An industrial process remaining time prediction method and device and a storage medium

Abstract

The application discloses an industrial process remaining time prediction method and device and a storage medium, and belongs to the field of data processing. The method comprises the following steps: acquiring a track prefix sequence and inputting the track prefix sequence into an activity completion model to output a complete track prefix sequence after activity completion; extracting a semantic feature vector and a time interval feature vector of each activity in the complete track prefix sequence, splicing the semantic feature vector and the time interval feature vector to generate a fusion feature vector, and arranging the fusion feature vector in an activity order to form a feature sequence; inputting the feature sequence into a remaining time prediction model to output a predicted remaining time of an industrial process instance; wherein the activity completion model is obtained based on a BERT architecture and combined with contrast learning training, and the remaining time prediction model is based on a Transformer architecture and introduces a time perception attention mechanism. The effective repair of missing industrial process activities, the deep fusion of multi-dimensional features and the accurate modeling of time dynamic characteristics are realized, and the accuracy of the predicted remaining time is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, and particularly relates to an industrial process residual time prediction method and device and a storage medium. BACKGROUND

[0002] With the continuous advancement of intelligent manufacturing and industrial informatization, enterprises have accumulated a large amount of high-dimensional and multi-source process data in the production process. How to accurately predict the residual time in the production process from these data has become a key problem in the field of predictive process monitoring. However, there are two problems in the industrial process event log: one is the activity missing problem, that is, due to equipment failure, operation error and other reasons, part of the activities in the event log are not recorded, which causes the model to lose key context information in the training process, thereby seriously affecting the accuracy of the prediction model; the other is the dynamic evolution problem of time characteristics, that is, the same activity has different execution times in different cases, and even the same trajectory prefix sequence has different residual times in different cases. If only the order of activities is considered in prediction and the dynamic evolution of time characteristics is ignored, the prediction error will increase and the prediction effect will be limited. Therefore, how to effectively complete the activities before prediction and fully perceive the dynamic evolution of time characteristics so as to improve the residual time prediction effect has become a key problem to be solved in the current industrial process monitoring field.

[0003] To address the issue of decreased prediction accuracy due to missing events in event logs, researchers have proposed various event log repair methods. For example, some researchers model event logs as token trajectories based on resources and queues, using linear programming to infer the timestamps of unrecorded events, mitigating performance analysis errors caused by incomplete recording in shared resource environments. Others have proposed repair methods based on automatic reordering and time allocation, focusing on the problem of multiple events within the same process instance having the same timestamp. While these methods can fill in missing events to some extent, they require high-quality process models and logs, making them difficult to handle noise and heterogeneous data in real-world industrial environments. Therefore, some researchers utilize statistical relationships and contextual information between activities to fill in missing events. For instance, some researchers transform activity repair into a clustering task, using an activity relationship matrix and the SOM algorithm for clustering to predict missing activities; others have proposed probability-based repair methods, repairing missing activities based on their occurrence frequency. The aforementioned methods do not rely on the quality of the process model and event log, but lack the ability to model the deep semantics of trajectory sequences. Deep learning methods, on the other hand, have significant advantages in representing complex nonlinear sequence patterns. Therefore, some researchers have successfully predicted missing activities and their timestamps by using LSTM networks to model the temporal dynamics of process trajectories. Others have proposed an image modeling method based on convolutional neural networks, representing event logs as image matrices and using spatial structure features to repair missing activities, thus eliminating dependence on the process model structure. Although deep learning methods improve the accuracy of activity completion, they still struggle to capture the deep dependencies between activities and their context.

[0004] To address the dynamic evolution of temporal features, various time-aware modeling methods have been proposed. For example, some researchers have proposed multi-channel adaptive time-aware LSTM, combining frequency measurement and cell state parameterization to process irregular time intervals in ICU data, thus improving the accuracy of patient risk prediction. Others have proposed a time-aware soft clustering algorithm based on EHR, combining time and clinical background to identify mixed subphenotypes of sepsis, providing important information for early warning. To further enhance the model's ability to perceive temporal dependencies, some studies have introduced attention mechanisms. For instance, some researchers have proposed a ship trajectory prediction method based on time-aware graph convolutional networks, combining self-attention mechanisms to improve the modeling accuracy of spatiotemporal relationships. In TSN hybrid traffic scheduling, some researchers have combined time-aware shapers with cyclic queuing and forwarding mechanisms, effectively improving bandwidth utilization and scheduling success rates. These methods alleviate the impact of dynamic changes in temporal features on downstream tasks to some extent. However, in actual industrial processes, activity rhythms change frequently and dependency patterns are complex, making it difficult for traditional modeling methods to fully capture the dynamic evolution of temporal features in industrial events.

[0005] There is currently no effective solution to the technical problems in the existing technology of industrial process event logs, which often have missing activities and are difficult to capture the dynamic evolution of time features, thus limiting the accuracy of prediction models. Summary of the Invention

[0006] The embodiments of this disclosure provide a method, apparatus, and storage medium for predicting the remaining time of an industrial process, in order to at least solve the technical problem in the prior art that industrial process event logs often have missing activities and are difficult to capture the dynamic evolution of time characteristics, which leads to limited accuracy of prediction models.

[0007] According to one aspect of the present disclosure, a method for predicting the remaining time of an industrial process is provided, comprising: obtaining a trajectory prefix sequence of an industrial process instance as the prediction object; wherein, a specific process activity performed in the industrial process instance is called an activity, and a specific execution of each activity is called an event, and the trajectory prefix sequence is an activity sequence composed of events that have occurred in the industrial process instance in chronological order of execution; inputting the trajectory prefix sequence into a pre-trained activity completion model, and outputting a complete trajectory prefix sequence after activity completion; wherein, the activity completion model is based on the BERT architecture and trained using contrastive learning, and is used to predict and complete missing activities in the trajectory prefix sequence; The semantic feature vector and time interval feature vector of each activity in the complete trajectory prefix sequence are taken, and the semantic feature vector and time interval feature vector of each activity are concatenated to generate a fused feature vector. The fused feature vector is then arranged in the order of the activities to form a feature sequence. The feature sequence is then input into a pre-trained remaining time prediction model to output the predicted remaining time of the industrial process instance. The remaining time prediction model is based on the Transformer architecture and introduces a time-aware attention mechanism. The predicted remaining time is an estimate of the remaining duration required from the timestamp of the last event in the trajectory prefix sequence until the industrial process instance is completed.

[0008] According to another aspect of the present disclosure, a storage medium is also provided, the storage medium including a stored program, wherein, when the program is executed, a processor performs any of the methods described above.

[0009] According to another aspect of the present disclosure, an industrial process remaining time prediction device is also provided, comprising: an acquisition module, configured to acquire a trajectory prefix sequence of an industrial process instance as the prediction object; wherein, a specific process activity performed in the industrial process instance is called an activity, and a specific execution of each activity is called an event, and the trajectory prefix sequence is an activity sequence composed of events that have occurred in the industrial process instance in chronological order of execution; and an activity completion module, configured to input the trajectory prefix sequence into a pre-trained activity completion model and output a complete trajectory prefix sequence after activity completion; wherein, the activity completion model is based on the BERT architecture and trained using contrastive learning, and is used to predict and complete missing activities in the trajectory prefix sequence; The feature extraction module is used to extract the semantic feature vector and time interval feature vector of each activity in the complete trajectory prefix sequence, concatenate the semantic feature vector and time interval feature vector of each activity to generate a fused feature vector, and arrange the fused feature vectors in the order of activities to form a feature sequence; and the prediction module is used to input the feature sequence into a pre-trained remaining time prediction model and output the predicted remaining time of the industrial process instance; wherein, the remaining time prediction model is based on the Transformer architecture and introduces a time-aware attention mechanism, and the predicted remaining time is an estimate of the remaining time required from the timestamp of the last event in the trajectory prefix sequence to the completion of the industrial process instance.

[0010] According to another aspect of the present disclosure, an industrial process remaining time prediction device is also provided, comprising: a processor; and a memory connected to the processor, configured to provide the processor with instructions for processing the following steps: acquiring a trajectory prefix sequence of an industrial process instance as the prediction object; wherein a specific process activity performed in the industrial process instance is called an activity, and a specific execution of each activity is called an event, and the trajectory prefix sequence is an activity sequence composed of events that have occurred in the industrial process instance in chronological order of execution; inputting the trajectory prefix sequence into a pre-trained activity completion model, and outputting a complete trajectory prefix sequence after activity completion; wherein the activity completion model is based on the BERT architecture and trained using contrastive learning, and is used for prediction. The missing activities in the trajectory prefix sequence are detected and completed; the semantic feature vector and time interval feature vector of each activity in the complete trajectory prefix sequence are extracted, and the semantic feature vector and time interval feature vector of each activity are concatenated to generate a fused feature vector, and the fused feature vector is arranged in the order of activities to form a feature sequence; and the feature sequence is input into a pre-trained remaining time prediction model to output the predicted remaining time of the industrial process instance; wherein, the remaining time prediction model is based on the Transformer architecture and introduces a time-aware attention mechanism, and the predicted remaining time is an estimate of the remaining time required from the timestamp of the last event in the trajectory prefix sequence to the completion of the industrial process instance.

