Method, system, and media for missing activity repair of business process based on missing awareness

Abstract

The present application relates to the technical field of electric digital data processing, in particular to a business process missing activity repair method and system based on missing perception and medium; the method encodes the order, concurrency and mutual exclusion relationship between activities in the event log into a two-dimensional relationship matrix, carries out multi-layer feature extraction processing based on different area convolution, obtains low-frequency features reflecting stable path patterns, medium-frequency features reflecting local branch structure features, and high-frequency features reflecting burst behavior patterns; then, taking the medium-frequency features as semantic anchor points, the multi-spectrum features are aligned and fused, the fused features are flattened into one-dimensional tensors after size adjustment and channel padding, and after one-dimensional convolution and global pooling, nonlinear processing is carried out to obtain missing activity prediction results; finally, based on the missing activity prediction results, the missing activities are filled in the event log to be repaired to obtain a repaired event log, and the method can significantly improve the activity missing repair rate.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of digital data processing, in particular to a missing activity repair method and system for business process based on missing perception and a medium. BACKGROUND

[0002] Under the background of accelerating digital transformation, business process management system (BPMS) has increasingly become an important support for enterprises to optimize process efficiency and strengthen governance capability. By standardizing process execution and monitoring business status in real time, BPMS provides data-driven decision support for management. As the core technology of BPMS, process mining extracts process models and behavior rules from system-generated event logs to build a bridge between data and process management.

[0003] Process mining mainly includes three tasks: process discovery, consistency checking and process enhancement, and the reliability of these tasks highly depends on the quality of event logs. Event logs record the name, time, resource and other information of each activity in the process execution process, and are the key basis for modeling process behavior. However, in actual business systems, due to system failure, cross-system synchronization delay or manual input error, etc., logs often have problems such as activity missing, time sequence disorder and noise. Among them, activity missing is particularly common and far-reaching - for example, the missing of the "diagnosis" step in medical logs may lead to modeling errors of clinical pathways; in financial approval, the missing of "risk assessment" may also cause serious decision bias.

[0004] For missing activity repair, existing methods are mainly divided into three categories. Model-based methods are limited in modeling flexibility and are difficult to adapt to dynamic process structure; model-free methods rely on statistical features and are difficult to capture deep semantics and context dependence; deep learning methods have automatic feature extraction capability, but there are still shortcomings in scale fusion and behavior relationship modeling, especially in high missing rate or complex structure scenarios. SUMMARY

[0005] The purpose of the present application is to provide a missing activity repair method and system for business process based on missing perception and a medium.

[0006] The technical scheme of the present application is as follows:

[0007] A missing activity repair method for business process based on missing perception, comprising the following operations:

[0008] S1, a plurality of historical complete event log samples are obtained, a plurality of missing event log samples are obtained through activity deletion processing based on a control activity deletion ratio, a plurality of historical complete event log samples and a plurality of missing event log samples form a training sample pair, an activity missing prediction model is trained, network parameters are updated through back propagation during the training process, the model learns the nonlinear mapping relationship between the missing pattern and the complete activity relationship, and an activity missing prediction training model is obtained;

[0009] S2, the missing activity prediction result is obtained by processing the to-be-repaired event log through the activity missing prediction training model.

[0010] The processing process of the activity missing prediction training model is: the activity relationship matrix of the to-be-repaired event log is processed through multi-layer feature extraction to obtain low-frequency features, medium-frequency features and high-frequency features; the low-frequency features, the medium-frequency features and the high-frequency features are fused to obtain fusion features; after one-dimensional convolution and global pooling, the fusion features are processed through non-linear processing to obtain the missing activity prediction result.

[0011] S3, based on the missing activity prediction result, the activity of the to-be-repaired event log is filled to obtain a repaired event log.

[0012] The event log in S1 is a set of multiple traces, a trace is a sequence of events arranged in time sequence in the same process instance, and an event is a specific execution instance of an activity in a business process.

[0013] In S2, the activity relationship matrix is constructed based on the relationship between activities in the to-be-repaired event log, and the relationship between activities includes sequential relationship, concurrent relationship, mutual exclusion relationship and missing relationship.

[0014] In S2, the operation of multi-layer feature extraction processing is: the activity relationship matrix of the to-be-repaired event log is processed through partial convolution to obtain low-frequency features; the activity relationship matrix of the to-be-repaired event log is processed through partial convolution and two-dimensional convolution to obtain medium-frequency features; the activity relationship matrix of the to-be-repaired event log is processed through partial convolution, two-dimensional convolution and three-dimensional convolution to obtain high-frequency features.

