Business process missing activity repairing method and system based on missing perception and medium

Abstract

The invention relates to the technical field of electric digital data processing, in particular to a business process missing activity repairing method and system based on missing perception and a medium. The method comprises the following steps of: firstly, coding sequences, concurrency and mutual exclusion relationships among activities in an event log into a two-dimensional relationship matrix, and performing multi-layer feature extraction processing based on different region convolution to obtain low-frequency features reflecting a stable path mode and intermediate-frequency features reflecting local branch structure features; reflecting the high-frequency characteristics of the emergent behavior mode; then, performing alignment and fusion of multi-spectrum features by taking the intermediate frequency features as semantic anchor points, flattening the fused features into a one-dimensional tensor after size adjustment and channel filling, and performing nonlinear processing after one-dimensional convolution and global pooling to obtain a missing activity prediction result; and finally, based on the missing activity prediction result, performing activity completion on the to-be-repaired event log to obtain a repaired event log, and the method can significantly improve the activity missing repair rate.

Description

Technical Field

[0001] This invention relates to the field of electronic digital data processing technology, specifically to a method, system, and medium for repairing missing activities in business processes based on missing awareness. Background Technology

[0002] Against the backdrop of accelerated digital transformation, Business Process Management Systems (BPMS) are increasingly becoming a crucial support for enterprises to optimize process efficiency and strengthen governance capabilities. By standardizing process execution and monitoring business status in real time, BPMS provides management with data-driven decision support. As a core technology of BPMS, process mining builds a bridge between data and process management by extracting process models and behavioral patterns from system-generated event logs.

[0003] Process mining primarily comprises three tasks: process discovery, consistency checks, and process enhancement. The reliability of these tasks highly depends on the quality of event logs. Event logs record information such as the name, time, and resources of each activity during process execution, serving as crucial evidence for modeling process behavior. However, in real-world business systems, logs often suffer from issues such as missing activities, outdated timing, and noise due to system failures, cross-system synchronization delays, or manual input errors. Among these, missing activities are particularly common and have far-reaching consequences—for example, the absence of the "diagnosis" step in medical logs can lead to errors in clinical pathway modeling; in financial approvals, a lack of "risk assessment" can also cause serious decision-making biases.

[0004] Existing methods for missing activity repair can be mainly divided into three categories. Model-based methods are limited in modeling flexibility and have difficulty adapting to dynamic process structures; model-free methods rely on statistical features and have difficulty capturing deep semantics and contextual dependencies; although deep learning methods have automatic feature extraction capabilities, they are still lacking in scale fusion and behavioral relationship modeling, especially in scenarios with high missing rates or complex structures where performance is unstable. Summary of the Invention

[0005] The purpose of this invention is to provide a method, system, and medium for repairing missing activities in business processes based on missing awareness.

[0006] The technical solution of this invention is as follows: A method for repairing missing activities in a business process based on missing activity awareness 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 missing activity 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, mid-frequency and high-frequency features; the low-frequency, mid-frequency and high-frequency features are fused into multi-spectral features 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. S3. Based on the missing activity prediction results, complete the activity in the event log to be repaired to obtain the repair event log.

[0007] 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.

[0008] 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.

[0009] In S2, the multi-layer feature extraction process is as follows: the activity relationship matrix of the event log to be repaired is processed by partial convolution 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.

[0010] The partial convolution operation is as follows: the activity relation 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 region features are concatenated, and after convolutional layers, batch normalization and non-linear processing are performed to obtain partial convolution features, which are used as low-frequency features or to perform two-dimensional convolution operations.

[0011] 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.

[0012] 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.

[0013] 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: 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 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. 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.

[0014] A missing activity repair device for a business process based on missing awareness includes a processor and a memory, wherein the processor implements the above-described missing activity repair method for a business process based on missing awareness when executing a computer program stored in the memory.