[0011] This application first obtains the trajectory prefix sequence of an industrial process instance as the prediction target. This sequence, composed of events that have occurred in chronological order, provides the current state context of the process execution. Then, the trajectory prefix sequence is input into an activity completion model trained based on the BERT architecture and a contrastive learning strategy. This activity completion model learns deep semantic dependencies between industrial activities through a masked language modeling task and enhances its ability to distinguish different trajectory sequences using a contrastive learning mechanism. This allows for accurate prediction and completion of missing activities in the sequence, generating a complete trajectory prefix sequence, effectively solving the problem of incomplete context information caused by missing data. Next, semantic feature vectors and time interval feature vectors are extracted and fused from the completed sequence. The resulting fused feature vectors are then arranged in order to form a feature sequence. This feature sequence simultaneously encodes the functional semantics of the activity and the dynamic characteristics of its execution time, providing a comprehensive and rich feature representation for remaining time prediction. Finally, the feature sequence is input into a remaining time prediction model based on the Transformer architecture, outputting a high-precision remaining time prediction value. The core of this remaining time prediction model lies in the introduction of a time-aware attention mechanism. By incorporating a moderating factor derived from time interval features into the self-attention calculation, the model can dynamically perceive and quantify the strength of time dependencies between activities, thereby accurately capturing the complex time evolution patterns in industrial processes. This enables effective repair of missing activities in industrial processes, deep fusion of multi-dimensional features, and accurate modeling of dynamic time characteristics, effectively improving the accuracy of the predicted remaining time. It solves the technical problem in existing technologies where missing activities and the difficulty in capturing the dynamic evolution of time features in industrial process event logs often limit the accuracy of prediction models. Attached Figure Description

[0012] The accompanying drawings, which are included to provide a further understanding of this disclosure and form part of this application, illustrate exemplary embodiments of this disclosure and are used to explain this disclosure, but do not constitute an undue limitation of this disclosure. In the drawings:

[0013] Figure 1 This is a hardware structure block diagram of a computing device for implementing the method according to Embodiment 1 of this disclosure;

[0014] Figure 2 This is a flowchart of the industrial process remaining time prediction method according to Embodiment 1 of this application;

[0015] Figure 3 This is a schematic diagram of the overall framework for training the activity completion model and the remaining time prediction model according to Embodiment 1 of this application;

[0016] Figure 4This is a schematic diagram of the framework of the activity completion module based on BERT model and contrastive learning as described in Embodiment 1 of this application;

[0017] Figure 5 This is a schematic diagram of the framework of the time-aware Transformer-based remaining time prediction module according to Embodiment 1 of this application;

[0018] Figure 6 This is a comparison chart of the activity completion accuracy of the Production dataset under different missing rates on the contrastive model according to Embodiment 1 of this application;

[0019] Figure 7 This is a comparison chart of the active completion accuracy of the WMFD dataset under different missing rates on the contrastive model according to Embodiment 1 of this application;

[0020] Figure 8 This is a comparison chart of the active completion accuracy of the Production dataset under different missing rates on the ablation model according to Embodiment 1 of this application;

[0021] Figure 9 This is a comparison chart of the active completion accuracy of the WMFD dataset under different missing rates on the ablation model according to Embodiment 1 of this application;

[0022] Figure 10 This is a comparison experiment result diagram based on the Production dataset described in Embodiment 1 of this application;

[0023] Figure 11 This is a comparison experiment result diagram based on the WMFD dataset described in Embodiment 1 of this application;

[0024] Figure 12 This is a diagram of the ablation experiment results based on the Production dataset described in Embodiment 1 of this application;

[0025] Figure 13 This is a diagram showing the ablation experiment results of the WFMD dataset described in Embodiment 1 of this application;

[0026] Figure 14 This is a schematic diagram of the industrial process remaining time prediction device according to Embodiment 2 of this application;

[0027] Figure 15 This is a schematic diagram of the industrial process remaining time prediction device according to Embodiment 3 of this application. Detailed Implementation

[0028] To enable those skilled in the art to better understand the technical solutions of this disclosure, the technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this disclosure, and not all embodiments. Based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this disclosure.

[0029] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this disclosure described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0030] Example 1

[0031] According to this embodiment, a method embodiment for predicting the remaining time of an industrial process is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0032] The method embodiments provided in this example can be executed on a server or similar computing device. Figure 1 A hardware block diagram of a computing device for implementing a method for predicting the remaining time of an industrial process is shown. Figure 1 As shown, a computing device may include one or more processors (processors may include, but are not limited to, microprocessors such as MCUs or programmable logic devices such as FPGAs), a memory for storing data, a transmission device for communication functions, and an input / output interface. The memory, transmission device, and input / output interface are connected to the processor via a bus. In addition, it may also include a display, keyboard, and cursor control device connected to the input / output interface. Those skilled in the art will understand that... Figure 1 The structure shown is for illustrative purposes only and does not limit the structure of the aforementioned electronic device. For example, a computing device may also include... Figure 1 The more or fewer components shown, or having the sameFigure 1 The different configurations shown.

[0033] It should be noted that the aforementioned one or more processors and / or other data processing circuits are generally referred to as "data processing circuits" in this application. The data processing circuit can be embodied, in whole or in part, as software, hardware, firmware, or any other combination thereof. Furthermore, the data processing circuit can be a single, independent processing module, or it can be integrated, in whole or in part, into any other element in a computing device. As involved in the embodiments of this disclosure, the data processing circuit serves as a processor control mechanism (e.g., selection of a variable resistor termination path connected to an interface).

[0034] The memory can be used to store software programs and modules of application software, such as the program instructions / data storage device corresponding to the industrial process remaining time prediction method in the embodiments of this disclosure. The processor executes various functional applications and data processing by running the software programs and modules stored in the memory, thereby realizing the industrial process remaining time prediction method of the aforementioned application. The memory may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory may further include memory remotely located relative to the processor, and these remote memories can be connected to the computing device via a network. Examples of the aforementioned networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0035] The transmission device is used to receive or send data via a network. Specific examples of the network described above may include a wireless network provided by the computing device's communication provider. In one example, the transmission device includes a Network Interface Controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In another example, the transmission device may be a Radio Frequency (RF) module used for wireless communication with the Internet.

[0036] The display can be, for example, a touchscreen liquid crystal display (LCD), which allows users to interact with the user interface of the computing device.

[0037] It should be noted here that, in some optional embodiments, the above... Figure 1 The computing device shown may include hardware elements (including circuitry), software elements (including computer code stored on a computer-readable medium), or a combination of both hardware and software elements. It should be noted that... Figure 1 This is only one instance of a specific particular instance, and is intended to illustrate the types of components that may exist in the aforementioned computing devices.

[0038] The relevant concepts involved in this embodiment are defined as follows:

[0039] (1) Definition of events and activities: Let A be the set of process activities in an industrial process, and activities It refers to the name of the specific process activity performed in an industrial process. Each specific execution of each activity is called an event. Each event Including the specific timestamp of the event execution ,Activity Case Study Event Number .

[0040] (2) Definition of trajectory, case, and event log: trajectory This represents a complete industrial process execution path, typically consisting of multiple consecutive events, i.e. ,in For trajectory The length of the variable, where each event occurs only once. Example This refers to an instance of an industrial process, where each event corresponds to a case, and each case contains a special attribute, namely the trajectory. Event Log It is a collection of trajectories, where each event appears only once in the event log. Table 1 shows a partial excerpt from the Production event log.

[0041] (3) Definition of trajectory prefix: trajectory prefix It is a trajectory The former An event sequence (also known as an activity sequence) consisting of 10 events, i.e. ,in .

[0042] (4) Definition of remaining time: Remaining time Refers to each trajectory prefix The timestamp of the last event and its corresponding trajectory The difference in timestamps of the last event represents the trajectory. How much longer will the corresponding industrial process take to complete?

[0043] Table 1 Production Event Log Fragments

[0044]

[0045] Under the above operating environment, according to the first aspect of this embodiment, a method for predicting the remaining time of an industrial process is provided.Figure 2 A flowchart illustrating the method is shown below. (Refer to...) Figure 2 As shown, the method includes:

[0046] S202: Obtain the trajectory prefix sequence of the industrial process instance as the prediction object; wherein, the specific process activity executed in the industrial process instance is called an activity, and each specific execution of an activity is called an event, and the trajectory prefix sequence is an activity sequence composed of the events that have occurred in the industrial process instance in the order of execution time.