[0015] The operation of partial convolution is: the activity relationship matrix is processed through padding in different directions to obtain different padded matrices; each padded matrix is convolved with a corresponding convolution kernel, and then sequentially processed through batch normalization and non-linear processing to obtain corresponding regional features; all regional features are spliced, and then processed through a convolution layer, batch normalization and non-linear processing to obtain partial convolution features, which are used as low-frequency features or perform two-dimensional convolution operations.

[0016] The multi-spectral fusion processing operation is as follows: the query feature of the mid-frequency feature, the key feature of the low-frequency feature, and the value feature of the high-frequency feature are subjected to attention processing to obtain the first reconstructed feature; the query feature of the mid-frequency feature, the key feature of the high-frequency feature, and the value feature of the low-frequency feature are subjected to attention processing to obtain the second reconstructed feature; the mid-frequency feature, the first reconstructed feature, and the second reconstructed feature are added element-wise and then processed by the ReLU activation function to obtain the fused feature.

[0017] During the training of the S1 missing activity prediction model, the loss function L The calculation formula is as follows:

[0018] ,

[0019] For the number of activity categories, C The total number of activity categories. For the first i One-hot encoding of a real label, For the first i The probability distribution of a missing activity log sample.

[0020] A missing activity repair system for business processes based on missing awareness, used to implement the above-mentioned missing activity repair method for business processes based on missing awareness, includes:

[0021] The activity missing prediction training model generation module is used to obtain multiple historical complete event log samples. After activity deletion processing based on the controlled activity deletion ratio, multiple missing event log samples are obtained. Multiple historical complete event log samples and multiple missing event log samples form training sample pairs to train the activity missing prediction model. During the training process, the network parameters are updated through backpropagation so that the model learns the nonlinear mapping relationship between missing patterns and complete activities, thus obtaining the activity missing prediction training model.

[0022] The missing activity prediction result generation module is used to process the event logs to be repaired by the missing activity prediction training model to obtain the missing activity prediction results. The processing of the missing activity prediction training model is as follows: the activity relationship matrix of the event logs to be repaired is processed by multi-layer feature extraction to obtain low-frequency features, mid-frequency features and high-frequency features; the low-frequency features, mid-frequency features and high-frequency features are fused by multi-spectral fusion to obtain fused features; the fused features are then processed by one-dimensional convolution and global pooling, and then non-linear processing is performed to obtain the missing activity prediction results.

[0023] The event log generation module is used to complete the event logs to be repaired based on the missing activity prediction results, thus obtaining the repaired event logs.

[0024] A missing-aware business process missing activity repairing device comprises a processor and a memory, wherein the processor implements the missing-aware business process missing activity repairing method as described above when executing the computer program stored in the memory.

[0025] A computer readable storage medium for storing a computer program, wherein the computer program is executed by a processor to implement the missing-aware business process missing activity repairing method as described above.

[0026] The present application has the following beneficial effects:

[0027] The missing-aware business process missing activity repairing method provided by the present application encodes the order, concurrency and mutual exclusion relationship between activities in the event log into a two-dimensional relationship matrix, performs multi-layer feature extraction processing based on different area convolutions, obtains low-frequency features reflecting stable path patterns, medium-frequency features reflecting local branch structure features, and high-frequency features reflecting burst behavior patterns, takes the medium-frequency features as semantic anchor points, performs alignment and fusion of multi-spectrum features, flattens the fused features into a one-dimensional tensor after size adjustment and channel padding, performs nonlinear processing after one-dimensional convolution and global pooling, and obtains missing activity prediction results, and finally, based on the missing activity prediction results, fills in activities for the event log to be repaired to obtain a repaired event log; the method can significantly improve the missing activity repairing rate, the accuracy rate reaches 80%-96%, and can improve the repairing performance stability in high missing rate or complex structure scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0028] The schemes and advantages of the present application will become clear to those skilled in the art by reading the following detailed description of the preferred embodiments. The accompanying drawings are only for the purpose of illustrating the preferred embodiments and are not considered as limiting the present application.

