Tool state prediction method and system based on digital twinning

Abstract

The application relates to the technical field of equipment state monitoring and digital twinning, and specifically discloses a tool state prediction method and system based on digital twinning. According to the application, whether the current tool working condition data needs to be marked and stored in a classified manner is determined according to the applicable probability of an expert network to a sensor feature vector and a machining condition vector, a gating entropy value and a wear state parameter. Secondly, model training is triggered according to the data caching condition in the classified cache area, and a network model to be updated is determined according to the data caching condition, so that the model does not need to be completely retrained, the retraining efficiency is improved, the network model to be updated is trained according to the cached data in the classified cache area, the processing capacity of the model for various types of working condition data is improved, and the tool state prediction accuracy is improved.

Description

Technical Field

[0001] This application relates to the fields of equipment condition monitoring and digital twin technology, and in particular to a tool condition prediction method and system based on digital twin. Background Technology

[0002] In automated manufacturing, accurate estimation of tool wear is crucial for ensuring machining quality, preventing catastrophic tool failures, and optimizing tool life. Tools operating under harsh conditions such as high-speed cutting and intermittent cutting are highly susceptible to degradation phenomena such as flank wear, crater wear, and chipping. Deviations in tool health can lead to inaccurate machining geometry, poor surface finish, and potentially damage to expensive workpieces, even resulting in safety risks and economic losses.

[0003] Existing tool wear monitoring methods can be broadly categorized into direct measurement and indirect inference. Direct measurement relies on offline methods such as microscopy or visual measurement, offering high accuracy but requiring machine downtime, making it unsuitable for online monitoring. Indirect inference methods infer wear conditions by collecting sensor data on vibration, force, torque, and current, offering advantages such as non-invasiveness and online applicability. Most indirect inference methods employ physics-based models and data-driven approaches. However, neither physics-based models nor data-driven approaches can cover a wide range of machining conditions, leading to a sharp drop in prediction accuracy during condition switching.

[0004] Therefore, improving the accuracy of tool condition prediction has become an urgent problem to be solved. Summary of the Invention

[0005] This application provides a tool condition prediction method and system based on digital twins to improve the accuracy of tool condition prediction.

[0006] In a first aspect, this application provides a tool state prediction method based on digital twins, applied to a digital twin prediction system. The digital twin prediction system includes a physical entity layer, a connection and data layer, a twin model layer, a service and application layer, and a synchronization and correction layer. The method includes:

[0007] The connection and data layer acquires first tool condition data from the physical entity layer through a data acquisition interface, and processes the first tool condition data to obtain sensor feature vectors and machining condition vectors.

[0008] The twin model layer processes the sensor feature vector and the processing condition vector to obtain the applicability probability, gate entropy value and first wear state vector of each expert network.

[0009] The service and application layer performs state estimation based on the first wear state vector to obtain wear state parameters;

[0010] When at least one of the applicable probability, the gate entropy value, and the wear state parameter meets a preset condition, the synchronization and correction layer tags the tool condition data and stores the tool condition data into a classification cache according to the tag.

[0011] The twin model layer determines the network model to be updated based on the data caching status of the classification cache, and trains the network model to be updated based on the cached data in the classification cache to obtain the latest twin model layer.

[0012] The synchronization and correction layer synchronizes the latest twin model layer and the physical entity layer so that the connection and data layer can obtain the second tool condition data from the physical entity layer through the data acquisition interface, and process the second tool condition data based on the latest twin model layer to obtain the second wear state vector.

[0013] This application discloses a tool condition prediction method and system based on digital twins. The method determines whether current tool condition data needs to be labeled and categorized based on the applicability probability of sensor feature vectors and machining condition vectors, the gate entropy value, and wear state parameters obtained from an expert network. Secondly, model training is triggered based on the data caching status in the categorization cache, and the network model to be updated is determined based on the data caching status, eliminating the need for complete model retraining and improving retraining efficiency. Training the network model to be updated based on the cached data in the categorization cache improves the model's ability to process various types of working condition data, thereby improving the accuracy of tool condition prediction. Attached Figure Description

[0014] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 This is a schematic diagram of the structure of a digital twin prediction system provided in an embodiment of this application;

[0016] Figure 2 This is a schematic flowchart of a first embodiment of a tool state prediction method based on digital twin provided by the embodiments of this application;

[0017] Figure 3 This is a schematic diagram of the structure of a hybrid expert module and gating network in the twin model layer of a digital twin prediction system provided in an embodiment of this application;

[0018] Figure 4This is a schematic diagram of the physical information module in the twin model layer of a digital twin prediction system provided in an embodiment of this application;

[0019] Figure 5 This is a schematic diagram of the operation flow of a tool state prediction method based on digital twin provided in an embodiment of this application;

[0020] Figure 6 This is a schematic flowchart of a second embodiment of a tool state prediction method based on digital twin provided in this application;

[0021] Figure 7 This is a schematic flowchart of a third embodiment of a tool state prediction method based on digital twin provided in this application;

[0022] Figure 8 This is a schematic flowchart of a fourth embodiment of a tool state prediction method based on digital twin provided in this application;

[0023] Figure 9 This is a schematic block diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0024] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0025] The flowchart shown in the attached diagram is for illustrative purposes only and does not necessarily include all content and operations / steps, nor does it necessarily have to be performed in the described order. For example, some operations / steps can be broken down, combined, or partially merged, so the actual execution order may change depending on the actual situation.

[0026] It should be understood that the terminology used in this application specification is for the purpose of describing particular embodiments only and is not intended to limit the application. As used in this application specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0027] It should also be further understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0028] This application provides a tool condition prediction method and system based on digital twins. The digital twin-based tool condition prediction method can be applied to a digital twin prediction system. It determines whether current tool condition data needs to be labeled and categorized based on the applicability probability of sensor feature vectors and machining condition vectors, the gate entropy value, and wear state parameters obtained by an expert network. Secondly, it triggers model training based on the data caching status in the categorization cache, and determines the network model to be updated based on the data caching status, eliminating the need for complete model retraining and improving retraining efficiency. Training the network model to be updated based on the cached data in the categorization cache improves the model's ability to process various types of working condition data, thereby improving the accuracy of tool condition prediction. For example, Figure 1 As shown, Figure 1 This is a schematic diagram of the structure of a digital twin prediction system provided in an embodiment of this application. The digital twin prediction system includes:

[0029] The physical entity layer includes processing equipment such as CNC machine tools, machining centers or industrial robots, as well as spindle / servo systems, cutting tools, workpieces and multi-source sensors (such as vibration sensors, force / torque sensors, spindle current sensors, temperature sensors, etc.).

[0030] The connection and data layer is equipped with a data acquisition interface, which is used to acquire multi-source sensor data and process parameters such as vibration, force, torque, and current from the physical entity layer through the data acquisition interface, and perform data cleaning, feature extraction and time alignment to obtain sensor feature vectors and processing condition vectors.

[0031] The twin model layer is configured with a gating network, a hybrid expert module, and a physical information module. It is used to route the received sensor feature vectors and processing condition vectors to the optimal expert network via the gating network, process the sensor feature vectors and processing condition vectors through the optimal expert network to obtain twin state feature vectors, and process the twin state feature vectors, sensor feature vectors, and processing condition vectors through the physical information module to obtain wear state vectors. It is also used to determine the network model to be updated based on the data caching status of the classification cache, and train the network model to be updated based on the cached data in the classification cache to obtain the latest twin model layer.