[0047] S204: Input the trajectory prefix sequence into a pre-trained activity completion model and output the complete trajectory prefix sequence after activity completion; wherein, the activity completion model is based on the BERT architecture and trained by contrastive learning, and is used to predict and complete the missing activities in the trajectory prefix sequence;

[0048] S206: Extract the semantic feature vector and time interval feature vector of each activity in the complete trajectory prefix sequence, concatenate the semantic feature vector and time interval feature vector of each activity to generate a fused feature vector, and arrange the fused feature vectors in the order of activities to form a feature sequence;

[0049] S208: Input the feature sequence into the pre-trained remaining time prediction model and output the predicted remaining time of the industrial process instance; wherein, the remaining time prediction model is based on the Transformer architecture and introduces a time-aware attention mechanism, and the predicted remaining time is an estimate of the remaining time required from the timestamp of the last event of the trajectory prefix sequence until the industrial process instance is completed.

[0050] Specifically, in the industrial process monitoring system, each industrial process instance (i.e., a production case, such as a wheel hub machining order) generates a series of event logs in real time. Each event records the execution timestamp, activity name, case ID, and other information of a specific process activity (such as "milling-turning compound - machine tool 5 processes" in Table 1 above). When it is necessary to predict the remaining time of an ongoing industrial process instance, the sequence of events that have occurred in the instance is first extracted from the event log database or real-time data stream to obtain the trajectory prefix sequence of the industrial process instance (corresponding to step S202). This trajectory prefix sequence is formed by arranging events strictly in chronological order of execution time. For example, if the instance has occurred k events, the trajectory prefix sequence... Each event Includes activity tags and timestamp The trajectory prefix sequence captures the executed path and historical context of the current process, providing basic data input for subsequent predictions. Furthermore, the acquisition process is implemented through real-time data interfaces or database queries, ensuring timely capture of the latest production progress to meet the high throughput and low latency requirements of industrial environments.

[0051] Due to equipment failure, communication interruptions, or human error in recording, industrial event logs often contain missing activity records. To address this issue, this embodiment uses a specially designed activity completion model to complete the acquired trajectory prefix sequence. This activity completion model is based on the BERT architecture and trained using contrastive learning, and is used to predict and complete missing activities in the trajectory prefix sequence. Specifically, this activity completion model is based on a deep neural network architecture and, through training on a large amount of historical industrial process data, learns the inherent correlation patterns and contextual dependencies between activities. When the input is a trajectory prefix sequence that may contain missing activities, the activity completion model can analyze the contextual information of known activities, infer the most likely missing activity type, and thus generate a complete trajectory prefix sequence after activity completion (corresponding to step S204). In this way, the integrity of the trajectory prefix sequence is ensured, laying the foundation for subsequent accurate prediction.

[0052] For the completed trajectory prefix sequence, it is necessary to extract multi-dimensional feature representations for each activity. First, based on the type information of each activity, it is converted into a dense semantic feature vector through an embedding layer to capture the semantic meaning and functional characteristics of the activity. Simultaneously, based on the timestamp information recorded in the activities, the time intervals between adjacent activities are calculated and encoded into time interval feature vectors, reflecting the temporal patterns and rhythmic changes of activity execution. Then, the semantic feature vectors and time interval feature vectors of each activity are concatenated to form a fused feature vector. This fusion method ensures that both the semantic and temporal characteristics of the activities are considered simultaneously. Finally, the fused feature vectors of all activities are arranged according to the actual execution order to form a complete feature sequence (corresponding to step S206). This feature sequence comprehensively characterizes the execution state and temporal pattern of the process instance, providing rich feature representations for the final time prediction.

[0053] Finally, the constructed feature sequence is input into a specially trained remaining time prediction model, which outputs the predicted remaining time for the industrial process instance (corresponding to step S208). This model, based on an advanced sequence processing architecture, effectively captures long-distance dependencies and complex time dynamics. By analyzing the semantic information and time patterns contained in the feature sequence, the model learns the execution rules and progress characteristics of the industrial process. Based on these learned patterns, the model predicts the remaining time required for the process instance to complete from the current time point. The prediction results are expressed in time units (such as minutes or hours), providing enterprises with accurate production progress estimates and supporting decision-making processes such as production scheduling, resource allocation, and delivery time management. Furthermore, by introducing a time-aware mechanism, the model can adapt to the dynamic changes in time characteristics in different cases, significantly improving prediction accuracy.

[0054] As described in the background section, industrial process event logs often suffer from two problems: first, the problem of missing activities, where some activities are not recorded due to equipment failure, operational errors, or other reasons, causing the model to lose key contextual information during training, thus severely affecting the accuracy of the prediction model; and second, the problem of dynamic evolution of time features, where the same activity has different execution times in different cases, and even the same trajectory prefix sequence has different remaining times in different cases. If only the order of activities is considered in the prediction, ignoring the dynamic evolution of time features, the prediction error will increase and the prediction effect will be limited.

[0055] In view of this, the technical solution of this application first obtains the trajectory prefix sequence of the industrial process instance as the prediction object. The sequence is composed of events that have occurred in chronological order, providing the current state context of the process execution. Then, the trajectory prefix sequence is input into an activity completion model trained based on the BERT architecture and a contrastive learning strategy. This activity completion model learns the deep semantic dependencies between industrial activities through a masked language modeling task and enhances its ability to distinguish different trajectory sequences using a contrastive learning mechanism, thereby accurately predicting and completing missing activities in the sequence and generating a complete trajectory prefix sequence, effectively solving the problem of incomplete context information caused by missing data. Next, semantic feature vectors and time interval feature vectors are extracted and fused from the completed sequence, and the resulting fused feature vectors are arranged in order to form a feature sequence. This feature sequence simultaneously encodes the functional semantics of the activity and the dynamic characteristics of its execution time, providing a comprehensive and rich feature representation for remaining time prediction. Finally, the feature sequence is input into a remaining time prediction model based on the Transformer architecture, outputting a high-precision remaining time prediction value. The core of this remaining time prediction model lies in the introduction of a time-aware attention mechanism. By incorporating a moderating factor derived from time interval features into the self-attention calculation, the model can dynamically perceive and quantify the strength of time dependencies between activities, thereby accurately capturing the complex time evolution patterns in industrial processes. This enables effective repair of missing activities in industrial processes, deep fusion of multi-dimensional features, and accurate modeling of dynamic time characteristics, effectively improving the accuracy of the predicted remaining time. It solves the technical problem in existing technologies where missing activities and the difficulty in capturing the dynamic evolution of time features in industrial process event logs often limit the accuracy of prediction models.

[0056] Optionally, the method trains the activity completion model through the following steps: acquiring an industrial event log dataset, which includes multiple complete historical process trajectories; wherein each process trajectory represents a complete industrial process execution path, consisting of multiple consecutive events; for the activity sequences in the historical process trajectories, randomly selecting positions according to a preset ratio for mask replacement to generate masked training samples; pre-training the BERT model using the masked training samples; for each historical process trajectory in the industrial event log dataset, constructing two different masked views as positive sample pairs, and sampling masked views of other historical process trajectories in the industrial event log dataset as negative sample pairs; and fine-tuning the pre-trained BERT model using the positive sample pairs and the negative sample pairs to obtain the activity completion model.

[0057] In this embodiment of the invention, combined with Figure 3 and Figure 4As shown, to address the issue of decreased prediction accuracy due to missing activities in industrial process event logs, an activity completion module based on the BERT model and contrastive learning (CL-BERT) is proposed. This activity completion module trains the activity completion model through the following steps:

[0058] For the process traces in the event log, the traditional 80%-10%-10% masking strategy in pre-trained task MLM is used. Several activity locations are randomly selected according to a fixed ratio for masking replacement. That is, for the selected activities, 80% are directly replaced with [MASK], 10% are replaced with any other activity, and 10% remain unchanged. Table 2 shows the masked case segments, where events e6 and e... 10 The activity in event e7 was replaced with [MASK], and the activity in event e7 was replaced with other activities. The BERT model was pre-trained using the masked trajectory sequence to enable it to understand the semantic information and contextual dependencies of the industrial activity sequence. The BERT model pre-trained on the industrial event logs was then applied to an activity completion task for fine-tuning, filling in missing activities in the event logs with [MASK].

