A tool wear state monitoring method and system based on digital twinning

Abstract

The application discloses a tool wear state monitoring method and system based on digital twinning, and the method comprises the following steps: acquiring the spindle speed, feed amount and milling force time sequence signal in the machine tool machining process, and preprocessing to obtain the measured milling force time sequence signal; a digital twin body following the machine tool is constructed, the cutting width and cutting depth are determined, and the workpiece material properties are acquired; a new tool theoretical calculation milling force time sequence signal is generated based on a new tool theoretical model; the measured milling force time sequence signal and the new tool theoretical calculation milling force time sequence signal are taken as inputs, the difference feature vector of the two in the multi-dimensional feature space is extracted through a contrast learning network, and a tool wear value is obtained through mapping; a pseudo sample incremental learning mechanism is introduced into the contrast learning network, which is used for predicting the tool wear value under a new working condition; when the tool wear value reaches a preset replacement threshold, the machine tool performs a tool replacement operation. The application can realize cross-working-condition generalization and efficient incremental learning of tool wear precise identification.

Description

Technical Field

[0001] This invention relates to the fields of digital twins and tool wear detection, and in particular, to a method and system for monitoring tool wear status based on digital twins. Background Technology

[0002] Tool wear monitoring is a critical aspect affecting machining quality. With the development of artificial intelligence technology, digital twins utilize virtual simulation to generate training data or provide benchmark references, opening up new avenues for monitoring tool condition under complex working conditions. However, traditional tool wear monitoring methods using digital twins suffer from poor generalization ability of the detection model, difficulty in obtaining sample data, and strong data dependence in complex machining environments. Therefore, higher demands are placed on accurately identifying tool wear conditions under complex working conditions.

[0003] While existing research on tool wear monitoring has made progress at the algorithm level, it faces the following bottlenecks in actual machining: First, frequent changes in cutting conditions lead to insufficient generalization ability of purely data-driven methods. Although physical models can reveal the wear mechanism, they are difficult to dynamically obtain the cutting width and depth based on the real-time relative position of the tool and workpiece, and cannot generate new tool theoretical calculation milling force in real time as a physical benchmark, making it difficult to embed into the machining process to achieve dynamic response. Second, in cross-condition generalization scenarios, existing systems often need to collect a large number of labeled samples from new conditions for pre-training, lacking the ability to continuously adapt to new conditions, which limits the flexibility and deployment efficiency of the monitoring system. Third, most existing data-driven and physical model fusion methods only use computational force as an auxiliary input, failing to establish a mapping relationship between the difference between the new tool theoretical calculation milling force and the measured milling force and the tool wear amount, and lacking a closed-loop training mechanism to update the model by backpropagation using the difference between the two. Fourth, most studies focus on the monitoring accuracy itself, paying little attention to tool management systems that achieve a closed loop of "monitoring-decision-control".

[0004] Therefore, a tool wear monitoring method is needed that can acquire cutting parameters in real time using digital twin technology and generate a time-series signal of the milling force calculated by the new tool in theory as a physical benchmark. The difference between the time-series signal of the milling force calculated by the new tool in theory and the time-series signal of the actual milling force is mapped to the tool wear value, and a loss update mechanism for backpropagation of force signal difference is established. At the same time, pseudo-sample incremental learning is introduced to achieve continuous adaptation to new working conditions under the condition of no sample under changing working conditions. Finally, it provides a decision basis for the system to control the machine tool tool change, forming a complete closed-loop machining monitoring. Summary of the Invention

[0005] The main objective of this invention is to provide a tool wear status monitoring method and system based on digital twins, which enables accurate identification of tool wear through cross-working-condition generalization and efficient incremental learning, and provides a decision basis for the system to control machine tool tool changing.

[0006] To achieve the above objectives, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for monitoring tool wear status based on digital twins, the method comprising: The overall construction of a digital twin can be divided into a physical layer, an edge computing layer, a digital twin layer, a deep learning model layer, and a cloud database. During the machining process, the milling force timing signal is collected in real time by the force measuring instrument, and the spindle speed and feed rate of the machine tool are acquired simultaneously. The edge computing terminal aggregates the spindle speed and feed rate through communication, and preprocesses the acquired milling force timing signal to obtain the preprocessed measured milling force timing signal; A digital twin that moves with the machine tool is constructed. Its front end determines the cutting width and depth of cut based on the envelope model of the workpiece and the tool and their real-time position. It receives the material properties specified by the user through an interactive interface and stores the spindle speed, feed rate, cutting width, cutting depth and material properties into a buffer according to time steps. The back end of the twin is embedded with a brand-new tool theory model and a milling force physical mechanism calculation model. It reads the above parameters from the buffer and generates a new tool theory calculation milling force timing signal in real time. The measured milling force time series signal and the milling force time series signal calculated by the new tool theory are used as inputs simultaneously. The idea of ​​contrastive learning is introduced to extract the difference feature vectors between the two in the multi-dimensional feature space. The predicted tool wear state, i.e. the tool wear value, is obtained by mapping the difference feature vectors and then backpropagating to update the network parameters. The comparative learning network incorporates a pseudo-sample incremental learning mechanism. When new cutting parameters for a new working condition are detected, the existing network branches for that working condition are fixed, and new network branches are created for learning the new working condition. Simultaneously, the digital twin system embeds a calculation model of the physical mechanism of milling force, retrieves the complete set of cutting process parameters for the new working condition stored in the cache, and uses existing mature milling force theoretical calculation formulas to complete the numerical solution of the milling force. Two types of milling force time-series signal data are generated in parallel: one type is a calculated milling force time-series signal considering tool wear, taking into account the progressive wear evolution law of the tool and wear characteristics, and generating corresponding wear labels. Another type is the milling force time-series signal calculated based on a brand-new tool theory model without considering tool wear. This is used to incrementally train and update the parameters of the new network branches using pseudo-samples. After the actual cutting samples and corresponding tool wear true values ​​generated at the machine tool end are uploaded to the cloud database, the batch of real sample datasets with true value annotations are retrieved, and the network branches corresponding to the current working conditions are updated with secondary refined parameters. The deviation caused by pseudo-samples is corrected based on the actual working condition measurement data, which further improves the model's wear prediction accuracy and working condition adaptation reliability, so that the model can continuously adapt to new machining conditions. All data generated at each stage of the system is uploaded to a cloud database for monitoring continuous model updates and tracing the processing.