[0015] A computer-readable storage medium for storing a computer program, wherein the computer program, when executed by a processor, implements the above-described method for repairing missing activities in a business process based on missing awareness.

[0016] The beneficial effects of this invention are as follows: This invention provides a method for repairing missing activities in business processes based on missing activity awareness. First, the order, concurrency, and mutual exclusion relationships between activities in the event log are encoded into a two-dimensional relationship matrix. Multi-layer feature extraction processing based on convolution in different regions is then performed to obtain low-frequency features reflecting stable path patterns, mid-frequency features reflecting local branch structure features, and high-frequency features reflecting sudden behavior patterns. Next, using the mid-frequency features as semantic anchors, multi-spectral features are aligned and fused. The fused features are then flattened into a one-dimensional tensor after size adjustment and channel filling. This tensor undergoes one-dimensional convolution and global pooling followed by non-linear processing to obtain the missing activity prediction result. Finally, based on the missing activity prediction result, activities are filled into the event log to be repaired, resulting in a repaired event log. This method can significantly improve the activity missing repair rate, achieving an accuracy of 80%–96%, and can improve the stability of repair performance in scenarios with high missing rates or complex structures. Attached Figure Description

[0017] The solutions and advantages of this application will become clear to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of the invention.

[0018] In the attached diagram: Figure 1 This is a flowchart illustrating the method of this embodiment. Figure 2 This is a comparison chart of repair success rates for different numbers of missing activities in the BPIC 2012 dataset, as shown in the example. Figure 3 This is a comparison chart of repair success rates under different percentages of missing activity in the BPIC 2013 dataset, as shown in the example. Detailed Implementation

[0019] Exemplary embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings.

[0020] Example 1 This embodiment provides a method for repairing missing activities in a business process based on missing activity awareness. (See also...) Figure 1 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 missing activity 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, mid-frequency and high-frequency features; the low-frequency, mid-frequency and high-frequency features are fused into multi-spectral features 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. S3. Based on the missing activity prediction results, complete the activity in the event log to be repaired to obtain the repair event log.

[0021] The specific steps are detailed below.

[0022] S1. Obtain multiple historical complete event log samples, and after activity deletion processing based on the controlled activity deletion ratio, obtain multiple missing activity log samples. Multiple historical complete event log samples and multiple missing event log samples form training sample pairs. Train the activity missing prediction model. During the training process, update the network parameters through backpropagation so that the model learns the nonlinear mapping relationship between missing patterns and complete activities, and obtain the activity missing prediction training model.

[0023] First, define the event and time logs as follows.

[0024] Events and Attributes. An event is a specific execution instance of an activity in a business process, including key information during execution. Each event consists of a set of attributes. An event is denoted as... Its attribute set is denoted as ,in, For case identification, As an event identifier, The name of the event. For timestamps, For resources.

[0025] Trace and Event Log. A trace is a sequence of events arranged chronologically within the same process instance, reflecting the complete execution path of a single process instance. An event log is a collection of multiple traces, covering the execution records of all instances of a specific business process. A trace is denoted as... ,in The length of the trace, Indicates the first One event. The event log is recorded as follows. ,in This represents the number of traces in the log. Indicates the first Traces.

[0026] Missing attributes and missing events. A missing attribute refers to an event where the value of a certain attribute is empty. If an event contains at least one missing attribute, then the event is called a missing event. For events... If it exists ,but It is a missing attribute, and For missing events, denoted as , This is the set of missing events.

[0027] Missing traces and missing logs. A missing trace is a trace that contains at least one missing event. An event log is called a missing log if it contains at least one missing trace. For traces... If it exists ,but For missing traces, denoted as , This is a set of missing traces. If the event log... satisfy ,but For missing log entries, record as , This is the set of missing logs.

[0028] Then, multiple complete historical event log samples were obtained, and after event deletion processing based on the controlled activity deletion ratio, multiple missing event log samples were obtained. The missing rate in the missing event log samples was 15%, 20%, or 30%, and the missing locations were randomly distributed to simulate different degrees of activity missing in real-world scenarios.

