A machining state monitoring and real-time error reconstruction visualization method

Abstract

The application belongs to the field of real-time monitoring and error prediction of precision machining process, and discloses a machining state monitoring and real-time error reconstruction visualization method. The method faces the real-time monitoring and error regulation requirements of the machining process, inputs the machining process data into the data twin model, classifies and extracts the spatial dimension features according to the physical structure of the real-time data. Then the sequence real-time twin data with spatial dimension features are transmitted to the LSTM layer to extract the time dimension features. The LSTM layer generates the prediction output of the twin system, realizes the real-time monitoring of the machining process of the manufacturing system, and realizes the accurate real-time evaluation of the workpiece machining process and the workpiece machining quality. The application overcomes the poor real-time performance, insufficient prediction accuracy and lack of interpretability of the existing digital twin technology, provides protection for improving the machining precision and machining efficiency of the workpiece, and lays a foundation for the accurate regulation of the manufacturing system.

Description

Technical Field

[0001] This invention belongs to the field of real-time monitoring and error prediction of precision machining processes, and relates to a machining status monitoring and real-time error reconstruction visualization method driven by digital twins and neural networks. Background Technology

[0002] In flexible manufacturing systems, the increasing complexity and scale of equipment make machine tool operation and processing more prone to machining errors. Machining errors often stem from the dynamic randomness, multi-source uncertainty, high coupling, and strong interference factors inherent in the machining process. Therefore, performing multi-source error data fusion and achieving rapid error diagnosis under various machining scenarios becomes extremely difficult. Traditional machining inspection methods involving machining, measurement, and feedback often lead to resource waste, extended R&D cycles, and high machining costs.

[0003] Digital twin (DT) technology leverages machine learning and artificial intelligence for in-depth modeling and maximized efficiency. These methods primarily utilize complex neural networks to extract hidden information from high-dimensional features, establish correlations between input data and predicted categories, and achieve end-to-end identification. Digital twin systems connect to the physical environment via communication, thus requiring reliable communication links and real-time bidirectional data flow between the physical entity and the virtual model through sensors. The simulation of the manufacturing process depends on the real-time evolution of the model, necessitating predictive models built based on physical laws / machine learning and sufficient prediction accuracy to ensure the accuracy and realism of the simulation. Finally, the digital twin system must be able to automatically optimize parameters or adjust control commands based on current and predicted processing states. This requires the integration and analysis of these different data streams to create a comprehensive and unified representation. Digital twins for manufacturing systems require the construction of a complete closed loop of "perception-computation-execution," which is also a core challenge of Industry 4.0 transformation.

[0004] Currently, in order to solve the above problems and achieve real-time and accurate monitoring of the processing process, existing technology literature 1 "Enhanced anomaly detection of industrial control systems via graph-driven spatio-temporal adversarial deep support vector data description", Li et al., 2025, 270, effectively captures spatial relationships by adopting a dual strategy of static and dynamic graph learning, and integrates attention-based temporal embedding and adversarial classification techniques. This method can solve the spatio-temporal complexity of multidimensional data and the impact of data imbalance on anomaly detection performance, thereby achieving efficient and accurate anomaly detection. However, this model currently lacks interpretability, and the complex combination of various technologies makes it difficult to understand its internal mechanism and the contribution of each component. Existing technology literature 2 "SCADA data-driven blade icing detection for wind turbines: an enhanced spatio-temporal feature learning approach", Jiang et al., 2023, 34(5), believes that real-time data in the processing process has complex time-varying characteristics, and there is a strong spatio-temporal correlation between different sensor information. Therefore, it is still challenging to extract effective processing features for accurate detection. Existing technical literature 3, "A novel framework for spatio-temporal prediction of environmental data using deep learning," Amato et al., 2020, 10(1): 22243, indicates that deep learning models have proven capable of capturing spatial, temporal, and spatiotemporal dependencies through feature learning. However, the problem of interpolating continuous spatiotemporal fields based on measurements from a set of irregular points in space remains insufficiently studied. Commonly used machine learning and artificial intelligence algorithms possess black-box properties; although these algorithms yield highly accurate results, the reasons for this are unclear, limiting their further promotion and application. Summary of the Invention

[0005] This invention aims to overcome the shortcomings and deficiencies of existing technologies for multi-source error data fusion and rapid error diagnosis in processing scenarios. Addressing the problems of resource waste, extended R&D cycles, and high processing costs associated with traditional processing and inspection methods involving processing, measurement, and feedback, this invention proposes a processing status monitoring and real-time error reconstruction visualization method driven by a combination of digital twins and neural networks. This method, designed for real-time monitoring and error control in the processing process, inputs processing data into a data twin model, classifies the real-time data according to its physical structure, and extracts spatial dimension features. Then, the real-time twin data sequence with spatial dimension features is passed to the prediction layer of a Long Short-Term Memory (LSTM) network to extract temporal dimension features. The LSTM utilizes its ability to remember important historical information to generate the predictive output of the twin system, ultimately achieving real-time monitoring of the manufacturing system's processing process and enabling accurate real-time evaluation of the workpiece processing process and quality. This overcomes the shortcomings and deficiencies of existing digital twin technologies, such as poor real-time performance, insufficient prediction accuracy, and lack of interpretability of prediction results, providing a guarantee for improving workpiece processing accuracy and efficiency, and laying the foundation for precise control of the manufacturing system.

[0006] The technical solution of the present invention:

[0007] A method for monitoring processing status and visualizing real-time error reconstruction is proposed. Addressing the significant temporal dependence and spatial correlation characteristics of real-time twin data during the processing, a digital twin model is established while ensuring data interpretability. This model achieves spatial dimension feature extraction of processing information and physical dimensionality reduction of the data scale. LSTM is used to extract temporal dimension features from the twin data, enabling prediction of processing accuracy while maintaining real-time performance. The specific steps of the method are as follows:

[0008] S1: Real-time construction of process-oriented digital twin systems (DT);

[0009] During the machining process, all material removal is achieved through high-precision servo motion of various moving parts of the machine tool. By collecting the input and output variables of the control process of each moving part, a digital twin system corresponding to the real-time machining process can be established. The manufacturing system completes the machining of the workpiece through the interaction between the tool and the workpiece, and the machining state is directly reflected in the "tool-workpiece" subsystem. The "tool-workpiece" subsystem under the digital twin system is constructed; the tools are grinding wheels, milling cutters, and turning tools, and the corresponding tool system is shown in the following equation:

[0010] (1)

[0011] Where G represents the grinding system, M represents the milling system, and L represents the turning system;

[0012] The physical structure correspondence of "machine tool-tool-workpiece" is constructed as follows: the tool is connected to the tool spindle, the tool spindle is connected to the machine tool guideway, and the machine tool guideway is connected to the machine tool bed; the workpiece is connected to the workpiece spindle, the workpiece spindle is connected to the machine tool guideway, and the machine tool guideway is connected to the machine tool bed; the machine tool is a CNC grinding machine (CNCMT), and the digital twin system representation of the physical structure of the CNC grinding machine is as follows:

[0013] (2)

[0014] Where X represents the X-axis feed system of the CNC grinding machine, Z represents the Z-axis feed system of the CNC grinding machine, C represents the C-axis rotation system of the CNC grinding machine, and T represents the tool system of the CNC grinding machine.

[0015] S2: Long Short-Term Memory (LSTM) Neural Network Model Oriented to Process Flow;

[0016] A qualified product is formed from raw materials through a series of processes, steps, and working strokes. Historical data during the processing will affect the current processing status of the workpiece. Processing history information helps to achieve fault diagnosis and improve diagnostic accuracy.

[0017] A fault diagnosis method for machining processes driven by digital twins and deep learning is proposed to realize fault diagnosis of workpiece machining processes. By simplifying the physical structure of CNC grinding machines with digital twin systems, machining process diagnosis based on historical machining data before time t and the current machining state at time t is realized through a long short-term memory neural network model (LSTM).

[0018] Based on the simplified physical structure of the CNC grinding machine according to the digital twin system, as shown in Equation (2), the real-time machining data of the physical structure of the CNC grinding machine is extracted through the digital twin system and used as the real-time input data s(t) of the long short-term memory neural network model:

[0019] (3)

[0020] Where D(t) is the command position corresponding to the digital twin system, δ(t) is the real-time error corresponding to the digital twin system, i(t) is the real-time current corresponding to the digital twin system, and P(t) is the real-time power corresponding to the digital twin system.