[0007] Furthermore, the physical layer includes: Parameters such as spindle speed, feed rate, spindle position and spindle power are obtained through physical entities, and the timing signal of actual milling force is collected by measuring instruments.

[0008] Furthermore, the edge computing layer includes: The parameters and signals obtained from the physical layer are acquired through a communication protocol, and the signals are preprocessed. The preprocessing includes: first, low-pass filtering is performed on the acquired milling force timing signal to remove high-frequency noise interference, then wavelet thresholding is performed for noise reduction, and finally, the signal is aligned according to the machining time step to obtain the preprocessed measured milling force timing signal.

[0009] Furthermore, the digital twin layer includes: Based on entity information obtained from the edge computing layer, a motion mapping between virtual and physical objects is constructed. An envelope model of the workpiece and tool is established according to the mapping relationship, thereby obtaining a cutting parameter model and calculating the cutting width and cutting depth in real time. Material properties are input through the user interface to finally realize the calculation of milling force.

[0010] The specific steps for calculating milling forces in a digital twin layer include: Establish communication between the digital twin system and the machine tool, and obtain spindle speed, feed rate, tool position and measured milling force through the edge computing layer to realize virtual motion mapping for synchronous operation of the system and the machine tool; At the front end of the digital twin system, an envelope model of the workpiece and the tool is constructed based on the relative positional relationship between the tool and the workpiece. The tool envelope volume is obtained from the envelope model, and then the cutting width and cutting depth are calculated in real time. By inputting material properties through the user interface, the spindle speed, feed rate, cutting width, cutting depth, and material properties are used as parameters required for calculating milling force. Based on a new tool theory model, a new tool theory-based milling force timing signal is generated. Under new working conditions, based on the milling force physical mechanism calculation model embedded in the digital twin, a milling force timing signal considering tool wear and the corresponding simulated tool wear label value are generated. Based on the new tool theory model, a corresponding new tool theory-based milling force timing signal is generated.

[0011] The calculation model of the physical mechanism of milling force includes a brand-new tool theory model. That is, when the wear value is zero, the calculation model of the physical mechanism of milling force is the brand-new tool theory model. It can be understood that the formula for calculating the milling force signal considering tool wear includes the formula for calculating the milling force signal using the new tool theory.

[0012] Furthermore, the deep learning model layers include: A contrastive learning architecture is constructed, using the milling force time-series signal calculated by the new tool theory in real time generated by the digital twin layer and the preprocessed measured milling force signal as common inputs. The contrastive learning network extracts the difference feature vectors between the two in the multi-dimensional feature space, and the tool wear value is obtained based on the difference feature vector mapping, thereby completing the tool wear state monitoring. The predicted tool wear value and the corresponding spindle speed, feed rate, cutting width, cutting depth and material properties are input into the milling force physical mechanism calculation model to calculate the reconstructed milling force signal in reverse. If there is no error, the reconstructed milling force signal should be the same as the measured force signal. The reconstructed milling force signal is compared with the preprocessed measured milling force signal to construct the force signal reconstruction loss. The wear mapping loss corresponding to the difference features and the force signal reconstruction loss are weighted and summed to obtain the total loss, which is used to update the model parameters. Specifically, during incremental training, the parameters of the new network branch are updated based on the wear mapping loss; during secondary training, the parameters of the new network branch are updated based on the total loss of the weighted sum of the wear mapping loss and the force signal reconstruction loss; after repeated iterative training, the model finally outputs the tool wear state, and the wear state is integrated into the digital twin layer. The system controls the machine tool to perform tool changing operations through G-code according to the wear state.

[0013] In this deep learning model, a pseudo-sample incremental learning mechanism is introduced into the contrastive learning network. Specifically, during real-time machining, when new cutting parameters are detected under new working conditions, the old network branches for existing working conditions are fixed, and new network branches are created for learning the new working conditions. Based on the spindle speed, feed rate, width at cut, depth of cut, and material properties under the new working conditions, and using the milling force physical mechanism calculation model embedded in the digital twin, a calculated milling force time-series signal considering tool wear and the corresponding simulated tool wear label value are generated under the new working conditions. Furthermore, a corresponding new tool theory calculated milling force time-series signal is generated based on a completely new tool theory model. The calculated milling force time-series signal considering tool wear under the new working conditions and the new tool theory calculated milling force time-series signal are then used to calculate the milling force. A pseudo-sample dataset without actual ground truth labels is constructed using sequence signals and simulated tool wear labels. This pseudo-sample dataset is then input into a new network branch. The calculated milling force time-series signal considering tool wear and the corresponding theoretically calculated milling force time-series signal for the new tool are used as inputs, with the simulated tool wear labels serving as the supervision signal. This updates the parameters of the new network branch, completing real-time incremental training. Once the actual milling force time-series signal and the corresponding ground truth value for tool wear under the new working condition are obtained, the new network branch is trained a second time using these signals. This yields a tool wear state monitoring model for the new working condition, used to predict tool wear values ​​under the new working condition.

