A method and system for disturbance prediction of a digital twin factory

Abstract

This invention discloses a disturbance prediction method and system for digital twin workshops. The method includes: acquiring physical workshop data and establishing a corresponding digital twin model of the workshop; extracting multimodal features from the physical workshop data and the digital twin model of the workshop to obtain physical workshop data features and workshop digital twin model features; constructing a disturbance feature library using the physical workshop data features, the workshop digital twin model features, and a historical feature database; setting a disturbance threshold based on standard disturbance feature cases in the disturbance feature library; inputting the physical workshop data features and the workshop digital twin model features into a pre-constructed disturbance prediction model and outputting the disturbance occurrence probability; comparing the disturbance occurrence probability with the disturbance threshold to obtain the disturbance prediction result.

Description

Technical Field

[0001] This invention relates to the field of industrial intelligent simulation technology, and in particular to a disturbance prediction method and system for a digital twin workshop. Background Technology

[0002] The production process in a workshop is subject to many uncertainties, which can cause various disturbances. These disturbances can significantly impact normal production and processing, leading to reduced production efficiency, delivery delays, and supply chain disruptions. The application of advanced manufacturing technologies helps workshops achieve real-time monitoring of the production process and predict the timing of disturbances. Current research on disturbances mainly falls into two categories: prediction before disturbances occur and response and rescheduling after disturbances occur. Responding and rescheduling after disturbances occur can cause production plan delays and reduced production efficiency. Therefore, achieving disturbance prediction is crucial to ensuring the efficient operation of workshop production.

[0003] A digital twin workshop is a virtual workshop created using digital twin technology. It achieves real-time synchronization and bidirectional mapping between the physical and virtual workshops through sensors, the Internet of Things (IoT), and other technologies. A digital twin workshop can reflect the real-time status of the actual workshop, including equipment operation, production progress, and environmental parameters. Therefore, the data generated by a digital twin workshop is mostly real-time data. Current technologies do not disclose how to fully utilize this data to uncover hidden information and achieve real-time prediction of potential disturbances. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a disturbance prediction method and system for digital twin workshops, which can more comprehensively reflect the workshop status and thus improve the accuracy of disturbance prediction.

[0005] To achieve the above objectives, the present invention is implemented using the following technical solution:

[0006] On one hand, the present invention provides a disturbance prediction method for a digital twin factory, comprising:

[0007] Acquire physical workshop data and establish corresponding digital twin models of the workshop;

[0008] Extract the multimodal features of the physical workshop data and the workshop digital twin model to obtain the physical workshop data features and the workshop digital twin model features;

[0009] Using the physical workshop data features, the workshop digital twin model features, and the historical feature database, a disturbance feature library is constructed; based on the standard disturbance feature cases in the disturbance feature library, a disturbance threshold is set;

[0010] The physical workshop data features and the workshop digital twin model features are input into a pre-built disturbance prediction model, and the disturbance occurrence probability is output.

[0011] By comparing the probability of the disturbance occurrence with the disturbance threshold, the disturbance prediction result is obtained;

[0012] The construction of the disturbance prediction model includes:

[0013] Obtain the original CNN-GRU-Attention model;

[0014] A memory network layer parallel to the CNN layer is added between the input layer and the GRU layer of the original CNN-GRU-Attention model to obtain the perturbation prediction model.

[0015] Optional, also includes:

[0016] Based on the disturbance prediction results, update the standard disturbance features of the digital twin workshop model and physical workshop data, as well as the physical workshop data and the disturbance feature library.

[0017] Optionally, before extracting features from the physical workshop data, the method further includes:

[0018] The physical workshop data undergoes preprocessing operations including data redundancy removal, data smoothing, and data noise reduction.

[0019] Optionally, the extraction of multimodal features from the physical workshop data includes:

[0020] For one-dimensional time-series data of physical workshop data, time-domain features, frequency-domain features, and time-frequency-domain features are extracted respectively.