[0032] The service and application layer is used to estimate the wear state based on the wear state vector output by the twin model layer and generate tool adjustment suggestions;

[0033] The synchronization and correction layer is used to maintain consistency between the virtual and the real world. It manages the synchronization cycle through event-triggered or time-triggered events, and combines prediction residual detection, gate entropy value, and responsibility probability threshold to determine whether there is consistency degradation. When consistency degradation is detected, it triggers incremental evolution update of the Siamese model layer.

[0034] The control feedback interface is used to transmit tool adjustment suggestions output from the service and application layers to the physical entity layer, determine the optimal suggestion execution mode, and execute the tool adjustment suggestions.

[0035] The following detailed description of some embodiments of this application is provided in conjunction with the accompanying drawings. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0036] Please see Figure 2 , Figure 2 This is a first schematic flowchart of a tool state prediction method based on digital twin provided in an embodiment of this application.

[0037] like Figure 2 As shown, the tool state prediction method based on digital twins specifically includes steps S101 to S106.

[0038] S101. The connection and data layer collects the first tool condition data from the physical entity layer through the data acquisition interface, and processes the first tool condition data to obtain the sensor feature vector and the machining condition vector.

[0039] In one embodiment, the physical entity layer includes processing equipment such as CNC machine tools, machining centers or industrial robots, as well as spindle / servo systems, cutting tools, workpieces and multi-source sensors (vibration sensors, force / torque sensors, spindle current sensors, temperature sensors, etc.) to generate tool condition data.

[0040] In one embodiment, the connection and data layer communicates with the PLC (Programmable Logic Controller) or CNC (Computer Numerical Control) controller via the OPC-UA (Open Platform Communications Unified Architecture) or MTConnect (Machine Tool Connectivity) protocol to acquire data from the physical entity layer, perform data cleaning, feature extraction, and time alignment to obtain sensor feature vectors and machining condition vectors, and then transmit the sensor feature vectors and machining condition vectors to the twin model layer.

[0041] In one embodiment, data is collected from the physical entity layer at a preset synchronization period, and state estimation is performed using a twin model. The synchronization period can be triggered in the following ways:

[0042] Event-triggered synchronization: Synchronization is triggered after each machining operation (such as drilling or milling pass) is completed;

[0043] Time-triggered synchronization: Synchronization is triggered periodically at preset time intervals ΔT;

[0044] Hybrid triggering synchronization: event triggering takes precedence, while time triggering serves as a fallback mechanism.

[0045] S102. The twin model layer processes the sensor feature vector and the processing condition vector to obtain the applicability probability, gate entropy value and first wear state vector of each expert network.

[0046] In a specific embodiment, for each data point passed in during a synchronization cycle Gating network Calculate the probability of expert applicability Hybrid Expert Module Calculate the twin state eigenvector h; frozen twin state prediction model use Generate wear state vector This includes the wear estimation state w and the wear rate. and uncertainty .

[0047] Where t represents the number of discrete operations performed (e.g., number of drillings, number of milling passes, etc.), used to quantify the cumulative usage of the tool.

[0048] This represents in-situ sensor feature vectors extracted directly from the machining process. These features are crucial for capturing the real-time state of the operation and may include, but are not limited to, signals from force gauges (such as cutting force, thrust), accelerometers (such as vibration features indicating chatter), acoustic emission sensors, spindle motor current and temperature sensors. Advanced signal processing and feature engineering techniques can be applied to the raw sensor data to extract these information-rich features.

[0049] The vector representing the current machining parameters and context conditions (i.e., the machining condition vector) includes controllable parameters such as spindle speed (Ω), feed rate (f), depth of cut, and tool geometry, as well as less influential but controllable factors such as workpiece material properties, coolant application, and ambient temperature.

[0050] Furthermore, the twin model layer includes a gating network, a hybrid expert module, and a physical information module. The hybrid expert module includes at least one expert network, and the physical information module includes a twin state prediction model.

[0051] In one embodiment, this application employs a novel Physical Information Evolutionary Hybrid Expert (PI-EMoE) architecture as the core model for digital twins. This architecture is not a single model, but rather consists of multiple specialized neural networks working together.

[0052] The PI-EMoE architecture comprises two main interactive neural networks. For example... Figure 3 As shown, the first interactive neural network is a gated network. and hybrid expert modules It is responsible for adaptively mapping the raw input data (sensor features and processing parameters) to a consistent and information-rich twin state feature representation for specific operating conditions, aiming to handle the variability of data distribution caused by different operating conditions.

[0053] The hybrid expert module is a network of K independent experts. A dynamically expandable set composed of elements is represented as Each expert Preferably, a neural network (such as an MLP, Multi-Layer Perceptron) is used, with the following parameters: Each expert Each expert is trained to handle specific, discrete operating conditions or data domains (e.g., drilling a specific alloy, using a certain tool coating, or operating within defined speed and feed ranges). Each expert corresponds to a specific physical machining condition domain, achieving a mapping to physical entities. Gated network. Used as an intelligent router or scheduler, it also supports consistency monitoring. The default is MLP, and its parameters are... The gating network receives the same input as the expert network. and Its output is a probability vector. Each of them and (Usually implemented using the softmax activation function of the output layer), probability Representative gating network for experts The entropy value output by the gating network is used to evaluate the applicability (i.e., applicability probability) of the current input data. and maximum applicability probability It is also used for consistency degradation detection.

[0054] In one embodiment, a target expert network is determined based on the applicability probability output by the gating network, and the target expert network is used to analyze the sensor feature vector. and processing condition vector The process is performed to obtain the twin state feature vector h.

[0055] In one embodiment, such as Figure 4 As shown, the second interactive neural network is the physical information module (twin state prediction model). This module constitutes the stable, physics-based core of the twin, responsible for... The module obtains the processed twin state feature vector h and generates the final twin state output while obeying the physical laws.

[0056] Structurally, Preferably, it has learnable parameters The MLP, with its architecture designed to be sufficiently expressive of wear functions, Modeling is performed. Regarding the input, Number of operations received (Physical time axis), sensor feature vector (Direct observation of physical state), processing condition vector (processing parameters) and twin state feature vectors ,Include Make It can interpret data from different operating conditions in a consistent manner. On the output side, this module estimates the output tool wear. and its estimated rate of change relative to the number of operations :

[0057]

[0058] direct estimation This is crucial for enforcing PDE constraints and analyzing wear trends.

[0059] In one embodiment, The training process includes physical information residuals. This residual loss... The degree to which the module output violates physical wear dynamics is quantified and defined as follows:

[0060]

[0061] in, and yes The direct output. This represents a set of known physical constants or parameters of established empirical or semi-empirical wear models (such as parameters in the Usui (physically based analytical model of tool wear) wear model or parameters in models that include chatter severity). If The specific form requires the spatial derivative of wear (e.g.) Then, automatic differentiation (AD) techniques can be used to analyze neural networks. By performing calculations, AD technology can accurately calculate these derivatives with respect to the network input, ensuring that the twin's output is consistent with the laws of physics.