[0014] Deep learning training methods include the following steps: A contrastive learning architecture is constructed, using the milling force time-series signal calculated by the new tool theory in real time by the digital twin system and the measured milling force time-series signal collected and preprocessed during the machine tool processing as common inputs. Feature extraction and feature fusion are performed on the input data to extract the difference feature vectors between the two in the multi-dimensional feature space. The difference feature vectors are used to map the actual tool wear value, and a deep learning model is initially constructed. The predicted tool wear value, along with the corresponding spindle speed, feed rate, width at cut, depth of cut, and material properties, are input into the milling force physical mechanism calculation model. The model then calculates the reconstructed milling force signal in reverse. If there are no errors, the calculated reconstructed milling force signal should be identical to the measured force signal. The deviation between the reconstructed force signal and the preprocessed measured milling force signal is used to construct the force signal reconstruction loss. Simultaneously, using the measured tool wear value as a label, wear mapping loss is calculated for the difference characteristics. The wear mapping loss and the force signal reconstruction loss are weighted and summed to obtain the total loss. Model optimization and validation: The model parameters are updated by backpropagation using the total loss. The optimized model is obtained through repeated iterative training. If the model does not meet the accuracy requirements, the model is retrained. During real-time machining on a machine tool, when the cutting parameters for a new working condition are detected during real-time cutting, the old network branches for the existing working condition are fixed, and a new network branch is created for learning the new working condition. The digital twin layer can generate a time-series signal of the milling force under this new working condition by real-time calculation, which includes the time-series signal of the milling force calculated by the new tool theory and the corresponding simulated wear label value. It is then processed and a pseudo-sample dataset without actual measured true value labeling is constructed. The pseudo-sample dataset is input into the newly created network branch. The calculated milling force time series signal considering tool wear is used to replace the measured milling force time series signal under the sample condition. This signal and the theoretically calculated milling force time series signal of the new tool are used as input together. The corresponding simulated wear label value in the pseudo-sample dataset is used to replace the real wear value under the sample condition. The parameters of the new network branch are updated based on the wear mapping loss to complete real-time incremental training. In this way, the model can complete the cross-condition parameter iterative update under the condition that there are no field measured labeled samples in the target working condition. After the actual cutting samples and corresponding tool wear true data generated at the machine tool end are uploaded to the cloud database, the model retrieves the batch of real sample datasets with true value annotations and performs a second fine-tuning parameter update on the network branch corresponding to the current working condition. During the second training, the network parameters are updated based on the total loss of the weighted sum of wear mapping loss and force signal reconstruction loss. The deviation caused by pseudo samples is corrected by relying on the actual measured data of the real working condition, further improving the model's wear prediction accuracy and working condition adaptation reliability. The iteratively saved model is used for real-time detection and intelligent guidance of tool wear during subsequent cutting processes under this working condition; The trained deep learning model is integrated into the digital twin system. The system determines whether the preset replacement threshold has been reached based on the tool wear status output by the model in real time. Once the threshold is reached, the machine tool is controlled to perform a tool change operation via G-code.

[0015] Furthermore, the cloud database layer includes: In the system constructed by this method, the data generated by the physical layer, edge computing layer, digital twin layer and deep learning model layer are all uploaded to the cloud database with time labels to realize the continuous updating of the tool wear monitoring model and the traceability of the machining process.

[0016] In a second aspect, the present invention provides a digital twin-based tool wear condition monitoring system for implementing the digital twin-based tool wear condition monitoring method described in the first aspect, the system comprising: The data acquisition and preprocessing module is used to acquire the spindle speed, feed rate and milling force timing signals during the machine tool machining process, and to preprocess the milling force timing signals to obtain the measured milling force timing signals. The digital twin module is used to acquire the machine tool's spindle speed, feed rate, tool position, and measured milling force timing signals in real time via communication protocols, realizing synchronous tracking mapping between the digital twin and the physical machine tool. At the front end of the twin system, the cutting width and cutting depth are dynamically calculated based on the tool and workpiece envelope model and their real-time relative positions. The system also receives the workpiece material properties specified by the user through a human-machine interface and stores the above process parameters in a time-series buffer. The back end of the twin system embeds a milling force physical mechanism model and a new tool theory model. It can read parameters from the buffer to generate new tool theory calculation milling force timing signals and calculation milling force timing signals considering tool wear in real time. The wear status and trend curves are visualized on the interactive interface, providing a unified platform for machining status traceability and reverse control tool changing. The deep learning module uses the milling force time-series signal calculated by the new tool theory and the preprocessed measured milling force time-series signal as common inputs. It employs a contrastive learning architecture to extract the difference vector between the two signals in a multi-dimensional feature space, and maps this vector to obtain the predicted tool wear value. During training, the predicted tool wear value and corresponding process parameters are input into the milling force physical mechanism calculation model. The model then reconstructs the milling force signal by calculating the difference between the reconstructed force and the preprocessed measured milling force, constructing a force signal reconstruction loss. The wear mapping loss corresponding to the difference features is weighted and summed with the force signal reconstruction loss to obtain the total loss, and the model parameters are iteratively updated. Under new working conditions, during incremental training, the parameters of the new network branch are updated based on the wear mapping loss. During secondary training, the parameters of the new network branch are updated based on the total loss of the weighted sum of the wear mapping loss and the force signal reconstruction loss. A pseudo-sample incremental learning module is used to introduce a pseudo-sample incremental learning mechanism into the contrastive learning network. This includes: during real-time machining on a machine tool, when new cutting parameters for a new working condition are detected, fixing the existing network branches for that working condition and creating new network branches for learning the new working condition; based on the spindle speed, feed rate, width at cut, depth of cut, and material properties under the new working condition, and using the milling force physical mechanism calculation model embedded in the digital twin, generating a calculated milling force time-series signal considering tool wear under the new working condition and the corresponding simulated tool wear label value, and generating a corresponding new tool theory calculated milling force time-series signal based on a completely new tool theory model; using the calculated milling force time-series signal considering tool wear under the new working condition and the new tool theory calculated milling force time-series signal... A pseudo-sample dataset without actual ground truth labels is constructed using cutting force timing signals and simulated tool wear label values. This pseudo-sample dataset is then input into a new network branch. The calculated milling force timing signal considering tool wear and the corresponding theoretically calculated milling force timing signal for the new tool are used as inputs, and the simulated tool wear label value is used as a supervision signal to update the parameters of the new network branch, completing real-time incremental training. Once the actual milling force timing signal and the corresponding ground truth value for the new working condition are obtained, the new network branch is trained a second time using these signals, resulting in a tool wear state monitoring model for the new working condition, used to predict tool wear values ​​under the new working condition. The cloud database module is used to upload all process data generated during the operation of the twin system, such as process parameters, calculated milling force, measured milling force, predicted wear value, model update records and tool change records, to the cloud database with time tags. This enables full traceability of machine tool status, workpiece status and cutting status, and provides historical data support for incremental learning. The tool change control module is used to determine whether the tool wear value output in real time by the deep learning model has reached the preset replacement threshold. Once it is reached, the twin system platform generates G-code instructions to control the machine tool to automatically perform the tool change operation, forming a closed-loop machining monitoring system of "perception-analysis-decision".