[0021] For multi-source heterogeneous data of physical workshop data, it is divided into multiple groups of one-dimensional data. The features of each group of one-dimensional data are extracted using a single channel. The extracted features of each group of one-dimensional data are combined into a feature matrix. The feature matrix is ​​then processed into grayscale to obtain a feature grayscale image. The feature grayscale image is used as the multimodal feature of the multivariate heterogeneous data.

[0022] Optionally, extracting the multimodal features of the workshop digital twin model includes extracting the time features, process features, and equipment features of the workshop digital twin model.

[0023] Optionally, the disturbance feature library is a three-layer analytical disturbance feature library consisting of "disturbance location - disturbance type - disturbance feature".

[0024] Optionally, the processing steps of the disturbance prediction model include:

[0025] The physical workshop data features and the workshop digital twin model features are input into a CNN layer for secondary feature extraction to obtain low-dimensional compressed features;

[0026] The physical workshop data features are input into the memory network layer, and historical perturbation data related to the physical workshop data features are retrieved to obtain historical perturbation features;

[0027] The low-dimensional compressed features and the historical perturbation features are sequentially input into the GRU layer and the Attention layer, and feature sequence processing, feature weight assignment and feature fusion are performed sequentially to obtain the fused features.

[0028] The fused features are sequentially input into the fully connected layer and the classifier layer for concatenation and classification to obtain the probability of perturbation occurrence.

[0029] Optionally, the training steps of the perturbation prediction model include:

[0030] Generative adversarial networks are used to generate perturbation datasets that are similar to real perturbation distributions.

[0031] The perturbation prediction model is trained by mixing real and spurious perturbation data as a new training dataset.

[0032] Secondly, the present invention provides a disturbance prediction system for a digital twin workshop, comprising:

[0033] The data acquisition module acquires physical workshop data and establishes a corresponding digital twin model of the workshop.

[0034] The feature extraction module extracts multimodal features from the physical workshop data and the workshop digital twin model to obtain physical workshop data features and workshop digital twin model features;

[0035] The threshold setting module constructs a disturbance feature library using the physical workshop data features, the workshop digital twin model features, and the historical feature database; and sets the disturbance threshold based on the standard disturbance feature cases in the disturbance feature library.

[0036] The physical workshop data features and the workshop digital twin model features are input into a pre-built disturbance prediction model, and the disturbance occurrence probability is output.

[0037] The construction of the disturbance prediction model includes:

[0038] Obtain the original CNN-GRU-Attention model;

[0039] Between the input layer and the GRU layer of the original CNN-GRU-Attention model, a memory network layer parallel to the CNN layer is added to obtain the perturbation prediction model.

[0040] The disturbance occurrence probability is compared with the disturbance threshold to obtain the disturbance prediction result.

[0041] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the disturbance prediction method for digital twin workshops described in the first aspect.

[0042] Beneficial effects

[0043] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:

[0044] This invention utilizes the multimodal features of physical workshop data and a workshop digital twin model for perturbation prediction, which can more comprehensively reflect the workshop state and thus improve the accuracy of perturbation prediction. It introduces a memory network to store and retrieve historical perturbation data and provides historical information related to the current workshop state, thereby helping the model better understand the changes in the workshop state and improving the model's prediction accuracy. It uses a generative adversarial network to generate spurious perturbation data with a distribution similar to the real perturbation data, expanding the training dataset, thereby solving the data imbalance problem and improving the model's generalization ability. Attached Figure Description

[0045] Figure 1 The diagram shown is a flowchart of one embodiment of the disturbance prediction method for the digital twin workshop of the present invention.

[0046] Figure 2 The diagram shown is a flowchart of one embodiment of the multimodal feature extraction method for physical workshop data according to the present invention.

[0047] Figure 3 The diagram shown is a flowchart of one embodiment of the multimodal feature extraction method for the workshop digital twin model of the present invention.

[0048] Figure 4 The diagram shown is a schematic representation of the attention mechanism of the present invention in one embodiment.