[0029] In the drawings:

[0030] Figure 1 For the embodiment, a flowchart of the method of the embodiment is shown;

[0031] Figure 2 For the embodiment, a comparison chart of repairing success rates under different missing activity quantities of the data set BPIC 2012 data set is shown;

[0032] Figure 3 For the embodiment, a comparison chart of repairing success rates under different missing activity percentages of the data set BPIC 2013 data set is shown. DETAILED DESCRIPTION

[0033] The exemplary embodiments of the present disclosure will be described in greater detail below with reference to the accompanying drawings.

[0034] Embodiment 1

[0035] The embodiment provides a missing activity repair method for a business process based on missing perception, and the method is described below with reference to Figure 1 , and includes the following operations:

[0036] S1, a plurality of historical complete event log samples are acquired, and a plurality of missing event log samples are obtained through activity deletion processing based on a control activity deletion ratio; the plurality of historical complete event log samples and the plurality of missing event log samples form a training sample pair, an activity missing prediction model is trained, network parameters are updated through back propagation in the training process, the model learns a nonlinear mapping relationship between a missing mode and a complete activity relationship, and an activity missing prediction training model is obtained;

[0037] S2, the missing activity prediction result is obtained through processing of the activity missing prediction training model on the event log to be repaired;

[0038] The processing process of the activity missing prediction training model is as follows: the activity relationship matrix of the event log to be repaired is processed through multi-layer feature extraction to obtain low-frequency features, medium-frequency features and high-frequency features; the low-frequency features, the medium-frequency features and the high-frequency features are subjected to multi-spectrum fusion to obtain fused features; after the fused features are subjected to one-dimensional convolution and global pooling, nonlinear processing is performed to obtain the missing activity prediction result;

[0039] S3, based on the missing activity prediction result, activity completion is performed on the event log to be repaired to obtain a repaired event log.

[0040] The specific steps are as follows.

[0041] S1, a plurality of historical complete event log samples are acquired, and a plurality of missing activity log samples are obtained through activity deletion processing based on a control activity deletion ratio; the plurality of historical complete event log samples and the plurality of missing event log samples form a training sample pair, an activity missing prediction model is trained, network parameters are updated through back propagation in the training process, the model learns a nonlinear mapping relationship between a missing mode and a complete activity relationship, and an activity missing prediction training model is obtained.

[0042] First, the event and time log are defined as follows.

[0043] Event and attribute. An event is a specific execution instance of an activity in a business process, including key information in the execution process, and each event is composed of a set of attributes. The event is denoted as , and the attribute set is denoted as , wherein is a case identifier, is an event identifier, is an activity name, is a timestamp, For resources.

[0044] Trace and event log. A trace is a sequence of events in a same process instance ordered by time, reflecting the complete execution path of a single process instance. An event log is a collection of multiple traces, covering all instance execution records of a certain business process. A trace is denoted as , where is the length of the trace, denotes the th event. An event log is denoted as , where is the number of traces in the log, denotes the th trace.

[0045] Missing attribute and missing event. A missing attribute refers to the value of a certain attribute in an event being empty. If an event contains at least one missing attribute, the event is called a missing event. For an event , if exists, then is a missing attribute, and is a missing event, denoted as , is a set of missing events.

[0046] Missing trace and missing log. A missing trace is a trace containing at least one missing event. If there exists at least one missing trace in an event log, the log is called a missing log. For a trace , if exists, then is a missing trace, denoted as , is a set of missing traces. If an event log satisfies , then is a missing log, denoted as , is a set of missing logs.

[0047] Then, a plurality of historical complete event log samples are obtained, and a plurality of missing event log samples are obtained through event deletion processing based on a control activity deletion ratio. The missing rate in the missing event log samples is 15%, or 20%, or 30%, and the missing positions are randomly distributed, so as to simulate different degrees of activity missing in actual scenarios.

[0048] Then, a plurality of historical complete event log samples (as labels) and a plurality of missing event log samples form a training sample pair, which is divided into a training set and a test set in a ratio of 7:3, a missing activity prediction model is trained, and network parameters are updated by back propagation optimization during the training process, so that the model learns the nonlinear mapping relationship between the missing pattern and the complete activity relationship, and a model with missing activity prediction capability, i.e., an activity missing prediction training model, is obtained.

[0049] The activity missing prediction model is composed of an input module, a multi-layer feature extraction module, a multi-spectrum fusion module, and a prediction output module in series, corresponding to input operation, multi-layer feature extraction operation, multi-spectrum fusion operation, and prediction output operation (one-dimensional convolution, global pooling, nonlinear processing-LogSoftmax activation function), and the corresponding processing process is described below.