[0021] (4)

[0022] In this context, the subscripts X, Z, C, and T represent the physical structure of the CNC grinding machine: the X-axis feed system, the Z-axis feed system, the C-axis rotary system, and the tool system, respectively.

[0023] The corresponding Long Short-Term Memory (LSTM) neural network model under the digital twin system architecture is represented as follows:

[0024] (5)

[0025] Among them, h DT (t-1) represents the fault diagnosis data corresponding to the real-time error output by the DT-LSTM system at the previous moment, and is processed through the forgetting unit f. DT (t) and memory update unit c DT (t) is reflected in real-time fault diagnosis data; forgetting unit f DT (t) is used to determine whether to use historical error data; memory update unit c DT (t) and candidate memory units DT (t) is updated in real time through gate activation functions and state activation functions; output unit i DT (t) determines the amount of information about the current state to retain and outputs it through the output gate. DT (t) Output; σ represents the gate activation function, tanh is the state activation function, and W s V s and b represent the learnable input weights, recurrent weights, and bias of the DT-LSTM system, respectively, and W si W sf W so and W sc V represents the input weights of the input gate, forget gate, output gate, and memory update unit, respectively. si V sf V so and V sc These represent the recurrent weights of the input gate, forget gate, output gate, and memory update unit, respectively. i b f b o and b c These represent the biases of the input gate, forget gate, output gate, and memory update unit, respectively; h DT (t-1) represents the fault diagnosis data corresponding to the real-time error output by the DT-LSTM system at the previous time step; s(t) represents the real-time input data of the Long Short-Term Memory neural network model; h DT (t) represents the fault diagnosis data corresponding to the real-time error output by the DT-LSTM system at the current moment:

[0026] (6)

[0027] Where p(t) is the current probability estimate of the component, and R(t) is the failure evolution probability, i.e., the impact of the current failure on the machining accuracy of subsequent workpieces:

[0028] (7)

[0029] And there are:

[0030] (8)

[0031] Where γ(t) represents the error caused by each factor;

[0032] Based on this, the computational flow of real-time twin data in the LSTM neural network in the digital twin model was established, and the real-time connection of processing data in different processes, steps, and work journeys based on historical processing information was obtained. This enabled the classification, location, and diagnosis of processing faults, and the predictive maintenance of the processing system.

[0033] S3: DT-LSTM collaborative driven machining error lead control strategy;

[0034] Compared to existing digital twin systems for machining processes, the DT-LSTM architecture established in this invention can effectively regulate the current machining state based on advanced predictive twin data. Traditional digital twin systems can achieve good real-time monitoring of the machining process and detect abnormalities in a timely manner. However, in this case, the twin data ht directly affects the workpiece, leading to a decrease in workpiece quality Q(t). This means the current workpiece is already scrapped, and adjusting the machining control parameters at this point only affects the workpiece quality at time t+1 and beyond, resulting in a significant waste of manpower, materials, and time. This invention, through its DT-LSTM network structure, can accurately predict future machining data h(t). By combining historical data and twin prediction data generated by LSTM, accurate prediction of the machining process can be achieved quickly. The CNC system can adjust control parameters in advance and directly affect the real-time twin data ht, thereby directly improving the machining quality of the workpiece and increasing the workpiece yield.

[0035] The DT-LSTM system constructed in step S2 effectively regulates the current processing state based on the advanced predictive twin data; the historical processing data s(t-1) before time t and the real-time processing data s(t) are transmitted to the digital twin system constructed by equation (2); at time step t, the LSTM network uses the fault diagnosis data h corresponding to the real-time error output by the DT-LSTM system at the previous time step. DT (t-1) and the fault diagnosis data h corresponding to the real-time error output by the DT-LSTM system at the current time of the sequence. DT (t) is used to calculate the output and the updated cell state h. DT(t+1); at the same time, the updated unit state at time step t+1 is transmitted to the digital twin system, and the corresponding predicted processing data s(t+1) is generated; based on this, the predicted processing data s(t+1) is compared with the real-time processing data s(t) to analyze the relative magnitude of the workpiece error. When the workpiece error exceeds the allowable range, a control strategy is adopted to adjust the relevant parameters of the current digital twin system and directly improve the processing quality of the workpiece, thereby increasing the yield of the workpiece, as shown in equation (9):

[0036] (9)

[0037] S4: A method for acquiring twin data of the manufacturing process in collaboration with a DT-LSTM system;

[0038] Based on the proposed digital twin-deep learning real-time architecture, the machining process can be monitored while simultaneously performing fault diagnosis and error localization. Through an LSTM neural network, the distribution patterns of machining errors and feature values ​​are quickly derived and compared with the characteristics of the main moving parts of the machine tool for fault diagnosis. Based on the digital twin architecture, the fault diagnosis results can be compared with the physical structure of the main moving parts to achieve error localization, reduce system errors during machining, and ensure stable operation of the product machining process.

[0039] During the workpiece processing, the establishment of the digital twin system is divided into two stages: model building and real-time monitoring. In the model building stage, the selected sampling frequency is greater than 2000Hz; in the real-time monitoring stage, the selected sampling frequency is less than 100Hz.

[0040] Based on the DT-LSTM system, the digital twin system directly acquires real-time machining data from the CNC grinding machine during workpiece processing, ensuring that the data acquisition method corresponds to the machining parameters. Since the real-time machining data is continuous over a long period, it is acquired in segments according to the workpiece machining principle. Each revolution of the C-axis is defined as a machining and acquisition unit for the digital twin system, corresponding to a training unit of the LSTM network. Therefore, the feed rate per revolution of the C-axis is defined as a spatial period, with the corresponding time period dt. h :

[0041] (10)

[0042] Where, n s Main spindle speed; one-dimensional twin data size N of a single moving part within one time period. s Represented as:

[0043] (11)

[0044] Among them, Fs The sampling frequency is set; multiple time periods are set as one time step:

[0045] (12)

[0046] Where n1 is the number of time periods, the total amount of twin data within one time step is expressed as:

[0047] (13)

[0048] Where m is the number of corresponding moving parts in the DT-LSTM system, and n is the type of real-time data for a single moving part;

[0049] It can be observed that the data size within a single time step of the DT-LSTM established in this invention can be determined by the number of basic time periods n1 and the sampling frequency F. s Adjustment is key. A higher sampling frequency results in a larger data volume and a longer system response time; a larger number of basic time periods leads to a longer single calculation time and higher theoretical prediction accuracy. In a digital twin system, the accuracy of the grating ruler determines the actual machining accuracy of the machine tool, while the data acquisition frequency depends on the system's response time and machining prediction accuracy. A higher sampling frequency results in a larger data volume, corresponding to higher computational costs and a longer system response time. The type of data acquired depends on the detection requirements, such as position data or power data. To balance data volume and computational cost, the data in the digital twin system structure can be dimensionality-reduced according to the above formula.

[0050] During the actual workpiece machining stage, due to the presence of workpiece material blank errors, installation errors, machine tool coordinate system alignment, and other factors inherent in the machining preparation stage, the data acquisition parameters can be set to the model building stage. As the pre-machining preparation work progresses, the DT-LSTM model of the machining process can be established simultaneously. When the workpiece begins full machining, the data acquisition parameters can be set to the real-time monitoring stage to achieve the detection and prediction of the machining process.

[0051] S5: A method for classifying and identifying machining errors driven by a DT-LSTM system;

[0052] Based on the physical meaning of the real-time twin data in the DT-LSTM system, the workpiece machining process is evaluated in real time using spatial dimension evaluation indicators in the same spatiotemporal coordinate system. In the machining of cylindrical workpieces, the spacing of the spiral lines of the machining trajectory is the feed per revolution of the C-axis, which directly corresponds to the grinding mark spacing of the cylindrical workpiece and affects the residual height of the material surface, greatly influencing the surface roughness of the workpiece, as shown in the following formula:

[0053] The feed rate (grid spacing) per revolution of the C-axis is expressed as:

[0054] (14)

[0055] Where, n s The spindle speed is F, and the feed rate is F. The workpiece surface machining residual height is calculated as follows:

[0056] (15)

[0057] Among them, R s R is the tool radius. h The surface roughness is expressed as the residual height of the workpiece surface after machining, and is:

[0058] (16)

[0059] Where l is the sampling length during the workpiece surface roughness measurement process, and f t (x) represents the actual contour curve function of the workpiece obtained by measurement; the motion error of the C-axis also directly corresponds to the roundness error of the workpiece, as shown in the following formula:

[0060] (17)

[0061] Among them, f max (x) and f min (x) represent the maximum and minimum profile curves of the same radial direction on the workpiece surface, respectively; the cylindricity of the workpiece is expressed as:

[0062] (18)

[0063] Among them, f ⊥ (x) and f ∥ f(x) represents two actual contour curves on the workpiece surface with a spatial angle difference of 90 degrees; f(x) represents the ideal target contour curve function of the workpiece; the Z-axis motion error directly corresponds to the taper of the workpiece, as shown in the following formula:

[0064] (19)

[0065] Among them, f z The taper ratio is represented by f1(x), the diameter of the large end of the workpiece is represented by f2(x), and the measurement length is represented by L.