[0062] In one embodiment, a twin state prediction model is trained using a nominal operating condition dataset. ,like Figure 4 As shown, the training of the twin state prediction model employs a method that incorporates data fidelity loss. PDE loss and monotonicity loss The composite loss function is used to optimize the parameters of the physical information module through gradient descent training. The composite loss function is expressed as:

[0063]

[0064] in, This indicates the loss of data fidelity, ensuring that the estimated wear matches the true value; This represents the PDE loss, penalizing deviations from the physical wear model to ensure that the twin output is consistent with physical laws. This indicates monotonic loss, ensuring that wear increases monotonically over time; , , These are the corresponding weight hyperparameters, which need to be adjusted through cross-validation or domain-specific knowledge.

[0065] Furthermore, before the twin model layer processes the sensor feature vector and the processing condition vector to obtain the applicability probability, gate entropy value, and first wear state vector of each expert network, it further includes: collecting a nominal working condition dataset; iteratively training the pre-trained twin state prediction model based on a preset composite loss function and the nominal working condition dataset to obtain the gradient norm of each loss term in the preset composite loss function; obtaining the current training stage, adjusting the weight coefficients of each loss term according to the current training stage, each gradient norm, and a preset weight adjustment formula to obtain the composite loss function for the next iteration; iteratively training according to the composite loss function for the next iteration until a preset condition is reached, then stopping the iterative training to obtain the twin state prediction model.

[0066] In one embodiment, such as Figure 5As shown, before the system runs, digital twin initialization is performed to establish a physical-twin mapping relationship. Specifically, multi-source sensor signals from the processing process are collected through the connection and data layers respectively. Features are extracted from signals such as vibration, force, torque, and current to form samples. Corresponding wear labels are obtained through offline measurement methods to obtain a nominal operating condition dataset. This establishes a physical-twin mapping relationship, allowing for the collection of a comprehensive nominal operating condition dataset under a single, well-defined nominal operating condition, thus synchronizing data from the physical entity layer to the connection and data layers. Nominal operating condition dataset. Includes: sensor feature vectors Synchronous time series and corresponding processing parameters Number of operations (e.g., the number of drill holes or milling passes performed) and the corresponding true value of tool wear obtained by a reliable method. (Obtained through direct or indirect measurement methods, such as optical examination using a microscope). For example... Figure 5 As shown, after obtaining the nominal working condition dataset, the nominal working condition dataset is used as the sample dataset to train the physical information module.

[0067] In one embodiment, tool wear in machining operations (such as drilling, milling, and turning) is a complex, progressive degradation phenomenon influenced by a variety of interacting factors, including cutting forces, tool-chip interface temperature, tool and workpiece material properties, cutting speed, feed rate, depth of cut, and the presence of dynamic instabilities such as chatter. Traditional methods use empirical models (such as Taylor's tool life equation) or simplified analytical models, but these models often lack the fidelity to capture the true details of wear under different conditions.

[0068] In this embodiment, the tool wear estimation state (denoted by w) is modeled as a function dependent on multiple variables, rather than just a single variable (such as cutting time). The tool wear estimation state w can be expressed as: , where f is a complex, possibly nonlinear, unknown function, and the data-driven component of the twin model in this application aims to approximate this function.

[0069] In one embodiment, the dynamics of tool wear evolution are governed by fundamental physical principles. Wear rate, i.e., the rate of change of wear relative to the number of operations. The wear rate (in partial derivative form) can be described by a governing physical relationship, which can be expressed as a partial differential equation (PDE). This PDE summarizes how wear evolves according to the current state and influencing factors, and the wear rate can be expressed as:

[0070]

[0071] in, It is a function representing the physical wear rate equation, which can depend on , , Current wear status and its spatial gradient relative to sensor features ( ) or spatial gradient of operating conditions ( ), and other related higher-order derivatives. It represents a set of known physical constants or parameters of established empirical or semi-empirical wear models (such as parameters in the Usui wear model or parameters in a model that includes flutter severity).

[0072] In one embodiment, in order to comply with physical constraints At the same time, it effectively estimates the wear state w, especially under dynamically changing operating conditions. This application adopts the PI-EMoE architecture as the core model of digital twin.

[0073] The entire PI-EMoE twin model minimizes a composite loss function. During training, this loss function aims to balance data fidelity with adherence to physical principles and desired behavior. The composite loss function is expressed as:

[0074]

[0075] In one embodiment, data fidelity loss Ensure that the estimated wear w matches the true wear value. The close matching, data fidelity loss function is expressed as:

[0076]

[0077] in, It is the number of data points with truth labels.

[0078] In one embodiment, PDE loss The penalty is for deviations from the physical wear model, thereby driving the system to learn physically reasonable wear dynamics. The PED loss function is expressed as:

[0079]

[0080] in, This is the number of configuration points used to evaluate the PDE residuals (these points are not necessarily required). Truth value).

[0081] In one embodiment, during the wear process, tool wear is a monotonically non-decreasing function of the number of operations t. This prior physical knowledge can be added as a soft constraint, forming a monotonic loss. For example, for continuous measurements under the same operating conditions. and (Where j represents the sequence number of the current measurement sample, and j+1 represents the next measurement sample,) (This indicates that the number of operations for the later sample is greater than that for the earlier sample).

[0082]

[0083] in, Penalty for estimated wear reduction It represents the number of time series paired samples (j and j+1).

[0084] In summary, the composite loss function It is a weighted sum of data fidelity loss, PED loss, and monotonicity loss:

[0085]

[0086] in, , and It is a hyperparameter that controls the relative importance of each loss term.

[0087] In one embodiment, the weights of the composite loss function during the training phase... , and It can adaptively adjust the weights. Specifically, during the initialization phase (e.g., the first 200 iterations), the weight coefficients for each loss are directly set, such as... , , Prioritize ensuring good data fit.

[0088] In the intermediate stage (e.g., 200-500 iterations): the weights are dynamically adjusted based on the gradient norm of each loss term. The larger the gradient norm, the larger the corresponding weight. The weight adjustment formula is as follows:

[0089]

[0090] in, Let represent the gradient norm of the j-th loss term.

[0091] Convergence phase (after 500 rounds), fixed To ensure the dominance of physical constraints, and The remaining weights are allocated in a 1:1 ratio.

[0092] In another embodiment, the adaptive adjustment of the weights in the composite loss function can also be achieved through adaptive adjustment based on operating conditions. Specifically, an operating condition complexity factor is introduced. (Based on the variance of processing parameters and the complexity of sensor signals), when >When the first preset complexity threshold (a default value set according to actual needs, such as 1.2) is reached, Adjust to the first preset weight value, such as 0.6, to enhance the stability support of physical constraints for complex working conditions; when When the second preset complexity threshold is reached, Adjust the weight to the second preset value, such as 0.7, to improve the prediction accuracy under simple working conditions.

[0093] In one embodiment, the composite loss function adaptively adjusts the weights while employing a weight constraint mechanism: setting upper and lower limits for the weights, for example... , , This avoids model bias caused by excessive weighting of a single element and ensures a balance between data fitting, physical constraints, and monotonicity.