[0049] Figure 5 The diagram shown is a structural schematic of the disturbance prediction model of the present invention in one embodiment;

[0050] Figure 6 The diagram shown is a schematic representation of the hardware device of the milling cutter wear disturbance prediction experimental platform of the present invention in one embodiment. Detailed Implementation

[0051] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0052] The term "and / or" simply describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. Additionally, the character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0053] Example 1

[0054] like Figure 1 As shown, this invention introduces a disturbance prediction method for a digital twin factory, which specifically includes the following steps:

[0055] Acquire physical workshop data and establish corresponding digital twin models of the workshop.

[0056] Physical workshop data includes data on equipment, personnel, materials, and production lines. This is primarily achieved through data acquisition methods, uploading equipment operating status data and workshop production parameter data to the next layer for processing. Data processing mainly includes three parts: data storage, data preprocessing, and feature extraction. Specifically, this includes:

[0057] First, after receiving multi-source heterogeneous data from the physical workshop layer, each sensing device stores it in a shared database to provide data support for building a digital twin model of the workshop and predicting disturbances.

[0058] Secondly, the physical workshop data undergoes preprocessing operations such as data redundancy removal, data smoothing, and data noise reduction.

[0059] Finally, the preprocessed physical workshop data and the workshop digital twin model are extracted to obtain the physical workshop data features and the workshop digital twin model features. On the one hand, multimodal features are extracted from the workshop operation status data and workshop production parameter data to provide usable data for disturbance prediction. On the other hand, multimodal features are extracted from the workshop digital twin model to complete the extraction of the state features of the digital twin workshop disturbance.

[0060] in,

[0061] like Figure 2 As shown, the physical workshop contains various types of data, and different types of data require different feature extraction methods. The extraction of multimodal features from the physical workshop data includes:

[0062] For one-dimensional time-series data of physical workshop data, time-domain features, frequency-domain features, and time-frequency-domain features are extracted respectively. The obtained feature data are then input into the convolutional neural network (CNN) in the perturbation prediction model for further feature extraction.

[0063] For two-dimensional image data of physical workshop data, since the CNN in the perturbation prediction model has the function of image feature extraction, the image data can be directly input into the CNN in the perturbation prediction model for feature extraction.

[0064] For multi-source heterogeneous data of physical workshop data, it is divided into multiple groups of one-dimensional data. The features of each group of one-dimensional data are extracted using a single channel. The extracted features of each group of one-dimensional data are combined to form a feature matrix. Since the feature matrix of the constructed multi-source heterogeneous data is large, the feature matrix is ​​processed into grayscale to obtain a feature grayscale image. The obtained feature grayscale image is used as the multimodal feature of multivariate heterogeneous data and input into the CNN in the perturbation prediction model for feature extraction.

[0065] like Figure 3 As shown, extracting the multimodal features of the workshop digital twin model includes extracting the time features, process features, and equipment features of the workshop digital twin model. During workshop production, changes in processing time and machine malfunctions can affect the workpiece processing time. Therefore, extracting the processing time of each stage of the workpiece is used to characterize the workshop production status. When there is a deviation between the actual processing process and the theoretical processing process, it may lead to inconsistencies and disturbances. In a digital twin workshop, workstations may issue error commands due to workpieces not arriving or arriving incorrectly. Therefore, by extracting the theoretical and actual information of the process, the difference between the two can be used to reflect disturbances caused by orders in the workshop. During equipment processing, the condition of components changes with the processing time. By extracting relevant feature data, the equipment condition can be predicted and monitored. For example, by collecting data such as the back face wear bandwidth and back face stress distribution of the milling cutter virtual model, disturbance prediction of the milling cutter can be achieved.

[0066] A disturbance feature library is constructed using the physical workshop data features, the workshop digital twin model features, and a historical feature database. The establishment of this library first requires classifying and categorizing workshop disturbances to clarify their characteristic manifestations. Based on the types of workshop data collected and potential problems during production, disturbance events are mainly categorized into equipment-related, task-related, and storage-related events. According to information such as the location and characteristics of workshop disturbances, disturbances are classified, and a three-layer analytical disturbance feature library is established: "Disturbance Location—Disturbance Type—Disturbance Feature."