[0066] This enables the evaluation of the workpiece's real-time C-axis feed per revolution, surface residual height, surface roughness, roundness, and taper error;

[0067] Workpiece errors, such as surface quality and contour errors, are directly related to the real-time motion parameters of machine tool moving parts. These often involve the measurement of different spatial characteristic parameters and are significantly affected by measurement errors, frequently requiring multiple measurements to obtain the maximum value. The DT-LSTM established by this invention can acquire machining process parameters in real time and perform real-time analysis and prediction. The relevant geometrical accuracy can be uniquely and accurately monitored by the digital twin architecture established by this invention. The real-time spatial twin data corresponding to the digital twin system avoids measurement errors and time losses introduced by offline measurements. Especially in the measurement of complex curved surface parts, its three major advantages—online measurement, measurement data within a unified machine tool coordinate system, and data directly derived from high-precision optical scales—significantly improve measurement accuracy and efficiency.

[0068] S6: A real-time monitoring method for the manufacturing process driven by a DT-LSTM system;

[0069] The DT-LSTM established by this invention achieves a twin mapping between complex three-dimensional spatial data and machine tool single-axis motion data. The DT-LSTM model established by this invention can effectively capture the corresponding twin data during workpiece machining. Since the machining scale of machine tools in three-dimensional space is on the order of millimeters, while the error data of machine tools is mostly on the order of sub-micrometers, the error data cannot be well reflected in the machining data. Therefore, the spatial dimension error data and position data of each axis are extracted separately, and time scale alignment is performed to obtain the temporal correspondence between real-time error data and machining data.

[0070] During the machining process of the same workpiece, its accuracy continuously improves with the execution of processes such as roughing, semi-finishing, and finishing. The ideal shape of the workpiece corresponding to different processes has spatial and temporal consistency, while the workpiece error varies under different processes. Therefore, real-time error data of the workpiece is more representative of the machining status than real-time position data.

[0071] The DT-LSTM architecture established according to this invention maps twin data to three-dimensional space, obtaining three-dimensional spatial error data corresponding to the virtual X, Y, and Z axes of the machine tool in standby mode. At this time, the machine tool spatial error data is in length units; with consistent units, longitudinal and lateral comparisons in the spatiotemporal dimensions can be performed. Furthermore, because this invention's digital twin system is based on physical components, each twin data point possesses significant physical meaning.

[0072] By comparing the error data of each axis before and after machining, and comparing the error data of each axis with the machining load, it is beneficial to select the optimal machining parameters and control parameters. The DT-LSTM model established in this invention can also clearly and intuitively illustrate the source of each twin signal and the motion state of the corresponding physical components. These functions are not available in previous conventional data twin structures.

[0073] Under the machine tool mechanical coordinate system, a machine tool spatial coordinate system and a digital twin system spatial coordinate system are constructed respectively, and the coordinate origins and X, Y and Z axes of the two correspond one-to-one;

[0074] In the spatial coordinate system of the digital twin system, spatial dimension error data and position data of the X, Z, and C axes are extracted respectively, and time scale alignment is performed to obtain the temporal and spatial correspondence between real-time error data and processing data, as shown in the following formula:

[0075] (20)

[0076] Where, x a (t), y a (t) and z a (t) represent the real-time coordinates in the three-dimensional space of the corresponding machine tool coordinate system, X DT (t), Z DT (t) and C DT (t) represents the real-time coordinates of each moving component of the digital twin system in the spatial coordinate system of the digital twin system;

[0077] The DT-LSTM system maps the twin data to the spatial coordinate system of the digital twin system, obtaining the three-dimensional spatial error data corresponding to the virtual X, Y, and Z axes of the machine tool in standby state, as shown in the following formula; at this time, the machine tool spatial error data is in length units, and the longitudinal and lateral comparisons of the spatiotemporal dimensions are performed under the condition that the units are consistent, and the digital twin system built based on physical components makes each twin data have significant physical meaning;

[0078] (twenty one)

[0079] Where, x e (t), y e (t) and z e (t) represents the real-time error in the three-dimensional space of the workpiece in the machine tool spatial coordinate system, x g (t), y g (t) and z g (t) represents the ideal position coordinates of the workpiece in the three-dimensional space of the machine tool coordinate system. Based on this, by comparing the error data of the X, Y, and Z axes before and after machining, the error data of each axis is compared with the machining load to select the best machining parameters and control parameters. The DT-LSTM system can also clearly and intuitively explain the source of the fault diagnosis data corresponding to the real-time error of the workpiece machining and the motion state of the corresponding physical components.

[0080] S7: A method for real-time extraction and visualization of machining errors driven by DT-LSTM system;

[0081] This invention establishes a DT-LSTM model that can map real-time error data into three-dimensional space, obtaining the spatial distribution of the error data. The DT-LSTM model proposed in this invention not only determines the correspondence between machining data in time and space, but also establishes the physical correspondence between workpiece accuracy and machine tool axis system, and can be intuitively visualized in three-dimensional space.

[0082] In ultra-precision machining, the surface morphology and shape errors of workpieces are mostly at the submicron scale. The digital twin machining system established based on this invention extracts the corresponding error data and maps it into three-dimensional space through real-time monitoring methods, thereby intuitively obtaining the spatial distribution of the corresponding machining errors, and at the same time obtaining the amplitude of the overall error data of the workpiece and the corresponding spatial distribution state.

[0083] The DT-LSTM system maps real-time workpiece machining data to the spatial coordinate system of a digital twin system, obtaining the spatial distribution of error data. The error data in the digital twin system's spatial coordinate system is then dimensionality-reduced and projected onto the xOz, xOy, and yOz planes of the digital twin system's spatial coordinate system using three views. The xOz and yOz planes represent the composite motion forms of the x-axis and y-axis with the z-axis, respectively, while the xOy plane represents the composite motion form of the x-axis and y-axis. Based on this, the two-dimensional spatial data is further reduced to three-axis motion forms represented by x, y, and z. The corresponding mapping relationship between the physical motion axes X, Z, and C in the machine tool's digital twin system and the virtual spatial axes x, y, and z is established, as shown in the following equation:

[0084] (twenty two)

[0085] According to equations (3) to (22), the established DT-LSTM system can collect machining process parameters in real time and perform real-time analysis and prediction of machining errors. The relevant accuracy can be uniquely and accurately monitored by the DT-LSTM established in this invention. The spatial twin data corresponding to the digital twin system is shown in the above equations, avoiding measurement errors and time losses introduced by offline measurement. Especially in the measurement process of complex curved surface parts, due to its three major advantages—online measurement, measurement data under the unified machine tool coordinate system, and data directly from high-precision grating rulers—real-time measurement accuracy and measurement efficiency are significantly improved.

[0086] The beneficial effects of this invention: Addressing the challenges of resource waste, prolonged R&D cycles, and high processing costs associated with traditional processing and inspection methods involving machining, measurement, and feedback, this invention proposes a processing status monitoring and real-time error reconstruction visualization method driven by a combination of digital twins and neural networks. To meet the needs of real-time monitoring and error control in the processing process, processing data is input into a data twin model. Based on the physical structure of the real-time data, it is classified and spatial dimension features are extracted. Then, the sequence of real-time twin data with spatial dimension features is passed to an LSTM layer to extract temporal dimension features. The LSTM utilizes its ability to remember important historical information to generate the predictive output of the twin system, ultimately achieving real-time monitoring of the manufacturing system's processing process and realizing accurate real-time evaluation of the workpiece processing process and quality. This method overcomes the shortcomings and deficiencies of existing digital twin technologies, such as poor real-time performance, insufficient prediction accuracy, and lack of interpretability of prediction results. It provides a guarantee for improving workpiece processing accuracy and efficiency and lays the foundation for precise control of the manufacturing system. Attached Figure Description

[0087] Figure 1 This is a flowchart of the processing status monitoring and real-time error reconstruction visualization method of the present invention.