[0050] During the training process, the loss function L The calculation formula is as follows:

[0051] ,

[0052] is the number of activity categories, C is the total number of activity categories, is the one-hot encoding of the true label, is the probability distribution of the model output, and the loss function can effectively guide the model to focus on the classification of missing activities, adapt to the LogSoftmax activation function of the model output layer, and improve the classification accuracy. The optimizer uses the Adam algorithm, and the initial learning rate is set to , and the regularization weight decay is combined to suppress the risk of overfitting, and the weight decay coefficient is set to .

[0053] In addition, during the training process, Batch Normalization is introduced after each convolution operation in the activity missing prediction model to standardize the feature distribution and accelerate convergence. To avoid overfitting of the model to the training data, a Dropout layer is added between the multi-spectrum fusion module and the output layer, and the dropout rate is set to 0.5. By randomly discarding part of the features, the generalization ability of the model is enhanced. The hyperparameters of the training process are set as follows: batch size 32 to balance the calculation efficiency and gradient stability; maximum iteration number 100, and early stopping mechanism is introduced, if the validation set accuracy does not improve for 10 consecutive rounds, the training is terminated, to avoid invalid iterations.

[0054] During the back propagation process, the gradient of the loss function with respect to the parameters of each layer is calculated by the chain rule, including the asymmetric convolution kernel weights of some convolutions in the multi-layer feature extraction module, the attention parameters of the multi-spectrum fusion module, and the weights of the fully connected layer, etc. Then the parameter update is completed by the Adam optimizer. During the training process, the gradient weighting processing is performed on the mask area of the missing position, and the weight coefficient is set to 1.2 to force the model to enhance the perception ability of the context of the missing neighborhood; at the same time, mixed precision calculation is used to accelerate the training process, and the numerical stability is ensured when processing high-dimensional sparse activity relationship matrix. After each round of training, the repair accuracy is evaluated on the validation set, and the learning rate is dynamically adjusted. Finally, the model parameters with the best performance on the validation set are saved for the missing activity prediction in the test phase.

[0055] S2, the event log to be repaired is processed by the activity missing prediction training model to obtain a missing activity prediction result.

[0056] The processing procedure of the activity missing prediction training model is as follows.

[0057] Step 1, the activity relationship matrix of the event log to be repaired is processed by multi-layer feature extraction to obtain low-frequency features, medium-frequency features and high-frequency features.

[0058] The activity relationship matrix is constructed based on the relationship between activities in the event log to be repaired, and is used to enhance the modeling ability of the model for the behavior structure of the business process. The relationship between activities includes sequential relationship, concurrent relationship, mutual exclusion relationship and missing relationship.

[0059] If activity always appears before activity and the two activities do not overlap in time, it is called a sequential relationship, denoted as ; if and appear together in the trajectory multiple times, and the order is not fixed or there is a possibility of concurrent execution, it is defined as a concurrent relationship, denoted as ; if and never appear together in all trajectories, it is considered that there is strong structural separation between them, and it is defined as a mutual exclusion relationship, denoted as In addition, in order to enhance the attention ability of the model to the missing area, if there is a missing activity, the relationship with other activities is defined as a missing relationship, and a missing mask mechanism is introduced to explicitly guide the model to focus on the neighborhood relationship of the missing position.

[0060] The operation of the above multi-layer feature extraction processing is that the activity relationship matrix of the event log to be repaired is processed by partial convolution to obtain low-frequency features reflecting stable path patterns; the activity relationship matrix of the event log to be repaired is processed by partial convolution and two-dimensional convolution to obtain medium-frequency features reflecting local branch structure features; the activity relationship matrix of the event log to be repaired is processed by partial convolution, two-dimensional convolution and three-dimensional convolution to capture burst behavior patterns, and high-frequency features are obtained.

[0061] In the business process log repair task, activity loss will cause structural rupture of the activity relationship matrix, and the traditional convolution is not sensitive to the direction of the missing position neighborhood, and it is difficult to capture the context dependence at the rupture. For example, after the loss of an activity, the order relationship between the activities before and after it needs to be inferred from multiple directions. Therefore, the embodiment designs a partial convolution processing method, and the core advantage is to accurately perceive the forward dependence of the missing position in the order relationship, the bidirectional interaction in the concurrent relationship and the exclusion boundary in the mutual exclusion relationship through asymmetric convolution kernels and multi-directional feature extraction, thereby enhancing the modeling ability of irregular missing activity patterns in the log.