[0059] Table 2: Case snippets after masking

[0060]

[0061] Specifically, in the fine-tuning phase, for each positive and negative sample pair constructed from the trajectory, they are input into the pre-trained BERT model to extract the sequence-level representation vectors output by the model. Then, the similarity between the two representation vectors in the positive sample pair is calculated, as well as its similarity with all negative sample representation vectors. Subsequently, a contrastive loss value is calculated based on these similarities using a contrastive loss function. Simultaneously, the cross-entropy loss value of the model on the masked language model task is calculated. Finally, the weighted sum of the contrastive loss value and the cross-entropy loss value is used as the total loss value, and the BERT model parameters are optimized using the backpropagation algorithm. Through multiple rounds of iterative training, the BERT model learns a more discriminative trajectory semantic representation while maintaining its ability to accurately predict missing activities, thus obtaining the final activity completion model.

[0062] Thus, through the training method of multi-task collaborative optimization using mask language modeling and contrastive learning, the model can not only accurately complete missing activities, but also deeply understand the semantic features of industrial process trajectories, significantly improving robustness and accuracy in the case of missing data.

[0063] Optionally, the operation of pre-training the BERT model using the masked training samples includes: for each activity sequence in the masked training samples, performing the following sub-steps: mapping each activity in the activity sequence to a learnable activity embedding vector; adding a learnable position embedding vector to each position in the activity sequence; adding the activity embedding vector and the position embedding vector at each position to generate the input representation of the position, and combining the input representations of all positions into an embedding matrix; inputting the embedding matrix into the encoder of the BERT model to obtain a context-aware encoding output matrix; wherein each output vector in the encoding output matrix represents the contextual semantic representation of the activity at the t-th position; for each masked position, inputting its corresponding output vector into a fully connected layer and a softmax function to predict the activity at that position; and optimizing the parameters of the BERT model with the goal of minimizing the cross-entropy loss between the prediction result at the masked position and the true activity label.

[0064] In this embodiment of the invention, the size of the event log activity set is assumed to be... Define the learnable embedding dimension as matrix Each activity Mapped to embedding vector At the same time, a position embedding matrix is ​​introduced. For each position Add location information This yields the input representation for each position. :

[0065] (1)

[0066] Thus, an embedding matrix is ​​formed. .in T This refers to the total number of activities in the event log activity set. M This represents the maximum length of the sequence, which is the maximum number of activities that a trajectory can possibly contain. d This represents the embedding dimension, that is, the size of the dimension of the vector representation at each position. Therefore, P It is M OK d A matrix of columns, where each row corresponds to a vector representing the position of a trajectory.

[0067] The purpose of introducing a position embedding matrix in this application is to overcome the limitation of self-attention mechanisms in modeling sequence order, so that each position... i Embedded vector and corresponding activities Binding enables both semantic and temporal modeling of trajectory sequences.

[0068] Embedding matrix It is fed into the pre-training process as the overall input. In the model, the context-aware encoded output is obtained:

[0069] (2)

[0070] in, Each vector in Indicates the first If the context of an activity at a given location is missing, then the activity location is also missing. The encoding is represented as .

[0071] Output vector of missing active locations Input to a fully connected projection layer, mapped to the active space dimension and used The function yields a normalized probability distribution:

[0072] (3)

[0073] in, This represents the predicted probability across all possible activities. It is the weight matrix of the output layer. It is the bias term of the output layer. v This represents the total number of all possible activities.

[0074] Prediction results Take the activity category with the highest probability:

[0075] (4)

[0076] During the training phase, the parameters of the BERT model are optimized to minimize the cross-entropy loss between the predicted results at the masked positions and the real active labels. The formula for calculating the cross-entropy loss function is:

[0077] (5)

[0078] in, For the set of all mask locations, For the first Real activity tags for each location, Predict loss using a mask.

[0079] It's important to note that activity completion is essentially a classification task, not a regression task. The model predicts the probability of each activity category at the masked position and ultimately selects the activity category with the highest probability as the completion result. For example, in Case 113 of Table 1, when the second activity position is masked, all possible activity categories include "milling-turning compound - machine tool 5th process", "milling-turning compound quality inspection process", and "laser marking - machine tool 7th process", etc. Based on the model's prediction, if "milling-turning compound - machine tool 5th process" has the highest probability, then this activity will be used as the completion value to fill in that position.

[0080] Thus, through the aforementioned pre-training process, the BERT model can gain a deep understanding of the contextual semantic dependencies of industrial activity sequences, laying a solid foundation for subsequent fine-tuning tasks.

[0081] Optionally, the operation of fine-tuning the pre-trained BERT model using the positive sample pairs and the negative sample pairs to obtain the activity completion model includes: inputting the first masked view sequence and the second masked view sequence from the positive sample pairs into the pre-trained BERT model respectively; extracting the activity prediction results corresponding to the mask positions generated by the pre-trained BERT model for each masked view sequence, as well as the sequence-level representation vector used to characterize the overall semantics of the masked view sequence; calculating the similarity between the two sequence-level representation vectors in the positive sample pairs, and calculating the similarity between any sequence-level representation vector in the positive sample pairs and the sequence-level representation vectors of all negative sample pairs; calculating the contrastive loss value using the contrastive learning loss function based on all the calculated similarity results; simultaneously, calculating the cross-entropy loss value based on the activity prediction results corresponding to the mask positions and the real activity labels; and optimizing and adjusting the parameters of the pre-trained BERT model using the weighted sum of the contrastive loss value and the cross-entropy loss value as the total loss value, finally completing the fine-tuning to obtain the activity completion model.

[0082] In this embodiment of the invention, to enhance the model's ability to model the overall semantic structure of the trajectory, a contrastive learning mechanism is introduced during the fine-tuning stage. Specifically, for each historical process trajectory in the dataset, two different masked view sequences are constructed through two independent random masking operations, denoted as... and The two are positive sample pairs. For example, for case 114 in Table 1, its first masked view sequence Possibly The second masked view sequence Possibly .

[0083] Subsequently, the two masked view sequences from the positive sample pair are input into the pre-trained BERT model, and the sequence-level representation vectors output by the model are extracted. Specifically, the output vector corresponding to the special marker [CLS] at the beginning of each masked view sequence is obtained after being encoded by the model, and is used as the overall semantic representation of that masked view sequence. Let the two representation vectors of the positive sample pair be... , Then they constitute a positive sample pair. Simultaneously, masked view sequences of other trajectories are used as negative samples. .

[0084] Based on the above positive and negative samples, define the InfoNCE contrast loss. as follows:

[0085] (6)

[0086] in, For cosine similarity, This is a temperature coefficient used to adjust the discrimination of difficult negative samples.

[0087] Simultaneously, the model predicts the activity at each masked position in the masked view sequence and calculates the cross-entropy loss between this prediction and the actual activity label. Ultimately, the total loss function in the fine-tuning phase is a weighted sum of the masked language model loss and the contrastive loss:

[0088] (7)

[0089] in, is the weighting coefficient, used to balance the contributions of the two losses.

[0090] By minimizing the total loss through backpropagation and jointly optimizing the model parameters, the model can learn to accurately complete missing activities while bringing the semantic representations of the same trajectory closer together and pushing the semantic representations of different trajectories further apart, thereby learning more discriminative trajectory feature representations.

[0091] It should be noted that in the contrastive learning mechanism, positive sample pairs must originate from two different masked views of the same original trajectory to ensure semantic consistency; while negative samples should originate from any other trajectory different from the current trajectory to provide discriminative signals. In this embodiment, negative samples are typically sampled from other trajectories within a training batch.

[0092] Thus, by jointly optimizing the mask language model loss and the contrastive learning loss, the trained activity completion model can not only accurately predict the missing specific activities, but also deeply understand the global semantics of the entire process trajectory, significantly improving the robustness and generalization ability of the model in complex industrial scenarios.

[0093] Optionally, the method extracts the semantic feature vector of each activity in the complete trajectory prefix sequence through the following steps: for each target activity in the complete trajectory prefix sequence, aggregate the semantic embedding vectors of all activities within its context; and calculate the average value of the semantic embedding vectors of all activities within the context, and use the average value as the semantic feature vector of the target activity.

[0094] In this embodiment of the invention, a continuous bag-of-words model is used to capture the contextual semantic information of activities. Specifically, a learnable dense vector (i.e., embedding vector) is assigned to each unique activity type in the dataset. For each target activity in the trajectory prefix sequence, the embedding vectors of all other activities within its context are aggregated. The context typically refers to all activities in the sequence other than the target activity itself. By calculating the average of these contextual activity embedding vectors, a semantic feature vector for the target activity is generated. Thus, a distributed representation rich in information about its execution environment is provided for the target activity, effectively characterizing the semantic role of the activity within a specific process path, and providing a crucial contextual semantic foundation for accurate prediction of remaining time.

[0095] Optionally, the method extracts the time interval feature vector of each activity in the complete trajectory prefix sequence through the following steps: for the complete trajectory prefix sequence, calculate the difference between the timestamp of each activity and the timestamp of its predecessor activity, as the time difference between adjacent activities of that activity; wherein, for the first activity in the complete trajectory prefix sequence, its time difference between adjacent activities is defined as 0; and use the Time2Vec encoder to encode the time difference between adjacent activities of each activity, and use the encoded vector as the time interval feature vector of that activity.