[0029] Next, multiple complete historical event log samples (as labels) and multiple missing event log samples were used to form training sample pairs, which were divided into training and test sets in a 7:3 ratio to train the event missing prediction model. During the training process, the network parameters were optimized and updated through backpropagation, so that the model learned the nonlinear mapping relationship between missing patterns and complete events, thus obtaining a model with the ability to predict missing events, namely the event missing prediction training model.

[0030] The missing activity prediction model consists of an input module, a multi-layer feature extraction module, a multi-spectral fusion module, and a prediction output module connected in series. These correspond to input operations, multi-layer feature extraction operations, multi-spectral fusion operations, and prediction output operations (one-dimensional convolution, global pooling, and non-linear processing - LogSoftmax activation function), respectively. The corresponding processing procedures will be described later.

[0031] During training, the loss function L The calculation formula is as follows: , For the number of activity categories, C The total number of activity categories. One-hot encoding of the real label. This loss function, representing the probability distribution of the model output, effectively guides the model to focus on classifying missing activities. It aligns well with the LogSoftmax activation function in the model's output layer, improving classification accuracy. The optimizer uses the Adam algorithm, with an initial learning rate set to... At the same time, combined Regularized weight decay suppresses the risk of overfitting; the weight decay coefficient is set to... .

[0032] Furthermore, during training, to standardize feature distribution and accelerate convergence, Batch Normalization is introduced after each convolutional operation in the activity missing prediction model. To avoid overfitting the model to the training data, a Dropout layer is added between the multi-spectral fusion module and the output layer, with a dropout rate set to 0.5, to enhance the model's generalization ability by randomly discarding some features. The hyperparameters for the training process are set as follows: batch size 32 to balance computational efficiency and gradient stability; maximum number of iterations 100; and an early stopping mechanism is introduced, terminating training if the validation set accuracy does not improve for 10 consecutive iterations to avoid invalid iterations.

[0033] During backpropagation, the gradient of the loss function with respect to the parameters of each layer is calculated using the chain rule, including the asymmetric convolutional kernel weights of some convolutions in the multi-layer feature extraction module, the attention parameters of the multi-spectral fusion module, and the weights of fully connected layers. The Adam optimizer then updates the parameters. During training, gradient weighting is applied to the masked regions at missing locations, with a weight coefficient set to 1.2, to force the model to enhance its perception of the missing neighborhood context. Simultaneously, mixed-precision computation is used to accelerate the training process and ensure numerical stability when dealing with high-dimensional sparse activity relation matrices. After each training round, 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 missing activity prediction in the testing phase.

[0034] S2. The event log to be repaired is processed by the activity missing prediction training model to obtain the missing activity prediction results.

[0035] The details of the processing steps for training the activity missing prediction model are as follows.

[0036] Step 1: 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.

[0037] The activity relationship matrix is ​​constructed based on the relationships between activities in the event log to be repaired, and is used to enhance the model's ability to model the behavioral structure of business processes. The relationships between activities include sequential relationships, concurrent relationships, mutual exclusion relationships, and missing relationships.

[0038] If the activity Always appearing at events Previously, if the two events did not overlap in time, then they were said to have an order relationship, denoted as . ;like and When two events appear together multiple times in a trajectory, and their order is not fixed or there is a possibility of concurrent execution, a concurrency relationship is defined as denoted as . ;like and If two trajectories never co-occur in any of the trajectories, then they are considered to have a strong structural separability, defined as a mutual exclusion relationship, denoted as . Furthermore, to enhance the model's ability to focus on missing regions, if a missing activity exists, its 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 relationships of the missing location.