[0062] The operation of the partial convolution is that the activity relationship matrix is respectively processed by different direction filling to obtain different filled matrices; each filled matrix is convolved with the corresponding convolution kernel, and then sequentially subjected to batch normalization and nonlinear processing to obtain the corresponding regional features; all regional features are spliced, subjected to convolution layer, batch normalization and nonlinear processing to obtain partial convolution features, which are used as low-frequency features or subjected to two-dimensional convolution operation.

[0063] The partial convolution is realized by the following formula:

[0064] ,

[0065] ,

[0066] ,

[0067] ,

[0068] ,

[0069] Y’ 、 Y 1 、 Y 2 、 Y 3 、 Y 4 respectively as partial convolution features, first regional features, second regional features, third regional features, fourth regional features, represents filling 1 0 in the left direction of the active relation matrix, filling 3 0 in the lower direction, and not filling in the right and upper directions, represents filling 3 0 in the right direction of the active relation matrix, filling 1 0 in the upper direction, and not filling in the left and upper directions, represents filling 3 0 in the left direction of the active relation matrix, filling 1 0 in the upper direction, and not filling in the right and lower directions, represents filling 1 0 in the right direction of the active relation matrix, filling 3 0 in the upper direction, and not filling in the left and lower directions, represents the direction convolution kernel, i = 1, 2, 3, c ’ is the number of output channels, The convolution kernel is used to extract the features of the horizontal direction and the lower left region; The convolution kernel is used to extract the features of the vertical direction and the lower right region, The convolution kernel is used to extract the features of the horizontal direction and the upper right region, The convolution kernel is used to extract the features of the vertical direction and the upper left region of the image, BN( ) is the batch normalization processing, SiLU( ) is SiLU The function processing, Concat( ) is the splicing processing, Conv fuse ( ) is the feature fusion and dimension adjustment of the spliced matrix through a convolution layer, which integrates the feature information of the four branches and adjusts the high channel number after splicing to adapt to the subsequent calculation, so that the features are more compact and effective.

[0070] Step 2, the low-frequency feature, the medium-frequency feature and the high-frequency feature are fused to obtain the fusion feature which provides context-sensitive multi-scale semantic representation for missing activity prediction.

[0071] The operation of multi-spectrum fusion processing is: the query feature of the medium-frequency feature is processed by attention with the key feature of the low-frequency feature and the value feature of the high-frequency feature to obtain the first reconstruction feature; the query feature of the medium-frequency feature is processed by attention with the key feature of the high-frequency feature and the value feature of the low-frequency feature to obtain the second reconstruction feature; after the element-by-element addition of the medium-frequency feature, the first reconstruction feature and the second reconstruction feature, the ReLU activation function processing is performed to obtain the fusion feature. The multi-spectrum fusion processing can significantly improve the accuracy and robustness of missing position positioning and activity prediction in multi-missing, multi-branch and high-noise logs by aligning the structure semantics of different trajectories.

[0072] The operation of multi-spectrum fusion processing corresponds to the following calculation formula:

[0073] ,

[0074] ,

[0075] ,

[0076] , , , are the fusion feature, the mid-frequency feature, the first reconstruction feature, the second reconstruction feature, respectively, , , , , are the query feature of the mid-frequency feature, the key feature of the low-frequency feature, the value feature of the low-frequency feature, the key feature of the high-frequency feature, the value feature of the high-frequency feature, respectively, is the number of channels, softmax( ) is processed by the softmax function, ReLU( ) is processed by the ReLU activation function.

[0077] Step 3, after one-dimensional convolution and global pooling of the fusion feature, nonlinear processing (which can be realized by the LogSoftmax activation function) is performed to obtain the missing activity prediction result.

[0078] S3, based on the missing activity prediction result, the activity of the to-be-repaired event log is filled in to obtain the repaired event log.

[0079] To verify the effectiveness of the method (hereinafter referred to as PMSF-Net) in the task of repairing missing activities in business processes, three groups of comparative experiments are designed. First, through the ablation experiment, the gain of multi-spectrum fusion on the performance of the model is verified; second, compared with the traditional model-free method, the advantages of the model in the single missing activity repair scene are evaluated; third, compared with the mainstream deep learning method, the robustness of the model in the complex process log and high missing rate scene is tested, and finally the practicability and advancement of PMSF-Net are comprehensively demonstrated.