[0067] Based on the standard perturbation feature cases in the perturbation feature library, set the perturbation threshold.

[0068] The physical workshop data features and the workshop digital twin model features are input into a pre-built disturbance prediction model, and the probability of disturbance occurrence is output.

[0069] The construction of the disturbance prediction model includes:

[0070] Obtain the original CNN-GRU-Attention model;

[0071] A memory network layer parallel to the CNN layer is added between the input layer and the GRU layer of the original CNN-GRU-Attention model to obtain the perturbation prediction model.

[0072] Generative adversarial networks are used to generate a perturbation dataset that is similar to the real perturbation distribution; the real perturbation data and the fake perturbation data are mixed as a new training dataset, and the perturbation prediction model is trained using the new training dataset.

[0073] like Figure 5 The diagram shows the structure of the perturbation prediction model. In this embodiment, the perturbation prediction model is a convolutional neural network (CNN-GRU-Attention) that includes a memory network and gated recurrent units, and uses an attention mechanism to dynamically allocate weights to the input features.

[0074] like Figure 4 The diagram illustrates the principle of the attention mechanism. The weighted approach of the attention mechanism involves three steps:

[0075] Step 1: Calculate the similarity between the query and the key. By employing different functions and calculation methods, the similarity between the query and a given key is calculated. The method in this embodiment includes:

[0076] The formula for calculating the dot product, cosine similarity, or using other neural networks to assess the similarity between two vectors is as follows:

[0077] Dot product:

[0078] ;

[0079] Cosine similarity:

[0080] ;

[0081] Feedforward Artificial Neural Network (MLP, Multilayer Perceptron):

[0082] .

[0083] Step 2: Normalize the scores. The calculated similarities are normalized using the Softmax function to transform them into a probability distribution, ensuring the sum of the weights is 1. The weight coefficients corresponding to the Value are then calculated, expressed by the formula:

[0084] ;

[0085] In the formula, Indicates the length of the input data.

[0086] Step 3: Perform a weighted sum of the weights and value. The attention weights, normalized using the Softmax function, are summed with the value to obtain the final attention value, expressed as:

[0087] .

[0088] The perturbation prediction model consists of five parts: an input layer, a CNN feature extraction layer, a memory network layer, a GRU-Attention layer, a fully connected layer, and an output layer.

[0089] The processing steps of the disturbance prediction model include:

[0090] Input layer: The physical workshop data features and the workshop digital twin model features are input into the CNN feature extraction layer; the physical workshop data features are input into the memory network layer.

[0091] CNN layer: Perform secondary feature extraction on the physical workshop data features and the workshop digital twin model features to obtain low-dimensional compressed features; through multiple convolution and pooling operations, the input data is represented as low-dimensional compressed data and input into the GRU-Attention layer.

[0092] Memory network layer: Retrieves historical disturbance data related to the physical workshop data characteristics to obtain historical disturbance characteristics.

[0093] The GRU-Attention layer sequentially processes the low-dimensional compressed features and the historical perturbation features through feature sequence processing, feature weight assignment, and feature fusion to obtain fused features. It fully learns the patterns between feature data and introduces weights to dynamically capture important information in the sequence, thereby better understanding the long-term dependencies of the workshop state and providing more accurate information for the prediction task. The specific processing flow of the GRU-Attention layer is as follows:

[0094] Step 1: Calculate the update and reset gates for the GRU. Calculate the update gate based on the input of the current time step and the hidden state of the previous time step. and reset door The activation value is represented as:

[0095] ;

[0096] ;

[0097] in, This represents the Sigmoid activation function, used for updating and resetting the gates, mapping input values ​​to the range of 0 to 1. This represents the weight matrix during the update process; Indicates time step The hidden state; This represents the input data of the sequence at time step t; The bias vector represents the update process; The weight matrix represents the reset process; This represents the bias vector during the reset process.