[0096] In this embodiment of the invention, to accurately capture the dynamic changing patterns of the execution rhythm of activities in an industrial process, a time interval feature vector is extracted for each activity. First, the timestamp difference between each activity in the sequence and its preceding activity is calculated; this difference reflects the interval duration or pace of the activity's execution. For the first activity in the sequence, since it has no preceding activity, its time difference is set to zero. Then, a Time2Vec encoder is used to encode each time difference. The Time2Vec encoder maps scalar time interval values ​​to a high-dimensional feature vector; its core design lies in using periodic functions to capture the inherent characteristics of time, such as periodicity and monotonicity. The high-dimensional vector output by the encoder serves as the time interval feature vector for that activity. Thus, this process transforms a single time interval value into a feature representation rich in temporal pattern information, enabling the model to perceive and learn the dynamic evolution patterns of time during activity execution, providing crucial information for solving the problem of "dynamic evolution of time features."

[0097] Optionally, the method trains the remaining time prediction model through the following steps: acquiring a complete industrial event log dataset with activity completion; dynamically generating all possible trajectory prefix sequence samples from each complete process trajectory in the complete industrial event log dataset using a sliding window method; constructing a corresponding remaining time label for each trajectory prefix sequence sample; wherein the label value of the remaining time label is the difference between the timestamp of the last event in the trajectory prefix sequence sample and the timestamp of the last event in the corresponding complete process trajectory; for each trajectory prefix sequence sample, extracting the semantic feature vector and time interval feature vector of each activity, concatenating them as the fusion feature vector of the activity, and arranging them in the order of activities to form a feature sequence; inputting the feature sequence into a pre-built neural network model for training; wherein the neural network model includes an encoder and linear layers based on the Transformer architecture, and a time-aware adjustment factor is introduced into the self-attention mechanism of the encoder to construct a time-aware attention mechanism; optimizing the parameters of the neural network model with the goal of minimizing the mean square error between the remaining time predicted by the model and the true remaining time label, and finally obtaining the remaining time prediction model.

[0098] Specifically, in industrial process prediction tasks, trajectory activities have complex contextual relationships, connections, and temporal dependencies, making it difficult for traditional static modeling methods to accurately capture their dynamic temporal evolution characteristics. To address this, this invention proposes a time-aware Transformer for Process Prediction (TAFormer) module, such as... Figure 5As shown. This module extracts the semantic feature vector of each activity in the complete trajectory prefix sequence through the following steps:

[0099] Given a complete event log with event completion , Representing the trajectory The first in One activity, For the trajectory length, Indicates the number of trajectories. Prefix sequences are dynamically generated using a sliding window. And construct the remaining time tag variable:

[0100] (8)

[0101] in, Indicates the first The trajectory of the th The timestamp of the event It is a trajectory prefix sequence The remaining time tag variable.

[0102] To encode the contextual semantic features of trajectory prefix sequences, an active embedding is constructed using the CBOW bag-of-words model. The active vocabulary size is [size missing]. Embedding dimension is (and the embedding dimension mentioned above) d (Same), the activity embedding matrix is Each activity The embedding vector is represented as , It is a trajectory prefix sequence The One activity.

[0103] For trajectory prefix sequence Each position The CBOW bag-of-words model aggregates its context representation as :

[0104] (9)

[0105] Define trajectory prefix sequence The time difference between adjacent activities is , For trajectory prefix sequence The Middle The timestamp of each activity, set .

[0106] For trajectory prefix sequence Time difference between each adjacent activity Time2vec encoding as follows:

[0107] (10)

[0108] in, and These are learnable parameters.

[0109] Final trajectory prefix sequence The input representation of the m-th activity for:

[0110] (11)

[0111] in, This indicates a splicing operation.

[0112] The input representations of all activities constitute the input sequence. A time-aware modulation factor is introduced into the self-attention mechanism of the Transformer encoder. :

[0113] (12)

[0114] in, It is a time feature, namely the time difference between adjacent activities (time interval feature vector) encoded by Time2vec as mentioned above. It is a multilayer perceptron based on time features. The diagonal matrix generated by the mapping , q It is the attention dimension. These are the parameters of a multilayer perceptron.

[0115] Constructing a time-aware attention mechanism:

[0116] (13)

[0117] in, , , These represent the query, key, and value matrices for the attention mechanism, respectively. This is the scaling factor.

[0118] The input sequence is fed into the TAFormer encoder (corresponding to...). Figure 5 In the Time-aware Transformer, the output sequence is obtained. Extract the representation vector of the last time step. Input it into a linear layer to predict the remaining time. :

[0119] (14)

[0120] in, and These are learnable parameters.

[0121] During the training phase, the loss function uses mean squared error (MSE) as the optimization objective:

[0122] (15)

[0123] in, It is a trajectory prefix sequence The actual value of the remaining time. It is a trajectory prefix sequence The predicted value of the remaining time, This represents the number of all track prefixes, where N represents the number of tracks in the event log. This is due to training losses.

[0124] The loss function is minimized using backpropagation and gradient descent algorithms, thus optimizing all parameters of the model.

[0125] Therefore, by introducing a time-aware attention mechanism, the core problem of capturing the dynamic evolution of time features, as proposed in the background technology, is effectively solved. This remaining time prediction model has the following significant effects:

[0126] (1) Precise time dynamic perception capability: By transforming the time interval features encoded by Time2Vec into attention adjustment factors, the model can adaptively focus on the activity dependencies under different time spans, accurately capturing the complex and ever-changing time evolution patterns in industrial processes. Compared with traditional static modeling methods, it significantly improves the ability to represent dynamic time features.

[0127] (2) Powerful contextual semantic modeling capability: The Transformer-based encoder architecture can effectively capture long-distance dependencies. Combined with the complete sequence information provided by the early activity completion module, it ensures that the model can make full use of rich contextual semantic information for prediction.

[0128] (3) Excellent prediction accuracy and generalization: Through end-to-end training, the prediction target is directly optimized with mean squared error, enabling the model to show higher prediction accuracy on both public datasets and real industrial datasets. At the same time, it shows good robustness to noisy data and heterogeneous data in various industrial processes.

[0129] It should be noted that the extraction processes for semantic feature vectors and time interval feature vectors employ the exact same computational methods and processing logic in both the model training and application phases. Based on the layout design of the technical solution in this application, this feature extraction process is explicitly defined as an independent data preprocessing step, and it is defined as one of two independent processing phases, separate from the remaining time prediction model.

[0130] To avoid repetition, this embodiment elaborates on the complete calculation steps and formulas for feature extraction as part of the training data preparation stage when describing the training process of the remaining time prediction model. However, when describing the application process after model deployment (i.e., step S206), based on the positioning of this independent step, the overall operation flow of "extracting the semantic feature vector and time interval feature vector of each activity in the complete trajectory prefix sequence" is only described in a textual summary.

[0131] Those skilled in the art will understand that the feature extraction operations actually performed in the application phase are exactly the same in detail as those described in the training data preparation process. The two share the same set of feature extraction logic and methods, thereby ensuring consistency between the training and application phases.

[0132] To verify the effectiveness of the method proposed in this embodiment, the experiment uses the publicly available Production dataset and the actual production dataset of a wheel hub machining factory in China for verification.

[0133] Dataset 1: The Production dataset consists of production process data for a portion of products from a production workshop from January to March 2012. This dataset contains 55 different process activities and 225 cases, meaning there are 225 different trajectories. Each trajectory represents a complete production process in the workshop. The dataset contains 4543 events, each event being a record of one process activity during production. Each trajectory contains a minimum of 1 event and a maximum of 175 events, with an average of 20 events per trajectory.

[0134] Dataset 2: Actual production dataset from a domestic wheel manufacturing plant (Wheel Machining Factory Dataset, WMFD). Wheel manufacturing involves multiple processes including die casting, heat treatment, rough turning, milling, painting, and laser engraving. Different types of wheel machining require different processes and production flows. This dataset contains 1538 wheel machining order samples, meaning there are 1538 cases and 1538 different trajectories. Each trajectory represents a complete wheel machining process. The dataset contains a total of 20305 data points, each representing a single machining activity within the wheel machining process.

[0135] In the activity completion module, the dataset is split into training and test sets at an 8:2 ratio, and the activities are randomly masked at ratios of 10%, 20%, and 30%, respectively. In the remaining time prediction module, the dataset is also split into training and test sets at an 8:2 ratio. A sliding window is used to dynamically construct the trajectory prefix sequence. On each trajectory, the length of the sliding window starts from 1 and gradually increases until the length equals the trajectory length.