[0088] Figure 2 This is a diagram of an LSTM neural network structure based on a digital twin architecture.

[0089] Figure 3 The DT-LSTM principle is compared with the traditional digital twin machining system. (a) is the machining process monitoring principle based on the conventional digital twin system; (b) is the machining quality prediction and control principle based on DT-LSTM.

[0090] Figure 4 It is the real-time machining data of the cylindrical workpiece within a single time step; where (a) is the change of X, Y, Z axis data and corresponding error data during the machining process; (b) is the X axis data and corresponding error data; (c) is the Y axis data and corresponding error data; and (d) is the Z axis data and corresponding error data.

[0091] Figure 5 It is the spatiotemporal scale alignment of the real-time machining data of the workpiece and the corresponding real-time error data within a single time step; where (a) is the real-time machining data of the workpiece within a single time step; and (b) is the real-time error data of the workpiece within a single time step.

[0092] Figure 6This refers to the real-time twin data changes of the machine tool under different states; among them, (a) is the error signal of the machine tool moving parts in the standby state; (b) is the corresponding three-dimensional spatial error signal of the machine tool moving parts in the standby state; (c) is the error signal of the machine tool moving parts in the machining state; (d) is the corresponding three-dimensional spatial error signal of the machine tool moving parts in the machining state; (e) is a comparison of the error signals of the machine tool moving parts before and after time t1 in the machining state; (f) is a comparison of the corresponding three-dimensional spatial error signals of the machine tool moving parts before and after time t1 in the machining state; and (g) is a comparison of the machine tool Z-axis error signals before and after time t1 in the machining state.

[0093] Figure 7 It is the real-time twin error data of the workpiece and its mapping relationship in three-dimensional space; where (a) is the spatiotemporal characteristics of the real-time twin error data; and (b) is the real-time error data of the virtual axis of the workpiece in the time dimension.

[0094] Figure 8 This describes the prediction and comparison verification of multi-axis twin data using the DT-LSTM architecture at different time steps. Specifically, (a) compares the prediction results of the multi-axis twin data with real-time data at the first time step; (b) compares the prediction results of the multi-axis twin data with real-time data at the eleventh time step; (c) compares the prediction results of the multi-axis twin data with real-time data at the twenty-first time step; (d) compares the prediction results of the multi-axis twin data with real-time data at the thirty-first time step; (e) compares the prediction results of the multi-axis twin data with real-time data at the forty-first time step; and (f) compares the prediction results of the multi-axis twin data with real-time data at the fifty-first time step.

[0095] Figure 9 This section describes the prediction and reconstruction of workpiece error data in the spatial dimension using DT-LSTM at different time steps. Among them, (a) is the prediction and reconstruction of roundness of workpiece machining data in the xOy plane in the first three time steps; (b) is the prediction and reconstruction of cylindricity of workpiece machining data in the spatial dimension in the first three time steps; (c) is the prediction and reconstruction of workpiece error data in the spatial dimension using DT-LSTM during workpiece machining; and (d) is the comparison and verification of predicted error data and actual error data in the spatial dimension using DT-LSTM during workpiece machining. Detailed Implementation

[0096] The specific embodiments of the present invention are described in detail below with reference to the technical solutions and accompanying drawings.

[0097] like Figure 1 The flowchart shown illustrates a method for monitoring processing status and visualizing real-time errors through the collaborative use of digital twins and neural networks. The specific steps are as follows:

[0098] S1: Real-time construction of digital twin (DT) systems;

[0099] During the machining process, all material removal is generated under the high-precision servo motion of various moving parts of the machine tool. A digital twin system corresponding to the real-time machining process is established by collecting the input and output variables of the control process of each moving part. This invention takes an ultra-precision grinding machine as an example. Milling machines, lathes, and other commonly used multi-axis machine tool structures differ from grinding machines only in the number and spatial position of moving parts, which can be used as a basis for further research. If the grinding axis is replaced with a milling axis, it becomes the digital twin system for an ultra-precision milling machine. Simplifying the grinding axis to a fixed tool, it can be considered the digital twin system for an ultra-precision lathe. The manufacturing system completes the machining of the workpiece through the interaction between the tool and the workpiece, and the machining state is directly reflected in the "tool-workpiece" subsystem. The "tool-workpiece" subsystem under the digital twin system is constructed; the tools are grinding wheels, milling cutters, and turning tools, and the corresponding tool system is shown in the following formula:

[0100] (twenty three)

[0101] Where G represents the grinding system, M represents the milling system, and L represents the turning system. For example... Figure 2 As shown, the physical structure correspondence of "machine tool-tool-workpiece" is constructed as follows: the tool is connected to the tool spindle, the tool spindle is connected to the machine tool guideway, and the machine tool guideway is connected to the machine tool bed; the workpiece is connected to the workpiece spindle, the workpiece spindle is connected to the machine tool guideway, and the machine tool guideway is connected to the machine tool bed; the corresponding digital twin system model of the CNC machine tool (CNCMT) can be represented as:

[0102] (twenty four)

[0103] S2: Long Short-Term Memory (LSTM) Neural Network Model Oriented to Process Flow;

[0104] A qualified product is formed from raw materials through a series of processes, steps, and working strokes. Historical data during the processing will affect the current processing status of the workpiece. Processing history information helps to achieve fault diagnosis and improve diagnostic accuracy.

[0105] A fault diagnosis method for machining processes driven by digital twins and deep learning is proposed to realize fault diagnosis of workpiece machining processes. By simplifying the physical structure of CNC grinding machines with digital twin systems, machining process diagnosis based on historical machining data before time t and the current machining state at time t is realized through a long short-term memory neural network model (LSTM).

[0106] Based on the simplified physical structure of the CNC grinding machine according to the digital twin system, as shown in Equation (24), the real-time machining data of the physical structure of the CNC grinding machine is extracted through the digital twin system and used as the real-time input data s(t) of the long short-term memory neural network model:

[0107] (25)

[0108] Where D(t) is the command position corresponding to the digital twin system, δ(t) is the real-time error corresponding to the digital twin system, i(t) is the real-time current corresponding to the digital twin system, and P(t) is the real-time power corresponding to the digital twin system.

[0109] (26)

[0110] In this context, the subscripts X, Z, C, and T represent the physical structure of the CNC grinding machine: the X-axis feed system, the Z-axis feed system, the C-axis rotary system, and the tool system, respectively.

[0111] The corresponding Long Short-Term Memory (LSTM) neural network model under the digital twin system architecture is represented as follows:

[0112] (27)

[0113] Among them, h DT (t-1) represents the fault diagnosis data corresponding to the real-time error output by the DT-LSTM system at the previous moment, and is processed through the forgetting unit f. DT (t) and memory update unit c DT (t) is reflected in real-time fault diagnosis data; forgetting unit f DT (t) is used to determine whether to use historical error data; memory update unit c DT (t) and candidate memory units DT (t) is updated in real time through gate activation functions and state activation functions; output unit i DT (t) determines the amount of information about the current state to retain and outputs it through the output gate. DT (t) Output; σ represents the gate activation function, tanh is the state activation function, and W s V s and b represent the learnable input weights, recurrent weights, and bias of the DT-LSTM system, respectively, and W si W sf W so and W sc V represents the input weights of the input gate, forget gate, output gate, and memory update unit, respectively. si V sf V so and V scThese represent the recurrent weights of the input gate, forget gate, output gate, and memory update unit, respectively. i b f b o and b c These represent the biases of the input gate, forget gate, output gate, and memory update unit, respectively; h DT (t-1) represents the fault diagnosis data corresponding to the real-time error output by the DT-LSTM system at the previous time step:

[0114] (28)

[0115] Where p(t) is the current probability estimate of the component, and R(t) is the failure evolution probability, i.e., the impact of the current failure on the machining accuracy of subsequent workpieces:

[0116] (29)

[0117] And there are:

[0118] (30)

[0119] Where γ(t) represents the error caused by environmental factors.