[0017] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects: First, by mapping the virtual motion of the workpiece and tool envelope model in the digital twin layer, the cutting width and cutting depth are dynamically obtained based on the real-time relative position of the tool and the workpiece. Combined with the material properties input by the interactive interface, the embedded milling force physical mechanism calculation model generates a milling force time-series signal that takes into account tool wear in real time. The new tool theory model generates a new tool theory calculation milling force time-series signal. Without the need to install additional depth-of-cut and width-of-cut sensors, convenient and accurate cutting parameter acquisition is achieved, providing a physical benchmark for wear monitoring.

[0018] Second, a contrastive learning architecture is introduced into the deep learning model layer to calculate the multidimensional feature difference between the milling force time series signal and the measured milling force time series signal to map the tool wear amount. Force signal reconstruction loss is constructed during basic training and secondary training. The deviation between the reconstructed force and the measured force is used as a constraint to update the model parameters, forming a closed-loop training mechanism. This makes the model physically interpretable and significantly improves the accuracy and robustness of wear state identification.

[0019] Third, the deep learning model employs a pseudo-sample incremental learning mechanism. When new cutting parameters are detected, the old network branches for existing working conditions are fixed, and new network branches are created for learning the new working conditions. Based on the spindle speed, feed rate, width of cut, depth of cut, and material properties under the new working conditions, and based on the milling force physical mechanism calculation model embedded in the digital twin, a calculated milling force time series signal considering tool wear and the corresponding simulated tool wear label value are generated under the new working conditions. A corresponding new tool theory calculated milling force time series signal is also generated based on a completely new tool theory model. Using the calculated milling force time series signal considering tool wear under the new working conditions, the new tool theory calculated milling force time series signal, and the simulated tool wear label value, a pseudo-sample dataset without actual measured ground truth labels is constructed. The pseudo-samples are then used to... The dataset is input into a new network branch, taking the calculated milling force time-series signal considering tool wear and the corresponding theoretically calculated milling force time-series signal of the new tool as inputs, and the simulated tool wear label value as a supervision signal to update the parameters of the new network branch, completing real-time incremental training. After obtaining the real milling force time-series signal and the corresponding true value of tool wear under the new working condition, the new network branch is trained a second time using the real milling force time-series signal under the new working condition, the corresponding theoretically calculated milling force time-series signal of the new tool, and the true value of tool wear, to obtain a tool wear state monitoring model under the new working condition. This model is used to predict the tool wear value under the new working condition, enabling the model to continuously adapt to the new working condition, avoiding catastrophic forgetting, and improving the accuracy and sustainable adaptability of the monitoring system under cross-working condition generalization conditions.

[0020] Fourth, all data generated at each layer of the system is uploaded to the cloud database, enabling full traceability of the machining process. At the same time, by monitoring wear status and tool change thresholds in real time, the twin system automatically generates G-codes to control machine tool tool changes, forming a closed-loop machining monitoring system of "perception-analysis-decision". Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating a method for monitoring tool wear status based on digital twins according to an embodiment of the present invention. Figure 2 This is a schematic diagram illustrating the setup of a tool wear status monitoring system based on digital twin according to an embodiment of the present invention; Figure 3This is a geometric schematic diagram of the cutting principle of the tool envelope and the workpiece according to an embodiment of the present invention; Figure 4 This is a flowchart of a deep learning model training process according to an embodiment of the present invention; Figure 5 This is a schematic diagram of a cloud database according to an embodiment of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. All other embodiments obtained by those skilled in the art based on the embodiments provided by this invention without inventive effort are within the scope of protection of this invention.

[0023] It should be noted that the accompanying drawings in the embodiments of this invention are only used to schematically illustrate the basic concept of the invention. Therefore, the drawings only show components related to the invention and are not drawn according to the actual number, shape, and size of the components in implementation. In actual implementation, the shape, number, and proportion of each component can be adjusted according to the actual situation, and its layout may also be more complex.

[0024] This invention provides a method and system for monitoring tool wear status based on digital twins. The method includes: real-time acquisition and preprocessing of spindle speed, feed rate, and milling force signals to obtain measured milling force time-series signals; constructing a digital twin that moves with the machine tool, determining the cutting width and depth based on the envelope model and the real-time tool-workpiece position, and generating a new tool theoretically calculated milling force time-series signal in real time by combining material properties and an embedded milling force physical mechanism calculation model, i.e., a brand-new tool theoretical model; using the new tool theoretically calculated milling force time-series signal and the measured milling force time-series signal as input, comparing and learning to extract the difference features to map wear values, constructing a force reconstruction loss based on the deviation between the reconstructed milling force and the measured milling force during training, and updating the model in conjunction with the wear mapping loss; simultaneously introducing a pseudo-sample incremental learning mechanism to achieve adaptive adaptation under changing working conditions by fixing old branches, opening new branches, and constructing pseudo-sample sets; uploading data to a cloud database for traceability, and automatically generating G-codes to change tools when wear exceeds a threshold. This invention uses milling force calculated by the new tool theory as a physical benchmark. Through force difference mapping of wear and reconstruction loss closed-loop update, it has the ability to generalize across working conditions and incremental learning, forming a "perception-analysis-decision" closed loop.

[0025] It should be noted that the milling force physical mechanism calculation model embedded in the digital twin backend actually includes a brand-new tool theory model. It can be understood that in the process of generating the calculated milling force signal that takes tool wear into account, a new tool theory calculation milling force signal is generated. However, for the sake of convenience, it is written in the article as the digital twin backend embedding a milling force physical mechanism calculation model and a brand-new tool theory model.