[0136] To evaluate the completion performance of the active completion model, this embodiment uses accuracy as the evaluation metric for the active completion task:

[0137] (16)

[0138] Where x represents the number of missing activities correctly completed, and X represents the total number of missing activities. A higher accuracy rate indicates better activity completion.

[0139] To evaluate the prediction performance of the remaining time prediction model, this embodiment uses the mean absolute error (MAE) as the evaluation metric for the prediction task:

[0140] (17)

[0141] in, For trajectory prefix sequence The actual value of the remaining time. For trajectory prefix sequence The MAE (Modified Remaining Time) is the predicted value of the remaining time. The smaller the MAE value, the smaller the error between the actual and predicted values ​​of the remaining time, and the higher the accuracy and precision of the prediction.

[0142] To verify the effectiveness of the activity completion model, a comparative experiment was conducted, selecting three methods as control groups:

[0143] (1) MissForest: This method is based on iterative completion of random forests and predicts missing activities through the collaborative prediction of multiple decision trees. The model trains a random forest using observed attributes, recursively infers missing values, and minimizes the mean squared error in each iteration until convergence or the iteration limit is reached, generating complete data consistent with the original distribution.

[0144] (2) VAIM: This method is based on adversarial generative completion using variational autoencoders, and is optimized collaboratively by a generator, a discriminator, and a learner. The generator encodes the missing trajectory and performs completion using a self-attention mechanism; the discriminator distinguishes between the original and generated data, guiding the completion result to be closer to the real distribution; the learner optimizes the generator parameters through meta-learning, enhancing the effectiveness of downstream tasks in the completed data.

[0145] (3) CNN: This method combines the temporal and behavioral attributes of activities, converts event logs into spatial data, and then into image matrices. CNNs are used to train the image matrices to predict missing activities. In the encoding stage, activity sequence and concurrent behavioral features are fused to capture complex behavioral relationships between activities.

[0146] In this experiment, activities with missing percentages of 10%, 20%, and 30% were selected in the Production and Wheel Machining datasets, respectively, and replaced with [MASK] simulated missing activities to obtain datasets with missing percentages of 10%, 20%, and 30%. The experiment will use the training set to train different comparative models, perform imputation on missing activities in the test set, and compare the error between the imputed activities and the real activities to evaluate the effectiveness of the imputation method.

[0147] Table 3. Results of the activity completion experiment under different missing rates.

[0148]

[0149] From Table 3, Figure 6 and Figure 7 It can be seen that CL-BERT performs best in handling activity completion tasks with different missing rates. The VAIM model performs well at low missing rates, but its accuracy gradually decreases as the missing rate increases. In contrast, CL-BERT's completion accuracy is relatively stable across different missing rates. This is because the CL-BERT model can effectively classify activities when there are severe missing items, thus improving the completion effect. The CNN model performs slightly better than VAIM, but still worse than CL-BERT. This is because CL-BERT can effectively extract context from trajectory sequences with high missing rates, resulting in better completion performance than the CNN model. The MissForest model's accuracy is lower than the CL-BERT model at all missing rates. This is because the CL-BERT model's ability to model the context of trajectory sequences is superior to that of Random Forest, resulting in better completion performance.

[0150] To further explore the contributions of each module in the CL-BERT model, this embodiment designed and conducted an ablation experiment, gradually removing key modules from the complete model and observing their impact on model performance. Three model versions were designed for the experiment: the CL-BERT-base model, the BERT model, and the complete CL-BERT model. Table 4 lists the activity completion accuracy of different models on two different datasets. The experimental results are as follows: Figure 8 and Figure 9 As shown, the following is a brief description of each ablation model:

[0151] (1) CL-BERT-base model: This model does not use industrial event logs to pre-train the BERT model, but directly uses the general BERT-base pre-trained model.

[0152] (2) BERT model: This model removes the contrastive learning module, uses industrial event logs to pre-train the BERT model, and directly adopts a fine-tuning strategy to predict missing activities.

[0153] (3) CL-BERT model (Ours): The BERT model is pre-trained using industrial event logs. After pre-training, a fine-tuning strategy is used to predict missing activities, and a contrastive learning task is introduced during the fine-tuning process.

[0154] Table 4. Experimental results of various ablation models under different missing rates.

[0155]

[0156] From Table 4, Figure 8 and Figure 9 As can be seen, the three ablation models exhibit significant differences in performance on the Production and WMFD datasets, with the CL-BERT model showing the best results on both datasets. Among the three ablation models, CL-BERT-base, which directly uses a general pre-trained model instead of industrial event logs, performs poorly. This is because industrial event logs differ significantly from natural language vocabulary, and the general pre-trained model may not accurately understand the semantic information in the industrial event logs, resulting in poor activity completion performance. The BERT model, which removes the contrastive learning module, outperforms the CL-BERT-base model, but still falls short of the CL-BERT model. This is because the CL-BERT model enhances its discriminative ability against different trajectories through contrastive learning, allowing it to better understand the contextual information of the current trajectory during activity completion, thus improving the activity completion effect.

[0157] To evaluate the effectiveness of the AC-TAFormer remaining time prediction model, three methods were selected for comparative experiments. A brief description of each model is as follows:

[0158] (1) HM: This method is an optimization combination strategy based on a hybrid model. It integrates annotation transformation system, multiple linear regression and random forest regression model. It optimizes the model combination weights through linear programming, continuously reduces the prediction error in multiple iterations until the optimal combination is reached, thereby constructing a remaining time prediction model.

[0159] (2) GGNN: This method uses graph neural network technology to construct the dependencies between activities into a graph structure, and combines one-hot encoding and time feature extraction to convert the event sequence into a numerical feature vector. The gated graph neural network is used to process the graph structure information, and after multiple iterations, the remaining time prediction result is output.

[0160] (3) Att-Bi-QRNN: This method constructs a remaining time prediction model based on a bidirectional quasi-recurrent neural network and an attention mechanism. Events are converted into vector representations using Word2vec, and the correlation between events is captured using a bidirectional quasi-recurrent neural network. An attention mechanism is introduced to highlight key event information, and a training iteration strategy with trajectory prefixes of different lengths is adopted to solve the problem of differences in the number of trajectory prefixes and improve prediction accuracy.

[0161] The specific experimental results of each model in the comparative experiment are shown in Table 5. Figure 10 and Figure 11 As shown, using mean absolute error (MAE) as the evaluation metric, Table 5 presents the mean absolute error values ​​for the remaining time predictions of each comparative model on the Production and WFMD datasets. Figure 10 and Figure 11 The results show the remaining time prediction performance of various comparative models on a randomly selected portion of the test set.

[0162] Table 5 Comparison of prediction results for each model

[0163]

[0164] Table 5 shows that the AC-TAFormer model achieves MAEs of 12.86 and 5.29 on the Production and WFMD datasets, respectively, both lower than other comparative models. Further analysis is needed. Figure 10 and Figure 11 The comparative experimental results show that the AC-TAFormer model outperforms the HM and Att-Bi-QRNN models in prediction. This is because the AC-TAFormer model completes missing activities and effectively models the trajectory prefix sequence, thus improving prediction performance. The GGNN model outperforms the HM and Att-Bi-QRNN models, but still falls short of the AC-TAFormer model. This is because the AC-TAFormer model has sufficient temporal awareness of the activity sequence; the same activity may have different temporal development trends in different trajectory sequences.

[0165] To further explore the contributions of each module in the AC-TAFormer model, this embodiment designed and conducted an ablation experiment, gradually removing key modules from the complete model and observing their impact on model performance. Three model versions were designed for the experiment: the TAFormer model, the AC-Transformer model, and the complete AC-TAFormer model. Table 6 lists the prediction results (MAE) of different models on two different production datasets. Figure 12 and Figure 13 The presentation shows the prediction results of randomly selected test set samples under different models. Below is a brief description of each ablation model:

[0166] (1) TAFormer model: This model removes the activity completion module and uses the direct deletion method for data preprocessing for missing activities instead of performing activity completion.

[0167] (2) AC-Transformer model: This model removes the time-aware mechanism and only uses the original Transformer encoder. It only encodes the trajectory prefix sequence into the encoder and does not encode the time difference between adjacent activities.

[0168] (3) AC-TAFormer model (Ours): This model completes the missing activities and fully realizes the time perception of the trajectory sequence during the training of the prediction model.

[0169] Table 6 Comparison of prediction results of various models in ablation experiments.

[0170]

[0171] Table 6 Figure 12 and Figure 13 Experimental results show significant differences in performance among different ablation models on the datasets. The AC-TAFormer model achieved the best results on the Production and WMFD datasets, respectively. The TAFormer model's prediction performance was significantly worse than the AC-TAFormer model, indicating that missing activities significantly impact the model's prediction performance, and the prediction model cannot effectively model trajectory sequences with many missing activities. While the AC-Transformer model's prediction performance was better than the TAFormer model, it was still worse than the AC-Transformer model. This suggests that completing missing activities can effectively improve prediction performance, but the AC-Transformer model lacks temporal awareness of the trajectory sequences and has not fully learned the temporal characteristics of activities in different contexts, thus limiting its prediction performance.