[0120] Based on this, the computational flow of real-time twin data in the LSTM neural network in the digital twin model was established, and the real-time connection of processing data in different processes, steps, and work journeys based on historical processing information was obtained. This enabled the classification, location, and diagnosis of processing faults, and the predictive maintenance of the processing system.

[0121] S3: DT-LSTM collaborative driven machining error lead control strategy;

[0122] Compared to existing digital twin systems for machining processes, the DT-LSTM architecture established in this invention can effectively regulate the current machining state based on advanced predictive twin data. Traditional digital twin systems can achieve good real-time monitoring of the machining process and detect abnormalities in a timely manner. However, in this case, the twin data ht directly affects the workpiece, leading to a decrease in workpiece quality Q(t). This means the current workpiece is already scrapped, and adjusting the machining control parameters at this point only affects the workpiece quality at time t+1 and beyond, resulting in a significant waste of manpower, materials, and time. This invention, through its DT-LSTM network structure, can accurately predict future machining data h(t). By combining historical data and twin prediction data generated by LSTM, accurate prediction of the machining process can be achieved quickly. The CNC system can adjust control parameters in advance and directly affect the real-time twin data ht, thereby directly improving the machining quality of the workpiece and increasing the workpiece yield.

[0123] The DT-LSTM system constructed in step S2 effectively regulates the current processing state based on the advanced predictive twin data; the historical processing data s(t-1) before time t and the real-time processing data s(t) are transmitted to the digital twin system constructed by equation (2); at time step t, the LSTM network uses the fault diagnosis data h corresponding to the real-time error output by the DT-LSTM system at the previous time step. DT (t-1) and the fault diagnosis data h corresponding to the real-time error output by the DT-LSTM system at the current time of the sequence. DT (t) is used to calculate the output and the updated cell state h. DT (t+1); at the same time, the updated unit state at time step t+1 is transmitted to the digital twin system, and the corresponding predicted processing data s(t+1) is generated; on this basis, the predicted processing data s(t+1) is compared with the real-time processing data s(t) to analyze the relative magnitude of the workpiece error. When the workpiece error exceeds the tolerance range, the control strategy is adopted to adjust the relevant parameters of the current digital twin system and directly improve the processing quality of the workpiece and increase the yield of the workpiece, as shown in equation (30):

[0124] (31)

[0125] S4: A process twin data acquisition method driven by the synergy of digital twins and neural networks (DT-LSTM);

[0126] To verify the DT-LSTM-driven machining status monitoring method proposed in this invention, precision grinding experiments oriented towards the machining process were conducted on a self-developed ultra-precision grinding machine. To reduce shape errors caused by grinding wheel wear, a 1000# metal-bonded diamond grinding wheel was selected. Before the experiment, it was ensured that the grinding wheel was dressed online, and the runout at the far end of the grinding wheel shank did not exceed 3 μm. The selected experimental parameters are shown in Table 1.

[0127] Table 1. Test parameters for DT-LSTM-driven machining status monitoring

[0128]

[0129] After the experiment, the workpiece was cleaned for 5 minutes using an ultrasonic cleaner and pure water to ensure surface cleanliness. After drying, the roundness and cylindricity of the workpiece were inspected using a roundness meter (Random NEX300, ACCRETECH, Tokyo, Japan). During inspection, the workpiece rotated along the axis of the cylindrical surface, and the instrument probe measured the circumferential data of the cylindrical surface. Samples were taken evenly along the machined area of ​​the cylindrical surface, and relevant data were measured at each sampling point; at least five sampling points were used.

[0130] Based on the proposed digital twin-deep learning real-time architecture, the machining process can be monitored while simultaneously performing fault diagnosis and error localization. Through an LSTM neural network, the distribution patterns of machining errors and feature values ​​are quickly derived and compared with the characteristics of the main moving parts of the machine tool for fault diagnosis. Based on the digital twin architecture, the fault diagnosis results can be compared with the physical structure of the main moving parts to achieve error localization, reduce system errors during machining, and ensure stable operation of the product machining process.

[0131] During the workpiece processing, the establishment of the digital twin system is divided into two stages: model building and real-time monitoring. In the model building stage, the selected sampling frequency is greater than 2000Hz; in the real-time monitoring stage, the selected sampling frequency is less than 100Hz.

[0132] Based on the DT-LSTM system, the digital twin system directly acquires real-time machining data from the CNC grinding machine during workpiece processing, ensuring that the data acquisition method corresponds to the machining parameters. Since the real-time machining data is continuous over a long period, it is acquired in segments according to the workpiece machining principle. Each revolution of the C-axis is defined as a machining and acquisition unit for the digital twin system, corresponding to a training unit of the LSTM network. Therefore, the feed rate per revolution of the C-axis is defined as a spatial period, with the corresponding time period dt. h :

[0133] (32)

[0134] The size of a one-dimensional twin of a single moving part within a basic time period can be represented as:

[0135] (33)

[0136] Set multiple complete basic time periods as a single time step:

[0137] (34)

[0138] in Given the number of basic time periods, the total amount of twin data within one time step is:

[0139] (35)

[0140] The data size within a single time step of the DT-LSTM established in this invention can be determined by the number of basic time periods n1 and the sampling frequency F. sAdjustment is key. A higher sampling frequency results in a larger data volume and a longer system response time; a larger number of basic time periods results in a longer single calculation time and higher theoretical prediction accuracy. In a digital twin system, the accuracy of the grating ruler determines the actual machining accuracy of the machine tool, while the data acquisition frequency depends on the system's response time and machining prediction accuracy. A higher sampling frequency leads to a larger data volume, which in turn increases computational cost and system response time. The type of data acquired depends on the detection requirements, such as position data or power data. To balance data volume and computational cost, the data in the digital twin system structure can be dimensionality-reduced according to the above formula. In this experiment, n1=3.

[0141] During the actual workpiece machining stage, due to the presence of workpiece material blank errors, installation errors, machine tool coordinate system alignment, and other factors inherent in the machining preparation stage, the data acquisition parameters can be set to the model building stage. As the pre-machining preparation work progresses, the DT-LSTM model of the machining process can be established simultaneously. When the workpiece begins full machining, the data acquisition parameters can be set to the real-time monitoring stage to achieve the detection and prediction of the machining process.

[0142] To ensure effective monitoring of the machining process by the DT-LSTM, the servo update rate of the PMAC system was set to 2.5 kHz. During the data acquisition phase, the sampling frequency of the grinding force signal was set to 200 Hz. Therefore, the acquisition settings for the DT-LSTM twin data in this example are shown in Table 2.

[0143] Table 2. Data Acquisition Settings for Real-Time Monitoring of the Manufacturing Process Driven by DTLSTM

[0144]

[0145] During machine tool configuration, to fully capture machining information, a DT-LSTM model is constructed by extracting temporal and spatial features. At this stage, the twin data sampling frequency is set to 2258.7 Hz. During product machining, to acquire grinding process signals over a sufficiently long period and improve data processing efficiency, the twin data sampling frequency is set to 75.3 Hz to reduce model processing time and improve the real-time response rate of the twin system. At this point, the machine tool sampling time is 13.28 ms, which meets the real-time processing and monitoring requirements of the twin system.

[0146] There is no significant difference between the model building phase and the real-time monitoring phase; the only difference lies in the purpose of data acquisition. During the actual workpiece machining phase, due to factors such as workpiece material blank errors, installation errors, and machine tool coordinate system alignment, the data acquisition parameters can be set to the model building phase. As the pre-machining preparation work progresses, the DT-LSTM model of the machining process can be built simultaneously. When the workpiece begins full machining, the data acquisition parameters can be set to the real-time monitoring phase to achieve the detection and prediction of the machining process.