[0039] The above 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 reflecting stable path patterns; the activity relationship matrix of the event log to be repaired is partially convolutional and two-dimensional convolutional to obtain mid-frequency features reflecting local branch structure features; the activity relationship matrix of the event log to be repaired is partially convolutional, two-dimensional convolutional and three-dimensional convolutional to capture sudden behavior patterns and obtain high-frequency features.

[0040] In business process log repair tasks, missing activities can lead to structural breaks in the activity relationship matrix. Traditional convolution is not sensitive enough to the direction of the neighborhood of the missing location, making it difficult to capture the contextual dependencies at the break. For example, after an activity is missing, the sequential relationship between its preceding and following activities needs to be inferred from multiple directions. Therefore, this embodiment designs a partial convolution processing method. Its core advantage lies in accurately perceiving the forward dependencies of the missing location in sequential relationships, the bidirectional interactions in concurrent relationships, and the exclusionary boundaries in mutually exclusive relationships through asymmetric convolution kernels and multi-directional feature extraction, thereby enhancing the ability to model irregular missing activity patterns in logs.

[0041] The partial convolution operation is as follows: the activity relation 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 region features are concatenated, and after convolutional layers, batch normalization and non-linear processing are performed to obtain partial convolution features, which are used as low-frequency features or to perform two-dimensional convolution operations.

[0042] Partial convolution is achieved using the following formula: , , , , , Y’ , Y 1 , Y 2 , Y 3 , Y 4 These are, respectively, partially convolutional features, first region features, second region features, third region features, and fourth region features. This indicates that the activity relation matrix is ​​padded with one zero on the left, three zeros on the top, and left unpadded on the right and top. This indicates that the activity relation matrix is ​​padded with three zeros to the right, one zero to the top, and left and top empty. This indicates that the activity relation matrix is ​​padded with three zeros to the left, one zero to the top, and nothing to the right or bottom. This indicates that the activity relation matrix is ​​padded with one zero to the right, three zeros to the top, and left and bottom empty. Denotes the directional convolution kernel, i=1,2,3, c ’ Number of output channels Convolutional kernels are used to extract features in the horizontal direction and the lower left region; Convolutional kernels are used to extract features in the vertical direction and the lower right region. Convolutional kernels are used to extract features in the horizontal direction and the upper right region. Convolutional kernels are used to extract features in the vertical direction and the upper left region of the image. BN() For batch normalization processing, SiLU( ) for SiLU Function processing, Concat() For splicing processing, Conv fuse ( ) To perform feature fusion and dimension adjustment on the concatenated matrix through a convolutional layer, this method integrates the feature information from the four branches and adjusts the number of high channels in the concatenated matrix to the number of channels for subsequent calculations, making the features more compact and effective.

[0043] Step 2: Perform multi-spectral fusion of low-frequency features, mid-frequency features and high-frequency features to obtain fused features that provide context-sensitive multi-scale semantic representations for missing activity prediction.

[0044] 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 then added element-wise and processed by the ReLU activation function to obtain the fused feature. By aligning the structural semantics of different trajectories, multi-spectral fusion processing can significantly improve the accuracy and robustness of missing location localization and activity prediction in logs with multiple missing elements, multiple branches, and high noise.

[0045] The calculation formulas corresponding to the multi-spectral fusion processing operation are as follows: , , , , , , These are the fusion features, mid-frequency features, first reconstruction features, and second reconstruction features, respectively. , , , , These are the query features of mid-frequency features, the key features of low-frequency features, the value features of low-frequency features, the key features of high-frequency features, and the value features of high-frequency features. For the number of channels, softmax() This is processed by the softmax function. ReLU() This is processed using the ReLU activation function.

[0046] Step 3: After the fused features are subjected to one-dimensional convolution and global pooling, non-linear processing is performed (which can be achieved through the LogSoftmax activation function) to obtain the missing activity prediction results.

[0047] S3. Based on the missing activity prediction results, complete the activity in the event log to be repaired to obtain the repair event log.