[0172] In summary, to address the challenges of missing activities and the difficulty in capturing the dynamic evolution of time features in industrial process event logs, this application proposes a method for predicting remaining time in industrial processes based on activity completion and time awareness. First, a BERT model is pre-trained using industrial event logs, and then a fine-tuning strategy is employed to predict missing activities, incorporating a contrastive learning strategy during the fine-tuning process. Second, a time-aware attention mechanism is introduced on top of the Transformer encoder, inputting the activity encoding and the time difference encoding between adjacent activities into the encoder to predict the remaining time. Finally, experiments are conducted on the publicly available Production dataset and a real-world production dataset from a domestic wheel manufacturing plant. The experiments demonstrate that the proposed method effectively improves the accuracy of remaining time prediction on industrial process datasets, achieving improvements of 6.5 percentage points and 9.8 percentage points on the two datasets, respectively.

[0173] In addition, refer to Figure 1 As shown, according to a second aspect of this embodiment, a storage medium is provided. The storage medium includes a stored program, wherein, when the program is executed, a processor performs any of the methods described above.

[0174] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the present invention is not limited to the described order of actions, because according to the present invention, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to the present invention.

[0175] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods according to the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0176] Example 2

[0177] Figure 14 An industrial process remaining time prediction device according to this embodiment is shown, which corresponds to the method described according to Embodiment 1. Reference Figure 14As shown, the industrial process remaining time prediction device includes: an acquisition module 1410, used to acquire the trajectory prefix sequence of an industrial process instance as the prediction object; wherein, the specific process activity executed in the industrial process instance is called an activity, and a specific execution of each activity is called an event, and the trajectory prefix sequence is an activity sequence composed of the events that have occurred in the industrial process instance in the order of execution time; an activity completion module 1420, used to input the trajectory prefix sequence into a pre-trained activity completion model and output the complete trajectory prefix sequence after activity completion; wherein, the activity completion model is based on the BERT architecture and combined with contrastive learning training, and is used to predict and complete the missing activities in the trajectory prefix sequence; and a feature extraction module 143. 0, used to extract the semantic feature vector and time interval feature vector of each activity in the complete trajectory prefix sequence, concatenate the semantic feature vector of each activity with the time interval feature vector to generate a fused feature vector, and arrange the fused feature vector in the order of activities to form a feature sequence; and prediction module 1440, used to input the feature sequence into a pre-trained remaining time prediction model and output the predicted remaining time of the industrial process instance; wherein, the remaining time prediction model is based on the Transformer architecture and introduces a time-aware attention mechanism, and the predicted remaining time is an estimate of the remaining time required from the timestamp of the last event in the trajectory prefix sequence to the completion of the industrial process instance.

[0178] According to this embodiment, the trajectory prefix sequence of the industrial process instance to be predicted is first obtained. This sequence consists of events that have occurred in chronological order, providing the current state context of the process execution. Then, the trajectory prefix sequence is input into an activity completion model trained based on the BERT architecture and a contrastive learning strategy. This activity completion model learns deep semantic dependencies between industrial activities through a masked language modeling task and enhances its ability to distinguish different trajectory sequences using a contrastive learning mechanism, thereby accurately predicting and completing missing activities in the sequence and generating a complete trajectory prefix sequence, effectively solving the problem of incomplete context information caused by missing data. Next, semantic feature vectors and time interval feature vectors are extracted and fused from the completed sequence, and the resulting fused feature vectors are arranged in order to form a feature sequence. This feature sequence simultaneously encodes the functional semantics of the activity and the dynamic characteristics of its execution time, providing a comprehensive and rich feature representation for remaining time prediction. Finally, the feature sequence is input into a remaining time prediction model based on the Transformer architecture, outputting a high-precision remaining time prediction value. The core of this remaining time prediction model lies in the introduction of a time-aware attention mechanism. By incorporating a moderating factor derived from time interval features into the self-attention calculation, the model can dynamically perceive and quantify the strength of time dependencies between activities, thereby accurately capturing the complex time evolution patterns in industrial processes. This enables effective repair of missing activities in industrial processes, deep fusion of multi-dimensional features, and accurate modeling of dynamic time characteristics, effectively improving the accuracy of the predicted remaining time. It solves the technical problem in existing technologies where missing activities and the difficulty in capturing the dynamic evolution of time features in industrial process event logs often limit the accuracy of prediction models.

[0179] Example 3

[0180] Figure 15 An industrial process remaining time prediction device according to this embodiment is shown, which corresponds to the method described according to Embodiment 1. Reference Figure 15As shown, the industrial process remaining time prediction device includes: a processor 1510; and a memory 1520 connected to the processor 1510, used to provide the processor 1510 with instructions to process the following steps: obtaining the trajectory prefix sequence of the industrial process instance as the prediction object; wherein, the specific process activity executed in the industrial process instance is called an activity, and a specific execution of each activity is called an event, and the trajectory prefix sequence is an activity sequence composed of the events that have occurred in the industrial process instance in the order of execution time; inputting the trajectory prefix sequence into a pre-trained activity completion model, and outputting the complete trajectory prefix sequence after activity completion; wherein, the activity completion model is based on the BERT architecture and trained by contrastive learning, and is used for prediction. The missing activities in the trajectory prefix sequence are completed; the semantic feature vector and time interval feature vector of each activity in the complete trajectory prefix sequence are extracted, and the semantic feature vector and time interval feature vector of each activity are concatenated to generate a fused feature vector. The fused feature vector is then arranged in the order of the activities to form a feature sequence. The feature sequence is then input into a pre-trained remaining time prediction model to output the predicted remaining time of the industrial process instance. The remaining time prediction model is based on the Transformer architecture and introduces a time-aware attention mechanism. The predicted remaining time is an estimate of the remaining time required from the timestamp of the last event in the trajectory prefix sequence until the industrial process instance is completed.

[0181] According to this embodiment, the trajectory prefix sequence of the industrial process instance to be predicted is first obtained. This sequence consists of events that have occurred in chronological order, providing the current state context of the process execution. Then, the trajectory prefix sequence is input into an activity completion model trained based on the BERT architecture and a contrastive learning strategy. This activity completion model learns deep semantic dependencies between industrial activities through a masked language modeling task and enhances its ability to distinguish different trajectory sequences using a contrastive learning mechanism, thereby accurately predicting and completing missing activities in the sequence and generating a complete trajectory prefix sequence, effectively solving the problem of incomplete context information caused by missing data. Next, semantic feature vectors and time interval feature vectors are extracted and fused from the completed sequence, and the resulting fused feature vectors are arranged in order to form a feature sequence. This feature sequence simultaneously encodes the functional semantics of the activity and the dynamic characteristics of its execution time, providing a comprehensive and rich feature representation for remaining time prediction. Finally, the feature sequence is input into a remaining time prediction model based on the Transformer architecture, outputting a high-precision remaining time prediction value. The core of this remaining time prediction model lies in the introduction of a time-aware attention mechanism. By incorporating a moderating factor derived from time interval features into the self-attention calculation, the model can dynamically perceive and quantify the strength of time dependencies between activities, thereby accurately capturing the complex time evolution patterns in industrial processes. This enables effective repair of missing activities in industrial processes, deep fusion of multi-dimensional features, and accurate modeling of dynamic time characteristics, effectively improving the accuracy of the predicted remaining time. It solves the technical problem in existing technologies where missing activities and the difficulty in capturing the dynamic evolution of time features in industrial process event logs often limit the accuracy of prediction models.