[0026] Example 1: Reference Figure 1 This embodiment provides a method for monitoring tool wear status based on digital twins, including the following steps: S1: Acquire the timing signals of spindle speed, feed rate and milling force during the machine tool cutting process, and preprocess them to obtain the measured milling force timing signal; S2: Construct a digital twin that follows the machine tool, determine the cutting width and depth, and obtain the workpiece material properties; S3: Generate milling force timing signals based on a new tool theory model; S4: Using the measured milling force time series signal and the milling force time series signal calculated by the new tool theory as input, the difference feature vector between the two in the multi-dimensional feature space is extracted through the contrastive learning network, and the tool wear value is obtained by mapping. S5: A pseudo-sample incremental learning mechanism is introduced into the contrastive learning network. During real-time machining of the machine tool, when new cutting parameters are detected under new working conditions, the old network branches for the existing working conditions are fixed, and new network branches are opened. A pseudo-sample dataset without actual measured ground truth labels is constructed. The calculated milling force time series signal considering tool wear and the corresponding theoretically calculated milling force time series signal for the new tool are used as inputs, and the simulated tool wear label value is used as the supervision signal. The parameters of the new network branch are updated based on the wear mapping loss to complete real-time incremental training. When the real milling force time series signal and the corresponding tool wear ground truth under the new working condition are obtained, the new network branch is trained a second time based on the total loss of the weighted sum of the wear mapping loss and the force signal reconstruction loss to obtain the tool wear state monitoring model under the new working condition. This model is used to predict the tool wear value under the new working condition to continuously monitor the tool wear value under the current working condition. S6: When the tool wear value reaches the preset replacement threshold, the machine tool is controlled to perform a tool change operation. All data generated by each link of the system is uploaded to the cloud database for continuous monitoring of model updates and traceability of the machining process.

[0027] The detection and determination of new working conditions can be carried out in the following ways: (1) The spindle speed, feed rate, cutting width, cutting depth and material properties are mapped to working condition IDs. When the model detects that the ID corresponding to the current working condition does not match the existing working condition ID, that is, when any process parameter changes, it is determined to be a new working condition; (2) Using process parameters and cutting force distribution as features, unsupervised clustering is performed on all working conditions, multidimensional representation vectors of each working condition are extracted, feature similarity between working conditions is calculated, and multiple working conditions with highly similar process parameters, cutting force and wear evolution laws are merged into the same working condition cluster. Only one network branch is assigned to a working condition cluster; when there is a measured value of cutting force, the measured milling force is used, and when there is no measured value, the calculated milling force considering tool wear is used; when the similarity between the newly appearing working condition and the existing working condition cluster is lower than the preset threshold, it is determined to be a new working condition, and a new working condition cluster and corresponding network branch are created for it.

[0028] The tool wear monitoring method provided in this embodiment is based on the following core principle: Through virtual motion mapping of the workpiece and tool envelope model in a digital twin layer, the cutting width and depth of cut are dynamically acquired according to the real-time relative position of the tool and workpiece. Combined with material properties input through the interactive interface, a calculated milling force time-series signal considering tool wear is generated in real time by an embedded milling force physical mechanism model. A new tool theory model generates a new tool theory calculated milling force time-series signal, eliminating the need for additional depth-of-cut and width-of-cut sensors. This enables convenient and accurate acquisition of cutting parameters, providing a physical benchmark for wear monitoring. At the deep learning model layer, the multidimensional feature difference between the calculated milling force time-series signal and the measured milling force time-series signal is used to map the tool wear amount, and a force signal reconstruction loss is constructed. The deviation between the reconstructed force and the measured force is used as a constraint to update the model parameters, forming a unique closed-loop training mechanism. Simultaneously, the incremental learning mechanism continuously adapts to new working conditions, significantly improving the accuracy and robustness of wear state identification. This method only requires the calculation of milling force as input to effectively predict the tool wear state under cross-working condition generalization, overcoming the problems of poor generalization ability, difficulty in obtaining samples and strong data dependence of existing methods.

[0029] Example 2: This embodiment provides a tool wear condition monitoring system based on digital twins, including: The data acquisition and preprocessing module is used to acquire the spindle speed, feed rate and milling force signals during the machine tool machining process, and to preprocess the milling force timing signal to obtain the measured milling force timing signal; The twin system module is used to build a twin model to obtain working conditions and tool parameters and calculate milling forces. It also provides a complete system platform for machining status traceability, human-machine interface, and system reverse control of machine tool tool changing.

[0030] The deep learning module is used to construct a deep learning model under dual signal comparison input, and input the model into the digital twin model to obtain a real-time predictive tool wear monitoring model. The pseudo-sample incremental learning module introduces a pseudo-sample incremental learning mechanism into the contrastive learning network. When new cutting parameters are detected, the old network branches for the existing working conditions are fixed, and a new network branch is created. A pseudo-sample dataset without actual ground truth labels is constructed. The calculated milling force time series signal considering tool wear and the corresponding theoretically calculated milling force time series signal for the new tool in the pseudo-sample dataset are used as inputs, and the simulated tool wear label value is used as the supervision signal. The parameters of the new network branch are updated based on the wear mapping loss to complete real-time incremental training. When the real milling force time series signal and the corresponding tool wear ground truth under the new working condition are obtained, the new network branch is trained again based on the total loss of the weighted sum of the wear mapping loss and the force signal reconstruction loss to update the tool wear state monitoring model, which is used to predict the tool wear value under the new working condition. The cloud database module is used to collect all the data obtained in the twin system and deep learning model layer, and store it in the cloud database with time as the label, so as to realize machine tool status tracking, workpiece status tracking, and cutting status tracking.

[0031] The tool change judgment module uses the tool wear monitoring model to determine whether the tool wear value has reached the replacement requirement, and controls the machine tool to change tools using G-code.

[0032] The digital twin system module acquires the cutting width and depth in real time through an envelope model within the digital twin system. Combined with spindle speed, feed rate, and user-input material properties, it generates a new tool theory model to calculate the milling force time-series signal in real time. This time-labeled dataset is then uploaded to a cloud database, providing physical benchmark data for subsequent wear prediction and tool change control. The deep learning module takes the milling force time-series signal calculated by the new tool theory and the pre-processed measured milling force time-series signal as input. It employs a contrastive learning architecture to extract differences and map them to tool wear values. During basic training, the model parameters are jointly optimized using force signal reconstruction loss and wear mapping loss. During incremental training, the network parameters are updated based on the wear mapping loss. During secondary training, the network parameters are updated based on the total loss of the weighted sum of wear mapping loss and force signal reconstruction loss, providing real-time output of the current tool wear status. When the predicted value exceeds a preset threshold, an early warning is triggered, and tool change is controlled via G-code, achieving closed-loop control.