[0094] In another embodiment, during the training process, a larger learning rate (e.g., 0.001) is first used to train the model. After the preset prediction accuracy is reached, a smaller learning rate (e.g., 0.0001) is used to fine-tune the network, which can make the network converge faster.

[0095] In one embodiment, once the twin state prediction model... To achieve the expected performance and capture the expected physical behavior (such as wear increasing monotonically over time), its parameters The freezing indicates that the physical information module is frozen, and thereafter, as a stable and immutable core component of the twin, it ensures that physical constraints are always enforced.

[0096] In one embodiment, such as Figure 5 As shown, after the physical information module is trained and frozen, the initial expert and gating network are trained to initialize the Siamese model layer. Specifically, the initial hybrid expert module is trained on nominal operating condition data. (Initially consisting of only one network of experts) and gated networks The goal is to enable experts Learning will input under nominal operating conditions Mapping to a twin state feature vector When the twin state feature vector is transferred to the frozen state In this way, wear and tear can be accurately predicted. Gated network It is trained to reliably route all nominal data to the optimal initial expert. The loss function at this stage is mainly... Because the laws of physics have been changed by coding.

[0097] S103. The service and application layer performs state estimation based on the first wear state vector to obtain wear state parameters;

[0098] In one embodiment, the service and application layer receives a first wear state vector transmitted from the twin model layer, performs twin services (including state estimation and suggestion generation) and virtual-real consistency checks based on the wear state vector, and outputs wear state parameters and tool adjustment suggestions.

[0099] In a specific embodiment, the service and application layers are based on wear state vectors. Provides twin services, including state estimation services (outputting wear state parameters, including current wear level). Wear rate and uncertainty ) and multi-model fusion decision recommendation services.

[0100] Furthermore, the digital twin prediction system also includes a control feedback interface. After the service and application layer estimates the state based on the first wear state vector and obtains the wear state parameters, it further includes: predicting the wear trend based on historical wear data and working condition change sequences using a time-series attention model to obtain wear prediction results; analyzing the wear state parameters and the wear prediction results based on a preset suggestion generation model to generate tool adjustment suggestions; and the control feedback interface feeds back the tool adjustment suggestions to the physical entity layer.

[0101] In one embodiment, the decision recommendation service is based on multi-model fusion optimization, employing a hierarchical AI model fusion architecture, including:

[0102] Basic decision-making layer: Retain the wear status parameters output by the original wear estimation service as the basis for tool changing and alarms;

[0103] Accurate prediction layer: Introducing the Temporal Attention Mechanism-Long Short-Term Memory (TA-LSTM) model, which predicts the wear trend of the future preset processing cycle based on historical wear data and working condition change sequences, and outputs the confidence interval of wear increment, thus solving the problem of insufficient short-term prediction accuracy of a single model.

[0104] Intelligent Optimization Layer: This layer integrates a reinforcement learning (RL) model to train a suggestion generation model. Using "achieving machining quality targets + maximizing tool life + minimizing energy consumption" as the joint reward function, the agent is trained to dynamically adjust process parameters. The reward function is defined as follows:

[0105]

[0106] in, Rewards will be given for meeting the dimensional accuracy standards. As a bonus for the remaining lifespan of the cutting tool, As a penalty for energy consumption, , , The weights are dynamic (calculated using the entropy weight method based on real-time operating conditions).

[0107] Anomaly Diagnosis Layer: Integrates an isolated forest (IF) model to perform real-time anomaly detection on sensor data, identify sudden faults such as flutter and material inclusions, trigger emergency alarms or process adjustment suggestions, and make up for the shortcomings of the original solution in responding to sudden anomalies.

[0108] In one embodiment, historical wear data includes the wear amount w(t) and wear rate over the entire tool lifecycle. The measurement timestamp or number of machining operations t, and the corresponding uncertainty u(t) (reflecting the reliability of the measurement or prediction). The operating condition change sequence is the time-series data of process parameters, including but not limited to changes in spindle speed, feed rate, depth of cut, and cooling flow rate over time. Multi-source sensor time-series characteristics include, but are not limited to, dynamic characteristic sequences such as vibration acceleration, cutting force, spindle current, and cutting temperature.

[0109] In a specific embodiment, the temporal attention model TA-LSTM structure adopts a hybrid architecture of LSTM (Long Short-Term Memory) and multi-head attention mechanism. The LSTM layer solves the problem of long sequence dependencies, learns the dynamic evolution law of wear over time, and outputs the hidden state at each time step. The multi-head attention layer generates attention weights by calculating the correlation between the hidden state at each historical time step and the current time step. The output layer predicts the wear prediction results for the future preset processing cycle.

[0110] Wear state parameters and wear prediction results are transmitted to a suggestion generation model trained by a reinforcement learning model. The suggestion generation model adopts a hybrid decision-making mechanism of rule engine + multi-model fusion to analyze wear state parameters and wear prediction results and generate final tool adjustment suggestions.

[0111] In one embodiment, the control feedback interface feeds tool adjustment suggestions (tool change command, alarm threshold, process parameter suggestions (speed, feed rate, cooling parameters) and tool compensation parameters) back to the physical entity layer.

[0112] In one embodiment, the execution modes of the control feedback interface are divided into suggestion mode (only suggestions are output, and manual confirmation is required for execution) and automatic execution mode (control commands are automatically issued within the safety boundary). The safety boundary is defined as the upper and lower limits of parameter adjustment. In abnormal situations, the system automatically degrades to suggestion mode, while critical operations require manual confirmation.

[0113] The control feedback interface adopts a model fusion decision-making mechanism of "weighted voting + confidence threshold filtering". Specifically, when the TA-LSTM predicts the wear increment exceeds the threshold, suggests the optimal parameter adjustment scheme output by the model, and there is no abnormal alarm in the isolated forest, the process parameter adjustment is automatically executed; when the conclusions of any two models conflict, the manual review process is triggered to ensure the reliability of the decision.

[0114] In another embodiment, the service and application layer also performs a twin health assessment, comprehensively considering three dimensions: virtual-real consistency error, model prediction uncertainty, and expert network activation stability, to determine the twin health indicators.

[0115]

[0116] Where a, b, and c represent the virtual-real consistency error, respectively. Model prediction uncertainty Stability of expert network activation The weighting coefficients.

[0117] When I is less than the preset health indicator threshold, twin calibration is triggered (including fine-tuning of the physical information module, reorganization of the hybrid expert network, and retraining of the gating network) to ensure the long-term reliability of the twin's operation.

[0118] S104. When at least one of the applicable probability, the gate entropy value, and the wear state parameter meets a preset condition, the synchronization and correction layer tags the tool condition data and stores the tool condition data into a classification cache according to the tag.

[0119] In one embodiment, such as Figure 5 As shown, during the online monitoring and consistency degradation detection phase, the physical entity and its twin maintain a real-time synchronization mode and perform consistency degradation detection to determine whether the virtual and physical entities are consistent or have degraded. Specifically, the synchronization and correction layer performs consistency degradation detection through threshold detection of three core indicators: applicability probability, gating entropy value, and wear state parameters, thereby determining whether the tool condition data needs to be labeled. For example, the preset condition could be: 1) the maximum applicability probability output by the gating network. 1) The confidence level is less than the preset confidence threshold; 2) The gating entropy value output by the gating network. 3) Predicted residuals are greater than the preset entropy threshold; The residual is greater than the preset residual threshold.