[0098] Step 2: Update the candidate hidden state. Using the reset gate and the input of the current time step, calculate the candidate hidden state for subsequent updates, represented as:

[0099] ;

[0100] in, Represents the hyperbolic tangent activation function; Represents the hidden weight matrix; This represents the hidden bias vector.

[0101] Step 3: Update the hidden state. Using the input gate and candidate hidden states, combined with the hidden state from the previous time step, calculate the hidden state for the current time step. , is represented as:

[0102] .

[0103] Step 4: After processing all time steps, finalize the hidden state. As the output of GRU, we obtain a vector set. And input it into the attention layer.

[0104] Step 5: In the attention layer, calculate the hidden state at each time step. Corresponding attention score , is represented as:

[0105] ;

[0106] in, The weight matrix representing attention; Represents a learnable weight vector; This represents the bias vector.

[0107] Step 6: Normalize the attention scores using the Softmax activation function to obtain the attention weights. , represented as:

[0108] ;

[0109] in, This represents the number of all attention mechanisms. If it's a dual-head attention mechanism, then... =2.

[0110] Step 7: Calculate the context attention vector using the obtained attention weights. Attention vector The input sequence is weighted and summed.

[0111] .

[0112] Fully connected layer: The fused features are concatenated, and a classifier is connected after the fully connected layer to classify the resulting feature vector into a target, obtaining the probability of perturbation. Since this embodiment deals with a binary classification problem (perturbation event = 1, normal event = 0), a sigmoid activation function is used to map the output value to a probability value between 0 and 1, which is the probability of perturbation. The Sigmoid activation function is expressed as:

[0113] .

[0114] Output layer: Sets the perturbation threshold based on standard perturbation feature cases in the constructed perturbation feature library. .

[0115] The disturbance occurrence probability is compared with the disturbance threshold to obtain the disturbance prediction result. If the event is normal, the output is 1 and it is determined to be a disturbance event; otherwise, the output is 0 and it is determined to be a normal event.

[0116] Based on the disturbance prediction results, update the standard disturbance features of the digital twin workshop model and physical workshop data, as well as the physical workshop data and the disturbance feature library.

[0117] Example 2

[0118] This embodiment introduces a disturbance prediction system for a digital twin factory, specifically including:

[0119] The data acquisition module acquires physical workshop data and establishes a corresponding digital twin model of the workshop.

[0120] The feature extraction module extracts multimodal features from the physical workshop data and the workshop digital twin model to obtain physical workshop data features and workshop digital twin model features;

[0121] The threshold setting module constructs a disturbance feature library using the physical workshop data features, the workshop digital twin model features, and the historical feature database; and sets the disturbance threshold based on the standard disturbance feature cases in the disturbance feature library.

[0122] The physical workshop data features and the workshop digital twin model features are input into a pre-built disturbance prediction model, and the disturbance occurrence probability is output.

[0123] The construction of the disturbance prediction model includes:

[0124] Obtain the original CNN-GRU-Attention model;

[0125] Between the input layer and the GRU layer of the original CNN-GRU-Attention model, a memory network layer parallel to the CNN layer is added to obtain the perturbation prediction model.

[0126] The disturbance occurrence probability is compared with the disturbance threshold to obtain the disturbance prediction result.

[0127] Example 3

[0128] like Figure 6 As shown in the figure, this embodiment introduces a specific experimental design for a disturbance prediction method in a digital twin factory, including the following steps:

[0129] The hardware equipment of the milling cutter wear disturbance prediction experimental platform mainly includes a CNC milling machine, vibration sensor, rotary cutting force gauge, tool appearance inspection instrument, and data acquisition card, etc. Figure 6 As shown, the selected equipment models are as follows: the CNC milling machine is a DMU-70V; the vibration sensor is a Vtall-S203L-1 three-axis vibration sensor, supporting Modbus and RS485 communication protocols, with a sampling frequency set to 20kHz; the rotary cutting force gauge is a 9171A rotary multi-component cutting force gauge, which only needs to be installed on the spindle to measure the force of the milling cutter in three axial directions. The sampling frequency was set to 20kHz; the tool appearance inspection instrument was model YK-D1800B-2, used for tool wear measurement; the data acquisition card was model PCI5650-D0, and the sampling frequency was set to 20kHz; the milling cutter material was high-speed steel with a hardness of 63; and the workpiece material was No. 45 steel.