[0033] This embodiment integrates functional modules such as signal acquisition, virtual acquisition of cutting parameters, wear prediction, pseudo-sample incremental learning and adaptive control, and machine tool regulation to construct a complete digital twin intelligent machining platform. During actual machining, the platform collects data in real time, calculates cutting parameters and milling forces through virtual mapping, predicts wear trends, and automatically generates G-codes to control tool changing when wear exceeds a threshold. This achieves a closed-loop operation throughout the entire process of data acquisition, virtual mapping, prediction and early warning, and precise control, significantly improving machining accuracy and process stability.

[0034] This embodiment also provides a more specific process for building a digital twin tool wear condition monitoring system, referring to... Figure 2 The specific implementation steps are as follows: First, based on 3D modeling software, the geometric, material, and physical properties of the machine tool entity in the physical world are simulated and reconstructed in virtual space, and then lightweighted. The completed 3D model is then imported into the Unity 3D platform to complete scene editing and switching, thus initially establishing a digital twin model.

[0035] To achieve real-time synchronization and accurate response to physical entities in virtual space, it is necessary to track the parameters of the physical system to ensure that the virtual model is consistent with the physical entity in terms of structural shape, size and system response.

[0036] Based on this, the data types required for calculating milling forces and deep learning modeling are analyzed. Calculating the milling force timing signal using the new tool theory and considering tool wear requires parameters such as spindle speed, feed rate, material properties, real-time tool wear value, depth of cut, and width of cut. Spindle speed and feed rate can be directly obtained from the physical machine tool, material properties are input through the user interface, cutting parameters are obtained in real-time through virtual motion mapping of the digital twin model, and the tool milling force timing signal is collected by installing a force gauge. Furthermore, the milling force timing signal calculated using the new tool theory does not require tool wear value or defaults to zero tool wear, while the milling force timing signal considering tool wear requires tool wear label values.

[0037] Subsequently, a digital twin model of the tool and workpiece is constructed to obtain information such as the geometric features, material parameters, and installation methods of the tool and workpiece, and a 1:1 reconstruction is achieved using 3D modeling software. The model is then imported into Unity 3D to ensure the real-time and accurate mapping of the relative positional relationship between the tool and workpiece, thereby constructing an envelope model of the workpiece and tool.

[0038] The collected data is transmitted to Unity 3D. Based on communication protocols and database technology, the physical model and virtual model are deeply integrated to complete the mapping of virtual and real actions and real-time interaction. The user interface is integrated to build a complete digital twin system.

[0039] By constructing a workpiece and tool envelope model through a digital twin layer, the tool envelope volume can be obtained, thus yielding a cutting parameter model. This allows for real-time calculation of the depth of cut and width of cut under the current working conditions. Spindle speed, feed rate, width of cut, depth of cut, and material properties are stored in a buffer with time tags. A new tool theory model is embedded in the backend of the twin, and the parameters are read from the buffer to generate a new tool theory calculation timing signal for milling forces in real time.

[0040] In the deep learning model layer, the time-series signal of milling force calculated by the new tool theory and the preprocessed measured milling force signal are input into the contrastive learning architecture. The difference vector between the two in the multi-dimensional feature space is extracted, and the predicted tool wear value is obtained by mapping it. During training, the predicted tool wear value, along with the corresponding spindle speed, feed rate, width at cut, depth of cut, and material properties, are input into the milling force physical mechanism calculation model. The milling force signal is reconstructed by reverse calculation, and the deviation between the reconstructed force and the preprocessed measured milling force is used to construct the force signal reconstruction loss. Then, the wear mapping loss corresponding to the difference features and the force signal reconstruction loss are weighted and summed to obtain the total loss, which is used to update the model parameters. After repeated iterative training, the model outputs the tool wear state.

[0041] The deep learning model introduces an incremental learning mechanism. During real-time machining on a machine tool, when new cutting parameters are detected under a new working condition, the old network branches for the existing working condition are fixed, and a new network branch is created for learning the new working condition. Based on the spindle speed, feed rate, width at cut, depth of cut, and material properties under the new working condition, and based on the milling force physical mechanism calculation model embedded in the digital twin, a calculated milling force time-series signal considering tool wear and the corresponding simulated tool wear label value are generated under the new working condition. A corresponding new tool theory calculated milling force time-series signal is also generated based on a new tool theory model. Using the calculated milling force time-series signal considering tool wear under the new working condition, the new tool theory calculated milling force time-series signal, and the simulated tool wear label value, a pseudo-sample dataset without actual measured values ​​is constructed. This pseudo-sample dataset is input into the new network branch to calculate the milling force considering tool wear. The cutting force timing signal and the corresponding theoretically calculated milling force timing signal of the new tool are used as inputs. The simulated tool wear label value is used as the supervision signal. The parameters of the new network branch are updated based on the wear mapping loss to complete real-time incremental training. After obtaining the real milling force timing signal and the corresponding true value of tool wear under the new working condition, the new network branch is trained a second time based on the total loss of the weighted sum of the wear mapping loss and the force signal reconstruction loss to obtain the tool wear state monitoring model under the new working condition, which is used to predict the tool wear value under the new working condition. Finally, the wear state is integrated into the digital twin layer. The system controls the machine tool to perform tool changing operation through G-code according to the wear state, forming a closed-loop machining monitoring system of intelligent perception of working condition, intelligent identification of wear, and intelligent decision-making for tool changing.

[0042] Example 3: Reference Figure 4 This embodiment provides a specific flowchart of a digital twin-based tool wear condition monitoring method, which includes: First, system setup and data preparation are carried out: a digital twin tool wear monitoring system is built, and a synchronous tracking mapping between the digital twin and the physical machine tool is established through a communication protocol; during the machining process, a force gauge collects the milling force timing signal in real time, and simultaneously acquires the machine tool's spindle speed and feed rate; the edge computing end performs low-pass filtering, wavelet threshold noise reduction, and alignment preprocessing on the collected milling force signal to obtain the preprocessed measured milling force timing signal; such as Figure 3 As shown, the front end of the digital twin dynamically calculates the cutting width and depth of cut based on the workpiece and tool envelope model and their real-time relative positions. It receives material properties through the interactive interface and stores the spindle speed, feed rate, cutting width, cutting depth, and material properties in the buffer according to the time sequence. The back end of the twin embeds a milling force physical mechanism calculation model (including a brand-new tool theory model) and reads the above parameters from the buffer to generate a new tool theory calculation milling force time sequence signal in real time.