[0120] In a specific embodiment, the entropy value output by the gating network is calculated. If the entropy value is too high, it indicates that the system has uncertainty about the current operating condition, and it is judged as consistency degradation; if the applicability probability of all existing experts for a given input is... All are low (e.g., maximum) Below the predefined confidence threshold If the output of the twin is inconsistent with the physical measurement value, then it is considered a consistency degradation; residual If the preset threshold is exceeded, a consistency check is triggered. If the residual drift exceeds the limit N times consecutively (this can be set according to actual needs), it is considered a consistency degradation. This mechanism is equivalent to a single-class classification task used to identify whether the current working condition is a learned working condition, with a confidence threshold. It can be adjusted through cross-validation according to specific application scenarios.

[0121] In one embodiment, such as Figure 5 As shown, when the consistency degradation detection result is consistency degradation, the process enters the twin model incremental evolution stage, executing consistency degradation response and data buffering, as well as the twin model incremental evolution strategy. Specifically, when any of the above conditions are met, the synchronization and correction layer, based on the combination of triggering conditions, adjusts the corresponding input data... Along with any available wear truth values Multi-dimensional composite labels are applied and categorized, stored in a buffer. Sufficient data is accumulated before incremental evolution of the twin model is performed. The buffer can employ a first-in, first-out (FIFO) queue or an importance-based sampling storage strategy to ensure the representativeness of the stored data. The labeling system includes the following core dimensions:

[0122] Working condition dimension classification: Based on the K-means clustering results of machining parameters (such as spindle speed, feed rate, and depth of cut), the new data is divided into three categories: known working condition variant data, similar new working condition data, and completely new working condition data. The clustering distance threshold is determined by the elbow rule.

[0123] Data quality is classified into three categories based on sensor data completeness (e.g., data completeness ≥95% is considered complete data) and noise intensity (e.g., signal-to-noise ratio ≥20dB is considered high-quality data): high-quality complete data, high-quality missing data, and low-quality data. Low-quality data is only used for anomaly trend analysis and is not used for expert training.

[0124] Wear stage classification: based on wear estimation status With failure threshold The ratio is divided into early wear, intermediate wear, and late wear, and expert networks are trained specifically for different wear stages. For example, early wear ( ), mid-term wear ( ), and later wear and tear ( Three categories.

[0125] Example label combination: [New working condition, Consistency degradation, High quality and complete, Mid-term wear] indicates that the data belongs to a new working condition that has not been learned, the twin prediction has a large deviation from the actual situation, the data quality is high, and the tool is in the mid-term wear stage.

[0126] In one embodiment, the synchronization and correction layer adopts a two-layer caching architecture of category-specific buffer + global shared buffer. Each data category corresponds to an independent buffer, which stores data according to the FIFO rule, and records metadata such as data acquisition time, working condition label, and quality rating, so as to realize fine-grained data management and efficient triggering of model updates.

[0127] S105. The twin model layer determines the network model to be updated based on the data caching status of the classification cache area, and trains the network model to be updated based on the cached data in the classification cache area to obtain the latest twin model layer.

[0128] In one embodiment, such as Figure 5 As shown, when consistency degradation is detected, the corresponding data is marked and stored in a buffer. Supra-model incremental evolution is then performed after sufficient data has accumulated. Specifically, the implementation strategy for the incremental evolution of the twin model is determined based on the data caching situation. Specifically, hierarchical triggering conditions are pre-set. When a certain type of new data reaches a preset quantity, the corresponding twin model incremental evolution strategy is triggered. During the twin model incremental evolution phase, new experts are instantiated and adaptively trained to expand the twin's processing capabilities and restore virtual-real consistency. For example: when the accumulated new operating condition data is ≥ the first preset threshold, the creation and training of a new expert network is triggered; when the accumulated similar new operating condition data is ≥ the second preset threshold or consistency degradation is detected for three consecutive synchronization cycles, a fine-tuning update of the similar expert network is triggered; when the accumulated known operating condition data is ≥ the third preset threshold, only the routing probability of the gating network is updated, and no new expert network is added.

[0129] In one embodiment, when it is confirmed that tagged data representing a new operating condition has been identified, the data will be sent to the hybrid expert module. Add a new expert network The new expert can be initialized with random weights or by cloning and fine-tuning the most similar existing expert to speed up convergence.

[0130] In one embodiment, new experts and gated networks The new expert is trained on a labeled dataset collected for the new operating condition, enabling it to specialize in handling the new operating condition data. The gating network is then updated to accurately route data from the new operating condition to the new expert without interfering with its routing decisions for previously known operating conditions.

[0131] Furthermore, during the adaptive training phase, the twin state prediction model Maintaining a freeze ensures that physical constraints are always enforced, and that new expert learning is generated in conjunction with... The expected input is consistent with the representation. The parameters of all previously trained expert networks are also kept frozen, which prevents catastrophic forgetting, i.e., the destruction of previously learned knowledge when learning a new task. Each expert retains its expertise in the previously learned task.

[0132] After training, the new expert will be fully integrated. In this process, the ability of twins to handle a wider range of operating conditions is effectively expanded, restoring consistency between the virtual and real systems. Then return to... Figure 5 The physical-twin real-time synchronization mode shown is used to update the twin model layer in preparation for recognizing and adapting to more novel operating conditions in the future.

[0133] S106. The synchronization and correction layer synchronizes the latest twin model layer and the physical entity layer so that the connection and data layer can obtain the second tool condition data from the physical entity layer through the data acquisition interface, and process the second tool condition data based on the latest twin model layer to obtain the second wear state vector.

[0134] In one embodiment, after the twin model layer is retrained, it is synchronized in real time with the calibration layer so that the connection and data layer can collect the second tool condition data through the data acquisition interface, and based on the latest twin model layer, perform the next wear state prediction to obtain the second wear state vector, as well as identify and adapt to more novel working conditions.

[0135] In the above embodiments, the expert network determines whether the current tool condition data needs to be labeled and classified for storage based on the applicability probability of the sensor feature vector and machining condition vector, the gate entropy value, and the wear state parameters. Secondly, model training is triggered based on the data caching status in the classification cache area, and the network model to be updated is determined based on the data caching status. This eliminates the need for complete retraining of the model, improving retraining efficiency. Training the network model to be updated based on the cached data in the classification cache area improves the model's ability to process various types of working condition data, thereby improving the accuracy of tool condition prediction.

[0136] Please see Figure 6 , Figure 6 This is a second schematic flowchart of a tool state prediction method based on digital twin provided in an embodiment of this application.

[0137] like Figure 6 As shown, the tool state prediction method based on digital twins specifically includes steps S201 to S202.

[0138] S201. When the amount of data in the data caching situation is greater than the first preset quantity threshold, the pre-trained general expert network and the gated network are obtained as the network model to be updated.

[0139] S202. Add the general expert network to the hybrid expert module, and train the general expert network and the gating network based on the preset learning rate dynamic adjustment strategy and the new working condition data of the classification cache to obtain the latest Siamese model layer.