[0048] To verify the effectiveness of the method in this embodiment (hereinafter referred to as PMSF-Net) in the task of repairing missing activities in business processes, three sets of comparative experiments were designed. First, ablation experiments were conducted to verify the performance gain of multi-spectral fusion. Second, compared with traditional model-free methods, the advantages of the model in single missing activity repair scenarios were evaluated. Third, compared with mainstream deep learning methods, the robustness of the model in complex process logs and high missing rate scenarios was tested. Finally, the practicality and advancement of PMSF-Net were comprehensively demonstrated.

[0049] Experimental Data. The experiment selected six event log datasets with domain representativeness and structural diversity, covering real-world business scenarios and manually constructed scenarios. Specifically, BPIC 2012 is a loan application process log from a Dutch financial institution; BPIC2013 is an event management log from Volvo's IT department; Roadfine comes from the Italian police traffic fine processing system; Hospital is a medical billing process log from a regional hospital; Help Desk is a work order log from an enterprise's IT service desk; and Log is a manually constructed dataset. Detailed information is shown in Table 1.

[0050] Table 1. Basic Statistical Information of the Dataset

[0051] To evaluate the performance of the method, the success rate was used in the experiment. success rate As a core evaluation indicator, it measures the proportion of successfully repaired missing activities to the total number of missing activities. Let... To restore the number of successful activities, Given the total number of missing activities in the dataset, the success rate is defined as follows: , In the experiment, all models were run under the same training and testing partitions, optimizers, and hyperparameters to eliminate interference from irrelevant variables and ensure the fairness of the metric comparison.

[0052] Experimental Results. To systematically verify the performance advantages of PMSF-Net, the experiments were conducted layer by layer from three dimensions: ablation experiments to verify the necessity of the module, comparison with traditional methods, and comparison with deep learning methods.

[0053] Table 2 Repair Capability Table

[0054] The ablation experiment used six datasets as test objects, with missing rates of 15%, 20%, and 30%, and missing locations were randomly distributed. The missing rate was calculated based on the total number of activities. The performance of the complete model and the baseline model were compared, and the results are shown in Table 2.

[0055] In regular processes, PMSF-Net outperforms PROELR by 84.8%-133.8%, consistently leading baseline models such as SRBA and Transformer, and maintaining a stable advantage even when the missing data rate increases.

[0056] In complex processes, PMSF-Net shows a more significant improvement, with a performance increase of 96.7%-133.2% compared to PROELR. It not only outperforms traditional baselines but also surpasses deep learning models such as CNN, and its stability against missing data is also better.

[0057] The core reason for this difference lies in the fact that multi-spectral fusion, through mid-frequency feature anchors, dynamically fuses low-scale stable process patterns with high-scale sudden behavioral details, effectively mitigating contextual breaks under high missing rates. In contrast, variant models, relying on single-scale features, struggle to capture cross-level activity correlations, ultimately leading to decreased accuracy in complex scene restoration, thus confirming the core benefit of multi-spectral fusion for complex process restoration.

[0058] For the BPIC 2012 dataset, with missing activity events set to 100, 150, 200, and 300, PMSF-Net was compared with traditional model-free methods. The results are as follows: Figure 2 As shown (x-axis represents different numbers of missing activities, y-axis represents repair success rate). Traditional model-free methods include PROELR based on trajectory similarity, SRBA based on activity dependency graphs, and Autoencoder based on autoencoder reconstruction. Among them, PROELR's repair accuracy is only 43.2%~46.7%, while SRBA and Autoencoder reach a maximum of about 80.7%, and their performance fluctuates significantly with the number of events. For example, SRBA's accuracy drops by 1.5% with 300 events compared to 150 events. In contrast, PMSF-Net's repair accuracy remains stable at 97.6%~98.2%. The limitation of traditional methods lies in their reliance on static rules or shallow feature reconstruction. For example, SRBA uses fixed dependency edges, and Autoencoder only relies on a single-dimensional error, making it impossible to dynamically model branch loops. PMSF-Net, on the other hand, captures the forward and backward dependencies of the missing location through missing-aware convolutions, achieving accurate repair of long sequence missing data.