[0043] Next, model training and incremental learning are performed: A comparative learning architecture is constructed using the milling force time-series signal calculated by the new tool theory and the preprocessed measured milling force signal as common inputs. The difference vector between the two in the multi-dimensional feature space is extracted, and the predicted tool wear value is obtained based on this vector. During training, the predicted tool wear value, along with the corresponding spindle speed, feed rate, width at cut, depth of cut, and material properties, are input into the milling force physical mechanism calculation model. The milling force signal is reconstructed by reverse calculation, and the deviation between the reconstructed force and the preprocessed measured milling force is used to construct the force signal reconstruction loss. The wear mapping loss corresponding to the difference features and the force signal reconstruction loss are weighted and summed as the total loss, and the model parameters are updated by backpropagation. After repeated iterative training to complete model optimization and iteration, model evaluation is performed. If the evaluation result does not meet the requirements, the model returns to the optimization and update stage for further iteration until the evaluation meets the standards, ultimately determining a tool anomaly prediction model that meets the accuracy requirements. When new cutting parameters are detected, the old network branches for existing working conditions are fixed, and new network branches are created for learning the new working conditions. Based on the spindle speed, feed rate, width of cut, depth of cut, and material properties under the new working conditions, and based on the milling force physical mechanism calculation model embedded in the digital twin, a calculated milling force time series signal considering tool wear and the corresponding simulated tool wear label value are generated under the new working conditions. A corresponding new tool theory calculated milling force time series signal is also generated based on a new tool theory model. Using the calculated milling force time series signal considering tool wear under the new working conditions, the new tool theory calculated milling force time series signal, and the simulated tool wear label value, a pseudo-sample dataset without actual measured values ​​is constructed. The new network branch is fed into a set of inputs, taking the calculated milling force time-series signal considering tool wear and the corresponding theoretically calculated milling force time-series signal of the new tool as inputs, and the simulated tool wear label value as the supervision signal. The parameters of the new network branch are updated based on the wear mapping loss to complete real-time incremental training. After obtaining the real milling force time-series signal and the corresponding true value of tool wear under the new working condition, the new network branch is trained a second time using the real milling force time-series signal under the new working condition, the corresponding theoretically calculated milling force time-series signal of the new tool, and the true value of tool wear. The total loss is calculated by weighting the wear mapping loss and the force signal reconstruction loss to obtain the tool wear state monitoring model under the new working condition, which is used to predict the tool wear value under the new working condition. Finally, online monitoring and decision-making are performed: During the machining process, the digital twin continuously generates a new tool theoretical calculation milling force time-series signal in real time, which is input together with the preprocessed measured milling force time-series signal into the trained deep learning model. The system outputs the tool wear status and wear trend curve in real time and displays it visually on the interactive interface. The system determines whether the preset replacement threshold has been reached based on the wear value. Once the threshold is reached, the digital twin platform automatically generates G-code instructions to control the machine tool to perform a tool change operation. If the threshold is not reached, machining continues and monitoring continues, forming a closed-loop machining monitoring system of "intelligent perception of working conditions, intelligent identification of wear, and intelligent decision-making for tool change".

[0044] Example 4: Reference Figure 5 This embodiment provides a cloud database storage solution, specifically including: In the system constructed using this method, all data generated by the physical layer, edge computing layer, digital twin layer, and deep learning model layer are uploaded to the cloud database with time tags. This includes, but is not limited to, spindle speed, feed rate, width at cut, depth of cut, material properties, unprocessed milling force time-series signals, preprocessed measured milling force time-series signals, theoretically calculated milling force time-series signals for new tools, predicted wear values, model update records, and tool change information. These data collectively constitute a complete record of the machine tool status, tool status, and cutting status at each time point, supporting machine tool status tracking, tool status tracking, and cutting status tracking, achieving full traceability of the machining process. Simultaneously, the historical data accumulated in the cloud database provides data support for incremental learning using real samples of new working conditions, enabling continuous updating and optimization of the tool wear monitoring model.

[0045] It should be noted that the method described in this embodiment is not only applicable to the prediction of tool wear conditions, but can also be extended to the monitoring of other tool conditions.

[0046] It is understood that the same or similar parts in the above embodiments can be referred to each other, and the contents not described in detail in some embodiments can be referred to the relevant descriptions in other embodiments.

[0047] The sequence numbers of the steps in the above embodiments do not represent the order of execution. The execution order of each step should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of this application.

[0048] Any process or method described in the flowchart or otherwise herein should be understood as comprising one or more executable instructions for implementing a particular logical function or step. The scope of this application also includes other implementations in which functions are performed substantially simultaneously or in reverse order, depending on the functionality involved, without being performed in the order shown or discussed, as will be understood by those skilled in the art.

[0049] Although embodiments of this application have been shown and described above, it should be understood that the above embodiments are merely illustrative and should not be construed as limiting this application. Any changes, modifications, substitutions, and variations made by those skilled in the art to the above embodiments within the scope of this application should fall within the protection scope of this application.

Claims

1. A method for monitoring tool wear status based on digital twins, characterized in that, The method includes: The spindle speed, feed rate and milling force timing signals during the machine tool machining process are acquired, and the milling force timing signals are preprocessed to obtain the measured milling force timing signals; Construct a digital twin that moves with the machine tool, determine the cutting width and depth of cut through the envelope model of the tool and the workpiece and their relative positional relationship, and obtain the workpiece material properties; Based on the spindle speed, feed rate, width of cut, depth of cut, and material properties, a new tool theory is generated to calculate the milling force timing signal. Using the measured milling force time series signal and the milling force time series signal calculated by the new tool theory as input, the difference feature vectors between the two in the multi-dimensional feature space are extracted through a contrastive learning network, and the tool wear value is obtained based on the mapping of the difference feature vectors. Introducing pseudo-sample incremental learning mechanisms into contrastive learning networks includes: During real-time machining on a machine tool, when new cutting parameters for a new working condition are detected, the old network branches for the existing working condition are fixed, and a new network branch is created for learning the new working condition. Based on the spindle speed, feed rate, cutting width, cutting depth and material properties under the new working conditions, and based on the milling force physical mechanism calculation model embedded in the digital twin, the calculated milling force timing signal considering tool wear under the new working conditions and the corresponding simulated tool wear label value are generated. Based on the new tool theory model, the corresponding new tool theory calculated milling force timing signal is generated. Using the milling force timing signal calculated under the new working condition considering tool wear, the milling force timing signal calculated theoretically under the new tool, and the simulated tool wear label value, a pseudo sample dataset without actual measured true value labeling is constructed. The pseudo-sample dataset is input into the new network branch. The time series signal of the calculated milling force considering tool wear and the corresponding time series signal of the calculated milling force of the new tool are used as inputs. The simulated tool wear label value is used as the supervision signal to update the parameters of the new network branch and complete the real-time incremental training. After obtaining the real milling force timing signal and the corresponding true value of tool wear under the new working condition, the milling force timing signal and the true value of tool wear are calculated using the real milling force timing signal and the corresponding new tool. The new network branch is then trained a second time to obtain the tool wear state monitoring model under the new working condition, which is used to predict the tool wear value under the new working condition. When the tool wear value reaches the preset replacement threshold, the machine tool is controlled to perform a tool change operation.