[0147] S5: A DT-LSTM collaborative method for classifying and identifying machining errors;

[0148] Based on the physical meaning of the real-time twin data in the DT-LSTM system, the workpiece machining process is evaluated in real time using spatial dimension evaluation indicators within the same spatiotemporal coordinate system. For example, the spacing of the machining trajectory spiral lines corresponds to the feed per revolution, directly affecting the workpiece's wear mark spacing and influencing the residual height on the material surface, thus significantly impacting the workpiece's surface roughness, as shown in the following formula:

[0149] The feed rate (grid spacing) per revolution of the C-axis is expressed as:

[0150] (36)

[0151] Where, n s Where is the spindle speed and F is the feed rate. The surface machining residual height can be calculated as follows:

[0152] (37)

[0153] Among them, R s R is the tool radius. h The surface roughness can be expressed as the residual height of the grinding marks on the workpiece surface:

[0154] (38)

[0155] Where l is the sampling length during the workpiece surface roughness measurement process, and f t (x) represents the actual contour curve function of the workpiece obtained by measurement; the motion error of the C-axis also directly corresponds to the roundness error of the workpiece, as shown in the following formula:

[0156] (39)

[0157] Among them, f max (x) and f min (x) represent the maximum and minimum profile curves of the same radial direction on the workpiece surface, respectively; the cylindricity of the workpiece is expressed as:

[0158] (40)

[0159] Among them, f ⊥ (x) and f ∥ f(x) represents two actual contour curves on the workpiece surface with a spatial angle difference of 90 degrees; f(x) represents the ideal target contour curve function of the workpiece; the Z-axis motion error directly corresponds to the taper of the workpiece, as shown in the following formula:

[0160] (41)

[0161] Among them, f z The taper ratio is represented by f1(x), the diameter of the large end is represented by f2(x), and the measurement length is represented by L.

[0162] This allows for the evaluation of the workpiece's real-time C-axis feed per revolution, surface residual height, surface roughness, roundness, and taper errors. Workpiece errors, such as surface quality and contour errors, are directly related to the real-time motion parameters of the machine tool's moving parts and often involve the measurement of different spatial characteristic parameters. They are significantly affected by measurement errors and often require multiple measurements to obtain the maximum value. The DT-LSTM established by this invention can acquire machining process parameters in real time and perform real-time analysis and prediction. The relevant geometrical accuracy can be uniquely and accurately monitored by the digital twin architecture established by this invention. The real-time spatial twin data corresponding to the digital twin system avoids measurement errors and time losses introduced by offline measurements. Especially in the measurement of complex curved surface parts, its three major advantages—online measurement, measurement data under a unified machine tool coordinate system, and data directly derived from a high-precision grating ruler—significantly improve measurement accuracy and efficiency.

[0163] S6: A real-time monitoring method for the manufacturing process driven by DT-LSTM;

[0164] The DT-LSTM established by this invention achieves a twin mapping between complex three-dimensional spatial data and machine tool single-axis motion data. The DT-LSTM model established by this invention can effectively capture the corresponding twin data during workpiece machining. Figure 4 This refers to real-time machining data within a single time step for the corresponding workpiece. It can be observed that the DT-LSTM model established in this invention can effectively capture the twin data corresponding to the workpiece machining process. [The data is then extracted separately.] Figure 4 The processing data and error data corresponding to points A, B, and C in (a) are used to obtain the error vector of that point, as shown below. Figure 4 As shown in (b), (c), and (d) in the figure. Since the machining scale of machine tools in three-dimensional space is on the order of millimeters, while the error data of machine tools is mostly on the order of sub-micrometers, the error data cannot be well reflected in the machining data. Based on this, the spatial dimension error data and position data of each axis are extracted respectively, and time scale alignment is performed to obtain the time correspondence between the real-time error data and the machining data.

[0165] In the spatial coordinate system of the digital twin system, spatial dimension error data and position data of the X, Z, and C axes are extracted respectively, and time scale alignment is performed to obtain the temporal and spatial correspondence between real-time error data and processing data, as shown in the following formula:

[0166] (42)

[0167] Where, x a (t), y a (t) and z a (t) represent the real-time coordinates in the three-dimensional space of the corresponding machine tool coordinate system, X DT (t), Z DT (t) and C DT (t) represents the real-time coordinates of each moving component of the digital twin system in the spatial coordinate system of the digital twin system;

[0168] The DT-LSTM system maps the twin data to the spatial coordinate system of the digital twin system, obtaining the three-dimensional spatial error data corresponding to the virtual X, Y, and Z axes of the machine tool in standby state, as shown in the following formula; at this time, the machine tool spatial error data is in length units, and the longitudinal and lateral comparisons of the spatiotemporal dimensions are performed under the condition that the units are consistent, and the digital twin system built based on physical components makes each twin data have significant physical meaning;

[0169] (43)

[0170] Where, x e (t), y e (t) and z e (t) represents the real-time error in the three-dimensional space of the workpiece in the machine tool spatial coordinate system, x g (t), y g (t) and z g (t) represents the ideal position coordinates of the workpiece in the three-dimensional space of the machine tool coordinate system. Based on this, by comparing the error data of the X, Y, and Z axes before and after machining, the error data of each axis is compared with the machining load to select the best machining parameters and control parameters. The DT-LSTM system can also clearly and intuitively explain the source of the fault diagnosis data corresponding to the real-time error of the workpiece machining and the motion state of the corresponding physical components.

[0171] Based on this, the spatial dimension error data and position data of each axis are extracted and aligned on a time scale to obtain the real-time machining data and corresponding real-time error data of the workpiece, as shown below. Figure 5As shown, a clear temporal correspondence exists between the real-time error data and the machining data. Simultaneously, the error data for the X, Y, and Z axes exhibit significant spatial fluctuations. Specifically, the maximum errors for the X and Y axes occur at the zero points of their corresponding real-time positions, such as t5 and t6 for the X-axis and t1 and t2 for the Y-axis. This is due to the correlation between the X and Y axis data and the grinding axis. Similarly, since the Z-axis is directly related to the workpiece axis, the error fluctuation range of the Z-axis often occurs in the contact area between the workpiece and the grinding wheel. In this example, it occurs in the positive value regions of the X and Y axes, such as t3 and t4. Therefore, it is reasonable to believe that the t3 and t4 regions precisely correspond to the machining times.

[0172] During the machining process of the same workpiece, its accuracy continuously improves with the execution of processes such as roughing, semi-finishing, and finishing. The ideal shape of the workpiece corresponding to different processes has spatial and temporal consistency, while the workpiece error varies under different processes. Therefore, real-time error data of the workpiece is more representative of the machining status than real-time position data. Figure 6 This describes the changes in the corresponding twin data within the digital twin architecture under different machine tool states. Figure 6 (a) in the diagram represents the error signal of the machine tool's moving parts in standby mode. It can be seen that in standby mode, due to the absence of external load, the errors of each moving part of the machine tool are extremely small. The error data for the X and Z axes fluctuate between ±0.015, while the error for the C axis, since it does not rotate, is close to 0. Because the X and Z axes are linear motion parts, the error data is in units of length, while the C axis is a rotary motion part, the error data is in units of angle. Therefore, the error data of the three cannot be directly compared.

[0173] The DT-LSTM architecture established according to this invention maps twin data to three-dimensional space, obtaining three-dimensional spatial error data corresponding to the virtual X, Y, and Z axes of the machine tool in standby mode, such as... Figure 6 As shown in (b) above. At this time, the machine tool spatial error data is in length units. Under the condition that the units are consistent, longitudinal and transverse comparisons in the spatiotemporal dimensions can be performed. Since the machine tool is in a static load state at this time, the errors of the X and Y axes are relatively small, reflecting that the standby error of the ultra-precision grinding machine does not exceed ±3 nm, and is mostly caused by external noise interference and random data fluctuations.

[0174] Error signals of moving parts of machine tools during machining, such as Figure 6As shown in (c), since the X, Z, and C axes of the machine tool have all moved at this time, the error of the moving parts of the machine tool in the machining state is significantly higher than that in the standby state. At time t1, the error signals of each moving part change abruptly, and after time t1, the error signals of each axis all show a significant increase. Based on this, it can be determined that time t1 is the starting point when the workpiece contacts the tool and machining begins. Before machining, in the no-load state, the X-axis error data fluctuates between ±0.15, followed by the Z-axis error data fluctuating between ±0.05, while the C-axis error is the smallest, fluctuating between ±0.02. Figure 6 As shown in (e) in the figure. It was also found that when the workpiece came into contact with the grinding wheel after machining began, the machine tool's X-axis error signal increased to fluctuate to ±0.35, the Z-axis error data fluctuated between ±0.2, while the C-axis error signal did not change much.