[0182] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0183] In the above embodiments of the present invention, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0184] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0185] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0186] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0187] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0188] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for predicting the remaining time of an industrial process, characterized in that, include: Obtain the trajectory prefix sequence of an industrial process instance that is the prediction object; wherein, the specific process activity executed in the industrial process instance is called an activity, and each specific execution of an activity is called an event, and the trajectory prefix sequence is an activity sequence composed of the events that have occurred in the industrial process instance in the order of execution time. The trajectory prefix sequence is input into a pre-trained activity completion model, which outputs a complete trajectory prefix sequence with activity completion. The activity completion model is trained based on the BERT architecture and combined with contrastive learning, and is used to predict and complete the missing activities in the trajectory prefix sequence. Extract the semantic feature vector and time interval feature vector of each activity in the complete trajectory prefix sequence, concatenate the semantic feature vector of each activity with the time interval feature vector to generate a fused feature vector, and arrange the fused feature vectors in the order of activities to form a feature sequence; The feature sequence is input into a pre-trained remaining time prediction model, which outputs the predicted remaining time of the industrial process instance. The remaining time prediction model is based on the Transformer architecture and introduces a time-aware attention mechanism. The predicted remaining time is an estimate of the remaining time required from the timestamp of the last event of the trajectory prefix sequence until the industrial process instance is completed. The activity completion model is trained through the following steps: Obtain an industrial event log dataset, which includes multiple complete historical process trajectories; each process trajectory represents a complete industrial process execution path, consisting of multiple consecutive events; For the activity sequence in the historical process trajectory, the mask is replaced at a randomly selected position according to a preset ratio to generate masked training samples; The BERT model is pre-trained using the masked training samples. For each historical process trajectory in the industrial event log dataset, two different masked views are constructed as positive sample pairs, and masked views of other historical process trajectories in the industrial event log dataset are sampled as negative sample pairs. The pre-trained BERT model is fine-tuned using the positive and negative sample pairs to obtain the active completion model.

2. The method according to claim 1, characterized in that, The operation of pre-training the BERT model using the masked training samples includes: For each activity sequence in the masked training samples, perform the following sub-steps: Map each activity in the activity sequence to a learnable activity embedding vector; add a learnable position embedding vector to each position in the activity sequence; add the activity embedding vector and the position embedding vector to each position to generate the input representation of the position, and combine the input representations of all positions into an embedding matrix; The embedding matrix is ​​input into the encoder of the BERT model to obtain a context-aware encoding output matrix; where each output vector in the encoding output matrix represents the contextual semantic representation of the activity at the t-th position; For each masked location, its corresponding output vector is input into a fully connected layer and a softmax function to predict the activity at that location. The parameters of the BERT model are optimized with the goal of minimizing the cross-entropy loss between the predicted results at the masked positions and the real active labels.

3. The method according to claim 1, characterized in that, The operation of fine-tuning the pre-trained BERT model using the positive and negative sample pairs to obtain the active completion model includes: The first and second mask view sequences from the positive sample pairs are respectively input into the pre-trained BERT model; Extract the activity prediction results corresponding to the mask position generated by the pre-trained BERT model for each mask view sequence, and use them as a sequence-level representation vector to characterize the overall semantics of the mask view sequence; Calculate the similarity between the two sequence-level representation vectors in the positive sample pair, and calculate the similarity between any sequence-level representation vector in the positive sample pair and the sequence-level representation vectors of all negative sample pairs; calculate the contrastive loss value using the contrastive learning loss function based on all the calculated similarity results; simultaneously, calculate the cross-entropy loss value based on the activity prediction result corresponding to the mask position and the real activity label; The parameters of the pre-trained BERT model are optimized and adjusted using the weighted sum of the contrast loss value and the cross-entropy loss value as the total loss value, and finally the fine-tuning is completed to obtain the active completion model.

4. The method according to claim 1, characterized in that, The semantic feature vector of each activity in the complete trajectory prefix sequence is extracted through the following steps: For each target activity in the complete trajectory prefix sequence, aggregate the semantic embedding vectors of all activities within its context. Calculate the average of the semantic embedding vectors of all activities within the context, and use the average as the semantic feature vector of the target activity.

5. The method according to claim 1, characterized in that, The time interval feature vector of each activity in the complete trajectory prefix sequence is extracted through the following steps: For the complete trajectory prefix sequence, the difference between the timestamp of each activity and the timestamp of its predecessor is calculated as the adjacent activity time difference of that activity; wherein, for the first activity in the complete trajectory prefix sequence, its adjacent activity time difference is defined as 0; The Time2Vec encoder is used to encode the time difference between adjacent activities for each activity, and the encoded vector is used as the time interval feature vector for that activity.

6. The method according to claim 1, characterized in that, The remaining time prediction model is trained using the following steps: Obtain the complete industrial event log dataset after activity completion; Using the sliding window method, all possible trajectory prefix sequence samples are dynamically generated from each complete process trajectory in the complete industrial event log dataset; Construct a corresponding remaining time label for each trajectory prefix sequence sample; wherein, the label value of the remaining time label is the difference between the timestamp of the last event in the trajectory prefix sequence sample and the timestamp of the last event in the corresponding complete process trajectory; For each trajectory prefix sequence sample, extract the semantic feature vector and time interval feature vector of each activity, concatenate the two to form the fusion feature vector of the activity, and arrange them in the order of the activities to form a feature sequence; The feature sequence is input into a pre-built neural network model for training; wherein the neural network model includes an encoder and a linear layer based on the Transformer architecture, and a time-aware adjustment factor is introduced into the self-attention mechanism of the encoder to construct a time-aware attention mechanism. With the goal of minimizing the mean squared error between the model's predicted remaining time and the actual remaining time label, the parameters of the neural network model are optimized to obtain the remaining time prediction model.

7. A storage medium, characterized in that, The storage medium includes a stored program, wherein, when the program is executed, a processor performs the method according to any one of claims 1 to 6.

8. An industrial process remaining time prediction device, characterized in that, include: The acquisition module is used to acquire the trajectory prefix sequence of an industrial process instance that is the prediction object; wherein, the specific process activity executed in the industrial process instance is called an activity, and a specific execution of each activity is called an event, and the trajectory prefix sequence is an activity sequence composed of the events that have occurred in the industrial process instance in the order of execution time. The activity completion module is used to input the trajectory prefix sequence into a pre-trained activity completion model and output the complete trajectory prefix sequence after activity completion; wherein, the activity completion model is based on the BERT architecture and trained by contrastive learning, and is used to predict and complete the missing activities in the trajectory prefix sequence; The feature extraction module is used to extract the semantic feature vector and time interval feature vector of each activity in the complete trajectory prefix sequence, concatenate the semantic feature vector of each activity with the time interval feature vector to generate a fused feature vector, and arrange the fused feature vectors in the order of activities to form a feature sequence. The prediction module is used to input the feature sequence into a pre-trained remaining time prediction model and output the predicted remaining time of the industrial process instance; wherein, the remaining time prediction model is based on the Transformer architecture and introduces a time-aware attention mechanism, and the predicted remaining time is an estimate of the remaining time required from the timestamp of the last event of the trajectory prefix sequence until the industrial process instance is completed. The activity completion model is trained through the following steps: Obtain an industrial event log dataset, which includes multiple complete historical process trajectories; each process trajectory represents a complete industrial process execution path, consisting of multiple consecutive events; For the activity sequence in the historical process trajectory, the mask is replaced at a randomly selected position according to a preset ratio to generate masked training samples; The BERT model is pre-trained using the masked training samples. For each historical process trajectory in the industrial event log dataset, two different masked views are constructed as positive sample pairs, and masked views of other historical process trajectories in the industrial event log dataset are sampled as negative sample pairs. The pre-trained BERT model is fine-tuned using the positive and negative sample pairs to obtain the active completion model.

9. An industrial process remaining time prediction device, characterized in that, include: processor; A memory, connected to the processor, for providing the processor with instructions to perform the following processing steps: Obtain the trajectory prefix sequence of an industrial process instance that is the prediction object; wherein, the specific process activity executed in the industrial process instance is called an activity, and each specific execution of an activity is called an event, and the trajectory prefix sequence is an activity sequence composed of the events that have occurred in the industrial process instance in the order of execution time. The trajectory prefix sequence is input into a pre-trained activity completion model, which outputs a complete trajectory prefix sequence with activity completion. The activity completion model is trained based on the BERT architecture and combined with contrastive learning, and is used to predict and complete the missing activities in the trajectory prefix sequence. Extract the semantic feature vector and time interval feature vector of each activity in the complete trajectory prefix sequence, concatenate the semantic feature vector of each activity with the time interval feature vector to generate a fused feature vector, and arrange the fused feature vectors in the order of activities to form a feature sequence; The feature sequence is input into a pre-trained remaining time prediction model, which outputs the predicted remaining time of the industrial process instance. The remaining time prediction model is based on the Transformer architecture and introduces a time-aware attention mechanism. The predicted remaining time is an estimate of the remaining time required from the timestamp of the last event of the trajectory prefix sequence until the industrial process instance is completed. The activity completion model is trained through the following steps: Obtain an industrial event log dataset, which includes multiple complete historical process trajectories; each process trajectory represents a complete industrial process execution path, consisting of multiple consecutive events; For the activity sequence in the historical process trajectory, the mask is replaced at a randomly selected position according to a preset ratio to generate masked training samples; The BERT model is pre-trained using the masked training samples. For each historical process trajectory in the industrial event log dataset, two different masked views are constructed as positive sample pairs, and masked views of other historical process trajectories in the industrial event log dataset are sampled as negative sample pairs. The pre-trained BERT model is fine-tuned using the positive and negative sample pairs to obtain the active completion model.

Images