aware Transformer encoding module, used to perform encoding operations based on a multi-head self-attention mechanism on the multi-scale local feature representation sequence and introduce a weight decay factor based on the time difference between events in the attention calculation to learn the global temporal dependencies in the event sequence prefix set and obtain a global feature representation sequence of the event sequence prefix set; and a remaining time prediction module, used to calculate the predicted remaining time value of the business process corresponding to the event sequence prefix set based on the global feature representation sequence.

[0024] To demonstrate the feasibility and superiority of the present invention, the following embodiments are provided.

[0025] To verify the effectiveness of the CNN-TF model in predicting remaining time in real-world business processes, the following implementation example is provided using the publicly available dataset BPI_13_Incidents (i.e., the Volvo IT incident management logs from the BPI Challenge 2013). This example simulates the incident ticket processing workflow of a large IT operations department.

[0026] The hardware configuration for this embodiment is as follows: Cluster environment: Intel i9-13900K CPU, NVIDIA RTX 4090 GPU; The software tools used in this embodiment are as follows: Process mining tool: ProM 6.11 (integrated with custom plugins, supporting clustering, sampling and visualization).

[0027] Embedding model training: Python 3.9 + PyTorch 2.2.2 deep learning framework.

[0028] The format of the public dataset BPI_13_Incidents is XES standard event log, that is, BPI_13_Incidents.xes. The public dataset BPI_13_Incidents is taken as an example event log, which contains a total of 7554 trajectories, a total of 67433 event activities, 7 types of event activity types, and a trajectory length ranging from 3 steps to 19 steps.

[0029] The above CNN-TF model is trained and evaluated on the BPI_13_Incidents dataset, and the MAE (mean absolute error) and RMSE (root mean square error) indicators are recorded and compared with traditional models. Table 1 lists the MAE and RMSE values of different models on the dataset. Among them, CNN is a convolutional neural network; Transformer is a sequence model based on attention mechanism; TF-CNN is a comparative model that first captures global context dependency by Transformer and then extracts local features by convolutional neural network; and CNN-TF is a hybrid prediction structure proposed by the application, which first extracts local patterns by convolutional neural network and then models long-range dependencies by Transformer.

[0030] Table 1 Comparison of MAE and RMSE values of different models on BPI_13_Incidents dataset .

[0031] As can be seen from Table 1, on the BPI_13_Incidents dataset, the MAE and RMSE of the CNN-TF model of the application are significantly better than those of the pure CNN, pure Transformer and TF-CNN models, verifying the effectiveness of the hybrid architecture design of the application. The method reduces the MAE index by 17.8% compared with the suboptimal model, further proving the effectiveness of CNN-TF in improving the accuracy of remaining time prediction.

[0032] In summary, the hybrid architecture of the application effectively fuses the local patterns and global context information of event sequences, and realizes accurate prediction of the remaining time of business processes. Through the processing of the above steps, the input event prefix is finally mapped to the remaining time prediction value. The experimental results verify the superior performance of the CNN-TF model on the BPI_13_Incidents dataset.

[0033] Of course, the above description is not a limitation of the present application, and the present application is not limited to the above examples. Changes, modifications, additions or substitutions made by those skilled in the art within the spirit and scope of the present application should also be included in the protection scope of the present application.

Claims

1. A hybrid architecture based business process remaining time prediction method, characterized in that, The mixed architecture comprises an event embedding model, a multi-scale one-dimensional convolutional neural network, a time-aware Transformer encoder, and a fully connected output layer. Step 1: Obtain the original event log of the business process, extract the prefix of each trace in the original event log, and generate a set of event sequence prefixes to be predicted. Step 2: Based on the event embedding model, each event in the event sequence prefix set is vectorized and embedded, and a position encoding vector and a time difference vector are introduced to obtain a time-aware embedding representation sequence that fuses position information and time information. Step 3: The time-aware embedding representation sequence is input into the multi-scale one-dimensional convolutional neural network for convolution operation, and multi-level local features are extracted through convolution operations with different kernel sizes, and then fused to obtain a multi-scale local feature representation sequence of the event sequence prefix set. Step 4: The multi-scale local feature representation sequence is input into the time-aware Transformer encoder to learn the global context dependency in the event sequence prefix set through the multi-head self-attention mechanism; a weight decay factor based on the time difference between events is introduced during attention calculation, and finally a global feature representation sequence containing time-dependent information is calculated. Step 5: Aggregate the global feature representation sequence and calculate the remaining time prediction value of the business process through the fully connected output layer.