[0080] Experimental data. Six event log datasets with domain representativeness and structural diversity are selected for the experiment, covering real business and artificially constructed scenarios. Among them, BPIC 2012 is the loan application process log of a Dutch financial institution, BPIC 2013 is the event management log of Volvo IT department, Roadfine comes from the traffic fine processing system of Italian police, Hospital is the medical billing process log of a regional hospital, Help Desk is the IT service desk work order log of a certain enterprise, and Log is an artificial dataset. The detailed information is shown in Table 1.

[0081] Table 1 Basic statistical information table of dataset

[0082]

[0083] To evaluate the performance of the method, the success rate is used in the experiment success Rate As the core evaluation index, it measures the proportion of the number of successfully repaired missing activities to the total number of missing activities. Let the number of successfully repaired activities, the total number of missing activities in the data set, the success rate is defined as follows:

[0084] ,

[0085] In the experiment, all models are ensured to run under the same training test division, optimizer and hyperparameter to eliminate irrelevant variable interference and ensure the fairness of the index comparison.

[0086] Experimental results. To verify the performance advantage of PMSF-Net, the experiment is carried out from three dimensions of ablation experiment to verify the necessity of the module, traditional method comparison and deep learning method comparison.

[0087] Table 2 Repair ability table

[0088]

[0089] The ablation experiment takes six data sets as test objects, sets the missing rate to 15%, 20% and 30%, and the missing position is randomly distributed. The missing proportion is calculated according to the total number of activities. The performance of the complete model and the baseline model is compared, and the results are shown in Table 2.

[0090] In the regular process, PMSF-Net improves by 84.8%-133.8% compared with PROELR, always leading the baseline models such as SRBA and Transformer, and still maintaining a stable advantage when the missing rate increases.

[0091] In the complex process, PMSF-Net has more prominent improvement effect, improving by 96.7%-133.2% compared with PROELR, not only leading the traditional baseline, but also being superior to the deep learning model such as CNN, and having better stability against data missing.

[0092] The core reason for this difference lies in the multi-spectrum fusion through the medium frequency feature anchor point, which dynamically fuses the low scale stable process rules and high scale sudden behavior details, effectively alleviating the context rupture under high missing rate. While the variant model relies on single scale features, it is difficult to capture cross-level activity correlation, which ultimately leads to a decrease in repair accuracy in complex scenarios, confirming the core gain of multi-spectrum fusion for complex process repair.

[0093] For the data set BPIC 2012, set the number of missing activity events to 100, 150, 200 and 300, and compare PMSF-Net with traditional non-model methods, and the results are as follows Figure 2(x-axis is the number of different missing activities, y-axis is the repair success rate). Traditional model-free methods include PROELR based on trajectory similarity, SRBA based on activity dependency graph, and Autoencoder based on self-encoding reconstruction. Among them, the PROELR repair accuracy is only 43.2%~46.7%, the SRBA and Autoencoder are about 80.7% at most, and the performance fluctuates significantly with the increase of the number of events, for example, the SRBA decreases by 1.5% at 300 events compared with 150 events; while the PMSF-Net repair accuracy is always stable at 97.6%~98.2%. The limitation of traditional methods lies in the dependence on static rules or shallow feature reconstruction, for example, SRBA adopts fixed dependent edges, and Autoencoder only relies on single-dimensional error, which cannot dynamically model branch cycles; while PMSF-Net captures the forward and backward dependencies of missing positions through missing perception convolution, achieving accurate repair of long sequence missing.

[0094] In the comparison of deep learning methods, first, six data sets are compared with CNN and Transformer, and the results are shown in Table 2. In the regular process, the PMSF-Net accuracy is 0.949 when the BPIC 2012 has a 30% missing rate, which is higher than that of CNN and Transformer. In the complex process, the advantage is more obvious, the PMSF-Net accuracy is 0.868 when the Help Desk has a 30% missing rate, which is higher than that of CNN 0.806; the PMSF-Net accuracy is 0.969 when the Log has a 15% missing rate, which is higher than that of CNN 0.928, which proves its processing ability for artificial constructed multi-type activity relationship.