[0140] In one embodiment, the new working condition data is completely unfamiliar working condition data (such as processing new materials, using new cutting tools, or significantly adjusting the range of process parameters) whose cosine similarity in feature space with the working condition data corresponding to the existing expert network is less than a preset similarity threshold.

[0141] In one embodiment, when the number of new working condition data in the classification cache exceeds a first preset threshold, a cross-industry pre-trained general expert network is loaded and used together with the gating network as the model to be updated. The general expert network is an expert network trained using cross-industry (machining, aerospace manufacturing, automotive parts processing) tool wear datasets to learn a general wear feature mapping.

[0142] A general expert network is added as a new expert network to the hybrid expert module, and the new expert network is fine-tuned. During the fine-tuning stage, only 10%-15% of the network parameters are updated, and the network converges quickly by combining new operating condition data.

[0143] In another embodiment, the training of new experts can also employ a three-stage learning rate strategy of warm-up, decay, and stabilization. Specifically, during the warm-up period (first 50 rounds), the learning rate is linearly increased from 0.0005 to 0.001 to stimulate parameter updates; during the decay period (rounds 50-200), cosine annealing is used to decay the learning rate and avoid local optima; during the stabilization period (after rounds 200), the learning rate is fixed at 0.0001 to finely adjust the parameters.

[0144] In one embodiment, during the training of the newly added general expert network and the gating network, the physical information module and the existing expert network are frozen throughout the process to ensure that the physical wear dynamics constraints are always in effect and do not destroy the learned mature operating condition knowledge.

[0145] Furthermore, the step of training the network model to be updated based on the cached data in the classification cache to obtain the latest Siamese model layer further includes: analyzing the new operating condition data based on the gating network to obtain the applicability probability of each expert network in the hybrid expert module to the new operating condition data, and performing state prediction on the new operating condition data based on each expert network to obtain state prediction results; performing a weighted average of each state prediction result based on the applicability probability to obtain a comprehensive predicted state; obtaining the actual wear state; constructing a distillation loss function based on the actual wear state and the comprehensive predicted state; and training the general expert network and the gating network based on the distillation loss function and the new operating condition data to obtain the latest Siamese model layer.

[0146] In one embodiment, the gated network analyzes new operating condition data and outputs the applicability probability vectors of each existing expert network. Each existing expert network predicts the wear state based on the new operating condition data, outputting its own state prediction result. Based on the applicability probability output by the gating network, the prediction results of each expert are weighted and averaged to obtain the comprehensive predicted state. The comprehensive prediction status includes the collective understanding of wear patterns among existing experts. Even if experts do not directly adapt to the new operating conditions, it can provide a general reference for wear trends, helping new experts to quickly establish reasonable feature mappings.

[0147] In one embodiment, a distillation loss function is constructed based on the actual wear state and the comprehensive predicted state, as shown in the following formula:

[0148]

[0149] in, For actual wear and tear Compared with predicted wear Cross-entropy loss function The weighting coefficients, To integrate the prediction results with the prediction results of the new expert network divergence loss function The weighting coefficients, and It can be set according to actual needs.

[0150] In one embodiment, minimizing the distillation loss function allows newly added general experts to quickly adapt to the unique characteristics of new operating conditions while inheriting general wear knowledge. Furthermore, during training, the physical information module and existing expert network parameters are frozen throughout, ensuring that physical wear dynamics constraints remain in effect without disrupting the learned, mature operating condition knowledge.

[0151] In another embodiment, the distribution difference between the intermediate features of the new expert network and the features of the existing expert network in the regenerating kernel Hilbert space is minimized by maximizing the mean difference, thereby constraining the consistency of the feature extraction direction and reducing the exploration cost of training the new expert.

[0152] In the above embodiments, by adopting a new expert network training strategy of pre-training general experts or knowledge distillation, the general wear law is transferred to the new working condition through the integrated knowledge distillation of existing experts, which reduces the expert training sample size requirement under the new working condition, improves the convergence speed, avoids the model to explore from scratch, and improves the prediction stability; the physical information module and the existing expert network are frozen throughout the process to ensure that the learning of the new working condition will not destroy the existing mature working condition adaptation capability.

[0153] Please see Figure 7 , Figure 7 This is a third schematic flowchart of a tool state prediction method based on digital twin provided in an embodiment of this application.

[0154] like Figure 7 As shown, the tool state prediction method based on digital twins specifically includes steps S301 to S303.

[0155] S301. When the data caching situation is such that the amount of similar new working condition data is greater than the second preset quantity threshold, the expert network corresponding to the similar working condition data is obtained as the benchmark expert network, wherein the similar new working condition data is new working condition data whose similarity to the similar working condition data is greater than the preset similarity threshold.

[0156] S302. Obtain the baseline expert network parameters, generate a new expert network, and use the gated network and the new expert network as the network model to be updated.

[0157] S303. Add the newly added expert network to the hybrid expert module, and fine-tune the gated network and the fine-tunable layer of the newly added expert network based on the preset difference constraint rules and the similar new working condition data to obtain the latest twin model layer.

[0158] In one embodiment, similar new working condition data refers to new working condition data whose cosine similarity in feature space to the corresponding working condition data of the existing expert network is greater than or equal to a first preset similarity threshold (such as processing parameters, sensor feature distributions that highly overlap with existing working conditions but have slight differences).

[0159] In one embodiment, when the cumulative number of similar new working condition data in the classification cache is greater than or equal to a second preset number, or when consistency degradation is detected for three consecutive synchronization cycles, the model training and update process described in the following embodiment is initiated.

[0160] In a specific embodiment, the feature space cosine similarity between the new working condition data and the corresponding working condition data of all existing expert networks is calculated, and the expert network with the highest similarity is selected as the benchmark expert network.

[0161] By cloning all parameters of the baseline expert network, a new expert network with the exact same structure as the baseline is generated. This allows the new expert network to inherit the mature feature extraction capabilities for similar operating conditions, avoiding training from scratch and significantly shortening the convergence time.

[0162] The newly added expert network and gating network are listed as network models to be updated. The physical information module and the parameters of the existing expert network are kept frozen to ensure that physical constraints are always effective, avoid catastrophic forgetting, and focus on adapting to new operating conditions.

[0163] In one embodiment, only the fine-tunable layers (output layer and the last two hidden layers) of the newly added expert network are fine-tuned, while the parameters of the remaining layers remain consistent with the baseline expert.

[0164] A domain-adaptive regularization term is introduced to constrain the range of parameter differences between new experts and baseline experts, avoiding deviations from the knowledge base of similar working conditions. The formula is as follows:

[0165]

[0166] in This is the cloning regularization coefficient (default 0.01). To allow for fine-tuning of the number of layers and ensure that new experts can quickly adapt to new scenarios while inheriting knowledge from similar working conditions, Add a parameter for the Lth level of the expert. Benchmark expert L-level parameters.

[0167] In one embodiment, during the training of the newly added expert and gating network, the loss function primarily focuses on data fidelity loss, combined with a difference constraint regularization term. The loss function is:

[0168]

[0169] in, A mean squared error loss function is provided for the newly added expert outputs the wear value and the actual value to ensure the prediction accuracy under the new working conditions.