2. The method of claim 1, wherein the remaining time of the business process is predicted based on the hybrid architecture. The specific process of step 1 is as follows: Step 1.1: Obtain the original event log from a public dataset website. The original event log is a structured file in XES format, containing event activity name, timestamp, event ID, event-associated performer, organizational role, resource type, event lifecycle identifier, and other optional business attributes. The original event log is a collection of several traces, and a trace is a sequence composed of several event activities. Define the original event log as: (1)； wherein, is the original event log; is the total number of traces; is the i-th trace, in particular: (2)； wherein, is the th event in the th trajectory; Each event is composed of an activity, a timestamp, and a set of context attribute embeddings, which are derived from the event context attributes recorded in the original event log, including event-associated performer, organizational role, resource type, event lifecycle identifier, and other optional business attributes. The context attribute embedding set is composed of multiple context attributes associated with the current event. When a certain type of attribute is missing, a default placeholder is used. Step 1.2, for each trajectory in the original event log, extract all prefixes of length of the trajectory in non-decreasing order of event timestamps, where ; is the number of events contained in the thtrajectory; The trace prefix is defined as: (3)； wherein, is a track prefix; is a first of a first event in a track, corresponding track prefix length; The event sequence prefix set is constituted by all prefixes of all trajectories : (4)。 3.The hybrid-architecture-based business process remaining time prediction method according to claim 2, characterized in that, The specific process of step 2 is as follows: Step 2.1: Use the event embedding model to vectorize and embed each event in the trace prefix to obtain an event embedding vector: (5)； wherein, is an embedding model for events; is an embedding model for events; is an activity of the event in the th trajectory, is an embedding model for events; is an embedding representation of the contextual attribute in the set of contextual attribute embeddings corresponding to the th event in the th trajectory, is an embedding vector for the th event in the th trajectory, is a set of contextual attribute embeddings corresponding to the th event in the th trajectory. Step 2.2, introduce position encoding for the event embedding vector, to obtain a position encoding vector, the first dimension and the second dimension of the position encoding vector are calculated as follows respectively: (6)； (7)； wherein, is the th event in the th track of the th position index, ; Step 2.3: Calculate the time interval between adjacent events and map it to a time difference vector through a time embedding function: (8)； (9)； wherein, is the th event in the th trajectory; , is the th event in the th trajectory; th event in the th trajectory; is the th event in the th trajectory; is a time embedding function; Step 2.4: Add the embedding vector of each event to its corresponding position encoding vector and time difference vector element by element to obtain an event embedding representation that fuses position information and time information: (10)； wherein, is the th event in the th trajectory; The embedded representations of all events constitute a sequence of time-aware embedded representations of trajectory prefixes The embedded representation of the th event corresponds to the th position in the sequence of time-aware embeddings.