[0059] In the comparison of deep learning methods, we first compared PMSF-Net with CNN and Transformer on six datasets, and the results are shown in Table 2. In regular processes, PMSF-Net achieved an accuracy of 0.949 with a 30% missing value in BPIC 2012, higher than CNN and Transformer. Its advantage is even more significant in complex processes. With a 30% missing value in Help Desk, PMSF-Net achieved an accuracy of 0.868, higher than CNN's 0.806; with a 15% missing value in Log, PMSF-Net achieved an accuracy of 0.969, higher than CNN's 0.928, demonstrating its ability to handle manually constructed multi-type activity relationships.

[0060] Focusing on BPIC 2013, further comparisons were made between PMSF-Net and CNN, Transformer, and BERT, with the following results: Figure 3 As shown (x-axis represents different missing percentages, y-axis represents repair success rate). At low missing rates, PMSF-Net achieves an accuracy of 0.850, outperforming CNN's 0.835 and BERT's 0.813, and significantly surpassing Transformer's 0.736. At high missing rates, PMSF-Net achieves an accuracy of 0.804, leading CNN by 8.1%, BERT by 7.2%, and Transformer by 15.7%, with the smallest performance drop of only 4.6%, compared to CNN's 11.2% drop. Analysis reveals that CNNs are limited by the receptive field of local convolutions, making it difficult to handle concurrent activities; Transformer and BERT rely on self-attention, making them sensitive to semantic breaks caused by missing data; while PMSF-Net perceives temporal relationships through direction-sensitive convolutions and combines multi-spectral fusion with multi-scale fusion, enhancing its tolerance to complex control flows.

[0061] In summary, the necessity of multi-spectral fusion, the efficient repair capability of PMSF-Net (the method in this embodiment) for rule processes, and its performance surpassing traditional and deep learning methods in complex scenarios, all jointly verify the effectiveness and robustness of the method in this embodiment for repairing missing activities in business processes.

[0062] This embodiment also provides 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, including: 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 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. 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.

[0063] This embodiment also provides a missing activity repair device for business processes based on missing awareness, including a processor and a memory, wherein the processor implements the above-described missing activity repair method for business processes based on missing awareness when executing a computer program stored in the memory.

[0064] This embodiment also provides a computer-readable storage medium for storing a computer program, wherein the computer program, when executed by a processor, implements the above-described method for repairing missing activities in a business process based on missing awareness.

[0065] This embodiment provides a missing activity repair method for business processes based on missing activity awareness. First, the order, concurrency, and mutual exclusion relationships between activities in the event log are encoded into a two-dimensional relationship matrix. Multi-layer feature extraction processing based on convolution in different regions is then performed to obtain low-frequency features reflecting stable path patterns, mid-frequency features reflecting local branch structure features, and high-frequency features reflecting sudden behavior patterns. Next, using the mid-frequency features as semantic anchors, multi-spectral features are aligned and fused. The fused features are then flattened into a one-dimensional tensor after size adjustment and channel padding. This tensor undergoes one-dimensional convolution and global pooling followed by non-linear processing to obtain the missing activity prediction result. Finally, based on the missing activity prediction result, activities are filled into the event log to be repaired, resulting in a repaired event log. This embodiment's method achieves a repair accuracy of 80%–96% under different missing rates on six public datasets, outperforming traditional methods overall. It performs particularly well in complex process logs with high concurrency and high missing rates, providing an efficient solution for missing activity repair.

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 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-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.

5. The method for repairing missing activities in a business process based on missing information as described in claim 4, characterized in that, The partial convolution operation is as follows: the activity relation 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.

6. 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.

7. 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.

8. 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.

9. 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 1-7.

10. 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 1-7.

Images