[0175] Three-dimensional spatial error data generated by the three axes of the machine tool during machining, such as Figure 6 As shown in (d), the three-axis error data have good comparability in the same spatial dimension, and the digital twin system established by this invention based on physical components gives each twin data a significant physical meaning. Firstly, the XY axes have larger errors due to the dynamic load of the high-speed rotation of the grinding axis, while the Z-axis dynamic load is mainly the low-speed rotation of the workpiece spindle. Therefore, the error is smaller, especially around time t1. When the tool and workpiece come into contact at time t1, there is a material removal effect, which is further applied as a dynamic load to the moving parts of the machine tool. It can be seen that because the initial dynamic load of the X and Y axes is large, the change in machining load is not obvious on the curve, as shown in Figure 1. Figure 6 As shown in (f). However, it can be judged by extracting the average value of the error data. The Z-axis error is relatively small, so the error increases significantly around time t1. When the tool and workpiece are not in contact, the Z-axis error signal fluctuates between ±10nm. When entering the machining state, the Z-axis error signal fluctuates between ±50nm, and the dynamic load increases significantly, as shown in (f). Figure 6 As shown in (g) in the diagram.

[0176] comprehensive Figure 6 As shown in (c) and (d), the grinding axis load is significantly higher than that of the workpiece spindle, which leads to significantly higher X and Y axis errors than Z axis errors during workpiece machining in three-dimensional space. Furthermore, by comparing the error data of each axis before and after machining, the error data of each axis can be compared with the machining load, which is beneficial for selecting the optimal machining and control parameters. For example, in this machining example, the X and Z axis dynamic loads > machining load > Z axis dynamic load, which corresponds one-to-one with the physical structure of the CNC grinding machine. Therefore, the DT-LSTM model established in this invention can clearly and intuitively explain the source of each twin signal and the motion state of the corresponding physical components. These functions are not available in previous conventional data twin structures.

[0177] S7: A method for real-time extraction and visualization of machining errors driven by DT-LSTM;

[0178] Although Figure 6 This reflects the distribution of three-dimensional spatial error data in the workpiece machining process over a time scale, but it cannot directly provide the workpiece's precision characteristics. Therefore, the DT-LSTM model established according to this invention can map real-time error data into three-dimensional space, obtaining the spatial distribution of the error data, such as... Figure 7 As shown, the error data of the x-axis and y-axis exhibit similarity. This is because the X and Y axes are virtual axes in DT (Data Transformer), and their error data is calculated through the machine tool's three physical motion axes X, Z, and C. Therefore, the errors of the X and Y axes originate from the machine tool's X and C axes. The Z-axis error data, however, directly corresponds to the machine tool's Z-axis. Thus, the DT-LSTM model proposed in this invention not only determines the temporal and spatial correspondence of machining data but also establishes the physical correspondence between workpiece accuracy and the machine tool's axis system, and can be intuitively visualized in three-dimensional space.

[0179] The DT-LSTM system maps real-time workpiece machining data to the spatial coordinate system of a digital twin system, obtaining the spatial distribution of error data. The error data in the digital twin system's spatial coordinate system is then dimensionality-reduced and projected onto the xOz, xOy, and yOz planes of the digital twin system's spatial coordinate system using three views. The xOz and yOz planes represent the composite motion forms of the x-axis and y-axis with the z-axis, respectively, while the xOy plane represents the composite motion form of the x-axis and y-axis. Based on this, the two-dimensional spatial data is further reduced to three-axis motion forms represented by x, y, and z. The corresponding mapping relationship between the physical motion axes X, Z, and C in the machine tool's digital twin system and the virtual spatial axes x, y, and z is established, as shown in the following equation:

[0180] (44)

[0181] Predicted and measured processing data were selected at time steps 1, 11, 21, ..., 51, respectively, as follows: Figure 8 As shown, the predicted processing data and the measured data have good consistency, indicating that the DT-LSTM structure established in this invention can effectively monitor and predict the processing process in real time. Based on this, the predicted and measured processing data are mapped into a three-dimensional space to achieve real-time visualization of the workpiece processing process, as shown below. Figure 9 As shown. In ultra-precision machining, the surface morphology and shape errors of workpieces are mostly at the sub-micron scale, which is... Figure 4-8 This cannot be intuitively demonstrated in the text. The digital twin processing system established based on this invention can solve this problem through real-time monitoring of twin data. (The text then abruptly shifts to a different topic: extracting...) Figure 8The error data corresponding to the prediction step is mapped to a three-dimensional space, such as... Figure 9 As shown in (a), the error data exhibits a fluctuating pattern along the radial direction, resembling a typical petal pattern. Mapping the corresponding spatial error data to a two-dimensional space, as shown... Figure 9 As shown in (b) above, the spatial distribution of the corresponding machining errors can be visually obtained. Based on this, the complete error data of the workpiece is mapped to three-dimensional space, as shown below. Figure 9 As shown in (c) above, the magnitude of the overall error data of the workpiece and its corresponding spatial distribution can be obtained. The predicted error value is compared with the measured value, as shown below. Figure 9 As shown in (d), the results show that the DT-LSTM machining process real-time monitoring and real-time error reconstruction visualization method proposed in this invention can accurately and intuitively monitor and quickly predict the machining process in real time, while displaying the real-time error of the workpiece in the form of spatial visualization.

[0182] This invention presents a method for monitoring and visualizing real-time errors in manufacturing processes through the collaborative drive of digital twins and neural networks. Addressing the challenges of traditional processing and inspection methods involving machining, measurement, and feedback, which result in resource waste, extended development cycles, and high processing costs, this method inputs processing data into a data twin model. The data is then classified based on its physical structure, and spatial dimension features are extracted. The sequence of real-time twin data with spatial dimension features is then passed to an LSTM layer to extract temporal dimension features. The LSTM leverages its ability to remember important historical information to generate the predictive output of the twin system, ultimately achieving real-time monitoring of the manufacturing system's processing and enabling accurate real-time evaluation of the workpiece processing process and quality. This method overcomes the shortcomings of existing digital twin technologies, such as poor real-time performance, insufficient prediction accuracy, and lack of interpretability of prediction results. It provides a guarantee for improving workpiece processing accuracy and efficiency and lays the foundation for precise control of manufacturing systems.