4. The method of claim 3, wherein the remaining time of the business process is predicted based on the hybrid architecture. The specific process of step 3 is as follows: Step 3.1, for the time-aware embedding representation sequence, a one-dimensional convolutional neural network with multiple different convolution kernel sizes is used for parallel convolution operation to extract multi-level local feature values; for the convolution operation of the first convolution kernel, the calculation formula is: ​ (11)； wherein, is a vector representation of a th event activity in the sequence of time-aware embedding representations; is a weight parameter of the th convolutional kernel at position ; is a local feature generated by the th convolutional kernel; is a non-linear activation function; is an embedding vector of a th event in the sequence of time-aware embedding representations based on the input, after being computed by the th convolutional kernel; is a local feature of the th convolutional kernel; Step 3.2, concatenate the local features generated by all convolution kernels of the first event to obtain a multi-scale local feature vector of the first event in the event sequence prefix set Step 3.2, concatenate the local features generated by all convolution kernels of the first event to obtain a multi-scale local feature vector of the first event in the event sequence prefix set Step 3.2, concatenate the local features generated by all convolution kernels of the first event to obtain a multi-scale local feature vector of the first event in the event sequence prefix set (12)； wherein, denotes the number of multi-scale convolution kernels; is an embedding vector representing the th event in the input sequence based on the time-aware embedding representation, and is a local feature obtained after the th convolution kernel calculation; is a multi-scale local feature vector after fusion of the th event. The multi-scale local feature vector of all events constitutes a multi-scale local feature representation sequence wherein is the length of the feature sequence of the convolution output.

5. The method of claim 4, wherein the remaining time of the business process is predicted based on the hybrid architecture. The specific process of step 4 is as follows: Step 4.1, obtaining a sequence of multi-scale local feature representations As the input of the Transformer encoder, the attention weights between positions in the sequence are calculated using the multi-head self-attention mechanism, and a weight decay factor of the time difference between events is introduced into the attention score, and the calculation formula is: (13)； wherein, is a multi-head self-attention mechanism; respectively represent the sequence of multi-scale local feature representations obtained query matrix, key matrix and value matrix; is a transpose symbol; is the dimension of the key vector for each attention head; is an inter-event time difference matrix; is a weight decay factor; denotes normalizing each row of the matrix; Step 4.2: The feature representation output by the multi-head self-attention mechanism is processed through a feedforward neural network, residual connections, and layer normalization to obtain a global feature representation sequence containing the event sequence prefix set with time-dependent information. ; for The Middle The global feature vector corresponding to each position The total number of positions; the specific process is as follows: The feature representation output by the multi-head self-attention mechanism constitutes a multi-head self-attention output matrix For the multi-head self-attention output matrix, the calculation formula through the feedforward neural network is: (14)； wherein is a feedforward neural network; , is a weight matrix of the feedforward neural network; , is a bias term of the feedforward neural network; The output of the feedforward neural network is further connected with a residual connection and subjected to layer normalization to obtain a global feature vector : (15)； wherein, denotes a layer normalization operation that normalizes each dimension of a vector; Finally, the global feature vectors corresponding to all positions form a global feature representation sequence .

6. The method of claim 5, wherein the remaining time of the business process is predicted based on the hybrid architecture. The specific process of step 5 is as follows: Step 5.1, computing global feature representation sequence performing an average pooling operation to obtain an aggregated global feature vector , the formula is: (16)； wherein, is the first global feature vector corresponding to the position, the position number corresponding to the event number; Step 5.2, the aggregated global feature vector Input fully connected output layer, calculate the remaining time prediction value of the business process The calculation formula is: (17)； wherein, is a weight vector for the fully connected output layer; is a bias term.

7. A hybrid architecture based business process remaining time prediction system, characterized in that, The method comprises the following steps: obtaining an event prefix set from an original event log of a business process; mapping each event activity in the event prefix set and its associated context attributes into a corresponding embedding vector, and introducing position encoding and time-aware encoding based on the time interval between events into each embedding vector to generate a time-aware embedding representation sequence containing business context, position information and time information; performing one-dimensional convolution operation with different convolution kernel sizes on the time-aware embedding representation sequence to extract multi-level local features, and fusing the multi-level local features to obtain a multi-scale local feature representation sequence of the event prefix set; performing encoding operation based on the multi-head self-attention mechanism on the multi-scale local feature representation sequence, and introducing a weight decay factor based on the time difference between events in the attention calculation to learn the global time sequence dependency in the event prefix set, thereby obtaining a global feature representation sequence of the event prefix set; and calculating the business process remaining time prediction value corresponding to the event prefix set according to the global feature representation sequence.

Images