[0170] After training is completed, the new expert network is added to the hybrid expert module, and the gated network updates the routing probability to achieve accurate adaptation to similar new working conditions.

[0171] Furthermore, the twin model layer determines the network model to be updated based on the data caching status of the classification cache area, and further includes: when the data caching status is that the amount of known working condition variant data is greater than a third preset quantity threshold, the gated network is used as the network model to be updated.

[0172] In one embodiment, when the known working condition variant data is data that perfectly matches the corresponding working condition of the existing expert network (i.e., similarity ≥ second similarity threshold (second similarity threshold is greater than first similarity threshold)) and the cumulative data volume ≥ third preset quantity threshold, only the gating network is updated. Specifically, when there is sufficient known working condition variant data, the gating network is retrained, the routing weights for the working condition data are optimized, and the routing probability parameters of the gating network are adjusted to ensure that the data is accurately allocated to the corresponding expert network without the need to add or fine-tune the expert network, thus reducing unnecessary retraining.

[0173] In the above embodiments, differentiated updates are performed for different types of working condition data (similar new working conditions and existing working conditions) to balance model adaptability and training cost. Secondly, the cloning + local fine-tuning mechanism under similar new working conditions allows the model to quickly adapt to new scenarios while inheriting existing knowledge, reducing the need for new working condition samples and improving training efficiency. The physical information module and existing expert network parameters are always frozen to ensure that the learned mature working condition knowledge is not destroyed when adapting to new working conditions.

[0174] Please see Figure 8 , Figure 8 This is a fourth schematic flowchart of a tool state prediction method based on digital twin provided in an embodiment of this application.

[0175] like Figure 8 As shown, the tool state prediction method based on digital twins specifically includes steps S401 to S403.

[0176] S401. Based on the gating network, process the sensor feature vector and the processing condition vector to obtain the applicability probability of each expert network and the gating entropy value;

[0177] S402. When the difference in applicability probability is greater than or equal to a preset difference threshold, the sensor feature vector and the processing condition vector are processed based on the expert network with the highest applicability probability to obtain a twin state feature vector, wherein the difference in applicability probability is the difference between the maximum applicability probability and the minimum applicability probability.

[0178] S403. Based on the twin state prediction model, process the sensor feature vector, the processing condition vector, and the twin state feature vector to obtain the first wear state vector.

[0179] In one embodiment, a gated network Used as an intelligent router or scheduler, it also supports consistency monitoring. The default is MLP, and its parameters are... The gating network receives sensor feature vectors. and processing condition vector It analyzes the sensor feature vector and processing condition vector, and outputs a probability vector. Each of them and probability Representing the gating network for experts The applicability assessment (i.e. applicability probability) of the current input data is processed, and the entropy value and maximum applicability probability output by the gating network are used simultaneously for consistency degradation detection.

[0180] In one embodiment, when the probability vector When the difference between the maximum and minimum applicability probabilities is greater than or equal to a preset threshold, it indicates that different expert networks have significantly different processing results for the current sensor feature vector and processing condition vector. In this case, the expert network with the highest applicability probability is selected to process the current sensor feature vector and processing condition vector to improve the accuracy of the results.

[0181] In a specific embodiment, the expert network with the highest applicability receives the sensor feature vector. and processing condition vector And convert it into a standardized twin state feature vector h.

[0182] In one embodiment, twin state prediction model The wear state vector is generated by analyzing the twin state feature vector, sensor feature vector, and processing condition vector.

[0183]

[0184] Further, after processing the sensor feature vector and the processing condition vector based on the gating network to obtain the applicability probability of each expert network and the gating entropy value, the method further includes: when the difference in applicability probability is less than the preset difference threshold, processing the sensor feature vector and the processing condition vector based on each expert network to obtain each state feature sub-vector; weighting and fusing each state feature sub-vector according to each applicability probability to obtain the twin state feature vector; and processing the sensor feature vector, the processing condition vector, and the twin state feature vector based on the twin state prediction model to obtain the first wear state vector.

[0185] In one embodiment, when the difference between the maximum applicability probability and the minimum applicability probability is less than a preset difference threshold, it indicates that the processing results of different expert networks for the current sensor feature vector and processing condition vector are relatively similar. In this case, the sensor feature vector is received by all K expert networks respectively. and processing condition vector And transform it into a standardized state feature subvector. .

[0186] In one embodiment, the final wear state vector The weighted sum of the applicability probabilities of all expert network outputs and gating network outputs:

[0187]

[0188] In one embodiment, twin state prediction model The wear state vector is generated by analyzing the twin state feature vector, sensor feature vector, and processing condition vector.

[0189] In the above embodiments, the optimal expert network for sensor feature vectors and processing condition vectors is determined based on the applicability probability of the expert network to the sensor feature vectors and processing condition vectors, so as to improve the accuracy of twin state feature vectors and thus improve the accuracy of final wear prediction.

[0190] It should be noted that although this invention is primarily described in the context of tool wear monitoring and prediction in automated machining (such as robotic drilling), the described digital twin framework has broad applicability. It can be applied to other problems in prediction and health management, where the system operates under dynamically changing conditions and the underlying physics is at least partially known. Application examples include bearing failure prediction, battery degradation modeling, or state estimation of other components in complex systems.

[0191] Please see Figure 9 , Figure 9 This is a schematic block diagram illustrating the structure of a computer device according to an embodiment of this application. The computer device may be a server.

[0192] See Figure 9 The computer device includes a processor, memory, and network interface connected via a system bus, wherein the memory may include non-volatile storage media and internal memory.

[0193] Non-volatile storage media can store operating systems and computer programs. These computer programs include program instructions that, when executed, cause the processor to perform any digital twin-based tool state prediction method.

[0194] The processor provides computing and control capabilities, supporting the operation of the entire computer device.

[0195] Internal memory provides an environment for the execution of computer programs in non-volatile storage media. When executed by a processor, the computer program enables the processor to perform any tool state prediction method based on digital twins.

[0196] This network interface is used for network communication, such as sending assigned tasks. Those skilled in the art will understand that... Figure 9 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0197] It should be understood that the processor can be a Central Processing Unit (CPU), but it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among these, a general-purpose processor can be a microprocessor or any conventional processor.

[0198] In one embodiment, the processor is configured to run a computer program stored in memory to perform the following steps:

[0199] The connection and data layer acquires first tool condition data from the physical entity layer through a data acquisition interface, and processes the first tool condition data to obtain sensor feature vectors and machining condition vectors.

[0200] The twin model layer processes the sensor feature vector and the processing condition vector to obtain the applicability probability, gate entropy value and first wear state vector of each expert network.

[0201] The service and application layer performs state estimation based on the first wear state vector to obtain wear state parameters;

[0202] When at least one of the applicable probability, the gate entropy value, and the wear state parameter meets a preset condition, the synchronization and correction layer tags the tool condition data and stores the tool condition data into a classification cache according to the tag.

[0203] The twin model layer determines the network model to be updated based on the data caching status of the classification cache, and trains the network model to be updated based on the cached data in the classification cache to obtain the latest twin model layer.