2. The tool wear condition monitoring method based on digital twin according to claim 1, characterized in that, The milling force timing signal is preprocessed to obtain the measured milling force timing signal, including: The milling force timing signal is first low-pass filtered to remove high-frequency noise, then subjected to wavelet threshold noise reduction, and finally aligned according to the machining time step to obtain the measured milling force timing signal.

3. The tool wear condition monitoring method based on digital twin according to claim 1, characterized in that, The cutting width and depth of cut are determined by the envelope model of the workpiece and the tool and their relative positional relationship, including: Establish communication between the digital twin and the machine tool to realize virtual motion mapping between the digital twin and the machine tool; Based on the relative positional relationship between the tool and the workpiece, an envelope model of the workpiece and the tool is constructed. The volume of the tool envelope is obtained from the envelope model, and the cutting width and cutting depth are calculated.

4. The tool wear condition monitoring method based on digital twin according to claim 1, characterized in that, The training process of a contrastive learning network includes: The predicted tool wear value and the corresponding spindle speed, feed rate, cutting width, cutting depth and material properties are input into the milling force physical mechanism calculation model, and the reconstructed milling force time sequence signal is calculated in reverse. Based on the deviation between the reconstructed milling force timing signal and the measured milling force timing signal, a force signal reconstruction loss is constructed; The wear mapping loss corresponding to the difference feature vector and the force signal reconstruction loss are weighted and summed to obtain the total loss, which is then used to update the parameters of the contrastive learning network through backpropagation. During incremental training, the parameters of the new network branch are updated based on the wear-out mapping loss; During the second training, the parameters of the new network branch are updated based on the total loss obtained by weighted summation of wear mapping loss and force signal reconstruction loss.

5. The tool wear condition monitoring method based on digital twin according to claim 1, characterized in that, The spindle speed, feed rate, cutting width, cutting depth, and material properties are stored in the buffer according to time sequence. The milling force physical mechanism calculation model embedded in the digital twin and the new tool theory model both read the process parameters under the corresponding working conditions from the buffer to generate the corresponding milling force timing signal.

6. The tool wear condition monitoring method based on digital twin according to claim 5, characterized in that, The method also includes: The process parameters, measured milling force timing signals, tool wear values, model update records, and tool change records during the machine tool machining process are uploaded to the cloud database with time as the tag to support the traceability of the machining process and the continuous updating of the model.

7. The tool wear condition monitoring method based on digital twin according to claim 1, characterized in that, The timing signal of milling force is acquired by a force measuring instrument.

8. The tool wear condition monitoring method based on digital twin according to claim 1, characterized in that, The system receives the workpiece material properties specified by the user through a human-computer interaction interface.

9. The tool wear condition monitoring method based on digital twin according to claim 1, characterized in that, The tool wear value and wear trend curve are visualized through a human-computer interaction interface.

10. A tool wear condition monitoring system based on digital twins for implementing the tool wear condition monitoring method based on digital twins as described in any one of claims 1 to 9, characterized in that, The system includes: The data acquisition and preprocessing module is used to acquire the spindle speed, feed rate and milling force timing signals during the machine tool processing, and to preprocess the milling force timing signals to obtain the measured milling force timing signals. The digital twin module is used to construct a digital twin that moves with the machine tool. It determines the cutting width and depth of cut through the envelope model of the tool and the workpiece and their relative positional relationship, and obtains the workpiece material properties. Based on the new tool theory model, it generates a new tool theory to calculate the milling force timing signal according to the spindle speed, feed rate, cutting width, cutting depth and material properties. The deep learning module is used to take the measured milling force time series signal and the milling force time series signal calculated by the new tool theory as input, extract the difference feature vectors between the two in the multi-dimensional feature space through a contrastive learning network, and obtain the tool wear value based on the difference feature vector mapping. A pseudo-sample incremental learning module is used to introduce a pseudo-sample incremental learning mechanism into the contrastive learning network. This includes: during real-time machining on a machine tool, when new cutting parameters for a new working condition are detected, fixing the existing network branches for that working condition and creating new network branches for learning the new working condition; based on the spindle speed, feed rate, width at cut, depth of cut, and material properties under the new working condition, and using the milling force physical mechanism calculation model embedded in the digital twin, generating a calculated milling force time-series signal considering tool wear under the new working condition and the corresponding simulated tool wear label value, and generating a corresponding new tool theory calculated milling force time-series signal based on a completely new tool theory model; using the calculated milling force time-series signal considering tool wear under the new working condition and the new tool theory calculated milling force time-series signal... A pseudo-sample dataset without actual ground truth labels is constructed using cutting force timing signals and simulated tool wear label values. This pseudo-sample dataset is then input into a new network branch. The calculated milling force timing signal considering tool wear and the corresponding theoretically calculated milling force timing signal for the new tool are used as inputs, and the simulated tool wear label value is used as a supervision signal to update the parameters of the new network branch, completing real-time incremental training. Once the actual milling force timing signal and the corresponding ground truth value for the new working condition are obtained, the new network branch is trained a second time using these signals, resulting in a tool wear state monitoring model for the new working condition, used to predict tool wear values ​​under the new working condition. The tool change control module is used to control the machine tool to perform a tool change operation when the tool wear value reaches the preset replacement threshold.

Images