[0095] Focusing on BPIC 2013, further comparison of PMSF-Net with CNN, Transformer, and BERT is made, and the results are shown in Table 3. Figure 3 (x-axis is different missing percentage, y-axis is repair success rate). At low missing rate, the PMSF-Net accuracy is 0.850, which is better than that of CNN 0.835, BERT 0.813, and significantly surpasses that of Transformer 0.736; at high missing rate, the PMSF-Net accuracy is 0.804, which leads CNN by 8.1%, BERT by 7.2%, and Transformer by 15.7%, and the performance decrease is the smallest, only 4.6%, while CNN decreases by 11.2%. Analysis shows that CNN is difficult to handle concurrent activities due to the limitation of local convolution receptive field; Transformer and BERT are sensitive to semantic rupture caused by missing due to the dependence on self-attention; while PMSF-Net enhances the fault tolerance of complex control flow by direction-sensitive convolution perception of temporal association and multi-scale fusion of multi-spectrum fusion.

[0096] In summary, the necessity of multi-spectrum fusion, the efficient repair ability of PMSF-Net (the method of the embodiment) to the regular process, and the performance surpassing of the traditional method and the deep learning method in complex scenarios jointly verify the effectiveness and robustness of the method of the embodiment in the missing activity repair task of the business process.

[0097] The embodiment also provides a missing activity repair system for a business process based on missing perception, which is used for implementing the missing activity repair method for a business process based on missing perception.

[0098] The activity missing prediction training model generation module is used for obtaining a plurality of historical complete event log samples, and obtaining a plurality of missing event log samples through activity deletion processing based on a control activity deletion ratio. The plurality of historical complete event log samples and the plurality of missing event log samples form a training sample pair, and the activity missing prediction model is trained. Network parameters are updated through back propagation in the training process, so that the model learns the nonlinear mapping relationship between the missing mode and the complete activity relationship, and the activity missing prediction training model is obtained.

[0099] The missing activity prediction result generation module is used for obtaining a missing activity prediction result through processing of the event log to be repaired by the activity missing prediction training model. The processing process of the activity missing prediction training model is as follows: the activity relationship matrix of the event log to be repaired is processed through multi-layer feature extraction to obtain low-frequency features, medium-frequency features and high-frequency features. The low-frequency features, the medium-frequency features and the high-frequency features are subjected to multi-spectrum fusion to obtain fusion features. After one-dimensional convolution and global pooling, the fusion features are subjected to nonlinear processing to obtain the missing activity prediction result.

[0100] The repair event log generation module is used for performing activity completion on the event log to be repaired based on the missing activity prediction result to obtain a repair event log.

[0101] The embodiment also provides a missing activity repair device for a business process based on missing perception, which comprises a processor and a memory. When the processor executes a computer program stored in the memory, the missing activity repair method for a business process based on missing perception is implemented.

[0102] The embodiment also provides a computer readable storage medium for storing a computer program. When the computer program is executed by a processor, the missing activity repair method for a business process based on missing perception is implemented.

[0103] The embodiment provides a business process missing activity repairing method based on missing perception, first encodes the order, concurrency and mutual exclusion relationship between activities in an event log into a two-dimensional relationship matrix, carries out multi-layer feature extraction processing based on different area convolution, acquires low-frequency features reflecting stable path patterns, medium-frequency features reflecting local branch structure features, and high-frequency features reflecting burst behavior patterns; then, taking the medium-frequency features as semantic anchor points, carries out alignment and fusion of multi-spectrum features, flattens the fused features into a one-dimensional tensor after size adjustment and channel filling, carries out nonlinear processing after one-dimensional convolution and global pooling, and obtains missing activity prediction results; finally, based on the missing activity prediction results, activities are supplemented in the event log to be repaired, and a repaired event log is obtained; the method has a repairing accuracy of 80%-96% under different missing rates of six public data sets, and has an overall performance superior to that of a traditional method, especially in a complex process log with high concurrency and high missing rate, and thus provides an efficient solution for missing activity repairing.