[0130] Based on the parameter ranges of the milling cutter and the milling machine, cutting tests were designed for three working conditions. The specific experimental parameters are shown in Table 1. During the experimental machining, climb milling was used, with a single tool path of 200mm. 200 tool passes were performed under each working condition. After each single-segment milling operation was completed, it was numbered, and the machine tool vibration signal, milling force data, and tool flank wear status were measured and recorded.

[0131] Table 1 Experimental parameters for prop wear

[0132]

[0133] Tool wear stages are generally divided into initial wear, normal wear, and severe wear. In actual machining, the wear rate of tools may vary under different working conditions, thus the division of wear stages will also differ. The division of wear stages is mainly based on the trend of the tool wear curve and the rate of change of wear. Therefore, this paper divides the wear stages according to the changing trend of the tool flank wear curve and the rate of change of the wear value before and after the curve. First, the approximate segmented region is determined by the wear curve, and then the position with the largest rate of change of the wear value before and after the curve is selected as the final stage division position, and severe wear is defined as a disturbance event. The wear types and sample sizes under the three working conditions are shown in Table 2.

[0134] Table 2. Wear types of cutting tools and corresponding sample numbers under three working conditions.

[0135]

[0136] The milling cutter wear disturbance prediction algorithm was run on a Windows 11 computer. The specific hardware and software parameters are shown in Table 3.

[0137] Table 3 Hardware and software parameters of the milling cutter wear disturbance prediction algorithm platform

[0138]

[0139] To verify the effectiveness of the perturbation prediction model proposed in this paper, the prediction results are compared with those of CNN, CNN-LSTM, and CNN-GRU models, as shown in Table 4. The table shows that the model in this embodiment performs best in all four performance metrics under the three different operating conditions. In particular, under operating condition OC3, the model in this embodiment achieves optimal performance, with accuracy, precision, recall, and value of 97.5%, 95.38%, 96.88%, and 96.12%, respectively. This corresponds to the lowest loss value under operating condition OC3. Compared with the CNN-GRU model without an attention mechanism, the prediction accuracy and precision are improved by an average of 0.84% ​​and 4.16%, respectively.

[0140] Table 4 Comparison of Experimental Results

[0141]

[0142] Example 3

[0143] This embodiment introduces a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the disturbance prediction method for the digital twin workshop described in Embodiment 1 or 3.

[0144] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0145] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0146] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0147] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0148] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A disturbance prediction method for a digital twin factory, characterized in that, include: Acquire physical workshop data and establish corresponding digital twin models of the workshop; Extract the multimodal features of the physical workshop data and the workshop digital twin model to obtain the physical workshop data features and the workshop digital twin model features; Using the physical workshop data features, the workshop digital twin model features, and the historical feature database, a disturbance feature library is constructed; based on the standard disturbance feature cases in the disturbance feature library, a disturbance threshold is set; The physical workshop data features and the workshop digital twin model features are input into a pre-built disturbance prediction model, and the disturbance occurrence probability is output. By comparing the probability of the disturbance occurrence with the disturbance threshold, the disturbance prediction result is obtained; The construction of the disturbance prediction model includes: Obtain the original CNN-GRU-Attention model; Between the input layer and the GRU layer of the original CNN-GRU-Attention model, a memory network layer parallel to the CNN layer is added to obtain the perturbation prediction model. The processing steps of the disturbance prediction model include: The physical workshop data features and the workshop digital twin model features are input into a CNN layer for secondary feature extraction to obtain low-dimensional compressed features; The physical workshop data features are input into the memory network layer, and historical perturbation data related to the physical workshop data features are retrieved to obtain historical perturbation features; The low-dimensional compressed features and the historical perturbation features are sequentially input into the GRU layer and the Attention layer, and feature sequence processing, feature weight assignment and feature fusion are performed sequentially to obtain the fused features. The fused features are sequentially input into the fully connected layer and the classifier layer, and then concatenated and classified to obtain the probability of perturbation occurrence.