Claims

1. A method for monitoring processing status and visualizing real-time error reconstruction, characterized in that, The steps are as follows: S1: Real-time construction of digital twin systems oriented towards the manufacturing process; The manufacturing system completes the machining of the workpiece through the interaction between the cutting tool and the workpiece, and the machining state is directly reflected in the "cutting tool-workpiece" subsystem. A "cutting tool-workpiece" subsystem is constructed under a digital twin system; the cutting tools are grinding wheels, milling cutters, and turning tools, and the corresponding cutting tool system is shown in the following equation: ； Where G represents the grinding system, M represents the milling system, and L represents the turning system; The physical structure correspondence of "machine tool-tool-workpiece" is constructed as follows: the tool is connected to the tool spindle, the tool spindle is connected to the machine tool guideway, and the machine tool guideway is connected to the machine tool bed; the workpiece is connected to the workpiece spindle, the workpiece spindle is connected to the machine tool guideway, and the machine tool guideway is connected to the machine tool bed; the machine tool is a CNC grinding machine, and the digital twin system of the physical structure of the CNC grinding machine is represented as follows: ； Where X represents the X-axis feed system of the CNC grinding machine, Z represents the Z-axis feed system of the CNC grinding machine, C represents the C-axis rotation system of the CNC grinding machine, and T represents the tool system of the CNC grinding machine. S2: Construct a long short-term memory neural network model oriented towards the processing procedure; A fault diagnosis method for machining process driven by digital twin and deep learning is proposed to realize fault diagnosis of workpiece machining process; by simplifying the physical structure of CNC grinding machine by digital twin system, machining process diagnosis based on historical machining data before time t and current machining state at time t is realized by long short-term memory neural network model; Based on the simplified physical structure of the CNC grinding machine according to the digital twin system, as shown in Equation (2), the real-time machining data of the physical structure of the CNC grinding machine is extracted through the digital twin system and used as the real-time input data s(t) of the long short-term memory neural network model: ； Where D(t) is the command position corresponding to the digital twin system, δ(t) is the real-time error corresponding to the digital twin system, i(t) is the real-time current corresponding to the digital twin system, and P(t) is the real-time power corresponding to the digital twin system. ； In this context, the subscripts X, Z, C, and T represent the physical structure of the CNC grinding machine: the X-axis feed system, the Z-axis feed system, the C-axis rotary system, and the tool system, respectively. The corresponding Long Short-Term Memory (LSTM) neural network model under the digital twin system architecture is represented as follows: ； Among them, h DT (t-1) represents the fault diagnosis data corresponding to the real-time error output by the DT-LSTM system at the previous moment, and is processed through the forgetting unit f. DT (t) and memory update unit c DT (t) is reflected in real-time fault diagnosis data; forgetting unit f DT (t) is used to determine whether to use historical error data; memory update unit c DT (t) and candidate memory units DT (t) is updated in real time through gate activation functions and state activation functions; output unit i DT (t) determines the amount of information about the current state to retain and outputs it through the output gate. DT (t) Output; σ represents the gate activation function, tanh is the state activation function, and W s V s and b represent the learnable input weights, recurrent weights, and bias of the DT-LSTM system, respectively, and W si W sf W so and W sc V represents the input weights of the input gate, forget gate, output gate, and memory update unit, respectively. hi V hf V ho and V hc These represent the recurrent weights of the input gate, forget gate, output gate, and memory update unit, respectively. i b f b o and b c represent the biases of the input gate, forget gate, output gate, and memory update unit, respectively; s(t) is the real-time input data of the long short-term memory neural network model; h DT (t) represents the fault diagnosis data corresponding to the real-time error output by the DT-LSTM system at the current moment: ； Where p(t) is the current probability estimate of the component, and R(t) is the failure evolution probability, i.e., the impact of the current failure on the machining accuracy of subsequent workpieces: ； And there are ； Where γ(t) represents the error caused by each factor; S3: DT-LSTM collaborative driven machining error lead control strategy; The DT-LSTM system constructed in step S2 effectively regulates the current processing state based on the advanced predictive twin data; the historical processing data s(t-1) before time t and the real-time processing data s(t) are transmitted to the digital twin system constructed by equation (2); at time step t, the LSTM network uses the fault diagnosis data h corresponding to the real-time error output by the DT-LSTM system at the previous time step. DT (t-1) and the fault diagnosis data h corresponding to the real-time error output by the DT-LSTM system at the current time of the sequence. DT (t) is used to calculate the output and the updated cell state h. DT (t+1); at the same time, the updated unit state at time step t+1 is transmitted to the digital twin system, and the corresponding predicted processing data s(t+1) is generated; based on this, the predicted processing data s(t+1) is compared with the real-time processing data s(t) to analyze the relative magnitude of the workpiece error. When the workpiece error exceeds the allowable range, a control strategy is adopted to adjust the relevant parameters of the current digital twin system and directly improve the processing quality of the workpiece, thereby increasing the yield of the workpiece, as shown in equation (9): ； S4: A method for acquiring twin data of the manufacturing process in collaboration with a DT-LSTM system; S5: A method for classifying and identifying machining errors driven by a DT-LSTM system; Based on the physical meaning of the real-time twin data in the DT-LSTM system, the workpiece machining process is evaluated in real time using spatial dimension evaluation indicators in the same spatiotemporal coordinate system. In the machining of cylindrical workpieces, the spacing of the spiral lines of the machining trajectory is the feed per revolution of the C-axis, which directly corresponds to the grinding mark spacing of the cylindrical workpiece and affects the residual height of the material surface, greatly influencing the surface roughness of the workpiece, as shown in the following formula: The feed rate per revolution of the C-axis is expressed as: ； Where F is the feed rate; the residual height of the workpiece surface is calculated as follows: ； Among them, R s R is the tool radius. h The surface roughness is expressed as the residual height of the workpiece surface after machining, and is: ； Where l is the sampling length during the workpiece surface roughness measurement process, and f t (x) represents the actual contour curve function of the workpiece obtained by measurement; the motion error of the C-axis also directly corresponds to the roundness error of the workpiece, as shown in the following formula: ； Among them, f max (x) and f min (x) represent the maximum and minimum profile curves of the same radial direction on the workpiece surface, respectively; the cylindricity of the workpiece is expressed as: ； Among them, f ⊥ (x) and f ∥ f(x) represents two actual contour curves on the workpiece surface with a spatial angle difference of 90 degrees; f(x) represents the ideal target contour curve function of the workpiece; the Z-axis motion error directly corresponds to the taper of the workpiece, as shown in the following formula: ； Among them, f z The taper ratio is represented by f1(x), the diameter of the large end of the workpiece is represented by f2(x), and the measurement length is represented by L. This enables the evaluation of the workpiece's real-time C-axis feed per revolution, surface residual height, surface roughness, cylindricity, roundness, and taper error; S6: A real-time monitoring method for the manufacturing process driven by a DT-LSTM system; Under the machine tool mechanical coordinate system, a machine tool spatial coordinate system and a digital twin system spatial coordinate system are constructed respectively, and the coordinate origins and X, Y and Z axes of the two correspond one-to-one; In the spatial coordinate system of the digital twin system, spatial dimension error data and position data of the X, Z, and C axes are extracted respectively, and time scale alignment is performed to obtain the temporal and spatial correspondence between real-time error data and processing data, as shown in the following formula: ； Where, x a (t), y a (t) and z a (t) represent the real-time coordinates in the three-dimensional space of the corresponding machine tool coordinate system, X DT (t), Z DT (t) and C DT (t) represents the real-time coordinates of each moving component of the digital twin system in the spatial coordinate system of the digital twin system; The DT-LSTM system maps the twin data to the spatial coordinate system of the digital twin system, obtaining the three-dimensional spatial error data corresponding to the virtual X, Y, and Z axes of the machine tool in standby state, as shown in the following formula; at this time, the machine tool spatial error data is in length units, and the longitudinal and lateral comparisons of the spatiotemporal dimensions are performed under the condition that the units are consistent, and the digital twin system built based on physical components makes each twin data have significant physical meaning; ； Where, x e (t), y e (t) and z e (t) represents the real-time error in the three-dimensional space of the workpiece in the machine tool spatial coordinate system, x g (t), y g (t) and z g (t) represents the ideal position coordinates of the workpiece in the three-dimensional space of the machine tool coordinate system. Based on this, by comparing the error data of the X, Y, and Z axes before and after machining, the error data of each axis is compared with the machining load to select the best machining parameters and control parameters. The DT-LSTM system can also clearly and intuitively explain the source of the fault diagnosis data corresponding to the real-time error of the workpiece machining and the motion state of the corresponding physical components. S7: A method for real-time extraction and visualization of machining errors driven by DT-LSTM system; The DT-LSTM system maps real-time workpiece machining data to the spatial coordinate system of a digital twin system, obtaining the spatial distribution of error data. The error data in the digital twin system's spatial coordinate system is then dimensionality-reduced and projected onto the xOz, xOy, and yOz planes of the digital twin system's spatial coordinate system using three views. The xOz and yOz planes represent the composite motion forms of the x-axis and y-axis with the z-axis, respectively, while the xOy plane represents the composite motion form of the x-axis and y-axis. Based on this, the two-dimensional spatial data is further reduced to three-axis motion forms represented by x, y, and z. The corresponding mapping relationship between the physical motion axes X, Z, and C in the machine tool's digital twin system and the virtual spatial axes x, y, and z is established, as shown in the following equation: 。 2. The processing status monitoring and real-time error reconstruction visualization method according to claim 1, characterized in that, S4: The specific method for acquiring twin data of the manufacturing process driven by the DT-LSTM system is as follows: During the workpiece processing, the establishment of the digital twin system is divided into two stages: model building and real-time monitoring. In the model building stage, the selected sampling frequency is greater than 2000Hz; in the real-time monitoring stage, the selected sampling frequency is less than 100Hz. Based on the DT-LSTM system, the digital twin system directly acquires real-time machining data from the CNC grinding machine during workpiece processing, ensuring that the data acquisition method corresponds to the machining parameters. Since the real-time machining data is continuous over a long period, it is acquired in segments according to the workpiece machining principle. Each revolution of the C-axis is defined as a machining and acquisition unit for the digital twin system, corresponding to a training unit of the LSTM network. Therefore, the feed rate per revolution of the C-axis is defined as a spatial period, with the corresponding time period dt. h : ； Where, n s Main spindle speed; one-dimensional twin data size N of a single moving part within one time period. s Represented as: ； Among them, F s The sampling frequency is set; multiple time periods are set as one time step: ； Where n1 is the number of time periods, the total amount of twin data within one time step is expressed as: ； Where m is the number of corresponding moving parts in the DT-LSTM system, and n is the type of real-time data for a single moving part.

Images