Claims

1. A method for repairing missing activities in a business process based on missing information awareness, characterized in that, This includes the following operations: S1. Obtain multiple historical complete event log samples, and after activity deletion processing based on the controlled activity deletion ratio, obtain multiple missing event log samples; multiple historical complete event log samples and multiple missing event log samples form training sample pairs, train the activity missing prediction model, and update the network parameters through backpropagation during the training process, so that the model learns the nonlinear mapping relationship between missing patterns and complete activities, and obtains the activity missing prediction training model. S2. The log of events to be repaired is processed by the missing activity prediction training model to obtain the missing activity prediction results; The process of training the activity missing prediction model is as follows: the activity relationship matrix of the event log to be repaired is processed through multi-layer feature extraction to obtain low-frequency features, mid-frequency features and high-frequency features; The multi-layer feature extraction process is as follows: the activity relationship matrix of the event log to be repaired is partially convolutional to obtain low-frequency features; The activity relationship matrix of the event log to be repaired is processed by partial convolution and two-dimensional convolution to obtain mid-frequency features; the activity relationship matrix of the event log to be repaired is processed by partial convolution, two-dimensional convolution and three-dimensional convolution to obtain high-frequency features; the partial convolution operation is as follows: the activity relationship matrix is ​​padded in different directions to obtain matrices with different padded values; each padded matrix is ​​convolved with the corresponding convolution kernel, and then batch normalization and non-linear processing are performed in sequence to obtain the corresponding region features; All regional features are concatenated, and after convolutional layers, batch normalization, and nonlinear processing, partial convolutional features are obtained, which can be used as low-frequency features or to perform two-dimensional convolution operations. The low-frequency, mid-frequency, and high-frequency features are fused using a multi-spectral method to obtain the fused features. The fused features are processed by one-dimensional convolution and global pooling, and then non-linear processing is performed to obtain the missing activity prediction results. S3. Based on the missing activity prediction results, complete the activity in the event log to be repaired to obtain the repair event log.

2. The method for repairing missing activities in a business process based on missing information as described in claim 1, characterized in that, In S1, the event log is a collection of multiple traces. A trace is a sequence of events arranged in chronological order within the same process instance, and an event is a specific execution instance of an activity in the business process.

3. The method for repairing missing activities in a business process based on missing information as described in claim 1, characterized in that, In S2, the activity relationship matrix is ​​constructed based on the relationships between activities in the event log to be repaired. The relationships between activities include sequential relationships, concurrent relationships, mutual exclusion relationships, and missing relationships.

4. The method for repairing missing activities in a business process based on missing information as described in claim 1, characterized in that, In S2, the multi-spectral fusion processing operation is as follows: The query features of the mid-frequency features, the key features of the low-frequency features, and the value features of the high-frequency features are combined with attention processing to obtain the first reconstructed features; The query features of the mid-frequency features, the key features of the high-frequency features, and the value features of the low-frequency features are combined with attention processing to obtain the second reconstructed features; The intermediate frequency features, the first reconstructed features, and the second reconstructed features are added element by element and then processed by the ReLU activation function to obtain the fused features.

5. The method for repairing missing activities in a business process based on missing information as described in claim 1, characterized in that, During the training of the S1 missing activity prediction model, the loss function L The calculation formula is as follows: ， For the number of activity categories, C The total number of activity categories. For the first i One-hot encoding of a real label, For the first i The probability distribution of a missing activity log sample.

6. A missing activity repair system for business processes based on missing awareness, used to implement the missing activity repair method for business processes based on missing awareness as described in claim 1, characterized in that, include: The activity missing prediction training model generation module is used to obtain multiple historical complete event log samples. After activity deletion processing based on the controlled activity deletion ratio, multiple missing event log samples are obtained. Multiple historical complete event log samples and multiple missing event log samples form training sample pairs to train the activity missing prediction model. During the training process, the network parameters are updated through backpropagation so that the model learns the nonlinear mapping relationship between missing patterns and complete activities, thus obtaining the activity missing prediction training model. The missing activity prediction result generation module is used to process the event log to be repaired by the missing activity prediction training model to obtain the missing activity prediction result. The processing process of the missing activity prediction training model is as follows: the activity relationship matrix of the event log to be repaired is processed by multi-layer feature extraction to obtain low-frequency features, mid-frequency features and high-frequency features. The low-frequency, mid-frequency, and high-frequency features are fused using a multi-spectral method to obtain the fused features. The fused features are processed by one-dimensional convolution and global pooling, and then non-linear processing is performed to obtain the missing activity prediction results. The event log generation module is used to complete the event logs to be repaired based on the missing activity prediction results, thus obtaining the repaired event logs.

7. A business process missing activity repair device based on missing awareness, characterized in that, It includes a processor and a memory, wherein when the processor executes a computer program stored in the memory, it implements the missing activity repair method for business processes based on missing awareness as described in any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, Used to store a computer program, wherein the computer program, when executed by a processor, implements the missing activity repair method for business processes based on missing awareness as described in any one of claims 1-5.

Images