2. The disturbance prediction method for digital twin workshops according to claim 1, characterized in that, Also includes: Based on the disturbance prediction results, update the standard disturbance features of the digital twin workshop model and physical workshop data, as well as the physical workshop data and the disturbance feature library.

3. The disturbance prediction method for digital twin workshops according to claim 1, characterized in that, Before extracting features from the physical workshop data, the process also includes: The physical workshop data undergoes preprocessing operations including data redundancy removal, data smoothing, and data noise reduction.

4. The disturbance prediction method for digital twin workshops according to claim 1, characterized in that, The multimodal features extracted from the physical workshop data include: For one-dimensional time-series data of physical workshop data, time-domain features, frequency-domain features, and time-frequency-domain features are extracted respectively. For multi-source heterogeneous data of physical workshop data, it is divided into multiple groups of one-dimensional data. The features of each group of one-dimensional data are extracted using a single channel. The extracted features of each group of one-dimensional data are combined into a feature matrix. The feature matrix is ​​then processed into grayscale to obtain a feature grayscale image. The feature grayscale image is used as the multimodal feature of the multivariate heterogeneous data.

5. The disturbance prediction method for digital twin workshops according to claim 1, characterized in that, The extraction of multimodal features from the workshop digital twin model includes extracting the time features, process features, and equipment features of the workshop digital twin model.

6. The disturbance prediction method for digital twin workshops according to claim 1, characterized in that, The disturbance feature library is a three-layer analytical disturbance feature library consisting of "disturbance location - disturbance type - disturbance feature".

7. The disturbance prediction method for digital twin workshops according to claim 1, characterized in that, The training steps for the perturbation prediction model include: Generative adversarial networks are used to generate perturbation datasets that are similar to real perturbation distributions. The perturbation prediction model is trained by mixing real and spurious perturbation data as a new training dataset.

8. A disturbance prediction system for a digital twin factory, characterized in that, include: The data acquisition module acquires physical workshop data and establishes a corresponding digital twin model of the workshop. The feature extraction module extracts multimodal features from the physical workshop data and the workshop digital twin model to obtain physical workshop data features and workshop digital twin model features; The threshold setting module constructs a disturbance feature library using the physical workshop data features, the workshop digital twin model features, and the historical feature database; and sets the disturbance threshold based on the standard disturbance feature cases in the disturbance feature library. The physical workshop data features and the workshop digital twin model features are input into a pre-built disturbance prediction model, and the disturbance occurrence probability is output. The construction of the disturbance prediction model includes: Obtain the original CNN-GRU-Attention model; Between the input layer and the GRU layer of the original CNN-GRU-Attention model, a memory network layer parallel to the CNN layer is added to obtain the perturbation prediction model. By comparing the probability of the disturbance occurrence with the disturbance threshold, the disturbance prediction result is obtained; The processing steps of the disturbance prediction model include: The physical workshop data features and the workshop digital twin model features are input into a CNN layer for secondary feature extraction to obtain low-dimensional compressed features; The physical workshop data features are input into the memory network layer, and historical perturbation data related to the physical workshop data features are retrieved to obtain historical perturbation features; The low-dimensional compressed features and the historical perturbation features are sequentially input into the GRU layer and the Attention layer, and feature sequence processing, feature weight assignment and feature fusion are performed sequentially to obtain the fused features. The fused features are sequentially input into the fully connected layer and the classifier layer, and then concatenated and classified to obtain the probability of perturbation occurrence.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the disturbance prediction method for the digital twin workshop as described in any one of claims 1-7.

Images