[0204] The synchronization and correction layer synchronizes the latest twin model layer and the physical entity layer so that the connection and data layer can obtain the second tool condition data from the physical entity layer through the data acquisition interface, and process the second tool condition data based on the latest twin model layer to obtain the second wear state vector.

[0205] The embodiments of this application also provide a computer-readable storage medium storing a computer program, the computer program including program instructions, and the processor executing the program instructions to implement any of the tool state prediction methods based on digital twins provided in the embodiments of this application.

[0206] The computer-readable storage medium may be an internal storage unit of the computer device described in the foregoing embodiments, such as the hard disk or memory of the computer device. The computer-readable storage medium may also be an external storage device of the computer device, such as a plug-in hard disk, SmartMedia Card (SMC), Secure Digital (SD) card, or Flash Card equipped on the computer device.

[0207] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A tool condition prediction method based on digital twin, characterized in that, The method is applied to a digital twin prediction system, which includes a physical entity layer, a connectivity and data layer, a twin model layer, a service and application layer, and a synchronization and correction layer. The connection and data layer acquires first tool condition data from the physical entity layer through a data acquisition interface, and processes the first tool condition data to obtain sensor feature vectors and machining condition vectors. The twin model layer processes the sensor feature vector and the processing condition vector to obtain the applicability probability, gate entropy value and first wear state vector of each expert network. The service and application layer performs state estimation based on the first wear state vector to obtain wear state parameters; When at least one of the applicable probability, the gate entropy value, and the wear state parameter meets a preset condition, the synchronization and correction layer tags the first tool condition data and stores the first tool condition data into a classification cache according to the tag. The twin model layer determines the network model to be updated based on the data caching status of the classification cache, and trains the network model to be updated based on the cached data in the classification cache to obtain the latest twin model layer. The synchronization and correction layer synchronizes the latest twin model layer and the physical entity layer so that the connection and data layer can obtain the second tool condition data from the physical entity layer through the data acquisition interface, and process the second tool condition data based on the latest twin model layer to obtain the second wear state vector.

2. The tool state prediction method based on digital twin according to claim 1, characterized in that, The twin model layer includes a gated network, a hybrid expert module, and a physical information module. The hybrid expert module includes at least one expert network, and the physical information module includes a twin state prediction model.

3. The tool state prediction method based on digital twin according to claim 2, characterized in that, The twin model layer determines the network model to be updated based on the data caching status of the classification cache, and trains the network model to be updated based on the cached data in the classification cache to obtain the latest twin model layer, including: When the data caching situation is such that the amount of new working condition data is greater than the first preset quantity threshold, the pre-trained general expert network and the gated network are obtained as the network model to be updated. The general expert network is added to the hybrid expert module, and the general expert network and the gating network are trained based on a preset learning rate dynamic adjustment strategy and the new working condition data of the classification cache to obtain the latest Siamese model layer.

4. The tool condition prediction method based on digital twin according to claim 3, characterized in that, The step of training the network model to be updated based on the cached data in the classification cache to obtain the latest Siamese model layer further includes: The new operating condition data is analyzed based on the gating network to obtain the applicability probability of each expert network in the hybrid expert module to the new operating condition data, and state prediction is performed on the new operating condition data based on each expert network to obtain the state prediction result. Based on the applicable probability, the prediction results of each state are weighted and averaged to obtain the comprehensive predicted state. Obtain the actual wear state, and construct a distillation loss function based on the actual wear state and the comprehensive predicted state; The general expert network and the gated network are trained based on the distillation loss function and the new operating condition data to obtain the latest twin model layer.

5. The tool state prediction method based on digital twin according to claim 2, characterized in that, The twin model layer determines the network model to be updated based on the data caching status of the classification cache, and trains the network model to be updated based on the cached data in the classification cache to obtain the latest twin model layer, and also includes: When the data caching situation is such that the amount of similar new working condition data is greater than the second preset quantity threshold, the expert network corresponding to the similar working condition data is obtained as the benchmark expert network, wherein the similar new working condition data is new working condition data whose similarity to the similar working condition data is greater than the preset similarity threshold. Obtain the baseline expert network parameters, generate a new expert network, and use the gated network and the new expert network as the network model to be updated. The newly added expert network is added to the hybrid expert module, and the gated network and the fine-tunable layer of the newly added expert network are fine-tuned based on the preset difference constraint rules and the similar new working condition data to obtain the latest twin model layer.

6. The tool state prediction method based on digital twin according to claim 1, characterized in that, The twin model layer determines the network model to be updated based on the data caching status of the classification cache, and also includes: When the data caching situation is such that the amount of known operating condition variant data is greater than a third preset threshold, the gated network is used as the network model to be updated.

7. The tool state prediction method based on digital twin according to claim 2, characterized in that, The twin model layer processes the sensor feature vector and the processing condition vector to obtain the applicability probability, gating entropy value, and first wear state vector of each expert network, including: The sensor feature vector and the processing condition vector are processed based on the gating network to obtain the applicability probability of each expert network and the gating entropy value. When the difference in applicability probability is greater than or equal to a preset difference threshold, the sensor feature vector and the processing condition vector are processed based on the expert network with the highest applicability probability to obtain a twin state feature vector, wherein the difference in applicability probability is the difference between the maximum applicability probability and the minimum applicability probability. Based on the twin state prediction model, the sensor feature vector, the processing condition vector, and the twin state feature vector are processed to obtain the first wear state vector.

8. The tool condition prediction method based on digital twin according to claim 7, characterized in that, After processing the sensor feature vector and the processing condition vector based on the gating network to obtain the applicability probability of each expert network and the gating entropy value, the method further includes: When the difference in applicable probabilities is less than the preset difference threshold, the sensor feature vector and the processing condition vector are processed based on each of the expert networks to obtain each state feature sub-vector; The twin state feature vectors are obtained by weighted fusion of the state feature subvectors according to the applicable probabilities. Based on the twin state prediction model, the sensor feature vector, the processing condition vector, and the twin state feature vector are processed to obtain the first wear state vector.

9. The tool state prediction method based on digital twin according to claim 2, characterized in that, Before the twin model layer processes the sensor feature vector and the processing condition vector to obtain the applicability probability, gate entropy value, and first wear state vector of each expert network, it also includes: Collect nominal operating condition dataset; Based on the preset composite loss function and the nominal working condition dataset, the pre-trained twin state prediction model is iteratively trained to obtain the gradient norm of each loss term in the preset composite loss function. Obtain the current training stage, and adjust the weight coefficients of each loss term according to the current training stage, each gradient norm, and the preset weight adjustment formula to obtain the composite loss function for the next iteration. Iterative training is performed according to the composite loss function of the next iteration until the preset conditions are met, at which point the iterative training stops and the twin state prediction model is obtained.

10. The tool state prediction method based on digital twin according to claim 1, characterized in that, The digital twin prediction system further includes a control feedback interface. After the service and application layer performs state estimation based on the first wear state vector and obtains the wear state parameters, it also includes: Wear trend prediction is performed on historical wear data and working condition change sequences based on a time-series attention model to obtain wear prediction results; Based on a preset suggestion generation model, the wear state parameters and wear prediction results are analyzed to generate tool adjustment suggestions; The control feedback interface feeds back the tool adjustment suggestions to the physical entity layer.

Images